An elevator failure early warning method based on a prediction model

Abstract

This invention provides an elevator fault early warning method based on a predictive model, belonging to the field of elevator fault monitoring technology. The method involves storing elevator fault information in a database; acquiring elevator state parameters; identifying elevator state keywords in the parameters using an ERNIE pre-trained language model; matching the identified elevator state keywords with elevator fault information in the database; and displaying the elevator fault information upon successful matching. Based on the pre-trained model, this invention, through model fine-tuning and centralized learning on elevator emergency response datasets, can quickly obtain suitable elevator fault text and match elevator fault information. Furthermore, the elevator fault prediction model based on ERNIE pre-training achieves a loss of 0.23 and an accuracy of 0.72, demonstrating its potential to improve elevator emergency response services through artificial intelligence technology and showing promising application prospects.

Description

Technical Field

[0001] This invention relates to the field of elevator fault monitoring technology, and in particular to an elevator fault early warning method based on a predictive model. Background Technology

[0002] With rapid urbanization, the number of elevators, a vertical transportation tool, continues to grow. Faced with a serious elevator safety management situation, various regions have successively established integrated elevator emergency response platforms, deployed hotlines or call systems, and assigned professional answering teams, providing crucial support for rapid response and proper handling of elevator safety accidents.

[0003] The integrated elevator emergency response platform enables elevator emergency response, including emergency command activities for elevator entrapment incidents and coordination activities for other incidents. Built upon information technologies such as communication, command and dispatch, geographic information systems, and the Internet of Things, the platform receives alarms for elevator entrapment and other malfunctions, and possesses functions such as emergency command, statistical analysis, consulting services, social supervision, and information dissemination, providing support for organizing, coordinating, and directing rescue operations.

[0004] However, with the increasing number of elevators and their aging, various elevator malfunctions are becoming more frequent, posing challenges to hotline access, problem identification, and response allocation. Currently, after receiving elevator status information, operators need to match the corresponding elevator fault from the database before reporting it. The mismatch between the number of operators and the number of elevators, as well as the insufficient professional experience of operators to handle the surge in elevator malfunctions, has become apparent, resulting in ineffective handling of reported information, delays in elevator fault resolution, and an inability to effectively eliminate elevator malfunctions. Summary of the Invention

[0005] This invention provides an elevator fault early warning method based on a predictive model. The method facilitates the identification of elevator faults, effectively matches fault information in the database, and helps to troubleshoot elevator faults in the later stage.

[0006] The methods include:

[0007] Step S101: Generate and store elevator fault information in the database;

[0008] Step S102: Obtain elevator state parameters and identify elevator state keywords in the elevator state parameters using the ERNIE pre-trained language model.

[0009] Step S103: Match the identified elevator status keywords with elevator fault information in the database. Once a match is found, display the elevator fault information.

[0010] Preferably, in step S101, elevator fault information classification conditions are set to classify elevator fault information and form an elevator fault classification table.

[0011] Label each elevator malfunction information.

[0012] Preferably, the labels for elevator fault information are matched with the elevator status keywords of the elevator status parameters;

[0013] The identified elevator status keywords are matched with elevator fault information tags. Once a match is successful, the elevator fault information is displayed.

[0014] Preferably, in step S102, an ERNIE pre-trained language model is constructed, so that the ERNIE pre-trained language model provides task embeddings to represent elevator state keywords in elevator state parameters.

[0015] Each recognition task is assigned a unique task embedding, and the corresponding elevator state parameters, recognition token, recognition statement, and recognition task are used as inputs to the ERNIE pre-trained language model.

[0016] Preferably, in step S102, the ERNIE pre-trained language model obtains the elevator state parameter sequence, embeds a [CLS] marker at the beginning of each elevator state parameter, and sets a separator [SEP] between the elevator state parameters;

[0017] The ERNIE pre-trained language model uses a multi-layer Transformer as the basic encoder and captures each token in the elevator state parameter sequence through a self-attention mechanism. It automatically assigns each elevator state parameter recognition task to different training stages to identify elevator state keywords.

[0018] Preferably, in step S102, the ERNIE pre-trained language model is evaluated using an AUC model. The AUC model retrieves evaluation samples, sorts the samples according to the probability prediction results generated by the classification model, and divides the sorted samples into two parts {x1, x2, ... x...} according to a certain cut-off point. m+1 x m+n The first m samples are classified as positive examples, and the second n samples are classified as negative examples. The AUC is defined as:

[0019]

[0020] Where f is the classification function; I is the indicator function, which is 1 when the expression in parentheses is true, and 0 otherwise. The AUC model takes values ​​between [0.5, 1].

[0021] Preferably, the AUC model uses precision, recall, and F1 score as evaluation metrics.

[0022] Where: precision represents the proportion of samples whose actual class and predicted class are both positive out of all samples whose predicted class is positive; recall represents the proportion of samples whose actual class and predicted class are both positive out of all samples whose actual class is positive; f1 score is the weighted harmonic mean of precision and recall; the calculation formulas are shown in (1)-(3);

[0023]

[0024]

[0025]

[0026] Wherein: TP indicates that the predicted class is A and the actual class is A; FP indicates that the predicted class is A but the actual class is not A; TN indicates that the predicted class is not A and the actual class is not A; FN indicates that the predicted class is not A but the actual class is A.

[0027] Preferably, the method also verifies the elevator state keywords identified by the ERNIE pre-trained language model in the elevator state parameters;

[0028] The validation method uses binary cross-entropy loss as the loss function for multi-label classification, which is defined as:

[0029]

[0030] Where N is the number of elevator status keywords in the elevator status parameters, K is the number of tags for elevator fault information, and y ij ∈{0,1} and Let represent the actual tag value and the probability prediction value of the j-th tag for the i-th elevator status keyword, respectively;

[0031] The skewed triangular learning rate (STLR) method is used, which involves first linearly increasing the learning rate and then linearly decreasing it according to the training cycle; specifically, it is expressed as:

[0032]

[0033]

[0034]

[0035] Where T is the total number of training iterations, cut fracThe learning rate increases as a percentage of the total training iterations; cut is the iteration number at which the learning rate inflects; p is the scaling ratio at which the learning rate increases or decreases; and ratio is the ratio between the minimum and maximum learning rates, η. max The ratio, η t It is the learning rate for the t-th iteration;

[0036] When cut frac When the ratio is equal to 0.1 and the ratio is equal to 32, it has the characteristics of short-term growth and long decay period. The ERNIE pre-trained language model identifies that the elevator state keywords in the elevator state parameters meet the requirements.

[0037] As can be seen from the above technical solutions, the present invention has the following advantages:

[0038] This invention summarizes elevator status parameters, facilitating access for monitoring personnel and effectively improving the efficiency of elevator operation monitoring. It also efficiently collects, stores, and processes elevator status parameters, identifying elevator status keywords and matching them to elevator fault information, which is then displayed. This improves the real-time performance and accuracy of elevator fault warnings, enabling timely detection and alerts to safety hazards during elevator operation. This enhances the management level and efficiency of elevator operation, controls risks during operation, and ultimately achieves timely and scientific supervision, management, and control of elevator operation.

[0039] This invention also applies the ERNIE pre-trained language model for recognition learning, enabling automatic classification and auxiliary reminders of fault types and causes of access hotlines, providing technical support for operators to better identify and handle elevator malfunctions. Attached Figure Description

[0040] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 The flowchart shows an elevator fault early warning method based on a predictive model.

[0042] Figure 2 This is a diagram illustrating the causes of elevator malfunctions. Detailed Implementation

[0043] The elevator fault early warning method based on a predictive model provided by this invention can acquire and process related data using artificial intelligence technology. Specifically, the elevator fault early warning method utilizes a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to obtain optimal results—a theory, method, technology, and application device.

[0044] Elevator malfunction early warning methods involve both hardware and software technologies. The fundamental technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operation / interaction systems, and mechatronics. The software technologies primarily include computer vision techniques, speech processing, natural language processing, and machine learning / deep learning. Elevator malfunction early warning methods also incorporate machine learning capabilities; the machine learning and deep learning methods used in this invention typically include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formulaic learning.

[0045] This elevator fault early warning method utilizes an ERNIE pre-trained language model, an AUC model, and a loss function to identify elevator state parameters and interactively reflect elevator fault information in real time, effectively matching the elevator's real-time status. It addresses the issue of inconsistent fault identification and timely handling caused by the diverse types of elevator anomalies and varying expressions of distress by those seeking help. Furthermore, it effectively solves the problem of inaccurate identification due to insufficient data in traditional methods.

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1 The diagram shows a flowchart of an elevator fault early warning method based on a prediction model in a specific embodiment. The method includes:

[0048] Step S101: Generate and store elevator fault information in the database;

[0049] This involves setting classification criteria for elevator malfunction information, classifying the elevator malfunction information to form an elevator malfunction classification table, and assigning a label to each elevator malfunction information.

[0050] In other words, to effectively manage common elevator malfunctions, elevator accident data should be categorized. These categories can include: violations of operating procedures or improper operation; equipment defects and safety component failures or protective device malfunctions; improper emergency rescue (self-rescue); inadequate safety management and maintenance; unlicensed operation; and failure to conduct timely inspections.

[0051] Based on the structural characteristics of elevators, common elevator accident causes are subdivided into nine categories, including door system and car-related causes. Figure 2 As shown, this provides support for the targeted initiation of corresponding response measures. Considering that the primary task of elevator emergency response is to protect human life and health safety, the fault types are set into two categories: entrapment and other, highlighting the rapid response to entrapment faults. This results in the elevator emergency response label data items, as shown in Table 1:

[0052] Table 1. Types and Causes of Elevator Faults

[0053]

[0054]

[0055] For example, the present invention also preprocesses elevator emergency response data. Data preprocessing can be performed on the elevator emergency response service platform's call handling record data, extracting key descriptions and labeling fault types and cause classification codes to obtain the historical emergency response dataset shown in Table 2.

[0056] Table 2

[0057]

[0058] Step S102: Obtain elevator state parameters and identify elevator state keywords in the elevator state parameters using the ERNIE pre-trained language model.

[0059] Among them, the labels of elevator fault information are matched with the elevator status keywords of elevator status parameters; the identified elevator status keywords are matched with the labels of elevator fault information, and the elevator fault information is displayed after a successful match.

[0060] In one exemplary embodiment, artificial intelligence technology is applied to intelligently process emergency calls to elevators, which falls under the text classification task in Natural Language Processing (NLP). Since it requires simultaneous fault type identification and fault cause classification, this task can be further defined as a multi-classification text task. Furthermore, due to the diverse types of elevator malfunctions and the varying expressive abilities of those seeking help, the NLP model needs to possess good generalization capabilities.

[0061] Pre-trained language models (PLMs or PTMs), also known as pre-trained language models or pre-trained models, belong to the transfer learning method. Natural language has relatively obvious inherent annotation features. Pre-trained models such as BERT and ERNIE, trained on large natural language corpora, exhibit outstanding characteristics in listening, speaking, reading, writing, translation, question answering, searching, summarizing, dialogue, and chat, demonstrating good generalization ability. Based on pre-trained models, through model fine-tuning and centralized learning on elevator emergency response datasets, models suitable for multi-classification tasks involving elevator malfunction text can be quickly obtained.

[0062] The ERNIE pre-trained language model involved in this invention is a semantic representation model proposed by Baidu. This model uses Transformer Encoder for semantic representation and can learn lexical, syntactic and semantic information at the same time. Compared with the BERT model, the ERNIE model improves the masking mechanism and directly models the prior semantic knowledge units, thereby enhancing the model's semantic representation ability.

[0063] In embodiments of the present invention, an ERNIE pre-trained language model is constructed, which provides task embedding to represent elevator state keywords in elevator state parameters.

[0064] Each recognition task is assigned a unique task embedding, and the corresponding elevator state parameters, recognition token, recognition statement, and recognition task are used as inputs to the ERNIE pre-trained language model.

[0065] The ERNIE pre-trained language model obtains the elevator state parameter sequence, embeds a [CLS] marker at the beginning of each elevator state parameter, and sets a [SEP] separator between the elevator state parameters;

[0066] The ERNIE pre-trained language model uses a multi-layer Transformer as the basic encoder and captures each token in the elevator state parameter sequence through a self-attention mechanism. It automatically assigns each elevator state parameter recognition task to different training stages to identify elevator state keywords.

[0067] In this way, the ERNIE pre-trained language model can acquire multiple elevator state parameters and establish an elevator state parameter sequence. It can simultaneously identify elevator state keywords in the elevator state parameters, ensuring recognition efficiency without forgetting previously trained elements.

[0068] Step S103: Match the identified elevator status keywords with elevator fault information in the database. If the match is successful, display the elevator fault information.

[0069] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0070] This invention aims to evaluate ERNIE pre-trained language models to ensure they meet recognition requirements.

[0071] The ERNIE pre-trained language model is evaluated using the AUC model. The AUC model retrieves evaluation samples, sorts the samples according to the probability prediction results generated by the classification model, and divides the sorted samples into two parts {x1, x2, ... x3} by a certain cut-off point. m+1 x m+n The first m samples are classified as positive examples, and the second n samples are classified as negative examples. The AUC is defined as:

[0072]

[0073] Where f is the classification function; I is the indicator function, which is 1 when the expression in parentheses is true, and 0 otherwise. The AUC model takes values ​​between [0.5, 1], and a larger value indicates better performance of the ERNIE pre-trained language model.

[0074] The AUC model uses precision, recall, and F1 score as evaluation metrics.

[0075] Where: precision represents the proportion of samples whose actual class and predicted class are both positive out of all samples whose predicted class is positive; recall represents the proportion of samples whose actual class and predicted class are both positive out of all samples whose actual class is positive; f1 score is the weighted harmonic mean of precision and recall; the calculation formulas are shown in (1)-(3);

[0076]

[0077]

[0078]

[0079] Wherein: TP indicates that the predicted class is A and the actual class is A; FP indicates that the predicted class is A but the actual class is not A; TN indicates that the predicted class is not A and the actual class is not A; FN indicates that the predicted class is not A but the actual class is A.

[0080] In embodiments of the present invention, the elevator state keywords identified by the ERNIE pre-trained language model in the elevator state parameters are also verified.

[0081] The validation method uses binary cross-entropy loss as the loss function for multi-label classification, which is defined as:

[0082]

[0083] Where N is the number of elevator status keywords in the elevator status parameters, K is the number of tags for elevator fault information, and y ij ∈{0,1} and Let represent the actual tag value and the probability prediction value of the j-th tag for the i-th elevator status keyword, respectively;

[0084] The skewed triangular learning rate (STLR) method is used, which involves first linearly increasing the learning rate and then linearly decreasing it according to the training cycle; specifically, it is expressed as:

[0085]

[0086]

[0087]

[0088] Where T is the total number of training iterations, cut frac The learning rate increases as a percentage of the total training iterations; cut is the iteration number at which the learning rate inflects; p is the scaling ratio at which the learning rate increases or decreases; and ratio is the ratio between the minimum and maximum learning rates, η. max The ratio, η t It is the learning rate for the t-th iteration;

[0089] When cut frac When the ratio is equal to 0.1 and the ratio is equal to 32, it has the characteristics of short-term growth and long decay period. The ERNIE pre-trained language model identifies that the elevator state keywords in the elevator state parameters meet the requirements.

[0090] This invention summarizes elevator status parameters, facilitating access for monitoring personnel and effectively improving the efficiency of elevator operation monitoring. It also efficiently collects, stores, and processes elevator status parameters, identifying elevator status keywords and matching them to elevator fault information, which is then displayed. This improves the real-time performance and accuracy of elevator fault warnings, enabling timely detection and alerts to safety hazards during elevator operation. This enhances the management level and efficiency of elevator operation, controls risks during operation, and ultimately achieves timely and scientific supervision, management, and control of elevator operation.

[0091] The units and algorithm steps of the various examples described in the elevator fault early warning method based on predictive models provided by this invention can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0092] In the elevator fault early warning method of the present invention, computer program code for performing the operations of the present disclosure can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or power server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (exemplarily using an Internet service provider for Internet connection).

[0093] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for early warning of elevator faults based on a predictive model, characterized in that, The methods include: Step S101: Generate and store elevator fault information in the database; Step S102: Obtain elevator state parameters and identify elevator state keywords in the elevator state parameters using the ERNIE pre-trained language model. Step S103: Match the identified elevator status keywords with elevator fault information in the database. Once elevator fault information is matched, display the elevator fault information. In step S102, the ERNIE pre-trained language model is evaluated using the AUC model. The AUC model retrieves evaluation samples, sorts the samples according to the probability prediction results generated by the classification model, and divides the sorted samples into two parts {x1, x2, ... x} according to a certain cut point. m+1 x m+n The first part of m samples is classified as positive examples, and the second part of n samples is classified as negative examples; the AUC is defined as: Where f is the classification function; I is the indicator function, which is 1 when the expression in parentheses is true, and 0 otherwise; the AUC model takes values ​​between [0.5, 1]. The AUC model uses precision, recall, and F1 score as evaluation metrics. Where: precision represents the proportion of samples whose actual class and predicted class are all positive out of all samples whose predicted class is positive; recall represents the proportion of samples whose actual class and predicted class are all positive out of all samples whose actual class is positive; f1score is the weighted harmonic mean of precision and recall; the calculation formulas are shown in (1)-(3); Where: TP indicates that the predicted class is A and the actual class is A; FP indicates that the predicted class is A but the actual class is not A; TN indicates that the predicted class is not A and the actual class is not A; FN indicates that the predicted class is not A but the actual class is A. In step S101, elevator fault information classification conditions are set to classify elevator fault information and form an elevator fault classification table. Label each elevator malfunction; The labels for elevator fault information are matched with the elevator status keywords in the elevator status parameters; The identified elevator status keywords are matched with elevator fault information tags. Once a match is successful, the elevator fault information is displayed. In step S102, an ERNIE pre-trained language model is constructed, which provides task embeddings to represent elevator state keywords in elevator state parameters. Each recognition task is assigned a unique task embedding, and the corresponding elevator state parameters, recognition token, recognition statement, and recognition task are used as inputs to the ERNIE pre-trained language model. The method also validates the elevator state keywords identified by the ERNIE pre-trained language model in the elevator state parameters; The validation method uses binary cross-entropy loss as the loss function for multi-label classification, which is defined as: Where N is the number of elevator status keywords in the elevator status parameters, K is the number of tags for elevator fault information, and y ij ∈{0,1} and Let represent the actual tag value and the probability prediction value of the j-th tag for the i-th elevator status keyword, respectively; The skewed triangular learning rate (STLR) method is used, which involves first linearly increasing the learning rate and then linearly decreasing it according to the training cycle; specifically, it is expressed as: Where T is the total number of training iterations, cut frac The learning rate increases as a percentage of the total training iterations; cut is the iteration number at which the learning rate inflects; p is the scaling ratio at which the learning rate increases or decreases; and ratio is the ratio between the minimum and maximum learning rates, η. max The ratio, η t It is the learning rate of the t-th iteration.

2. The elevator fault early warning method based on a predictive model according to claim 1, characterized in that, In step S102, the ERNIE pre-trained language model obtains the elevator state parameter sequence, embeds a [CLS] marker at the beginning of each elevator state parameter, and sets a separator [SEP] between the elevator state parameters; The ERNIE pre-trained language model uses a multi-layer Transformer as the basic encoder and captures each token in the elevator state parameter sequence through a self-attention mechanism. It automatically assigns each elevator state parameter recognition task to different training stages to identify elevator state keywords.

3. The elevator fault early warning method based on a predictive model according to claim 2, characterized in that, When cut frac When the ratio is equal to 0.1 and the ratio is equal to 32, it has the characteristics of short-term growth and long decay period. The ERNIE pre-trained language model identifies that the elevator state keywords in the elevator state parameters meet the requirements.

Images