Method and device for predicting residual service life of equipment based on isomorphic space-time fusion and causal expansion convolution

Abstract

The invention discloses an equipment remaining service life prediction method and device based on isomorphic space-time fusion and causal expansion convolution. The method comprises the following steps: firstly, acquiring an original vibration signal of the whole life cycle of equipment, screening data with learning characteristics, and performing sliding time window segmentation and normalization preprocessing; the preprocessed sequential sequence is subjected to causal convolution processing and then input into a time channel and a space channel, the receptive field is expanded by using expansion convolution with the same expansion rate, and spatial-temporal feature cross fusion is realized through a gating activation unit and a weight cross fusion mechanism. And performing global information aggregation and key feature extraction by means of a residual connection and information aggregation module, and finally outputting the residual life of the equipment through linear regression. According to the method, the spatial-temporal feature extraction capability is improved, the long-term dependency relationship modeling effect is enhanced, and an efficient and accurate prediction means is provided for health management of industrial equipment.

Description

Technical Field

[0001] The present invention relates to a method and device for predicting the remaining useful life of equipment based on isomorphic spatio-temporal fusion and causal dilated convolution, and belongs to the technical field of equipment health management and remaining useful life prediction. Background Art

[0002] Maintenance management plays a very important role in the operation of modern large-scale mechanical equipment. With the rapid development of modern instruments and measurement technologies, it is possible to obtain condition monitoring data from operating mechanical equipment. Condition-based maintenance (CBM) is a maintenance method that requires measuring various parameters of the equipment and reflecting the actual state of the equipment at any time. Compared with traditional preventive maintenance strategies, CBM is more effective in reality because it utilizes real-time system health information, and thus is widely applied to the maintenance of modern industrial equipment. CBM involves predicting the remaining useful life (RUL) and potential failures of the equipment based on the real-time operating state. On this basis, maintenance decisions can be made as needed according to the prediction information. Obviously, RUL prediction is one of the most critical technologies for the effective implementation of CBM. If the RUL of a mechanical equipment is predicted based on current or historical operating information, the failure time can be accurately known. Therefore, RUL prediction is very important for researchers in the field of CBM.

[0003] Generally speaking, RUL prediction methods can be roughly divided into traditional model-based methods, data-driven methods, and hybrid methods. Traditional model-based methods require accurate dynamic modeling of mechanical equipment or components to describe the degradation trend of the components. However, the structures of modern industrial large-scale equipment are becoming increasingly complex, and there are various nonlinear relationships among different systems and components. Therefore, it is unrealistic to establish an accurate model.

[0004] Currently, the popularity of data-driven methods in RUL prediction has surged, mainly due to their powerful nonlinear representation and commendable generalization ability. The goal of data-driven RUL prediction methods is to establish a mapping relationship between RUL and the characteristics of the target equipment. It does not require extensive expertise and physical modeling of complex mechanical equipment. In the literature, some traditional machine learning algorithms have been used for RUL prediction, such as support vector regression (SVR), random forest (RF), and extreme learning machine (ELM). However, these methods rely on cumbersome feature engineering. In contrast, deep learning-based methods can automatically extract valuable features from raw CBM data and obtain better prediction performance. Therefore, deep learning-based RUL prediction methods have a wider range of applicability and have recently received increasing attention.

[0005] In the field of deep learning, common RUL prediction algorithms include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Attention Mechanism (AM), and Transformer, etc. Among them, recurrent neural networks and their variants perform well in time series prediction tasks and can effectively understand the patterns in time series, but they are not the optimal choice. In contrast, Convolutional Neural Network (CNN) and Generative Autoencoder have become one of the most widely used deep learning models for RUL prediction recently. The core advantage of CNN lies in its ability to capture spatial invariant features in images and patterns in texts, and it performs well in image classification tasks through effective pattern learning. In addition, many deep learning models based on CNN have also been developed to capture the patterns and structures in time series. However, when dealing with long time series, CNN-based methods need to continuously increase the size of convolutional kernels to expand the receptive field in order to obtain more sequence information. This indicates that CNN has certain limitations in capturing long-term dependency information in sequence data.

[0006] In addition, the powerful representation ability of deep learning provides new solutions for the processing of time series data in complex equipment monitoring. However, when existing prediction methods are used to process the RUL prediction of aeroengines, they often fail to fully mine the long-term, potential, and deep features in the data and are difficult to take into account both local and global degradation information at the same time. Multivariate equipment monitoring data usually contains multi-dimensional features, such as vibration signals in different directions, etc., and these features contain rich spatio-temporal patterns. Most existing data-driven methods extract time and space features through independent modules respectively, so they face limitations in flexibility and expression ability at the feature fusion stage. Most of the current mainstream feature fusion strategies adopt simple addition or concatenation operations, which are difficult to fully capture the interaction relationships between different features and cannot effectively distinguish the importance of features. This limitation not only reduces the prediction performance of the model but also weakens its robustness in practical applications. Therefore, developing a flexible and efficient feature fusion method has become a key research direction. Summary of the Invention

[0007] The object of the present invention is to provide a method and device for predicting the remaining useful life of equipment based on isomorphic spatio-temporal fusion and causal dilated convolution. By introducing causal dilated convolution, the receptive field of convolutional feature extraction is expanded. At the same time, an isomorphic spatio-temporal fusion layer is introduced to flexibly achieve feature fusion of spatio-temporal dual channels.

[0008] To achieve the above object, the present invention is implemented by adopting the following technical solutions.

[0009] On the one hand, the present invention provides a method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution, including:

[0010] Collect the original vibration signals during the entire life cycle of the device to be studied and perform preprocessing to obtain a time series;

[0011] Input the time series into a pre-constructed neural network to obtain the prediction result of the remaining useful life of the device. The data processing process in the neural network is as follows:

[0012] Perform a causal convolution operation on the input, and use an adaptive zero-padding strategy to supplement the dimensions of the input data during the causal convolution operation;

[0013] Extract spatio-temporal features from the output result of the causal convolution through a stacked spatio-temporal fusion module;

[0014] Aggregate the output of the stacked spatio-temporal fusion module through an information aggregation module for information aggregation and key feature extraction, and input it into the prediction module;

[0015] Output the prediction result of the remaining useful life of the device by the prediction module;

[0016] The extracting spatio-temporal features from the output result of the causal convolution through a stacked spatio-temporal fusion module includes:

[0017] Input the output result of the causal convolution into both the time and space channels of the first spatio-temporal fusion layer for dilated convolution processing with the same dilation rate;

[0018] Apply gated activation units to the dilated convolution results of both the time and space channels respectively, and perform spatio-temporal cross-fusion;

[0019] On the one hand, transfer the result of the spatio-temporal cross-fusion to the next spatio-temporal fusion layer through a residual connection, and on the other hand, output it to the information aggregation module.

[0020] Preferably, the collecting the original vibration signals during the entire life cycle of the device to be studied and performing preprocessing to obtain a time series includes:

[0021] Collect the original vibration signals of the device during the entire life cycle through different types of sensors, and screen according to the characteristic distribution of the collected original vibration signals to screen out the original vibration signals with learning characteristics;

[0022] Perform sliding time window segmentation and normalization processing on the screened original vibration signals to obtain a time series.

[0023] Preferably, the performing a causal convolution operation on the input includes:

[0024] Add zero padding to the front end and the back end of the input simultaneously to form a vector with a length of ; where is the length of the input sequence, and is the receptive field size.

[0025] Preferably, the stacked spatio-temporal fusion module includes N spatio-temporal fusion layers with the same structure;

[0026] Each spatio-temporal fusion layer includes a temporal channel and a spatial channel;

[0027] Each channel includes: an initial convolutional layer, a linear layer, a stacked dilated convolutional module, a gated activation unit, a weight-based cross-fusion layer, and a residual connection;

[0028] The initial convolutional layer is used to adjust the dimension of the input features to adapt to the computational requirements of the stacked dilated convolutional layer.

[0029] Preferably, the stacked dilated convolutional module expands the receptive field by introducing dilated convolutions layer by layer. The receptive field of the stacked dilated convolutional module is calculated as follows:

[0030] ,

[0031] where is the receptive field of the stacked dilated convolution, represents the dilation rate of the th layer, represents the stride of the convolution in the th layer, represents the kernel size of the convolution in the th layer, and is the number of layers of the dilated convolutional layer.

[0032] Preferably, the gated activation unit is used to enhance the selectivity of the features extracted by the stacked dilated convolution.

[0033] The gated activation unit includes two parallel linear mapping paths: a main path and a gated path;

[0034] The main path is used to perform a non-linear feature transformation on the output of the stacked dilated convolutional module. The gated path is used to generate a gating weight through the Sigmoid function. The outputs of the two paths are fused through element-wise multiplication to achieve adaptive screening of features;

[0035] After the features fused by element-wise multiplication are layer-normalized, the Dropout mechanism is used to suppress overfitting to obtain the output of the gated activation unit.

[0036] Preferably, the data processing process of the weight-based cross-fusion layer is as follows:

[0037] The output of the time channel and the space channel passing through the gated activation unit are respectively adjusted in terms of the feature map dimension by a linear layer, and then spatio-temporal cross-fusion is achieved through weighted summation, which is expressed as:

[0038] ,

[0039] ,

[0040] wherein, and respectively represent the output of the space channel and the time channel passing through the gated activation unit, and respectively represent the output after spatio-temporal cross-fusion of the space channel and the time channel, and respectively represent the linear transformation operations of the space channel and the time channel, and are the fusion weights.

[0041] Preferably, the output of the stacked spatio-temporal fusion module is subjected to information aggregation and key feature extraction through an information aggregation module, including:

[0042] Global weighted accumulation of the spatio-temporal features extracted by each layer of the stacked spatio-temporal fusion module is achieved through skip connection. After passing through the ReLU activation function, max pooling is performed.

[0043] Preferably, the prediction module is used to perform a regression operation on the input features to obtain the final prediction result of the remaining useful life of the device;

[0044] The prediction module includes: two 1×1 convolutional layers, a ReLU activation function, and a Softmax.

[0045] On the other hand, the present invention provides a device for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution, which is used to implement the method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution. The device includes:

[0046] A data preprocessing module, which is used to collect the original vibration signals during the whole life cycle of the device to be studied and perform preprocessing to obtain a time series;

[0047] A neural network, which is used to take the time series obtained by the data preprocessing module as input and output the prediction result of the remaining useful life of the device;

[0048] The neural network includes:

[0049] A causal convolutional layer, which is used to perform a causal convolution operation on the input, and an adaptive zero-padding strategy is adopted to supplement the dimension of the input data during the causal convolution operation;

[0050] Stacked spatio-temporal fusion module, which is used to extract spatio-temporal dual-channel features from the output of the causal convolution layer, uses dilated convolution to expand the receptive field, and adopts a gated activation unit and a weight-based cross-fusion mechanism for spatio-temporal cross-fusion, and stacks multiple layers for gradual extraction;

[0051] Information aggregation module, which is used to aggregate information and extract key features from the output of the stacked spatio-temporal fusion module, and output to the prediction module;

[0052] Prediction module, which is used to perform a regression operation on the input features and output the predicted result of the remaining useful life of the device;

[0053] The stacked spatio-temporal fusion module includes N spatio-temporal fusion layers with the same structure;

[0054] Each spatio-temporal fusion layer includes a time channel and a space channel;

[0055] Each channel includes: an initial convolution layer, a linear layer, a stacked dilated convolution module, a gated activation unit, a weight-based cross-fusion layer, and a residual connection.

[0056] Compared with the prior art, the beneficial effects achieved by the present invention:

[0057] The present invention proposes a method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution. By introducing causal dilated convolution, the receptive field of convolutional feature extraction is significantly expanded, thereby enhancing the ability to capture long-term dependencies in time series data. At the same time, the present invention innovatively introduces an isomorphic spatio-temporal fusion layer to achieve flexible fusion of time and space dual-channel features, effectively solving the problem of feature fusion limitations in existing methods when dealing with complex multi-variable device monitoring data. In addition, the present invention combines an efficient gated activation unit and a weight cross-fusion mechanism to further optimize the feature extraction process, and through residual connection and information aggregation module, improves the training stability and prediction accuracy of the model. This method not only has innovation in theory, but also shows higher computational efficiency and prediction performance in practical applications, providing a more accurate and efficient solution for the health management of modern industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Shown is a flowchart of a method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution provided in an embodiment of the present invention;

[0059] Figure 2 Shown is a schematic diagram of a sliding time window in an embodiment of the present invention;

[0060] Figure 3The figure shows a schematic diagram of a neural network for a method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution provided in an embodiment of the present invention;

[0061] Figure 4 The figure shows a schematic diagram of the spatio-temporal fusion layer structure in an embodiment of the present invention;

[0062] Figure 5 The figure shows a schematic diagram of stacked dilated convolution in an embodiment of the present invention;

[0063] Figure 6 The figure shows a schematic diagram of the gated activation unit structure in an embodiment of the present invention. Detailed implementation manners

[0064] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present invention are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0065] Here, it should also be noted that in order to avoid obscuring the present invention due to unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less related to the present invention are omitted.

[0066] It should be emphasized that the term "including / comprising" when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0067] It should be emphasized here that the step labels mentioned below are not limitations on the order of the steps. Instead, it should be understood that the steps can be executed in the order mentioned in the embodiments, can be different from the order in the embodiments, or several steps can be executed simultaneously.

[0068] It should also be noted that the term "and / or" is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " generally represents an "or" relationship between the associated objects before and after.

[0069] Embodiment 1

[0070] This embodiment provides a method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution. Refer to Figure 1 , including:

[0071] Collect the original vibration signals during the whole life cycle of the equipment to be studied, screen them according to the characteristic distribution of the data of each sensor, select the sensor data with learning characteristics, and perform sliding time window segmentation and normalization processing on the selected original vibration signals to obtain a time series;

[0072] Input the preprocessed time series data into a pre-constructed neural network to obtain the prediction result of the remaining service life of the equipment. Among them, the neural network data processing process is as follows:

[0073] Perform causal convolution operations and adopt an adaptive zero-padding strategy to ensure the causality of the time series and eliminate boundary effects;

[0074] Extract spatio-temporal features from the output result of the causal convolution through a stacked spatio-temporal fusion module;

[0075] After aggregating information and extracting key features from the features extracted by the stacked spatio-temporal fusion module through an information aggregation module, input them into the prediction module;

[0076] The prediction module outputs the prediction result of the remaining service life of the equipment.

[0077] Among them, extracting spatio-temporal features from the output result of the causal convolution through a stacked spatio-temporal fusion module includes:

[0078] Input the output result of the causal convolution into the time and space channels of the first spatio-temporal fusion layer for dilated convolution processing with the same dilation rate;

[0079] Apply gated activation units to the dilated convolution results of the time and space channels respectively, and introduce a weight cross-fusion mechanism to achieve cross-fusion of spatio-temporal channel data;

[0080] On the one hand, transfer the result of spatio-temporal cross-fusion to the next spatio-temporal fusion layer through a residual connection, and on the other hand, output it to the information aggregation module for further feature extraction.

[0081] In this embodiment, for the problem that some sensors have no learning characteristics in the original vibration signals collected by different types of sensors during the whole life cycle of the equipment, the sensor data with learning characteristics are selected by checking the characteristic distribution of the sensor data.

[0082] In this embodiment, the original vibration signals after sensor screening are preprocessed by the sliding time window method, which is commonly used for data segmentation preprocessing. See Figure 2 , including:

[0083] Assume that the original time series data is , , where is the total length of the time series, and the size of the time sliding window is , and the step size is . By processing with a sliding time window, the original time series can be segmented into multiple subsequences, and the length of each subsequence is .

[0084] In this embodiment, the data processed by the sliding window is normalized. The data from different sensors with different units and scales is normalized by the min-max scaling method. For the condition-based maintenance (CBM) data , the processing process is shown in Equation (1):

[0085] (1)

[0086] where represents the minimum value in the time series data, represents the maximum value in the time series data, represents the value after normalization.

[0087] In this embodiment, the constructed neural network structure is as shown in Figure 3 , including: a feature extractor and a prediction module. Among them, the feature extractor includes: a causal convolutional layer, a stacked spatio-temporal fusion module, and a max pooling layer.

[0088] In this embodiment, aiming at the forward zero-padding strategy used in the original causal convolution, there are boundary effects in time series prediction, and there are certain limitations in processing time-step sequential sequences. An improved bidirectional zero-padding strategy is proposed, including:

[0089] Assume that the input data is , with a length of , a filter size of , a receptive field size of . The traditional unilateral causal padding method adds zero padding at the front end of the input, thus forming a vector with a length of . The output dimension is shown in Equation (2):

[0090] (2)

[0091] While the adaptive zero-padding strategy adds zero padding at both the front end and the back end of the input, forming a vector with a length of . The output dimension is shown in Equation (3):

[0092] (3)

[0093] In this embodiment, the stacked spatio-temporal fusion module includes N spatio-temporal fusion layers with the same structure. The output of the upper spatio-temporal fusion layer will be used as the input of the next spatio-temporal fusion layer, and the same operation is performed on each layer. The spatio-temporal fusion layer is as Figure 4 shown, including:

[0094] Adopt an isomorphic feature extraction structure, which consists of a time channel and a space channel. Both use the same computing module, and only differ in the input dimension and parameter matrix;

[0095] Each channel consists of the following modules: an initial convolutional layer, a linear layer, a stacked dilated convolutional module, a gated activation unit, a weight-based cross-fusion layer, and a residual connection, and finally outputs through a feature fusion layer;

[0096] Among them, the initial convolutional layer is used to adjust the dimension of the input features to adapt to the computational requirements of the dilated convolution. Specifically, the input dimension of the spatial features and the input dimension of the time features are mapped to and through the initial convolution, where is the batch size, is the spatial dimension size, is the time step, is the dimension required for the dilated convolution, ensuring that the input feature dimension matches the computational structure of the subsequent dilated convolution.

[0097] In this embodiment, the stacked dilated convolution is used to expand the receptive field by introducing dilated convolution layer by layer. The example of the change in the receptive field and dilation factor is as Figure 5 shown. The derivation process of the receptive field size result includes:

[0098] Define a convolutional neural network, whose input signal is , is the number of dilated convolution layers, and the final output is , the output height of the th layer is , the width is , and the depth is , , and its output representation is shown in Equation (4):

[0099] (4)

[0100] In each layer , the kernel size is , the left padding is , the right padding is and the stride size is . Therefore, the receptive field size of the final output of each layer is defined as shown in formula (5):

[0101] (5)

[0102] in, Indicates the The receptive field size of the layer convolution, Indicates the The stride of the layer convolution, Indicates the The convolution kernel size of the layer convolution;

[0103] Expression (5) can be rewritten as (6):

[0104] (6)

[0105] Formula (6) is a first-order nonhomogeneous recursive equation with variable coefficients, which can be obtained by multiplying it by To solve, we get formula (7):

[0106] (7)

[0107] when ,and ,but , expression (7) can be derived as formula (8):

[0108] (8)

[0109] Transform expression (8) from arrive Sum and convert into formula (9):

[0110] (9)

[0111] because , , then the calculation result of the receptive field of the full convolutional network is shown in formula (10):

[0112] (10)

[0113] Expression (10) can be converted into the following expression (11):

[0114] (11)

[0115] In stacked dilated convolutions, the dilation rate is introduced to effectively increase the receptive field. By constructing a multi-layer dilated convolutional network, the dilation factor increases linearly with the layer depth. The dilation rate ensures that the filter skips the processed input at a constant rate instead of directly applying the filter to the input sequentially. Therefore, compared with conventional stacked convolutions, the receptive field of the stacked dilated convolution is calculated as shown in Equation (12):

[0116] (12)

[0117] where represents the dilation rate of the th layer.

[0118] In this embodiment, the selectivity of feature extraction by the dilated convolution is enhanced through a gated activation unit. The structure is as Figure 6 shown, including:[[]]

[0119] It consists of two parallel linear transformation branches, which are respectively used for feature transformation and gated weight calculation;

[0120] Specifically, the input data first passes through two independent linear mapping paths. Among them, the main path is responsible for non-linear feature transformation, while the gated path generates a gated weight through the Sigmoid function to adjust the passing ratio of the information flow. Finally, the outputs of these two branches are fused through element-wise multiplication to achieve adaptive screening of features. The calculation process is as shown in Equation (13):

[0121] (13)

[0122] where is the input feature after dilated convolution, and respectively represent the learnable parameters of the two branches, and respectively represent the bias terms of the two branches, represents the non-linear activation function of the main branch. Here, ReLU is selected, is the Sigmoid function, which is used to generate the gated signal, represents element-wise multiplication;

[0123] To improve the stability of the model and optimize the training process, layer normalization is further introduced after the calculation of this gated activation unit to mitigate the problem of gradient vanishing, and the Dropout mechanism is adopted to suppress overfitting. The final output calculation is as shown in Equation (14):

[0124] (14)

[0125] where Is the output of the gating activation unit, Indicates the Dropout operation, Indicates the layer normalization operation.

[0126] In this embodiment, a feature dynamic exchange is achieved in the width direction of the network through a weight-based cross-fusion mechanism, which is similar to the information transfer mechanism of the residual connection in the depth direction, including:

[0127] During the cross-fusion process, the spatial and temporal features first pass through the gating activation unit ( and ) for screening to suppress redundant information and enhance the expression ability of key features;

[0128] To ensure the dimensional consistency of the fused features, linear layers ( and ) are respectively used to adjust the feature map dimensions;

[0129] Cross-fusion is achieved through weighted summation, and its mathematical formulas are shown in Eqs. (15), (16), (17), and (18):

[0130] (15)

[0131] (16)

[0132] (17)

[0133] (18)

[0134] Wherein, and respectively represent the outputs of the spatial and temporal channels after passing through the gating activation unit, and respectively represent the outputs after cross-fusion of the spatial and temporal channels, and respectively represent the input matrices of the spatial and temporal channels after dilated convolution, and respectively represent the gating activation unit operations of the spatial and temporal channels, and respectively represent the linear transformation operations of the spatial and temporal channels, while and serve as fusion weights to adaptively adjust the influence of cross features.

[0135] In this embodiment, since spatio-temporal feature extraction is performed by stacking spatio-temporal fusion layers, the cross-fusion result of each layer is output to the next layer through residual connection processing, specifically as follows:

[0136] The input information is retained through the identity mapping and added to the transformation results of subsequent layers, thereby alleviating the vanishing gradient problem and promoting the training of deeper networks. Its calculation formula is shown in Equation (19):

[0137] (19)

[0138] where, represents the initial data input to the current layer, represents the result obtained after a series of operations on the input such as dilated convolution, gated activation unit, cross fusion, etc., represents the output result after residual connection.

[0139] In this embodiment, the results of the stacked spatio-temporal fusion module are processed by an information aggregation module. The information aggregation module includes a skip connection layer, ReLU activation, and a max pooling layer. The data processing process is as follows: The skip connection is used to achieve the global weighted accumulation of the spatio-temporal features extracted by the stacked spatio-temporal fusion module. The expression is shown in Equation (20):

[0140] (20)

[0141] where, represents the aggregated features after skip connection, is a hierarchically adaptive weight matrix (usually the convolution kernel), is the output feature of the th layer. This mechanism can effectively retain the feature representations of different scales and improve the utilization rate of deep-layer information;

[0142] The ReLU activation function is used to ensure that the model can learn more complex feature relationships and suppress irrelevant or negatively influential features;

[0143] The max pooling layer (Max Pooling) is used to further extract key features and reduce redundant information. The calculation method is shown in Equation (21):

[0144] (21)

[0145] where, represents the input after ReLU processing, represents the stride of the pooling layer, is the pooling window size, is the result after pooling by the th pooling unit, , Indicates the length after pooling. The introduction of the max pooling layer can effectively highlight the most significant features within the local window, reduce noise interference, and improve the robustness of the model.

[0146] In this embodiment, for the results of feature extraction, the prediction module performs a regression operation on the extracted features to obtain the predicted result of the remaining useful life of the device. The prediction module is the Figure 3 regression layer in, including: two 1×1 convolutional layers, a ReLU activation function, and a Softmax. Among them, the 1×1 convolution is used to adjust the feature dimension to match the output dimension, and the activation function is used for non-linear mapping to enhance the feature expression ability.

[0147] Embodiment 2

[0148] Based on the same inventive concept, Embodiment 2 of the present invention provides a device for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution, which is used to implement the method of Embodiment 1 above. The device includes:

[0149] A data preprocessing module, which is used to perform data screening with learning features, sliding window processing, and normalization processing on the original signals generated by the device under study during its entire life cycle to obtain preprocessed data that meets the requirements;

[0150] A causal convolution module, which is used to perform causal convolution operations on the input to ensure the causality of the time series and eliminate boundary problems, and uses an adaptive zero-padding strategy to supplement the dimensions of the preprocessed data;

[0151] A stacked spatio-temporal fusion module, which is used to perform spatio-temporal dual-channel feature extraction on the causal data, uses dilated convolution to expand the receptive field, and uses a gated unit and a weight cross-fusion mechanism to achieve cross-fusion of the time and space channels, and stacks multiple layers for gradual extraction;

[0152] An information aggregation module, which is used to aggregate information and further extract features from the spatio-temporal fusion results, realizes global weighted accumulation of features through skip connections, and uses a max pooling layer to extract key information;

[0153] A prediction module, which is used to perform a regression operation on the extracted features to obtain the final result of the remaining useful life of the device.

[0154] It should be noted that this device embodiment corresponds to the above method embodiment. The implementation methods of the above method embodiment are all applicable to this device embodiment and can achieve the same or similar technical effects, so they will not be elaborated here.

[0155] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0156] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0157] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that realize the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0158] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0159] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit of the present invention and the scope protected by the claims. These all fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining useful life of equipment based on isomorphic spatio-temporal fusion and causal dilated convolution, characterized in that, Including: Collecting the original vibration signals during the whole life cycle of the device to be studied and performing preprocessing to obtain a time series; Inputting the time series into a pre-constructed neural network to obtain the predicted result of the remaining service life of the device. The data processing process in the neural network is as follows: Performing causal convolution operation on the input, and using an adaptive zero-padding strategy to supplement the dimension of the input data in the causal convolution operation; Performing spatio-temporal feature extraction on the output result of the causal convolution through a stacked spatio-temporal fusion module; Aggregating information and extracting key features from the output of the stacked spatio-temporal fusion module through an information aggregation module, and inputting it into the prediction module; Outputting the predicted result of the remaining service life of the device by the prediction module; The performing spatio-temporal feature extraction on the output result of the causal convolution through a stacked spatio-temporal fusion module includes: Inputting the output result of the causal convolution into both the time and space channels of the first spatio-temporal fusion layer, and performing dilated convolution processing with the same dilation rate; Applying gated activation units to the dilated convolution results of both the time and space channels respectively, and performing spatio-temporal cross-fusion; On the one hand, transmitting the result of the spatio-temporal cross-fusion to the next spatio-temporal fusion layer through a residual connection, and on the other hand, outputting it to the information aggregation module.

2. The method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 1, characterized in that, The collecting the original vibration signals during the whole life cycle of the device to be studied and performing preprocessing to obtain a time series includes: Collecting the original vibration signals of the device during the whole life cycle through different types of sensors, screening according to the characteristic distribution of the collected original vibration signals, and screening out the original vibration signals with learning characteristics; Performing sliding time window segmentation and normalization processing on the screened original vibration signals to obtain a time series.

3. A method for predicting the remaining useful life of equipment based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 1, characterized in that, The performing causal convolution operation on the input includes: Add zero-padding to both the front end and the back end of the input to form a vector of length ; where is the length of the input sequence, and is the receptive field size.

4. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 1, characterized in that, The stacked spatio-temporal fusion module includes N spatio-temporal fusion layers with the same structure; Each spatio-temporal fusion layer includes a time channel and a space channel; Each channel includes: an initial convolution layer, a linear layer, a stacked dilated convolution module, a gated activation unit, a weight-based cross-fusion layer, and a residual connection; The initial convolution layer is used to adjust the dimension of the input features to adapt to the calculation requirements of the stacked dilated convolution layer.

5. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 4, characterized in that, The stacked dilated convolution module expands the receptive field by introducing dilated convolution layer by layer. The calculation of the receptive field of the stacked dilated convolution module is as follows: ， Among them, is the receptive field of the stacked dilated convolution, represents the dilation rate of the -th layer, represents the stride of the -th layer convolution, represents the convolution kernel size of the -th layer convolution, is the number of layers of the dilated convolution layer.

6. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 4, characterized in that, The gated activation unit is used to enhance the selectivity of the features extracted by the stacked dilated convolution, The gated activation unit includes two parallel linear mapping paths: a main path and a gating path; The main path is used to perform non-linear feature transformation on the output of the stacked dilated convolution module, and the gating path is used to generate gating weights through the Sigmoid function. The outputs of the two paths are fused through element-wise multiplication to achieve adaptive screening of features; After performing layer normalization on the features fused through element-wise multiplication, the Dropout mechanism is used to suppress overfitting to obtain the output of the gated activation unit.

7. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 6, characterized in that The data processing process of the weight-based cross-fusion layer is: Adjusting the feature map dimensions of the outputs of the time channel and the space channel through the gated activation unit respectively by using a linear layer, and then realizing spatio-temporal cross-fusion through weighted summation, expressed as: ， ， Among them, and respectively represent the outputs of the spatial channel and the temporal channel after being gated and activated by the gating activation unit, and respectively represent the outputs after the spatial channel and the temporal channel are cross - fused, and respectively represent the linear transformation operations of the spatial channel and the temporal channel, and are the fusion weights.

8. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to claim 1, characterized in that The information aggregation and key feature extraction of the output of the stacked spatio-temporal fusion module through the information aggregation module includes: Using skip connections to achieve global weighted accumulation of spatio-temporal features extracted by each layer of the stacked spatio-temporal fusion module. After passing through the ReLU activation function, max pooling is performed.

9. A method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution, characterized in that, The prediction module is used to perform a regression operation on the input features to obtain the final predicted result of the remaining useful life of the device; The prediction module includes: two 1×1 convolutional layers, a ReLU activation function, and a Softmax.

10. A device remaining useful life prediction device based on isomorphic spatio-temporal fusion and causal dilated convolution, characterized in that, For implementing the method for predicting the remaining useful life of a device based on isomorphic spatio-temporal fusion and causal dilated convolution according to any one of claims 1 to 9, the device includes: A data preprocessing module, configured to collect the original vibration signals during the entire life cycle of the device to be studied and perform preprocessing to obtain a time series; A neural network, configured to take the time series obtained by the data preprocessing module as input and output the predicted result of the remaining useful life of the device; The neural network includes: A causal convolutional layer, configured to perform a causal convolution operation on the input, and in the causal convolution operation, an adaptive zero-padding strategy is used to supplement the dimensions of the input data; A stacked spatio-temporal fusion module, configured to perform spatio-temporal dual-channel feature extraction on the output of the causal convolutional layer, use dilated convolution to expand the receptive field, and use a gated activation unit and a weight-based cross-fusion mechanism for spatio-temporal cross-fusion, and stack multiple layers for gradual extraction; An information aggregation module, configured to perform information aggregation and key feature extraction on the output of the stacked spatio-temporal fusion module and output it to the prediction module; A prediction module, configured to perform a regression operation on the input features and output the predicted result of the remaining useful life of the device; The stacked spatio-temporal fusion module includes N spatio-temporal fusion layers with the same structure; Each spatio-temporal fusion layer includes a time channel and a space channel; Each channel includes: an initial convolutional layer, a linear layer, a stacked dilated convolution module, a gated activation unit, a weight-based cross-fusion layer, and a residual connection.

Images