Real-time elevator operation data acquisition and fault prediction system based on the Internet of Things

Abstract

This invention discloses an IoT-based real-time elevator operation data acquisition and fault prediction system, belonging to the field of fault detection technology. It includes: an operation data acquisition module for collecting elevator operation data based on sensor devices, including operation status, fault records, and maintenance records; a data storage module for classifying and compressing the elevator operation data; an operation data processing module for processing the stored elevator operation data to obtain processed data; and a fault prediction module for analyzing the processed data, establishing an elevator safety risk prediction model, and obtaining elevator fault detection results. This invention solves the technical problems of existing technologies, such as high cost of elevator operation data acquisition and storage, uncontrollable compression accuracy, reliance on manual experience for fault feature extraction, limited accuracy of prediction models, and lack of a complete closed-loop capability from fault detection to fault location.

Description

Technical Field

[0001] This invention belongs to the field of fault detection technology, specifically relating to an IoT-based system for real-time acquisition of elevator operation data and fault prediction. Background Technology

[0002] With the accelerating pace of urbanization, elevators, as an indispensable vertical transportation tool in high-rise buildings, are experiencing rapid and continuous growth in their number. The safety and reliability of elevator operation are directly related to the safety of people's lives and property; therefore, real-time monitoring of elevator operation and fault early warning have significant practical importance and application value.

[0003] Currently, the main technical solutions for elevator operation data collection and fault detection are as follows: A common approach is to rely on the fault diagnosis function built into the elevator's original equipment manufacturer (OEM) control system. This approach uses a programmable logic controller (PLC) or microprocessor in the elevator control cabinet to collect elevator operating parameters, triggering a fault alarm when the parameters exceed preset thresholds. However, this approach has significant shortcomings: first, the interface protocols of control systems from different manufacturers and models are incompatible, making it difficult for the data acquisition system to achieve unified access for elevators of different brands; second, the OEM control system only records instantaneous data at the moment of the fault, lacking the ability to store operating data long-term and analyze trends, thus failing to provide early warning of faults.

[0004] Another approach is to install independent sensor acquisition devices on the outside of the elevator. This approach involves installing vibration and temperature sensors in the elevator car, traction machine, and other components to collect physical signals during elevator operation and upload them to a monitoring platform for analysis. However, existing solutions often employ fixed sampling frequencies and compression strategies, failing to adequately consider the differences in data characteristics during different elevator operation phases (starting, constant speed, braking), resulting in massive data storage and high transmission and storage costs. Furthermore, existing solutions often sacrifice data accuracy for sensor data compression, making the compressed data quality insufficient for the accuracy requirements of fault prediction models.

[0005] In terms of data processing, existing technical solutions mostly use simple threshold discrimination or traditional statistical analysis (such as time-domain indicators such as mean, variance, and peak factor) for fault identification. Although these methods are simple to calculate and have good real-time performance, they are difficult to capture the complex time-series dependencies and nonlinear characteristics in elevator operation data. They are also insufficient in identifying early, minor faults (such as initial bearing wear or slight imbalance of the traction sheave), and are prone to missed or false alarms.

[0006] In terms of fault prediction, although some studies in recent years have introduced machine learning methods (such as support vector machines and random forests) for elevator fault classification, these methods usually require manual feature design. The quality of these features is highly dependent on the experience and knowledge of domain experts, and the models have limited ability to model long-term dependencies in time-series data. In addition, existing solutions mostly focus on the analysis of single data sources (such as vibration data or only operational status data), failing to fully utilize the correlation information between multi-source heterogeneous data (such as maintenance records, fault history, maintenance manual text, etc.), making it difficult to further improve fault location accuracy and prediction accuracy.

[0007] In summary, the existing elevator operation data acquisition and fault detection technologies have the following technical defects: (1) The sensor data acquisition and compression strategies lack adaptability, resulting in high storage and transmission costs; (2) The accuracy loss during data compression is uncontrollable, making it difficult to balance compression rate and reconstruction accuracy; (3) Fault feature extraction relies on manual experience and lacks a systematic quantitative indicator system; (4) The prediction model is insufficient in its ability to model the long-range dependence and nonlinear features of time series data, resulting in limited prediction accuracy; (5) Multi-source heterogeneous data has not been effectively integrated and utilized, lacking a complete closed-loop capability from fault detection to fault location. Summary of the Invention

[0008] The purpose of this invention is to provide an IoT-based real-time elevator operation data acquisition and fault prediction system to solve the technical problems in the prior art, such as high cost of elevator operation data acquisition and storage, uncontrollable compression accuracy, reliance on manual experience for fault feature extraction, limited accuracy of prediction models, and lack of a complete closed-loop capability from fault detection to fault location.

[0009] To achieve the above objectives, the present invention provides the following solution: an elevator operation data real-time acquisition and fault prediction system based on the Internet of Things, comprising: an operation data acquisition module, a data storage module, an operation data processing module, and a fault prediction module; The operation data acquisition module is used to collect elevator operation data based on sensor devices. The elevator operation data includes: operation status, fault records, and maintenance records. The data storage module is used to classify and compress the elevator operation data for storage; The operation data processing module is used to process the stored elevator operation data to obtain processed data; The fault prediction module is used to analyze the processed data, establish an elevator safety risk prediction model, and obtain elevator fault detection results.

[0010] More preferably, the elevator operation data is divided into three non-overlapping subsets: an enumerated feature set, an integer feature set, and a floating-point feature set. 。

[0011] More preferably, the data storage module uses a layered compression strategy to compress the elevator operation data to obtain compressed data, which is then stored and queried by the operation data processing module. The hierarchical compression strategy includes: edge dead zone lossy compression, cloud field-level adaptive encoding compression, prediction model-based residual compression, and feature-level dimensionality reduction compression.

[0012] More preferably, the edge dead zone lossy compression adopts an adaptive dead zone compression method and is implemented for floating-point features; For the d A floating-point feature, dead zone radius >0, when a new data point arrives Compared to the previous retained data point The absolute value of the deviation is less than the dead zone radius. If the current point is not found, discard it; otherwise, retain the current data point and update it. ; The criteria for retention are: ; In the formula, Indicates an indicator function.

[0013] More preferably, the dead zone radius includes: ; In the formula, Indicates the first d The scaling factor for each floating-point feature; Indicated by t i Local sliding window centered W load Inner d Standard deviation of each floating-point feature; Indicates the first d Minimum dead zone radius of a floating-point feature; in, ; ; In the formula, Indicates the local mean; L win This indicates the length of the local sliding window.

[0014] More preferably, the elevator safety risk prediction model includes an encoder and a decoder; The encoder employs a two-layer bidirectional LSTM structure to extract deep features from time-series data. ; ; ; In the formula, x t For the first t The input feature vector at time step h; t For bidirectional LSTM in t Concatenating the hidden states at any given moment; , They represent the first t The hidden states of the forward LSTM and the backward LSTM at each time step; , These represent the state update functions for the forward LSTM and the backward LSTM, respectively.

[0015] More preferably, the total loss function is: ; in, ; ; In the formula, LMSE and LCE These represent the mean squared error reconstruction loss and the weighted cross-entropy classification loss, respectively. β Indicates the classification loss weight coefficient; l Represents the L2 regularization coefficient; , These represent the predicted first and second halves of the series. i The sample belongs to the first c The probability of the type of fault and its true label; T Indicates the total number of sampling points; Indicates the first c Category weight coefficients for different types of faults; C Indicates the total number of fault types; This indicates a reconstructed sequence.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a five-layer hierarchical adaptive compression strategy, consisting of edge-end dead-zone lossy compression, cloud-based field-level adaptive encoding compression, prediction model-based residual compression, feature-level dimensionality reduction compression, and lossless entropy encoding. Edge-end dead-zone compression utilizes the local sliding window standard deviation to adaptively adjust the dead-zone radius, dynamically adjusting the compression intensity based on data fluctuation characteristics at different operational stages. The cloud-based strategy employs differentiated compression strategies such as dictionary encoding, differential encoding, and second-order differential encoding for different data types, including enumeration, integer, and floating-point types. Prediction model-based residual compression utilizes a lightweight BiLSTM network to capture data change trends, storing only the prediction residuals, significantly reducing information entropy. Actual measurements show that this invention achieves an overall compression rate of 3.2% (compression ratio approximately 31:1), which is 3 to 6 times higher than traditional time-series database compression methods (compression ratio approximately 5:1 to 10:1), significantly reducing the storage and transmission costs of elevator IoT systems. This invention embeds a precision guarantee mechanism into the multi-layer compression process. Edge-end dead-zone compression introduces a maximum permissible distortion constraint, adjusting the dead-zone radius parameter in reverse when the reconstruction error exceeds the limit. In residual compression, when the predicted residual exceeds a preset threshold, the original value is directly stored and a marker bit is added to ensure that sudden fault data does not suffer any accuracy loss. The peak signal-to-noise ratio (PSNR) of the reconstructed signal throughout the compression-decompression process is no less than 40dB, ensuring that the input data quality of the fault prediction module is not significantly affected while maintaining a high compression ratio. This invention adopts a VMD-BILSTM-AEAM hybrid prediction model, using variational mode decomposition (VMD) to adaptively denoise the noisy signal and using the Tuna Swarm Optimization (TSO) algorithm to automatically search for the optimal decomposition parameters. The encoder uses a two-layer bidirectional LSTM to extract deep bidirectional features from the time-series data, and the sliding window attention mechanism effectively captures local temporal dependencies. The autoencoder (AE) architecture identifies anomalies through reconstruction errors, reducing the dependence on labeled fault samples. At the same time, a class weight mechanism is introduced into the loss function to effectively alleviate the problem of imbalanced fault data. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the structural framework of the IoT-based elevator operation data real-time acquisition and fault prediction system provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1: like Figure 1 As shown, this embodiment provides an IoT-based real-time elevator operation data acquisition and fault prediction system, including: an operation data acquisition module, a data storage module, an operation data processing module, and a fault prediction module; the operation data acquisition module is used to collect elevator operation data based on sensor devices, and the elevator operation data includes: operation status, fault records, and maintenance records; the data storage module is used to classify and compress the elevator operation data for storage; the operation data processing module is used to process the stored elevator operation data to obtain processed data; the fault prediction module is used to analyze the processed data, establish an elevator safety risk prediction model, and obtain elevator fault detection results.

[0022] The operational data acquisition module uses various sensors to collect monitoring data from the traction system, car status, door operator system, and safety circuits. This includes: traction machine bearing temperature, traction machine housing vibration acceleration, brake shoe clearance, traction sheave speed, car triaxial vibration acceleration, car load, car leveling accuracy, door opening / closing speed, door motor current, door lock closure status, safety switch status, and speed governor operation records. The data collected from different sensors is aggregated, preliminarily cleaned, time-series encapsulated, and then encrypted before being transmitted to the data storage module.

[0023] In this embodiment, the data storage module employs a hierarchical compression strategy to compress elevator operation data, resulting in compressed data. This compressed data is then stored and queried by the operation data processing module. The hierarchical compression strategy includes: edge dead zone lossy compression, cloud-based field-level adaptive encoding compression, prediction model-based residual compression, and feature-level dimensionality reduction compression, finally culminating in lossless entropy encoding. Each layer can work independently or collaboratively.

[0024] Specifically, the elevator operation dataset collected by the data acquisition module is as follows: ,in, N This represents the total number of data points, and each data point... It is D dimensional vector,D Indicates the total number of feature dimensions; i Represents a time series index; Indicates the first i The sampling time of the first sampling moment d The original values ​​of each feature.

[0025] Will D The operational data features are divided into three non-overlapping subsets: enumerated feature set. F enum Integer feature set F int and floating-point feature sets F float The enumerated feature set includes elevator travel direction, floor number, and fault codes, with a limited value space. The integer feature set includes the number of door openings and closings, and the number of start-stop cycles, with values ​​being non-negative integers. The floating-point feature set includes temperature, vibration acceleration, and load value, with values ​​being real numbers.

[0026] Lossy compression with dead zones at the edges employs an adaptive dead-zone compression method, which is only applied to floating-point features; enumerated and integer features are not processed at this stage. For the first... d A floating-point feature, its dead zone radius >0, when a new data point arrives Compared to the previous retained data point The absolute value of the deviation is less than the dead zone radius. If the current point is not found, discard it; otherwise, retain the current data point and update it. .

[0027] The retention criteria are: ; In the formula, This indicates an indicator function; a value of 1 retains the current data point, while a value of 0 discards it.

[0028] A fixed dead zone radius is insufficient to accommodate vibration amplitude variations during periods when the elevator is not in operation. Therefore, this embodiment employs a relative dead zone radius: ; In the formula, Indicates the first d The scaling factor for each floating-point feature; Indicated by t i Local sliding window centered W load Inner d Standard deviation of each floating-point feature; Indicates the first dThe minimum dead zone radius of each floating-point feature prevents excessive compression and loss of detail when the signal amplitude is too small.

[0029] in, ; ; In the formula, Indicates the local mean; L win This indicates the length of the local sliding window.

[0030] Definition of the first d The number of data points for each floating-point feature after the first layer of compression is: The compression ratio of this layer is: .

[0031] The overall compression rate of the first layer is: .

[0032] To ensure the accuracy of subsequent fault prediction, a maximum permissible distortion constraint is introduced: ; In the formula, This represents an approximate value reconstructed through linear interpolation; Indicates the first d The maximum permissible mean square error of each floating-point feature is determined by the sensitivity analysis of the fault prediction model.

[0033] Maximum permissible distortion constraint is used for inverse adjustment. When the reconstruction error exceeds the limit, reduce To reduce the dead zone radius.

[0034] Cloud-based field-level adaptive encoding compression is used to employ differentiated lossless / near-lossless compression strategies for different types of features. Specifically, for enumerated features... d ∈ F enum Construct a dictionary mapping table: This maps each original string or category value to an integer codeword. The encoded codewords use variable-length encoding: each codeword is statistically analyzed. m Frequency of occurrence p d ( m The prefix code is generated using the Huffman algorithm, and the average code length satisfies: ; In the formula, l d ( m ) indicates the first m The length of Huffman code for one codeword; MThis indicates the number of bits in the original fixed-length code.

[0035] The compression gain is: .

[0036] For integer features d ∈ F int First-order differential coding is used: ; In the formula, Indicates the first d One feature in i The first-order difference value at time t.

[0037] for i =1, retain the original value. As an initial reference. Difference sequence The numerical range of the sequence is much smaller than that of the original sequence, so it can be represented with fewer bits.

[0038] Furthermore, the Zigzag mapping is used to convert potentially negative difference values ​​into non-negative integers: ; In the formula, This represents the non-negative integer encoded value after Zigzag mapping.

[0039] Zigzag mapping ensures that small absolute value differences are mapped to small numbers, which facilitates subsequent packaging and storage.

[0040] Floating-point data is both a key focus and a challenge in compression. This involves addressing the characteristics of floating-point data. d ∈ F float It adopts a three-level processing flow of second-order difference + precision scaling + LZ4 block compression.

[0041] Because raw sensor data is typically retained to 3-6 decimal places, but fault detection requires far less precision. (Definition of the first...) d Quantization step size of each floating-point feature ,in, k d The number of decimal places to retain (determined by the importance of the feature, such as 1 decimal place for temperature and 3 decimal places for vibration acceleration).

[0042] The quantized integer value is: .

[0043] Quantization error satisfies .

[0044] For quantized integer sequences Calculate the first-order difference: .

[0045] Next, calculate the second-order difference: .

[0046] For an elevator moving at a constant speed, its speed, vibration, and other characteristics are approximately constant, and the second-order difference values ​​are largely concentrated near 0, exhibiting extremely high compression potential. Only the following needs to be stored during storage: , as well as .

[0047] The second-order difference sequence is divided into fixed block sizes L. block The data is divided into 1024 blocks, and the LZ4 lossless compression algorithm is applied independently to each block. Because the data distribution within a block is similar, LZ4 can effectively identify recurring patterns and compress them significantly.

[0048] The timestamp sequence uses Delta-of-Delta encoding.

[0049] The second layer compresses the correlation between unused data. Therefore, this layer uses the VMD-BILSTM-AEAM model in the fault prediction module to perform prediction-residual encoding on floating-point features.

[0050] For the d ∈ F float A floating-point characteristic, utilizing its previous P Predicting the current value from historical values: ; In the formula, It is a lightweight BiLSTM prediction network with only one unidirectional or bidirectional LSTM layer and 32-64 hidden layer units. The initial weights are the weights of the first few layers after the encoder part has converged, and then frozen or fine-tuned. P This is the length of the history window.

[0051] Define the prediction residual as: .

[0052] Because the above prediction process captures the main trends in data change, the residual sequence The variance is much smaller than that of the original sequence. Its information entropy H ( r The information entropy is significantly lower than that of the original sequence. H ( x ).

[0053] The residual sequence is entropy encoded using exponential Golomb coding. Coding parameters. kEG Determined adaptively by the residual distribution: ; In the formula, L EG ( r , k ) indicates the use of parameters k Time-coded residual value r Number of bits required.

[0054] When prediction error Exceeding the preset threshold When the prediction fails (which may correspond to a sudden failure event), the residual is not stored. Instead, the original value is stored directly with a flag bit attached to ensure that the failure data is not destroyed by lossy compression.

[0055] The final data stored in the third layer is: ; In the formula, FLAG For specific flag bits.

[0056] The fourth layer utilizes the optimal feature subset selected by the ACDR algorithm to fundamentally reduce data dimensionality. D Let all D The correlation coefficient matrix among the features is , of which p Line number q Column elements r pq Features p With features q The Pearson correlation coefficient.

[0057] For the d Each feature is used to calculate its discriminant value. r d and representative d d : ; ; In the formula, This represents the correlation threshold.

[0058] Right now d d Features d The absolute value of the minimum correlation coefficient between it and all features with higher discriminative power.

[0059] Overall rating: R d = r d × δdAll features are sorted by R d Sort by largest to smallest, then select the first... K These features are used as the retained feature set. F keep ,in K Determined by the cumulative explained variance ratio: ; in R (d) For the sorted number d Each rating.

[0060] Only retain when storing F keep In K Features ( K D ), discard the remainder D K There are redundant features. Data with discarded features can be approximately recovered by retaining the regression model between features (but not as the compressed formal storage content).

[0061] The compression ratio of the fourth layer dimension is defined as follows: or (4) = K / D It is usually between 0.3 and 0.5.

[0062] After the first four layers of processing, all data has been converted into integer bitstreams or bytestreams. The final layer uses the Zstandard (zstd) algorithm for global lossless compression, utilizing its built-in dictionary training function to create a dedicated dictionary for different elevator models and device IDs, further improving the compression ratio.

[0063] Define the total number of bytes of the original data as Size raw The total number of bytes after compression is Size comp The corresponding compression ratio is: .

[0064] The overall compression ratio is the product of the compression ratios of each layer: or overall = or (1) × or (2) × or (3) × or (4) × or (5) ; In the formula, or (2) This represents the average compression ratio of each sub-coding method in the second layer; or (3) For residual coding compression ratio; or (5) This is the zstd lossless compression ratio.

[0065] Define the data to be reconstructed after decompression Compared with the original data x i Peak signal-to-noise ratio (PSNR) PSNR Used to evaluate compression quality: ; in, ; In the formula, MAX This represents the maximum possible value of the signal. MSE This represents the mean square error between the reconstructed signal and the original signal.

[0066] exist PSNR Under the constraint of ≥40dB, the compression method provided in this embodiment can stably achieve a compression ratio of 31:1, significantly reducing storage and transmission costs, while ensuring that the quality of input data to the fault prediction module is not significantly affected.

[0067] In the data processing module, the Isolation Forest algorithm is first used to identify outliers in the sensor data. An isolation tree is constructed by randomly selecting features and segmentation values; outliers have shorter path lengths because they are easily isolated. The outlier score calculation method includes: ; in, h ( x ) as a sample x Path length in the isolation tree; E ( h ( x The mean of all tree path lengths is denoted as ). c ( n ) represents the given number of samples n The average path length normalization factor under the given conditions. When s ( x , n If the value is greater than 0.6, the data point is considered an outlier and is removed.

[0068] Next, missing values ​​are handled. For consecutive missing ≤3 sampling points, linear interpolation is used to fill in the missing values; for consecutive missing >3 sampling points, the data is marked as missing and is not included in the feature calculation.

[0069] Finally, standardization is performed to unify the dimensions and sampling frequencies of different sensors. In this embodiment, Z-score standardization is used for floating-point features.

[0070] Each elevator operation is divided into three stages: starting, constant speed, and braking. The vibration characteristics of each stage have different fault indications. Starting stage: from zero speed to 90% of the rated speed; Constant speed stage: the speed is maintained at 90%-110% of the rated speed; Braking stage: from the speed decreasing to 90% of the rated speed to zero speed.

[0071] Vibration anomaly features were extracted, and the amplitude variation coefficient during the starting phase of each elevator operation was calculated. A i and amplitude frequency coefficient F i : ; ; in, a i,k For the first i The first phase of operation startup k Vertical vibration acceleration at each sampling point; and s i These represent the mean and standard deviation of the vibration acceleration during the start-up phase, respectively.

[0072] Vibration anomaly coefficient during startup for: ; Similarly, calculate the vibration anomaly coefficient during the braking phase. .

[0073] The vibration consistency coefficient due to traction sheave imbalance during each elevator operation is obtained by calculating the dynamic time warping (DTW) distance between the vibration acceleration data during the constant speed phase and the start-up and braking phases. C i : ; In the formula, , , They represent the first i The vertical vibration acceleration train of the second starting phase, the first i The vertical vibration acceleration sequence of the second uniform phase, the first i Vertical acceleration sequence during the secondary braking phase.

[0074] The DTW distance is calculated using a dynamic programming algorithm: ; In the formula, dπ Indicates the alignment path between time series X and Y; This indicates Euclidean distance.

[0075] No. i Vertical vibration anomaly during the second run It integrates the vibration characteristics of the three stages: .

[0076] The larger the vertical vibration anomaly value, the more significant the vertical vibration anomaly, which may indicate potential faults in the traction system (bearing wear, traction wheel imbalance) or guide rail system (uneven guide rail joints, guide shoe wear).

[0077] Modal decomposition was performed on the horizontal vibration acceleration of components such as the car, guide rails, and traction machine to extract the natural frequencies of each modal component. The natural frequencies of each component were then arranged according to their modal component sequence numbers to obtain a natural frequency sequence. The mean of the maximum mutual information coefficients between any two component natural frequency sequences was used as the approximate natural frequency coefficients for the horizontal vibration acceleration during elevator operation. H t .

[0078] Using a preset duration (e.g., 1 hour) as a monitoring period, all load data, vertical vibration anomalies, and natural frequency approximation coefficients within each monitoring period are arranged in ascending order to obtain the load sequence. L t Vertical vibration anomaly sequence Approximate sequence of natural frequencies H t .

[0079] No. t Elevator fault significance values ​​during each monitoring period S t for: ; in, This is the Pearson correlation coefficient.

[0080] S t The smaller the value in (0,1), the weaker the correlation between vibration anomaly and load change (i.e., vibration anomaly is independent of load change and is more likely to indicate mechanical failure), and the higher the risk of failure.

[0081] Next, multi-source data fusion was performed, including time-series + vibration fusion and time-series + document association. The vibration spectrum curve was converted into a time-series vector, and similarity matching was performed using Dynamic Time Warping (DTW) distance with a historical fault sample database to obtain the similarity matching results. 。

[0082] Sim ∈(0,1], the closer the value is to 1, the more similar the current vibration mode is to historical fault samples.

[0083] When a significant value for a specific fault is detected to be abnormal, the corresponding maintenance suggestion section in the maintenance manual is automatically associated with the SQL query to generate auxiliary decision-making information.

[0084] To address the feature redundancy problem in the original elevator operating parameter set, this embodiment uses the Attribute Correlation Density Ranking (ACDR) algorithm for feature selection to eliminate redundant parameters.

[0085] Based on the correlation coefficient between parameters d ij Calculate the discrimination of each parameter. r i and representative d i : ; ; In the formula, i Relevance threshold; Discrimination r i Used to measure parameters i It is highly correlated with many other parameters, and is representative. d i Measured parameters i Minimum distance to a higher discrimination parameter.

[0086] Overall score for each parameter r i = r i × d i .according to r i Sort by largest to smallest, then select the first... B Each parameter is used as the optimal feature subset.

[0087] Elevator speed curves are affected by random factors such as passengers entering and exiting the car, and contain a large amount of high-frequency noise. The Tuna Swarm Optimization (TSO) algorithm is used to optimize the variational mode decomposition (VMD) parameters, achieving adaptive denoising of the speed sequence.

[0088] VMD will use the original signal f ( t It can be decomposed into B intrinsic mode functions (IMFs). u b ( t The optimization problem can be expressed as: ; ; In the formula, oh b For the first b The center frequency of each IMF; This represents the Dirac unit impulse function.

[0089] The TSO algorithm automatically searches for the optimal decomposition level by simulating the foraging behavior of tuna groups. B and penalty factor α Combinatorial approach, using the minimization of envelope entropy as the fitness function: ; in, E b For the first b The envelope entropy of each IMF component. After TSO optimization, high-frequency noise components are removed, and the denoised velocity signal is reconstructed.

[0090] This embodiment uses a VMD-BILSTM-AEAM (variational mode decomposition-bidirectional long short-term memory-autoencoder-attention mechanism) hybrid model as the elevator safety risk prediction model. The structure includes an encoder and a decoder.

[0091] The encoder employs a two-layer bidirectional LSTM (BILSTM) structure to extract deep features from time-series data. ; ; ; In the formula, x t For the first t The input feature vector at time step h (after ACDR selection and TSO-VMD denoising) t For bidirectional LSTM in t Concatenating the hidden states at any given moment; , They represent the first t The hidden states of the forward LSTM and the backward LSTM at each time step; , These represent the state update functions for the forward LSTM and the backward LSTM, respectively.

[0092] A sliding window attention mechanism is introduced between the encoder and decoder to capture local temporal dependencies: ; In the formula, c t Represents the attention context vector; W To adjust the sliding window size; at,i This represents the attention weight.

[0093] Attention weight α t,i Calculated using Softmax normalization: ; Among them, W a This is a learnable attention weight matrix.

[0094] The decoder uses an LSTM network to process the attention context vector c. t Compared with the previous time step output y of the target sequence t-1 The concatenated data is used as input: ; In the formula, LSTMdec This represents the LSTM state update function of the decoder.

[0095] Reconstructed sequence Output from the fully connected layer: ; In the formula, W d b d These represent the weight matrix and bias vector of the fully connected output layer, respectively.

[0096] The total loss function is: ; in, ; ; In the formula, LMSE and LCE These represent the mean squared error reconstruction loss and the weighted cross-entropy classification loss, respectively. β Indicates the classification loss weight coefficient; l Represents the L2 regularization coefficient; , These represent the predicted first and second halves of the series. i The sample belongs to the first c The probability of the type of fault and its true label; T Indicates the total number of sampling points; Indicates the first c Category weight coefficients for different types of faults; CThis indicates the total number of fault types.

[0097] The patience value is set to 20 rounds. If the validation set loss does not decrease for 20 consecutive rounds, training is terminated and the optimal model weights are restored. The Adam optimizer is used, with an initial learning rate of... or =10 3 By combining the strong Wolfe condition with step size search, we can ensure convergence to a local optimum.

[0098] Strong Wolfe conditions include: conditions for decreasing function value. f (x k + α k d k )≤ f (x k )+ c 1 α k f (x k ) T d k Strong curvature condition: | f (x k + α k d k ) T d k |≤ c 2 | f (x k ) T d k |; where x k d k These represent the model parameter vector and the search direction at the k-th iteration, respectively; α k This represents the step size factor for the k-th iteration; f (x k ) represents the loss function in x k Gradient vector at ; 0 < c 1< c 2 < 1.

[0099] When the optimization process gets stuck in a local optimum, a reactivation mechanism is introduced to restart the optimization process and avoid serious deviations in the solution. L reactive = L total + c 1 (stagnation); in, c 1 is the amplification factor (slightly greater than 0), and 1(stagnation) is the stagnation indicator function.

[0100] For each test sample, calculate the reconstruction error vector: .

[0101] The reconstruction errors of each feature sequence are fused using a SoftMax classifier to output the fault type classification result.

[0102] For the first t Significant fault values ​​for each monitoring period S t Perform normalization and set fault diagnosis thresholds. t :like ≤ t The elevator is operating normally; if > t An elevator malfunction has triggered an alarm. Threshold t Determined through cross-validation; express S t Normalized fault significance value after normalization.

[0103] The output includes: elevator health score, fault warning notification, and maintenance suggestions. The elevator health score is as follows: The higher the score, the better the health status. Fault warning push notifications include: predicting potential fault types and possible time windows (e.g., "abnormal growth trend in brake shoe clearance, expected to exceed limits within 5 days"). Maintenance recommendations include: automatically linking to the corresponding chapters in the maintenance manual and pushing specific maintenance operation suggestions.

[0104] For elevators with detected faults, an enhanced recurrent graph convolutional network (e-RGCN) is further employed to locate the faulty components. The vibration signal time series is transformed into a graph structure. Two vertices v i and v j If the following conditions are met: Then there exists an edge; where, The recursion radius threshold is given by R. The recursion matrix R is... ij for: ; In the formula, Θ( ) is the Heaviside step function.

[0105] In this embodiment, the e-RGCN classification network consists of two GCN layers, one fully connected layer, and a Softmax classification layer. The GCN layer extracts graph features from the association graph, and the fully connected layer acts as a classifier, outputting the fault type and possible faulty components. Through this network, the source of the fault (traction machine bearing, guide rail, gantry crane, safety circuit, etc.) can be more accurately located based on fault detection, significantly improving maintenance efficiency.

[0106] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A real-time elevator operation data acquisition and fault prediction system based on the Internet of Things, characterized in that, include: It includes a data acquisition module, a data storage module, a data processing module, and a fault prediction module. The operation data acquisition module is used to collect elevator operation data based on sensor devices. The elevator operation data includes: operation status, fault records, and maintenance records. The data storage module is used to classify and compress the elevator operation data for storage; The operation data processing module is used to process the stored elevator operation data to obtain processed data; The fault prediction module is used to analyze the processed data, establish an elevator safety risk prediction model, and obtain elevator fault detection results.

2. The elevator operation data real-time acquisition and fault prediction system based on the Internet of Things as described in claim 1, characterized in that, The elevator operation data is divided into three non-overlapping subsets: an enumerated feature set, an integer feature set, and a floating-point feature set. 。 3. The real-time elevator operation data acquisition and fault prediction system based on the Internet of Things as described in claim 2, characterized in that, The data storage module uses a hierarchical compression strategy to compress elevator operation data, obtains compressed data, stores the compressed data, and provides it for the operation data processing module to query. The hierarchical compression strategy includes: edge dead zone lossy compression, cloud field-level adaptive encoding compression, prediction model-based residual compression, and feature-level dimensionality reduction compression.

4. The real-time elevator operation data acquisition and fault prediction system based on the Internet of Things as described in claim 3, characterized in that, The edge dead zone lossy compression adopts an adaptive dead zone compression method and is implemented for floating-point features; For the d A floating-point feature, dead zone radius >0, when a new data point arrives Compared to the previous retained data point The absolute value of the deviation is less than the dead zone radius. If the current point is not found, discard it; otherwise, retain the current data point and update it. ; The criteria for retention are: ； In the formula, Indicates an indicator function.

5. The real-time elevator operation data acquisition and fault prediction system based on the Internet of Things as described in claim 4, characterized in that, The dead zone radius includes: ； In the formula, Indicates the first d The scaling factor for each floating-point feature; Indicated by t i Local sliding window centered W load Inner d Standard deviation of each floating-point feature; Indicates the first d Minimum dead zone radius of a floating-point feature; in, ； ； In the formula, Indicates the local mean; L win This indicates the length of the local sliding window.

6. The real-time elevator operation data acquisition and fault prediction system based on the Internet of Things as described in claim 1, characterized in that, The elevator safety risk prediction model includes an encoder and a decoder; The encoder employs a two-layer bidirectional LSTM structure to extract deep features from time-series data. ； ； ； In the formula, x t For the first t The input feature vector at time step; h t For bidirectional LSTM in t Concatenating the hidden states at any given moment; , They represent the first t The hidden states of the forward LSTM and the backward LSTM at each time step; , These represent the state update functions for the forward LSTM and the backward LSTM, respectively.

7. The real-time elevator operation data acquisition and fault prediction system based on the Internet of Things as described in claim 6, characterized in that, The total loss function is: ； in, ； ； In the formula, LMSE and LCE These represent the mean squared error reconstruction loss and the weighted cross-entropy classification loss, respectively. β Indicates the classification loss weight coefficient; λ Represents the L2 regularization coefficient; , These represent the predicted first... i The sample belongs to the first c The probability of the type of fault and its true label; T Indicates the total number of sampling points; Indicates the first c Category weight coefficients for different types of faults; C Indicates the total number of fault types; This indicates a reconstructed sequence.

Images