CN116578858A - Air compressor fault prediction and health degree evaluation method and system based on graphic neural network - Google Patents
Air compressor fault prediction and health degree evaluation method and system based on graphic neural network Download PDFInfo
- Publication number
- CN116578858A CN116578858A CN202310108821.6A CN202310108821A CN116578858A CN 116578858 A CN116578858 A CN 116578858A CN 202310108821 A CN202310108821 A CN 202310108821A CN 116578858 A CN116578858 A CN 116578858A
- Authority
- CN
- China
- Prior art keywords
- air compressor
- health
- time sequence
- scale
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000036541 health Effects 0.000 title claims abstract description 111
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 51
- 238000011156 evaluation Methods 0.000 title claims description 28
- 238000000034 method Methods 0.000 claims abstract description 44
- 238000010586 diagram Methods 0.000 claims abstract description 39
- 238000012512 characterization method Methods 0.000 claims abstract description 29
- 238000012545 processing Methods 0.000 claims abstract description 19
- 230000002776 aggregation Effects 0.000 claims abstract description 18
- 238000004220 aggregation Methods 0.000 claims abstract description 18
- 230000007246 mechanism Effects 0.000 claims abstract description 12
- 239000011159 matrix material Substances 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 11
- 230000009466 transformation Effects 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 10
- 239000011664 nicotinic acid Substances 0.000 claims description 4
- 230000002123 temporal effect Effects 0.000 claims description 4
- 230000008569 process Effects 0.000 abstract description 9
- 238000001514 detection method Methods 0.000 abstract description 7
- 230000000306 recurrent effect Effects 0.000 description 8
- 230000001537 neural effect Effects 0.000 description 7
- 239000000498 cooling water Substances 0.000 description 6
- 238000003062 neural network model Methods 0.000 description 6
- 238000006116 polymerization reaction Methods 0.000 description 6
- 238000009499 grossing Methods 0.000 description 5
- 238000012544 monitoring process Methods 0.000 description 5
- 238000011160 research Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 description 4
- 239000003921 oil Substances 0.000 description 4
- 238000005457 optimization Methods 0.000 description 4
- 238000012549 training Methods 0.000 description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 239000010687 lubricating oil Substances 0.000 description 3
- 210000005036 nerve Anatomy 0.000 description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000003592 biomimetic effect Effects 0.000 description 2
- 238000007418 data mining Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000013210 evaluation model Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000000547 structure data Methods 0.000 description 2
- 238000000714 time series forecasting Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000004378 air conditioning Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000002431 foraging effect Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000005461 lubrication Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F04—POSITIVE - DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS FOR LIQUIDS OR ELASTIC FLUIDS
- F04B—POSITIVE-DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS
- F04B51/00—Testing machines, pumps, or pumping installations
Landscapes
- Engineering & Computer Science (AREA)
- Mechanical Engineering (AREA)
- General Engineering & Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a method and a system for predicting faults and evaluating health of an air compressor based on a graph neural network, wherein the method and the system are used for carrying out convolution processing on different receptive fields on a multivariate time sequence to obtain a multiscale embedded feature; establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing diagram convolution to obtain an aggregation representation with instantaneous intervals, and solving the aggregation representation by utilizing a normal differential equation to obtain multi-scale information characterization; predicting a multivariate time sequence according to the obtained multi-scale information characterization; and calculating the health index of the air compressor according to the prediction sequence and the health degree model, and further performing fault prediction and health degree assessment on the air compressor. According to the invention, the dependency relationship among the variables in the whole evolution process is considered by establishing the dynamic time sequence diagram, the long-range dynamic relationship among the variables in the time sequence is captured, the detection variable value at the future moment is predicted, the fault of the air compressor system is detected, and the health degree of the air compressor is evaluated.
Description
Technical Field
The invention belongs to the technical field of combined research of Neural Networks (NNs) and air compressor fault diagnosis in Machine Learning (Machine Learning), relates to air compressor fault prediction based on the Neural Networks, and particularly relates to an air compressor fault prediction and health evaluation method and system based on a graph Neural network. The invention mainly utilizes Deep Learning (Deep Learning) and a graph neural network model (Graph Neural Network Model) to construct a Learning model, the model combines two technologies of a normal differential equation (Ordinary Differential Equations, ODEs) and a graph neural network (Graph Neural Network, GNN), long-range interior and interrelation of a time embedded sequence in time sequence prediction can be captured, multi-scale characterization (Multi-scale Representations) can be effectively utilized to form continuous evolution dynamic, and a health degree model is established to perform fault prediction and health degree assessment on predicted data. In the task of fault detection of the air compressor of the medium-low pressure air system, the method can learn rich relations and modes shared among different air compressor characteristic scales, and the robustness and accuracy of a prediction result are enhanced.
Background
In the current running air compressor system, most of equipment is designed only by considering the rated range of the machine, but not the system data range in normal running, and the system has no detection alarm function, and the related problems can be analyzed by very relying on field inspection equipment of operation maintenance personnel and collecting data. The current system running condition monitoring still adopts the traditional manual management mode, a very large lifting space still exists in the aspect of intelligent detection, meanwhile, the current detection index and the evaluation mode of the system are too single or scattered, the evaluation mode is behind, dynamic adjustment can not be carried out through the running condition, and the evaluation result is too dependent on the skill level of related staff.
Since multivariate time series is ubiquitous in life, it has a significant impact on many aspects of daily life, and predictions for multivariate time series play a significant role in a number of modern applications such as climate analysis, traffic and urban flow, the power industry and the financial market. A great deal of research effort has been directed to improving the performance of multivariate time series predictions. In air compressor systems, there are also a number of multivariate time series features, such as: exhaust pressure and exhaust temperature in the gas path system; the oil temperature and the oil pressure of the lubrication oil in the oil circuit system; cooling water temperature, cooling water pressure and the like in the waterway system. The method is characterized in that the fault attribute of the air compressor is successfully predicted according to the multivariable characteristics in the air compressor system, and the method plays an important role in health assessment of the air compressor system.
In the existing literature, multivariate time series predictions are based on an autoregressive model that has the ability to capture characteristics of the sequence. For example, models based on ARIMA, recurrent neural networks (Recurrent Neural Network, RNN) and gaussian processes have been used for various real world time series data. However, these models either ignore global information, fail to explicitly consider relationships between different variables, or lack the interpretability of the model. Recent advances in deep learning, particularly the rapid development of GNNs, stimulated some research into modeling multivariate time series by taking advantage of the ability of GNNs to learn a neighborhood context.
While existing GNN-based work has achieved encouraging results, there are problems that limit their performance: (1) Previous work has typically utilized the shallow layers of the GNN model, and most people achieve the best performance with a layer 2 network. This is because deep stacking of layers can lead to significant degradation in performance (also known as oversmooth phenomenon). However, in a layer 2 design, each node in the graph will only aggregate two-hop neighbors, ignoring long-range associations with other nodes, which typically contain expression characteristics that facilitate prediction; some efforts have been made to introduce infinite neural elements to solve this problem, but doing so makes it extremely difficult for the model to converge while learning the features of the graph; (2) Because of the above limitations, most GNN-based methods consider only static graphs, they employ a time-invariant structure between variables, and build graphs without dynamic relationships in different time steps; however, such designs ignore interactions between local information and global evolution; (3) Time series in the real world typically show various patterns in multi-scale observations; existing methods of extracting multi-scale characterizations tend to fuse them together simply by series or linear transformations. In general, such operations are difficult to provide interpretability, and also difficult to learn information-rich cross-scale interactions.
The invention overcomes the problem of excessive smoothing by incorporating the neural ordinary differential equation into the graph neural network, captures the long-distance semantic association, and enables the model to learn global co-evolution and interdependence relationships at the same time. And adjusting the continuous evolution track according to the control differential equation theory, and realizing successful prediction of the fault attribute of the air compressor system according to the multivariable characteristic attribute of the air compressor.
Disclosure of Invention
The invention aims at solving the problems in the prior art, and provides a method and a system for predicting air compressor faults and evaluating health degree based on a graph neural network, which are used for modeling a time sequence in multi-scale characterization by using a graph neural network model (Graph Neural Network Model) and a neural ordinary differential equation (Ordinary Differential Equations, ODEs), and utilizing potential connection assistance between a graph structure and the differential equation to infer and predict time evolution, and simultaneously simulate the dynamics of a multi-element time sequence in a continuous space, so as to predict the multi-variable time sequence in an air compressor system, and analyze the prediction attribute to evaluate the health of the air compressor system according to the fault attribute.
The idea of the invention is to construct a novel continuous graph multivariable prediction framework (Continuous Graph Multivariate Forecasting, CGMF) designed by combining a graph neural network and a neural ordinary differential equation, which is used for multivariable time sequence prediction in an air compressor system and fault detection by combining prediction variables. The framework provides a straightforward method to simulate the dynamics of a multi-element time series in continuous space. Specifically, (1) converting a multivariate time sequence obtained by monitoring in an air compressor system into multi-scale embedding for subsequent reasoning; (2) Establishing a dynamic multivariable timing diagram structure embedded by each scale, and incorporating ODE into GNN, so that the model can learn global co-evolution and mutual dependency simultaneously; (3) Fully utilizing the multi-scale information to generate a time sequence prediction result; (4) And establishing a health degree model, evaluating the predicted data, and judging the failure rate of the predicted data to complete the health evaluation of the air compressor system.
Based on the thought of the invention, the invention provides a method for predicting the faults and evaluating the health of an air compressor based on a graph neural network, which comprises the following steps:
s1, carrying out convolution processing on multiple variable time sequences with different receptive fields to obtain multi-scale embedded features;
S2, establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing diagram convolution to obtain an aggregation representation with instantaneous intervals, and solving the aggregation representation by utilizing a normal differential equation to obtain multi-scale information characterization;
s3, predicting a multivariate time sequence according to the obtained multi-scale information representation;
s4, calculating the health index of the air compressor according to the prediction sequence and the following health degree model, and further performing fault prediction and health degree assessment on the air compressor:
wherein ,Wn A weight for health representing a given nth observation variable, n=1, 2, …, N; n represents the number of observed variables;an air compressor health index representing the n-th attribute variable of the predicted τ times;Andrespectively representing the maximum value and the minimum value of an nth observation variable of the input time sequence, namely the maximum value and the minimum value of the observation variable n when the air compressor works normally;The standard value of the nth variable representing the input time series is the average value of the nth observed variable of the input time series.
In the step S1, modeling is performed on a plurality of scales, and the multi-scale embedded feature is obtained by inputting the multi-scale time series. Given a multivariate time series Wherein T represents a multivariable time series length, N represents N observed variables at time T (the observed variables of the air compressor include exhaust pressure, exhaust temperature, lubricating oil pressure, cooling water temperature, etc.), and M represents an observed dimension of each variable. The goal of this step is to create a series of embedded features +.>For further processing. Each of which is wherein S＝{s1 ,s 2 ,…,s k Is represented by }The set of all dimensions is such that,representing a time step of scale s (in particular a time sequence of length T divided into T/s data segments on the scale of s), d representing the embedding dimension,/->Representing a rounding down. In implementation, the present invention employs a plurality of two-dimensional convolution units to achieve this conversion, in particular, convolution units with different receptive fields (i.e., different convolution kernels) are utilized to generate a multi-scale embedded representation as follows:
the input and output channels of the two-dimensional convolution are M and d, respectively, corresponding to the dimensions of the input features and the output embedded characterizations. The convolution kernel is 1 x s in size, that is, the time series of each variable is embedded with a scale s. With such an arrangement, a multi-scale embedded representation can be ultimately generatedAnd used for subsequent reasoning.
In the step S2, the multi-scale embedded token H obtained in S1 is first utilized, and the adjacency matrix of the dynamic timing diagram is calculated by a multi-head attention mechanism algorithm to construct a dynamic diagram structure.
Firstly, a dynamic time sequence diagram is constructed by utilizing multi-scale embedded features, the time sequence diagram consists of T multiplied by N nodes, and represents variables in time sequences with all time steps of T, and an adjacency matrix is formedThe relationship between them is recorded. The relationship of the node to the neighbor node is then learned by GNN. Classical GNNs update the representation of each node by iteratively aggregating their own and neighbor information. Given a dynamic timing diagram, an s-scale discrete information propagation layer can be described as follows:
wherein ,information embedded representation representing the first aggregation, < >>Representing an adjacency matrix with self-loop operation, I representing an identity matrix,Representation->F (·) represents a feed forward network, typically consisting of dense layers and an activation function, L is the number of discrete propagation layers. However, a large number of discrete propagation layers l→infinity will lead to excessive smoothing problems. To address this problem, we incorporate the nerve ODE into CGMF, allowing it to aggregate almost infinitely to capture long-distance internals and interrelationships and overcome the transition smoothing problem.
We build a dynamic timing diagram and connect variables by a method based on a multi-headed attention mechanism because it can focus on the most relevant and important parts of the variables in the time series. Multi-scale temporal embedding of each variableFirstly, remolding them into a two-dimensional matrix, and embedding H into the matrix multi-scale time 2d The expression is as follows:
they are then mapped into higher order query and key matrices by linear transformation, with C parallel heads, expressed as follows:
wherein all of themRepresent learning parameters d a Representing dimensions of the query and key matrices; then, point multiplication is applied to calculate the attention weight between every two embeddings, with the formula:
wherein ,is an adjacency matrix of the dynamic timing diagram. Finally, the first p weights are kept and the other weights are reset to 0, ordered by matrix element size, since the few dot product weights contribute to the main attention, while the other weights have negligible impact.
Thus, this limitation can be overcome by performing a large number of polymerizations in several discrete propagation steps. Specifically, let L '=l/K denote intermediate variables between K and L, and denote the aggregation interval, L' =1 in a typical discrete GNN. K is the aggregate number, indicating that information within K-hop (i.e., range K) neighbors can be considered. Based on the above settings, the polymerization with a spacing l' on one s-scale can be expressed by:
We useTo simply represent. Now we can set L to a very small constant. When K is → is 0. We obtain a continuous information propagation process with instantaneous spacing, when l=0, we obtain an aggregate representation with instantaneous spacing according to equation (1.6) above +.>
And solving the multi-scale information representation by using a normal differential equation (a nerve ODE solver)The method comprises the following steps:
wherein ,representing the embedding at each scale, g s (. Cndot.) can be any input +.>Is provided. Here let->No additional learning parameters are required. This not only increases the computational efficiency of the integral approximation, but also aggregates the linear features in all nodes. By infinite polymerization, I>Long-distance correlation between nodes in the graph structure is included, and excessive smoothing phenomenon does not occur.
Subsequently, the pattern of the multivariate sequence can be further extracted and introduced into the nonlinearity. The multi-scale information characterization of the node is updated by the biomimetic transformation and activation function, represented as follows:
wherein, sigma (·) represents the PReLU activation function,and b represents the learned parameters.
Multi-scale information characterization of all nodes on the s-scale
In the step S3, a multivariate time series is predicted. Multi-scale information characterization in obtaining equation (1.8) Then, a basic Multi-Layer Perceptron (MLP) is used as a predictor to generate predicted Multi-variable time series predictions of τ time steps.
Where θ represents a learnable parameter. For model optimization, general optimizers such as Adam and RMSprop can be employed to calculate gradients and update all learning parameters. And uses MSE loss as an optimization objective because the prediction task can be regarded as a linear regression task.
In the step S4, after the multivariate time series prediction value of S3 is obtained, a health degree model is built for the prediction data to perform fault prediction and health degree evaluation. The health index is a concept proposed for describing the operation condition of the power equipment. The basic idea is to reflect the health of the device being evaluated in digital form, based on a comprehensive mathematical transformation of each parameter being evaluated and the calculation of the quantitative index. The specific calculation of the health index for a single variable is as follows:
wherein ,the air compressor health index of the n-th variable of the predicted tau time; andRespectively representing the maximum value and the minimum value of an nth observation variable of the input time sequence, namely the maximum value and the minimum value of the observation variable n when the air compressor works normally; / >The standard value of the nth variable representing the input time series.
The running health index of the air compressor is defined as follows:
wherein ,Wn Representing the weight of a given nth variable with respect to health, n=1, 2, …, N.
Calculated HV T+τ The result of (2) is between 0 and 1. The research shows that the relation between the operation health index of the air compressor and the failure rate is as follows: when the running health index is 0.7-1, the equipment health degree is healthy, and the fault occurrence rate is very low; when the running health index is between 0.6 and 0.7 (excluding the end values of 0.6 and 0.7), the equipment health degree is sub-health, and the failure occurrence rate is low; when the running health index is 0-0.6, the equipment health degree is fault, and the fault occurrence rate is high. Thus, we can predict the resulting data fromAnd obtaining fault information and health degree of the air compressor system.
The method for predicting the multivariate time series based on the graph neural network and the neural ODE can overcome the problem of excessive smoothness, capture long-distance semantic association, learn global co-evolution and mutual dependency relationships at the same time, successfully predict multivariate time series monitoring data in an air-conditioning system, and realize fault judgment and health judgment by constructing a health degree model. It should be noted that the method of the present invention is not only aimed at this problem, but can be widely applied to time series scenes in the real world, and the data type is a multi-variable time series prediction problem.
Based on the inventive thought, the invention provides an air compressor fault prediction and health evaluation system based on a graph neural network, which comprises the following steps:
a multivariate time series prediction model, the model comprising:
the multi-scale embedded feature acquisition module is used for carrying out convolution processing on different receptive fields on the multi-variable time sequence to obtain multi-scale embedded features;
the multi-scale information characterization acquisition module is used for establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing diagram convolution to obtain an aggregation representation with instantaneous intervals, and solving the aggregation representation by utilizing a normal differential equation to obtain multi-scale information characterization;
the multivariable time sequence prediction module is used for predicting the multivariable time sequence according to the obtained multi-scale information representation;
the air compressor fault prediction and health degree assessment model is used for calculating the health index of the air compressor according to the following health degree model according to a prediction sequence, and further carrying out fault prediction and health degree assessment on the air compressor:
wherein ,Wn A weight for health representing a given nth observation variable, n=1, 2, …, N; n represents the number of observed variables;an air compressor health index representing the n-th attribute variable of the predicted τ times; / > andRespectively representing the maximum value and the minimum value of the nth observation variable of the input time sequence;The standard value of the nth variable representing the input time series is the average value of the nth observed variable of the input time series.
The multi-scale embedded feature acquisition module performs an operation according to the step S1; it comprises the following steps:
the multi-variable time sequence dividing unit is used for dividing the input multi-variable sequence according to different scales to obtain time sequences of a plurality of time periods;
and the convolution unit is used for carrying out convolution processing on the time sequence of each divided time period to generate a multi-scale embedded representation.
The convolution units have different receptive fields (i.e., convolution kernel sizes) for different scales. For scale s, the convolution kernel size of the convolution unit is 1×s.
The multi-scale information characterization acquisition module executes operation according to the step S2; it comprises the following steps:
a dynamic time sequence diagram establishing unit for establishing a dynamic time sequence diagram based on a multi-head attention mechanism;
the graph convolution processing unit is used for discretizing the dynamic time sequence graph to obtain an aggregation representation with instantaneous intervals;
the ODE solver is used for solving the aggregation representation with the instantaneous interval to obtain the multi-scale information representation.
The dynamic time chart creation unit obtains the adjacency matrix a of the dynamic time chart according to the above formula (1.5).
The graph convolution processing unit obtains an aggregate representation with temporal spacing according to the above equation (1.6)
The ODE solver solves according to the formula (1.7) to obtain the multi-scale information characteristic representation
The multi-scale information characterization acquisition module further comprises an updating unit, wherein the updating unit is used for updating the multi-scale information characteristics through the bionic transformation and the activation function. Specifically, the multiscale information characteristic representation obtained by the ODE solver is updated according to a formula (1.8).
The multivariate time series prediction module uses a Multi-Layer Perceptron (MLP).
The multivariate time series prediction model further comprises an optimizer for optimizing model parameters. Optimizers such as Adam and RMSprop employed in the present invention calculate gradients and update all learning parameters; and takes MSE loss as an optimization target.
Compared with the prior art, the air compressor fault prediction and health evaluation method and system based on the graphic neural network provided by the invention have the following beneficial effects:
1. the invention is based on the neural network of the graph, combine the neural equation of often differentiation, has proposed a method (CGMF) used for studying the prediction of the time series of continuous multivariable, it considers the dependency between variables in the whole evolution process through setting up the dynamic time sequence chart, catch the long-range dynamic relation between every variable in the time series; detecting variable values at future time are predicted by utilizing multielement monitoring data in the air compressor system, including data such as exhaust pressure, exhaust temperature, lubricating oil pressure, cooling water temperature and the like, detecting faults of the air compressor system by analyzing and judging the predicted data, and evaluating health degree;
2. Modeling a multi-variable time sequence on multiple scales, exploring a rich mode on each scale, constructing a dynamic time sequence diagram, dynamically extracting cross-scale knowledge from the dynamic time sequence diagram and generating multi-variable prediction;
3. compared with the traditional manual experience judging method, the method provided by the invention realizes the intellectualization of the detection system and the accuracy of fault prediction, and has a guiding effect on the intelligent evaluation of the normal operation and the health degree of the air compressor system.
4. Based on the method, the method is verified on a time series data set of the attributes of the multivariable air compressor collected in the real world; experimental results show that compared with the existing most advanced model, the prediction method of the invention has more excellent CGMF performance.
Drawings
Fig. 1 is a schematic flow chart of a method for predicting air compressor faults and evaluating health degree based on a graph neural network.
Interpretation of the terms
RNN is an abbreviation for recurrentneuroalnetwork, which means "recurrent neural network". This is a type of recurrent neural network that takes sequence (e.g., time-series) data as input, performs recursion in the evolution direction of the sequence, and all nodes are chained. The method can capture the relation of the sequence data before and after, and plays an important role in tasks such as natural language processing, time sequence prediction and the like. The theoretical basis of which can be referred to in the literature [ A.Sagher and M.Kotb.Unsuupervised Pre-Training of A Deep Lstm-based Stacked Autoencoder for Multivariate Time Series Forecasting Problems.scientific Reports,2019,9 (1): 1-16 ]
GNN is an abbreviation of Graph Neural Network, and represents a "graph neural network", which is an algorithm generic term for learning graph structure data by using a neural network, extracting and exploring features and modes in the graph structure data, and meeting the requirements of graph learning tasks such as clustering, classifying, predicting, dividing, generating and the like; the theory of relevance can be referred to in the literature: [ F.Scarselli, M.Gori, A.C.Tsoi, et al, graph Neural Network model IEEE Transactions on Neural Networks,2008,20 (1): 61-80.)
ODEs are abbreviations for Ordinary Differential Equations, representing the "ordinary differential equation": the ordinary differential equation parameterizes the derivative of the hidden state by using a neural network, rather than the discrete sequence of hidden layers used in traditional models (such as ResNet and RNN), and has the characteristic of balancing numerical accuracy and computation, while being capable of significantly saving memory costs; the theoretical basis can be referred to in the literature [ R.T.Chen, Y.Rubanova, J.Bettencourt, and D.Duvenaud. Neal ordinary differential equivalents. In NeurIPS,2018, pp.6572-6583 ]
Detailed Description
The invention will be further described with reference to the accompanying drawings.
Example 1
The method for predicting air compressor faults and evaluating health based on the graph neural network provided in the embodiment, as shown in fig. 1, includes:
S1, carrying out convolution processing on the multivariate time sequence with different receptive fields to obtain the multiscale embedded features.
Given a multivariate time seriesWherein T represents a multivariable time series length, N represents N observed variables at time T (the observed variables of the air compressor include exhaust pressure, exhaust temperature, lubricating oil pressure, cooling water temperature, etc.), and M represents an observed dimension of each variable.
For a multivariate time series X at different scales s 1:T Dividing intoA data segment, for any time period t s Is convolved by a convolution unit, and is expressed as follows:
the input and output channels of the two-dimensional convolution are M and d, respectively, corresponding to the dimensions of the input features and the output embedded features, and the convolution kernel size is 1×s.
Therefore, each time period is sequentially processed, and the multi-scale embedded characterization corresponding to the scale s can be generated
Based on different dimensions s=s 1 ,s 2 ,…,s k Processing according to the above operation results in a series of embedded features
S2, establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing graph convolution to obtain an aggregate representation with instantaneous intervals, and solving the aggregate representation by utilizing a normal differential equation to obtain multi-scale information characterization.
First, multi-scale temporal embedding on a per-variable basisRemolding them into a two-dimensional matrix, and embedding H in a multi-scale time after matrixing 2d The expression is as follows:
they are then mapped into higher order query and key matrices by linear transformation, with C parallel heads, expressed as follows:
wherein all of themRepresent learning parameters d a Representing dimensions of the query and key matrices; then, point multiplication is applied to calculate the attention weight between every two embeddings, with the formula:
wherein ,is an adjacency matrix of the dynamic timing diagram. Finally, the first p weights are kept and the other weights are reset to 0, ordered by matrix element size, since the few dot product weights contribute to the main attention, while the other weights have negligible impact.
Thus, this limitation can be overcome by performing a large number of polymerizations in several discrete propagation steps. Specifically, let L' =l/K denote intermediate variables between K and L, and denote the aggregation interval. Based on the above settings, the polymerization with a spacing l' on one s-scale can be expressed by:
we useTo simply represent. Now we can set L to a very small constant. When K is → is 0. We obtain a continuous information propagation process with instantaneous spacing, when l=0, we obtain an aggregate representation with instantaneous spacing according to the above formula +. >
And solving the multi-scale information representation by using a normal differential equation (a nerve ODE solver)The method comprises the following steps:
wherein ,representing the embedding at each scale, g s (. Cndot.) can be any input +.>Is provided. Here let->No additional learning parameters are required. This not only increases the computational efficiency of the integral approximation, but also aggregates the linear features in all nodes. By infinite polymerization, I>Long-distance correlation between nodes in the graph structure is included, and excessive smoothing phenomenon does not occur.
Subsequently, the pattern of the multivariate sequence can be further extracted and introduced into the nonlinearity. The multi-scale information characterization of the node is updated by the biomimetic transformation and activation function, represented as follows:
wherein, sigma (·) represents the PReLU activation function,and b representsThe learned parameters.
Multi-scale information characterization of all nodes on the s-scale
S3, predicting the multivariate time sequence according to the obtained multi-scale information characterization.
In this step, a multi-layer perceptron is used as a predictor to generate predicted multivariate time series predictions for the τ time steps.
Where θ represents a learnable parameter.
And S4, calculating the health index of the air compressor according to the prediction sequence and the health degree model, and further performing fault prediction and health degree assessment on the air compressor.
The specific calculation of the health index for a single variable is as follows:
wherein ,an air compressor health index representing the n-th attribute variable of the predicted τ times;Andrespectively representing the maximum value and the minimum value of an nth observation variable of the input time sequence, namely the maximum value and the minimum value of the observation variable n when the air compressor works normally;Representing the inputInto the standard value of the nth variable of the time series.
The running health index of the air compressor is as follows:
wherein ,Wn Representing the weight of a given nth variable with respect to health, n=1, 2, …, N.
Calculated HV T+τ The result of (2) is between 0 and 1. The research shows that the relation between the operation health index of the air compressor and the failure rate is as follows: when the running health index is 0.7-1, the equipment health degree is healthy, and the fault occurrence rate is very low; when the running health index is between 0.6 and 0.7 (excluding the end values of 0.6 and 0.7), the equipment health degree is sub-health, and the failure occurrence rate is low; when the running health index is 0-0.6, the equipment health degree is fault, and the fault occurrence rate is high. Therefore, fault information and health degree of the air compressor system can be obtained according to the data obtained through prediction.
Example 2
The embodiment provides an air compressor fault prediction and health evaluation system based on a graph neural network, which comprises a multivariate time sequence prediction model and an air compressor fault prediction and health evaluation model.
The multi-variable time sequence prediction model comprises a multi-scale embedded feature acquisition module, a multi-scale information characterization acquisition module, a multi-variable time sequence prediction module and an optimizer.
And the multiscale embedded feature acquisition module is used for carrying out convolution processing on different receptive fields on the multivariate time sequence to obtain multiscale embedded features.
The module performs the operation according to step S1 in embodiment 1; it comprises the following steps:
the multi-variable time sequence dividing unit is used for dividing the input multi-variable sequence according to different scales to obtain time sequences of a plurality of time periods.
A multivariable time series dividing unit for dividing the multivariable time series in different scales sX 1:T Dividing intoData segments.
And the convolution unit is used for carrying out convolution processing on the time sequence of each divided time period to generate a multi-scale embedded representation.
The convolution units have different receptive fields (i.e., convolution kernel sizes) for different scales. For scale s, the convolution kernel size of the convolution unit is 1×s.
And (3) carrying out convolution processing on the time sequence of each time period of the convolution processing of the time sequence divided under the corresponding scale by utilizing a convolution unit according to a formula (1.1) to generate a corresponding multi-scale embedded representation.
The multi-scale information characterization acquisition module is used for establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing graph convolution to obtain an aggregate representation with instantaneous intervals, and solving the aggregate representation by utilizing a normal differential equation to obtain multi-scale information characterization.
The module executes the operation according to the step S2; it comprises the following steps:
a dynamic time sequence diagram establishing unit for establishing a dynamic time sequence diagram based on a multi-head attention mechanism; specifically, the adjacency matrix A of the dynamic time sequence diagram is obtained according to the formula (1.5).
The graph convolution processing unit is used for discretizing the dynamic time sequence graph to obtain an aggregation representation with instantaneous intervals; in particular, an aggregate representation with instantaneous spacing is obtained according to the above formula (1.6)
An ODE solver for solving the aggregate representation with the instantaneous interval to obtain a multi-scale information representation; solving according to the formula (1.7) to obtain the multi-scale information characteristic representation
The updating unit is used for updating the multi-scale information characteristics through the bionic transformation and the activation function; specifically, the multiscale information characteristic representation obtained by the ODE solver is updated according to a formula (1.8).
And the multivariable time sequence prediction module is used for predicting the multivariable time sequence according to the obtained multi-scale information characterization.
The module uses a Multi-Layer Perceptron (MLP).
And the optimizer is used for optimizing the model parameters. The optimizer adopted in the embodiment is Adam algorithm, so as to calculate the gradient and update all learning parameters; and takes MSE loss as an optimization target.
The air compressor fault prediction and health degree assessment model is used for calculating the health index of the air compressor according to the following health degree model according to a prediction sequence, and further carrying out fault prediction and health degree assessment on the air compressor:
wherein ,Wn A weight for health representing a given nth observation variable, n=1, 2, …, N; n represents the number of observed variables;an air compressor health index representing the n-th attribute variable of the predicted τ times;Andrespectively representing the maximum value and the minimum value of the nth observation variable of the input time sequence;The standard value of the nth variable representing the input time series is the average value of the nth observed variable of the input time series.
Application example
The air compressor fault prediction and health evaluation method CGMF based on the graph neural network provided in embodiment 1 is adopted to perform experiments on a time series data set of a multivariate air compressor collected in the real world, and the multivariate observation variables include: current (unit: a), voltage (unit: volt), exhaust pressure (unit: kpa), exhaust temperature (unit: celsius), cooling water pressure (unit: kpa), and the like, and the granularity of each data is 1 ms/time. Meanwhile, we compared the present invention with 5 different baseline models, respectively: a statistical model ARIMA and a neural network model (GRU, latentODE, MTGNN, STODE).
In order to meet the requirement of the deep learning model learning problem, the data set required by each test is divided into a training set, a verification set and a test set, and the ratio is 6:2:2. The training set and the verification set are used for carrying out the test on the parameters and the adjustment super parameters of the multivariate time series prediction model provided in the embodiment 2, and the test set is used for carrying out the test on the basis of the air compressor fault prediction and health evaluation system based on the graph neural network provided in the embodiment 2 according to the method provided in the embodiment 1. Meanwhile, in order to verify and predict the effectiveness of failure and health degree evaluation of the air compressor system in different time periods, values of various tau are set. Taking the significance of time into consideration, taking three future time periods of 1 hour, 1 day and 1 week to predict and report the result.
The relative square root error (Root Relative Squared Error, RSE) and empirical correlation coefficients (Empirical Correlation Coefficient, CORR) widely used in time series prediction are used as evaluation indices (the lower the RSE value, the better the model method and the higher the CORR value, the better the model). Where Ours represents the results of the model of the invention, the best results are indicated in bold.
Table 1: prediction of future τ time across all datasets in an application example
The rest of the methods in the table are described as follows:
ARIMA: is a typical statistical-based time series prediction model, which consists of three parts: autoregressive models, moving average models, and differential orders. It only needs the endogenous variable and does not need to make use of other external variables to complete the prediction. It is capable of making predictions of water flow based on historical observations, as described in the process references [ Wang, wen-chuan, et al, "Improving forecasting accuracy of annual runoff time series using ARIMA based on EEMD Decomposition." Water Resources Management 29.8.8 (2015): 2655-2675 ]
GRU: a variant based on Recurrent Neural Network (RNN) improvement. The addition of several gating logic units to the GRU compared to the original recurrent neural network enables the network to capture longer dependencies and process long-term sequential events. And the problems of gradient explosion, gradient disappearance and the like of the original circulating neural network in the training process can be relieved to a certain extent. Such models are widely used in timing prediction problems, and the implementation process can be respectively referred to in the literature [ Yang, shuyu, et al, "Real-time reservoir operation using recurrent neural networks and inflow forecast from adistributed hydrological model" Journal of Hydrology 579 (2019): 124229 ] [ Apaydin, halit, et al, "Comparative analysis of recurrent neural network architectures for reservoir inflow foraging", "Water 12.5 (2020): 1500 ]
Latentde: it extends the discrete RNN to continuous time concealment dynamics defined by the ODE. It treats the potential representation as a time series variable in the RNN, being able to handle any time interval between observations. Its implementation procedure and details reference [ Y.Rubanova, R.T.Chen, and D.Duvenaud. Latent odes for irregularly sampled time series. In NeurIPS,2019 ]
STODE: an improvement on neural ordinary differential equations is provided for time-space data by providing a continuous graph neural network with restart distribution to capture long-distance spatial correlation of multivariable traffic flow predictions. Its implementation procedure and details reference [ Z.Fang, et al spatial-temporal graph ode networks for traffic flow shaping.proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery&Data Mining.2021 ]
MTGNN: predicting the multivariable time sequence based on a graph neural network, and capturing the spatial and time dependence in the time sequence by extracting the unoriented relation between the variable graphs and combining the graphs and the time convolution module to further provide a novel mixed jump propagation layer and an expanded initial layer. Its implementation procedure and details reference [ Z.Wu, S.Pan, G.Long, et al, C.connection the dots: multivariate time series forecasting with graph neural networks of the 26th ACM SIGKDD international conference on knowledge discovery&data mining,2020,pp.753-763 ]
As can be seen from the experimental results in Table 1, the air compressor fault prediction and health degree evaluation method based on the graph neural network provided by the invention can be used for the prediction task of the multivariate time sequence of the air compressor system, and after accurately predicting the data of the future time period, the prediction attribute is converted into the health degree index by establishing the health degree evaluation model, so that the fault can be successfully predicted. Compared with other deep learning methods, the method has the advantage that the performance of the method on two indexes of relative square root error and experience correlation coefficient is obviously improved.
Therefore, the air compressor fault prediction and health evaluation method based on the graphic neural network is successfully applied to multivariable fault prediction and health evaluation of an air compressor system, and integrates long-range internal and interrelationships among variables and trans-scale time information with expressive force. By inputting the multivariable time sequence monitoring data of the air compressor system, the running state value of the air compressor in the future time period can be accurately predicted, and the fault analysis and the health degree evaluation are carried out. Meanwhile, the invention also provides a general model in a time sequence prediction task.
Those of ordinary skill in the art will recognize that the embodiments described herein are for the purpose of aiding the reader in understanding the principles of the present invention and should be understood that the scope of the invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific modifications and combinations from the teachings of the present disclosure without departing from the spirit thereof, and such modifications and combinations remain within the scope of the present disclosure.
Claims (10)
1. The air compressor fault prediction and health evaluation method based on the graph neural network is characterized by comprising the following steps of:
s1, carrying out convolution processing on multiple variable time sequences with different receptive fields to obtain multi-scale embedded features;
s2, establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing diagram convolution to obtain an aggregation representation with instantaneous intervals, and solving the aggregation representation by utilizing a normal differential equation to obtain multi-scale information characterization;
s3, predicting a multivariate time sequence according to the obtained multi-scale information representation;
s4, calculating the health index of the air compressor according to the prediction sequence and the following health degree model, and further performing fault prediction and health degree assessment on the air compressor:
wherein ,Wn A weight for health representing a given nth observation variable, n=1, 2, …, N; n represents the number of observed variables;an air compressor health index representing the n-th attribute variable of the predicted τ times; andRespectively representing the maximum value and the minimum value of the nth observation variable of the input time sequence;The standard value of the nth variable representing the input time series.
2. The air compressor failure prediction and health evaluation method based on the neural network according to claim 1, wherein in step S1, a multivariate time sequence is givenWherein T represents the length of a multivariate time sequence, N represents N observed variables at the moment T, and M represents the observed dimension of each variable;
for a multivariate time series X at different scales s 1:T Dividing intoA data segment, for any time period t s Is convolved by a convolution unit, and is expressed as follows:
the input and output channels of the two-dimensional convolution are M and d, respectively, corresponding to the dimensions of the input features and the output embedded features, and the convolution kernel size is 1×s.
Thus for each timeThe segments are sequentially processed to generate a multi-scale embedded representation corresponding to the scale s
Based on different dimensions s=s 1 ,s 2 ,…,s k Processing according to the above operation results in a series of embedded features
3. The air compressor failure prediction and health evaluation method based on a neural network according to claim 1, wherein first, multi-scale time embedding based on each variableRemolding them into a two-dimensional matrix, and embedding H in a multi-scale time after matrixing 2d The expression is as follows:
they are then mapped into higher order query and key matrices by linear transformation, with C parallel heads, expressed as follows:
wherein all of themRepresent learning parameters d a Representing dimensions of the query and key matrices; then, point multiplication is applied to calculate the attention weight between every two embeddings, with the formula:
wherein ,is an adjacency matrix of the dynamic timing diagram;
then, the aggregation with a spacing l' on one s-scale can be expressed by:
by usingTo simply represent; when l=0, an aggregate representation with temporal spacing is obtained according to the above formula +.>
And solving the multi-scale information representation by utilizing a normal differential equationThe method comprises the following steps:
wherein ,representing the embedding at each scale, g s (. Cndot.) can be any input +.>Is a feed forward network of (2); here, the
4. The air compressor failure prediction and health evaluation method based on a neural network according to claim 3, wherein the first p weights are reserved according to the matrix element size ordering, and the other weights are reset to 0.
5. The air compressor fault prediction and health evaluation method based on the graphic neural network according to claim 3 or 4, wherein the multi-scale information characterization of the nodes is further updated by a bionic transformation and activation function, and is represented as follows:
wherein, sigma (·) represents the PReLU activation function,and b represents a learned parameter;
multi-scale information characterization of all nodes on the s-scale
6. The method for predicting air compressor failure and evaluating health based on a neural network according to claim 5, wherein in step S3, a multi-layer perceptron is used as a predictor to generate predicted multivariate time series predicted values of τ time steps.
Where θ represents a learnable parameter.
7. The air compressor fault prediction and health evaluation method based on the graph neural network according to claim 1, wherein in step S4, the relationship between the operation health index of the air compressor and the fault rate thereof is as follows: when the running health index is 0.7-1, the equipment health degree is healthy, and the fault occurrence rate is very low; when the running health index is between 0.6 and 0.7 (excluding the end values of 0.6 and 0.7), the equipment health degree is sub-health, and the failure occurrence rate is low; when the running health index is 0-0.6, the equipment health degree is fault, and the fault occurrence rate is high.
8. An air compressor machine fault prediction and health degree evaluation system based on picture neural network, characterized by comprising:
a multivariate time series prediction model, the model comprising:
the multi-scale embedded feature acquisition module is used for carrying out convolution processing on different receptive fields on the multi-variable time sequence to obtain multi-scale embedded features; the module comprises:
the multi-variable time sequence dividing unit is used for dividing the input multi-variable sequence according to different scales to obtain time sequences of a plurality of time periods;
the convolution unit is used for carrying out convolution processing on the time sequence of each divided time period to generate a multi-scale embedded representation;
the multi-scale information characterization acquisition module is used for establishing a dynamic time sequence diagram based on a multi-head attention mechanism; discretizing the dynamic time sequence diagram by utilizing diagram convolution to obtain an aggregation representation with instantaneous intervals, and solving the aggregation representation by utilizing a normal differential equation to obtain multi-scale information characterization; the module comprises:
a dynamic time sequence diagram establishing unit for establishing a dynamic time sequence diagram based on a multi-head attention mechanism;
the graph convolution processing unit is used for discretizing the dynamic time sequence graph to obtain an aggregation representation with instantaneous intervals;
An ODE solver for solving the aggregate representation with the instantaneous interval to obtain a multi-scale information representation;
the multivariable time sequence prediction module is used for predicting the multivariable time sequence according to the obtained multi-scale information representation; the module uses a multi-layer perceptron;
the air compressor fault prediction and health degree assessment model is used for calculating the health index of the air compressor according to the following health degree model according to a prediction sequence, and further carrying out fault prediction and health degree assessment on the air compressor:
wherein ,Wn A weight for health representing a given nth observation variable, n=1, 2, …, N; n represents the number of observed variables;an air compressor health index representing the n-th attribute variable of the predicted τ times; andRespectively representing the maximum value and the minimum value of the nth observation variable of the input time sequence;The standard value of the nth variable representing the input time series.
9. The air compressor fault prediction and health evaluation system based on the graphic neural network according to claim 8, wherein the multi-scale information characterization acquisition module further comprises an updating unit for updating the multi-scale information features through a bionic transformation and an activation function.
10. The air compressor fault prediction and health assessment system based on a graph neural network of claim 8, wherein the multivariate time series prediction model further comprises an optimizer for optimizing model parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310108821.6A CN116578858A (en) | 2023-02-14 | 2023-02-14 | Air compressor fault prediction and health degree evaluation method and system based on graphic neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310108821.6A CN116578858A (en) | 2023-02-14 | 2023-02-14 | Air compressor fault prediction and health degree evaluation method and system based on graphic neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116578858A true CN116578858A (en) | 2023-08-11 |
Family
ID=87544148
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310108821.6A Pending CN116578858A (en) | 2023-02-14 | 2023-02-14 | Air compressor fault prediction and health degree evaluation method and system based on graphic neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116578858A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556331A (en) * | 2024-01-11 | 2024-02-13 | 广东艾林克能源装备有限公司 | AI-enhancement-based air compressor maintenance decision method and system |
-
2023
- 2023-02-14 CN CN202310108821.6A patent/CN116578858A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117556331A (en) * | 2024-01-11 | 2024-02-13 | 广东艾林克能源装备有限公司 | AI-enhancement-based air compressor maintenance decision method and system |
| CN117556331B (en) * | 2024-01-11 | 2024-04-26 | 广东艾林克能源装备有限公司 | AI-enhancement-based air compressor maintenance decision method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yu et al. | LSTM-EFG for wind power forecasting based on sequential correlation features | |
| Chen et al. | Learning graph structures with transformer for multivariate time-series anomaly detection in IoT | |
| Wang et al. | A compound framework for wind speed forecasting based on comprehensive feature selection, quantile regression incorporated into convolutional simplified long short-term memory network and residual error correction | |
| CN113313947B (en) | Road condition evaluation method of short-term traffic prediction graph convolution network | |
| CN108445752B (en) | Random weight neural network integrated modeling method for self-adaptively selecting depth features | |
| Ma et al. | Histgnn: Hierarchical spatio-temporal graph neural network for weather forecasting | |
| Zhang et al. | Multistep speed prediction on traffic networks: A graph convolutional sequence-to-sequence learning approach with attention mechanism | |
| CN113780684A (en) | Intelligent building user energy consumption behavior prediction method based on LSTM neural network | |
| Zou et al. | Deep non-crossing probabilistic wind speed forecasting with multi-scale features | |
| CN111340282A (en) | DA-TCN-based method and system for estimating residual service life of equipment | |
| Konstantakopoulos et al. | Design, benchmarking and explainability analysis of a game-theoretic framework towards energy efficiency in smart infrastructure | |
| Qi et al. | ADGCN: An asynchronous dilation graph convolutional network for traffic flow prediction | |
| CN117175588B (en) | Space-time correlation-based electricity load prediction method and device | |
| Awad et al. | Prediction of water demand using artificial neural networks models and statistical model | |
| CN115965160A (en) | Data center energy consumption prediction method and device, storage medium and electronic equipment | |
| CN113128666A (en) | Mo-S-LSTMs model-based time series multi-step prediction method | |
| CN116578858A (en) | Air compressor fault prediction and health degree evaluation method and system based on graphic neural network | |
| Fan et al. | Online learning-empowered smart management for A2O process in sewage treatment processes | |
| Yu et al. | A novel discussion on two long-term forecast mechanisms for hydro-meteorological signals using hybrid wavelet-NN model | |
| Lv et al. | A transportation Revitalization index prediction model based on Spatial-Temporal attention mechanism | |
| CN117354172A (en) | Network traffic prediction method and system | |
| Zhang et al. | Adaptive spatio-temporal convolutional network for traffic prediction | |
| Liu et al. | Prediction of dam horizontal displacement based on CNN-LSTM and attention mechanism | |
| Shen et al. | Long-term multivariate time series forecasting in data centers based on multi-factor separation evolutionary spatial–temporal graph neural networks | |
| Zhang et al. | Ensemble optimization approach based on hybrid mode decomposition and intelligent technology for wind power prediction system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |