Transformer-based predictive model for advance error probability assessment system for remote services of mechanical parking facilities

Abstract

A system for pre-determining error probability using a transformer-based prediction model for remote service of mechanical parking facilities is provided, comprising a user terminal that provides facility data of mechanical parking facilities, a collection unit that collects facility data from the user terminal, a feature unit that generates a feature vector in which the facility data is arranged in a time series after normalizing the facility data, an inference unit that inputs the feature vector into a pre-established prediction model, and a predictive maintenance server that performs a risk diagnosis based on the predicted result when the prediction model predicts a risk.

Description

Technology Field

[0001] The present invention relates to a system for pre-determining the probability of an error using a transformer-based prediction model for remote services of mechanical parking facilities. It provides a system that probabilistically determines the probability of an error occurring in advance using a transformer-based prediction model trained to recognize abnormal signs in advance by reflecting the relative state differences and temporal change patterns between mechanical parking facilities. Background Technology

[0002] Mechanical parking systems are becoming widespread in various forms to efficiently utilize urban parking spaces; however, due to their structural characteristics involving the complex interoperability of machinery, sensors, and control devices, failures have a significant impact on user safety and operational efficiency. Currently, maintenance methods primarily rely on reactive responses based on threshold exceedances or simple alarm logs. However, this approach has limitations, as it fails to adequately reflect differences in equipment characteristics and the diversity of operating environments, leading to frequent false positives and false negatives. In particular, since normal operation patterns vary even within the same type of equipment depending on installation location, usage frequency, time-based load, and operating conditions, it is difficult to reliably assess the likelihood of failure using only a single standard or judgment methods based on absolute values.

[0003] At this time, methods for managing mechanical parking lots in an integrated manner or performing predictive maintenance of facilities have been researched and developed. In this regard, prior art Korean Published Patent No. 2022-0028685 (published March 8, 2022) and Korean Published Patent No. 2025-0178342 (published December 29, 2025) disclose, respectively, a configuration for predicting the possibility of failure through rule-based analysis or a database-based prediction model based on sensor information and past failure history to manage safety diagnosis, failure prediction, and parking reservation for mechanical parking lots in an integrated manner, and a configuration for preprocessing time-series data including temperature, vibration, pressure, and current collected from precision machining equipment at an edge device and predicting the possibility of failure of precision machining equipment using a machine learning model or a deep learning model.

[0004] However, regarding the former, while the focus is on monitoring the current status of parking facilities and improving management functions, predictive models that learn long-term patterns from time-series data or the relative normal state between facilities are not disclosed. Regarding the latter, although the configurations for predictive maintenance and optimal processing condition settings are disclosed, along with absolute sensor value analysis and retrospective anomaly detection at the facility level, they do not disclose configurations for relative comparison between facilities, cluster-based baseline setting, and learning of temporal deviation structures. These are necessary to resolve the false positive and false negative issues inherent in absolute-standard-based predictions caused by variations between facilities and environments, and to enable reliable failure prediction in actual operating environments. Accordingly, research and development of a system for an intelligent analysis framework capable of recognizing anomaly signs in advance by comprehensively considering operational data from multiple mechanical parking facilities and reflecting relative state differences and temporal change patterns among them are required. The problem to be solved

[0005] One embodiment of the present invention can provide a system for pre-determining error probability using a transformer-based prediction model for remote services of mechanical parking facilities, which can determine the possibility of error in advance through the collection and learning of facility data of mechanical parking facilities and provide a minimum time-based recommendation for the order of entry and exit according to customer requests; collect, normalize, and group facility data of mechanical parking facilities to form a peer baseline, and train a prediction model based on the peer baseline, wherein the model automatically learns deviation structures with different time zones or operating conditions by training the model through self-directed learning and contrastive learning regarding how similar and different the peer baseline and the facility data are; calculate a risk score after inputting facility data from a user terminal into the completed prediction model; and sound an alarm and recommend diagnosis and measures based on the risk score. However, the technical problem that this embodiment aims to solve is not limited to the technical problem described above, and other technical problems may exist. means of solving the problem

[0006] As a technical means for achieving the aforementioned technical task, one embodiment of the present invention includes a predictive maintenance server comprising: a user terminal providing equipment data of a mechanical parking facility; a collection unit collecting equipment data from the user terminal; a feature unit generating a feature vector in which the equipment data is arranged in a time series after normalizing the equipment data; an inference unit inputting the feature vector into a pre-established prediction model; and a diagnosis unit performing a risk diagnosis based on the prediction result when the prediction model predicts a risk. Effects of the invention

[0007] According to any one of the means for solving the problem of the present invention described above, by analyzing equipment data collected from mechanical parking facilities in combination with peer baselines between facilities, problems of false positives and false negatives caused by absolute value deviations due to unique characteristics of each facility or differences in installation environments can be effectively reduced. Furthermore, by learning the temporal change structure of relative deviations by referencing peer group-based peer baselines, it is possible to detect subtle abnormal signs in advance at the pre-failure stage, rather than limiting detection to post-failure detection after the occurrence of a failure. Additionally, it is possible to flexibly respond to various operating conditions and changes in time zones through self-directed learning and comparative learning utilizing unlabeled actual equipment data. Since stable error probability prediction is possible without separate rule settings or threshold adjustments for individual mechanical parking facilities, maintenance costs and the risk of operational interruption can be reduced, and reliable preventive maintenance and safety enhancement in a remote management environment can be achieved simultaneously. Brief explanation of the drawing

[0008] FIG. 1 is a diagram illustrating an error probability pre-determination system using a transformer-based prediction model for remote service of mechanical parking facilities according to an embodiment of the present invention. Figure 2 is a block diagram illustrating a predictive maintenance server included in the system of Figure 1. FIGS. 3 and 4 are drawings for illustrating an embodiment in which an error possibility pre-determination solution according to an embodiment of the present invention is implemented. FIG. 5 is an operation flowchart illustrating a method for providing a solution for pre-determining the possibility of an error according to an embodiment of the present invention. Specific details for implementing the invention

[0009] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0010] Throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are "directly connected" but also cases where they are "electrically connected" with other elements interposed between them. Furthermore, when a part is described as "including" a component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components, and it should be understood that this does not preclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0011] Terms such as “about,” “substantially,” etc., used throughout the specification, are used to mean at or near the stated value when inherent manufacturing and material tolerances are presented in the stated meaning, and are used to prevent unscrupulous infringers from unfairly exploiting the disclosure in which precise or absolute values ​​are mentioned to aid in understanding the invention. Terms such as “step” or “step of” used throughout the specification of the invention do not mean “step for”.

[0012] In this specification, the term "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Additionally, one unit may be realized using two or more pieces of hardware, and two or more units may be realized by one piece of hardware. Meanwhile, "part" is not limited to software or hardware, and "part" may be configured to reside in an addressable storage medium or configured to run on one or more processors. Accordingly, as an example, "part" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts." In addition, the components and '~parts' may be implemented to play one or more CPUs within the device or secure multimedia card.

[0013] Some of the operations or functions described herein as being performed by a terminal, device, or device may instead be performed by a server connected to said terminal, device, or device. Likewise, some of the operations or functions described as being performed by a server may also be performed by a terminal, device, or device connected to said server.

[0014] In this specification, some of the operations or functions described as mapping or matching with a terminal may be interpreted as meaning mapping or matching the terminal's unique number or personal identification information, which is the terminal's identifying data.

[0015] The present invention will be described in detail below with reference to the attached drawings.

[0016] FIG. 1 is a diagram illustrating an error probability pre-determination system using a transformer-based prediction model for remote service of mechanical parking facilities according to an embodiment of the present invention. Referring to FIG. 1, the error probability pre-determination system (1) using a transformer-based prediction model for remote service of mechanical parking facilities may include at least one user terminal (100), a predictive maintenance server (300), and at least one information provision server (400). However, since the error probability pre-determination system (1) using a transformer-based prediction model for remote service of mechanical parking facilities of FIG. 1 is merely an embodiment of the present invention, the present invention is not to be interpreted as being limited through FIG. 1.

[0017] At this time, each component of FIG. 1 is generally connected through a network (Network, 200). For example, as shown in FIG. 1, at least one user terminal (100) can be connected to a predictive maintenance server (300) through the network (200). And, the predictive maintenance server (300) can be connected to at least one user terminal (100) and at least one information providing server (400) through the network (200). Also, at least one information providing server (400) can be connected to the predictive maintenance server (300) through the network (200).

[0018] Here, a network refers to a connection structure capable of exchanging information among individual nodes, such as multiple terminals and servers. Examples of such networks include Local Area Networks (LANs), Wide Area Networks (WANs), the World Wide Web (WWW), wired and wireless data networks, telephone networks, and wired and wireless television networks. Examples of wireless data communication networks include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), 5G NR (New Radio), 6G (6th Generation of Cellular Networks), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth network, NFC (Near-Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc.

[0019] In the following, the term "at least one" is defined as a term including both singular and plural forms, and it will be obvious that even if the term "at least one" does not exist, each component may exist in a singular or plural form and may mean singular or plural. Furthermore, whether each component is provided in a singular or plural form may be changed according to the embodiment.

[0020] At least one user terminal (100) may be a terminal of a user who is an operator, manager, or person in charge of a mechanical parking facility and receives fault prediction by inputting facility data using a web page, app page, program, or application related to an error possibility pre-determination solution.

[0021] Here, at least one user terminal (100) may be implemented as a computer capable of connecting to a remote server or terminal via a network. Here, the computer may include, for example, a navigation system, a laptop equipped with a web browser, a desktop, a laptop, etc. At this time, at least one user terminal (100) may be implemented as a terminal capable of connecting to a remote server or terminal via a network. At least one user terminal (100) may include all kinds of handheld-based wireless communication devices, such as navigation, PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminal, smartphone, smartpad, tablet PC, etc.

[0022] The predictive maintenance server (300) may be a server that provides a solution web page, app page, program, or application for pre-determining the possibility of an error. Additionally, the predictive maintenance server (300) may be a server that, after modeling a prediction model, predicts a failure based on equipment data input from a user terminal (100) and provides the prediction result to the user terminal (100). Here, the predictive maintenance server (300) may be implemented as a computer capable of connecting to a remote server or terminal via a network. Here, the computer may include, for example, a laptop, desktop, or laptop equipped with a navigation system or a web browser.

[0023] At least one information providing server (400) may be a server that provides facility data of a mechanical parking facility, with or without using a web page, app page, program, or application related to an error possibility pre-determination solution. Here, at least one information providing server (400) may be implemented as a computer capable of connecting to a remote server or terminal via a network. Here, the computer may include, for example, a navigation system, a laptop, a desktop, a laptop equipped with a web browser.

[0024] FIG. 2 is a block diagram for explaining a predictive maintenance server included in the system of FIG. 1, and FIG. 3 and FIG. 4 are drawings for explaining an embodiment in which an error possibility pre-determination solution according to an embodiment of the present invention is implemented.

[0025] Referring to FIG. 2, the predictive maintenance server (300) may include a collection unit (310), a characterization unit (320), an inference unit (330), a diagnosis unit (340), a standard generation unit (350), a relative standard learning unit (360), a measure recommendation unit (370), and an improvement unit (380).

[0026] When a predictive maintenance server (300) or another server (not shown) operating in conjunction with an embodiment of the present invention transmits an error possibility pre-determination solution application, program, app page, web page, etc. to at least one user terminal (100), the at least one user terminal (100) may install or open the error possibility pre-determination solution application, program, app page, web page, etc. Additionally, a service program may be executed on at least one user terminal (100) using a script executed in a web browser. Here, a web browser refers to a program that enables the use of web (WWW: World Wide Web) services and receives and displays hypertext described in HTML (Hyper Text Mark-up Language), and includes, for example, Chrome, Microsoft Edge, Safari, Firefox, Whale, UC Browser, etc. Additionally, an application refers to an application program on a terminal, and includes, for example, an app executed on a mobile terminal (smartphone).

[0027] The reference numerals for each component of FIG. 2 were assigned in the order of the claims. However, since the reference numerals were assigned in the order of importance of the components of the claims, they do not correspond to the actual order. Therefore, the components of FIG. 2 will be described with reference to FIG. 3a to 3c according to the actual order of execution.

[0028] <Learning Stage>

[0029] <Data Collection>

[0030] In order to train a prediction model, materials to be used for training must first be collected. To this end, equipment data of a mechanical parking facility can be collected from at least one information providing server (400). First, a mechanical parking facility refers to a facility that efficiently utilizes parking space by using a mechanical device to lift, move, and load vehicles instead of a person driving and parking the vehicles directly, as shown in FIGS. 4a to 4e.

[0031] Mechanical parking facility explanation Elevator A method of parking by lifting the vehicle up and down. Planar reciprocating A method of arranging in empty spaces by moving left, right, forward, and backward. Multi-stage circulation A system where vehicles move in a circular motion like a Ferris wheel Tower style High-rise structural system using vertical loading via elevator

[0032] In this context, equipment data of a mechanical parking facility refers to information indicating the operating status, physical characteristics, control signals, and operational history of the mechanical parking facility, and means all quantitative and qualitative data used to determine whether the facility is operating normally and to identify signs of abnormality.

[0033] type definition example ① Sensor Data (Physical Sensor Data) Data that directly measures the physical condition of the equipment ▶ The most critical material for signs of mechanical abnormalities Motor current, motor temperature, vibration RMS, hydraulic pressure, hydraulic fluid temperature, lifting height encoder value, rotational speed ② Operational KPI Data This is an indicator calculated from the operation results of the equipment. ▶ Quantifies perceived issues such as "it has slowed down" or "it is lagging." Door opening / closing time, elevator cycle time, turntable rotation time, waiting time, number of repetition delays ③ State / Event Data Discrete state changes occurring in equipment control systems ▶ Direct signal of abnormal operation Door interlock ON / OFF, vehicle presence sensor detection, position switch signal, emergency stop occurrence, manual operation intervention ④ Error / Alarm Log Data Anomaly history recorded on controllers or operating servers ▶ Very important for learning past failure precursor patterns Door interlock unreleash error, sensor chatter alarm, pressure drop alarm, communication error, overcurrent protection operation ⑤ Maintenance / Operation History Data Data recording human intervention ▶ The basis for the model connecting "what to do next" Chain lubrication history, switch alignment history, parts replacement history, inspection date and time, maintenance result memo ⑥ Operational Context Data Needed to explain the "pattern that becomes abnormal only at this time" even though the sensor values ​​are normal. Day of the week, time of day, peak hours, usage frequency, weather conditions

[0034] In this way, time synchronization can be performed on equipment data collected from sensors, motors, and controllers of mechanical parking facilities, and metadata can be assigned.

[0035] Normalization and Characterization

[0036] Next, if the collected equipment data differs in units or scales, it can be converted into a unified format, and after processing missing values ​​or outliers, time-series derived features, i.e., feature vectors, can be generated. First, for normalization, the raw equipment data collected from mechanical parking facilities can be transformed into analyzable normalized data by using an ETL pipeline, missing value interpolation algorithms, and outlier removal logic to correct unit inconsistencies and missing intervals, and to remove abnormal values ​​caused by sensor noise. Then, by processing the normalized equipment data in time-window units using moving average and variance calculation techniques, feature vectors containing time-series characteristics such as motor current changes, vibration variability, and cycle time delays can be generated. For example, the following feature vectors can be generated for each time point t.

[0037]

[0038] Of course, the features included in the feature vector are not limited to those listed in Equation 1.

[0039] <Formation of Pier Baselines>

[0040] The reference generation unit (350) can generate a cluster of mechanical parking facilities with similar facility data of at least one item constituting at least one mechanical parking facility in order to build training data for modeling a prediction model, and can generate a peer baseline based on the mean and variance of facility data of at least one item within the cluster.

[0041] First, there is not just one type of mechanical parking facility, but various types as shown in FIGS. 4a to 4e. Facility data is collected from various types of mechanical parking facilities; however, if all of this data is fed into a prediction model for training, the characteristics specific to each type will not appear, resulting in unnecessary training or training with reduced predictive power. Furthermore, even if the type of mechanical parking facility is the same, there is a problem where the normality is judged differently even with the same absolute sensor value due to differences in installation environment, usage frequency, and age of each facility. Accordingly, in one embodiment of the present invention, a peer reference line is established, and the extent to which the pattern deviates from this reference line can be determined.

[0042] At this time, referring to Fig. 3b, the peer baseline refers to a representative value or statistical standard of a normal state calculated by grouping equipment data collected from multiple mechanical parking facilities with similar structure, capacity, and operating conditions. It is a reference standard used to determine the extent and direction of deviation of an individual facility's current state compared to a group of similar facilities. Accordingly, the reason for establishing the peer baseline is to resolve the problem where the normality of a facility is judged differently even with the same absolute sensor value due to differences in installation environment, usage frequency, and age among mechanical parking facilities, even if they are identical. It is also established to more precisely identify signs of abnormality in individual facilities based on the relative state deviation within a group of similar facilities. Accordingly, the system is set to predict failure by observing how much the output value deviates from that of similar mechanical parking facilities, rather than the absolute value.

[0043] Accordingly, mechanical parking facilities with similar facility types, capacities, and usage patterns can be clustered. For this clustering, any one of the following unsupervised clustering models may be used: K-Means, Hierarchical Clustering, Gaussian Mixture Model, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), or HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise), but is not limited to these. The reason for using unsupervised clustering models is that facility data lacks ground truth labels (normal / failure), similarities between mechanical parking facilities are difficult to define clearly in advance, and installation environments and operating conditions vary widely. Therefore, unsupervised learning that enables the facility data to group autonomously is essential.

[0044] Next, mean and variance-based baselines, i.e., peer baselines, are generated within the same cluster. The process for this may be as shown in Table 3 below.

[0045] step explanation purpose Step 1: Select Peer Facility Group Group facilities satisfying at least some of the following conditions into peer groups: - Facility structure type (vertical circulation, elevator type, etc.) - Capacity and specifications - Control logic version - Installation age range - Average usage frequency The purpose is to group only the facilities that can be compared. Step 2: Selecting data from normal operation periods Not all data should be used for the baseline ▶ Excluded data: Sections where emergency stop occurred, Sections immediately after error code generation, Sections where instability occurs immediately after maintenance ▶ Included data: Sections where repeated operation occurred without errors, Sections where normal cycles were continuous To avoid distorting the standard of normal state Step 3: Separation by Time and Operation Conditions -Mechanical parking facilities vary in normal operation depending on the time- Separate data as follows: *By day of the week *By time of day *Peak / non-peak *Season or weather conditions This is because the morning peak and the night peak are different. Step 4: Calculation of Statistical Reference Values Calculate actual peer baseline * average baseline (Equation 2) → Average value for each feature item * Range of variation (covariance) (Equation 3) → Include correlation between sensors This is the basic pier baseline. Step 5: Outlier Removal and Stabilization To further stabilize the baseline, the following can be applied: IQR-based outlier removal, Hampel filter, and moving average-based smoothing. Prevents the baseline from shaking due to temporary spikes Step 6: Save and Fix the Baseline The generated peer baseline is stored along with: - Peer group ID - Time zone condition - Feature vector dimension ▶ Subsequently * Training phase: Reference baseline * Inference phase: Used as a fixed baseline

[0046]

[0047] Mathematical formula 2 represents the average value for each feature item.

[0048]

[0049] This is the basic peer baseline, and it can be further refined by performing outlier removal and stabilization in step 5. Then, in step 6, the peer baseline is saved and fixed; it serves as a reference standard for training during the training phase and as a fixed standard that is retrieved and used during the inference phase.

[0050] Once the peer baseline has been established in this way, the peer baseline deviation Δt must now be calculated for the equipment data, which can be expressed as Equation 4.

[0051]

[0052] x (site) t represents the feature vector of the equipment data of a specific mechanical parking facility at time t, and μ (peer) t represents the average feature vector of the equipment data of a homogeneous mechanical parking facility group.

[0053] In this context, "Site" refers to a physical location or management unit where one or more mechanical parking facilities are installed and operated, and includes units where the facilities are managed under the same operating entity or environmental conditions, such as apartment complexes, officetels, commercial buildings, and public parking facilities. "Peer" refers to a unit subject to comparison that constitutes a reference group for the relative comparison of the status of individual mechanical parking facilities, comprising mechanical parking facilities or sites where such facilities are installed that have identical or similar structures, capacities, and operational characteristics.

[0054] Accordingly, a relative deviation vector is calculated by quantifying how much the target mechanical parking facility deviates from peer mechanical parking facilities using mathematical formula 4.

[0055] Relative-based learning

[0056] The relative standard learning unit (360) can train a prediction model through self-supervised learning, which learns on its own even without labels for equipment data of at least one item within a cluster, and contrastive learning, which learns by contrasting equipment data of at least one item within a cluster that is similar or different. The prediction model may be a transformer-based prediction model. That is, it is a method that allows the model to learn on its own based on whether the data is similar or different, even without a human attaching the correct label.

[0057] Self-directed learning refers to learning autonomously by creating fake problems using unlabeled data. For example, this involves predicting the state of the next minute using sensors from the last 10 minutes, or learning the time flow of normal operating patterns and then predicting anomalies when the pattern is not normal. Contrastive learning involves learning by contrasting the same or different categories. For instance, if the weekday morning data and weekday afternoon data for the same mechanical parking facility A are similar (i.e., similar facility data appears in both morning and afternoon), it can be determined that the data from mechanical parking facility A and mechanical parking facility B are different. This is a learning method that places similar items close together in the embedding space and different items far apart.

[0058] Accordingly, referring to Fig. 3c, the method involves learning a representation space using facility data, which is unlabeled time-series data collected from multiple mechanical parking facilities, so that normal operation patterns between identical or similar mechanical parking facilities are represented closely, while different mechanical parking facilities or abnormal states are represented far apart. Through this, the prediction model learns, as a pattern in the representation space, how facility data changes over time relative to a peer baseline to be considered normal, and what patterns of change accumulate to lead to an error.

[0059] Learning Target explanation note Temporal variation pattern of relative deviation Learning whether the relative deviation—whether it increases suddenly, accumulates gradually, or appears in a specific order; that is, learning "what shape the flow of deviation from the baseline looks like" rather than "how much it deviates from the baseline." Example: *Normal → Deviation fluctuates but returns to the average* *Failure Precursor → Deviation continuously accumulates in one direction A representation space capable of distinguishing between normal and abnormal peers *Normal sections of the same equipment → close to each other* *Normal sections of peer equipment → close to each other* *Other equipment types or abnormal sections → far apart What the model learns is a relative distance structure: "This state is of the same class as / different class from that state." Directionality of patterns leading to error precursors -Learning that going in this direction is dangerous- In other words, the relative deviation vector learns 'which direction to move in to become dangerous' in the representation space. Example: *Vibration↑→Current↑→Cycle Delay↑→Dangerous Direction* *Vibration↑→Return Immediately→Normal Direction

[0060] In this case, the representation space is a vector space trained such that when the equipment data of the mechanical parking facility is converted into numerical vectors, [similar states are placed close together, and different states are placed far apart].

[0061] For example, equipment data such as a motor current of 11.8A, vibration of 0.21, cycle time of +1.2 seconds, and a few event logs is difficult to compare with one another, and it is also difficult to define normal / dangerous boundaries. Therefore, the prediction model transforms the equipment data once more, and the space in which the result of this transformation lies is the representation space. The inputs for this are the equipment data (time series data) and the deviation from the peer baseline (relative deviation). The output is a fixed-dimensional vector zt, and the space where these zts are collected is the representation space.

[0062]

[0063] Learning in the expression space refers to Table 5 below.

[0064] Learning Target explanation ① Normal state -Normal section of the same facility-Normal section of the peer facility ▶ Arranged close to each other in the representation space ② Warning signs of danger and error -Section where deviations from the peer baseline accumulate- Pattern immediately preceding failure ▶ Arranged to move away from the normal state ③ Key Points Predictive models do not remember "what this value is" but learn "which cluster this state belongs to" from the representation space.

[0065] To use an intuitive analogy, the representation space can be assumed to be a point plotted on a map, where the normal state is neighborhood A, the dangerous state is neighborhood B, and the failure precursor is a point moving from the normal neighborhood to the dangerous neighborhood. At this time, the prediction model learns which neighborhood this point is currently in and which neighborhood it is heading to. Accordingly, during inference, the equipment data of the mechanical parking facility input from the user terminal (100) is converted into a vector zt, and after confirming the location in the representation space, the distance and direction from the danger cluster are calculated to calculate a risk score. This is the prediction of the possibility of error in the inference step described later. Here, Xt is as shown in Equation 6 below.

[0066]

[0067] In this case, as explained in Equation 1, xt is a feature vector (current, vibration, pressure, cycle time, etc.) that quantifies the equipment data of the mechanical parking facility at time t, t is the reference time point for judgment, and Xt represents not just the current state, but the entire flow of state changes up to the present. In other words, Xt is a bundle of time-series data (equipment data) input into the prediction model. Xt is important because failures usually do not occur all at once, but rather arise when deviations accumulate over time. L is the length of the input sequence. That is, if xt is a snapshot of the mechanical parking facility at this very moment, Xt can be likened to an image of the facility over the past few minutes. Accordingly, the prediction model makes a judgment based on the image (Xt) rather than a single photograph (xt).

[0068] <Inference Stage>

[0069] The collection unit (310) can collect equipment data from the user terminal (100). The user terminal (100) can provide equipment data of the mechanical parking facility.

[0070] The feature unit (320) can generate a feature vector in which the equipment data is arranged in a time series after normalizing the equipment data. The feature vector generated by the feature unit (320) can be converted into a relative deviation relative to a pre-set peer baseline and input into a prediction model. At this time, the relative deviation refers to Δt. At this time, the peer baseline is generated based on the mean and variance of the equipment data of at least one item within the cluster, after mechanical parking facilities with similar equipment data of at least one item constituting at least one mechanical parking facility are generated as a cluster.

[0071] At this point, the inference stage follows the same process as the learning stage, [data collection - normalization & feature]. Next, the peer baseline undergoes a selection process rather than formation. This is because the peer baselines have already been generated by type during the learning stage. Accordingly, in the inference stage, the process of selecting the already generated peer baselines is carried out according to the type, capacity, etc. of the mechanical parking facility of the user terminal (100). This is because data to be included in the prediction model is prepared in this way. The learning stage and the inference stage of the artificial intelligence model, including the prediction model, have the same process. That is, while the form of the process is the same, there is a difference in that model parameters are updated in the learning stage, whereas in the inference stage, only the prediction result is produced while the model parameters remain fixed. Accordingly, if the learning stage goes through [data collection - normalization & feature - peer baseline formation - relative standard learning], the inference stage goes through [data collection - normalization & feature - peer baseline selection - Transformer prediction].

[0072] division Learning stage Inference stage Peer cluster generation use Pier baseline forming Select & Apply Relative standard learning commitment Do not perform Calculation of relative deviation commitment commitment Prediction model (Transformer-based) Learning inference

[0073] The inference unit (330) can input a feature vector into a pre-established prediction model. A prediction model according to one embodiment of the present invention calculates the probability of error using the deviation between the facility data, which is time-series data collected from a mechanical parking facility, and the peer baseline, and the calculation flow is performed in the following order.

[0074] <Time Encoding>

[0075] A prediction model according to one embodiment of the present invention directly reflects the difference between normal and abnormal patterns according to operational contexts such as day of the week, time zone, and peak time as time information. To this end, positional encoding is performed, which is as shown in Equation 7.

[0076]

[0077] t is the time index in the time series, i is the embedding dimension index, and d represents the total embedding dimension. In other words, it represents the size of the model representation space. PE(t,·) is a time encoding vector, which represents a position representation containing time information.

[0078] <Input Embedding and QKV Generation>

[0079] First, the input embedding is as shown in Equation 8 below.

[0080]

[0081] As mentioned above, Xt is a bundle of equipment data up to time point t, and PEt is a time encoding vector. X't is an input embedding containing time information. In other words, Xt contains only the value itself and does not contain time or location information, but with the addition of PEt, it becomes a combined representation of [value + time / location]. Accordingly, the prediction model understands the order and flow.

[0082] The next step is the process of calculating Q, K, and V for the attention mechanism. Q (Query), K (Key), and V (Value) are three vector representations generated to compare and reference the input equipment data; they are internal representations used to calculate how related information at a given point in time is to information at another point in time or from other sensors.

[0083] In this context, Q serves as a query asking what state I want to know; K represents index information indicating that the sensor or point in time possesses specific characteristics; and V represents the actual information being transmitted, indicating that the real information of that point in time or sensor is what it is. For example, Q represents the motor current change or cycle delay state at the present or a specific point in time, K represents vibration, hydraulic pressure, switch status, etc., at a past point in time, and V represents the entire actual equipment data for that point in time. In other words, it is a structure that numerically calculates which past vibration or pressure patterns are most closely related to the current abnormal motor current.

[0084] To summarize, Q·K·V is an internal expression that separates [what to compare based on (Q)], [what to compare with (K)], and [what information to reflect as a result of the comparison (V)].

[0085]

[0086] Q, K, and V are generated from X't, meaning that only time-reflecting inputs can be used for attention calculation. In this case, WQ, WK, and WV are learning parameters for generating Query, Key, and Value. Accordingly, representation vectors for attention calculation are generated from the equipment data.

[0087] Attention

[0088] Attention operations such as Equation 10 reflect the correlation and importance between sensors in the form of weights, for example, by quantitatively reflecting which sensor signal (vibration, pressure, etc.) a change in motor current appears together with, thereby highlighting the fault precursor signal.

[0089]

[0090] dk represents the dimension of the Key vector. Att(Q,K,T) is the attention output, where Q is the feature vector of the object of interest, such as motor current, vibration, and response time, and K is the feature vector of the comparison criterion, such as sensor state, switch signal, and hydraulic pressure.

[0091] Parallel Learning

[0092] Next, parallel learning is performed to learn multiple sensor combinations and fault signs in parallel, and the Multi-Head Attention (MHA) for this is as shown in Equation 11 below.

[0093]

[0094] headh ​​represents the attention result calculated from different perspectives, and H represents the number of attention heads. In this case, headh ​​is the h-th attention result vector (Equation 10) calculated from Q·K·V, which is calculated by projecting the same input X't onto different weight matrices.

[0095] Concat stands for Concatenation, a vector linkage operation, and WO is an output weight matrix consisting of learning parameters automatically determined during the training phase. Through this, different fault indicators—such as door opening / closing, vehicle presence sensors, hydraulic pressure, and motor vibration—can be learned in parallel, thereby preventing false positives based on a single indicator.

[0096] High-dimensional feature extraction

[0097] Next, the following mathematical formula 12 can be used to enhance the high-dimensional feature representation by non-linearly transforming the attention results.

[0098]

[0099] FFN(x) represents a high-dimensional non-linear feature as an output vector, x is an input vector (attention-processed feature vector), W1 and W2 are weight matrices, b1 and b2 are bias vectors (offset correction), and σ(·) is an activation function, for example, ReLU or GELU. The meaning of Equation 12 is to non-linearly amplify the in / out cycle time, iteration period, and minute delay pattern to enable learning or inference of even subtle changes prior to failure.

[0100] In this context, a Position-wise Feed Forward Network refers to a neural network module that performs a non-linear transformation by independently applying a Fully Connected Neural Network with identical parameters to the output vector of each position generated through attention operations in a Transformer structure, without mixing information between positions.

[0101] <Prediction Model Output>

[0102] After passing through several encoder layers, the output from the prediction model encoder may be the following mathematical formula 13.

[0103]

[0104] Then, the risk score finally output from the prediction model output layer (head) is as shown in Equation 14 below.

[0105]

[0106] Wr and br are risk weights (risk judgment parameters), and r(t) is the risk score.

[0107] In other words, the prediction model combines learned temporal anomaly patterns and peer baseline deviations to output the probability of an error occurring as a value. At this stage, it becomes possible to determine whether to trigger an alert or not.

[0108] The diagnosis unit (340) can perform a risk diagnosis based on the predicted result when the prediction model predicts a risk. For example, if the risk score is 0.7 and the value separating the risk-urgent from normal-risk-urgent is 0.69, a warning can be sounded as an emergency situation. At this time, the risk diagnosis is to determine what kind of risk it is, why it is dangerous, and whether immediate action is required. It can determine the level of risk, estimate the type of risk, determine whether there is a precursor to failure, and determine the necessity of action.

[0109] Risk Assessment explanation Risk Level Assessment Determines stepwise risk levels such as [Normal-Caution-Warning-Danger] based on the predicted risk score Estimation of risk type Analyze which relative deviation is the primary cause—Predominant current deviation—Predominant oscillation deviation—Predominant cycle time delay ▶ Specify the nature of the risk, rather than a "vague risk." Determining whether there are signs of a failure -Distinguish between short-term anomalies and cumulative warning signs over time ▶ Decide between immediate alarm vs. maintaining observation Determination of the necessity of measures - Whether immediate suspension is necessary - Recommended inspection level - Maintain simple monitoring ▶ Generate a decision result to proceed to the next step

[0110] For example, the prediction model may output a risk score of 0.82, and the risk diagnosis may be [a pattern in which motor current and vibration deviations increase simultaneously, indicating a high probability of failure due to increased chain friction].

[0111] The action recommendation unit (370) can analyze the cause of the error based on the prediction results and provide a guideline for action measures to eliminate the cause of the error. By using a multimodal analysis module, a RAG (Retrieval-Augmented Generation) based diagnostic engine, and a rule engine, the cause of the error can be derived by comprehensively analyzing the prediction results, equipment data, past failure cases, and maintenance knowledge, and a guideline for action measures to eliminate the cause of the error can be generated.

[0112] The improvement unit (380) can retrain the prediction model by collecting equipment data of the mechanical parking facility and performing a performance evaluation to verify whether the prediction result is correct.

[0113] <Recommended for minimum usage>

[0114] In addition, a platform according to one embodiment of the present invention may further provide a minimum time usage recommendation algorithm for the order of inbound and outbound operations based on customer requests. The predictive maintenance server (300) may proceed with ① the request and equipment status processing and collection step. That is, it collects multiple inbound and outbound requests received from a user, current status data of the mechanical parking facility, and operational context information. Then, the predictive maintenance server (300) proceeds with ② the inbound and outbound time prediction step, in which a transformer-based prediction model can predict the expected processing time and cycle time for each inbound and outbound request by using equipment status data and time zone / load information as input. Then, the predictive maintenance server (300) proceeds with ③ the relative performance correction step, in which a peer baseline-based relative deviation calculation logic compares the predicted time with the average performance of similar equipment to calculate a correction time that reflects the possibility of performance degradation or delay of the current equipment.

[0115] Then, the predictive maintenance server (300) proceeds with the ④ sequence optimization step, and can determine the execution order of inbound and outbound requests so that the total processing time is minimized based on the corrected inbound and outbound time using a sequence optimization algorithm and a cost function-based scheduling logic. In addition, the predictive maintenance server (300) proceeds with the ⑤ recommendation and guidance step, and the recommendation result generation module and the user interface linkage module provide the determined inbound and outbound order to the user or operation manager as a recommendation result based on the minimum waiting time.

[0116] <Recommended In / Out Order Based on License Plate Number>

[0117] A platform according to one embodiment of the present invention can further recommend an entry and exit order using vehicle number-based usage pattern analysis. The predictive maintenance server (300) can learn the average entry time, average exit time, and repeated usage pattern of a specific vehicle by analyzing the past entry and exit history of the mechanical parking facility based on the vehicle number. Additionally, when multiple exit requests occur at the same time, the predictive maintenance server (300) can minimize the total waiting time by recommending an entry and exit order to prioritize the processing of vehicles that had a short exit time in their past operation history or vehicles that were frequently used during that time. For the recommendation, the predictive maintenance server (300) may use a time-series statistical model to calculate the average exit time, variance, and frequency by time period for each vehicle, a clustering algorithm such as K-Means or a Gaussian Mixture Model to cluster vehicles with similar usage patterns, and a weight-based priority calculation algorithm to consider past average processing time and recent usage frequency.

[0118] Additionally, ② the order of inbound and outbound operations can be optimized in advance by reflecting the reservation status. The predictive maintenance server (300) can predict the flow of inbound and outbound requests within a certain time interval by analyzing the inbound and outbound reservation information registered in advance and the current equipment status together. Based on the predicted inbound and outbound request timing and the estimated processing time for each request, the predictive maintenance server (300) can determine the order of inbound and outbound operations in advance so that the total processing time among multiple reservation requests is minimized, and can provide the order to the user or management system. In order to minimize the total waiting time by considering future inbound and outbound requests in this way, the predictive maintenance server (300) can use a time-based request prediction model such as LSTM or Transformer, a scheduling optimization algorithm such as SPT (Shortest Processing Time) or EDF (Earliest Deadline First), and a cost function-based greedy algorithm as an objective function for minimizing the total processing time.

[0119] In addition, ③ a scenario based on past actual operation cases may be recommended. The predictive maintenance server (300) stores actual inbound and outbound operation cases performed under the same or similar equipment conditions and usage conditions in the past in a database and can search for operation cases that are highly similar to the current situation. By applying the inbound and outbound order applied in the searched past operation cases to the current request situation, the predictive maintenance server (300) can recommend the shortest time inbound and outbound order based on an already verified operation scenario. In order to reapply a method that worked well in the past to the current situation, the predictive maintenance server (300) may use similarity calculation algorithms such as Case-Based Reasoning (CBR), Cosine Similarity or Dynamic Time Warping (DTW), Top-K nearest case selection logic, etc.

[0120] Additionally, ④ the order may be dynamically changed in consideration of equipment condition deterioration. When the predictive maintenance server (300) determines through a prediction model that the cycle time of a specific facility is delayed relative to the peer baseline, it may place the mechanical parking facility in a lower priority and dynamically adjust the order of entry and exit so that facilities in relatively good condition are utilized first. In order to place mechanical parking facilities with signs of failure in a lower priority in the operation order, the predictive maintenance server (300) may use a Transformer-based prediction model according to an embodiment of the present invention for predicting cycle time and delay possibility, a peer baseline-based relative performance score to use a Δt-based performance degradation index, and a dynamic weighted scheduling algorithm for reordering the order according to the condition score.

[0121] Hereinafter, the operation process according to the configuration of the predictive maintenance server of FIG. 2 described above will be explained in detail with reference to FIG. 3 and FIG. 4. However, it is obvious that the embodiment is merely one of the various embodiments of the present invention and is not limited thereto.

[0122] Referring to FIGS. 3a through 3c, for the learning stage, data is collected, preprocessed, and feature vectors are generated. Then, groups (clusters) are created through clustering among at least one type of mechanical parking facility, and peer baselines are formed for each cluster. Subsequently, to train a prediction model, feature vectors of facility data from individual mechanical parking facilities, relative deviation vectors relative to peer baselines, and time operation context information are input as training data. This allows the model to learn the flow when the peer baseline evacuation deviation is normal, the order in which deviations accumulate when a failure is a precursor, and how the deviation structure changes according to time zones and load conditions. The learning methods used are self-directed learning and contrastive learning. Through self-directed learning, deviations at the next time point can be predicted or reconstructed, and through contrastive learning, the representations of normal and deviation states can be separated. Here, the parameters of the prediction model, i.e., the transformer, are updated.

[0123] In the subsequent inference stage, the real-time error probability is calculated through the learned prediction model. Real-time equipment data is collected from the user terminal (100), preprocessing and feature vectors are generated (same as in the learning stage), and relative deviations are calculated based on the already generated peer baseline. Since the peer baseline was formed in the learning stage, it is only used in the inference stage and is not newly generated. Then, the equipment data is configured as a time series input and fed into the prediction model to start the inference operation. That is, the encoder of the learned prediction model is executed to generate a high-dimensional representation vector representing the equipment status, and after calculating the risk score, an alarm and a recommendation are output.

[0124] FIGS. 4a to 4e are drawings included to illustrate types of mechanical parking facilities. The drawings of the mechanical parking facilities in FIGS. 4a to 4e were extracted from the catalog of the applicant (Otek Otis Parking System Co., Ltd.).

[0125] As for the details regarding the method for providing a solution for pre-determining the possibility of error shown in FIGS. 2 to 4 that are not described, they are identical to or can be easily inferred from the details described above regarding the method for providing a solution for pre-determining the possibility of error shown in FIG. 1, so further explanation will be omitted.

[0126] FIG. 5 is a diagram illustrating the process of data transmission and reception between each component included in the error probability pre-determination system using a transformer-based prediction model for remote service of mechanical parking facilities of FIG. 1 according to an embodiment of the present invention. Hereinafter, an example of the process of data transmission and reception between each component will be described through FIG. 5, but the present invention is not to be interpreted as being limited to such an embodiment, and it is obvious to those skilled in the art that the process of data transmission and reception shown in FIG. 5 may be changed according to various embodiments described above.

[0127] Referring to FIG. 5, the predictive maintenance server collects equipment data from a user terminal (S5100), normalizes the equipment data, and then generates a feature vector in which the equipment data is arranged in a time series (S5200).

[0128] Additionally, the predictive maintenance server inputs the feature vector into a pre-established prediction model (S5300), and when the prediction model predicts a risk, it performs a risk diagnosis based on the predicted result (S5400).

[0129] The order of the steps described above (S5100~S5400) is merely an example and is not limited thereto. That is, the order of the steps described above (S5100~S5400) may vary, and some of these steps may be executed simultaneously or deleted.

[0130] As for the details regarding the method for providing a solution for pre-determining the possibility of error shown in Fig. 5 that are not described, they are identical to or can be easily inferred from the details described above regarding the method for providing a solution for pre-determining the possibility of error shown in Figs. 1 to 4, so further explanation will be omitted.

[0131] The method for providing a solution for pre-determining the possibility of an error according to one embodiment described through FIG. 5 may also be implemented in the form of a recording medium containing instructions executable by a computer, such as an application or program module executed by a computer. A computer-readable medium may be any available medium accessible by a computer and includes both volatile and non-volatile media, and both removable and non-removable media. Additionally, a computer-readable medium may include all computer storage media. A computer storage medium includes both volatile and non-volatile, removable and non-removable media implemented by any method or technique for storing information such as computer-readable instructions, data structures, program modules, or other data.

[0132] The method for providing a solution for pre-determining the possibility of an error according to one embodiment of the present invention described above may be executed by an application that is basically installed on a terminal (which may include a program included in a platform or operating system, etc., that is basically installed on the terminal), or by an application (i.e., a program) that is directly installed by a user on a master terminal through an application providing server, such as an application store server, an application, or a web server related to the service. In this sense, the method for providing a solution for pre-determining the possibility of an error according to one embodiment of the present invention described above may be implemented as an application (i.e., a program) that is basically installed on a terminal or directly installed by a user, and may be recorded on a computer-readable recording medium such as a terminal.

[0133] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0134] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention.

Claims

Claim 1 A system for pre-determining error probability using a transformer-based prediction model for remote services of mechanical parking facilities, comprising: a user terminal providing facility data of a mechanical parking facility; a collection unit that collects facility data from the user terminal; a feature unit that generates a feature vector by normalizing the facility data and arranging the facility data in a time series; an inference unit that inputs the feature vector into a pre-established prediction model; and a diagnosis unit that performs a risk diagnosis based on the predicted result when the prediction model predicts a risk, wherein the feature vector generated by the feature unit is converted into a relative deviation relative to a pre-set peer baseline and input into the prediction model, and the peer baseline is generated based on the mean and variance of the facility data of at least one item within the cluster after mechanical parking facilities with similar facility data of at least one item constituting at least one mechanical parking facility are generated as a cluster. Claim 2 delete Claim 3 A system for pre-determining error probability using a transformer-based prediction model for remote services of mechanical parking facilities, characterized in that, in claim 1, the predictive maintenance server further comprises a reference generation unit that, in order to construct training data for modeling the prediction model, generates a cluster of mechanical parking facilities where the facility data of at least one item constituting at least one mechanical parking facility is similar, and generates a peer baseline based on the mean and variance of the facility data of at least one item within the cluster. Claim 4 In claim 3, the predictive maintenance server further comprises a relative standard learning unit that trains the prediction model through self-supervised learning, which learns even without labels on equipment data of at least one item within the cluster; and contrastive learning, which learns by contrasting equipment data of at least one item within the cluster with similar or different ones. This characterizes a system for pre-determining error probability using a transformer-based prediction model for remote services of mechanical parking facilities. Claim 5 A system for pre-determining the possibility of an error using a transformer-based prediction model for remote service of mechanical parking facilities, characterized in that, in claim 1, the predictive maintenance server further includes a measure recommendation unit that analyzes the cause of an error based on the prediction result and provides a measure to eliminate the cause of the error as a guideline. Claim 6 A system for pre-determining error probability using a transformer-based prediction model for remote service of mechanical parking facilities, characterized in that, in claim 1, the predictive maintenance server further includes an improvement unit that collects facility data of the mechanical parking facility, performs a performance evaluation to verify whether the prediction result is correct, and retrains the prediction model. Claim 7 A system for pre-determining error probability using a transformer-based prediction model for remote service of mechanical parking facilities, characterized in that, in claim 1, the prediction model is a transformer-based prediction model.

Images