Bearing device residual life prediction model and method based on tcn and decay map

Abstract

The application discloses a bearing device residual life prediction model and method based on TCN and decay graph, and belongs to the technical field of industrial equipment state monitoring and fault prediction and diagnosis. The model comprises sequentially connected input processing units, local time period coding modules, TCN modules based on differential graphs, decay graph modules and prediction output modules. The input processing units receive information collected by sensors and splice input data. The local time period coding modules are used for time period division and linear coding and output local time period set coding results. Decay characteristics between local time periods are output through the TCN modules based on differential graphs. Adjacency matrices of final time periods are output through the decay graph modules. Non-negative integer bearing residual life is output through the prediction output modules. The bearing device residual life prediction model and method based on TCN and decay graph are used to form an integrated dynamic modeling framework from signal input to life output, adapt to working condition changes and strengthen early decay identification.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and fault prediction and diagnosis technology, and in particular to a bearing device remaining life prediction model and method based on TCN and decay diagram. Background Technology

[0002] Bearings are core transmission components in industrial equipment, and their operating status directly affects the equipment's efficiency and safety. Regarding the prediction of the remaining life of bearings, various deep learning-based prediction models have been proposed in related technical fields. These models generally extract the decay characteristics of individual sensor signals through time-series modules, then use methods such as graph neural networks (GNNs) to model the joint decay relationships between multiple sensors, and finally complete the remaining life prediction based on the extracted features. Among them, HAD-GNN is one of the most representative models in this direction. However, existing deep learning models, represented by HAD-GNN, still have technical shortcomings in practical applications, such as insufficient discrimination of decay features and the inability of the graph structure to adapt to dynamic correlation changes. This results in prediction accuracy that cannot meet the high-precision requirements of industrial scenarios, thus limiting the application effect of existing technologies in the prediction of the remaining life of bearings. Summary of the Invention

[0003] The purpose of this invention is to provide a bearing device remaining life prediction model and method based on TCN and decay diagram, thereby solving the above-mentioned technical problems.

[0004] To achieve the above objectives, the present invention provides a bearing device remaining lifetime prediction model and method based on TCN and decay map, comprising an input processing unit, a local time period encoding module, a TCN module based on difference map, a decay map module and a prediction output module connected in sequence.

[0005] The input processing unit is a data preprocessing unit used to receive the sensor signal sequence of bearing-type devices. Bearing usage time information and will and Concatenate as input data ;

[0006] The local time-segment encoding module is the time-segment feature extraction unit of the model, used for... Perform time segmentation and linear coding, and output the coding results of the local time segment set. ;

[0007] The TCN module based on the difference graph serves as the decay feature extraction unit for the model, used for... Process the data and output the decay characteristics within a local time period. ;

[0008] The decay diagram module is a dynamic correlation modeling unit used for... Process the data and output the adjacency matrix for the final time period. ;

[0009] The prediction output module is the model's lifetime calculation unit, used for... Perform flattening and linear calculations to output the remaining bearing life in non-negative integer form. .

[0010] Preferably, the time segmentation of the local time segmentation encoding module is achieved through a sliding window, and the window length of the sliding window is [missing information]. Step size is and After segmentation, local time-segment sets are obtained. Then, for local time sets Perform linear encoding output The formulas for time segmentation and linear coding are as follows:

[0011] ;

[0012] in, , , , The number of local time periods. Let be the set of real numbers. The length of the signal sequence acquired by the sensor. For the number of sensors, For learnable weight matrix, For learnable bias vectors, For encoding dimensions, This is a time segmentation function based on a sliding window.

[0013] Preferably, the TCN module based on the difference graph consists of It consists of stacked processing layers, each containing a gating mechanism unit, a region-differential TCN unit, a GCN unit, and a differential pooling unit.

[0014] Preferably, in the first In each processing layer, the gating mechanism unit controls... Layer normalization is performed to obtain Then calculate the first-order difference sequence. The splicing fragments are added using the Padding operation. The starting position, The formula is:

[0015] ;

[0016] in, For learning parameters, This is a first-order difference operation. For copy and splice operations, For sequence padding operations;

[0017] Calling TCN branches with non-shared parameters in region-difference-based TCN units and ,extract Original temporal characteristics and Differential time series characteristics ,and Its formula is:

[0018] ;

[0019] Then process using the tanh activation function sigmoid activation function processing The temporal features of the fused differential information are obtained by element-wise multiplication. , The formula is:

[0020]

[0021] in, For element-wise multiplication, The tanh activation function is used. It is the sigmoid activation function.

[0022] Preferably, calculation is performed using differential pooling units. Index of the maximum absolute value over time ,index The formula is:

[0023] ;

[0024] Based on index from The period of drastic fluctuation was selected as , The previous period was , and The formula is:

[0025] ;

[0026] in, Indicates index-based Select sequence The element at the corresponding position in the middle, and , ;

[0027] Will and The feature matrix is ​​concatenated and then computed using an attention mechanism. Key matrix Value matrix Its formula is:

[0028] ;

[0029] in, , , All are learnable weight matrices. , , All are learnable bias vectors;

[0030] Finally, the time-transient decay characteristics are calculated. ,and , The formula is:

[0031] ;

[0032] in, This is the softmax activation function.

[0033] Preferably, the GCN unit is based on a learnable adjacency matrix. right Perform graph convolution calculation and output. Then Replication along the time dimension This time received and will and The features are concatenated, and the concatenated result is then processed through a linear layer to output the final result. , The formula is:

[0034] ;

[0035] in, The weight matrix is ​​a learnable weight matrix;

[0036] No. The output of each processing layer is , The formula is:

[0037] ;

[0038] right The outputs of each processing layer are aggregated to obtain the decay characteristics. The formula for the polymerization process is:

[0039] ;

[0040] in, This is the hidden layer state. For learnable weight matrix, For learnable bias vectors, The output is the decay characteristic result. For learnable weight matrix, It is a learnable bias vector.

[0041] Preferably, the decay graph module consists of multiple circularly connected G-RNN graph units, each of which integrates gating and self-attention mechanisms. The Middle Characteristics of each time period First, calculate the Query matrix of the self-attention mechanism. With Key Matrix Its formula is:

[0042] ;

[0043] in, , For learnable weight matrix, , A learnable bias vector;

[0044] Then to Layer normalization is performed, and then a similarity map is obtained by applying the sigmoid activation function. ,and , The formula is:

[0045] ;

[0046] Next, calculate the gating parameters using the following formula:

[0047] ;

[0048] in, For learnable weight vectors, For learnable bias vectors, For input gate, For the Gate of Oblivion It is a vector consisting entirely of 1s;

[0049] Finally, update the adjacency matrix. ,and Its formula is:

[0050] ;

[0051] in, For the first The adjacency matrix for each time period.

[0052] Preferably, the initial adjacency matrix of the first G-RNN graph unit in the decay graph module Set as the identity matrix or a randomly initialized learnable matrix. After processing the first... Characteristics of each time period Finally, the adjacency matrix output by the last G-RNN graph unit is .

[0053] Preferably, the prediction output module is for Flattening the vector yields the vector. Then, the remaining lifetime is obtained through linear calculation and absolute value operation. , The formula is:

[0054] ;

[0055] in, The flattening operation is used to convert a matrix into a one-dimensional vector. For learnable weight matrix, For learnable weight vectors, For learnable bias vectors, for The transpose of .

[0056] Preferably, it includes the following steps:

[0057] S1. Acquire bearing sensor signal sequence Information on bearing usage time The input data is obtained by concatenating the data according to the feature dimensions through the input processing unit. ;

[0058] S2, will The local time-segmentation encoding module of the input prediction model is used to obtain the local time-segmentation set through time segmentation. Through linear encoding Perform a linear transformation to obtain the local time-segment encoding result. ;

[0059] S3, will The input is a TCN module based on a difference graph. The module sequentially performs layer normalization, first-order difference calculation, TCN branch feature extraction, difference pooling, and aggregation processing, outputting decay features. ;

[0060] S4, will The input decay graph module processes the data through its G-RNN graph units, and outputs the adjacency matrix for the final time interval. ;

[0061] S5, will The input prediction output module processes the data through flattening, linear calculation, and absolute value operations, outputting the bearing's remaining life in non-negative integer form. .

[0062] Therefore, the bearing device remaining life prediction model and method based on TCN and decay diagrams adopted in this invention have the following beneficial effects:

[0063] 1. The TCN module based on the difference map incorporates the idea of ​​regional difference into the TCN combined with gating, which can effectively distinguish the characteristics of different sensor states based on the difference information between time periods.

[0064] 2. The decay graph module uses the graph structure to capture the correlation between sensors, and can make dynamic adjustments based on the current state of the sensors.

[0065] 3. By encoding the sensor signal sequence in local time periods, the model focuses on the signal differences between local time periods, and subsequent modules consider the differences between local time periods when extracting decay features, thereby achieving more accurate remaining lifetime prediction.

[0066] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0067] Figure 1 This invention relates to a deep learning network structure. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0069] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0070] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0071] like Figure 1 As shown, this invention provides a bearing device remaining lifetime prediction model and method based on TCN and decay map, including an input processing unit, a local time period encoding module, a TCN module based on a difference map, a decay map module, and a prediction output module connected in sequence; wherein, the input processing unit is a data preprocessing unit, used to receive the sensor acquisition signal sequence of bearing devices. Bearing usage time information and will and Concatenate as input data The local time-segment encoding module is the time-segment feature extraction unit of the model, used for... Perform time segmentation and linear coding, and output the coding results of the local time segment set. This allows subsequent modules to focus on signal differences between local time periods; the TCN module based on the difference map serves as the model's decay feature extraction unit, used for... Process the data and output the decay characteristics within a local time period. This module can distinguish between the sensor's health state and decay state based on inter-time period differences, without requiring additional signal denoising preprocessing; the decay map module is a dynamic correlation modeling unit used for... Process the data and output the adjacency matrix for the final time period. This matrix records the overall decay process of the correlation between sensors; the prediction output module is the lifetime calculation unit of the model, used to... Perform flattening and linear calculations to output the remaining bearing life in non-negative integer form. .

[0072] Specifically, the time segmentation of the local time segmentation module is implemented through a sliding window, with a window length of [missing information]. Step size is and After segmentation, local time-segment sets are obtained. Then, for local time sets Perform linear encoding output The formulas for time segmentation and linear coding are as follows:

[0073] ;

[0074] in, , , , The number of local time periods. Let be the set of real numbers. The length of the signal sequence acquired by the sensor. For the number of sensors, For learnable weight matrix, For learnable bias vectors, For encoding dimensions, This is a time-segmentation function based on a sliding window, used to aggregate time-step information in local areas and reduce interference from single-time-step signal fluctuations.

[0075] The TCN module based on difference graphs consists of It consists of stacked processing layers. The integers are positive. Each processing layer contains a gating mechanism unit, a region-difference-based TCN unit, a GCN unit, and a differential pooling unit, which can collaboratively achieve original temporal feature extraction, differential temporal feature enhancement, cross-temporal decay effect modeling, and global correlation fusion. In the... In each processing layer, are positive integers and Through the gating mechanism unit Layer normalization is performed to obtain Then calculate the first-order difference sequence. The splicing fragments are added using the Padding operation. The starting position, The formula is:

[0076] ;

[0077] in, For learning parameters, This is a first-order difference operation. For copy and splice operations, For sequence padding operations; call the TCN branch in the region-difference-based TCN unit where parameters are not shared. and ,extract Original temporal characteristics and Differential time series characteristics ,and Its formula is:

[0078] ;

[0079] Then process using the tanh activation function sigmoid activation function processing The temporal features of the fused differential information are obtained by element-wise multiplication. , The formula is:

[0080]

[0081] in, For element-wise multiplication, The tanh activation function is used. It is the sigmoid activation function.

[0082] Differential pooling units are used to model the effects of sensor degradation over time. The calculations performed using differential pooling units... Index of the maximum absolute value over time ,index The formula is:

[0083] ;

[0084] Based on index from The period of drastic fluctuation was selected as , The previous period was The selection method has cross-time characteristics and can capture key nodes of sensor decay. and The formula is:

[0085] ;

[0086] in, Indicates index-based Select sequence The element at the corresponding position in the middle, and , ;Will and The feature matrix is ​​concatenated and then computed using an attention mechanism. Key matrix Value matrix Its formula is:

[0087] ;

[0088] in, , , All are learnable weight matrices. , , All are learnable bias vectors; finally, the time-dependent decay features are calculated. ,and , The formula is:

[0089] ;

[0090] in, The activation function is softmax. GCN units are based on a learnable adjacency matrix. right Perform graph convolution calculation and output. Then Replication along the time dimension This time received and will and The features are concatenated, and the concatenated result is then processed through a linear layer to output the final result. , The formula is:

[0091] ;

[0092] in, The learnable weight matrix; the first The output of each processing layer is , The formula is:

[0093] ;

[0094] right The outputs of each processing layer are aggregated to obtain the decay characteristics. The formula for the polymerization process is:

[0095] ;

[0096] in, This is the hidden layer state. For learnable weight matrix, For learnable bias vectors, The output is the decay characteristic result. For learnable weight matrix, It is a learnable bias vector.

[0097] The decay graph module consists of multiple circularly connected G-RNN graph units. Each G-RNN graph unit is a graph processing unit that integrates gating and self-attention mechanisms. Its hidden state can dynamically track the evolution of the adjacency matrix. Each G-RNN graph unit integrates gating and self-attention mechanisms. The Middle Characteristics of each time period First, calculate the Query matrix of the self-attention mechanism. With Key Matrix Its formula is:

[0098] ;

[0099] in, , For learnable weight matrix, , For learnable bias vectors, then for Layer normalization is performed, and then a similarity map is obtained by applying the sigmoid activation function. ,and , The formula is:

[0100] ;

[0101] Next, calculate the gating parameters using the following formula:

[0102] ;

[0103] in, For learnable weight vectors, For learnable bias vectors, For input gate, For the Gate of Oblivion It is a vector of all 1s; finally, the adjacency matrix is ​​updated. ,and Its formula is:

[0104] ;

[0105] in, For the first The adjacency matrix for each time period. Additionally, the initial adjacency matrix of the first G-RNN graph unit in the decay graph module. Set as the identity matrix or a randomly initialized learnable matrix. After processing the first... Characteristics of each time period Finally, the adjacency matrix output by the last G-RNN graph unit is Then, the output prediction module is used to... Flattening the vector yields the vector. Then, the remaining lifetime is obtained through linear calculation and absolute value operation. , The formula is:

[0106] ;

[0107] in, The flattening operation is used to convert a matrix into a one-dimensional vector. For learnable weight matrix, For learnable weight vectors, For learnable bias vectors, for The transpose of .

[0108] The prediction model for the remaining life of bearing-type devices includes the following steps: S1, acquiring bearing sensor signal sequences. Information on bearing usage time The input data is obtained by concatenating the data according to the feature dimensions through the input processing unit. S2, will The local time-segmentation encoding module of the input prediction model is used to obtain the local time-segmentation set through time segmentation. Through linear encoding Perform a linear transformation to obtain the local time-segment encoding result. S3, will The input is a TCN module based on a difference graph. The module sequentially performs layer normalization, first-order difference calculation, TCN branch feature extraction, difference pooling, and aggregation processing, outputting decay features. S4, will The input decay graph module processes the data through its G-RNN graph units, and outputs the adjacency matrix for the final time interval. S5, will The input prediction output module processes the data through flattening, linear calculation, and absolute value operations, outputting the bearing's remaining life in non-negative integer form. .

[0109] To verify the remaining lifetime prediction performance of the PGNet model in classic industrial scenarios, the CMAPSS dataset was selected. PGNet was compared with mainstream baseline models such as LeNet, LSTM, and Transformer on the FD001, FD002, FD003, and FD004 subsets and their average values, using the root mean square error as the benchmark. and fractions The evaluation indicators and comparison results are shown in Table 1:

[0110] Table 1 Comparison of Remaining Lifetime Estimations between PGNet and Baseline Models

[0111]

[0112] Experimental results show that the PGNet model of this invention improves the prediction accuracy in the task of predicting the remaining life of bearing devices, and its performance is significantly better than that of mainstream baseline models.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A bearing device residual life prediction model based on TCN and decay map, characterized in that: The method comprises the following steps: The input processing unit is a data preprocessing unit configured to receive a sensor collected signal sequence of the bearing type device , bearing use length information , and concatenate and as input data ; The local time period encoding module is a time period feature extraction unit of the model, configured to perform time period segmentation and linear encoding on the time period feature sequence to output a local time period set encoding result ;​ The TCN module based on the difference map is a decay feature extraction unit of the model, used for processing the decay feature between the local time periods and outputting the decay feature between the local time periods The module is composed of a plurality of stacked processing layers, each processing layer comprising a gating mechanism unit, a TCN unit based on a region difference, a GCN unit, and a difference pooling unit. In the In each processing layer, the gating mechanism unit controls... Layer normalization is performed to obtain Then calculate the first-order difference sequence. The splicing fragments are added using the Padding operation. The starting position, The formula is: ； wherein, is a learning parameter, is a first difference operation, is a copy-paste operation, is a sequence padding operation; Calling a tcn unit based on regional difference parameter not sharing tcn branch and extracting the original timing features of and the differential timing features of , and the formula is: ； Reusing tanh activation function processing , sigmoid activation function processing , element multiplication to obtain fusion differential information of time sequence characteristics , The formula is: ； wherein, is an element-wise multiplication, is a tanh activation function, is a sigmoid activation function; The decay diagram module is a dynamic correlation modeling unit used for... Process the data and output the adjacency matrix for the final time period. This module consists of multiple circularly connected G-RNN graph units. Each G-RNN graph unit integrates gating and self-attention mechanisms. The Middle Characteristics of each time period First, calculate the Query matrix of the self-attention mechanism. With Key Matrix Its formula is: ; wherein, , is a learnable weight matrix, , is a learnable bias vector; Then the layer normalization processing is performed on , and the similarity graph is obtained through a sigmoid activation function , and , The formula is: ； The gating parameter is recalculated, and the formula is as follows: ； wherein, is a learnable weight vector, is a learnable bias vector, is an input gate, is a forget gate, is an all-ones vector; last update adjacency matrix , and The formula is: ； wherein is the adjacency matrix for the first time period; The prediction output module is a life calculation unit of the model, configured to perform flattening and linear calculation on the bearing remaining life in the non-negative integer form and output the bearing remaining life in the non-negative integer form .

2. The TCN and decay map based bearing component remaining useful life prediction model according to claim 1, wherein: The time interval division of the local time interval coding module is realized by a sliding window, the window length of the sliding window is , the step is , and , and the local time interval set is obtained after the division. Then, the local time interval set is linearly coded and outputted. The formula of the time interval division and the linear coding is: ; wherein, , , , is a number of local time segments, is a set of real numbers, is a length of a sensor collected signal sequence, is a number of sensors, is a learnable weight matrix, is a learnable bias vector, is an encoding dimension, is a sliding window based time segmenting function.

3. The TCN and decay map based bearing component remaining useful life prediction model according to claim 2, wherein: Calculate by difference pooling unit Maximum value index of absolute value in time dimension , index The formula is: ； Based on the index From The period of violent fluctuations is selected as , The previous period of , And The formula is: ； wherein represents the element at the position corresponding to the index selecting a sequence corresponding position in the sequence , ; Will and In the feature dimension splicing, the Query matrix is calculated by attention mechanism , Key matrix , Value matrix The formula is: ； wherein, , , are all learnable weight matrices, , , are all learnable bias vectors; Finally, the cross-time decay characteristics are calculated , and , The formula is: ； wherein, is a softmax activation function.

4. The TCN and decay map based bearing component remaining useful life prediction model according to claim 3, wherein: GCN units are based on learnable adjacency moments. right Perform graph convolution calculation and output. Then Replication along the time dimension This time received and will and The features are concatenated, and the concatenated result is then processed through a linear layer to output the final result. , The formula is: ； wherein, is a learnable weight matrix; No. The output of each processing layer is , The formula is: ； right The outputs of each processing layer are aggregated to obtain the decay characteristics. The formula for the polymerization process is: ； wherein, is a hidden layer state, is a learnable weight matrix, is a learnable bias vector, is an output decay feature result, is a learnable weight matrix, is a learnable bias vector.

5. The TCN and decay map based bearing component remaining useful life prediction model according to claim 4, wherein: initial adjacency matrix of the first G-RNN graph unit in the decay graph module learnable matrix initialized as identity matrix or randomly, when the features of the period are processed, the adjacency matrix output by the last G-RNN graph unit is .​ 6. The TCN and decay map based bearing component remaining useful life prediction model according to claim 5, wherein: The prediction output module is flattened to obtain a vector Then, the remaining life is obtained by linear calculation and absolute value operation. The formula is: ； wherein, is a flattening operation to convert the matrix into a one-dimensional vector, is a learnable weight matrix, is a learnable weight vector, is a learnable bias vector, is is a transposed vector.

7. The method of predicting the remaining life of a bearing device based on a TCN and a decay map according to any one of claims 1 to 6, characterized in that: The method comprises the following steps: S1, collect bearing sensor signal sequence With bearing use time length information , through the input processing unit, splice by feature dimension, get input data ; S2, the input prediction model local period encoding module, and through the period segmentation to get local period set , by linear coding to linear transformation, get local period set encoding results ; S3, will The input is a TCN module based on a difference graph. The module sequentially performs layer normalization, first-order difference calculation, TCN branch feature extraction, difference pooling, and aggregation processing, outputting decay features. ; S4, will The input decay graph module processes the data through its G-RNN graph units, and outputs the adjacency matrix for the final time interval. ; S5, will The input prediction output module processes the data through flattening, linear calculation, and absolute value operations, outputting the bearing's remaining life in non-negative integer form. .

Images