Remaining life prediction method and device based on multi-scale decomposition enhancement

Abstract

The present invention discloses a method and device for predicting remaining life based on multi-scale decomposition enhancement. Three different scales are randomly selected as pooling kernel sizes. Average pooling features and maximum pooling features are obtained based on a parameter data set, and fused features are generated. Three fused features are then sorted from large to small according to the corresponding scales to obtain three enhanced features. Three enhanced features corresponding to the three fused features are then determined. A Bi-LSTM model is used to obtain the three enhanced features and their resulting feature representation set. Target gating weights are determined based on the feature representation set, and prediction results are determined based on the three obtained feature representations. The prediction results are weighted and aggregated using the target gating weights to obtain intermediate prediction results corresponding to the three scales. A target predicted life is determined based on the intermediate prediction results. This method accurately predicts the remaining life of a device, given that features at different scales in the parameter data set may have significant degradation characteristics.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment life prediction, and in particular to a method and device for predicting remaining life based on multi-scale decomposition enhancement. Background Art

[0002] With the continuous development of industry, approaches such as cyber-physical systems (CPS), the Internet of Things (IoT), the Internet of Services (IOS), and data analytics have effectively improved manufacturing efficiency and helped industries successfully address the economic, social, and environmental challenges of industrial production. Condition-Based Maintenance (CBM) involves executing systematic maintenance and repair routines based on industrial production needs, while Prognostics and Health Management (PHM) involves monitoring the evolution of component wear through detection, diagnosis, and prediction. This is a proactive approach to CBM. PHM approaches primarily predict remaining useful life (RUL) from historical and actual operating conditions. Remaining Useful Life (RUL) is one of the most important parameters for predicting component failure. While current data-driven approaches have achieved promising results in equipment life prediction, most fail to consider the degradation process from a multi-scale perspective. Furthermore, existing multi-scale approaches often directly fuse features from multiple scales into a single feature representation for analysis. However, features at different scales may have distinct degradation characteristics, increasing the difficulty of prediction and reducing the accuracy of life prediction. Summary of the Invention

[0003] Based on this, it is necessary to propose a remaining life prediction method and device based on multi-scale decomposition enhancement to address the above problems.

[0004] A method for predicting the remaining life of used equipment based on multi-scale decomposition enhancement is provided. The method comprises:

[0005] Acquire a parameter data set of the used device, and preprocess the parameter data set to obtain a target data set;

[0006] Randomly select three different scales as pooling kernel sizes, perform pooling operation on the target data set to obtain average pooling features and maximum pooling features;

[0007] The average pooling feature and the maximum pooling feature are spliced ​​in the channel dimension to obtain a spliced ​​feature; the spliced ​​feature is operated by a convolutional neural network and a sigmoid activation function to obtain an initial gating weight; the fusion feature is determined according to the average pooling feature, the maximum pooling feature and the initial gating weight;

[0008] The three scales corresponding to the corresponding scales are sorted from large to small, and are defined as the first fused feature, the second fused feature, and the third fused feature; the first fused feature, the second fused feature, and the third fused feature are multi-level fused through the MLP layer to update the second fused feature to obtain the second updated fused feature, and the third fused feature is updated to obtain the third updated fused feature; the first fused feature, the third fused feature, and the third updated fused feature are respectively subjected to attention enhancement operations to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature;

[0009] Obtaining, through the Bi-LSTM model, a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation;

[0010] Determining a target gating weight based on the feature representation set, and determining a prediction result based on the first feature representation, the second feature representation, and the third feature representation;

[0011] The prediction results are weighted and aggregated by the target gating weight to obtain intermediate prediction results corresponding to three scales; and the target prediction life is determined according to the intermediate prediction results.

[0012] In one embodiment, performing a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature includes:

[0013] Performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature;

[0014] Performing a maximum pooling operation on the target data set to obtain the maximum pooling feature.

[0015] In one embodiment, determining the fusion feature according to the average pooling feature and the maximum pooling feature includes:

[0016] Concatenate the average pooling feature and the maximum pooling feature in the channel dimension to obtain a concatenated feature;

[0017] The concatenated features are operated by a convolutional neural network and a sigmoid activation function to obtain an initial gating weight;

[0018] The fusion feature is determined according to the average pooling feature, the maximum pooling feature and the initial gating weight.

[0019] In one embodiment, determining the first enhanced feature, the second enhanced feature, and the third enhanced feature corresponding to the first fusion feature, the second fusion feature, and the third fusion feature after being enhanced includes:

[0020] Performing multi-level fusion through the MLP layer based on the first fused feature, the second fused feature, and the third fused feature to update the second fused feature to obtain a second updated fused feature, and updating the third fused feature to obtain a third updated fused feature;

[0021] An attention enhancement operation is performed on the first fused feature, the third fused feature and the third updated fused feature respectively to obtain the first enhanced feature, the second enhanced feature and the third enhanced feature.

[0022] In one embodiment,

[0023] The acquisition of the parameter data set of the used device and the preprocessing of the parameter data set to obtain the target data set are achieved by the following expression:

[0024] X

[0025] Among them, X is the target data set, X in is the parameter data set, is the mean of the parameter data set, is the standard deviation of the parameter data set;

[0026] Performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature; performing a maximum pooling operation on the target data set to obtain the maximum pooling feature is achieved by the following expression:

[0027]

[0028]

[0029] in, is the average pooling feature, is the maximum pooling feature, AvgPool1D represents the one-dimensional average pooling operation, and MaxPool1D represents the maximum pooling operation.

[0030] In one embodiment, the average pooling feature and the maximum pooling feature are spliced ​​in the channel dimension to obtain a spliced ​​feature; the spliced ​​feature is operated by a convolutional neural network and a sigmoid activation function to obtain an initial gating weight; and the fusion feature is determined according to the average pooling feature, the maximum pooling feature and the initial gating weight by the following expression:

[0031] ∈

[0032]

[0033]

[0034] in, Concat is a concatenation feature. is the average pooling feature, is the maximum pooling feature, is the initial gating weight, Represents the Sigmoid activation function, Conv1D represents the one-dimensional convolution operation, For fusion features.

[0035] In one embodiment, the multi-level fusion of the first fused feature, the second fused feature, and the third fused feature is performed through the MLP layer to update the second fused feature to obtain a second updated fused feature, and the third fused feature is updated to obtain a third updated fused feature; and the attention enhancement operation is performed on the first fused feature, the third fused feature, and the third updated fused feature to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively, which are implemented by the following expressions:

[0036] , i=2, 3

[0037] , i=2, 3

[0038]

[0039] When i is 2, That is , is the first mapping feature, That is , is the first fusion feature; when i is 3, That is , is the second mapping feature, That is , is the second fusion feature; when i is 2 or 3 They correspond to the second updated fusion feature and the third updated fusion feature respectively; when i takes the value of 1, 2 and 3, They correspond to the first enhancement feature, the second enhancement feature and the third enhancement feature respectively, For data format conversion operations, For attention-enhancing operations.

[0040] In one embodiment, the Bi-LSTM model is used to obtain the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation; determining the target gating weight according to the feature representation set, and determining the prediction result according to the first feature representation, the second feature representation, and the third feature representation are implemented by the following expression:

[0041]

[0042]

[0043]

[0044] A={ 、 、 }

[0045]

[0046]

[0047]

[0048] , i=1, 2, 3

[0049] , i=1, 2, 3

[0050] , i=1, 2, 3

[0051] in, is the local time feature, is the first enhancement feature, the second enhancement feature, or the third enhancement feature; is the global temporal feature of the i-th scale, is the first feature representation, the second feature representation or the third feature representation, A is the feature representation set, is the expert weight, is the learnable parameter matrix, is the first bias vector, is the mask matrix, is the target gating weight, is the transformation feature, is the mapping transformation feature, To predict the results, is the weight matrix, is the second bias vector.

[0052] In one embodiment, performing a weighted aggregation operation on the prediction results using the target gating weights to obtain intermediate prediction results corresponding to three scales; and determining the target predicted lifespan based on the intermediate prediction results is achieved by the following expression:

[0053]

[0054]

[0055] in, is the intermediate prediction result, is the gating weight of the prediction result of the s-th expert on the i-th scale feature, represents the prediction result of the s-th expert for the i-th scale feature, represents the weight of the i-th scale, S represents the number of scales, Predict lifespan for a target.

[0056] A remaining life prediction device based on multi-scale decomposition enhancement, the device comprising:

[0057] An acquisition module acquires a parameter data set of a used device, preprocesses the parameter data set to obtain a target data set, and randomly selects three different size scales as the pooling kernel size of the operation module;

[0058] An operation module, configured to perform a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature;

[0059] A first determination module is configured to concatenate the average pooling feature and the maximum pooling feature in the channel dimension to obtain a concatenated feature; operate the concatenated feature through a convolutional neural network and a sigmoid activation function to obtain an initial gating weight; and determine the fused feature based on the average pooling feature, the maximum pooling feature, and the initial gating weight;

[0060] a sorting module, configured to sort the fused features corresponding to the three different scales from large to small according to the corresponding scales, and define them as a first fused feature, a second fused feature, and a third fused feature; perform multi-level fusion through an MLP layer based on the first fused feature, the second fused feature, and the third fused feature to update the second fused feature to obtain a second updated fused feature, and update the third fused feature to obtain a third updated fused feature; perform an attention enhancement operation on the first fused feature, the third fused feature, and the third updated fused feature, respectively, to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature;

[0061] An enhancement module, configured to obtain, through a Bi-LSTM model, a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation;

[0062] a second determination module, configured to determine a target gating weight according to the feature representation set, and determine a prediction result according to the first feature representation, the second feature representation, and the third feature representation;

[0063] The third determination module is used to perform a weighted aggregation operation on the prediction results through the target gating weight to obtain intermediate prediction results corresponding to three scales; and determine the target predicted lifespan according to the intermediate prediction results.

[0064] The present invention preprocesses the parameter data set, randomly extracts three scales of different sizes as the pooling kernel size, and obtains the average pooling feature and the maximum pooling feature. The average pooling feature and the maximum pooling feature are combined to generate a fusion feature, which is sorted from large to small according to the corresponding scale and defined as the first fusion feature, the second fusion feature, and the third fusion feature; and the first enhanced feature, the second enhanced feature, and the third enhanced feature corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature are determined respectively; the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature are obtained through a Bi-LSTM model; the target gating weight is determined according to the feature representation set, and the prediction result is determined according to the first feature representation, the second feature representation, and the third feature representation; the prediction result is weighted and aggregated by the target gating weight to obtain the intermediate prediction results corresponding to the three scales; and the target predicted life is determined according to the intermediate prediction result. Since the features of different scales in the parameter data set may have obvious degradation characteristics, the remaining life of the equipment is accurately predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0066] in:

[0067] Figure 1FIG2 is an application environment diagram of a remaining useful life prediction method based on multi-scale decomposition enhancement in one embodiment;

[0068] Figure 2 Flowchart of a method for predicting remaining useful life based on multi-scale decomposition enhancement in one embodiment;

[0069] Figure 3 FIG1 is a framework diagram of an MSP-Moe algorithm of a remaining useful life prediction method based on multi-scale decomposition enhancement in one embodiment;

[0070] Figure 4 A comparison chart of the predicted value and label of the entire equipment life of the test set in data set 1 in one embodiment;

[0071] Figure 5 A comparison chart of the predicted value and label of the entire equipment life of the test set in data set 2 in one embodiment;

[0072] Figure 6 The degradation prediction process of data set 1 of device No. 34 in one embodiment is shown;

[0073] Figure 7 This is a degradation prediction process for data set 2 of device No. 21 in one embodiment;

[0074] Figure 8 Part of the experimental results of data set 1 in one embodiment;

[0075] Figure 9 Another part of the experimental results of data set 1 in one embodiment;

[0076] Figure 10 1 is a structural block diagram of a remaining life prediction device based on multi-scale decomposition enhancement in one embodiment;

[0077] Figure 11 FIG. 1 is a structural block diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0078] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0079] Figure 1 FIG. 1 is an application environment diagram of a method for predicting remaining life based on multi-scale decomposition enhancement in one embodiment. Figure 1, the remaining life prediction method based on multi-scale decomposition enhancement is applied to the remaining life prediction system based on multi-scale decomposition enhancement. The remaining life prediction system based on multi-scale decomposition enhancement includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server 120 can be implemented as an independent server or a server cluster composed of multiple servers. The terminal 110 is used to obtain the parameter data set of the used device, pre-process the parameter data set to obtain the target data set; randomly extract three different scales as the pooling kernel size, and perform a pooling operation on the target data set to obtain the average pooling feature and the maximum pooling feature. The server 120 is used to determine a fusion feature based on the average pooling feature and the maximum pooling feature; sort the fusion features corresponding to the three different scales from large to small according to the corresponding scales, and define them as a first fusion feature, a second fusion feature, and a third fusion feature; and determine the first enhanced feature, the second enhanced feature, and the third enhanced feature corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively after being enhanced; obtain the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation through a Bi-LSTM model; determine a target gating weight based on the feature representation set, and determine a prediction result based on the first feature representation, the second feature representation, and the third feature representation; perform a weighted aggregation operation on the prediction result through the target gating weight to obtain an intermediate prediction result corresponding to the three scales; and determine a target predicted lifespan based on the intermediate prediction result.

[0080] With the continuous development of industry, approaches such as cyber-physical systems (CPS), the Internet of Things (IoT), the Internet of Services (IOS), and data analytics have effectively improved manufacturing efficiency and helped industries successfully address the economic, social, and environmental challenges inherent in industrial production. Condition-Based Maintenance (CBM) involves executing systematic maintenance and repair routines based on industrial production needs, while Prognostics and Health Management (PHM) involves monitoring the evolution of component wear through detection, diagnosis, and prediction. This is a proactive approach to CBM. PHM approaches primarily predict remaining useful life (RUL) from historical and actual operating conditions. Remaining Useful Life (RUL) is one of the most important parameters for predicting component failure. While current data-driven approaches have achieved promising results in equipment life prediction, most fail to consider the degradation process from a multi-scale perspective. Furthermore, existing multi-scale approaches often directly fuse features from multiple scales into a single feature representation for analysis. However, features at different scales may have distinct degradation characteristics, increasing the difficulty of prediction and reducing the accuracy of life prediction. In order to solve the above technical problems, this application provides a method for predicting the remaining life of used equipment based on multi-scale decomposition enhancement, such as Figure 2 As shown, the method includes:

[0081] S10: Acquire a parameter data set of the used device, and preprocess the parameter data set to obtain a target data set;

[0082] S20: Randomly select three different scales as pooling kernel sizes, and perform a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature;

[0083] S30: Determine a fusion feature according to the average pooling feature and the maximum pooling feature;

[0084] S40: Sort the fusion features corresponding to the three different scales from large to small according to the corresponding scales, and define them as a first fusion feature, a second fusion feature, and a third fusion feature; and determine a first enhanced feature, a second enhanced feature, and a third enhanced feature corresponding to the first fusion feature, the second fusion feature, and the third fusion feature after being enhanced, respectively;

[0085] S50: Obtaining, through a Bi-LSTM model, a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation;

[0086] S60: Determine a target gating weight according to the feature representation set, and determine a prediction result according to the first feature representation, the second feature representation, and the third feature representation;

[0087] S70: performing a weighted aggregation operation on the prediction results using the target gating weights to obtain intermediate prediction results corresponding to three scales; and determining a target predicted lifespan based on the intermediate prediction results.

[0088] The present invention preprocesses the parameter data set, randomly extracts three different scales as the pooling kernel size, obtains the average pooling feature and the maximum pooling feature, generates a fusion feature by combining the average pooling feature and the maximum pooling feature, and sorts them from large to small according to the corresponding scale, and defines them as the first fusion feature, the second fusion feature, and the third fusion feature; and determines the first enhanced feature, the second enhanced feature, and the third enhanced feature corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively; obtains the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively through the Bi-LSTM model; determines the target gating weight according to the feature representation set, and determines the prediction result according to the first feature representation, the second feature representation, and the third feature representation; performs a weighted aggregation operation on the prediction result by the target gating weight to obtain the intermediate prediction results corresponding to the three scales; and determines the target predicted life according to the intermediate prediction result. Since the features of different scales in the parameter data set may have obvious degradation characteristics, the remaining life of the equipment is accurately predicted.

[0089] In one embodiment, performing a pooling operation on the target data set in step S20 to obtain an average pooling feature and a maximum pooling feature includes:

[0090] S201: performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature;

[0091] S202: Perform a maximum pooling operation on the target data set to obtain the maximum pooling feature.

[0092] In one embodiment, determining the fusion feature according to the average pooling feature and the maximum pooling feature in step S30 includes:

[0093] S301: Concatenate the average pooling feature and the maximum pooling feature in the channel dimension to obtain a concatenated feature;

[0094] S302: operating the splicing features through a convolutional neural network and a sigmoid activation function to obtain an initial gating weight;

[0095] S303: Determine the fusion feature according to the average pooling feature, the maximum pooling feature and the initial gating weight.

[0096] In one embodiment, the first enhanced feature, the second enhanced feature, and the third enhanced feature corresponding to the first fusion feature, the second fusion feature, and the third fusion feature determined in step S40 after enhancement respectively include:

[0097] S401: performing multi-level fusion through the MLP layer according to the first fused feature, the second fused feature, and the third fused feature to update the second fused feature to obtain a second updated fused feature, and update the third fused feature to obtain a third updated fused feature;

[0098] S402: Performing an attention enhancement operation on the first fused feature, the third fused feature, and the third updated fused feature respectively to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature.

[0099] In one embodiment, for the step S10 of obtaining the parameter data set of the used device, preprocessing the parameter data set to obtain the target data set is achieved by the following expression:

[0100] X (1)

[0101] Among them, X is the target data set, X in is the parameter data set, is the mean of the parameter data set, is the standard deviation of the parameter data set;

[0102] In step S20, performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature; performing a maximum pooling operation on the target data set to obtain the maximum pooling feature is achieved by the following expression:

[0103] (2)

[0104] (3)

[0105] in, is the average pooling feature, is the maximum pooling feature, AvgPool1D represents the one-dimensional average pooling operation, and MaxPool1D represents the maximum pooling operation.

[0106] Specifically, the device mentioned in this application may be a sensor, and the parameter data set is specifically X in ∈ , where L represents the sequence length, which is the time series X in The length of C represents the number of feature channels; X in =[fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan inlet pressure, bypass duct pressure, high-pressure compressor outlet pressure, fan physical speed, core engine physical speed]. The present invention uses a kernel size k∈[5, 11, 17]. The three numbers are all medium-sized prime numbers and have no common divisors with each other. This can avoid periodic overlap in feature extraction at different scales and can capture information at different granularities.

[0107] In one embodiment, in steps S301 to S303, the average pooling feature and the maximum pooling feature are concatenated in the channel dimension to obtain a concatenated feature; the concatenated feature is operated by a convolutional neural network and a sigmoid activation function to obtain an initial gating weight; and the fusion feature is determined according to the average pooling feature, the maximum pooling feature, and the initial gating weight by the following expression:

[0108] ∈ (4)

[0109] (5)

[0110] (6)

[0111] in, Concat is a concatenation feature. is the average pooling feature, is the maximum pooling feature, is the initial gating weight, Represents the Sigmoid activation function, Conv1D represents the one-dimensional convolution operation, For fusion features.

[0112] Specifically, in this embodiment, a gating mechanism is introduced to learn a gating weight Control the contribution of average pooling features and maximum pooling features, so as to adaptively fuse the average pooling features and maximum pooling features. Specifically, firstly, the average pooling features and the maximum pooling features are spliced ​​in the channel dimension. Then, the convolutional neural network and the sigmoid activation function are used to realize the gated weight. The output of To fuse the average pooling features obtained by the one-dimensional average pooling operation and the maximum pooling features obtained by the maximum pooling operation. Through the above gating mechanism, the model can adaptively select the contribution of the average pooling features and the maximum pooling features to obtain a more expressive feature representation. At the same time, in order to reduce the boundary effect brought by the pooling operation, that is, the loss of boundary information, the input data is filled with reflection filling before the pooling operation.

[0113] In one embodiment, the multi-level fusion of the first fused feature, the second fused feature, and the third fused feature is performed through the MLP layer to update the second fused feature to obtain a second updated fused feature, and the third fused feature is updated to obtain a third updated fused feature; and the attention enhancement operation is performed on the first fused feature, the third fused feature, and the third updated fused feature to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature respectively, which are implemented by the following expressions:

[0114] , i=2, 3 (7)

[0115] , i=2, 3 (8)

[0116] (9)

[0117] When i is 2, That is , is the first mapping feature, That is , is the first fusion feature; when i is 3, That is , is the second mapping feature, That is , is the second fusion feature; when i is 2 or 3 They correspond to the second updated fusion feature and the third updated fusion feature respectively; when i takes the value of 1, 2 and 3, They correspond to the first enhancement feature, the second enhancement feature and the third enhancement feature respectively, For data format conversion operations, For attention-enhancing operations.

[0118] Specifically, in this embodiment, in order to better fuse features of different scales, the mechanism uses multiple MLP layers to gradually fuse features of different scales from coarse to fine. Specifically, first, the fusion features of the three scales are (i.e. the first fusion feature ), (i.e. the second fusion feature ), (i.e. the third fusion feature ); Sort by size from high to low , that is, starting from the coarsest features, step by step fusion to smaller scales, that is, more refined features, that is, Corresponding to the fusion feature of the largest scale, for the solution of the i-th scale, first use an MLP layer to process the previous fusion feature. Through this multi-level fusion method, the data can be gradually transferred from the coarse scale to the fine scale, thereby achieving more effective feature fusion, and when performing the next level of fusion, the scale fusion result of the previous level can be directly used. The purpose of the spatiotemporal feature extraction module is to process multi-scale feature data from the perspective of time and space. This module uses the attention mechanism and LSTM to extract temporal features and spatial features of features at multiple scales respectively. The first is the intra-scale feature attention layer, which enhances the attention of the features at each scale. Specifically, a multi-head attention mechanism with shared parameters is used to capture the relationship between different channels within each scale. For a given feature of a scale, where B represents the batch size, L represents the sequence length, and C represents the number of feature channels, first After transposition, the input is fed into the multi-head attention mechanism. It is worth noting that the multi-head attention layers used for data of different scales share parameters. This design is intended to ensure that, despite the different data scales, the dependencies between sensors, and therefore between channels, should be shared. Furthermore, data of different scales can implicitly interact through parameter sharing in the multi-head attention layers along the feature dimension.

[0119] In one embodiment, in steps S50 to S70, the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, are obtained by using the Bi-LSTM model; a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation is obtained; determining the target gating weight according to the feature representation set, and determining the prediction result according to the first feature representation, the second feature representation, and the third feature representation is achieved by the following expression:

[0120] (10)

[0121] (11)

[0122] (12)

[0123] A={ 、 、 } (13)

[0124] (14)

[0125] (15)

[0126] (16)

[0127] , i=1, 2, 3 (17)

[0128] , i=1, 2, 3 (18)

[0129] , i=1, 2, 3 (19)

[0130] in, is the local time feature, is the first enhancement feature, the second enhancement feature, or the third enhancement feature; is the global temporal feature of the i-th scale, is the first feature representation, the second feature representation or the third feature representation, A is the feature representation set, is the expert weight, is the learnable parameter matrix, is the first bias vector, is the mask matrix, is the target gating weight, is the transformation feature, is the mapping transformation feature, To predict the results, is the weight matrix, is the second bias vector.

[0131] Specifically, the temporal features of each scale are extracted. First, Bi-LSTM is used to extract the local features of the input data. For the input of the i-th scale , using bidirectional LSTM processing, the output of each Bi-LSTM The input is fed into the corresponding temporal multi-head attention submodule (TimeMHA) to further extract global temporal dependencies. The TimeMHA module maps the input into multiple subspaces and computes attention weights in parallel, capturing information from different representation subspaces. After obtaining the TimeMHA output, it is element-wise added to the Bi-LSTM output to fuse local and global temporal dependency information. The fused features are then average-pooled along the temporal dimension for dimensionality reduction, yielding the final feature representation at that scale.

[0132] In one embodiment, in step S70, performing a weighted aggregation operation on the prediction results using the target gating weights to obtain intermediate predicted lifespans corresponding to three scales; and determining the target predicted lifespan based on the intermediate prediction results is achieved by the following expression:

[0133] (20)

[0134] (twenty one)

[0135] in, is the intermediate prediction result, is the gating weight of the prediction result of the s-th expert on the i-th scale feature, represents the prediction result of the s-th expert for the i-th scale feature, represents the weight of the i-th scale, S represents the number of scales, Predict lifespan for a target.

[0136] In summary, this paper proposes a method for predicting remaining useful life based on enhanced multiscale decomposition. First, a multiscale feature extraction module decomposes the time series data of the device degradation process into multiple scales to obtain feature representations at different scales. In the spatiotemporal feature extraction layer, a feature attention layer with multiscale shared parameters is used to extract dependencies between channels while implicitly establishing connections between multiple scales. Subsequently, a bidirectional LSTM and temporal multi-head attention are used to learn local features and global long-term dependency features in the temporal dimension. Finally, a hybrid expert model is used to enhance the utilization of multiscale information. Gating is used to select appropriate predictors for information at different scales to obtain prediction results.

[0137] The present invention also provides a remaining life prediction device based on multi-scale decomposition enhancement, such as Figure 10 As shown, the device includes:

[0138] An acquisition module 10 acquires a parameter data set of the used device and preprocesses the parameter data set to obtain a target data set;

[0139] An operation module 20 is configured to randomly select three different scales as pooling kernel sizes, and perform a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature;

[0140] A first determining module 30 is configured to concatenate the average pooling feature and the maximum pooling feature in the channel dimension to obtain a concatenated feature; operate the concatenated feature using a convolutional neural network and a sigmoid activation function to obtain an initial gating weight; and determine the fused feature based on the average pooling feature, the maximum pooling feature, and the initial gating weight;

[0141] The sorting module 40 is configured to sort the fused features corresponding to the three different scales from large to small according to the corresponding scales, and define them as a first fused feature, a second fused feature, and a third fused feature; perform multi-level fusion through an MLP layer based on the first fused feature, the second fused feature, and the third fused feature to update the second fused feature to obtain a second updated fused feature, and update the third fused feature to obtain a third updated fused feature; perform an attention enhancement operation on the first fused feature, the third fused feature, and the third updated fused feature, respectively, to obtain the first enhanced feature, the second enhanced feature, and the third enhanced feature;

[0142] An enhancement module 50 is configured to obtain, through a Bi-LSTM model, a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation;

[0143] A second determination module 60 is configured to determine a target gating weight based on the feature representation set, and determine a prediction result based on the first feature representation, the second feature representation, and the third feature representation;

[0144] The third determination module 70 is configured to perform a weighted aggregation operation on the prediction results using the target gating weights to obtain intermediate prediction results corresponding to three scales; and determine a target predicted lifespan based on the intermediate prediction results.

[0145] The multiscale decomposition-enhanced remaining useful life prediction model (MSC-MoE) designed in this paper was thoroughly evaluated and validated on the C-MAPSS dataset. The C-MAPSS dataset consists of multiple sub-datasets. Two of these datasets, labeled Dataset 1 and Dataset 2, were selected for this study's experiments. Dataset information is shown in Table 1. Each sub-dataset contains a training set, a test set, and actual RUL values. The training set contains complete engine data from operation to failure, while the test set contains partial data from before failure. Each data set includes three operating settings and monitoring data from 21 sensors, which vary over time. In addition, three environmental parameters determine the specific operating conditions under which the engine operates. In the training set, each engine is initially healthy, but then a failure occurs at a random time step. The failure gradually worsens, leading to engine degradation and complete failure. The training set includes the operation of multiple engines, each in different degradation states: some engines are initially in a degraded state, while others are initially in a normal state.

[0146] Table 1 Dataset information

[0147]

[0148] (1) Data preprocessing: This paper uses a sliding time window to convert the original data sequence into a series of fixed-length subsequences. Each subsequence contains sensor data and corresponding RUL tags within a period of time. For each engine that runs to failure, a fixed-length time window is used to slide from the start point to the end point of the sequence. Specifically, assuming that the total operating cycle length of an engine is , set the length of the time window to T. The step size of the window sliding is fixed to 1. Each time it slides, the time window moves forward by one time unit. The remaining service life of the equipment within a time step can be expressed as formula (22):

[0149] (twenty two)

[0150] in , The time step is The remaining service life of each device can be expressed as formula (23):

[0151] (twenty three)

[0152] Most of the data in the test set are partial data during the device degradation process. Therefore, the data of the entire device may be less than the length of the time window. At this time, the data of the initial time step is used to fill the data from the head to generate a sample.

[0153] After using a time window to segment the device data into samples, the samples need to be segmented and modeled. This paper uses a linear degradation model to process the label portion of the data set, which can help the model converge quickly. The specific operation is as follows: a constant value Rmax is set to represent the remaining service life of a healthy engine. In the initial stage, the engine's RUL remains constant and then begins to decrease linearly from the time point when the remaining service life reaches Rmax. In this experiment, Rmax is set to 125, and the RUL calculation can be expressed as formula (24):

[0154] (twenty four)

[0155] (2) Model training: The overall training process of the model is as follows Figure 3 As shown, the MSP-MoE loss function consists of two parts: prediction loss and auxiliary loss, where the prediction loss measures the difference between the predicted lifespan value Y and the actual lifespan The auxiliary loss is the load balancing loss of the hybrid expert model. This study uses Huber Loss to calculate the prediction loss.

[0156] Huber Loss combines the advantages of MSE (mean square error) and MAE (mean absolute error). It is less sensitive to outliers than MSE, while maintaining a faster convergence speed when the error is small. For a single sample, the calculation of Huber Loss is shown in formula (25):

[0157] (25)

[0158] in, is the predicted loss, Y is the predicted life of MSP-MoE, which is a scalar, and is the true remaining life value, Is a hyperparameter that controls the threshold at which Huber Loss switches between MSE and MAE. When the absolute value of the prediction error is less than or equal to When the loss function is the mean square error, when the absolute value of the prediction error is greater than When , the loss function is the mean absolute error. The total loss It is the weighted sum of the prediction loss (i.e. Huber Loss) and the auxiliary loss, with the weight being , calculated as shown in formula (26):

[0159] (26)

[0160] (3) Case analysis: In order to more intuitively demonstrate the performance of the MSP-MoE proposed in this paper on C-MAPSS, Figure 4 and Figure 5 This report compares the actual and predicted remaining lifespans of individual devices on the Dataset 1 and Dataset 2 test sets. It shows that the error between the actual and predicted lifespans is within 20 cycles for most devices. Furthermore, when the remaining lifespan of a device is shorter, more accurate predictions are generally achieved, which helps to more accurately identify devices in poor health in real-world scenarios. Figure 6 and Figure 7 The figure shows the predicted and actual lifespans of device No. 34 in Dataset 1 and device No. 21 in Dataset 2 throughout their degradation processes. The MSP-MoE model demonstrates good performance in predicting the lifespan of individual devices throughout their degradation process. The degradation process for a single device exhibits similar patterns to the individual predictions for multiple devices. In the initial stages of normal operation, the device's operating status is relatively stable, and the lifespan predictions are relatively accurate. However, in the early stages of device degradation, the error between the predicted and actual lifespans increases due to changes in data characteristics, but the predicted and actual lifespan trends remain largely consistent. Towards the end of degradation, the error between the predicted and actual values ​​decreases, demonstrating that the model can accurately identify devices with serious issues. Furthermore, a comparison of the lifespan graphs for the entire degradation process using individual models shows that most predicted lifespans are shorter than the actual lifespans. This can be beneficial in real-world scenarios, enabling early replacement of aging equipment to prevent losses caused by premature failures.

[0161] In combination with the solution of the present invention, the experimental analysis is carried out as follows:

[0162] (1) Experimental setup: The MSP-MoE method was implemented using Python 3.11, PyTorch 2.2.2, and CUDA 12.2. All experiments were performed on an NVIDIA GeForce RTX 4090. The pooling kernel sizes were uniformly set to 17, 11, and 5, i.e., three sizes of pooling kernels were used for scale decomposition. The number of experts was 6, and the TopK was set to 2, i.e., two experts from the six experts were selected each time for lifespan prediction. The number of Bi-LSTM layers was set to 2, and the dimension of the hidden layer was set to 128. In the spatiotemporal feature extraction layer, the number of multi-scale shared feature multi-head attention layers was 10, and the number of temporal multi-head attention layers was 4. In this experiment, when splitting the training and test data, the time window size was set to 30 cycles, i.e., one sample was generated every 30 time steps, and the window step size was 1, i.e., each time a sample was split out, the time window was moved forward by one time step to obtain the next sample. Then, the maximum remaining lifespan Rmax was set to 125 and the data was normalized. After data preprocessing, the training data was divided into a training set and a validation set in a ratio of 8:2. In addition, this experiment used an early stopping strategy with a tolerance of 10. That is, if the model performance did not improve for 10 consecutive epochs on the validation set, early stopping would be triggered and training would be stopped, thus avoiding overfitting to a certain extent.

[0163] (2) Metrics: The present invention uses the root mean square error (RMSE) and the C-MAPSS dataset standard score (Score) as the evaluation indicators of the turbofan engine remaining useful life (RUL) prediction model. The calculation formulas are shown in formulas (27) and (28):

[0164] (27)

[0165] (28)

[0166] in It corresponds to Real lifespan, correspond The final prediction result, n is the number of test samples. In the scoring function, a smaller penalty is given to the case of early prediction, and the denominator is set to 13, while a larger penalty is given to the case of delayed prediction, and the denominator is set to 10, thereby encouraging the model to predict the remaining life more conservatively. In practical terms, it can encourage the prediction of shorter life, allowing production personnel to update equipment earlier and prevent losses caused by equipment damage.

[0167] (3) Experimental analysis: The present invention uses parameter sensitivity experiments to verify the influence of model parameters on the life prediction task. Using comparative experiments, the performance of the MSP-Moe model on the C-MAPSS dataset is compared with the SCTA-LSTM, IMDSSN, EAPN, and DVGTforemr baseline models. It can be seen that the present invention performs well on all evaluation indicators on the dataset. The effectiveness of each module is proved through ablation experiments, and the complete model can achieve the best performance.

[0168] 1) Parameter sensitivity experiment: In order to further explore the impact of MSP-Moe model parameters on life prediction tasks, this paper conducted a parameter sensitivity experiment on the model on Dataset1. The research objects were the number of bidirectional LSTM stacking layers in the spatiotemporal feature extraction layer and the total number of experts in the hybrid expert prediction module. The results are shown in Figure 2. Figure 8 and Figure 9 As shown in the figure, the performance of the model with 1, 2, 4, and 6 layers was studied. As the number of LSTM layers increased, the prediction error began to decrease, indicating that more LSTM layers can extract deeper degenerate features. The best model performance was achieved with 2 LSTM layers, but the prediction error continued to rise when the number of LSTM layers exceeded 2. This is primarily because LSTMs are used in the model to extract basic time series features, while the subsequent multi-head attention in the time dimension has better long-range dependency capabilities. Excessively deep LSTM layers are redundant, and more complex models based on time and channel dimensions not only increase the difficulty of training but also increase the risk of overfitting. Furthermore, for the hybrid expert prediction module, this study examined the impact of different total numbers of experts on model performance, while maintaining a TopK of 2. Too few experts fails to cover degenerate features at multiple scales, while too many experts results in a large number of idle parameters. Consequently, insufficient training data for each expert leads to overfitting, negatively impacting model performance.

[0169] 2) Comparative Experiments: To verify the performance of the MSP-MoE model, comparative experiments were conducted using the baseline models and our proposed model. The results are shown in Table 2, which shows the performance of MSP-MoE on Dataset 1 and Dataset 2 from CMAPSS. The bolded data indicates the best method, while the underlined data indicates the suboptimal method. MSP-MoE performed well on all evaluation metrics for both datasets.

[0170] Table 2 Performance comparison of each model on Dataset1 and Dataset2

[0171]

[0172] While the pure attention structure used by DAST and MTSTN can effectively capture global dependencies, it lacks explicit modeling of local temporal correlations. In the C-MAPSS dataset, short-term mutations in vibration signals during engine degradation are often critically linked to the degradation process, limiting the model's capabilities. The multi-scale blocks of MSIDSN only use sliding averages across different time windows and do not dynamically fuse high-frequency detail features and low-frequency trend features, which can cause high- and low-frequency information to interfere with each other. MSP-MoE's multi-scale interaction layer uses a gating mechanism and cascaded MLPs to fuse information at different scales. IMDSSN's use of two sparse strategies can potentially lead to the loss of local fine-grained features and compromise the integrity of multi-scale features. EAPN uses two RNN variants to extract temporal features, resulting in functional overlap. The single graph structure in DVGfomer cannot represent multi-scale spatial relationships. MLEAN processes features at each scale independently and ignores the connections between scale features. The MSP-MoE proposed in this study first explicitly fuses multiple scales from coarse-grained to fine-grained through cascading MLPs, and implicitly performs multi-scale interactions through a shared channel feature attention layer and an expert mixture model, which can obtain more accurate lifespan predictions from the connections between multiple scales.

[0173] Figure 11 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 11 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement the remaining life prediction method based on multi-scale decomposition enhancement. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement the remaining life prediction method based on multi-scale decomposition enhancement. It will be understood by those skilled in the art that Figure 11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0174] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0175] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0176] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for predicting the remaining life of used equipment based on multi-scale decomposition enhancement, characterized in that: The method comprises: Obtain the parameter data set of the used device, the used device is a turbofan engine, and preprocess the parameter data set to obtain a target data set; the parameter data set is specifically X in ∈ , X in = [Fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan inlet pressure, bypass duct pressure, high-pressure compressor outlet pressure, fan physical speed, core engine physical speed]; Randomly select three different scales as pooling kernel sizes, perform pooling operation on the target data set to obtain average pooling features and maximum pooling features; The average pooling feature and the maximum pooling feature are spliced ​​in the channel dimension to obtain the splicing feature; through the convolutional neural network and sigmoid An activation function operates the splicing feature to obtain an initial gating weight; a fusion feature is determined based on the average pooling feature, the maximum pooling feature and the initial gating weight; The fusion features corresponding to the three different scales are sorted from large to small according to the corresponding scales, and are defined as the first fusion feature, the second fusion feature and the third fusion feature; MLP The layer performs multi-level fusion according to the first fusion feature, the second fusion feature and the third fusion feature to update the second fusion feature to obtain the second updated fusion feature, and updates the third fusion feature to obtain the third updated fusion feature; the first fusion feature, the second updated fusion feature and the third updated fusion feature are respectively subjected to attention enhancement operation to obtain the first enhanced feature, the second enhanced feature and the third enhanced feature; the expression is as follows: ， i =2、3 ， i =2、3 in, i When the value is 2 ， That is , is the first mapping feature, That is , is the first fusion feature; i When the value is 3 ， That is , is the second mapping feature, That is , is the second fusion feature; i When the value is 2 or 3 They correspond to the second updated fusion feature and the third updated fusion feature respectively; i When the value is 1, 2, and 3, They correspond to the first enhancement feature, the second enhancement feature and the third enhancement feature respectively, For data format conversion operations, To enhance operations for attention; pass Bi-LSTM The model obtains a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation; Determining a target gating weight based on the feature representation set, and determining a prediction result based on the first feature representation, the second feature representation, and the third feature representation; The prediction results are weighted and aggregated by the target gating weights to obtain intermediate prediction results corresponding to three scales; and the target predicted life of the turbofan engine is determined based on the intermediate prediction results.

2. The method for predicting remaining useful life based on multi-scale decomposition enhancement according to claim 1, characterized in that: The performing a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature includes: Performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature; Performing a maximum pooling operation on the target data set to obtain the maximum pooling feature.

3. The method for predicting remaining useful life based on multi-scale decomposition enhancement according to claim 2, characterized in that: The acquisition of the parameter data set of the used device and the preprocessing of the parameter data set to obtain the target data set are achieved by the following expression: X in, X is the target data set, X in is the parameter data set, is the mean of the parameter data set, is the standard deviation of the parameter data set; Performing a one-dimensional average pooling operation on the target data set to obtain the average pooling feature; performing a maximum pooling operation on the target data set to obtain the maximum pooling feature is achieved by the following expression: in, is the average pooling feature, is the maximum pooling feature, AvgPool1D represents a one-dimensional average pooling operation, MaxPool1D Represents the max pooling operation.

4. The method for predicting remaining useful life based on multi-scale decomposition enhancement according to claim 1, characterized in that: The average pooling feature and the maximum pooling feature are spliced ​​in the channel dimension to obtain the splicing feature; through the convolutional neural network and sigmoid The activation function operates on the splicing feature to obtain the initial gating weight; the fusion feature is determined according to the average pooling feature, the maximum pooling feature and the initial gating weight by the following expression: ∈ in, is the splicing feature, Concat Represents the splicing operation, is the average pooling feature, is the maximum pooling feature, is the initial gating weight, represents the Sigmoid activation function, Conv1D represents a one-dimensional convolution operation, For fusion features.

5. The method for predicting remaining useful life based on multi-scale decomposition enhancement according to claim 1, characterized in that: Said through Bi-LSTM The model obtains the first feature representation, the second feature representation, and the third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation; determines the target gating weight according to the feature representation set, and determines the prediction result according to the first feature representation, the second feature representation, and the third feature representation through the following expression: A={ 、 、 } ， i =1、2、3 ， i =1、2、3 ， i =1、2、3 in, is the local time feature, is the first enhancement feature, the second enhancement feature, or the third enhancement feature; For the i The global temporal characteristics of the scale, is the first feature representation, the second feature representation or the third feature representation, A is the feature representation set, is the expert weight, is the learnable parameter matrix ， is the first bias vector, is the mask matrix, is the target gating weight, is the transformation feature, is the mapping transformation feature, To predict the results, is the weight matrix, is the second bias vector.

6. The method for predicting remaining useful life based on multi-scale decomposition enhancement according to claim 5, characterized in that: The weighted aggregation operation of the prediction results by the target gating weight is performed to obtain intermediate prediction results corresponding to three scales; and the target predicted life span is determined according to the intermediate prediction results by the following expression: in, is the intermediate prediction result, For the s Experts on i The gating weight of the prediction result of the scale feature, Representative s An expert on i The prediction results of scale features, Indicates the i The weight of the scale, S represents the number of scales, Predict lifespan for a target.

7. A remaining life prediction device based on multi-scale decomposition enhancement, characterized in that: The device comprises: An acquisition module acquires a parameter data set of a used device, wherein the used device is a turbofan engine, and pre-processes the parameter data set to obtain a target data set; the parameter data set is specifically X in ∈ , X in = [Fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan inlet pressure, bypass duct pressure, high-pressure compressor outlet pressure, fan physical speed, core engine physical speed]; An operation module, configured to randomly select three different scales as pooling kernel sizes, and perform a pooling operation on the target data set to obtain an average pooling feature and a maximum pooling feature; The first determination module is used to splice the average pooling feature and the maximum pooling feature in the channel dimension to obtain a spliced ​​feature; through a convolutional neural network and sigmoid An activation function operates the splicing feature to obtain an initial gating weight; a fusion feature is determined based on the average pooling feature, the maximum pooling feature and the initial gating weight; A sorting module is used to sort the fusion features corresponding to the three different scales from large to small according to the corresponding scales, and define them as the first fusion feature, the second fusion feature and the third fusion feature; MLP The layer performs multi-level fusion according to the first fusion feature, the second fusion feature and the third fusion feature to update the second fusion feature to obtain the second updated fusion feature, and updates the third fusion feature to obtain the third updated fusion feature; the first fusion feature, the second updated fusion feature and the third updated fusion feature are respectively subjected to attention enhancement operation to obtain the first enhanced feature, the second enhanced feature and the third enhanced feature; the expression is as follows: ， i =2、3 ， i =2、3 in, i When the value is 2 ， That is , is the first mapping feature, That is , is the first fusion feature; i When the value is 3 ， That is , is the second mapping feature, That is , is the second fusion feature; i When the value is 2 or 3 They correspond to the second updated fusion feature and the third updated fusion feature respectively; i When the value is 1, 2, and 3, They correspond to the first enhancement feature, the second enhancement feature and the third enhancement feature respectively, For data format conversion operations, To enhance operations for attention; Enhancement module for Bi-LSTM The model obtains a first feature representation, a second feature representation, and a third feature representation corresponding to the first enhanced feature, the second enhanced feature, and the third enhanced feature, respectively, and a feature representation set consisting of the first feature representation, the second feature representation, and the third feature representation; a second determination module, configured to determine a target gating weight according to the feature representation set, and determine a prediction result according to the first feature representation, the second feature representation, and the third feature representation; The third determination module is used to perform a weighted aggregation operation on the prediction results through the target gating weight to obtain intermediate prediction results corresponding to three scales; and determine the target prediction result of the turbofan engine based on the intermediate prediction results.

Images