An elevator dangerous situation active recognition and intelligent rescue system

Abstract

The present invention discloses an elevator dangerous situation active recognition and intelligent rescue system, belonging to the technical field of elevator monitoring, which solves the problems that the existing system does not perform pre-recognition processing on multi-source heterogeneous data, resulting in an increase in the data processing volume of the data processing module and insufficient accuracy of adverse risk evaluation. The system includes a data acquisition end, a cloud processing platform, a response feedback module, and a visualization terminal. The cloud processing platform identifies and analyzes the monitoring information set based on a dangerous situation recognition model, and calculates the elevator dangerous situation value based on dangerous situation risk factors. In the present invention, the data acquisition end performs pre-recognition on elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, so as to realize distributed pre-recognition processing of data, which is applicable to elevator monitoring scenarios with multi-source heterogeneity and possibly involving data of different subjects, improves the efficiency and accuracy of data processing of the dangerous situation recognition model, and thus ensures the active recognition response speed and comprehensive evaluation accuracy of elevator dangerous situations.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator monitoring, and particularly relates to an elevator dangerous situation active recognition and intelligent rescue system. Background Technique

[0002] With the acceleration of the urbanization process and the increasing number of high-rise buildings, elevators / lifts, as an indispensable vertical transportation means in modern society, their safety and reliability have received extensive attention. However, elevators still face many potential risks during operation, and these risks may cause passengers to be trapped, injured or even endanger their lives.

[0003] When an elevator / lift is operating, common risk types include mechanical failure risks, electrical failure risks, human factor risks, and environmental factor risks. At present, elevator risk identification mainly relies on regular manual inspection and monitoring equipment. Manual inspection has problems such as strong subjectivity, low efficiency, and difficulty in comprehensive coverage, and some potential risks may not be discovered in time.

[0004] The invention patent application with the publication number CN115818388A discloses an elevator operation condition poor risk monitoring system. The elevator control module collects risk assessment data in real time during the process of controlling the elevator operation; the data upload module uploads the risk assessment data to the database, and the data processing module uses the risk assessment data in the database to identify whether there are poor risks and when there are poor risks, pushes the poor risk prompt information to the maintenance terminal, and also identifies whether the poor risks are eliminated according to the maintenance operation information fed back by the maintenance terminal and the risk assessment data after maintenance is completed. However, the existing system only uploads the risk assessment data to the database and does not perform pre-identification processing on multi-source heterogeneous data, resulting in an increase in the data processing volume of the data processing module and insufficient accuracy in judging poor risks. In view of the above problems, an elevator dangerous situation active recognition and intelligent rescue system is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide an elevator dangerous situation active recognition and intelligent rescue system for the deficiencies of the existing technology, and solve the problems that the existing system only uploads the risk assessment data to the database and does not perform pre-identification processing on multi-source heterogeneous data, resulting in an increase in the data processing volume of the data processing module and insufficient accuracy in judging poor risks.

[0006] The present invention is realized as follows. An elevator dangerous situation active recognition and intelligent rescue system, the system includes:

[0007] A data acquisition end, which is used to obtain multi-source heterogeneous elevator monitoring data in real time, integrate and normalize the elevator monitoring data, pre-identify the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and upload the pre-identified monitoring information set of the data to the cloud processing platform;

[0008] A cloud processing platform, which is used to identify and store a monitoring information set, update an elevator operation database based on the monitoring information set, construct and train a danger identification model based on the elevator operation database, the danger identification model identifies and analyzes the monitoring information set, outputs danger risk factors, and calculates an elevator danger value based on the danger risk factors;

[0009] A response feedback module, which responds to the elevator danger value, indexes the danger risk factors in the monitoring information set based on the elevator danger value, triggers an alarm feedback signal based on the danger risk factors, and traverses the elevator operation database to generate an intelligent rescue plan corresponding to the alarm feedback signal;

[0010] A visualization terminal, which is communicatively connected to the data acquisition end, the cloud processing platform, and the response feedback module respectively, and is used to visually present the pre-identification results of elevator monitoring data, danger risk factors, and intelligent rescue plans.

[0011] Preferably, the data acquisition end includes:

[0012] An elevator sensor group, which is distributed in the elevator operation environment, and the elevator sensor group is used to obtain elevator mechanical operation data, electrical operation data, and environmental monitoring data in real time;

[0013] A video monitoring terminal, which is used to obtain real-time monitoring data inside and outside the elevator in real time;

[0014] A front-end controller, which is communicatively connected to the elevator sensor group and the video monitoring terminal, and the front-end controller is used to collect elevator mechanical operation data, electrical operation data, environmental monitoring data, and internal and external real-time monitoring data and process them into elevator monitoring data, and integrate and normalize the elevator monitoring data;

[0015] A multi-layer perception module, which pre-identifies the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and uploads the pre-identified monitoring information set to the cloud processing platform.

[0016] Preferably, the method for pre-identifying the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism includes:

[0017] Load the integrated and normalized elevator monitoring data, assign weights to the data types of the elevator monitoring data by using the principal component analysis method, and decompose and process the elevator monitoring data based on the sliding window method, where the sliding window width is 1.5 seconds and the sliding step is 0.5 seconds;

[0018] Perform spectral analysis on the elevator monitoring data within each sliding window to determine the noise position, quantity, amplitude, and bandwidth in the signal, and judge whether the weight value of the data type corresponding to the signal exceeds the preset filtering threshold. Determine the data filtering algorithm based on the weight value of the data type corresponding to the signal;

[0019] If the weight value of the data type corresponding to the signal does not exceed the preset filtering threshold, perform discrete wavelet decomposition on the elevator monitoring data using the Daubechies filtering algorithm of order DB6;

[0020] If the weight value of the data type corresponding to the signal exceeds the preset filtering threshold, decompose and reconstruct the elevator monitoring data based on the biorthogonal wavelet basis algorithm;

[0021] Integrate the elevator monitoring data after filtering processing, and perform feature type clustering on the elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration to obtain at least one set of sub-cluster centers;

[0022] Among them, when performing feature type clustering on the elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration, the optimization objective function is expressed as:

[0023] (3);

[0024] Among them, represents the optimization objective function of feature type clustering, is the elevator monitoring data and the membership degree matrix of the sub-cluster center , the matrix is order matrix, is the number of sub-cluster centers , is the elevator monitoring data and the Euclidean distance of the sub-cluster center , is the fuzzy exponent;

[0025] The constraint condition of the membership degree matrix is:

[0026] (4);

[0027] Calculate the closeness of the elevator monitoring data and the sub-cluster center based on the relational attention mechanism, and judge whether the elevator monitoring data belongs to the sub-cluster center based on the preset closeness threshold;

[0028] Among them, the closeness of the elevator monitoring data and the sub-cluster center is calculated by the following formula:

[0029] (5);

[0030] Among them, represents the proximity degree, represents the elevator monitoring data the membership degree matrix for the sub-cluster center ; represents the elevator monitoring data for the data type the membership degree matrix;

[0031] Calculate the clustering entropy value of the elevator monitoring data within the sub-cluster center, use the mean of the clustering entropy as the clustering difference judgment threshold between the elevator monitoring data and the sub-cluster center to extract the pre-identified abnormal data in the sub-cluster center and integrate the elevator monitoring data in the sub-cluster center and the pre-identified abnormal data to obtain the monitoring information set. Preferably, when decomposing and reconstructing the elevator monitoring data based on the biorthogonal wavelet basis algorithm, select the CDF5 / 3 wavelet basis and the decomposition level to decompose the elevator monitoring data, perform convolution operations on the two sets of filters through the CDF5 / 3 wavelet basis, and perform upsampling processing. Sequentially decompose the elevator monitoring data of each decomposition layer into approximation coefficients and detail coefficients, integrate at least one set of detail coefficients, determine the shrinkage threshold of the detail coefficients based on the cross-validation criterion combined with the random undersampling algorithm, and set the detail coefficients with coefficient amplitudes less than the shrinkage threshold to 0 to complete the reconstruction of the elevator monitoring data;

[0032]

[0033]

[0034] Among them, the shrinkage threshold is determined by the following formula:

[0035] (1);

[0036] (2); Among them, represents the shrinkage threshold, represents the weight value corresponding to the data type ; is the maximum value of the detail coefficient, is the mean value of the detail coefficient, is the standard deviation of the detail coefficient, is the number of noises in the signal, is the decomposition level, is the initial shrinkage threshold, represents the detail correction value, The undersampling magnification of the noise quantity.

[0037] Preferably, the cloud processing platform includes:

[0038] A database management unit, configured to identify and store the monitoring information set, and update the elevator operation database based on the monitoring information set;

[0039] A model construction unit, configured to construct and train a danger identification model based on the elevator operation database;

[0040] A danger value determination unit, configured to load the danger identification model, identify and analyze the monitoring information set based on the danger identification model, output a danger risk factor, and calculate an elevator danger value based on the danger risk factor.

[0041] Preferably, when the model construction unit constructs and trains a danger identification model based on the elevator operation database, an extreme learning machine model is used as the initial model. The initial model includes an input layer, a hidden layer, and an output layer. The hidden layer is frozen, and an improved Transformer architecture is used to replace the hidden layer. The improved Transformer architecture includes an autoencoder, a feature extraction layer, a feature fusion layer, and a decoder. Both the autoencoder and the decoder are composed of three layers of self-attention mechanism, a feedforward neural network, a global pooling layer, and an average pooling layer. The entropy weight method is introduced into the autoencoder to obtain the objective weight of the feature. The game theory weighting method is introduced into the feature fusion layer to calculate the comprehensive weight of each index. A multi-layer perception mechanism and a nearest neighbor propagation clustering algorithm are introduced into the decoder to calculate the elevator danger value. The feature extraction layer includes 1 BiLSTM layer, 3 attention layers, and 1 CFCs layer, and the feature fusion layer is a federated learning architecture including a softmax layer.

[0042] Preferably, the method for constructing and training a danger identification model based on the elevator operation database includes:

[0043] Load a pre-constructed danger identification model, and preset the connection weights between the input layer, the improved Transformer architecture, and the output layer in the danger identification model, as well as the number of neurons, the activation function, and the loss function;

[0044] Traverse the elevator operation database, import training samples from the elevator operation database, and divide the training samples into a training set and a test set;

[0045] Adopt a feature optimization strategy to preprocess the training set, input the preprocessed training set into the Transformer architecture, iteratively train the Transformer architecture, calculate the risk factor identification accuracy and the danger value accuracy of the Transformer architecture under different activation functions, and select the optimal risk factor identification accuracy and the danger value accuracy as the optimal activation function of the Transformer architecture;

[0046] With the optimal activation function of the Transformer architecture as a prerequisite, set the initial learning rate and the number of training epochs, select the adam optimizer to adaptively adjust the hyperparameters of the Transformer architecture, and output a converged danger recognition model;

[0047] Obtain a test set, use the test set to test the danger recognition model, output the test results, and determine whether the test results meet the preset test threshold. If they meet the preset test threshold, output the converged danger recognition model.

[0048] Preferably, the method for the danger recognition model to identify and analyze the monitoring information set includes:

[0049] Load the monitoring information set. The input layer extracts the pre-identified abnormal data in the monitoring information set. The autoencoder objectively assigns weights to the feature factors of the pre-identified abnormal data based on the entropy weight method to obtain the pre-identified abnormal data with weighted abnormal feature factors;

[0050] The pre-identified abnormal data with weighted abnormal feature factors is propagated through the BiLSTM layer, attention layer, and CFCs layer in the feature extraction layer. The BiLSTM layer and attention layer extract the feature vectors of the pre-identified abnormal data, and the CFCs is used to reduce the feature extraction calculation amount and speed up the inference speed, and output a set of feature vectors;

[0051] Load the set of feature vectors. The feature fusion layer calculates the comprehensive weights of each abnormal feature factor based on the game theory weight assignment method. The federated learning architecture globally and dynamically aggregates the abnormal feature factors based on the comprehensive weights of the abnormal feature factors, and extracts the danger risk factors through the softmax layer, and outputs a feature fusion set containing the danger risk factors;

[0052] Obtain the feature fusion set, input the feature fusion set into the decoder, and the decoder calculates the elevator danger value based on the multi-layer perception mechanism and the nearest neighbor propagation clustering algorithm.

[0053] Preferably, the elevator danger value is calculated by the following formula:

[0054] (6);

[0055] (7);

[0056] Where, is the elevator danger value, represents the activation function of the decoder multi-layer perception mechanism, represents the current danger risk factor 's input representation, represents the current danger risk factor 's comprehensive weight, Indicates the current danger risk factor is the bias term of where the number of danger risk factors Indicates the current danger risk factor based on the affinity propagation clustering algorithm is the cumulative distribution function of is the input mean of the danger risk factor Indicates the damping coefficient of the affinity propagation clustering algorithm Indicates the maximum number of iterations of the affinity propagation clustering algorithm

[0057] Compared with the prior art, the embodiments of the present application mainly have the following beneficial effects:

[0058] In the present invention, the data acquisition end pre-identifies the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, so as to realize the distributed pre-identification processing of the data, which is applicable to the elevator monitoring scenario with multi-source heterogeneous and possibly involving data of different entities, improves the efficiency and accuracy of the data processing of the danger identification model, and the cloud processing platform identifies and analyzes the monitoring information set based on the danger identification model, thus ensuring the active identification response speed and comprehensive evaluation accuracy of the elevator danger, and realizing the distributed pre-identification processing of the data

[0059] In the present invention, when pre-identifying the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, the filtering algorithm is dynamically selected according to the weight value of the data type, and the most suitable filtering method can be adopted for data with different importance levels. This not only improves the filtering effect, but also avoids the limitations of a single filtering algorithm. By flexibly adjusting the filtering strategy, it is ensured that key data is processed more finely, while secondary data is efficiently filtered, improving the overall data processing efficiency

[0060] In the present invention, a danger identification model and a training method are provided. Taking the extreme learning machine model as the initial model, using its advantages of fast learning and high efficiency, an improved Transformer architecture is used to replace the hidden layer. The extreme learning machine has advantages in processing large-scale data and complex non-linear problems, while the Transformer architecture performs well in processing sequence data and capturing long-sequence dependencies. This combination gives full play to the advantages of both, enabling the model to quickly learn data features and effectively process the time series information in the elevator operation data, improving the accuracy of danger identification. The entropy weight method is introduced into the autoencoder to obtain the objective weight of the features. The entropy weight method can determine the weight of the features according to the information entropy of the data, avoiding the influence of subjective factors and making the feature weights more objective and accurate. The game theory weighting method is introduced into the feature fusion layer to calculate the comprehensive weight of each index. The game theory weighting method can consider the mutual relationship and influence between different indexes, and determine the optimal weight of each index through the game, so as to more comprehensively and accurately reflect the features and risk factors of the elevator operation data

[0061] In the present invention, an intelligent rescue plan is automatically generated according to the real-time monitored danger value, which avoids the subjectivity and uncertainty of manual judgment, ensures that the rescue measures match the actual danger of the elevator, improves the effectiveness and success rate of the rescue, and maximally guarantees the life safety of passengers and the normal operation of elevator equipment. The full-process automated processing from data acquisition, pre-identification, danger analysis to rescue plan generation greatly shortens the response time, can promptly respond to elevator dangers, reduces the time that passengers are trapped due to elevator failures, reduces the losses and impacts caused by accidents, and improves the user experience and social satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a schematic structural diagram of an elevator danger active identification and intelligent rescue system provided by the present invention;

[0063] Figure 2 shows a schematic implementation flow diagram of a method for pre-identifying elevator monitoring data based on a federated learning algorithm combined with a relational attention mechanism;

[0064] Figure 3 shows a schematic implementation flow diagram of a method for constructing and training a danger identification model based on an elevator operation database;

[0065] Figure 4 shows a schematic implementation flow diagram of a method for the danger identification model to identify and analyze a monitoring information set.

[0066] In the figure: 100, data acquisition end; 110, elevator sensor group; 120, video monitoring terminal; 130, front-end controller; 140, multi-layer perception module; 200, cloud processing platform; 210, database management unit; 220, model construction unit; 230, danger value determination unit; 300, response feedback module; 400, visualization terminal. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of this application or the above drawings are used to distinguish different objects and not to describe a specific order.

[0068] The existing system only uploads risk assessment data to the database and does not perform pre-identification processing on multi-source heterogeneous data, resulting in an increased data processing volume of the data processing module and insufficient accuracy in judging bad risks. In view of the above problems, an elevator emergency active identification and intelligent rescue system is proposed. The system includes a data acquisition terminal 100, a cloud processing platform 200, a response feedback module 300, and a visualization terminal 400. When working, first, the data acquisition terminal 100 acquires multi-source heterogeneous elevator monitoring data in real time, integrates and normalizes the elevator monitoring data, and pre-identifies the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism. The cloud processing platform 200 constructs and trains an emergency identification model based on the elevator operation database. The emergency identification model identifies and analyzes the monitoring information set, outputs the emergency risk factors, calculates the elevator emergency value based on the emergency risk factors, and finally, the response feedback module 300 responds to the elevator emergency value, indexes the emergency risk factors in the monitoring information set based on the elevator emergency value, triggers an alarm feedback signal based on the emergency risk factors, and traverses the elevator operation database to generate an intelligent rescue plan corresponding to the alarm feedback signal.

[0069] In the embodiment of the present invention, the data acquisition terminal 100 pre-identifies the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, thereby realizing distributed pre-identification processing of data, which is applicable to elevator monitoring scenarios with multi-source heterogeneity and may involve data of different entities, improving the efficiency and accuracy of data processing of the emergency identification model. Moreover, the cloud processing platform 200 identifies and analyzes the monitoring information set based on the emergency identification model, thereby ensuring the response speed and comprehensive evaluation accuracy of the active identification of elevator emergencies and realizing distributed pre-identification processing of data. It overcomes the problems of the existing system that only uploads risk assessment data to the database and does not perform pre-identification processing on multi-source heterogeneous data, resulting in an increased data processing volume of the data processing module and insufficient accuracy in judging bad risks.

[0070] The embodiment of the present invention provides an elevator emergency active identification and intelligent rescue system. Figure 1 The implementation process schematic diagram of the elevator emergency active identification and intelligent rescue system is shown. The elevator emergency active identification and intelligent rescue system includes:

[0071] A data acquisition terminal 100, configured to acquire multi-source heterogeneous elevator monitoring data in real time, integrate and normalize the elevator monitoring data, pre-identify the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and upload the pre-identified monitoring information set to the cloud processing platform 200;

[0072] Wherein, the data acquisition terminal 100 includes:

[0073] Elevator sensor group 110, which is distributed in the elevator operating environment. The elevator sensor group 110 is used to obtain elevator mechanical operation data, electrical operation data, and environmental monitoring data in real time;

[0074] It should be noted that the elevator sensor group 110 includes, but is not limited to, acceleration sensors, gravity sensors, vibration sensors, temperature sensors, humidity sensors, wind speed sensors, etc. The elevator sensor group 110 integrates multiple monitoring functions. In addition to monitoring the basic operating parameters of the elevator (such as speed, position, door status, etc.), it can also diagnose and analyze elevator faults. It can judge whether the elevator is in a normal operating state according to preset algorithms and rules, and take timely measures, such as alarming, stopping operation, etc., so as to obtain multi-source heterogeneous elevator mechanical operation data, electrical operation data, and environmental monitoring data.

[0075] Video monitoring terminal 120, which is used to obtain real-time monitoring data inside and outside the elevator in real time. It should be noted that the video monitoring terminal 120 is installed in positions such as the elevator car and the elevator shaft, and monitors the situation inside the elevator and the opening and closing status of the elevator door in real time. Through video image analysis technology, some abnormal behaviors can also be identified, such as passengers fighting, whether there are foreign objects in the car blocking the elevator door from closing, etc. The video monitoring terminal 120 can be a high-definition network camera;

[0076] Front-end controller 130, which is respectively communicatively connected to the elevator sensor group 110 and the video monitoring terminal 120. The front-end controller 130 is used to collect elevator mechanical operation data, electrical operation data, environmental monitoring data, and internal and external real-time monitoring data, process them into elevator monitoring data, and integrate and normalize the elevator monitoring data;

[0077] Multi-layer perception module 140, which is used to pre-identify elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and upload the pre-identified monitoring information set to the cloud processing platform 200.

[0078] Cloud processing platform 200, which is used to identify and store the monitoring information set, update the elevator operation database based on the monitoring information set, build and train a danger recognition model based on the elevator operation database. The danger recognition model identifies and analyzes the monitoring information set, outputs danger risk factors, and calculates the elevator danger value based on the danger risk factors;

[0079] Response feedback module 300, the response feedback module 300 responds to the elevator danger value, indexes the danger risk factors in the monitoring information set based on the elevator danger value, triggers an alarm feedback signal based on the danger risk factors, and traverses the elevator operation database to generate an intelligent rescue plan corresponding to the alarm feedback signal;

[0080] It should be noted that the response feedback module 300 first receives the elevator danger value calculated by the cloud processing platform 200. The elevator danger value is a quantitative representation of the current operation risk of the elevator. The elevator danger value is calculated based on the analysis and processing of the danger identification and monitoring information set in the cloud processing platform 200 and in combination with the danger risk factors. The response feedback module 300 will judge whether the elevator is in a dangerous state and the severity of the danger according to the elevator danger value. The danger risk factors are output by the danger identification model of the cloud processing platform 200 after analyzing the monitoring information set. The danger risk factors include various key information and characteristics related to the elevator danger. By indexing the danger risk factors, the response feedback module 300 can clarify the specific factors leading to the elevator danger, provide a basis for subsequent alarm and rescue plan generation. According to the indexed danger risk factors and the corresponding elevator danger value, the response feedback module 300 will trigger an alarm feedback signal. The triggering of the alarm feedback signal is based on preset thresholds and rules. While triggering the alarm feedback signal, the response feedback module 300 will traverse the elevator operation database. A large amount of elevator operation data, historical fault records, maintenance information, and corresponding rescue measures are stored in the elevator operation database. The response feedback module 300 will search for matching historical data and rescue experience in the elevator operation database according to the indexed danger risk factors, so as to generate an intelligent rescue plan corresponding to the current danger situation.

[0081] In the embodiment of the present invention, an intelligent rescue plan is automatically generated according to the real-time monitored elevator danger value, which avoids the subjectivity and uncertainty of manual judgment, ensures that the rescue measures match the actual elevator danger, improves the effectiveness and success rate of the rescue, and maximally guarantees the life safety of passengers and the normal operation of elevator equipment. The full-process automated processing from data acquisition, pre-identification, danger analysis to rescue plan generation greatly shortens the response time, can promptly respond to the elevator danger, reduces the trapped time of passengers caused by elevator failures, reduces the losses and impacts caused by accidents, and improves the user experience and social satisfaction.

[0082] Visualization terminal 400, the visualization terminal 400 is respectively communicatively connected to the data acquisition end 100, the cloud processing platform 200, and the response feedback module 300, and is used to visually present the pre-identification results of elevator monitoring data, danger risk factors, and intelligent rescue plans.

[0083] It should be noted that the visualization terminal 400 can be a terminal with communication functions such as a smart phone or a tablet computer. The visualization terminal 400 is communicatively connected to the data acquisition end 100, the cloud processing platform 200, and the response feedback module 300 through a local area network, 5G, or Bluetooth.

[0084] In the embodiment of the present invention, the data acquisition end 100 pre-identifies the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, so as to realize the distributed pre-identification processing of the data. It is applicable to the elevator monitoring scenario with multi-source heterogeneous and possibly different subject data, improves the efficiency and accuracy of the data processing of the danger identification model, and the cloud processing platform 200 identifies and analyzes the monitoring information set based on the danger identification model, thus ensuring the active identification response speed and comprehensive evaluation accuracy of the elevator danger, and realizing the distributed pre-identification processing of the data.

[0085] The embodiment of the present invention provides a method for pre-identifying elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism. Figure 2 The schematic diagram of the implementation process of the method for pre-identifying elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism is shown. The method for pre-identifying elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism includes:

[0086] Step S101: Load and integrate the normalized elevator monitoring data, assign weights to the data types of the elevator monitoring data by using the principal component analysis method, and decompose the elevator monitoring data based on the sliding window method, where the width of the sliding window is 1.5 seconds and the sliding step is 0.5 seconds;

[0087] In this embodiment, the elevator monitoring data includes but is not limited to basic operation data, running direction, speed, position, floor information, load capacity, fault code, fault time, temperature, humidity, maintenance time / content, top collision, bottom squat fault data, trapped person data, and real-time monitoring data. Through the integration and normalization process, the multi-source heterogeneous elevator monitoring data is converted into a unified format and dimension. This helps to eliminate the differences between different data sources, improve the consistency and comparability of the data, and provide a more reliable data basis for subsequent analysis and processing. Assigning weights to the data types by using the principal component analysis method can allocate different weights according to the importance and relevance of the data. This helps to highlight the key data features, reduce the interference of noise data, and improve the efficiency and accuracy of data processing.

[0088] Step S102: Perform spectrum analysis on the elevator monitoring data in each sliding window to determine the noise position, quantity, amplitude, and bandwidth in the signal;

[0089] It should be noted that by setting an appropriate sliding window width and step size (such as a 1.5-second window width and a 0.5-second step size), real-time monitoring and dynamic analysis of the elevator operation status can be achieved, abnormal situations can be detected in a timely manner, and by performing spectral analysis on the data within each sliding window, the noise position, quantity, amplitude, and bandwidth in the signal can be accurately determined. This helps to identify and separate noise signals and provides a basis for subsequent data filtering.

[0090] Step S103: Determine whether the weight value of the signal corresponding data type exceeds a preset filtering threshold. The preset filtering threshold can be 0.2 - 0.5, and determine the data filtering algorithm based on the weight value of the signal corresponding data type;

[0091] In the embodiment of the present invention, when pre-identifying elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, the filtering algorithm is dynamically selected according to the weight value of the data type, and the most suitable filtering method can be adopted for data of different importance levels. This not only improves the filtering effect but also avoids the limitations of a single filtering algorithm. By flexibly adjusting the filtering strategy, it ensures more refined processing of key data, while efficiently filtering secondary data, thereby improving the overall data processing efficiency.

[0092] Step S104: If the weight value of the signal corresponding data type does not exceed the preset filtering threshold, perform discrete wavelet decomposition on the elevator monitoring data using the Daubechies filtering algorithm of order DB6;

[0093] When the weight value of the signal corresponding data type does not exceed the preset filtering threshold, perform discrete wavelet decomposition on the elevator monitoring data using the Daubechies filtering algorithm of order DB6. Daubechies wavelets have good compact support and orthogonality, and can effectively remove noise while better retaining the local features and time-frequency information of the data. The selection of order DB6 is a balance between computational complexity and filtering effect, which can not only ensure effective suppression of noise but also avoid losing important data features due to overprocessing, thus providing a high-quality data basis for subsequent feature extraction and analysis.

[0094] Step S105: If the weight value of the signal corresponding data type exceeds the preset filtering threshold, decompose and reconstruct the elevator monitoring data based on the biorthogonal wavelet basis algorithm;

[0095] When the weight value of the signal corresponding data type exceeds the preset filtering threshold, decompose and reconstruct the elevator monitoring data based on the biorthogonal wavelet basis algorithm. Selecting the CDF5 / 3 wavelet basis has the advantages of simple calculation, fast speed, and symmetry, and can improve the calculation efficiency while ensuring the filtering effect, especially suitable for elevator monitoring scenarios with high real-time requirements.

[0096] Among them, when decomposing and reconstructing elevator monitoring data based on the biorthogonal wavelet basis algorithm, the CDF5 / 3 wavelet basis and the decomposition level are selected to decompose the elevator monitoring data. The CDF5 / 3 wavelet basis is used for convolution operation of two groups of filters, and upsampling processing is performed. The elevator monitoring data of each decomposition level is decomposed into approximation coefficients and detail coefficients in turn. At least one group of detail coefficients is integrated. Based on the cross-validation criterion and the random undersampling algorithm, the shrinkage threshold of the detail coefficients is determined. The detail coefficients with coefficient amplitudes less than the shrinkage threshold are set to 0 to complete the reconstruction of the elevator monitoring data;

[0097] Among them, the shrinkage threshold is determined by the following formula:

[0098] (1);

[0099] (2);

[0100] Among them, represents the shrinkage threshold; represents the weight value of the corresponding data type ; is the maximum value of the detail coefficients; is the mean value of the detail coefficients; is the standard deviation of the detail coefficients; is the number of noises in the signal; is the decomposition level, which can be 2 - 5; is the initial shrinkage threshold, which can be preset to 0.1 - 0.5; represents the detail correction value; is the undersampling multiple of the number of noises.

[0101] Step S106: Integrate the elevator monitoring data after filtering processing, and perform feature type clustering on the elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration to obtain at least one group of sub-cluster centers;

[0102] Among them, when performing feature type clustering on the elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration, the optimization objective function is expressed as:

[0103] (3);

[0104] Among them, represents the optimization objective function of feature type clustering, is the elevator monitoring data and the sub-cluster center 's membership matrix, and the matrix is order matrix, is the sub-cluster center 's quantity, is the elevator monitoring data The Euclidean distance from the sub-cluster center is, where is the fuzziness index, which can be set to 2 - 4 in this embodiment;

[0105] The membership matrix is subject to the constraint:

[0106] (4);

[0107] In the embodiment of the present invention, fuzzy clustering can handle the uncertainty and fuzziness of data, making the clustering results more in line with the actual situation. By this way, at least one set of sub-cluster centers can be obtained, which can more accurately reflect the characteristic distribution of elevator monitoring data and provide a more reliable basis for subsequent abnormal data identification.

[0108] Step S107: Calculate the closeness between the elevator monitoring data and the sub-cluster center , and judge whether the elevator monitoring data belongs to the sub-cluster center based on a preset closeness threshold;

[0109] Among them, the closeness between the elevator monitoring data and the sub-cluster center is calculated by the following formula:

[0110] (5);

[0111] Among them, represents the closeness, represents the membership matrix of the elevator monitoring data to the sub-cluster center , represents the membership matrix of the elevator monitoring data to the data type ;

[0112] Step S108: Calculate the clustering entropy value of the elevator monitoring data inside the sub-cluster center , use the mean value of the clustering entropy as the clustering difference judgment threshold between the elevator monitoring data and the sub-cluster center , extract the pre-identified abnormal data in the sub-cluster center , and integrate the elevator monitoring data and the pre-identified abnormal data in the sub-cluster center to obtain a monitoring information set.

[0113] In the embodiment of the present invention, by calculating the cluster entropy value of the data in the sub-cluster center and using the cluster entropy mean as the judgment threshold, the pre-identified abnormal data can be effectively identified. This judgment method based on cluster differences can improve the accuracy of abnormal data extraction. By calculating the cluster entropy value, the distribution characteristics of the data can be quantified, so as to more accurately identify abnormal data that is significantly different from normal data. Integrating the data in the sub-cluster center and the abnormal data to form a monitoring information set provides more comprehensive and accurate data support for subsequent crisis identification and analysis.

[0114] On the other hand, when pre-identifying elevator monitoring data, the introduction of the relational attention mechanism can dynamically adjust the weights between data features, highlight key features, and reduce noise interference. In elevator monitoring, this means that features related to elevator safety, such as elevator running speed, acceleration, vibration signals, etc., can be more accurately identified. In combination with federated learning and the relational attention mechanism, multi-source heterogeneous data can be effectively extracted and fused without the need to concentrate all data on a central node for processing. This distributed processing method makes full use of the computing resources of each node, realizes parallel processing of data, greatly shortens the time for data pre-identification, and improves processing efficiency. At the same time, when facing large-scale elevator monitoring data, distributed processing can effectively avoid single point failures and data transmission bottlenecks, ensure the stability and reliability of the system, and by integrating a variety of advanced technologies such as federated learning algorithms, relational attention mechanisms, and biorthogonal wavelet basis algorithms, the model can mine personalized features of elevator monitoring data from different angles. The federated learning algorithm can learn the distributed features of different elevator data, the relational attention mechanism can focus on key features, and the biorthogonal wavelet basis algorithm can refine the data. The synergy of these technologies enables the model to understand the operating status of the elevator more comprehensively and deeply, providing strong support for personalized pre-identification.

[0115] In this embodiment, the cloud processing platform 200 includes:

[0116] A database management unit 210, used to identify and store monitoring information sets, and update the elevator operation database based on the monitoring information sets;

[0117] A model building unit 220, used to build and train a danger recognition model based on an elevator operation database;

[0118] The danger value determination unit 230 is used to load the danger recognition model, identify and analyze the monitoring information set based on the danger recognition model, output the danger risk factor, and calculate the elevator danger value based on the danger risk factor.

[0119] It should be noted that the database management unit 210, the model construction unit 220, and the critical value determination unit 230 are communicatively connected via Bluetooth or a local area network, and the elevator operation database stores elevator maintenance logs, operation logs, monitoring information, and intelligent rescue plans.

[0120] The embodiment of the present invention provides a method for constructing and training a critical situation recognition model based on an elevator operation database. Figure 3 The schematic diagram of the implementation process of the method for constructing and training a critical situation recognition model based on an elevator operation database is shown. The method for constructing and training a critical situation recognition model based on an elevator operation database includes:

[0121] Step S201: Load a pre-constructed critical situation recognition model, the connection weights between the input layer, the improved Transformer architecture, and the output layer in the preset critical situation recognition model, as well as the number of neurons, activation function, and loss function.

[0122] Step S202: Traverse the elevator operation database, import training samples from the elevator operation database, and divide the training samples into a training set and a test set. The ratio of the training set to the test set can be 4:1.

[0123] Step S203: Preprocess the training set using a feature selection strategy, input the preprocessed training set into the Transformer architecture, iteratively train the Transformer architecture, calculate the risk factor recognition accuracy and critical value accuracy of the Transformer architecture under different activation functions, and select the optimal risk factor recognition accuracy and critical value accuracy as the optimal activation function of the Transformer architecture.

[0124] Step S204: Based on the optimal activation function of the Transformer architecture as a prerequisite, set the initial learning rate and the number of training epochs. In this embodiment, the optimal activation function of the Transformer architecture can be a hard threshold function, the initial learning rate is set to 0.01, and the number of training epochs is set to 100 - 150 times. Select the adam optimizer to adaptively adjust the hyperparameters of the Transformer architecture, and output a converged critical situation recognition model. This optimization strategy can dynamically adjust the learning rate and hyperparameters according to the training situation of the model, improving the convergence speed and stability of the model. The Adam optimizer combines the advantages of Adagrad and Adadelta, can adaptively adjust the learning rate of each parameter, avoids the cumbersome and difficult manual adjustment of the learning rate, and can also converge to the optimal solution faster, outputting a converged critical situation recognition model, improving the practicality and reliability of the model.

[0125] Step S205: Obtain the test set, use the test set to test the critical situation recognition model, and output the test results.

[0126] Step S206: Determine whether the test result meets a preset test threshold, where the test threshold can be 0.85 - 0.9;

[0127] Step S207: If the preset test threshold is met, output the converged danger recognition model.

[0128] If it is determined that the test result does not meet the preset test threshold, return to step S203 to continue iteratively training the model. This test and evaluation mechanism can promptly discover the problems and deficiencies existing in the model and ensure that the performance of the model meets the expected requirements. If the test result does not meet the preset threshold, return to step S203 to continue iteratively training the model. This iterative optimization process can continuously improve and perfect the model, gradually enhancing the accuracy and generalization ability of the model until the preset test threshold is met, thereby obtaining a danger recognition model with excellent performance.

[0129] In this embodiment, when the model construction unit 220 constructs and trains the danger recognition model based on the elevator operation database, an extreme learning machine model is used as the initial model. The initial model includes an input layer, a hidden layer, and an output layer. The hidden layer is frozen, and an improved Transformer architecture is used to replace the hidden layer. The improved Transformer architecture includes an autoencoder, a feature extraction layer, a feature fusion layer, and a decoder. Both the autoencoder and the decoder are composed of three layers of self-attention mechanisms, feed-forward neural networks, global pooling layers, and average pooling layers. The self-attention mechanism can automatically focus on the important parts in the data, capture the long-sequence dependencies of the data, and improve the model's understanding and processing ability for complex data. The feed-forward neural network can perform non-linear transformations on the data to enhance the model's expressive ability. The global pooling layer and the average pooling layer can perform dimensionality reduction and feature extraction on the data, reduce the data volume, and improve the computational efficiency and generalization ability of the model. The entropy weight method is introduced into the autoencoder to obtain objective feature weights, the game theory weighting method is introduced into the feature fusion layer to calculate the comprehensive weights of each index, and a multi-layer perception mechanism and a nearest neighbor propagation clustering algorithm are introduced into the decoder to calculate the elevator danger value. The feature extraction layer includes 1 BiLSTM layer, 3 attention layers, and 1 CFCs layer, and the feature fusion layer is a federated learning architecture including a softmax layer.

[0130] In the embodiments of the present invention, a danger recognition model and a training method are provided. Taking the extreme learning machine model as the initial model, making use of its advantages of fast learning and high efficiency, an improved Transformer architecture is used to replace the hidden layer. The extreme learning machine has advantages in dealing with large-scale data and complex non-linear problems, while the Transformer architecture performs excellently in dealing with sequence data and capturing long-sequence dependency relationships. This combination fully exploits the advantages of both, enabling the model to not only quickly learn data features but also effectively process the temporal information in elevator operation data, improving the accuracy of danger recognition. The entropy weight method is introduced into the autoencoder to obtain objective weights of features. The entropy weight method can determine the weights of features according to the information entropy of the data, avoiding the influence of subjective factors and making the feature weights more objective and accurate. The game theory weighting method is introduced into the feature fusion layer to calculate the comprehensive weights of each index. The game theory weighting method can consider the mutual relationships and influences between different indexes and determine the optimal weights of each index through a game, so as to more comprehensively and accurately reflect the characteristics and risk factors of elevator operation data.

[0131] The embodiments of the present invention provide a method for the danger recognition model to identify and analyze a monitoring information set. Figure 4 The schematic diagram of the implementation process of the method for the danger recognition model to identify and analyze a monitoring information set is shown. The method for the danger recognition model to identify and analyze a monitoring information set includes:

[0132] Step S301: Load the monitoring information set. The input layer extracts the pre-identified abnormal data in the monitoring information set. The autoencoder assigns objective weights to the feature factors of the pre-identified abnormal data based on the entropy weight method to obtain the pre-identified abnormal data with weighted abnormal feature factors.

[0133] Step S302: The pre-identified abnormal data with weighted abnormal feature factors is propagated through the BiLSTM layer, attention layer, and CFCs layer in the feature extraction layer. The BiLSTM layer and the attention layer extract the feature vectors of the pre-identified abnormal data. The CFCs layer is used to reduce the computational amount of feature extraction and accelerate the inference speed, and outputs a set of feature vectors.

[0134] In this embodiment, the feature extraction layer includes 1 BiLSTM layer, 3 attention layers, and 1 CFCs layer. The BiLSTM layer can process bidirectional sequence data and capture the time dependency relationships and context information in elevator operation data. The attention layer can automatically focus on the important parts in the data and improve the model's attention to key features. The CFCs layer can further enhance the model's feature extraction ability, extract richer and more representative features, and provide strong support for subsequent danger recognition.

[0135] Step S303: Load the feature vector set. The feature fusion layer calculates the comprehensive weights of each abnormal feature factor based on the game theory weighting method. The federated learning architecture globally and dynamically aggregates the abnormal feature factors based on the comprehensive weights of the abnormal feature factors, and extracts the crisis risk factors through the softmax layer, outputting a feature fusion set containing the crisis risk factors. Extracting the crisis risk factors through the softmax layer can convert the information after feature fusion into specific crisis risk indicators, providing a quantitative basis for subsequent crisis assessment;

[0136] Step S304: Obtain the feature fusion set and input the feature fusion set into the decoder. The decoder calculates the elevator crisis value based on the multi-layer perception mechanism and the affinity propagation clustering algorithm.

[0137] It should be noted that the multi-layer perception mechanism can perform multi-level feature extraction and representation on data, enhancing the learning ability and generalization ability of the model. The affinity propagation clustering algorithm can cluster data according to the similarity between data, thereby more accurately identifying the crisis state of the elevator and providing strong support for subsequent rescue and decision-making.

[0138] In this embodiment, the elevator crisis value is calculated by the following formula:

[0139] (6);

[0140] (7);

[0141] Where, is the elevator crisis value, represents the activation function of the decoder multi-layer perception mechanism, represents the input representation of the current crisis risk factor , represents the comprehensive weight of the current crisis risk factor , represents the bias term of the current crisis risk factor , is the number of crisis risk factors, represents the cumulative distribution function of the current crisis risk factor based on the affinity propagation clustering algorithm, is the input mean of the crisis risk factor, represents the damping coefficient of the affinity propagation clustering algorithm, represents the maximum number of iterations of the affinity propagation clustering algorithm. In this embodiment, the damping coefficient can be 0.05 - 0.2, and the maximum number of iterations can be 5 - 8.

[0142] In the embodiments of the present invention, the multi-layer perception mechanism in the decoder can perceive and analyze the operating state of the elevator from multiple levels, further improving the ability to identify elevator emergencies. The affinity propagation clustering algorithm can cluster and classify data according to the similarity between data, enabling the model to accurately judge the similarity between the current elevator state and known emergency patterns, and thus calculate a more accurate elevator emergency value. The elevator emergency value calculated in this way can intuitively reflect the safety status of the elevator and provide an important reference basis for the maintenance and management of the elevator.

[0143] In summary, the present invention provides an elevator emergency active identification and intelligent rescue system. In the embodiments of the present invention, the data acquisition end 100 pre-identifies elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, thereby realizing distributed pre-identification processing of data. It is applicable to elevator monitoring scenarios with multi-source heterogeneous data that may involve data of different entities, improving the efficiency and accuracy of data processing of the emergency identification model. And the cloud processing platform 200 identifies and analyzes the monitoring information set based on the emergency identification model, thereby ensuring the active identification response speed and comprehensive evaluation accuracy of elevator emergencies and realizing distributed pre-identification processing of data.

[0144] It should be noted that for the foregoing embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps may be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the protection scope of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still, without conflict, make non-creative efforts to combine, add, delete or make other adjustments to the features in the embodiments of the present invention according to the situation, so as to obtain different technical solutions that are essentially not divorced from the concept of the present invention, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. An elevator dangerous situation active recognition and intelligent rescue system, characterized in that The system includes: A data acquisition terminal, which is used to acquire multi-source heterogeneous elevator monitoring data in real time, integrate and normalize the elevator monitoring data, pre-identify the elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and upload the pre-identified monitoring information set to the cloud processing platform; A cloud processing platform, which is used to identify and store the monitoring information set, update the elevator operation database based on the monitoring information set, construct and train a danger identification model based on the elevator operation database, the danger identification model identifies and analyzes the monitoring information set, outputs danger risk factors, and calculates the elevator danger value based on the danger risk factors; A response feedback module, which responds to the elevator danger value, indexes the danger risk factors in the monitoring information set based on the elevator danger value, triggers an alarm feedback signal based on the danger risk factors, and traverses the elevator operation database to generate an intelligent rescue plan corresponding to the alarm feedback signal; A visualization terminal, which is communicatively connected to the data acquisition terminal, the cloud processing platform, and the response feedback module respectively, and is used to visually present the pre-identification results of the elevator monitoring data, the danger risk factors, and the intelligent rescue plan; When the model construction unit constructs and trains a danger identification model based on the elevator operation database, it uses the extreme learning machine model as the initial model. The initial model includes an input layer, a hidden layer, and an output layer. The hidden layer is frozen, and the improved Transformer architecture is used to replace the hidden layer. The improved Transformer architecture includes an autoencoder, a feature extraction layer, a feature fusion layer, and a decoder. Both the autoencoder and the decoder are composed of three layers of self-attention mechanism, a feed-forward neural network, a global pooling layer, and an average pooling layer. The entropy weight method is introduced into the autoencoder to obtain the objective weight of the features. The game theory weighting method is introduced into the feature fusion layer to calculate the comprehensive weight of each index. The multi-layer perception mechanism and the affinity propagation clustering algorithm are introduced into the decoder to calculate the elevator danger value. The feature extraction layer includes 1 BiLSTM layer, 3 attention layers, and 1 CFCs layer. The feature fusion layer is a federated learning architecture including a softmax layer; The method for the danger identification model to identify and analyze the monitoring information set includes: Loading the monitoring information set, the input layer extracts the pre-identified abnormal data in the monitoring information set, and the autoencoder assigns objective weights to the feature factors of the pre-identified abnormal data based on the entropy weight method to obtain the pre-identified abnormal data with weighted abnormal feature factors; The pre-identified abnormal data with weighted abnormal feature factors is propagated through the BiLSTM layer, the attention layer, and the CFCs layer in the feature extraction layer. The BiLSTM layer and the attention layer extract the feature vectors of the pre-identified abnormal data and output a set of feature vectors; Loading the set of feature vectors, the feature fusion layer calculates the comprehensive weight of each abnormal feature factor based on the game theory weighting method. The federated learning architecture globally and dynamically aggregates the abnormal feature factors based on the comprehensive weight of the abnormal feature factors, and extracts the danger risk factors through the softmax layer, and outputs a feature fusion set containing the danger risk factors; Obtaining the feature fusion set, inputting the feature fusion set into the decoder, and the decoder calculates the elevator danger value based on the multi-layer perception mechanism and the affinity propagation clustering algorithm.

2. The elevator dangerous situation active recognition and intelligent rescue system according to claim 1, wherein: The data acquisition terminal includes: An elevator sensor group, which is distributed in the elevator operating environment. The elevator sensor group is used to obtain elevator mechanical operation data, electrical operation data, and environmental monitoring data in real time; A video monitoring terminal, which is used to obtain real-time monitoring data inside and outside the elevator in real time; A front-end controller, which is communicatively connected to the elevator sensor group and the video monitoring terminal. The front-end controller is used to collect elevator mechanical operation data, electrical operation data, environmental monitoring data, and real-time monitoring data inside and outside, process them into elevator monitoring data, and integrate and normalize the elevator monitoring data; A multi-layer perception module, which pre-identifies elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism, and uploads the pre-identified monitoring information set to the cloud processing platform.

3. The elevator dangerous situation active recognition and intelligent rescue system according to claim 2, characterized in that: The method for pre-identifying elevator monitoring data based on the federated learning algorithm combined with the relational attention mechanism includes: Loading the integrated and normalized elevator monitoring data, assigning weights to the data types of the elevator monitoring data using the principal component analysis method, and decomposing the elevator monitoring data based on the sliding window method, where the sliding window width is 1.5 seconds and the sliding step is 0.5 seconds; Performing spectral analysis on the elevator monitoring data within each sliding window, determining the noise position, quantity, amplitude, and bandwidth in the signal, judging whether the weight value of the signal corresponding data type exceeds the preset filtering threshold, and determining the data filtering algorithm based on the weight value of the signal corresponding data type; If the weight value of the signal corresponding data type does not exceed the preset filtering threshold, perform discrete wavelet decomposition on the elevator monitoring data using the Daubechies filtering algorithm of order DB6; If the weight value of the signal corresponding data type exceeds the preset filtering threshold, decompose and reconstruct the elevator monitoring data based on the biorthogonal wavelet basis algorithm; Integrate the filtered elevator monitoring data, and perform feature type clustering on the elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration to obtain at least one group of sub-cluster centers; Among them, when performing feature type clustering on elevator monitoring data based on the federated learning algorithm combined with fuzzy clustering iteration, the optimization objective function is expressed as: （3）； Among them, represents the optimization objective function for feature type clustering, is the elevator monitoring data and the membership degree matrix with the sub-cluster center . The matrix is an - order matrix, is the number of sub-cluster centers , is the Euclidean distance between the elevator monitoring data and the sub-cluster center , is the fuzzy index; Membership matrix The constraint conditions are as follows: （4）； Calculating elevator monitoring data based on relational attention mechanism and the sub-cluster center closeness, and judging whether the elevator monitoring data belongs to the sub-cluster center based on a preset closeness threshold ; Among them, the elevator monitoring data and the sub-cluster center The proximity degree is calculated by the following formula: （5）； Among them, represents the closeness degree, represents the elevator monitoring data to the sub-cluster center membership degree matrix, represents the elevator monitoring data to the data type membership degree matrix; Calculate the sub-cluster center Internal elevator monitoring data The clustering entropy value of, and use the mean value of the clustering entropy as the elevator monitoring data And the sub-cluster center The clustering difference judgment threshold of, and extract the sub-cluster center The pre-identified abnormal data in, and integrate the sub-cluster center The elevator monitoring data in The pre-identified abnormal data to obtain the monitoring information set.

4. The elevator dangerous situation active recognition and intelligent rescue system according to claim 3, characterized in that: When decomposing and reconstructing elevator monitoring data based on the biorthogonal wavelet basis algorithm, select the CDF5 / 3 wavelet basis and the decomposition level to decompose the elevator monitoring data, perform convolution operations on two groups of filters through the CDF5 / 3 wavelet basis, and perform upsampling processing. Sequentially decompose the elevator monitoring data at each decomposition layer into approximation coefficients and detail coefficients, integrate at least one group of detail coefficients, determine the shrinkage threshold of the detail coefficients based on the cross-validation criterion combined with the random undersampling algorithm, and set the detail coefficients with coefficient amplitudes less than the shrinkage threshold to 0 to complete the reconstruction of the elevator monitoring data; Among them, the shrinkage threshold is determined by the following formula: （1）； （2）； Among them, represents the shrinkage threshold, represents the weight value corresponding to the data type ; is the maximum value of the detail coefficient, is the mean value of the detail coefficient, is the standard deviation of the detail coefficient, is the number of noises in the signal, is the decomposition level, is the initial shrinkage threshold, represents the detail correction value, is the undersampling multiple of the number of noises. It should be noted that there was a semicolon (;) missing in the original Chinese text at line 6 which has been added in the translation for correct grammar. If this is not allowed according to the strict rules, please adjust accordingly.

5. The elevator dangerous situation active recognition and intelligent rescue system according to claim 1, characterized in that: The cloud processing platform includes: A database management unit, which is used to identify and store the monitoring information set, and update the elevator operation database based on the monitoring information set; A model construction unit, which is used to construct and train a danger identification model based on the elevator operation database; A danger value determination unit is configured to load a danger recognition model, identify and analyze a monitoring information set based on the danger recognition model, output danger risk factors, and calculate an elevator danger value based on the danger risk factors.

6. The elevator dangerous situation active recognition and intelligent rescue system according to claim 5, characterized in that: A method for constructing and training a danger recognition model based on an elevator operation database includes: Loading a pre-constructed danger recognition model, presetting connection weights between an input layer, an improved Transformer architecture, and an output layer in the danger recognition model, as well as the number of neurons, activation functions, and loss functions; Traversing the elevator operation database, importing training samples from the elevator operation database, and dividing the training samples into a training set and a test set; Preprocessing the training set using a feature optimization strategy, inputting the preprocessed training set into the Transformer architecture, iteratively training the Transformer architecture, calculating the risk factor recognition accuracy and danger value accuracy of the Transformer architecture under different activation functions, and selecting the optimal risk factor recognition accuracy and danger value accuracy as the optimal activation function of the Transformer architecture; Based on the optimal activation function of the Transformer architecture as a prerequisite, setting an initial learning rate and the number of training epochs, selecting an adam optimizer to adaptively adjust the hyperparameters of the Transformer architecture, and outputting a converged danger recognition model; Obtaining the test set, testing the danger recognition model using the test set, outputting a test result, and determining whether the test result meets a preset test threshold. If it meets the preset test threshold, outputting a converged danger recognition model.

7. The elevator dangerous situation active recognition and intelligent rescue system according to claim 6, characterized in that: The elevator danger value is calculated by the following formula: （6）； （7）； Among them, is the elevator emergency value, represents the activation function of the decoder multi-layer perception mechanism, represents the input representation of the current emergency risk factor , represents the comprehensive weight of the current emergency risk factor , represents the bias term of the current emergency risk factor , is the number of emergency risk factors, represents the cumulative distribution function of the current emergency risk factor based on the affinity propagation clustering algorithm , is the input mean of the emergency risk factor, represents the damping coefficient of the affinity propagation clustering algorithm, represents the maximum number of iterations of the affinity propagation clustering algorithm.

Images