A method, medium and device for predicting the remaining service life of a flight conveyor

Abstract

The application discloses a kind of scraper conveyer remaining useful life prediction method, medium and equipment, it is related to remaining life prediction technical field, including the different sensor data of the scraper conveyer obtained, constructs multiple original graph data;Through convolution layer, the graph data constructed after pre-processing is used as the input of model, feature is extracted, the dynamic dependency between sensors is obtained through graph structure learning layer, and the spatial correlation of sensor is fused;The spatial features of aggregated nodes are updated in the graph attention network in the space-time fusion layer, and then the dependency between sequences is extracted using the time attention mechanism and bidirectional gated recurrent unit, and finally the fused space-time features are input into the fully connected layer to determine the final remaining useful life, to realize the life prediction of scraper conveyer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of residual life prediction, and particularly relates to a residual service life prediction method, medium and device for a scraper conveyor. BACKGROUND

[0002] The scraper conveyor is an important transportation equipment in the coal mining face in the coal mine, and its reliability and service life directly affect the efficiency and safety of the coal mine production. Due to the poor environment in the coal mine, the scraper conveyor is prone to various faults in long-term operation, and once a fault or maintenance problem occurs, it will cause huge economic losses, so it is of great significance to predict the residual service life thereof.

[0003] There are mainly two kinds of deep learning networks applied to residual useful life (RUL) prediction: convolutional neural network (CNN) and recurrent neural network (RNN). The RUL prediction model based on CNN can extract local and global features of original data, and realize the fusion of features through multi-layer convolution and pooling operation. Yang et al. proposed a method based on a double-CNN model architecture, which uses the first CNN to identify early fault points and the second CNN as a regression model to predict RUL. However, the size of the convolution kernel limits the ability to extract long-term correlation information in the time series. The RUL prediction model based on RNN can learn features in the time scale and use internal states to store historical information. However, the method based on RNN ignores the relationship between data variables. In addition, some scholars combine CNN and RNN to form a hybrid prediction model. Ma et al. designed a convolution-based long short-term memory (CLSTM) network to predict the RUL of rotating machinery by mining vibration data.

[0004] However, in the above RUL prediction model, the potential spatial interaction relationship between sensors is ignored, and some scholars try to obtain the spatial correlation. Liu et al. introduced an attention mechanism to adaptively weight the features of each time point of the sensor, improving the performance of the prediction. Zhang et al. stacked sensors and then used 2D-CNN to capture spatial dependencies. Song et al. designed an RUL prediction model based on a distributed attention mechanism to weight different sensors and time steps. The above DL models have achieved good results in processing Euclidean data, but many data in practical application scenarios are generated from non-Euclidean space. Therefore, when facing more complex industrial equipment, the existing technology may not be able to accurately predict RUL, resulting in inaccurate prediction of the degradation trend of the service life of the scraper conveyor. SUMMARY

[0005] The application provides a scraper conveyor residual service life prediction method, medium and equipment to solve the above problems in the prior art, i.e., how to accurately predict the life degradation trend of the scraper conveyor. The application provides a scraper conveyor residual service life prediction method, which comprises the following steps:

[0006] Different sensor data of the acquired scraper conveyor are used to construct multiple original graph data; multi-scale convolution operation is used to extract features from the original graph data to obtain feature maps of different scales; and the feature maps of different scales are fused to obtain fused graph data.

[0007] By acquiring the spatial installation layout of the sensor, a prior adjacency matrix is determined; based on the fused graph data, an initial adjacency matrix is constructed by using a KNN algorithm, and the initial adjacency matrix is reconstructed by using a graph auto-encoder to obtain a similar adjacency matrix; and the prior adjacency matrix and the similar adjacency matrix are fused by using an adaptive mechanism to obtain an adaptive adjacency matrix.

[0008] The graph data and the adaptive adjacency matrix are fused by using a graph attention mechanism to obtain updated node spatial features; the updated node spatial features are weighted by using a time attention mechanism to determine global time features; and the global time features are fused with the original time sequence to obtain new time sequence features.

[0009] The fused new time sequence features are encoded by using a bidirectional gated recurrent unit to obtain fused spatiotemporal features; and the fused spatiotemporal features are input into a full connection layer to determine the final residual service life, thereby predicting the life of the scraper conveyor.

[0010] Optionally, the different sensor data of the acquired scraper conveyor are used to construct multiple original graph data, and the method specifically comprises the following steps.

[0011] The acquired different sensor data of the scraper conveyor are normalized, and the normalized data is obtained by using the following formula:

[0012]

[0013] wherein x i,j is the i-th original data of the j-th sensor, is the normalized data, and are the minimum value and the maximum value of the original data of the j-th sensor, respectively.

[0014] The normalized data is divided by using a sliding window technology to obtain the original graph data of the sensor.

[0015] Optionally, the prior adjacency matrix is determined by acquiring the spatial installation layout of the sensor, the initial adjacency matrix is constructed based on the fused graph data by using a KNN algorithm, and the dynamic similar adjacency matrix is obtained by reconstructing the initial adjacency matrix through a graph auto-encoder, and the method specifically comprises the following steps:

[0016] The prior adjacency matrix is determined by using the following formula:

[0017]

[0018] Wherein, A' is the prior adjacency matrix, indicates that the node X i and the node X j exist interaction;

[0019] Based on the fused graph data, the distance between each node and all nodes is determined by using a KNN algorithm, the nearest k nodes are selected, and the neighbor node set N i is determined, and the initial adjacency matrix is obtained by using the following formula:

[0020]

[0021] Wherein, A t indicates whether the node X i in the tth graph data is adjacent to the node X n ;

[0022] The initial adjacency matrix is reconstructed by using the following formula to obtain the similar adjacency matrix:

[0023] Z=GCN(h,A t ),

[0024] A t ″=σ(ZZ T ),

[0025] Wherein, GCN(·) represents the update process of GCN, h is the node feature, sigma is the sigmoid activation function, Z is the embedding vector of the node, and A″ t is the tth similar adjacency matrix.

[0026] Optionally, the graph data and the adjacency matrix are fused by using a graph attention mechanism to obtain the updated spatial node feature, and the method specifically comprises the following steps:

[0027] The graph data and the adjacency matrix are fused by using a graph attention mechanism, the similarity coefficient between adjacent nodes is obtained through a nonlinear activation function LeakyReLU, and the normalized attention coefficient is obtained after the Softmax function is activated, and the updated node spatial feature is obtained through the multi-head attention attention weight and the output vector.

[0028] Optionally, the non-linear activation function LeakyReLU obtains the similarity coefficient between adjacent nodes, and then obtains the normalized attention coefficient after the Softmax function activation, specifically including:

[0029] The similarity coefficient between adjacent nodes and the normalized attention coefficient are obtained by the following formula:

[0030]

[0031]

[0032] wherein, is the feature of the i-th node, is the feature of the adjacent node, F is the original feature dimension of each node, e ij is the similarity coefficient between adjacent nodes, a ij is the normalized attention coefficient, a∈R 2F′ is a learnable weight vector, W∈R F×F′ is a learnable shared weight matrix, and F' represents the feature dimension of the node output.

[0033] Optionally, the updated spatial node features are weighted by using the time attention mechanism to determine the global time feature, and the global time feature is fused with the original time sequence to obtain new time sequence features, specifically including:

[0034] The updated spatial node features are weighted by using the time attention mechanism, and the original weight is determined by the following formula:

[0035] ω t =o T tanh(h′ t ),

[0036] The normalized weight is obtained by the following formula:

[0037]

[0038] The global time feature is determined according to the normalized weight, and the specific formula is as follows:

[0039]

[0040] The global time feature is fused with the original time sequence to obtain new time sequence features μ t :

[0041] μ t =[h′ t ||S],

[0042] Where h′ t is the updated spatial node feature, o∈R N×F′ is a learnable vector, F′ is the feature dimension of the node output, ω t is the original weight, θ t is the normalized weight, S∈R T It is the representation of global time features.

[0043] Optionally, the bidirectional gated recurrent unit is used to encode the fused new time series features to obtain fused spatiotemporal features, specifically including:

[0044] The forward hidden state and the reverse hidden state are obtained using the following formula:

[0045]

[0046]

[0047] The forward hidden state and the reverse hidden state are concatenated using the following formula to obtain the fused spatiotemporal features:

[0048]

[0049] Among them, μ t is the new time series feature after fusion, and Represent the forward and reverse hidden states at time t, Y t is the output after fusing the bidirectional hidden values, GRU(·) represents the GRU update process, u t , v t , b t are the corresponding learnable parameters.

[0050] Optionally, the sensor data specifically includes:

[0051] Material level sensor data, zero speed sensor data and speed sensor data for speed regulation.

[0052] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method for predicting the remaining service life of a scraper conveyor is implemented.

[0053] The present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for predicting the remaining service life of a scraper conveyor is implemented.

[0054] Compared with the prior art, the application has the following beneficial effects: the application provides a residual useful life prediction method of a scraper conveyor, which mines hidden features of original data at multiple scales by using 1D convolution kernels of different sizes, and more comprehensive features provide support for capturing the correlation between sensors; a new correlation graph construction method is designed to realize the combination of prior static graph structure and dynamic similar graph structure, wherein a new dynamic similar graph structure construction method KNN-GAE is proposed, the graph structure is dynamically adjusted based on the degradation characteristics of data, an adaptive fusion mechanism is proposed to adaptively update the spatial coupling relationship between sensors from two aspects; a new RUL prediction model based on GAT and BiGRU is proposed, the spatial features between sensors are aggregated by GAT, the local and global feature fusion between sequences is realized by using the attention mechanism, the time dependence between sequences is captured by BiGRU, and the spatio-temporal features after fusion improve the RUL prediction performance. BRIEF DESCRIPTION OF DRAWINGS

[0055] The drawings incorporated into the specification and forming part of the specification, show embodiments consistent with the application, and together with the specification serve to explain the principles of the application.

[0056] Figure 1 A residual useful life prediction method of a scraper conveyor provided by an embodiment of the application is provided.

[0057] Figure 2 A sliding window graph data graph provided by the embodiment of the application is provided.

[0058] Figure 3 A multi-scale convolution graph provided by the embodiment of the application is provided.

[0059] Figure 4 A dynamic similar graph provided by the embodiment of the application is provided.

[0060] Figure 5 A graph attention mechanism network graph provided by the embodiment of the application is provided.

[0061] Figure 6 A sequence local and global feature fusion graph provided by the embodiment of the application is provided.

[0062] Figure 7 A computer device schematic diagram of a residual useful life prediction method of a scraper conveyor provided by the embodiment of the application is provided. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0064] During the operation of the scraper conveyor, multiple sensors are needed to monitor the operating state of the scraper conveyor, and there may be potential dependency between sensors with similar positions.

[0065] However, the prior art cannot fully exploit the spatial interaction relationship between sensors. The correlation graph constructed using prior knowledge or feature similarity ignores the dynamic evolution relationship between sensors over time, and a single-angle correlation graph is not enough to fully reflect the spatial interaction relationship between sensors, limiting their RUL prediction performance, resulting in inaccurate prediction of the life degradation trend of the scraper conveyor.

[0066] To solve the above problems, the present application provides a scraper conveyor remaining useful life prediction method, which fully utilizes the spatial structure characteristics of the scraper conveyor and the dynamic interaction relationship between sensors during the degradation process, fully captures the dynamic coupling relationship between sensors, and thus can accurately predict the life degradation trend of the scraper conveyor.

[0067] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below in combination with the drawings.

[0068] Embodiment 1

[0069] Figure 1 is a whole framework diagram of a scraper conveyor remaining useful life prediction method provided by the embodiment of the present application, which maps the sensors as nodes in the graph, contains a section of feature data, and each edge represents whether there is interaction between the sensors. The set of edges is used as an adjacency matrix. As shown in Figure 1 The scraper conveyor remaining useful life prediction method provided by the embodiment of the present application includes:

[0070] S101: Constructing multiple original graph data for the acquired different sensor data of the scraper conveyor.

[0071] Optionally, the acquired different sensor data of the scraper conveyor is normalized, and the normalized data is obtained by using the following formula:

[0072]

[0073] wherein x i,j is the i-th raw data of the j-th sensor, is the normalized data, and are the minimum and maximum values of the j-th sensor raw data, respectively.

[0074] Exemplarily, the non-sensor data of the plate conveyor can include, for example, the material level sensor data, the zero speed sensor data and the speed sensor data for speed regulation.

[0075] Exemplarily, as shown in Figure 2 , the normalized data can be divided by using a sliding window technology to obtain the raw graph data of the sensor.

[0076] S102: A multi-scale convolution operation is performed on the raw graph data to extract features, different scale feature maps are obtained, and the different scale feature maps are fused to obtain fused graph data.

[0077] Exemplarily, a one-dimensional convolution kernel of different sizes is used to perform a plurality of convolution operations on the input graph data, and then a 1x1 convolution kernel is used to fuse the features in different channel directions to form a more complete feature representation, and the convolution process is as shown in Figure 3 .

[0078] In this step, the convolution layer is used to extract features at different scales, which are used as the initial features of the nodes, in order to obtain information at different levels.

[0079] S103: By obtaining the spatial installation layout of the sensor, a prior adjacency matrix is determined, based on the fused graph data, an initial adjacency matrix is constructed by using a KNN algorithm, and the initial adjacency matrix is reconstructed by using a graph auto-encoder to obtain a similar adjacency matrix, and the prior adjacency matrix and the similar adjacency matrix are fused by using an adaptive mechanism to obtain an adaptive adjacency matrix.

[0080] Exemplarily, sensors that are physically adjacent or functionally similar can be assigned a higher interaction probability. For example, the transition groove and the middle groove, and the machine head frame are physically connected, so it is considered that the sensors monitoring them have an interaction relationship; at the same time, when the speed of the scraper chain increases, it will cause the temperature of the plate in the middle groove to rise, so it is considered that the sensors monitoring them also have an interaction relationship.

[0081] Optionally, the prior adjacency matrix is determined by using the following formula:

[0082]

[0083] Wherein, A' is the prior adjacency matrix, The node X i and the node X j There is an interaction.

[0084] Based on the fused graph data, as shown in the formula (1), the KNN algorithm is used to determine the distance between each node and all nodes, and the distance vector D(i, n) ∈ R N is obtained, the nearest k nodes are selected, the neighbor node set N i is determined, and the initial adjacency matrix A t ∈ R N×N is obtained. Figure 4

[0085]

[0086] Wherein, A t represents whether the node X i in the tth graph data is adjacent to the node X n .

[0087] The initial adjacency matrix is reconstructed by using the graph auto-encoder to obtain the similar adjacency matrix, and the specific formula is as follows:

[0088] Z=GCN(h,A t )。

[0089] A t ″=σ(ZZ T )。

[0090] Wherein, GCN(·) represents the update process of GCN, h is the node feature, σ is the sigmoid activation function, Z is the embedding vector of the node, and A″ t is the tth similar adjacency matrix.

[0091] The gating mechanism process is as follows:

[0092] gate=σ(A′W g1 +A″ t W g2 +b g )。

[0093] A″′ t =gate#A′+(1-gate)A″ t 。

[0094] Wherein, A″′ t is the adaptive fused graph structure, W g1 , W g2 ∈ R N×N is a learnable weight, and b g ∈ R N ​is a learnable bias.

[0095] The gating mechanism can adaptively adjust the weight of the graph structure according to different inputs, so that more rich interactions and dependencies between features can be obtained.

[0096] S104: The updated node space features are obtained by fusing the graph data and the adaptive adjacency matrix through the graph attention mechanism, the global time features are determined by weighting the updated node space features through the time attention mechanism, and the new time sequence features are obtained by fusing the global time features and the original time sequence.

[0097] Optionally, the graph data and the adjacency matrix are fused through the graph attention mechanism, the similarity coefficient between adjacent nodes is obtained through the nonlinear activation function LeakyReLU, and the normalized attention coefficient is obtained after the Softmax function activation, and the updated space node features are obtained through the multi-head attention attention weight and the output vector.

[0098] Illustratively, the specific formula of the similarity coefficient between adjacent nodes is as follows:

[0099]

[0100] The specific formula of the normalized attention coefficient is as follows:

[0101]

[0102] wherein, is the feature of the i th node, is the feature of the adjacent node, F is the original feature dimension of each node, e ij is the similarity coefficient between adjacent nodes, as shown in (a) of Figure 5 , a ij is the normalized attention coefficient, a∈R 2F′ is a learnable weight vector, W∈R F×F′ is a learnable shared weight matrix, and F' represents the feature dimension of the node output.

[0103] The sparse attention coefficient is obtained after random dropping the attention coefficient, and the specific calculation formula is as follows:

[0104] a i ′ j =Droupout(a ij ).

[0105] The feature of the node i after aggregating the neighborhood node information is represented as:

[0106]

[0107] As Figure 5 shown in (b), the output multi-channel feature map is averaged to obtain new node features using the following formula:

[0108]

[0109]

[0110] where K is the number of attention mechanisms introduced, and || represents the vector concatenation operation.

[0111] After GAT fusion, a new set of node features is generated is the updated new feature of the node, and the potential spatial dependency between sensors is fused into the feature h'.

[0112] Optionally, as Figure 6 shown in (c), a time attention mechanism is used to weight the updated node spatial feature sequence, and the original weight is determined using the following formula:

[0113] ω t = o T tanh(h′ t ).

[0114] The normalized weight is obtained using the following formula:

[0115]

[0116] The global time feature is determined according to the normalized weight, and the specific formula is as follows:

[0117] S = [θ1||θ2…||θ T ].

[0118] The global time feature is fused with the original time sequence to obtain a new time sequence feature μ t :

[0119] μ t = [h′ t ||S].

[0120] where h′ t is the updated node spatial feature, o∈R N×F′ is a learnable vector, F′ is the feature dimension of the node output, ω t is the original weight, θ t is the normalized weight, S∈R T is the representation of the global time feature.

[0121] S105: encode the fused new time series feature by using a bidirectional gated recurrent unit, obtain a fused spatio-temporal feature, and input the fused spatio-temporal feature into a fully connected layer to determine a final remaining useful life and perform life prediction of the scraper conveyor.

[0122] Optionally, the forward hidden state and the reverse hidden state are obtained by using the following formula:

[0123]

[0124]

[0125] The forward hidden state and the reverse hidden state are spliced by using the following formula to obtain the fused spatio-temporal feature:

[0126]

[0127] wherein μ t is the fused new time series feature, and represent the forward and reverse hidden states at t, respectively, Y t is an output after fusing the bidirectional hidden values, GRU(·) represents a GRU updating process, u t , v t , and b t are corresponding learnable parameters.

[0128] In this step, the fused new time series feature is encoded by using a bidirectional gated recurrent unit, so as to better capture the dynamic changes and long-term dependencies of the time series. Finally, the outputs of the integrated spatial and temporal features are transmitted to a fully connected layer to realize the final remaining life prediction task.

[0129] The embodiment provides a method for predicting the remaining useful life of a scraper conveyor. The preprocessed graph data is taken as the input of a model by using a convolutional layer, hidden features at different scales can be extracted and taken as initial features of nodes, then a dynamic dependency relationship between sensors is captured by using device structure characteristics and the features after convolution by using a graph structure learning layer, so that the correlation of the graph data is comprehensively fused, spatial features of aggregated nodes are updated by using a GAT network through a spatio-temporal fusion layer, on this basis, a time attention module and a BiGRU are used to extract dependencies between sequences, finally, the fused spatio-temporal feature is input into a fully connected layer to obtain a final RUL, so that the life degradation trend of the scraper conveyor can be accurately predicted.

[0130] Embodiment 2

[0131] In this example, the effectiveness of the method is verified by selecting full life cycle degradation data of four different scraper conveyors in a coal mine.

[0132] Optionally, the performance of the model is evaluated by using the root mean square error (RMSE) and the score function (Score), and the specific formula is as follows:

[0133]

[0134]

[0135] wherein, and y i are the predicted and true RUL, respectively, and N is the number of samples in the test set. The RMSE function measures the deviation between the predicted value and the true value, and early and late predictions are treated equally.

[0136] In contrast, the Score function has a greater penalty for late predictions, avoiding the more serious consequences of lagging predictions. The lower the value of the two indicators, the better the prediction performance of the model. In order to eliminate the influence of randomness, the model is tested ten times on each scraper conveyor data, and the average value is taken as the final evaluation result.

[0137] As shown in Table 1, the scores of the scraper conveyor of the method used in the application and other methods are compared.

[0138] Table 1 Comparison of scores of different methods for scraper conveyor:

[0139]

[0140] As can be seen from Table 1, the machine learning-based methods ELM, RF, and DBN perform poorly, which is due to the insufficient nonlinear expression capability of traditional machine learning methods when dealing with high-dimensional and massive data. The deep learning-based methods DCNN and MS-DCNN extract local features of sensor data, but MS-DCNN enhances the prediction performance by fusing multi-scale local features. However, the CNN-based model ignores the time dependence of the data. The deep learning-based methods Bi-LSTM and ATT-LSTM use recurrent networks to extract the time correlation of the data, and ATT-LSTM improves the prediction performance to some extent by introducing an attention mechanism. However, the RNN-based model ignores the relationship between variables. The hybrid CNN-RNN method BLCNN is generally better than single-type methods, and the deep learning-based method 2D-CNN and AGCNN that considers the correlation of sensors generally perform better; the method based on transformer Transformer is powerful in capturing global dependencies in data, but may ignore the order of time series data.

[0141] The GNN-based method integrates time and space features and achieves more significant results. In terms of graph structure construction, STFA relies on prior structural knowledge of devices for construction, but it ignores the dynamic relationship between sensors in degradation. GGCN uses cosine similarity between sensor features to construct, but does not fully utilize prior knowledge of devices. CDSG utilizes Hadamard product to construct prior graph structure and designed recursive graph structure, but it is difficult to achieve perfect fusion of the two angles of graph structure through Hadamard product. In view of the above problems, the application proposes a new adaptive fusion method to construct a graph structure, that is, more weights are added to important graph structures, which makes up for the one-sidedness of a single correlation graph structure and more richly reflects the interaction between sensors. The experimental results verify the prediction performance of the method.

[0142] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the scraper conveyor residual service life prediction method provided above.

[0143] The application further provides Figure 7 The structure diagram of the computer device is shown in the figure, and the computer device includes a processor, an internal bus, a network interface, a memory and a non-volatile memory. Figure 7 As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory and a non-volatile memory, and of course can also include other hardware required by the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to realize the scraper conveyor residual service life prediction method provided in the above embodiment.

[0144] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Wherein, any reference to memory, storage, database or other medium used in each embodiment of the application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

[0145] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features described in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered to be within the scope of the present application.

Claims

1. A method for predicting the remaining service life of a scraper conveyor, characterized in that: include: Based on the different sensor data of the scraper conveyor, multiple raw graph data are constructed. Multi-scale convolution operations are used to extract features from the raw graph data to obtain feature graphs of different scales. The feature graphs of different scales are fused to obtain fused graph data. By obtaining the spatial installation layout of sensors, the prior adjacency matrix is ​​determined. Based on the fused graph data, the initial adjacency matrix is ​​constructed using the KNN algorithm. The initial adjacency matrix is ​​reconstructed using a graph autoencoder to obtain a similar adjacency matrix. The prior adjacency matrix and the similar adjacency matrix are fused through an adaptive mechanism to obtain an adaptive adjacency matrix. The method obtains the spatial installation layout of the sensors, determines the prior adjacency matrix, constructs the initial adjacency matrix based on the fused graph data by using the KNN algorithm, and reconstructs the initial adjacency matrix by using the graph autoencoder to obtain the similarity adjacency matrix, specifically including: The following formula is used to determine the prior adjacency matrix: ， in, is the prior adjacency matrix, Representation node and nodes There is interaction; Based on the fused graph data, the KNN algorithm is used to determine the distance between each node and all nodes, and the closest nodes, determine the set of neighbor nodes , the initial adjacency matrix is ​​obtained using the following formula: ， in, Indicates the Nodes in the graph Is it related to the node adjacent; The following formula is used to reconstruct the initial adjacency matrix to obtain the similar adjacency matrix: ， ， in, Represents the update process of GCN, is the node feature, is the sigmoid activation function, is the embedding vector of the node, It is Similarity adjacency matrix; By adopting the graph attention mechanism to fuse the graph data and the adaptive adjacency matrix, the updated node spatial features are obtained. The updated node spatial features are weighted by the time attention mechanism to determine the global time features. The global time features are spliced ​​and fused with the original time series to obtain the new time series features. The time attention mechanism is used to weight the updated spatial node features to determine the global time features, and the global time features are spliced ​​and fused with the original time series to obtain new time series features, specifically including: The time attention mechanism is used to weight the updated node spatial features, and the original weight is determined by the following formula: ， The normalized weight is obtained using the following formula: ， The global time feature is determined based on the normalized weights. The specific formula is as follows: ， The global time feature is fused with the original time series, and the new time series feature is obtained using the following formula : ， in, is the updated spatial node feature, is a learnable vector, is the feature dimension of the node output, is the original weight, is the normalized weight, is the representation of global time features; A bidirectional gated recurrent unit is used to encode the fused new time series features to obtain the fused spatiotemporal features. The fused spatiotemporal features are input into the fully connected layer to determine the final remaining service life and predict the life of the scraper conveyor.

2. The method for predicting the remaining service life of a scraper conveyor according to claim 1, wherein: The method of constructing a plurality of original graph data based on the different sensor data of the scraper conveyor obtained specifically includes: The obtained scraper conveyor sensor data are normalized and the normalized data are obtained using the following formula: ， in, It is The first sensor The original data, is the normalized data, and They are The minimum and maximum values ​​of the raw data of each sensor; The sliding window technology is used to divide the normalized data to obtain the original image data of the sensor.

3. The method for predicting the remaining service life of a scraper conveyor according to claim 1, wherein: The updated node spatial features are obtained by fusing the graph data and the adjacency matrix using the graph attention mechanism, which specifically includes: By adopting the graph attention mechanism to fuse the graph data and the adjacency matrix, the nonlinear activation function Get the similarity coefficient between adjacent nodes, and then After the function is activated, the normalized attention coefficient is obtained, and the updated node space feature is obtained through the attention weight and output vector.

4. The method for predicting the remaining service life of a scraper conveyor according to claim 3, wherein: The nonlinear activation function Get the similarity coefficient between adjacent nodes, and then After the function is activated, the normalized attention coefficient is obtained, including: The similarity coefficient between adjacent nodes and the normalized attention coefficient are obtained using the following formula: ， ， in, It is The characteristics of the nodes, is the feature of its adjacent nodes, , , is the original feature dimension of each node, is the similarity coefficient between adjacent nodes, is the normalized attention coefficient, is a learnable weight vector, is a learnable shared weight matrix, Represents the feature dimension of the node output.

5. The method for predicting the remaining service life of a scraper conveyor according to claim 1, wherein: The bidirectional gated recurrent unit is used to encode the fused new time series features to obtain fused spatiotemporal features, which specifically include: The forward hidden state and the reverse hidden state are obtained using the following formula: ， ， The forward hidden state and the reverse hidden state are concatenated using the following formula to obtain the fused spatiotemporal features: ， in, is the new time series feature after fusion, and Respectively The hidden state of forward and reverse at all times, is the output after fusing the bidirectional hidden values, Represents the GRU update process, , , are the corresponding learnable parameters.

6. The method for predicting the remaining service life of a scraper conveyor according to claim 1, wherein: The sensor data specifically includes: Material level sensor data, zero speed sensor data and speed sensor data for speed regulation.

7. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by the processor, the method for predicting the remaining service life of a scraper conveyor according to any one of claims 1 to 6 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting the remaining service life of a scraper conveyor as claimed in any one of claims 1 to 6 is implemented.

Images