Escalator operation and maintenance knowledge recommendation method fusing knowledge graph and deep learning

Abstract

An escalator operation and maintenance knowledge recommendation method fusing a knowledge graph and deep learning comprises the steps that firstly, an escalator fault knowledge graph is constructed, knowledge graph embedding is carried out, then data preprocessing is carried out on escalator vibration original data, 16 time domain indexes, 5 frequency domain indexes and 4 entropy indexes are extracted, and normalization operation is carried out; inputting the preprocessing result and the knowledge graph embedded data into an SKCNN layer for feature extraction; secondly, processing feature sequence data through an FA-BiGRU layer, and capturing internal rules and features of the sequence; and finally outputting the fault type of the escalator according to a feature processing result. Before the escalator equipment fails, possible potential faults and corresponding processing methods are analyzed. According to the invention, the intelligence and digitization degrees of escalator fault diagnosis are improved, and convenience is provided for the operation and maintenance of the escalator.

Description

Technical Field

[0001] The present invention belongs to the field of escalator operation and maintenance knowledge recommendation, and relates to an escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning. Background Art

[0002] As a complex mechanical system, the escalator needs to ensure its reliability during the design phase and also requires regular maintenance to ensure normal operation. An efficient operation and maintenance solution can reduce escalator operation failures, reduce operation and maintenance costs, and improve elevator safety. Since the concept of knowledge graph was born, many researchers have been committed to combining it with recommendation systems, and many excellent knowledge recommendation models have been produced. According to the tasks undertaken by the knowledge graph in the recommendation process, these models can be basically divided into three categories: ontology-based recommendation, representation learning-based recommendation, and hybrid recommendation that takes both into account. At present, knowledge graphs have played an important role in the Internet field through search engines, recommendation systems, etc., and have gradually emerged in the fields of law and medicine. However, in the domestic manufacturing industry, especially the escalator industry, the understanding and application of knowledge graphs are still relatively small. Studies have shown that the application research on knowledge graphs is concentrated in the fields of fault diagnosis and design process generation. At present, despite the existence of safety technical specifications and management methods related to escalators, the safety and reliability of escalators in the operation and maintenance stage cannot be guaranteed due to the following three reasons:

[0003] (1) The escalator is not running smoothly, and it is difficult to find hidden faults through periodic monitoring. Some faults are caused by equipment performance degradation, which requires long-term, dynamic monitoring to detect the problem.

[0004] (2) The maintenance personnel only dealt with the surface problems of the faults, but did not analyze the failure mechanism of the components. They did not solve the problem fundamentally, resulting in a vicious cycle of maintenance and damage.

[0005] (3) There is a lag in fault handling, and it is often not until a fault occurs that the escalator is discovered to be in a poor operating state. Summary of the invention

[0006] In order to overcome the shortcomings of the prior art, the present invention provides an escalator operation and maintenance knowledge recommendation method that integrates knowledge graph and deep learning. First, the escalator fault knowledge graph is constructed and embedded in the knowledge graph. Then, the escalator vibration raw data is preprocessed to extract 16 time domain indicators, 5 frequency domain indicators, and 4 entropy indicators and normalize them. Then, the preprocessing results and the knowledge graph embedded data are input into the SKCNN layer for feature extraction. Secondly, the feature sequence data is processed through the FA-BiGRU layer to capture the inherent laws and characteristics of the sequence. Finally, the feature processing results output the escalator fault type. Before the escalator equipment fails, the possible potential faults and corresponding processing methods are analyzed.

[0007] The technical solution adopted by the present invention to solve the technical problem is:

[0008] An escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning, the method comprising the following steps:

[0009] Step 1: The SPN model extracts escalator fault data. The process is as follows:

[0010] Step (1.1) Escalator fault data encoding

[0011] Using the Bert model as a sentence encoder, the input sentence is segmented with tokens through byte pair encoding and then input into the Bert model to obtain the context-aware representation of each fault feature in the escalator fault data platform;

[0012] Step (1.2) non-autoregressive decoding of escalator fault data;

[0013] Step 2: Build the knowledge graph of escalator failure. The process is as follows:

[0014] Step (2.1) Escalator Fault Mode Layer Construction

[0015] Step (2.2) embeds the knowledge graph through the TransD model to obtain the vectorized data of the escalator fault knowledge graph;

[0016] Step 3: Preprocessing of escalator vibration data. The process is as follows:

[0017] Step (3.1) Time domain index extraction: The time domain characteristics of the vibration data are obtained by statistically analyzing various time domain parameters and index calculations of the vibration data. The 16 time domain indexes of the escalator vibration data, including mean, variance, standard deviation, mean square value, maximum value, minimum value, peak-to-peak value, kurtosis, skewness, peak index, pulse index, margin index, kurtosis index, skewness index and waveform index, are extracted;

[0018] Step (3.2) Frequency domain index extraction:

[0019] By analyzing the frequency domain characteristics of escalator vibration data, the spectrum information of the data is accurately represented. Five frequency domain indicators, including center of gravity frequency, frequency variance, frequency standard deviation, mean square frequency, and root mean square frequency, are extracted from the escalator vibration data;

[0020] Step (3.3) Entropy index extraction: The entropy index evaluates the degree of disorder of the escalator vibration data, and extracts four entropy indexes, namely, approximate entropy, sample entropy, fuzzy entropy, and permutation entropy, from the escalator vibration data;

[0021] Step (3.4) Data normalization: After extracting the time domain, frequency domain and entropy indicators of the escalator operation data, the data is normalized and scaled between [0, 1]. After the normalization operation, the influence of the data dimension is not considered, which is convenient for processing between different indicators;

[0022] Step 4: Input the escalator operation fault data into the SKCNN feature extraction layer for feature extraction. The process is as follows:

[0023] Step (4.1) adopts the escalator fault extraction method: the time domain, frequency domain and entropy indicators of the escalator operation vibration data are used to replace the original time series data for feature extraction, so as to better express the characteristics of the escalator vibration data in the time domain, frequency domain and entropy indicators;

[0024] After data preprocessing, the feature extraction input matrix is:

[0025]

[0026] Where N^ is the sum of the number of time domain indicators, frequency domain indicators and entropy indicators; M^ is the number of samples in the data set; each sample size of the input is N^*1, t^ 1 Represents the first indicator after preprocessing of escalator vibration data samples. Similarly represents the N^th index after preprocessing of escalator vibration data samples;

[0027] Step (4.2) uses the convolution kernel automatic selection operation, that is, selective kernel convolution SK is performed between multiple convolution kernels of different kernel sizes;

[0028] Step 5: Input the features extracted by the SKCNN layer into the FA-BiGRU layer for feature learning.

[0029] Further, the process of the fifth step is as follows:

[0030] Step (5.1) uses the feature extraction result feature-Result as the input of the FA-BiGRU layer;

[0031] Step (5.2) The forward FA-GRU consists of two gating units, r t To control the reset gate, z t To control the update gating, the calculation formula is as follows:

[0032]

[0033] Among them, h t-1 is the state of the previous node output, is the input of the current node, σ^ is the Sigmoid function, W r is the weight matrix for resetting the gate;

[0034]

[0035] Among them, W z To update the gating weight matrix;

[0036] Step (5.3) obtains the gate signal z t After that, use the reset gate r t To get new data and current input x t$ Splicing, after tanh activation function scaling to between -1 and 1, the calculation formula is as follows:

[0037]

[0038] in, is the candidate state of the current time step, tanh is the hyperbolic tangent function;

[0039] Among them, · is the point multiplication operation, * is the matrix multiplication;

[0040] Step (5.4) updates the memory phase, performing both the forgetting and memorizing steps simultaneously, using the previously obtained update gate z t Equivalent to the forget gate, 1-z t Equivalent to the input gate, the updated expression is:

[0041]

[0042] in, is the forward hidden layer state at the current moment, is the forward hidden state of the previous moment;

[0043] Among them, the hidden layer state represents the state information at the current time, that is, the fault characteristic data of the escalator vibration;

[0044] Step (5.5) According to steps (5.1)-(5.4), input feature-Result into the backward FA-GRU in reverse order to obtain the backward hidden layer state

[0045] Among them, the structure and calculation process of the backward FA-GRU are the same as those of the forward FA-GRU;

[0046] The final hidden layer state of step (5.6) is the concatenation of the forward hidden layer state and the backward hidden layer state by position, forming the output:

[0047]

[0048] in, It means concatenating the two hidden states together to get the output of the current time step, h t Represents the overall hidden layer state;

[0049] Step (5.7) takes the final output of feature-Result and FA-BiGRU as the input of the feature attention mechanism and sends it to the multi-layer perceptron for calculation. The formula is as follows:

[0050]

[0051] in, is the vibration feature weight value at the current moment, m* is the feature dimension, V e , W e , U e is the weight parameter of the multilayer perceptron, b e is the bias parameter of the multilayer perceptron, V e T Indicates V e The transposed matrix of

[0052] Step (5.8) will find The final output of FA-BiGRU is obtained by normalizing with Softmax.

[0053] Furthermore, the process of step (1.2) is as follows:

[0054] Step (1.2.1) Predict escalator fault triple relationship

[0055] The non-autoregressive decoder is composed of N identical Transformer layers stacked together. In each Transformer layer, a multi-head self-attention mechanism is used to calculate the relationship between escalator fault triplets.

[0056] P(Yzj′|Xz,q)=p L (n′|Xz)∏p(Yz i′ |Xz,Y j′≠i′ ;θ)

[0057] Among them, Transfrmer is a commonly used deep network model for processing sequence data. L (n′|Xz) is the number of possible fault triplets, p(Y Zi′ |Xz,Y zj′≠i′ ; θ) represents the target escalator fault triplet Yz i′ Not only related to the given sentence Xz, but also to other triples Yz j′≠i′ For, i′, j′ are the i′th row and j′th column of the decoder respectively;

[0058] Step (1.2.2) decodes the predicted escalator failure triples into relations and entities

[0059] p r =softmax(W r h d )

[0060] Among them, p r is the relationship between the predicted faults, W r is the trainable parameter matrix, h d It is the output representation of the multi-layer Transformer, and softmax is the normalization function;

[0061] Step (1.2.3) distinguishes the parent and child entities of the predicted relationship

[0062]

[0063] Where T is the number of relationship types, H e is the output of the Bert model, tanh is the hyperbolic tangent function, p is the predicted probability, s and o are two different classifiers, star is the escalator fault parent node, end is the escalator fault child node, v 1 、v 2 、v 3 、v 4 are four trainable parameter matrices, W 1 , W 2 , W 3 , W 4 , W 5 , W 6 , W 7 , W 8 It is the weight parameter for predicting the head and tail nodes of the parent and child entities.

[0064] Furthermore, the process of step (2.1) is as follows:

[0065] Step (2.1.1) Escalator Fault Modeling

[0066] Escalator faults are modeled using escalator fault subsystem locations, escalator entity numbers, and escalator fault descriptions;

[0067] Among them, the escalator subsystem includes: steps, handrails, drive, traction, comb plate; escalator faults are described as specific fault types: broken pedal combs, broken bracket shaft holes, too large gap with skirt, step deflection, handrail jamming, drive escalator damage, abnormal noise from the drive, elongation of the traction chain, broken combs;

[0068] Step (2.1.2) escalator fault semantic modeling, supplementing the entity relationship and attribute relationship between escalator fault data to form escalator fault ontology;

[0069] Among them, the entity relationship is the association relationship between escalator fault data, the attribute relationship is the attribute of the escalator fault node, and the ontology is a semantic model used to describe the collection of relationships in the field. Entity and attribute are two elements in the ontology.

[0070] Preferably, the process of step (2.1.2) is as follows:

[0071] Step (2.1.2.1) starts from the description of the escalator fault and adds three attributes to the fault description, namely the escalator fault type, fault time and fault source;

[0072] Step (2.1.2.2) finds the corresponding solution for the fault description through the Solve to relationship

[0073] Step (2.1.2.3) locates the position of the specific faulty component in the escalator subsystem through the Locate to relationship;

[0074] Step (2.1.2.4) adds the subsystem number to the fault subsystem through the Belong to relationship and adds the fault number to the fault type through the ID to relationship;

[0075] Step (2.1.2.5) uses the Contain to relationship to store the fault number in the specific number of the escalator equipment, and extracts the maintenance history, operation data and location information of the escalator equipment in the database to which the equipment belongs through the specific number of the equipment.

[0076] Furthermore, in step (3.1), the mean reflects the average value of the escalator vibration data samples, which is expressed by the following formula:

[0077]

[0078] in, is the average value of the escalator vibration data samples, N is the number of escalator vibration samples, x i is the vibration data of the ith escalator, Indicates that N x i The operation of addition;

[0079] The variance reflects the average of the square value of the difference between a sample and the average of all sample values ​​in the escalator vibration data, representing the dynamic component of the escalator vibration energy, and is expressed by the following formula:

[0080]

[0081] in, is the variance of the escalator vibration data sample;

[0082] The standard deviation reflects the degree of dispersion of escalator vibration data and is expressed by the following formula:

[0083]

[0084] Among them, X r is the standard deviation, |x i |x i The absolute value of

[0085] The mean square value reflects the average value of the squares of all escalator vibration data and is expressed by the following formula:

[0086]

[0087] in, is the absolute value of the average value of escalator vibration data;

[0088] The RMS value is the arithmetic square root of the mean square value and is expressed as follows:

[0089]

[0090] Among them, X rms is the root mean square value;

[0091] The peak-to-peak value reflects the difference between the maximum and minimum values ​​of the escalator vibration data, which is expressed by the following formula:

[0092] X pp =max(x i )-min(x i )

[0093] Among them, X pp is the peak-to-peak value, max(x i ) represents the maximum value of the escalator vibration data, min(x i ) represents the minimum value of escalator vibration data;

[0094] The kurtosis reflects the distribution characteristics of escalator vibration data and is expressed by the following formula:

[0095]

[0096] Where beta is the kurtosis of the escalator vibration data;

[0097] The skewness reflects the skewness direction and degree of the escalator vibration data distribution, which is expressed by the following formula:

[0098]

[0099] Among them, α is the skewness;

[0100] The peak index reflects the quotient of the maximum absolute value of the escalator vibration data and the root mean square value, which is expressed by the following formula:

[0101]

[0102] Among them, C f is the peak index, X max is the maximum absolute value of the escalator vibration data;

[0103] The pulse index reflects the quotient of the maximum absolute value of the escalator vibration data and the absolute average value, which is expressed by the following formula:

[0104]

[0105] Among them, I f It is a pulse indicator;

[0106] The margin index reflects the proportion of impact signals in the escalator vibration data and is expressed by the following formula:

[0107]

[0108] Among them, CL f is the margin index, X r is the square root amplitude;

[0109] The kurtosis index reflects the impact characteristics of escalator vibration data and is expressed by the following formula:

[0110]

[0111] Among them, K v is the kurtosis index; Represents the fourth power of the root mean square value;

[0112] The skewness index reflects the symmetry of the escalator vibration data distribution and is expressed by the following formula:

[0113]

[0114] Among them, S k is the skewness indicator;

[0115] The waveform index reflects the quotient of the root mean square value and the absolute average value, which is expressed by the following formula:

[0116]

[0117] Among them, S f It is a waveform indicator.

[0118] In the step (3.2), the center of gravity frequency can describe the frequency of the signal component with a larger component in the spectrum of the escalator vibration data, and reflect the distribution of the power spectrum of the escalator vibration signal. The calculation formula is:

[0119]

[0120] Among them, FC is the center of gravity frequency, f is the amplitude of the escalator vibration frequency at the sampling point, and P(f) is the corresponding escalator vibration power spectrum density;

[0121] The frequency variance is the square of the frequency standard deviation, which is another metric to measure the energy dispersion of the escalator vibration power spectrum. The calculation formula is:

[0122]

[0123] Where VF is the frequency variance;

[0124] The frequency standard deviation is the inertia radius centered on the center of gravity frequency, which describes the dispersion of the energy distribution of the escalator vibration power spectrum. The calculation formula is:

[0125]

[0126] Among them, RVF is the frequency standard deviation;

[0127] The mean square frequency is the root mean square frequency squared and is calculated as:

[0128]

[0129] Wherein, MSF is mean square frequency;

[0130] The RMS frequency is the arithmetic square root of the mean square frequency and is calculated as:

[0131]

[0132] Here, RMSF is the root mean square frequency.

[0133] In the step (3.3), the approximate entropy is used to measure the probability of generating a new pattern in the escalator vibration signal, and the calculation formula is:

[0134]

[0135] in is the approximate entropy, m is the dimension of the escalator vibration data sample, r is the similarity tolerance, which is the degree of difference or mismatch that can be accepted when calculating the target item, and φ m+1 (r) is φ of dimension m+1 m (r),φ m (r) is the intermediate variable for calculating approximate entropy, and the calculation formula is:

[0136]

[0137] in, is the number of escalator vibration data within similar tolerance, N # is the dimension of the escalator vibration time series, m # is a commonly used parameter for calculating approximate entropy, used to represent the dimension of the reconstructed vector, i # Indicates the i # escalator vibration data;

[0138] Sample entropy is improved on the basis of approximate entropy. Both of them measure the complexity of the time series and the probability of the series generating new patterns when the dimension changes. The calculation formula is:

[0139] SampEn(m,r)=lnB m (r)-lnB m+1 (r)

[0140] Among them, SampEn(m,r) is the sample entropy, ln is the natural logarithm, B m (r) is the ratio of the number of samples within the similar tolerance to the peak-to-peak value of the escalator vibration samples, B m+1 (r) represents B with dimension m+1 m (r) value;

[0141] Fuzzy entropy is an improved sample entropy obtained by introducing fuzzy membership function on the basis of sample entropy. The calculation formula is:

[0142] FuzzyEn(m,n # ,r)=lnφ m (n # ,r)-lnφ m+1 (n # ,r)

[0143] Among them, FuzzyEn is the fuzzy entropy, n# is the number of escalator vibration data sample points, φ m It is an intermediate variable defined after the fuzzy membership function is introduced in the fuzzy entropy calculation process. The calculation formula is:

[0144]

[0145] in, is the similarity between the vibration data of the i-th and j-th escalator in the sample;

[0146] Permutation entropy is a method for detecting dynamic mutations and randomness of time series. It can quantitatively evaluate the random noise contained in the signal sequence. The calculation formula is:

[0147] Among them, H p is the permutation entropy, P j represents the number of permutations of the jth data, and k^ represents the dimension of the escalator vibration data.

[0148] Furthermore, the process of step (4.2) is as follows:

[0149] Step (4.2.1) Feature separation

[0150] For any given feature map, two transformations with convolutional sizes of 3×3 and 5×5 are first performed;

[0151]

[0152] in, and They are composed of deep convolution, batch normalization, and ReLu functions. is the convolution result of the two branches, H', W', C' are the sizes of the convolution kernels. Odd-sized convolution kernels are a common method in CNN, and their purpose is to align pixels and better extract image features. Therefore, 3×3 and 5×5 are the most commonly used convolution kernel sizes. In addition, the 5×5 convolution kernel can be replaced by a 3×3 convolution kernel and a diffusion convolution with an expansion size of 2.

[0153] Step (4.2.2) feature fusion process is as follows:

[0154] Step (4.2.2.1) first uses the element-wise summation method to effectively merge the output results of the two branches:

[0155]

[0156] Among them, U is the overall convolution information after fusing the convolution results of the two branches;

[0157] Step (4.2.2.2) then averages the features of each channel to embed the global information;

[0158] Step (4.2.2.3) creates a compact escalator fault feature through a fully connected layer;

[0159] Step (4.2.3) Feature selection

[0160] Add attention mechanism to information from different channels:

[0161]

[0162] Among them, A,B∈R C'×d' , C', d' represent the number of rows and columns of escalator operation fault features, a, b represent two attention weights, A c ∈R C×d represents the cth row of A, a c is the cth element b of a c is the cth element of b, and Z is a trainable parameter matrix;

[0163] Step (4.2.4) replaces a in step (4.2.3) c 、b c As the feature extraction result feature-Result.

[0164] The beneficial effects of the present invention are mainly manifested in: a unified representation of faults is performed by constructing an escalator fault knowledge graph, and the escalator signal fault characteristics are obtained through the adaptive convolution kernel convolution network SKCNN feature extraction layer; the temporal attention bidirectional recurrent neural network FA-BiGRU is used to more comprehensively capture the fault information in the vibration characteristics, thereby improving the intelligence and digitization of escalator fault diagnosis and providing convenience for the operation and maintenance of escalators. BRIEF DESCRIPTION OF THE DRAWINGS

[0165] Figure 1 This is the SPN model structure diagram.

[0166] Figure 2 It is the structure diagram of the escalator operation and maintenance knowledge graph.

[0167] Figure 3 This is the SKCNN model structure diagram. DETAILED DESCRIPTION

[0168] The present invention will be further described below in conjunction with the accompanying drawings.

[0169] Reference Figure 1 to Figure 3 , an escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning, comprising the following steps:

[0170] Step 1: SPN (Sequence Predict Network) model extracts escalator fault data. The process is as follows:

[0171] Step (1.1) Escalator fault data encoding

[0172] Using the Bert model as a sentence encoder, the input sentence is byte-pair encoded and tokenized to obtain Emb, which is then input into the Bert model to obtain the context-aware representation of each fault feature in the escalator fault data platform;

[0173] Among them, the Bert model is a commonly used neural network model for data encoding, and Emb is the embedding process of converting text into vectors;

[0174] Step (1.2) Non-autoregressive decoding of escalator fault data, the process is as follows:

[0175] Step (1.2.1) Predict escalator fault triple relationship

[0176] The non-autoregressive decoder is composed of N identical Transformer layers stacked together. In each Transformer layer, there is a multi-head self-attention mechanism and a feed forward neural network (FFN).

[0177] P(Yzj′|Xz,q)=p L (n′|Xz)∏p(Yz i′ |Xz,Y j′≠i′ ;θ)

[0178] Among them, the forward network is a classic neural network structure. There is no feedback signal in the network, and it only propagates in one direction. Transfrmer is a commonly used deep network model for processing sequence data. L (n′|Xz) is the number of possible fault triplets, p(Y Zi′ |Xz,Y zj′≠i′ ; θ) represents the target escalator fault triplet Yz i′ Not only related to the given sentence Xz, but also to other triples Yz j′≠i′ For, i′, j′ are the i′th row and j′th column of the decoder respectively;

[0179] Step (1.2.2) decodes the predicted escalator failure triples into relations and entities

[0180] p r =softmax(W r h d )

[0181] Among them, pr is the relationship between the predicted faults, W r is the trainable parameter matrix, h d It is the output representation of multi-layer Transformer, and softmax is a commonly used normalization function;

[0182] Step (1.2.3) distinguishes the parent and child entities of the predicted relationship

[0183]

[0184] Where T is the number of relationship types, H e is the output of the Bert model, tanh is the hyperbolic tangent function, p is the predicted probability, s and o are two different classifiers, star is the escalator fault parent node, end is the escalator fault child node, v 1 、v 2 、v 3 、v 4 are four trainable parameter matrices, W 1 , W 2 , W 3 , W 4 , W 5 , W 6 , W 7 , W 8 To predict the weight parameters of the head and tail nodes of the parent and child entities;

[0185] Step 2: Build the knowledge graph of escalator failure. The process is as follows:

[0186] Step (2.1) The escalator fault mode layer is constructed as follows:

[0187] Step (2.1.1) Escalator Fault Modeling

[0188] Escalator faults are modeled using escalator fault subsystem locations, escalator entity numbers, and escalator fault descriptions;

[0189] Among them, the escalator subsystem includes steps, handrails, drive, traction and comb plates; escalator faults are described as specific fault types: broken pedal combs, broken bracket shaft holes, too large gaps with skirts, step deflection, handrails stuck, escalator drive damage, abnormal noise from the drive, elongation of the traction chain and broken combs;

[0190] Step (2.1.2) Semantic modeling of escalator faults

[0191] Supplement the entity relationship and attribute relationship between escalator fault data to form the escalator fault ontology;

[0192] Among them, the entity relationship is the association relationship between escalator fault data, the attribute relationship is the attribute of the escalator fault node, and the ontology is a semantic model used to describe the collection of relationships in the field. Entity and attribute are two commonly used elements in ontology;

[0193] Step (2.1.2.1) starts from the description of the escalator fault and adds three attributes to the fault description, namely the escalator fault type, fault time and fault source;

[0194] Step (2.1.2.2) finds the corresponding processing method for the fault description through the Solve to relationship;

[0195] Step (2.1.2.3) locates the position of the specific faulty component in the escalator subsystem through the Locate to relationship;

[0196] Step (2.1.2.4) adds the subsystem number to the faulty subsystem through the Belong to relationship, and adds the fault number to the fault type through the ID to relationship;.

[0197] Step (2.1.2.5) uses the Contain to relationship to store the fault number into the specific number of the escalator equipment, and extracts the maintenance history, operation data and location information of the escalator equipment from the database to which the equipment belongs through the specific number of the equipment;

[0198] Step (2.2) embeds the knowledge graph through the TransD model to obtain the vectorized data of the escalator fault knowledge graph;

[0199] Among them, the TransD model is a commonly used knowledge graph embedding model;

[0200] Step 3: Preprocessing of escalator vibration data. The process is as follows:

[0201] Step (3.1) Time domain index extraction: The time domain characteristics of the vibration data are obtained by statistically analyzing various time domain parameters and index calculations of the vibration data. The 16 time domain indexes of the escalator vibration data, including mean, variance, standard deviation, mean square value, maximum value, minimum value, peak-to-peak value, kurtosis, skewness, peak index, pulse index, margin index, kurtosis index, skewness index and waveform index, are extracted;

[0202] The mean reflects the average value of the escalator vibration data samples and is expressed as follows:

[0203]

[0204] in, is the average value of the escalator vibration data samples, N is the number of escalator vibration samples, xi is the vibration data of the ith escalator, Indicates that N x i The operation of addition;

[0205] The variance reflects the average of the square value of the difference between a sample and the average of all sample values ​​in the escalator vibration data, representing the dynamic component of the escalator vibration energy, and is expressed by the following formula:

[0206]

[0207] in, is the variance of the escalator vibration data sample;

[0208] The standard deviation reflects the degree of dispersion of escalator vibration data and is expressed by the following formula:

[0209]

[0210] Among them, X r is the standard deviation, |x i |x i The absolute value of

[0211] The mean square value reflects the average value of the squares of all escalator vibration data and is expressed by the following formula:

[0212]

[0213] Where, |X| is the absolute value of the average value of escalator vibration data;

[0214] The RMS value is the arithmetic square root of the mean square value and is expressed as follows:

[0215]

[0216] Among them, X rms is the root mean square value;

[0217] The peak-to-peak value reflects the difference between the maximum and minimum values ​​of the escalator vibration data, which is expressed by the following formula:

[0218] X pp =max(x i )-min(x i )

[0219] Among them, X pp is the peak-to-peak value, max(x i ) represents the maximum value of the escalator vibration data, min(x i ) represents the minimum value of escalator vibration data;

[0220] The kurtosis reflects the distribution characteristics of escalator vibration data and is expressed by the following formula:

[0221]

[0222] Where beta is the kurtosis of the escalator vibration data;

[0223] The skewness reflects the skewness direction and degree of the escalator vibration data distribution, which is expressed by the following formula:

[0224]

[0225] Among them, α is the skewness;

[0226] The peak index reflects the quotient of the maximum absolute value of the escalator vibration data and the root mean square value, which is expressed by the following formula:

[0227]

[0228] Among them, C f is the peak index, X max is the maximum absolute value of the escalator vibration data;

[0229] The pulse index reflects the quotient of the maximum absolute value of the escalator vibration data and the absolute average value, which is expressed by the following formula:

[0230]

[0231] Among them, I f It is a pulse indicator;

[0232] The margin index reflects the proportion of impact signals in the escalator vibration data and is expressed by the following formula:

[0233]

[0234] Among them, CL f is the margin index, X r is the square root amplitude;

[0235] The kurtosis index reflects the impact characteristics of escalator vibration data and is expressed by the following formula:

[0236]

[0237] Among them, K v is the kurtosis index; Represents the fourth power of the root mean square value;

[0238] The skewness index reflects the symmetry of the escalator vibration data distribution and is expressed by the following formula:

[0239]

[0240] Among them, S k is the skewness indicator;

[0241] The waveform index reflects the quotient of the root mean square value and the absolute average value, which is expressed by the following formula:

[0242]

[0243] Among them, S f It is a waveform indicator;

[0244] Step (3.2) Frequency domain index extraction:

[0245] By analyzing the frequency domain characteristics of escalator vibration data, the spectrum information of the data is accurately represented. Five frequency domain indicators, including center of gravity frequency, frequency variance, frequency standard deviation, mean square frequency, and root mean square frequency, are extracted from the escalator vibration data;

[0246] The center of gravity frequency can describe the frequency of the signal component with a larger component in the spectrum of the escalator vibration data, and reflect the distribution of the power spectrum of the escalator vibration signal. The calculation formula is:

[0247]

[0248] Among them, FC is the center of gravity frequency, f is the amplitude of the escalator vibration frequency at the sampling point, and P(f) is the corresponding escalator vibration power spectrum density;

[0249] The frequency variance is the square of the frequency standard deviation, which is another metric to measure the energy dispersion of the escalator vibration power spectrum. The calculation formula is:

[0250]

[0251] Where VF is the frequency variance;

[0252] The frequency standard deviation is the inertia radius centered on the center of gravity frequency, which describes the dispersion of the energy distribution of the escalator vibration power spectrum. The calculation formula is:

[0253]

[0254] Among them, RVF is the frequency standard deviation;

[0255] The mean square frequency is the root mean square frequency squared and is calculated as:

[0256]

[0257] Wherein, MSF is mean square frequency;

[0258] The RMS frequency is the arithmetic square root of the mean square frequency and is calculated as:

[0259]

[0260] Where RMSF is the root mean square frequency;

[0261] Step (3.3) Entropy index extraction: The entropy index evaluates the degree of disorder of the escalator vibration data, and extracts four entropy indexes, namely, approximate entropy, sample entropy, fuzzy entropy, and permutation entropy, from the escalator vibration data;

[0262] Approximate entropy is used to measure the probability of generating new patterns in escalator vibration signals. The more complex the time series, the larger the corresponding approximate entropy, and the more regular the time series, the smaller the corresponding approximate entropy. The calculation formula is:

[0263]

[0264] in is the approximate entropy, m is the dimension of the escalator vibration data sample, r is the similarity tolerance, which is the degree of difference or mismatch that can be accepted when calculating the target item, and φ m+1 (r) is φ of dimension m+1 m (r),φ m (r) is the intermediate variable for calculating approximate entropy, and the calculation formula is:

[0265]

[0266] in, is the number of escalator vibration data within similar tolerance, N # is the dimension of the escalator vibration time series, m # is a commonly used parameter for calculating approximate entropy, used to represent the dimension of the reconstructed vector, i # Indicates the i # escalator vibration data;

[0267] Sample entropy is improved on the basis of approximate entropy. Both of them measure the complexity of time series and the probability of a new pattern in the series when the dimension changes. The greater the probability of generating a new pattern, the higher the complexity of the sequence, and the greater the entropy value. The calculation formula is:

[0268] SampEn(m,r)=lnB m (r)-lnB m+1 (r)

[0269] Among them, SampEn(m,r) is the sample entropy, ln is the natural logarithm, B m (r) is the ratio of the number of samples within the similar tolerance to the peak-to-peak value of the escalator vibration samples, B m+1 (r) represents B with dimension m+1m (r) value;

[0270] Fuzzy entropy is an improved sample entropy obtained by introducing an exponential function, fuzzy membership function, on the basis of sample entropy. The calculation formula is:

[0271] FuzzyEn(m,n # ,r)=lnφ m (n # ,r)-lnφ m+1 (n # ,r)

[0272] Among them, FuzzyEn is the fuzzy entropy, n # is the number of escalator vibration data sample points, φ m It is an intermediate variable defined after the fuzzy membership function is introduced in the fuzzy entropy calculation process. The calculation formula is:

[0273]

[0274] in, is the similarity between the vibration data of the i-th and j-th escalator in the sample;

[0275] Permutation entropy is a method for detecting dynamic mutations and randomness of time series. It can quantitatively evaluate the random noise contained in the signal sequence. The calculation formula is:

[0276] Among them, H p is the permutation entropy, P j represents the number of permutations of the jth data, and k^ represents the dimension of the escalator vibration data;

[0277] Step (3.4) Data normalization: After extracting the time domain, frequency domain and entropy indicators of the escalator operation data, the data is normalized and scaled between [0,1]. After the normalization operation, the impact of the data dimension is not considered, which is convenient for processing between different indicators. For neural networks, the advantage of normalization is to speed up the convergence speed and suppress gradient explosion. The normalization formula is as follows:

[0278]

[0279] Among them, x * is the normalized data, max is the maximum value in the escalator vibration data sample, min is the minimum value in the escalator vibration data sample, x i This is the escalator vibration data before normalization;

[0280] Step 4: Input the escalator operation fault data into the SKCNN feature extraction layer for feature extraction. The process is as follows:

[0281] Step (4.1) uses the time domain, frequency domain and entropy indicators of the escalator operation vibration data to replace the original time series data for feature extraction, so as to better express the characteristics of the escalator vibration data in the time domain, frequency domain and entropy indicators;

[0282] After data preprocessing, the feature extraction input matrix is:

[0283]

[0284] Where N ^ is 25, which is the sum of the number of time domain indicators, frequency domain indicators and entropy indicators; M^ is the number of samples in the data set; each sample size of the input is 25*1, t^ 1 Represents the first indicator after preprocessing of escalator vibration data samples. Similarly Represents the Nth escalator vibration data sample after preprocessing ^ indicators;

[0285] Step (4.2) uses the automatic convolution kernel selection operation, that is, selective kernel convolution (Selective Kernel, SK) is performed between multiple convolution kernels of different kernel sizes;

[0286] Step (4.2.1) Feature separation

[0287] For any given feature map, two transformations with convolutional sizes of 3×3 and 5×5 are first performed;

[0288]

[0289] in, and They are composed of deep convolution, batch normalization, and ReLu functions. is the convolution result of two branches, H', W', C' are the sizes of convolution kernels. Odd-sized convolution kernels are a common method in CNN, and their purpose is to align pixels and better extract image features. Therefore, 3×3 and 5×5 are the most commonly used convolution kernel sizes. In addition, the 5×5 convolution kernel can be replaced by a 3×3 convolution kernel and a diffusion convolution with an expansion size of 2. This process ensures that the receptive field of the model remains unchanged while reducing the amount of calculation;

[0290] Step (4.2.2) Feature Fusion

[0291] Step (4.2.2.1) first uses the element-wise summation method to effectively merge the output results of the two branches:

[0292]

[0293] Among them, U is the overall convolution information after fusing the convolution results of the two branches;

[0294] Step (4.2.2.2) then averages the features of each channel to embed the global information;

[0295] Step (4.2.2.3) creates a compact escalator fault feature through a fully connected layer;

[0296] Among them, the fully connected layer is a basic structure in deep learning neural networks. Its function is to connect each input node with each node in the next layer. Each connection has a corresponding weight parameter, so it can learn the complex mapping relationship between input and output;

[0297] Step (4.2.3) Feature selection

[0298] Add attention mechanism to information from different channels:

[0299]

[0300] Among them, the attention mechanism is a feature extraction processing method to better extract the fault features A, B∈R of the escalator operation C'×d' , C', d' represent the number of rows and columns of escalator operation fault features, a, b represent two attention weights, A c ∈R C×d represents the cth row of A, a c is the cth element b of a c is the cth element of b, and Z is a trainable parameter matrix;

[0301] Step (4.2.4) replaces a in step (4.2.3) c 、b c As feature extraction result feature-Result;

[0302] Step 5: Input the features extracted by the SKCNN layer into the FA-BiGRU layer for feature learning. The process is as follows:

[0303] The FA-BiGRU adopted in the present invention adds a temporal attention mechanism and a multi-layer perceptron to the traditional BiGRU to enhance the learning ability of different fault characteristics.

[0304] Step (5.1) uses the feature extraction result feature-Result as the input of the FA-BiGRU layer;

[0305] Step (5.2) The forward FA-GRU consists of two gating units, rt To control the reset gate, z t To control the update gating, the calculation formula is as follows:

[0306]

[0307] Among them, h t-1 is the state of the previous node output, is the input of the current node, σ^ is the Sigmoid function, W r is the weight matrix for resetting the gate;

[0308]

[0309] Among them, W z To update the gating weight matrix;

[0310] Step (5.3) obtains the gate signal z t After that, use the reset gate r t To get new data and current input x t$ Splicing, after tanh activation function scaling to between -1 and 1, the calculation formula is as follows:

[0311]

[0312] in, is the candidate state of the current time step, tanh is the hyperbolic tangent function;

[0313] Among them, · is the point multiplication operation, * is the matrix multiplication;

[0314] Step (5.4) updates the memory phase, performing both the forgetting and memorizing steps simultaneously, using the previously obtained update gate z t Equivalent to the forget gate, 1-z t Equivalent to the input gate, the updated expression is:

[0315]

[0316] in, is the forward hidden layer state at the current moment, is the forward hidden state of the previous moment;

[0317] Among them, the hidden layer state represents the state information at the current time, that is, the fault characteristic data of the escalator vibration;

[0318] Step (5.5) According to steps (5.1)-(5.4), input feature-Result into the backward FA-GRU in reverse order to obtain the backward hidden layer state

[0319] Among them, the structure and calculation process of the backward FA-GRU are the same as those of the forward FA-GRU;

[0320] The final hidden layer state of step (5.6) is the concatenation of the forward hidden layer state and the backward hidden layer state by position, forming the output:

[0321]

[0322] in, It means concatenating the two hidden states together to get the output of the current time step, h t Represents the overall hidden layer state;

[0323] Step (5.7) takes the final output of feature-Result and FA-BiGRU as the input of the feature attention mechanism and sends it to the multi-layer perceptron for calculation. The formula is as follows:

[0324]

[0325] in, is the vibration feature weight value at the current moment, m* is the feature dimension, V e , W e , U e is the weight parameter of the multilayer perceptron, b e is the bias parameter of the multilayer perceptron, V e T Indicates V e The transposed matrix of

[0326] Step (5.8) will find The final output of FA-BiGRU is obtained by normalizing with Softmax.

[0327] in is the final hidden layer state of FA-BiGRU, which contains the escalator failure information; Softmax is a common normalization function used to map the outputs of multiple neurons to the (0,1) interval to represent their probabilities.

[0328] The method of this embodiment uniformly represents the faults by constructing an escalator fault knowledge graph, and obtains the escalator signal fault characteristics through the adaptive convolution kernel convolution network SKCNN feature extraction layer; the temporal attention bidirectional recurrent neural network FA-BiGRU is used to more comprehensively capture the fault information in the vibration characteristics, thereby improving the intelligence and digitization of escalator fault diagnosis and providing convenience for the operation and maintenance of escalators.

[0329] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A method for recommending escalator operation and maintenance knowledge by integrating knowledge graph and deep learning, characterized in that: The method comprises the following steps: Step 1: The SPN model extracts escalator fault data. The process is as follows: Step (1.1) Escalator fault data encoding Using the Bert model as a sentence encoder, the input sentence is segmented with tokens through byte pair encoding and then input into the Bert model to obtain the context-aware representation of each fault feature in the escalator fault data platform; Step (1.2) non-autoregressive decoding of escalator fault data; Step 2: Build the knowledge graph of escalator failure. The process is as follows: Step (2.1) Escalator Fault Mode Layer Construction Step (2.2) embeds the knowledge graph through the TransD model to obtain the vectorized data of the escalator fault knowledge graph; Step 3: Preprocessing of escalator vibration data. The process is as follows: Step (3.1) Time domain index extraction: The time domain characteristics of the vibration data are obtained by statistically analyzing various time domain parameters and index calculations of the vibration data. The 16 time domain indexes of the escalator vibration data, including mean, variance, standard deviation, mean square value, maximum value, minimum value, peak-to-peak value, kurtosis, skewness, peak index, pulse index, margin index, kurtosis index, skewness index and waveform index, are extracted; Step (3.2) Frequency domain index extraction: By analyzing the frequency domain characteristics of escalator vibration data, the spectrum information of the data is accurately represented, and five frequency domain indicators, including center of gravity frequency, frequency variance, frequency standard deviation, mean square frequency, and root mean square frequency, are extracted from the escalator vibration data. Step (3.3) Entropy index extraction: The entropy index evaluates the degree of disorder of the escalator vibration data, and extracts four entropy indexes, namely, approximate entropy, sample entropy, fuzzy entropy, and permutation entropy, from the escalator vibration data; Step (3.4) Data normalization: After extracting the time domain, frequency domain and entropy indicators of the escalator operation data, the data is normalized and scaled between [0, 1]. After the normalization operation, the influence of the data dimension is not considered, which is convenient for processing between different indicators; Step 4: Input the escalator operation fault data into the SKCNN feature extraction layer for feature extraction. The process is as follows: Step (4.1) adopts the escalator fault extraction method: the time domain, frequency domain and entropy indicators of the escalator operation vibration data are used to replace the original time series data for feature extraction, so as to better express the characteristics of the escalator vibration data in the time domain, frequency domain and entropy indicators; After data preprocessing, the feature extraction input matrix is: Where N^ is the sum of the number of time domain indicators, frequency domain indicators and entropy indicators; M^ is the number of samples in the data set; each sample size of the input is N^*1, t^ 1 represents the first index after the escalator vibration data sample is preprocessed. Similarly, t^ N^ represents the N^th index after preprocessing of escalator vibration data samples; Step (4.2) uses the convolution kernel automatic selection operation, that is, selective kernel convolution SK is performed between multiple convolution kernels of different kernel sizes; Step 5: Input the features extracted by the SKCNN layer into the FA-BiGRU layer for feature learning.

2. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 is characterized in that: The process of the fifth step is as follows: Step (5.1) uses the feature extraction result feature-Result as the input of the FA-BiGRU layer; Step (5.2) The forward FA-GRU consists of two gating units, r t To control the reset gate, z t To control the update gating, the calculation formula is as follows: Among them, h t-1 is the state of the previous node output, is the input of the current node, σ^ is the Sigmoid function, W r is the weight matrix for resetting the gate; Among them, W z To update the gating weight matrix; Step (5.3) obtains the gate signal z t After that, use the reset gate r t To get new data and current input Splicing, after tanh activation function scaling to between -1 and 1, the calculation formula is as follows: in, is the candidate state of the current time step, tanh is the hyperbolic tangent function; Among them, · is the point multiplication operation, * is the matrix multiplication; Step (5.4) updates the memory phase, performing both the forgetting and memorizing steps simultaneously, using the previously obtained update gate z t Equivalent to the forget gate, 1-z t Equivalent to the input gate, the updated expression is: in, is the forward hidden layer state at the current moment, is the forward hidden state of the previous moment; Among them, the hidden layer state represents the state information at the current time, that is, the fault characteristic data of the escalator vibration; Step (5.5) According to steps (5.1)-(5.4), input feature-Result into the backward FA-GRU in reverse order to obtain the backward hidden layer state Among them, the structure and calculation process of the backward FA-GRU are the same as those of the forward FA-GRU; The final hidden layer state of step (5.6) is the concatenation of the forward hidden layer state and the backward hidden layer state by position, forming the output: in, It means concatenating the two hidden states together to get the output of the current time step, h t Represents the overall hidden layer state; Step (5.7) takes the final output of feature-Result and FA-BiGRU as the input of the feature attention mechanism and sends it to the multi-layer perceptron for calculation. The formula is as follows: in, is the vibration feature weight value at the current moment, m* is the feature dimension, V e , W e , U e is the weight parameter of the multilayer perceptron, b e is the bias parameter of the multilayer perceptron, Indicates V e The transposed matrix of Step (5.8) will find The final output of FA-BiGRU is obtained by normalizing with Softmax.

3. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: The process of step (1.2) is as follows: Step (1.2.1) Predict escalator fault triple relationship The non-autoregressive decoder is composed of N identical Transformer layers stacked together. In each Transformer layer, a multi-head self-attention mechanism is used to calculate the relationship between escalator fault triplets. P(Yzj′|Xz,q)=p L (n′|Xz)∏p(Yz i′ |Xz,Y j′≠i′ ;θ) Among them, Transfrmer is a commonly used deep network model for processing sequence data. L (n′|Xz) is the number of possible fault triplets, p(Y Zi′ |Xz,Y zj′≠i′ ; θ) represents the target escalator fault triplet Yz i′ Not only related to the given sentence Xz, but also to other triples Yz j′≠i′ With respect to, i′, j′ are the i′th row and j′th column of the decoder respectively; Step (1.2.2) decodes the predicted escalator failure triples into relations and entities p r =softmax(W r h d ) Among them, p r is the relationship between the predicted faults, W r is the trainable parameter matrix, h d It is the output representation of the multi-layer Transformer, and softmax is the normalization function; Step (1.2.3) distinguishes the parent and child entities of the predicted relationship Where T is the number of relationship types, H e is the output of the Bert model, tanh is the hyperbolic tangent function, p is the predicted probability, s and o are two different classifiers, star is the escalator fault parent node, end is the escalator fault child node, v1, v2, v3, v4 are four trainable parameter matrices, W1, W2, W3, W4, W5, W6, W7, W8 are the weight parameters for predicting the head and tail nodes of the parent and child entities.

4. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: The process of step (2.1) is as follows: Step (2.1.1) Escalator Fault Modeling Escalator faults are modeled using escalator fault subsystem locations, escalator entity numbers, and escalator fault descriptions; Among them, the escalator subsystem includes: steps, handrails, drive, traction, comb plate; escalator faults are described as specific fault types: broken pedal combs, broken bracket shaft holes, too large gap with skirt, step deflection, handrail jamming, drive escalator damage, abnormal noise from the drive, elongation of the traction chain, broken combs; Step (2.1.2) escalator fault semantic modeling, supplementing the entity relationship and attribute relationship between escalator fault data to form escalator fault ontology; Among them, the entity relationship is the association relationship between escalator fault data, the attribute relationship is the attribute of the escalator fault node, and the ontology is a semantic model used to describe the collection of relationships in the field. Entity and attribute are two elements in the ontology; Preferably, the process of step (2.1.2) is as follows: Step (2.1.2.1) starts from the description of the escalator fault and adds three attributes to the fault description, namely the escalator fault type, fault time and fault source; Step (2.1.2.2) finds the corresponding solution for the fault description through the Solve to relationship Step (2.1.2.3) locates the position of the specific faulty component in the escalator subsystem through the Locate to relationship; Step (2.1.2.4) adds the subsystem number to the fault subsystem through the Belong to relationship and adds the fault number to the fault type through the ID to relationship; Step (2.1.2.5) uses the Contain to relationship to store the fault number in the specific number of the escalator equipment, and extracts the maintenance history, operation data and location information of the escalator equipment in the database to which the equipment belongs through the specific number of the equipment.

5. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: In the step (3.1), the mean reflects the average value of the escalator vibration data samples, which is expressed by the following formula: in, is the average value of the escalator vibration data samples, N is the number of escalator vibration samples, x i is the vibration data of the ith escalator, Indicates that N x i The operation of addition; The variance reflects the average of the square value of the difference between a sample and the average of all sample values ​​in the escalator vibration data, representing the dynamic component of the escalator vibration energy, and is expressed by the following formula: in, is the variance of the escalator vibration data sample; The standard deviation reflects the degree of dispersion of escalator vibration data and is expressed by the following formula: Among them, X r is the standard deviation, |x i |x i The absolute value of The mean square value reflects the average value of the squares of all escalator vibration data and is expressed by the following formula: in, is the absolute value of the average value of escalator vibration data; The RMS value is the arithmetic square root of the mean square value and is expressed as follows: Among them, X rms is the root mean square value; The peak-to-peak value reflects the difference between the maximum and minimum values ​​of the escalator vibration data, which is expressed by the following formula: X pp =max(x i )-min(x i ) Among them, X pp is the peak-to-peak value, max(x i ) represents the maximum value of the escalator vibration data, min(x i ) represents the minimum value of escalator vibration data; The kurtosis reflects the distribution characteristics of escalator vibration data and is expressed by the following formula: Where beta is the kurtosis of the escalator vibration data; The skewness reflects the skewness direction and degree of the escalator vibration data distribution, which is expressed by the following formula: Among them, α is the skewness; The peak index reflects the quotient of the maximum absolute value of the escalator vibration data and the root mean square value, which is expressed by the following formula: Among them, C f is the peak index, X max is the maximum absolute value of the escalator vibration data; The pulse index reflects the quotient of the maximum absolute value of the escalator vibration data and the absolute average value, which is expressed by the following formula: Among them, I f It is a pulse indicator; The margin index reflects the proportion of impact signals in the escalator vibration data and is expressed by the following formula: Among them, CL f is the margin index, X r is the square root amplitude; The kurtosis index reflects the impact characteristics of escalator vibration data and is expressed by the following formula: Among them, K v is the kurtosis index; Represents the fourth power of the root mean square value; The skewness index reflects the symmetry of the escalator vibration data distribution and is expressed by the following formula: Among them, S k is the skewness indicator; The waveform index reflects the quotient of the root mean square value and the absolute average value, which is expressed by the following formula: Among them, S f It is a waveform indicator.

6. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: In the step (3.2), the center of gravity frequency can describe the frequency of the signal component with a larger component in the spectrum of the escalator vibration data, and reflect the distribution of the power spectrum of the escalator vibration signal. The calculation formula is: Among them, FC is the center of gravity frequency, f is the amplitude of the escalator vibration frequency at the sampling point, and P(f) is the corresponding escalator vibration power spectrum density; The frequency variance is the square of the frequency standard deviation, which is another metric to measure the energy dispersion of the escalator vibration power spectrum. The calculation formula is: Where VF is the frequency variance; The frequency standard deviation is the inertia radius centered on the center of gravity frequency, which describes the dispersion of the energy distribution of the escalator vibration power spectrum. The calculation formula is: Among them, RVF is the frequency standard deviation; The mean square frequency is the root mean square frequency squared and is calculated as: Wherein, MSF is mean square frequency; The RMS frequency is the arithmetic square root of the mean square frequency and is calculated as: Here, RMSF is the root mean square frequency.

7. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: In the step (3.3), the approximate entropy is used to measure the probability of generating a new pattern in the escalator vibration signal, and the calculation formula is: in is the approximate entropy, m is the dimension of the escalator vibration data sample, r is the similarity tolerance, which is the degree of difference or mismatch that can be accepted when calculating the target item, and φ m+1 (r) is φ of dimension m+1 m (r),φ m (r) is the intermediate variable for calculating approximate entropy, and the calculation formula is: in, is the number of escalator vibration data within similar tolerance, N # is the dimension of the escalator vibration time series, m # is a commonly used parameter for calculating approximate entropy, used to represent the dimension of the reconstructed vector, i # Indicates the i # escalator vibration data; Sample entropy is improved on the basis of approximate entropy. Both of them measure the complexity of the time series and the probability of the series generating new patterns when the dimension changes. The calculation formula is: SampEn(m,r)lnB m (r)-lnB m+1 (r) Among them, SampEn(m,r) is the sample entropy, ln is the natural logarithm, B m (r) is the ratio of the number of samples within the similar tolerance to the peak-to-peak value of the escalator vibration samples, B m+1 (r) represents B with dimension m+1 m (r) value; Fuzzy entropy is an improved sample entropy obtained by introducing fuzzy membership function on the basis of sample entropy. The calculation formula is: FuzzyEn(m,n # ,r)=lnφ m (n # ,r)-lnφ m+1 (n # ,r) Among them, FuzzyEn is the fuzzy entropy, n # is the number of escalator vibration data sample points, φ m It is an intermediate variable defined after the fuzzy membership function is introduced in the fuzzy entropy calculation process. The calculation formula is: in, is the similarity between the vibration data of the i-th and j-th escalator in the sample; Permutation entropy is a method for detecting dynamic mutations and randomness of time series. It can quantitatively evaluate the random noise contained in the signal sequence. The calculation formula is: Among them, H p is the permutation entropy, P j represents the number of permutations of the jth data, and k^ represents the dimension of the escalator vibration data.

8. The escalator operation and maintenance knowledge recommendation method integrating knowledge graph and deep learning as claimed in claim 1 or 2, characterized in that: The process of step (4.2) is as follows: Step (4.2.1) Feature separation For any given feature map, two transformations with convolutional sizes of 3×3 and 5×5 are first performed; in, and They are composed of deep convolution, batch normalization, and ReLu functions. is the convolution result of the two branches, H', W', C' are the sizes of the convolution kernels. Odd-sized convolution kernels are a common method in CNN, and their purpose is to align pixels and better extract image features. Therefore, 3×3 and 5×5 are the most commonly used convolution kernel sizes. In addition, the 5×5 convolution kernel can be replaced by a 3×3 convolution kernel and a diffusion convolution with an expansion size of 2. Step (4.2.2) feature fusion process is as follows: Step (4.2.2.1) first uses the element-wise summation method to effectively merge the output results of the two branches: Among them, U is the overall convolution information after fusing the convolution results of the two branches; Step (4.2.2.2) then averages the features of each channel to embed the global information; Step (4.2.2.3) creates a compact escalator fault feature through a fully connected layer; Step (4.2.3) Feature selection Add attention mechanism to information from different channels: Among them, A,B∈R C'×d' , C', d' represent the number of rows and columns of escalator operation fault features, a, b represent two attention weights, A c ∈R C×d represents the cth row of A, a c is the cth element b of a c is the cth element of b, and Z is a trainable parameter matrix; Step (4.2.4) replaces a in step (4.2.3) c 、b c As the feature extraction result feature-Result.

Images