Fault Detection Method for Elevator Traction System Based on Artificial Intelligence

Abstract

The present invention discloses a fault detection method for an elevator traction system based on artificial intelligence, belonging to the technical field of intelligent elevator operation and maintenance. The method includes collecting elevator traction system data, enhancing elevator traction system data, constructing a fault detection model, and detecting faults in the elevator traction system. The present invention uses a recursive generative adversarial network model for data enhancement, which can better capture the temporal dependence of data, and make the synthetic data closer to the actual data in statistical characteristics, improving the authenticity of the synthetic data, thereby reducing the bias during model training, enhancing the accuracy and reliability of elevator traction system fault detection; uses a parallel deep convolutional coupled graph convolutional network to construct a fault detection model for the elevator traction system, fully extracting local and global features, effectively processing multi-scale features, and capturing the complex relationships between data through graph structure modeling, thereby improving the accuracy of elevator traction system fault detection and providing more accurate data support for intelligent elevator operation and maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent elevator operation and maintenance, and specifically refers to a fault detection method for an elevator traction system based on artificial intelligence. Background Art

[0002] For the fault detection of an elevator traction system based on artificial intelligence, artificial intelligence technology is used to monitor and analyze the data generated during the operation of the elevator traction system, and faults are identified in a timely manner. The aim is to automatically analyze the state of the elevator traction system, achieve intelligent fault detection, thereby improving the elevator maintenance efficiency and safety, and promoting the intelligent development of elevator operation and maintenance.

[0003] However, in the existing elevator traction system fault detection process, there are technical problems as follows: the fault data of the elevator traction system is scarce, resulting in the model detection being prone to bias towards normal data, affecting the detection accuracy. In addition, the elevator traction system data has temporal characteristics, while the traditional generative adversarial network mainly processes static data and is difficult to effectively generate high-quality temporal data, resulting in poor quality of synthetic data and further reducing the detection accuracy; the faults of the elevator traction system are usually caused by the collaborative action of multiple components, so the global modeling ability of the detection model is required to be high, and the elevator traction system fault detection involves various sensor data and has multi-modal complexity. The traditional fault detection model is difficult to fully explore the complex relationships between multi-modal data and has insufficient multi-scale feature extraction ability, thus affecting the detection accuracy. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides an elevator traction system fault detection method based on artificial intelligence. In the process of fault detection of the existing elevator traction system, there is a shortage of fault data of the elevator traction system, resulting in the model detection being prone to bias towards normal data and affecting the detection accuracy. In addition, the data of the elevator traction system has temporal characteristics, while the traditional generative adversarial network mainly processes static data and is difficult to effectively generate high-quality temporal data, resulting in poor quality of synthetic data and further reducing the detection accuracy. To solve this technical problem, the present solution creatively uses a recursive generative adversarial network model for data augmentation. By introducing a recurrent neural network based on long short-term memory units, it can better capture the temporal dependence of the data. And by improving the model loss function, the synthetic data is made closer to the actual data in statistical characteristics, improving the authenticity of the synthetic data, thereby reducing the bias during model training and enhancing the accuracy and reliability of elevator traction system fault detection. In the process of fault detection of the existing elevator traction system, the faults of the elevator traction system are usually caused by the collaborative action of multiple components, so the global modeling ability of the detection model is required to be high. And the fault detection of the elevator traction system involves various sensor data and has multi-modal complexity. Traditional fault detection models are difficult to fully explore the complex relationships between multi-modal data and have insufficient multi-scale feature extraction ability, which further affects the detection accuracy. To solve this technical problem, the present solution creatively uses a parallel deep convolutional coupled graph convolutional network to construct an elevator traction system fault detection model, fully extract local and global features, effectively process multi-scale features, and capture the complex relationships between data through graph structure modeling, thereby improving the accuracy of elevator traction system fault detection and providing more accurate data support for intelligent elevator operation and maintenance.

[0005] The technical solution adopted by the present invention is as follows: The elevator traction system fault detection method based on artificial intelligence provided by the present invention includes the following steps:

[0006] Step S1: Collect elevator traction system data;

[0007] Step S2: Augment elevator traction system data;

[0008] Step S3: Construct a fault detection model;

[0009] Step S4: Detect faults in the elevator traction system.

[0010] Further, in step S1, the collection of elevator traction system data is specifically to install sensors on the elevator's traction machine, motor bearings, and wire rope parts, collect elevator traction system data and perform data annotation to obtain elevator traction system status labels, and then through data preprocessing of the elevator traction system data, obtain the actual data of the elevator traction system.

[0011] Further, in step S2, the enhanced elevator traction system data is used to alleviate the data imbalance problem. Specifically, based on the elevator traction system status labels and the actual elevator traction system data, a recursive generative adversarial network model is used for data enhancement to obtain the enhanced elevator traction system data;

[0012] The recursive generative adversarial network model includes a generator and a discriminator;

[0013] The steps for enhancing the elevator traction system data include the following:

[0014] Step S21: Construct a generator for synthesizing data. Specifically, receive the minority class data in the actual elevator traction system data as the actual system data, introduce a recurrent neural network based on long short-term memory units in the generator to learn the temporal dependence features in the data, and generate the synthesized elevator traction system data;

[0015] Step S22: Construct a discriminator for distinguishing between actual data and synthesized data. Specifically, receive the minority class data in the actual elevator traction system data as the actual system data, and receive the synthesized elevator traction system data. Introduce a recurrent neural network based on long short-term memory units in the discriminator to judge the authenticity of the data and obtain the discrimination result;

[0016] Step S23: Improve the model loss function. Specifically, construct the generator loss function by combining the moment matching loss function, and construct the discriminator loss function by introducing a gradient penalty term;

[0017] The moment matching loss function is used to retain the statistical characteristics of the actual data, and its calculation formula is:

[0018] ;

[0019] In the formula, L mm is the moment matching loss function, t max is the maximum time step, which is used to represent the temporal step length of the data, is the square of the L2 norm, t is the time step index, is the mean of the actual system data at the t-th time step, is the mean of the synthesized elevator traction system data at the t-th time step, is the standard deviation of the actual system data at the t-th time step, is the standard deviation of the synthesized elevator traction system data at the t-th time step, G(z) is the synthesized elevator traction system data, and z is the noise data point, specifically referring to the data point randomly sampled from the noise distribution;

[0020] The calculation formula of the generator loss function is:

[0021] ;

[0022] In the formula, L RGAN is the generator loss function, is the first weight of the generator, is the expectation of the synthetic data of the elevator traction system, which is used to measure the authenticity of the synthetic data of the elevator traction system, p z is the noise distribution, and D(·) is the discriminator scoring function, which is used to represent the authenticity of the synthetic data of the elevator traction system, is the second weight of the generator;

[0023] The calculation formula of the discriminator loss function is:

[0024] ;

[0025] In the formula, L D is the discriminator loss function, is the expectation of the actual data of the system, p x is the distribution of the actual data of the system, x is the actual data point, specifically referring to the data point randomly sampled from the distribution of the actual data of the system, is the weight of the gradient penalty term, is the expectation of the interpolated data, is the distribution of the interpolated data, and the distribution of the interpolated data specifically refers to the data distribution obtained by linearly interpolating the actual data of the system and the synthetic data of the elevator traction system, is the interpolated data point, is the gradient of the discriminator to the interpolated data, is the L2 norm;

[0026] Step S24: Train the recursive generative adversarial network model, specifically by alternately training the generator and the discriminator to perform recursive generative adversarial network model training to obtain the recursive generative adversarial network model;

[0027] Step S25: Data augmentation, specifically by using the recursive generative adversarial network model to synthesize the minority class data in the actual data of the elevator traction system to generate the synthetic data of the elevator traction system, and then merging the synthetic data of the elevator traction system and the actual data of the elevator traction system to obtain the enhanced data of the elevator traction system.

[0028] Furthermore, in step S3, the construction of the fault detection model is used to construct the fault detection model of the elevator traction system. Specifically, based on the state labels of the elevator traction system and the enhanced data of the elevator traction system, a parallel deep convolutional coupled graph convolutional network is used to construct the fault detection model of the elevator traction system;

[0029] The parallel deep convolutional coupled graph convolutional network includes an input layer, a parallel deep convolutional subnet, a feature fusion attention block, a parallel graph convolutional subnet, and a classification output layer;

[0030] The parallel deep convolutional subnet is used to capture multi-scale features in parallel;

[0031] The feature fusion attention block is used to fuse multi-scale features and focus on key features;

[0032] The parallel graph convolutional subnet is used to further capture the relationships between features;

[0033] The steps for constructing the fault detection model are as follows:

[0034] Step S31: Construct an input layer, specifically, receive elevator traction system enhanced data through the input layer to construct a fault detection input sample;

[0035] Step S32: Construct a parallel deep convolutional subnet, specifically, set a first deep convolutional subnet and a second deep convolutional subnet in the parallel deep convolutional subnet. The first deep convolutional subnet and the second deep convolutional subnet are in a parallel structure. Extract multi-scale features of the fault detection input sample through the parallel deep convolutional subnet to obtain a fault detection global feature and a fault detection local feature, including the following steps:

[0036] Step S321: Construct a first deep convolutional subnet, specifically, set three convolutional layers in the first deep convolutional subnet, and set a batch normalization layer after each convolutional layer. Set the convolutional kernel size in the first deep convolutional subnet to 7×1, and extract the fault detection global feature through the first deep convolutional subnet;

[0037] Step S322: Construct a second deep convolutional subnet, specifically, set three convolutional layers in the second deep convolutional subnet, and set a batch normalization layer after each convolutional layer. Set the convolutional kernel size in the second deep convolutional subnet to 3×1, and extract the fault detection local feature through the second deep convolutional subnet;

[0038] Step S33: Construct a feature fusion attention block, specifically, set a feature fusion layer, an attention mechanism layer, and a fully connected layer in the feature fusion attention block. Process the fault detection global feature and the fault detection local feature through the feature fusion attention block to obtain a fault detection fusion enhanced feature, including the following steps:

[0039] Step S331: Construct a feature fusion layer, specifically, in the feature fusion layer, obtain a fault detection fusion feature through a dot product operation on the fault detection global feature and the fault detection local feature;

[0040] Step S332: Construct an attention mechanism layer. Specifically, in the attention mechanism layer, apply the channel attention mechanism to perform weighted calculation on the fault detection fusion features to obtain the fault detection attention enhanced features;

[0041] Step S333: Construct a fully connected layer. Specifically, perform feature dimensionality reduction on the fault detection attention enhanced features through the fully connected layer to obtain the fault detection fusion enhanced features;

[0042] Step S34: Construct a parallel graph convolutional subnet, including the following steps:

[0043] Step S341: Construct a graph structure. Specifically, use the fault detection fusion enhanced features of each sample as a node feature to establish a node feature matrix. By measuring the similarity between each node feature, establish an adjacency matrix and perform a normalization operation to obtain the fault detection adjacency matrix;

[0044] Step S342: Construct parallel first-order, second-order, and third-order graph convolutional layers to process the fault detection adjacency matrix in parallel, extract multi-level information, splice the extraction results, and perform feature dimensionality reduction through a fully connected layer to obtain the fault detection multi-level features;

[0045] Step S35: Construct a classification output layer. Specifically, set a fully connected layer and a softmax activation function in the classification output layer to classify the fault detection multi-level features to obtain the model output result;

[0046] Step S36: Model training. Specifically, construct a parallel deep convolutional coupled graph convolutional network by the constructed input layer, the constructed parallel deep convolutional subnet, the constructed feature fusion attention block, the constructed parallel graph convolutional subnet, and the constructed classification output layer, and perform model training to obtain an elevator traction system fault detection model.

[0047] Further, in step S4, the elevator traction system fault detection is specifically to perform elevator traction system fault detection through the elevator traction system fault detection model to obtain the elevator traction system state. According to the elevator traction system state, generate an elevator traction system fault detection report for assisting the management personnel to maintain the elevator traction system.

[0048] The beneficial effects achieved by the present invention using the above solution are as follows:

[0049] (1)In the process of detecting faults in the existing elevator traction system, there is a scarcity of fault data in the elevator traction system, which leads to the model detection being prone to bias towards normal data and affects the detection accuracy. In addition, the data of the elevator traction system has temporal characteristics, while traditional generative adversarial networks mainly process static data and are difficult to effectively generate high-quality temporal data, resulting in poor quality of synthetic data and further reducing the detection accuracy. To solve this technical problem, this solution creatively uses a recursive generative adversarial network model for data augmentation. By introducing a recurrent neural network based on long short-term memory units, it can better capture the temporal dependence of the data. And by improving the model loss function, the synthetic data is made closer to the actual data in statistical characteristics, improving the authenticity of the synthetic data, thereby reducing the bias during model training and enhancing the accuracy and reliability of elevator traction system fault detection.

[0050] (2)In the process of detecting faults in the existing elevator traction system, faults in the elevator traction system are usually caused by the collaborative action of multiple components, so a high global modeling ability of the detection model is required. Moreover, elevator traction system fault detection involves various sensor data and has multi-modal complexity. Traditional fault detection models are difficult to fully explore the complex relationships between multi-modal data and have insufficient multi-scale feature extraction ability, thus affecting the detection accuracy. To solve this technical problem, this solution creatively uses a parallel deep convolutional coupled graph convolutional network to construct an elevator traction system fault detection model, fully extract local and global features, effectively process multi-scale features, and capture the complex relationships between data through graph structure modeling, thereby improving the accuracy of elevator traction system fault detection and providing more accurate data support for intelligent elevator operation and maintenance. Brief Description of the Drawings

[0051] Figure 1 It is a schematic flowchart of the method for detecting faults in an elevator traction system based on artificial intelligence provided by the present invention;

[0052] Figure 2 It is a schematic flowchart of step S2;

[0053] Figure 3 It is a schematic flowchart of step S3;

[0054] Figure 4 It is a schematic flowchart of step S33.

[0055] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. Detailed Description of the Invention

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation to the present invention.

[0058] Embodiment 1. Refer to Figure 1 , the method for detecting faults in an elevator traction system based on artificial intelligence provided by the present invention includes the following steps:

[0059] Step S1: Collect data of the elevator traction system;

[0060] Step S2: Enhance the data of the elevator traction system;

[0061] Step S3: Build a fault detection model;

[0062] Step S4: Detect faults in the elevator traction system.

[0063] Embodiment 2. Refer to Figure 1 , based on the above embodiment, in step S1, the collection of the data of the elevator traction system is specifically to install sensors on the traction machine, motor bearing, and wire rope part of the elevator, collect the data of the elevator traction system and perform data annotation to obtain the state label of the elevator traction system. Then, through data preprocessing on the data of the elevator traction system, the actual data of the elevator traction system is obtained;

[0064] The state labels of the elevator traction system include normal operation, current overload, abnormal vibration, abnormal temperature, abnormal acceleration, and abnormal speed;

[0065] The data preprocessing includes noise filtering, data cleaning, Z-score standardization, data alignment, and data dimensionality reduction;

[0066] The noise filtering filters data noise through wavelet transform;

[0067] The data cleaning includes removing duplicate values, eliminating outliers, and filling in missing values;

[0068] The data alignment aligns the timestamps in the data through dynamic time warping;

[0069] The data dimensionality reduction reduces the data dimensionality through principal component analysis;

[0070] The data of the elevator traction system includes vibration signals, current signals, voltage signals, temperature, noise, acceleration, and the rotational speed of the traction machine.

[0071] Example 3, refer to Figure 1 and Figure 2 , this example is based on the above example. In step S2, the enhanced elevator traction system data is used to alleviate the data imbalance problem. Specifically, according to the elevator traction system status label and the actual data of the elevator traction system, a recursive generative adversarial network model is used for data enhancement to obtain the enhanced data of the elevator traction system;

[0072] The recursive generative adversarial network model includes a generator and a discriminator;

[0073] The steps for enhancing the elevator traction system data include the following:

[0074] Step S21: Construct a generator for synthesizing data. Specifically, receive the minority class data in the actual data of the elevator traction system as the actual system data, introduce a recurrent neural network based on long short-term memory units in the generator to learn the temporal dependence features in the data, and generate the synthetic data of the elevator traction system;

[0075] Step S22: Construct a discriminator for distinguishing actual data and synthetic data. Specifically, receive the minority class data in the actual data of the elevator traction system as the actual system data, and receive the synthetic data of the elevator traction system. Introduce a recurrent neural network based on long short-term memory units in the discriminator to judge the authenticity of the data and obtain the discrimination result;

[0076] Step S23: Improve the model loss function. Specifically, construct the generator loss function by combining the moment matching loss function, and construct the discriminator loss function by introducing a gradient penalty term. The gradient penalty term is used to constrain the discriminator to avoid overfitting;

[0077] The moment matching loss function is used to preserve the statistical characteristics of the actual data, and the calculation formula is:

[0078] ;

[0079] In the formula, L mm is the moment matching loss function, t max is the maximum time step, which is used to represent the temporal step length of the data, is the square of the L2 norm, t is the time step index, is the mean of the actual system data at the t-th time step, is the mean of the synthetic data of the elevator traction system at the t-th time step, is the standard deviation of the actual system data at the t-th time step, is the standard deviation of the synthetic data of the elevator traction system at the t-th time step, G(z) is the synthetic data of the elevator traction system, and z is the noise data point, specifically referring to the data point randomly sampled from the noise distribution;

[0080] The calculation formula of the generator loss function is:

[0081] ;

[0082] In the formula, L RGAN is the generator loss function, is the first weight of the generator, is the expectation of the synthetic data of the elevator traction system, used to measure the authenticity of the synthetic data of the elevator traction system, p z is the noise distribution, D(·) is the discriminator scoring function, used to represent the authenticity of the synthetic data of the elevator traction system, is the second weight of the generator;

[0083] The calculation formula of the discriminator loss function is:

[0084] ;

[0085] In the formula, L D is the discriminator loss function, is the expectation of the actual system data, p x is the actual system data distribution, x is the actual data point, specifically referring to the data point randomly sampled from the actual system data distribution, is the gradient penalty term weight, is the expectation of the interpolated data, is the interpolated data distribution, and the interpolated data distribution specifically refers to the data distribution obtained by linearly interpolating the actual system data and the synthetic data of the elevator traction system, is the interpolated data point, is the gradient of the discriminator to the interpolated data, is the L2 norm;

[0086] Step S24: Train the recursive generative adversarial network model, specifically by alternately training the generator and the discriminator to perform recursive generative adversarial network model training to obtain the recursive generative adversarial network model;

[0087] Step S25: Data augmentation, specifically, the minority class data in the actual data of the elevator traction system is synthesized through a recursive generative adversarial network model to generate synthetic data of the elevator traction system. Then, the synthetic data of the elevator traction system and the actual data of the elevator traction system are merged to obtain enhanced data of the elevator traction system;

[0088] Table 1 is a comparison table of the performance of the elevator traction system fault detection model before and after enhancing the data of the elevator traction system in Embodiment 4 of the present invention. As shown in the table, the F1 score is the harmonic mean of the precision and recall. The higher the F1 score, the better the model performance;

[0089] Table 1 Comparison table of the performance of the elevator traction system fault detection model before and after enhancing the data of the elevator traction system

[0090]

[0091] By performing the above operations, in the existing elevator traction system fault detection process, there is a shortage of elevator traction system fault data, resulting in the model detection being prone to bias towards normal data and affecting the detection accuracy. In addition, the elevator traction system data has temporal characteristics, while traditional generative adversarial networks mainly process static data and are difficult to effectively generate high-quality temporal data, resulting in poor quality of synthetic data and further reducing the detection accuracy. To solve this technical problem, this solution creatively uses a recursive generative adversarial network model for data augmentation. By introducing a recurrent neural network based on long short-term memory units, it can better capture the temporal dependence of the data. And by improving the model loss function, the synthetic data is made closer to the actual data in statistical characteristics, improving the authenticity of the synthetic data, thereby reducing the bias during model training and enhancing the accuracy and reliability of elevator traction system fault detection.

[0092] Embodiment 4, refer to Figure 1 、 Figure 3 and Figure 4 In this embodiment, based on the above embodiment, in step S3, the construction of the fault detection model is used to construct an elevator traction system fault detection model. Specifically, according to the elevator traction system status labels and the enhanced data of the elevator traction system, a parallel deep convolutional coupled graph convolutional network is used to construct an elevator traction system fault detection model;

[0093] The parallel deep convolutional coupled graph convolutional network includes an input layer, a parallel deep convolutional subnet, a feature fusion attention block, a parallel graph convolutional subnet, and a classification output layer;

[0094] The parallel deep convolutional subnet is used to capture multi-scale features in parallel;

[0095] The feature fusion attention block is used to fuse multi-scale features and focus on key features;

[0096] The parallel graph convolutional subnet is used to further capture the relationships between features;

[0097] The construction of the fault detection model includes the following steps:

[0098] Step S31: Construct an input layer, specifically, receive the elevator traction system enhanced data through the input layer to construct a fault detection input sample;

[0099] Step S32: Construct a parallel deep convolutional subnet. Specifically, set a first deep convolutional subnet and a second deep convolutional subnet in the parallel deep convolutional subnet. The first deep convolutional subnet and the second deep convolutional subnet are in a parallel structure. Extract multi-scale features of the fault detection input sample through the parallel deep convolutional subnet to obtain the fault detection global feature and the fault detection local feature, including the following steps:

[0100] Step S321: Construct a first deep convolutional subnet. Specifically, set three convolutional layers in the first deep convolutional subnet, and set a batch normalization layer after each convolutional layer. Set the convolutional kernel size in the first deep convolutional subnet to 7×1, and extract the fault detection global feature through the first deep convolutional subnet;

[0101] Step S322: Construct a second deep convolutional subnet. Specifically, set three convolutional layers in the second deep convolutional subnet, and set a batch normalization layer after each convolutional layer. Set the convolutional kernel size in the second deep convolutional subnet to 3×1, and extract the fault detection local feature through the second deep convolutional subnet;

[0102] Step S33: Construct a feature fusion attention block. Specifically, set a feature fusion layer, an attention mechanism layer, and a fully connected layer in the feature fusion attention block. Process the fault detection global feature and the fault detection local feature through the feature fusion attention block to obtain the fault detection fusion enhanced feature, including the following steps:

[0103] Step S331: Construct a feature fusion layer. Specifically, in the feature fusion layer, perform a dot product operation on the fault detection global feature and the fault detection local feature to obtain the fault detection fusion feature;

[0104] Step S332: Construct an attention mechanism layer. Specifically, in the attention mechanism layer, apply a channel attention mechanism to perform weighted calculation on the fault detection fusion feature to obtain the fault detection attention enhanced feature;

[0105] Step S333: Construct a fully connected layer. Specifically, perform feature dimensionality reduction on the fault detection attention enhanced feature through the fully connected layer to obtain the fault detection fusion enhanced feature;

[0106] Step S34: Construct a parallel graph convolutional subnet, including the following steps:

[0107] Step S341: Construct a graph structure. Specifically, take the fault detection fusion enhanced feature of each sample as a node feature, establish a node feature matrix, establish an adjacency matrix by measuring the similarity between each node feature, and perform a normalization operation to obtain a fault detection adjacency matrix;

[0108] The formula for establishing the adjacency matrix by measuring the similarity between each node feature, performing a normalization operation, and obtaining the fault detection adjacency matrix is:

[0109] ;

[0110] In the formula, A is the fault detection adjacency matrix, Nor(·) is the normalization operation, specifically referring to the symmetric normalization operation, M is the node feature matrix, the size of the node feature matrix is N×U, N is the number of nodes, the number of nodes is equal to the number of samples, U is the dimension of each node feature, T is the transpose operation, which is used to measure the similarity between each node feature;

[0111] Step S342: Construct parallel first-order, second-order, and third-order graph convolutional layers to process the fault detection adjacency matrix in parallel, extract multi-level information, splice the extraction results, and perform feature dimensionality reduction through a fully connected layer to obtain fault detection multi-level features. The calculation formula is:

[0112] ;

[0113] ;

[0114] In the formula, F k is the extraction result of the k-th order graph convolutional layer, LeakyReLU(·) is the leaky ReLU activation function, k is the aggregation order index, when k = 1, it represents aggregating first-order neighbor information, W k is the learning weight of the k-th order graph convolutional layer, F final is the fault detection multi-level feature, FC(·) is the fully connected layer function, and concat(·) is the feature splicing operation;

[0115] The first-order graph convolutional layer is used to aggregate first-order neighbor information;

[0116] The second-order graph convolutional layer is used to aggregate second-order neighbor information;

[0117] The third-order graph convolutional layer is used to aggregate third-order neighbor information;

[0118] Step S35: Construct a classification output layer. Specifically, a fully connected layer and a softmax activation function are set in the classification output layer to classify the multi-level features of fault detection and obtain the model output result;

[0119] Step S36: Model training. Specifically, by constructing the input layer, the parallel deep convolutional subnet, the feature fusion attention block, the parallel graph convolutional subnet, and the classification output layer, a parallel deep convolutional coupled graph convolutional network is constructed and model training is performed to obtain an elevator traction system fault detection model;

[0120] By performing the above operations, in the existing elevator traction system fault detection process, the faults of the elevator traction system are usually caused by the collaborative action of multiple components, so the global modeling ability of the detection model is required to be high. And the elevator traction system fault detection involves various sensor data and has multi-modal complexity. Traditional fault detection models are difficult to fully explore the complex relationships between multi-modal data and have insufficient multi-scale feature extraction ability, which in turn affects the detection accuracy. In response to this technical problem, this solution creatively adopts a parallel deep convolutional coupled graph convolutional network to construct an elevator traction system fault detection model, fully extract local and global features, effectively process multi-scale features, and capture the complex relationships between data through graph structure modeling, thereby improving the accuracy of elevator traction system fault detection and providing more accurate data support for intelligent elevator operation and maintenance.

[0121] Example 5, refer to Figure 1 , based on the above example, in step S4, the elevator traction system fault detection is specifically to perform elevator traction system fault detection through the elevator traction system fault detection model to obtain the state of the elevator traction system, and generate an elevator traction system fault detection report based on the state of the elevator traction system for assisting the management personnel to maintain the elevator traction system;

[0122] The elevator traction system fault detection report includes the detection date, elevator information, elevator traction system state, and maintenance suggestions.

[0123] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0124] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0125] The present invention and its embodiments have been described above. Such description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural forms and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. An elevator traction system fault detection method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: Collect elevator traction system data to obtain the elevator traction system status label and the elevator traction system actual data; Step S2: enhancing the elevator traction system data, using a recursive generative adversarial network model to perform data enhancement according to the elevator traction system state label and the actual data of the elevator traction system, to obtain the elevator traction system enhanced data; the recursive generative adversarial network model includes a generator and a discriminator; the generator loss function is constructed by combining the moment matching loss function, and the discriminator loss function is constructed by introducing a gradient penalty term; Step S3: construct a fault detection model, based on the elevator traction system state label and the elevator traction system enhanced data, using a parallel deep convolution coupled graph convolutional network to construct an elevator traction system fault detection model; The parallel deep convolution coupled graph convolution network includes an input layer, a parallel deep convolution subnet, a feature fusion attention block, a parallel graph convolution subnet and a classification output layer; The parallel deep convolutional subnetwork is used to capture multi-scale features in parallel; The feature fusion attention block is used to fuse multi-scale features and focus on key features; The parallel graph convolutional subnetwork is used to further capture the relationship between features; Step S4: Elevator traction system fault detection.

2. The elevator traction system fault detection method based on artificial intelligence according to claim 1 is characterized in that: In step S3, the construction of the fault detection model is used to construct the elevator traction system fault detection model, including the following steps: Step S31: constructing an input layer, specifically receiving the elevator traction system enhanced data through the input layer to construct a fault detection input sample; Step S32: construct a parallel deep convolutional subnetwork, specifically, setting a first deep convolutional subnetwork and a second deep convolutional subnetwork in the parallel deep convolutional subnetwork, wherein the first deep convolutional subnetwork and the second deep convolutional subnetwork are parallel structures, and extracting multi-scale features of the fault detection input samples through the parallel deep convolutional subnetwork to obtain global features and local features of fault detection; Step S33: construct a feature fusion attention block, specifically, setting a feature fusion layer, an attention mechanism layer and a fully connected layer in the feature fusion attention block, and processing the global features and local features of fault detection through the feature fusion attention block to obtain the fault detection fusion enhanced features; Step S34: Constructing a parallel graph convolutional subnetwork, including the following steps: Step S341: constructing a graph structure, specifically, taking the fault detection fusion enhancement feature of each sample as a node feature, establishing a node feature matrix, establishing an adjacency matrix by measuring the similarity between each node feature, and performing a normalization operation to obtain a fault detection adjacency matrix; Step S342: constructing a parallel first-order graph convolution layer, a second-order graph convolution layer, and a third-order graph convolution layer, processing the fault detection adjacency matrix in parallel, extracting multi-level information, performing feature concatenation on the extracted results, and performing feature dimension reduction through a fully connected layer to obtain multi-level features for fault detection; Step S35: constructing a classification output layer, specifically setting a fully connected layer and a softmax activation function in the classification output layer, classifying multi-level features of fault detection, and obtaining a model output result; Step S36: model training, specifically, constructing a parallel deep convolutional coupled graph convolutional network by constructing the input layer, constructing the parallel deep convolutional subnet, constructing the feature fusion attention block, constructing the parallel graph convolutional subnet and constructing the classification output layer, and performing model training to obtain the elevator traction system fault detection model.

3. The elevator traction system fault detection method based on artificial intelligence according to claim 2 is characterized in that: In step S32, the construction of a parallel deep convolutional subnetwork includes the following steps: Step S321: constructing a first deep convolutional subnet, specifically, setting three convolutional layers in the first deep convolutional subnet, setting a batch normalization layer after each convolutional layer, setting the convolution kernel size in the first deep convolutional subnet to 7×1, and extracting global features for fault detection through the first deep convolutional subnet; Step S322: construct a second deep convolutional subnetwork, specifically, set three convolutional layers in the second deep convolutional subnetwork, set a batch normalization layer after each convolutional layer, set the convolution kernel size in the second deep convolutional subnetwork to 3×1, and extract local features for fault detection through the second deep convolutional subnetwork.

4. The elevator traction system fault detection method based on artificial intelligence according to claim 3 is characterized in that: Step S33: construct a feature fusion attention block, specifically, setting a feature fusion layer, an attention mechanism layer and a fully connected layer in the feature fusion attention block, processing the global features and local features of fault detection through the feature fusion attention block, and obtaining the fault detection fusion enhanced features, including the following steps: Step S331: constructing a feature fusion layer, specifically, in the feature fusion layer, performing a dot product operation on the global features of fault detection and the local features of fault detection to obtain a fusion feature of fault detection; Step S332: constructing an attention mechanism layer, specifically, in the attention mechanism layer, applying a channel attention mechanism to perform weighted calculation on the fault detection fusion feature to obtain the fault detection attention enhancement feature; Step S333: construct a fully connected layer, specifically, perform feature dimension reduction on the fault detection attention enhancement feature through the fully connected layer to obtain the fault detection fusion enhancement feature.

5. The method for detecting elevator traction system faults based on artificial intelligence according to claim 4, characterized in that: In step S2, the elevator traction system data is enhanced to alleviate the data imbalance problem, specifically, based on the elevator traction system state label and the actual data of the elevator traction system, a recursive generative adversarial network model is used to perform data enhancement to obtain the elevator traction system enhanced data; The recursive generative adversarial network model includes a generator and a discriminator; The method of enhancing the elevator traction system data comprises the following steps: Step S21: constructing a generator for synthesizing data, specifically receiving minority class data in the actual data of the elevator traction system as the actual system data, introducing a recursive neural network based on long short-term memory units into the generator, learning the temporal dependency features in the data, and generating the synthesized data of the elevator traction system; Step S22: constructing a discriminator for distinguishing actual data from synthetic data, specifically receiving minority class data in the actual data of the elevator traction system as the actual data of the system, and receiving the synthetic data of the elevator traction system, introducing a recursive neural network based on long short-term memory units into the discriminator to judge the authenticity of the data, and obtaining a discrimination result; Step S23: improving the model loss function, specifically, constructing a generator loss function by combining the moment matching loss function, and constructing a discriminator loss function by introducing a gradient penalty term; The moment matching loss function is used to retain the statistical characteristics of the actual data, and the calculation formula is: ； Where, L mm is the moment matching loss function, t max is the maximum time step, used to represent the time series step of the data, is the square of the L2 norm, t is the time step index, is the mean of the actual data of the system at the tth time step, is the mean value of the synthetic data of the elevator traction system at the tth time step, is the standard deviation of the actual data of the system at time step t, is the standard deviation of the synthetic data of the elevator traction system at the tth time step, G(z) is the synthetic data of the elevator traction system, and z is a noise data point, specifically a data point randomly sampled from the noise distribution; The calculation formula of the generator loss function is: ； Where, L RGAN is the generator loss function, is the first weight of the generator, is the expectation of the synthetic data of the elevator traction system, which is used to measure the authenticity of the synthetic data of the elevator traction system. z is the noise distribution, D(·) is the discriminator scoring function, which is used to represent the authenticity of the synthetic data of the elevator traction system. is the second weight of the generator; The calculation formula of the discriminator loss function is: ； Where, L D is the discriminator loss function, is the expectation of the actual data of the system, p x is the actual data distribution of the system, x is the actual data point, specifically refers to the data point randomly sampled from the actual data distribution of the system, is the gradient penalty weight, is the expectation of the interpolated data, is an interpolation data distribution, which specifically refers to a data distribution obtained by linearly interpolating the actual system data and the synthetic data of the elevator traction system, are interpolated data points, is the gradient of the discriminator with respect to the interpolated data, is the L2 norm; Step S24: training the recursive generative adversarial network model, specifically, training the recursive generative adversarial network model by alternately training the generator and the discriminator to obtain the recursive generative adversarial network model; Step S25: data enhancement, specifically synthesizing the minority class data in the actual data of the elevator traction system through a recursive generative adversarial network model to generate elevator traction system synthetic data, and then merging the elevator traction system synthetic data and the elevator traction system actual data to obtain the elevator traction system enhanced data.

6. The method for detecting elevator traction system faults based on artificial intelligence according to claim 5, characterized in that: In step S1, the elevator traction system data is collected, specifically by installing sensors at the traction machine, motor bearings, and wire ropes of the elevator, collecting the elevator traction system data and labeling the data to obtain the elevator traction system status label, and then preprocessing the elevator traction system data to obtain the elevator traction system actual data.

7. The method for detecting elevator traction system faults based on artificial intelligence according to claim 6, characterized in that: In step S4, the elevator traction system fault detection is specifically to perform elevator traction system fault detection through an elevator traction system fault detection model to obtain the elevator traction system status, and generate an elevator traction system fault detection report based on the elevator traction system status to assist management personnel in maintaining the elevator traction system.

Images