Elevator operation state monitoring and early warning method and platform based on multi-source data fusion

Abstract

The application relates to the technical field of data processing, and discloses an elevator operation state monitoring and early warning method and platform based on multi-source data fusion. The method comprises the following steps: collecting elevator operation data through a multi-source sensor, introducing an improved entropy weight method to fuse data, constructing a hybrid deep learning model to predict the elevator state, analyzing parameter abnormalities based on a correlation rule, generating a hierarchical early warning and fault diagnosis report, and realizing intelligent monitoring and early warning of the elevator. The application realizes accurate collection, deep analysis and fault prediction of the elevator operation data, so that the scientificity and accuracy of elevator maintenance are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to an elevator operation state monitoring and early warning method and platform based on multi-source data fusion. BACKGROUND

[0002] With the acceleration of urbanization and the continuous increase of building height, as an indispensable vertical transportation tool in modern buildings, the safe operation of elevators is crucial to people's daily life and work. Traditional elevator operation state monitoring mainly relies on periodic manual inspection and simple parameter monitoring, and maintenance personnel assess the operation state and potential failure risk of the elevator through experience and fixed period inspection. Most existing elevator monitoring systems use single sensor data acquisition and simple threshold alarm mode, which is difficult to fully capture the complex operation characteristics and potential failure signals of the elevator.

[0003] The existing elevator monitoring technology has many limitations: the data acquisition method is single, and cannot realize the deep fusion of multi-source data; the failure diagnosis relies on experience and lacks data-driven intelligent analysis; the predictive maintenance capability is weak, and cannot accurately locate and warn potential failures; the maintenance scheme recommendation lacks systematicness and scientificity, and the maintenance efficiency and accuracy are low. These technical deficiencies lead to high elevator maintenance cost and difficulty in effectively controlling failure risk, which seriously restricts the improvement of elevator operation safety and service life. SUMMARY

[0004] The present application provides an elevator operation state monitoring and early warning method and platform based on multi-source data fusion, which is used to realize accurate collection, deep analysis and failure prediction of elevator operation data, so as to improve the scientificity and accuracy of elevator maintenance.

[0005] In a first aspect, the present application provides an elevator operation state monitoring and early warning method based on multi-source data fusion, which comprises: collecting and preprocessing vibration signals, noise, speed fluctuation, motor temperature, load data, door machine parameters and guide rail deformation data in the elevator operation process to obtain normalized multi-source data; based on the improved entropy weight method, the normalized multi-source data is fused and processed, the weight distribution is adjusted by introducing a sensitivity correction factor to obtain an elevator operation state fusion index; the elevator operation state fusion index and the original multi-source data are input into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism and a convolutional neural network for training and processing to obtain an elevator state prediction model; according to the output result of the elevator state prediction model, the parameter state and the failure type are analyzed for correlation to obtain a set of correlation rules of parameter abnormal combination and failure type; based on the set of correlation rules, the elevator operation state is evaluated to obtain graded early warning information, and a failure diagnosis report is generated based on the graded early warning information.

[0006] In a second aspect, the application provides an elevator operation state monitoring and early warning platform based on multi-source data fusion, comprising:

[0007] a processing module, configured to collect and preprocess vibration signals, noise, speed fluctuations, motor temperature, load data, door machine parameters and guide rail deformation data in the elevator operation process to obtain normalized multi-source data;

[0008] a fusion module, configured to perform fusion processing on the normalized multi-source data based on an improved entropy weight method, and adjust weight distribution by introducing a sensitivity correction factor to obtain an elevator operation state fusion index;

[0009] an input module, configured to input the elevator operation state fusion index and original multi-source data into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism and a convolutional neural network for training and processing to obtain an elevator state prediction model;

[0010] an analysis module, configured to perform correlation analysis on parameter states and fault types according to an output result of the elevator state prediction model to obtain a correlation rule set of parameter abnormal combinations and fault types;

[0011] an evaluation module, configured to evaluate the elevator operation state based on the correlation rule set to obtain graded early warning information, and generate a fault diagnosis report based on the graded early warning information.

[0012] In a third aspect, a computer device is provided, comprising a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the computer device to perform the above-mentioned elevator operation state monitoring and early warning method based on multi-source data fusion.

[0013] In a fourth aspect, a computer readable storage medium is provided, the computer readable storage medium storing instructions, when running on a computer, causing the computer to perform the above-mentioned elevator operation state monitoring and early warning method based on multi-source data fusion.

[0014] The technical scheme provided in the application, through innovative integration of multi-source sensor data collection, improved entropy weight data fusion, mixed deep learning model construction and intelligent fault correlation analysis technical features, significantly improves the scientificity and accuracy of the elevator operation state monitoring. At the data collection level, through the cooperative work of various sensors such as acceleration sensors, sound sensors and speed sensors, the comprehensive capture of multi-dimensional characteristics such as vibration signals, noise, speed fluctuations, motor temperature and load data during the elevator operation is realized, laying a solid data foundation for subsequent deep analysis. The improved entropy weight method introduces a sensitivity correction factor, breaking through the limitations of the traditional entropy weight method weight distribution, and can more dynamically and accurately adjust the weight of different data sources, effectively balancing the importance of various sensor data. The innovation of the mixed deep learning model lies in the integration of bidirectional long short-term memory network, attention mechanism and convolutional neural network, which fully gives play to the advantages of artificial intelligence algorithms in complex time series data processing. Specifically, the bidirectional LSTM network captures long-term dependencies, the attention mechanism realizes dynamic weight distribution of key time nodes, and the one-dimensional convolutional neural network extracts local feature patterns, and the organic combination of the three algorithms significantly improves the accuracy and robustness of the elevator operation state prediction. The introduction of the association rule mining technology realizes the deep correlation analysis between the parameter abnormal state and the fault type, and through the setting of the dynamic threshold and the time sequence information weighting, an intelligent early warning system of the elevator fault is constructed. The feature vector similarity matching and multi-dimensional scoring mechanism are adopted in the fault case retrieval and maintenance scheme recommendation link, breaking through the traditional experience-driven maintenance mode, and establishing a data-driven scientific maintenance paradigm. This multi-source data fusion and artificial intelligence algorithm empowerment technical path significantly improves the accuracy and early warning ability of the elevator operation state monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical scheme of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 An embodiment schematic diagram of the elevator operation state monitoring and early warning method based on multi-source data fusion in the embodiments of the application;

[0017] Figure 2 An embodiment schematic diagram of the elevator operation state monitoring and early warning platform based on multi-source data fusion in the embodiments of the application;

[0018] Figure 3 A structural schematic block diagram of a computer device in the embodiments of the application. DETAILED DESCRIPTION

[0019] The embodiment of the present application provides a kind of based on multi-source data fusion's elevator running state monitoring and early warning method and platform.The terms "first", "second", "third", "fourth" and the like (if exist) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence.It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below, please refer to Figure 1 One embodiment of the elevator running state monitoring and early warning method based on multi-source data fusion in the embodiment of the present application includes:

[0021] Step S101, vibration signal, noise, speed fluctuation, motor temperature, load data, door machine parameter and guide rail deformation data in the process of elevator running are collected and preprocessed, and normalized multi-source data is obtained;

[0022] Step S102, the normalized multi-source data is fused and processed based on improved entropy weight method, and the weight distribution is adjusted by introducing sensitivity correction factor, to obtain elevator running state fusion index;

[0023] Step S103, the elevator running state fusion index and original multi-source data are input into the hybrid deep learning model composed of bidirectional long short-term memory network, attention mechanism and convolutional neural network for training processing, to obtain elevator state prediction model;

[0024] Step S104, according to the output result of elevator state prediction model, parameter state and fault type are associated and analyzed, to obtain the association rule set of parameter abnormal combination and fault type;

[0025] Step S105, based on the association rule set, the running state of the elevator is evaluated, to obtain graded warning information, and fault diagnosis report is generated based on the graded warning information.

[0026] It can be understood that the execution subject of the present application can be an elevator running state monitoring and early warning platform based on multi-source data fusion, and can also be a terminal or a server, and the specific place is not limited.The embodiment of the present application takes the server as the execution subject for example.

[0027] Specifically, the elevator operation state data is acquired by multiple sensors. Acceleration sensors are installed at the bottom, top and around the elevator car, with a sampling frequency of 200 Hz, recording the XYZ three-axis vibration acceleration values; the temperature sensor records the motor surface temperature every 10 seconds; the speed sensor records the elevator up and down speed changes with a sampling frequency of 50 Hz; the door machine control system records the time and current changes of each door opening and closing cycle; the load sensor monitors the weight change in the car in real time; the sound sensor records the car interior noise decibel value with a sampling frequency of 20 Hz; the displacement sensor measures the gap change between the guide rail and the car. The collected raw data is preprocessed, first by wavelet transform method to eliminate environmental interference, then by linear interpolation or forward filling method to complete the missing data, and then the different types of data are normalized to the [-1, 1] interval to eliminate the dimension effect, and the normalized multi-source data is obtained. The normalized multi-source data is fused. A standardized data matrix is constructed, which is processed according to the positive and negative index properties respectively, the proportion of each index at each sampling point is calculated, and then the information entropy is calculated. The difference coefficient is calculated by the information entropy, the larger the difference coefficient, the greater the effect of the index. On the basis of the traditional entropy weight method, a sensitivity correction factor is introduced to adjust the weight distribution, and the sensitivity correction factor is calculated by the ratio of the index standard deviation to the maximum standard deviation plus 1. After the final index weight is corrected by the correction factor, the elevator operation state fusion index of each sampling point is calculated, with a value range of [0, 1], and the closer the value is to 1, the better the elevator operation state. Then the elevator operation state fusion index and the original multi-source data are input into the hybrid deep learning model for training. The input window length is set to 24 hours and the prediction window length is set to 8 hours. The hybrid deep learning model includes a bidirectional long short-term memory network layer, an attention mechanism layer, a one-dimensional convolutional neural network layer and a fully connected layer. The bidirectional long short-term memory network layer captures the long-term dependence of time series data; the attention mechanism layer identifies important time points and features; the one-dimensional convolutional neural network layer extracts local feature patterns; the fully connected layer fuses the extracted features. The mean square error is used as the loss function in the training process, and the Adam optimizer is used for parameter optimization. When the validation set loss does not decrease for 10 consecutive cycles, the training is stopped in advance to prevent overfitting. Based on the trained elevator state prediction model, the parameter state and fault type are analyzed for correlation. The predicted parameter values are discretized into three levels of normal state, slight abnormal state and severe abnormal state. A transaction database is established combining with historical fault records, an improved Apriori algorithm is used to find frequent item sets, and association rules of parameter abnormal combinations and fault types are formed. The strength of the rules is evaluated by calculating the support, confidence and lift indexes, and a time sequence information weighting mechanism is introduced to give different weights to data at different time points, forming a set of association rules of parameter abnormal combinations and fault types. Based on the association rule set, the elevator operation state is evaluated. An adaptive dynamic threshold system is constructed, and the threshold calculation considers factors such as elevator type, service life and load state.The elevator operation state is divided into four levels of normal, attention, warning, and danger according to the matching results of the association rules, forming a hierarchical early warning information. The probabilities of each potential fault type are calculated, the development trend of high-probability faults is analyzed, and the fault occurrence time is estimated. Similar historical cases are retrieved from the fault knowledge base, and maintenance experience is extracted to form a reference maintenance plan. Finally, the early warning level, fault probability, time estimation result, and maintenance plan are integrated into a fault diagnosis report.

[0028] Taking an elevator in a shopping mall as an example, the XYZ three-axis vibration acceleration values collected by the acceleration sensor show abnormal fluctuations, and the vibration frequency is 40% higher than the normal value; the temperature sensor shows that the motor temperature continues to rise to 65℃; the noise sensor records that the noise during the car operation reaches 75 decibels; the displacement sensor measures that the guide rail gap becomes larger. After preprocessing and fusion of these data, the elevator operation state fusion index decreases to 0.37, which is in the "poor" level. The prediction model predicts that the fusion index will further decrease to 0.22 within 24 hours. The association rule analysis shows that the current parameter combination has a high matching degree with the "guide rail wear" fault type, with a support degree of 0.045, a confidence degree of 0.86, and an improvement degree of 12.3. The early warning platform generates a "warning" level early warning, and suggests in the fault diagnosis report that the guide rail state should be checked first and maintenance should be arranged.

[0029] In the embodiments of the present application, through the innovative integration of multi-source sensor data acquisition, improved entropy weight data fusion, mixed deep learning model construction, and intelligent fault correlation analysis, the scientificity and accuracy of elevator operation state monitoring are significantly improved. At the data acquisition level, through the cooperative work of various sensors such as acceleration sensors, sound sensors, and speed sensors, multi-dimensional characteristics such as vibration signals, noise, speed fluctuations, motor temperature, and load data during elevator operation are comprehensively captured, laying a solid data foundation for subsequent in-depth analysis. The improved entropy weight method introduces a sensitivity correction factor, breaking through the limitations of traditional entropy weight distribution, and can more dynamically and accurately adjust the weights of different data sources, effectively balancing the importance of various sensor data. The innovation of the mixed deep learning model lies in the integration of bidirectional long short-term memory network, attention mechanism, and convolutional neural network, fully leveraging the advantages of artificial intelligence algorithms in complex time series data processing. Specifically, the bidirectional LSTM network captures long-term dependencies, the attention mechanism realizes dynamic weight distribution for key time nodes, and the one-dimensional convolutional neural network extracts local feature patterns. The organic combination of the three algorithms significantly improves the accuracy and robustness of elevator operation state prediction. The introduction of association rule mining technology realizes deep correlation analysis between parameter abnormal state and fault type, and through the setting of dynamic threshold and time series information weighting, an intelligent early warning system for elevator faults is constructed. The fault case retrieval and maintenance scheme recommendation link uses feature vector similarity matching and multi-dimensional scoring mechanism, breaking through the traditional experience-driven maintenance mode and establishing a data-driven scientific maintenance paradigm. This multi-source data fusion and artificial intelligence algorithm empowerment technology path significantly improves the accuracy and early warning capability of elevator operation state monitoring.

[0030] In a specific embodiment, the process of step S101 can specifically include the following steps:

[0031] (1) Collecting original operation data through various sensor devices arranged in the elevator system, including collecting vibration signals through acceleration sensors, collecting noise data through sound sensors, collecting speed fluctuation values through speed sensors, collecting motor temperature through temperature sensors, collecting load data through load sensors, collecting door machine parameters through door machine control systems, and collecting guide rail deformation data through displacement sensors, to obtain elevator operation original monitoring data;

[0032] (2) According to the elevator operation original monitoring data, the wavelet transform method is used to eliminate environmental interference, and the directional noise reduction algorithm is used to separate effective signals and noise, to obtain elevator operation noise reduction data;

[0033] (3) Based on the elevator operation noise reduction data, the data missing is supplemented by linear interpolation or forward filling method, and the data blank points are filled according to the time sequence continuity principle, so that the complete data of elevator operation with a completeness of not less than 99.5% is obtained;

[0034] (4) The complete data of elevator operation is synchronously processed by time stamp alignment function, and each sensor data stream is reconstructed according to a unified sampling period, so that the time sequence aligned elevator monitoring data is obtained;

[0035] (5) The 3σ principle is used to identify and replace the abnormal value points of the time sequence aligned elevator monitoring data, and the data points deviating from the normal range are screened out based on the statistical distribution characteristics, so that the elevator monitoring data after abnormal value processing is obtained;

[0036] (6) The elevator monitoring data after abnormal value processing is converted to the interval [-1, 1] according to the maximum and minimum value normalization method, so as to eliminate the influence of different physical dimensions, and the normalized multi-source data is obtained.

[0037] Specifically, a plurality of sensor devices are arranged in the elevator system, including acceleration sensors, sound sensors, speed sensors, temperature sensors, load sensors, door machine control systems and displacement sensors. The acceleration sensors are installed at the bottom, top and periphery of the elevator car, with a sampling frequency of 200 Hz, recording XYZ three-axis vibration acceleration values to form vibration time sequence data; the sound sensor is installed in the car, with a sampling frequency of 20 Hz, continuously recording the noise decibel value in the car; the speed sensor is installed in the elevator drive system, with a sampling frequency of 50 Hz, monitoring the up and down speed of the elevator and its fluctuation; the temperature sensor is attached to the surface of the motor, recording temperature data every 10 seconds; the load sensor is installed at the bottom of the car, monitoring the weight change in the car in real time; the door machine control system records the time and current change of each opening and closing of the door; the displacement sensor measures the gap change between the guide rail and the car. These sensors work simultaneously and continuously collect data to form multi-dimensional and multi-angle elevator operation original monitoring data. After obtaining the elevator operation original monitoring data, noise reduction processing is required to eliminate environmental interference. Wavelet transform method is used for noise reduction. Wavelet transform has time-frequency localization characteristics and is suitable for processing non-stationary signals. The specific steps include selecting appropriate wavelet basis function, decomposing the original signal into wavelet coefficients of different frequencies, setting threshold to screen noise coefficients, retaining effective signal coefficients, and then reconstructing the signal. According to the characteristics of different types of sensors, appropriate wavelet basis function and threshold strategy are selected. Daubechies wavelet is selected for elevator vibration signal, Symlet wavelet is selected for noise data, and Coiflet wavelet is selected for temperature and other slowly varying signals. The signal is split into multiple frequency subbands by wavelet decomposition, soft threshold or hard threshold method is applied to high frequency subbands to suppress noise, low frequency subband information is retained, and finally the elevator operation noise reduction data is obtained by reconstruction.

[0038] During the collection process, data may be missing due to temporary sensor failure, communication interruption, or power fluctuations. Linear interpolation or forward filling method is used to process the missing values in the elevator operation noise reduction data. Linear interpolation is suitable for cases where the data changes smoothly before and after the missing point, and the missing value is calculated by the known data points on both sides of the missing point. Forward filling method is suitable for sudden missing cases, and the nearest valid value before the missing point is filled. For different types of data, choose the appropriate filling method. Linear interpolation is used for vibration signals and noise data due to their rapid changes. Forward filling is used for temperature and load data due to their slow changes. Through the principle of time series continuity, the time correlation of the data is analyzed, and the possible values of the missing data points are reasonably inferred to form elevator operation complete data with a completeness of not less than 99.5%. Due to the different sampling frequencies of each sensor, there are differences in the time points of data collection, so the elevator operation complete data needs to be time-synchronized. Through the timestamp alignment function, the data from different sensors is reorganized according to a unified time reference. First, determine the synchronization reference time point, then resample each sensor data to make it conform to the unified sampling period. For high sampling rate vibration signals and speed data, downsample; for low sampling rate temperature and door machine parameters, interpolate. The least common multiple principle is used to determine the unified sampling period, and then all sensor data is mapped to this unified time axis to form time-aligned elevator monitoring data.

[0039] The time-aligned elevator monitoring data is processed for outliers. The 3σ principle in statistics is used, which means that under normal distribution, the probability of data falling within the range of mean μ ± 3σ is 99.73%, and data points beyond this range are considered outliers. For each type of sensor data, calculate its mean μ and standard deviation σ, set the upper and lower threshold μ ± 3σ, and detect data points beyond the threshold. For the identified outliers, use the median replacement method or local mean replacement method for processing. The median replacement method is suitable for cases with many outliers, and the local mean replacement is suitable for isolated outliers. Through outlier processing, unreasonable data disturbances are removed, and data reflecting the true running state of the elevator are retained to form outlier-processed elevator monitoring data. The elevator monitoring data after outlier processing is normalized. Since different sensor data has different physical dimensions and large numerical range differences, it needs to be unified to the same scale. The maximum and minimum value normalization method is used to linearly map each type of data to the [-1, 1] interval. Through normalization processing, the influence of different physical dimensions is eliminated, and all data is compared and fused on the same scale to form normalized multi-source data.

[0040] Taking the passenger elevator in a high-rise building as an example, the vibration data collected by the acceleration sensor contains noise interference of the building itself vibration. The Daubechies4 wavelet is used for 5-layer decomposition, and the soft threshold method is used to process the high-frequency coefficients. The pure elevator vibration signal is obtained after reconstruction. During data collection, the temperature sensor data is intermittently missing due to poor contact. The missing values of 10 consecutive time points are filled by linear interpolation, and the data integrity reaches 99.7%. The sampling frequencies of various sensors are different from 20Hz to 200Hz. By setting a uniform 100ms sampling interval, all data are resampled and time-aligned. The abnormal value in the motor temperature data suddenly rises to 120℃, which is obviously beyond the normal working temperature range (30-80℃). The mean value is 45℃ and the standard deviation is 5℃. The abnormal value exceeding 60℃ is replaced by the average value of the previous and next time points. The vibration data is normalized to the range of [-1, 1] from the range of ±20g, the temperature data is normalized to the range of [-1, 1] from the range of 30-80℃, and the load data is normalized to the range of [-1, 1] from the range of 0-1600kg. A standardized multi-source data set is formed.

[0041] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0042] (1) Construct a standardized data matrix from the normalized multi-source data, and obtain a standardized processing matrix through normalized calculation of positive and negative indicators;

[0043] (2) Calculate the proportion and information entropy of each indicator at each sampling point according to the standardized processing matrix, and obtain the information entropy of each indicator through the proportion multiplied by the logarithm and then summed;

[0044] (3) Calculate the difference coefficient and initial weight based on the information entropy of each indicator, obtain the difference coefficient by subtracting the information entropy from the constant 1, obtain the initial weight by dividing the difference coefficient by the sum, and obtain the initial weight configuration;

[0045] (4) Calculate the sensitivity correction factor by the ratio of the standard deviation of each indicator to the maximum standard deviation, and obtain the weight adjustment parameter by adding 1 to the ratio as the correction factor;

[0046] (5) Modify the initial weight using the weight adjustment parameter, and obtain the final weight by multiplying the correction factor by the initial weight and normalizing it;

[0047] (6) Based on the weighted sum of the final weight and the standardized data, establish a sub-scene evaluation system for elevator operation under different working conditions, and obtain the elevator operation state fusion index.

[0048] Specifically, the normalized multi-source data is fused, and the comprehensive evaluation index of the elevator operation state is constructed by improving the entropy weight method. First, the normalized multi-source data is constructed into a standardized data matrix, and the elevator operation parameters are divided into two categories of positive indicators and negative indicators. The positive indicators are parameters whose values the larger the better, such as door machine operation stability; the negative indicators are parameters whose values the smaller the better, such as vibration amplitude and noise decibel value. For the positive indicators, the formula is used for standardization processing:

[0049]

[0050] For the negative indicators, the formula is used for standardization processing:

[0051]

[0052] where x ij represents the jth index value of the ith sampling point, min(x j ) and max(x j ) represent the minimum and maximum values of the jth index respectively, and x' ij represents the standardized value. Through standardization processing, all index values are mapped to the interval [0, 1] to form a standardization processing matrix.

[0053] Then, the proportion of each index at each sampling point is calculated according to the standardization processing matrix. The proportion calculation formula is:

[0054]

[0055] where p ij represents the proportion of the jth index at the ith sampling point, and S represents the total number of sampling points. Then, the information entropy of each index is calculated. The information entropy is an index for measuring data uncertainty, and the calculation formula is:

[0056]

[0057] where E j represents the information entropy of the jth index, κ = 1 / ln(S) is a normalization constant, which ensures that the value of E j is between 0 and 1. When p ij = 0, p ij ln(p ij ) = 0 is defined. The greater the information entropy, the lower the discrimination of the index to the elevator state.

[0058] The difference coefficient is calculated based on the information entropy of each index. The difference coefficient represents the discrimination ability of the index, and the calculation formula is:

[0059] d j = 1-E j

[0060] where d j represents the difference coefficient of the jth index. The larger the difference coefficient, the greater the role of the index in evaluation. Further calculate the initial weight, and the initial weight calculation formula is:

[0061]

[0062] where w j represents the initial weight of the jth index, and N represents the total number of indexes. After the initial weight is configured, each index obtains a corresponding weight value.

[0063] The traditional entropy weight method only considers the data distribution characteristics, and lacks the consideration of index sensitivity, so a sensitivity correction factor is introduced for adjustment. The sensitivity correction factor calculation formula is:

[0064]

[0065] where a j represents the sensitivity correction factor of the jth index, and s j represents the standard deviation of the jth index, and s max represents the maximum value of the standard deviation of all indexes. The index with large standard deviation is more sensitive to data changes, and is given a higher weight.

[0066] The initial weight is corrected by using the sensitivity correction factor, and the correction formula is:

[0067]

[0068] where w j ' represents the maximum weight of the jth index. Through sensitivity correction, the index sensitive to the change of elevator state is highlighted, and the sensitivity of the fusion index is improved.

[0069] Finally, based on the maximum weight and the standardized data, the weighted sum is calculated to calculate the elevator operation state fusion index, and the calculation formula is:

[0070]

[0071] where F i represents the elevator operation state fusion index of the ith sampling point, and the value range is [0, 1]. The value closer to 1 indicates that the elevator operation state is better, and the value closer to 0 indicates that the potential fault risk is higher.

[0072] For different operating conditions of the elevator, such as empty uplink, full uplink, empty downlink, full downlink, door opening and closing process, etc., the sub-scenarios are divided and the fusion index is constructed to form a complete elevator operation state evaluation system.

[0073] Taking the elevator of a certain shopping mall as an example, seven types of index data such as vibration signal, noise, speed fluctuation, motor temperature, load data, door machine parameter and guide rail deformation are obtained through multi-source data collection. Among them, the vibration amplitude, noise decibel value, speed fluctuation rate, motor temperature and guide rail deformation belong to negative indicators, and the load stability and door machine smoothness belong to positive indicators. The indicators are standardized, such as the vibration amplitude is standardized to the interval of 0-1 from 0.1g to 0.5g. The information entropy of each index is calculated, and it is found that the information entropy of the motor temperature is 0.45, the information entropy of the vibration signal is 0.32, and the information entropy of the noise data is 0.38, indicating that the vibration signal has the highest discrimination. The difference coefficient is calculated, the difference coefficient of the vibration signal is 0.68, and the difference coefficient of the motor temperature is 0.55, to obtain the initial weight configuration. The standard deviation of the vibration signal is 0.08, and the standard deviation of the motor temperature is 0.05. The sensitivity correction factor is calculated to correct the initial weight. Finally, the fusion index of the elevator running state is obtained by weighted summation, which is 0.82, indicating that the elevator running state is good. When the elevator enters the full load uplink working condition, the load data weight increases, and the fusion index decreases to 0.76, entering the "good" level. When the door machine parameter is detected to be abnormal, the corresponding weight increases, and the fusion index further decreases to 0.63, triggering the "attention" level warning.

[0074] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0075] (1) The historical time series data is divided into a sliding window, and the input time length of 24 hours and the prediction time length of 8 hours are set to form a training sample set;

[0076] (2) The elevator running state fusion index and the original multi-source data in the training sample set are input into the bidirectional long short-term memory network layer, the long-term dependence relationship is captured through forward propagation and back propagation, and the sequence feature representation is obtained;

[0077] (3) The attention weight coefficient is constructed based on the sequence feature representation, the importance distribution of each time point is calculated through the softmax function, and the weighted feature representation is obtained;

[0078] (4) The weighted feature representation is subjected to local feature extraction through a one-dimensional convolutional neural network, the local mode change is captured by using a sliding convolution kernel, and the feature mapping result is obtained;

[0079] (5) The feature mapping result is input into the full connection layer for feature fusion, and the dropout mechanism is introduced to control the risk of overfitting, and the prediction output vector is obtained;

[0080] (6) The mean square error loss is calculated according to the prediction output vector and the true label, the parameters are updated through the Adam optimizer, the training is stopped when the loss of the validation set no longer decreases, and the elevator state prediction model is obtained.

[0081] Specifically, the historical time series data is divided into sliding windows, and the input window length is set to 24 hours and the prediction window length is set to 8 hours. The sliding window is a time series processing technology that cuts continuous time series into multiple overlapping data segments by sliding a fixed-size window on the time axis. In specific implementation, 24 hours of data is taken as input features and 8 hours of data in the future is taken as prediction targets on the continuously collected elevator operation data. The sliding step can be set to 1 hour, that is, a new training sample is generated every time the window moves forward by 1 hour. In this way, a large number of training samples can be generated from long-term historical data, such as more than 8000 training samples from one year of data, forming a training sample set containing input features and prediction targets. The elevator operation state fusion index and the original multi-source data in the training sample set are input into the bidirectional long short-term memory network layer for processing. The bidirectional long short-term memory network (BiLSTM) is a special recurrent neural network that can consider both past and future information, and is suitable for processing time series data. BiLSTM is composed of forward LSTM and reverse LSTM, and the forward LSTM processes the sequence from left to right, and the reverse LSTM processes the sequence from right to left, and the long-term dependence relationship in the time series data is captured through this bidirectional structure. In elevator state prediction, the BiLSTM layer receives an input tensor with a dimension of [batch size x time step x feature number], and each time step corresponds to one hour of data, including the fusion index and the original multi-source data. Through the gating mechanism (including the input gate, the forgetting gate and the output gate), the information is screened and updated, the input gate controls the degree of new information entering the cell state, the forgetting gate controls the degree of old information retention, and the output gate controls the degree of cell state transmission to the output. The BiLSTM layer is set to two layers, with 128 hidden units, and the parameters are learned through the forward propagation and backward propagation algorithms, and finally the sequence feature representation is output, capturing the time series pattern and long-term trend of the elevator operation data.

[0082] The sequence feature representation output by BiLSTM is used to construct attention weight coefficients. Attention mechanism is a technique that allows the model to selectively focus on important parts of the input sequence. In elevator state prediction, data at different time points contribute differently to the prediction result, and the attention mechanism can dynamically adjust the weights of data at each time point. In specific implementation, first, the hidden state output by BiLSTM is linearly transformed, then it is nonlinearly mapped using the tanh activation function, and then the softmax function is used to normalize to get the attention weight of each time point. The softmax function converts the input into a probability distribution, ensuring that the sum of all weights is 1. The attention weight reflects the importance of data at each time point to the prediction, and the larger the value, the greater the influence of the time point on the prediction result. Through the weighted sum of the attention weight and the original feature, the weighted feature representation is obtained, which highlights the information of important time points and suppresses the noise interference of irrelevant time points.

[0083] The weighted feature representation is subjected to local feature extraction by a one-dimensional convolutional neural network. One-dimensional convolutional neural network (1D-CNN) is suitable for processing sequence data, and captures local pattern changes by moving the convolution kernel in the time dimension. In elevator state prediction, the 1D-CNN layer sets the number of convolution kernels to 64 and the size of the convolution kernel to 3, meaning that 3 consecutive time points of data are considered at each time. The convolution operation calculates the dot product between the convolution kernel and the local region of the input data to extract local time series features. The sliding convolution kernel moves from the start of the sequence to the end, and a convolution calculation is performed at each step to obtain a set of feature maps. The ReLU activation function is used to increase nonlinearity and enhance the model's expression ability. 1D-CNN can effectively extract local patterns in elevator operation data, such as short-term vibration trends and temperature fluctuation patterns, to generate feature map results.

[0084] The feature map results are input into a fully connected layer for feature fusion. The fully connected layer flattens the feature maps extracted by the previous layers and connects all neurons through a weight matrix to realize the fusion of different features. In the elevator state prediction model, two fully connected layers are set with node numbers of 64 and 32 respectively, and the ReLU activation function is used. The dropout mechanism is also introduced to control the risk of overfitting. Dropout is a regularization technique that randomly discards a portion of neurons during training to prevent the model from relying too much on certain features. In the elevator state prediction model, the dropout rate is set to 0.3, meaning that 30% of neurons are randomly discarded each time. The fully connected layer finally outputs the fusion index prediction value and the future value of each key parameter for the next 8 hours, forming the prediction output vector.

[0085] The mean squared error loss is calculated according to the predicted output vector and the true label. Mean squared error (MSE) is a commonly used loss function in regression problems, which calculates the average of the square of the difference between the predicted value and the true value. In elevator state prediction, the loss value is obtained by taking the average of the square of the difference between the model's predicted future 8-hour fusion index and the actual observation value. Parameter updating is performed by the Adam optimizer, which combines the advantages of the momentum method and the adaptive learning rate method, and dynamically adjusts the learning rate according to the gradient. The initial learning rate is set to 0.001, and the learning rate is reduced by 10% every 50 epochs. To prevent overfitting, when the loss on the validation set does not decrease for 10 consecutive epochs, the early stopping mechanism is started, and the training is stopped. After the complete training process, the average prediction error of the elevator state prediction model on the test set is not more than 3.5%, and the correlation coefficient is not less than 0.92, ensuring the reliability of the prediction results.

[0086] Taking the elevators in a residential community as an example, two years of operation data were collected, and about 15000 training samples were obtained by sliding window division. Each sample contains 24 hours of input data (fusion index and 7 types of original parameters) and corresponding 8-hour prediction target. The bidirectional LSTM layer processes the time series information, capturing the long-term correlation between elevator load and vibration, as well as the periodic pattern of temperature change. The attention mechanism analysis found that the data 4-6 hours before the occurrence of vibration anomalies is most critical to prediction, and the attention weight at the corresponding time point is significantly higher than that at other time points. The one-dimensional CNN extracts local feature patterns such as rapid rise of motor temperature and vibration frequency change. The fully connected layer fuses these features to generate a 8-hour state prediction. The model can accurately predict the vibration anomalies caused by elevator guide rail wear and give a 6-hour early warning. When the model detects a continuous rise in motor temperature and an abnormal vibration frequency, it predicts that the fusion index will drop from 0.75 to 0.45 in the next 4 hours, triggering a "warning" level warning, and maintenance personnel replace the worn parts in advance to avoid elevator failure.

[0087] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0088] (1) Discretize the fusion index prediction value and parameter prediction data output by the elevator state prediction model, and divide them into three levels of normal state, slight abnormal state and severe abnormal state to obtain parameter state discrete data;

[0089] (2) Establish a transaction database according to the parameter state discrete data and historical fault records, record the association between parameter state combinations and corresponding fault types, and obtain a state-fault mapping dataset;

[0090] (3) Set minimum support threshold and minimum confidence threshold for the state-failure mapping dataset, filter out single parameter abnormal states that frequently occur through support calculation, and obtain frequent item sets;

[0091] (4) Perform self-connection operation based on the frequent item sets, construct multi-parameter combined frequent item sets through iterative connection and pruning, and obtain parameter combination feature sets;

[0092] (5) Extract association rules from the parameter combination feature sets, and obtain preliminary association rules of parameter abnormalities and failure types by calculating support, confidence and lift;

[0093] (6) Introduce a time sequence information weighting mechanism for the preliminary association rules, assign different weight values according to the data time, and obtain the association rule set of parameter abnormal combination and failure type according to the weighted confidence and lift.

[0094] Specifically, the fusion index prediction value output by the elevator state prediction model and each parameter prediction data are discretized. Discretization is the process of converting continuous numerical values into discrete states, which facilitates subsequent association rule mining. For the fusion index, the value range [0, 1] is divided into five levels: excellent [0.8, 1.0], good [0.6, 0.8), general [0.4, 0.6), poor [0.2, 0.4) and dangerous [0, 0.2]. For each parameter data, according to the equipment operation standard and historical statistical results, the parameter value is divided into three levels of normal state, slight abnormal state and serious abnormal state. For example, the motor temperature is below 60°C for normal state, 60-75°C for slight abnormal state, and above 75°C for serious abnormal state; the vibration acceleration is less than 0.2g for normal state, 0.2-0.4g for slight abnormal state, and greater than 0.4g for serious abnormal state. Through this discretization, the continuous monitoring data of elevator operation is converted into parameter state discrete data for analysis.

[0095] Then, a transaction database is established according to the parameter state discrete data and historical failure records. The transaction database is the basis for association rule mining, and each record contains parameter state combination and corresponding failure type. For each failure case in history, the parameter state discrete data within a certain period of time before the failure (such as 24 hours before the failure) is extracted and associated with the failure type. For example, a record may contain "{vibration increase = serious abnormal, motor temperature rise = slight abnormal, noise increase = serious abnormal, failure type = guide rail failure}". By organizing the relevant data of all historical failure cases, a complete state-failure mapping dataset is formed, providing data support for subsequent association rule mining.

[0096] After obtaining the state-fault mapping dataset, the improved Apriori algorithm is applied for association rule mining. Apriori algorithm is a classical association rule mining algorithm, which is used to discover the frequent itemsets and their association relationships in the dataset. First, the minimum support threshold and the minimum confidence threshold are set. The support represents the proportion of records containing a certain itemset in the total records, and the confidence represents the proportion of records containing itemset X that also contain itemset Y. In elevator fault diagnosis, the minimum support is usually set low (e.g. 0.03) because some important faults may occur less frequently; the minimum confidence is set high (e.g. 0.75) to ensure the reliability of the rules. By calculating the frequency of single parameter abnormal state in the dataset, parameter states with support not lower than the minimum support are selected to form frequent 1-itemsets.

[0097] Based on the frequent 1-itemsets, self-connection operation is performed to construct higher-order frequent itemsets. Self-connection is the process of connecting frequent k-itemsets with themselves to generate candidate k+1-itemsets. During the connection process, it is required that two k-itemsets have k-1 identical items. For example, from the frequent 1-itemsets {vibration increase = severe abnormality} and {motor temperature rise = slight abnormality}, the candidate 2-itemset {vibration increase = severe abnormality, motor temperature rise = slight abnormality} can be generated. After generating the candidate itemsets, pruning operation is needed to delete those candidate k-itemsets whose any k-1 subsets are not frequent itemsets. By iteratively performing connection and pruning operations, higher-order frequent itemsets are gradually constructed until no new frequent itemsets can be generated. In this way, all frequent itemsets are obtained, forming the parameter combination feature set, which contains the features of various parameter abnormality combinations in elevator operation.

[0098] From the parameter combination feature set, association rules are extracted to form the relationship description between parameter abnormalities and fault types. For each frequent itemset, all its possible non-empty proper subsets are enumerated as antecedents, and the remaining part is taken as the consequent to form candidate association rules. Then the support, confidence and lift of each rule are calculated. The support represents the proportion of the rule in the total dataset; the confidence represents the probability of the occurrence of the consequent when the antecedent occurs; the lift represents the ratio of the frequency of the occurrence of the consequent in the records containing the antecedent to the frequency of the occurrence of the consequent in the total records, which is used to measure the effectiveness of the rule. Lift greater than 1 indicates that the occurrence of the antecedent increases the likelihood of the occurrence of the consequent, and the rule has positive correlation. Rules that meet the minimum support and minimum confidence requirements at the same time are selected to obtain the preliminary association rules between parameter abnormalities and fault types.

[0099] Finally, the time sequence information weighting mechanism is introduced to the preliminary association rules to improve the timeliness of the rules. The running state of the elevator is dynamically changing, and the recent data is more indicative of the current state than the long-term data. The time sequence information weighting mechanism gives different weights according to the distance of the data generation time, and the recent data has a higher weight. The weight calculation uses an exponential decay function to give higher importance to the recent data. The rules are reordered by the weighted confidence and lift to form the final association rule set of the parameter abnormal combination and the fault type. These rules express the relationship of "if a certain parameter abnormal combination occurs, a certain type of fault may occur", which provides knowledge support for the elevator fault diagnosis and early warning.

[0100] Taking an office building elevator as an example, the transaction database containing 2000 records is established by analyzing the running data and fault records in the past six months. The continuous monitoring data is converted into state descriptions through discretization processing, such as converting the vibration data 0.45g into "vibration increase = severe abnormality". The minimum support threshold is set to 0.03, and the minimum confidence threshold is set to 0.75. The frequent item sets are mined by the Apriori algorithm, such as the frequent 1-item set containing "vibration increase = severe abnormality" (support 0.06) and "noise increase = severe abnormality" (support 0.05). The multi-parameter combination frequent item sets are constructed through self-connection and pruning of the frequent item sets, such as {vibration increase = severe abnormality, motor temperature rise = slight abnormality, noise increase = severe abnormality}. The association rules are extracted from them, such as "{vibration increase = severe abnormality, motor temperature rise = slight abnormality, noise increase = severe abnormality} → {guide rail fault}", which has a support of 0.045, a confidence of 0.86, and a lift of 12.3, indicating that this parameter abnormal combination is highly related to the guide rail fault. Through time sequence weighting processing, the data in the recent two weeks has a weight of 1.5 times that of the long-term data, further improving the timeliness of the rules. The final association rule set can effectively guide the elevator fault diagnosis. When the parameter abnormal combination conforming to the rule antecedent is detected, the early warning platform will timely issue the corresponding type of fault warning.

[0101] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0102] (1) An adaptive dynamic threshold system is constructed to obtain the elevator running parameter alarm threshold through the comprehensive calculation of the parameter historical mean, standard deviation, and elevator type, service life, and load state;

[0103] (2) According to the matching results of the association rule set, the elevator running state is divided into four levels of normal, attention, warning, and danger through comparative analysis of the parameter abnormal state and the rule antecedent, and the graded early warning information is obtained;

[0104] (3) Calculate the probability of each potential fault type based on the hierarchical early warning information, and obtain the fault probability distribution by multiplying the matching degree of the confidence and lift of the association rule and the current parameter state;

[0105] (4) Perform trend analysis on the fault type data with high fault probability distribution, and obtain the fault time estimation result by inferring the fault development speed through the exponential smoothing prediction method;

[0106] (5) Retrieve similar historical cases from the fault knowledge base according to the fault type, and obtain the reference maintenance scheme by extracting relevant maintenance experience through fault feature similarity calculation;

[0107] (6) Integrate the hierarchical early warning information, fault probability distribution, fault time estimation result, and reference maintenance scheme into structured data to obtain the fault diagnosis report.

[0108] Specifically, an adaptive dynamic threshold system is constructed. Traditional fixed thresholds cannot adapt to the differences of different elevator types and different operating conditions, while adaptive dynamic thresholds can dynamically adjust the alarm boundaries according to the characteristics of the elevator. The adaptive threshold calculation considers the historical mean, standard deviation of the parameters and the specific situation of the elevator, and uses the formula: threshold = parameter historical mean ± adaptive coefficient × parameter historical standard deviation. The adaptive coefficient is determined by factors such as elevator type, service life, load state, etc., and the calculation method is: adaptive coefficient = base coefficient × [1 + life influence factor × log(usage life + 1) + load influence factor × load rate + frequency influence factor × running frequency]. The base coefficient is set according to the type of elevator, usually between 2.5 and 3.5; the life influence factor is about 0.15, the longer the service life, the wider the threshold range; the load influence factor is about 0.1, the higher the load rate, the higher the upper limit of the threshold; the frequency influence factor is about 0.08, the higher the running frequency, the greater the allowed threshold fluctuation. Through this dynamic calculation method, the alarm threshold of each elevator operating parameter is obtained, providing a reference for state evaluation.

[0109] Then, according to the matching results of the set of association rules, the elevator operation state is evaluated in stages. The elevator operation state is divided into four levels: normal, attention, warning, and danger. The division is based on the matching of the current parameter abnormal state and the rule antecedent. The normal state means that all parameters are within the normal range and no association rule is triggered. The attention state means that a parameter has a slight abnormality or a rule with a low confidence level is triggered. The warning state means that multiple parameters have abnormalities or a rule with a high confidence level is triggered. The danger state means that a key parameter has a serious abnormality or a rule with an extremely high confidence level is triggered. The specific judgment process is to compare the current elevator operation parameter state with the antecedent of each rule in the set of association rules, calculate the matching degree, and determine the elevator operation state level according to the highest matching degree and the confidence level of the corresponding rule to form a staged early warning information.

[0110] Based on the staged early warning information, the occurrence probability of each potential fault type is calculated. For each possible fault type, the probability is evaluated through all rules related to the fault in the set of association rules. The fault probability calculation formula is: fault probability = sum of (confidence x lift x current state matching degree) of all related rules. The confidence reflects the reliability of the rule, the lift reflects the correlation strength of the rule, and the current state matching degree reflects the matching degree of the current situation and the rule condition. For example, the confidence of a rule "vibration increase + temperature rise → guide rail fault" is 0.85, the lift is 12, and the matching degree of the current state and the rule antecedent is 0.9. Therefore, the contribution of the rule to the probability of guide rail fault is 0.85 x 12 x 0.9 = 9.18. The probability score of guide rail fault is obtained by adding the contributions of all rules related to guide rail fault. In this way, the probability scores of various faults are calculated to form a fault probability distribution, indicating the most likely fault type to occur.

[0111] The trend of the higher fault type in the fault probability distribution is analyzed to estimate the fault occurrence time. The exponential smoothing prediction method is adopted, which is suitable for processing time series data with trends. Exponential smoothing prediction gives higher weight to recent data and lower weight to long-term data, and predicts future trends through weighted average. The specific calculation process is to use the parameter degradation trend to infer the time point when the parameter reaches the fault threshold. For example, by analyzing the temperature rise rate of the motor, the time required for the temperature to reach the fault threshold of 80°C is predicted; by analyzing the vibration amplitude growth trend, the time required for the vibration to reach the danger threshold of 0.6g is predicted. By integrating the prediction results of various related parameters, the time range when the fault may occur is obtained to form a fault time estimation result, providing a time reference for maintenance planning.

[0112] The historical case similar to the current early warning fault type is retrieved from the fault knowledge base to provide reference for maintenance. The fault knowledge base records historical fault cases, including fault description, parameter characteristics, fault cause, maintenance measures, and maintenance effect, etc. The most relevant case is found by calculating the similarity between the current fault characteristics and the historical cases. The similarity calculation is based on the Euclidean distance or cosine similarity of fault parameter characteristics. The specific process is to compare the parameter characteristic vector of the current fault with the characteristic vector of the historical cases. For example, the current fault characteristics include increased vibration frequency, increased motor temperature, and increased noise decibel. After being represented by the characteristic vector, the similarity with the cases in the knowledge base is calculated, and the historical cases with similarity higher than the threshold are screened out. The maintenance experience is extracted from these similar cases, including fault cause analysis, maintenance method suggestion, required tools and materials, etc., to form a reference maintenance scheme.

[0113] Finally, the hierarchical early warning information, fault probability distribution, fault time estimation result, and reference maintenance scheme are integrated into structured data to generate a fault diagnosis report. The report includes the following parts: the basic information part records the basic data such as elevator number, model, installation location, and service life; the operation status overview part shows the recent operation time, passenger capacity, operation frequency, and other statistical information; the abnormal parameter analysis part lists all parameters that exceed the normal range, including parameter name, normal range, current value, deviation degree, and change trend; the fault diagnosis result part gives the possible fault type, fault probability, fault cause analysis, etc.; the reference case part provides historical similar fault cases for reference; the maintenance suggestion part proposes specific maintenance suggestions according to the fault type and severity, including maintenance items, priority, estimated working hours, and required materials; the prevention measures part proposes suggestions for long-term prevention of similar faults. The fault diagnosis report is managed according to the severity level, and different levels of reports use different delivery methods and processing procedures to ensure that important information is handled in a timely manner.

[0114] Taking a high-rise residential elevator as an example, the vibration data of the elevator is found to be abnormal through multi-source data fusion monitoring. The acceleration value gradually increases from the normal 0.15g to 0.38g, the noise increases to 72 decibels, and the motor temperature rises to 68℃. The adaptive threshold system considers that the elevator has been used for 8 years, the average load rate is 60%, and calculates the vibration warning threshold as 0.35g, the noise warning threshold as 70 decibels, and the motor temperature warning threshold as 65℃. According to the matching results of the association rule set, the matching degree of the current state and the rule "{vibration increase = slight abnormality, noise increase = slight abnormality, motor temperature rise = slight abnormality}→{guide rail fault}" is 0.85, the rule confidence is 0.82, and the "warning" level alarm is triggered. Calculate the probability of each fault type, the probability of guide rail fault is the highest, which is 78%, the probability of steel wire rope wear is 15%, and the probability of other faults is relatively low. Through exponential smoothing prediction, the vibration value will reach the dangerous threshold of 0.5g within 72 hours, and it is recommended to repair within 72 hours. Three similar cases are retrieved from the fault knowledge base, all of which are vibration abnormalities caused by guide rail wear, and the maintenance experience is extracted to form a reference maintenance plan, including checking whether the guide rail is deformed, checking whether the guide shoe is worn, adjusting the guide shoe gap, etc. A structured fault diagnosis report is generated and submitted to maintenance personnel for targeted repair, avoiding the occurrence of elevator failure.

[0115] In a specific embodiment, the process of extracting relevant maintenance experience through fault feature similarity calculation can specifically include the following steps:

[0116] (1) Construct a feature vector representation for the fault type, convert the key parameter abnormal pattern, equipment running state description and fault phenomenon into a structured representation through coding to obtain the target fault feature vector;

[0117] (2) Extract historical fault case data from the fault knowledge base, extract the fault features, maintenance measures and maintenance effects of each case through structure, and obtain the historical case feature set;

[0118] (3) Calculate the similarity based on the target fault feature vector and the historical case feature set, quantify the matching degree of the fault mode through the cosine similarity method, and obtain a case similarity ranking list;

[0119] (4) Filter historical cases with similarity higher than the threshold value from the case similarity ranking list, set a minimum similarity value to ensure the relevance of the reference cases, and obtain relevant historical maintenance cases;

[0120] (5) Cluster analysis of maintenance measures in relevant historical maintenance cases, group maintenance methods based on the relevance of fault causes and component positions, and obtain maintenance scheme categories;

[0121] (6) Based on the maintenance scheme category and historical success rate data, the effectiveness index and resource consumption score of each scheme are calculated to form a priority list, and the reference maintenance scheme is obtained.

[0122] Specifically, the multi-dimensional data collected by various sensors during the operation of the elevator needs to be finely extracted and coded. Specifically, the feature vector construction includes the structural representation of key parameters such as the frequency spectrum characteristics of the vibration signal, the motor temperature change curve, and the load fluctuation trend. For example, for traction system failure, the vibration signal will be decomposed into frequency components, amplitude characteristics and time characteristics, and the motor temperature change will extract abnormal points, change slope and other key indicators. These data are converted into multi-dimensional feature vectors through standardized coding. The extraction of fault knowledge base data follows strict structured principles. Each historical fault case is regarded as a complete knowledge unit, and the extraction process not only collects basic fault information, but also includes detailed maintenance technical routes, types of spare parts used, maintenance personnel operation details, and performance recovery of the equipment after maintenance. For elevator control system failure, the knowledge base records detailed fault diagnosis process, replacement control board model, debugging parameters, fault duration and other detailed information, ensuring that each case can provide comprehensive technical reference.

[0123] The similarity calculation adopts the cosine similarity algorithm as the core quantitative method. This algorithm accurately assesses the matching degree of the fault mode by calculating the cosine value of the target fault feature vector and the historical case feature set in the multi-dimensional feature space. The calculation process involves mapping high-dimensional feature vectors to a unified feature space, normalizing the features in each dimension, and eliminating the influence of dimension and scale differences. For example, for brake system failures, the system compares the similarity of vibration frequency abnormalities, temperature change characteristics, current fluctuations, and other dimensions in the target fault feature vector with historical cases. Case selection achieves precise matching by setting a similarity threshold. The threshold is set based on expert experience and big data statistical analysis to ensure that only highly relevant historical maintenance cases are included in the reference range. The screening process not only considers the similarity score, but also analyzes factors such as the timeliness of the case and the matching degree of the equipment type. For elevator traction system failures, the system may set a similarity threshold of 0.75 and only select historical maintenance cases with high feature vector matching. Maintenance measure clustering analysis is based on fault causes and component locations. The clustering algorithm classifies historical maintenance cases according to fault locations, maintenance technical routes, and the use of spare parts. For elevator brake systems, maintenance solution categories such as electrical control, mechanical component replacement, and lubrication and debugging may be formed. Each category represents a typical maintenance processing logic, reflecting the professionalism and systematicness of elevator maintenance. The solution priority ranking considers the maintenance effectiveness index and resource consumption score. The effectiveness index is calculated based on historical case maintenance success rate, equipment recovery time, and other indicators; the resource consumption score includes spare parts cost, labor cost, downtime, and other dimensions. Through multi-dimensional weighted evaluation, a maintenance solution priority list is generated.

[0124] The above describes the elevator operation state monitoring and early warning method based on multi-source data fusion in the embodiments of the application. The elevator operation state monitoring and early warning platform based on multi-source data fusion in the embodiments of the application is described below. Please refer to Figure 2 An embodiment of the elevator operation state monitoring and early warning platform based on multi-source data fusion in the embodiments of the application includes:

[0125] The processing module is configured to collect and preprocess vibration signals, noise, speed fluctuations, motor temperatures, load data, door machine parameters, and guide rail deformation data during the operation of the elevator to obtain normalized multi-source data.

[0126] The fusion module is configured to fuse and process the normalized multi-source data based on the improved entropy weight method, adjust the weight distribution by introducing a sensitivity correction factor, and obtain an elevator operation state fusion index.

[0127] The input module is configured to input the elevator operation state fusion index and original multi-source data into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism and a convolutional neural network for training and processing to obtain an elevator state prediction model.

[0128] The analysis module is configured to perform correlation analysis on parameter states and fault types according to an output result of the elevator state prediction model to obtain a set of correlation rules of parameter abnormal combinations and fault types.

[0129] The evaluation module is configured to evaluate the elevator operation state based on the set of correlation rules to obtain hierarchical early warning information and generate a fault diagnosis report based on the hierarchical early warning information.

[0130] Through the cooperation of the above components, the scientificity and accuracy of the elevator operation state monitoring are significantly improved through the innovative integration of multi-source sensor data acquisition, improved entropy weight data fusion, hybrid deep learning model construction and intelligent fault correlation analysis and other technical features. At the data acquisition level, through the cooperative work of various sensors such as acceleration sensors, sound sensors and speed sensors, multi-dimensional characteristics such as vibration signals, noise, speed fluctuations, motor temperature and load data during the operation of the elevator are comprehensively captured, laying a solid data foundation for subsequent in-depth analysis. The improved entropy weight method introduces a sensitivity correction factor, breaking through the limitations of traditional entropy weight distribution, and can more dynamically and accurately adjust the weights of different data sources, effectively balancing the importance of various sensor data. The innovation of the hybrid deep learning model lies in the integration of bidirectional long short-term memory network, attention mechanism and convolutional neural network, which fully utilizes the advantages of artificial intelligence algorithms in complex time series data processing. Specifically, the bidirectional LSTM network captures long-term dependencies, the attention mechanism realizes dynamic weight distribution of key time nodes, and the one-dimensional convolutional neural network extracts local feature patterns. The organic combination of the three algorithms significantly improves the accuracy and robustness of elevator operation state prediction. The introduction of the association rule mining technology realizes the deep correlation analysis between parameter abnormal states and fault types, and through the setting of dynamic threshold and time sequence information weighting, an intelligent early warning system for elevator faults is constructed. The feature vector similarity matching and multi-dimensional scoring mechanism are adopted in the fault case retrieval and maintenance scheme recommendation link, breaking through the traditional experience-driven maintenance mode and establishing a data-driven scientific maintenance paradigm. This multi-source data fusion and artificial intelligence algorithm empowerment technology path significantly improves the accuracy and early warning ability of elevator operation state monitoring.

[0131] Reference Figure 3 In the embodiment of the present application, a computer device, which can be a server, is also provided, and the internal structure thereof can be as shown in Figure 3The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store corresponding data in the embodiment. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the above method.

[0132] Those skilled in the art can understand that, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied.

[0133] The computer readable storage medium in the embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0134] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiment methods. Any reference to the memory, storage, database or other medium provided by the present application and used in the embodiments can include non-volatile and / or volatile memory. The non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM, etc.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described here.

[0136] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0137] The above-described and above-mentioned embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An elevator operation state monitoring and early warning method based on multi-source data fusion, characterized in that, The elevator operation state monitoring and early warning method based on multi-source data fusion comprises: Collect and preprocess vibration signals, noises, speed fluctuations, motor temperatures, load data, door machine parameters, and guide rail deformation data during elevator operation to obtain normalized multi-source data; Fuse the normalized multi-source data based on an improved entropy weight method, adjust the weight distribution by introducing a sensitivity correction factor, and obtain an elevator operation state fusion index; Input the elevator operation state fusion index and the original multi-source data into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism, and a convolutional neural network for training and processing to obtain an elevator state prediction model; According to the output results of the elevator state prediction model, analyze the correlation of parameter states and fault types to obtain a set of correlation rules of parameter abnormal combinations and fault types; Based on the set of correlation rules, evaluate the elevator operation state to obtain graded warning information, and generate a fault diagnosis report based on the graded warning information.

2. The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The collection and preprocessing of vibration signals, noises, speed fluctuations, motor temperatures, load data, door machine parameters, and guide rail deformation data during elevator operation to obtain normalized multi-source data comprises: Collecting original operation data through various sensors installed in the elevator system, including collecting vibration signals through an acceleration sensor, collecting noise data through a sound sensor, collecting speed fluctuation values through a speed sensor, collecting motor temperature through a temperature sensor, collecting load data through a load sensor, collecting door machine parameters through a door machine control system, and collecting guide rail deformation data through a displacement sensor to obtain elevator operation original monitoring data; According to the elevator operation original monitoring data, eliminate environmental interference through wavelet transform method, separate effective signals and noises according to directional noise reduction algorithm to obtain elevator operation denoising data; Based on the elevator operation denoising data, supplement the data missing part through linear interpolation or forward filling method, fill in the data blank points according to the time series continuity principle to obtain elevator operation complete data with a completeness of not less than 99.5%; Synchronize the elevator operation complete data through timestamp alignment function, reconstruct each sensor data stream according to a unified sampling period to obtain time series aligned elevator monitoring data; Identify and replace abnormal value points of the time series aligned elevator monitoring data using the 3σ principle, exclude data points deviating from the normal range based on statistical distribution characteristics to obtain abnormal value processed elevator monitoring data; Convert the abnormal value processed elevator monitoring data to the [-1, 1] interval according to the maximum and minimum value normalization method to eliminate the influence of different physical dimensions to obtain the normalized multi-source data. 3.The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The fusion processing of the normalized multi-source data based on the improved entropy weight method, adjusting the weight distribution by introducing a sensitivity correction factor, and obtaining an elevator operation state fusion index comprises: Construct a standardized data matrix for the normalized multi-source data, calculate the normalized values of positive and negative indicators to obtain a standardized processing matrix; According to the standardized processing matrix, the proportion and information entropy of each index at each sampling point are calculated, the proportion is multiplied by the logarithm and then summed to obtain the information entropy of each index; Based on the information entropy of each index, the difference coefficient and the initial weight are calculated, the difference coefficient is obtained by subtracting the information entropy from the constant 1, and the initial weight is obtained by dividing the difference coefficient by the sum, to obtain the initial weight configuration; The sensitivity correction factor is calculated by the ratio of the standard deviation of each index to the maximum standard deviation, and the ratio plus 1 is taken as the correction factor to obtain the weight adjustment parameter; The initial weight is modified using the weight adjustment parameter, and the final weight is obtained by multiplying the correction factor by the initial weight and normalizing it; Based on the weighted sum of the final weight and the standardized data, a sub-scene evaluation system for elevator operation under different working conditions is established to obtain the elevator operation state fusion index.

4. The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, The elevator operation state fusion index and the original multi-source data are input into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism and a convolutional neural network for training and processing to obtain an elevator state prediction model, including: The historical time series data is divided into sliding windows, and the input time length is set to 24 hours and the prediction time length is set to 8 hours to form a training sample set; The elevator operation state fusion index and the original multi-source data in the training sample set are input into a bidirectional long short-term memory network layer to capture long-term dependencies through forward propagation and back propagation to obtain sequence feature representation; Based on the sequence feature representation, an attention weight coefficient is constructed, and the importance distribution of each time point is calculated by a softmax function to obtain a weighted feature representation; The weighted feature representation is subjected to local feature extraction by a one-dimensional convolutional neural network to capture local pattern changes using a sliding convolution kernel to obtain a feature mapping result; The feature mapping result is input into a fully connected layer for feature fusion, and a dropout mechanism is introduced to control the risk of overfitting to obtain a prediction output vector; The mean square error loss is calculated according to the prediction output vector and the true label, and the parameters are updated by an Adam optimizer. Training is stopped when the loss on the validation set no longer decreases to obtain the elevator state prediction model.

5. The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 1, characterized in that, According to the output results of the elevator state prediction model, the parameter state and the fault type are analyzed for relevance to obtain a set of association rules between parameter abnormal combinations and fault types, including: The fusion index prediction value and the parameter prediction data output by the elevator state prediction model are discretized, and are divided into three levels of normal state, slight abnormal state and severe abnormal state to obtain parameter state discrete data; A transaction database is established based on the parameter state discrete data and historical fault records, and parameter state combinations are associated with corresponding fault types to obtain a state-fault mapping dataset; The state-fault mapping dataset is set with a minimum support threshold and a minimum confidence threshold, and the single parameter abnormal state that frequently appears is selected by support calculation to obtain a frequent item set; Based on the frequent item set, a self-connection operation is performed to construct a multi-parameter combination frequent item set by iterative connection and pruning to obtain a parameter combination feature set; extracting association rules from the parameter combination feature set, obtaining preliminary association rules of parameter abnormalities and fault types by calculating three indexes of support, confidence and lift; introducing a time sequence information weighting mechanism to the preliminary association rules, giving different weight values according to the data time, and obtaining an association rule set of the parameter abnormal combination and the fault type according to the weighted confidence and lift.

6. The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 5, characterized in that, The elevator operation state is evaluated based on the association rule set to obtain hierarchical warning information, and a fault diagnosis report is generated based on the hierarchical warning information, including: An adaptive dynamic threshold system is constructed to obtain the elevator operation parameter alarm threshold through the comprehensive calculation of parameter historical mean value, standard deviation, elevator type, service life and load state; According to the matching result of the association rule set, the elevator operation state is divided into four levels of normal, attention, warning and danger through comparative analysis of the parameter abnormal state and the rule antecedent, and the hierarchical warning information is obtained; The probability of each potential fault type is calculated based on the hierarchical warning information, and the fault probability distribution is obtained through the product operation of the confidence, lift and matching degree of the current parameter state of the association rule; Trend analysis is performed on the fault type data with high fault probability distribution, and the fault time estimation result is obtained by using the exponential smoothing prediction method to infer the fault development speed; Similar historical cases of the fault type are retrieved from the fault knowledge base, and relevant maintenance experience is extracted through fault feature similarity calculation to obtain a reference maintenance scheme; The hierarchical warning information, fault probability distribution, fault time estimation result and reference maintenance scheme are integrated into structured data to obtain the fault diagnosis report.

7. The elevator operation state monitoring and early warning method based on multi-source data fusion according to claim 6, characterized in that, The similar historical cases of the fault type are retrieved from the fault knowledge base, and relevant maintenance experience is extracted through fault feature similarity calculation to obtain a reference maintenance scheme, including: A feature vector representation is constructed for the fault type, which is converted into a structured representation through key parameter abnormal mode, equipment operation state description and fault phenomenon coding to obtain a target fault feature vector; Historical fault case data is extracted from the fault knowledge base, and the fault features, maintenance measures and maintenance effects of each case are extracted through structure to obtain a historical case feature set; Similarity calculation is performed based on the target fault feature vector and the historical case feature set, and the matching degree of the fault mode is quantified through the cosine similarity method to obtain a case similarity ranking list; High-similarity historical cases are selected from the case similarity ranking list, and the relevance of the reference cases is ensured by setting a minimum similarity threshold to obtain relevant historical maintenance cases; Cluster analysis is performed on the maintenance measures in the relevant historical maintenance cases, and the maintenance methods are grouped according to the relevance of fault causes and component positions to obtain maintenance scheme categories; Based on the maintenance scheme categories and historical success rate data, the effectiveness index and resource consumption score of each scheme are calculated to form a priority list through comprehensive sorting, and the reference maintenance scheme is obtained.

8. An elevator operation state monitoring and early warning platform based on multi-source data fusion, used to implement the elevator operation state monitoring and early warning method based on multi-source data fusion according to any one of claims 1-7, characterized in that, The elevator operation state monitoring and warning platform based on multi-source data fusion includes: The processing module is configured to collect and preprocess vibration signals, noise, speed fluctuations, motor temperature, load data, door machine parameters, and guide rail deformation data in the elevator operation process to obtain normalized multi-source data. The fusion module is configured to perform fusion processing on the normalized multi-source data based on an improved entropy weight method, adjust weight distribution by introducing a sensitivity correction factor, and obtain an elevator operation state fusion index. The input module is configured to input the elevator operation state fusion index and the original multi-source data into a hybrid deep learning model composed of a bidirectional long short-term memory network, an attention mechanism, and a convolutional neural network for training and processing to obtain an elevator state prediction model. The analysis module is configured to perform correlation analysis on parameter states and fault types based on an output result of the elevator state prediction model to obtain a set of correlation rules of parameter abnormal combinations and fault types. The evaluation module is configured to evaluate the elevator operation state based on the set of correlation rules, obtain hierarchical early warning information, and generate a fault diagnosis report based on the hierarchical early warning information.

9. A computer device, comprising: A memory and a processor are included, and the memory stores a computer program that can run on the processor. When the processor executes the computer program, the processor implements the elevator operation state monitoring and early warning method based on multi-source data fusion according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, causes the processor to perform the elevator operation state monitoring and early warning method based on multi-source data fusion according to any one of claims 1 to 7.

Images