Equipment residual life prediction method based on broadcast attention mechanism

Abstract

The invention discloses an equipment residual life prediction method based on a broadcast attention mechanism, and belongs to the technical field of equipment residual life prediction, and the method comprises the steps: collecting equipment operation state data; preprocessing the collected equipment operation state data; training a preset equipment residual life prediction model by using the preprocessed equipment operation state data; wherein the equipment residual life prediction model comprises a broadcast attention mechanism; and realizing equipment residual life prediction by using the trained equipment residual life prediction model. According to the method, the accuracy, the calculation efficiency and the robustness of the residual life prediction of the equipment can be improved, so that intelligent operation and maintenance and predictive maintenance can be better supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment residual life prediction, in particular to an equipment residual life prediction method based on a broadcast attention mechanism. BACKGROUND

[0002] In the whole life cycle of equipment, it usually goes through the state from normal to hidden degradation, obvious degradation, and then gradually deepens to the state of being unable to use or failure. By classifying and sorting different states, a system performance degradation state sequence is formed. Then, different coping schemes can be carried out according to different degradation states to reduce costs and improve stability. If the specific state of the equipment can be effectively evaluated, the state can be coped with, for example, no maintenance is needed for complete normal, regular inspection is needed for hidden degradation, and maintenance is needed for obvious degradation. These states are before the failure occurs, so the degradation evaluation can effectively prevent the occurrence of unexpected equipment failure, and the occurrence of serious failure also reduces the maintenance cost such as overhaul.

[0003] The residual life prediction method based on data needs to analyze the basic model of system operation data, and can accurately predict the residual life length on the basis of the prediction model. The reliability of this method is inseparable from a large amount of historical operation data, which needs to be obtained as a specific sample, and the accuracy of the prediction model is repeatedly adjusted. If the data-driven approach is adopted, this life prediction method can be divided into two common types of parameter method and non-parameter method. Among them, the parameter method needs to assume a certain distribution form first, then fits the historical data to obtain the parameters of the distribution, and uses these parameters to predict the future life. The non-parametric method directly analyzes the historical data without assuming a distribution form, and establishes a prediction model through statistical analysis of the historical data to predict the future life.

[0004] In order to predict the residual service life of the servo system equipment, a large number of sensors are usually used to collect temperature, vibration, current, voltage and other signal parameters, and these signal quantities are time series with time dependence. At present, long short-term memory network (LSTM) is widely used to capture long-term dependence in time series and has achieved great success in time series data analysis. For example, 3D convolution is integrated into LSTM, and CNN-based modules are used to decouple physical dynamics. However, although the convolution method is widely used, due to its local characteristics, there are inherent limitations in capturing spatiotemporal dependencies, which reduces the efficiency of capturing spatiotemporal dependencies, and there are still problems of unbalanced feature extraction, feature loss and low learning rate.

[0005] In summary, the existing equipment residual life prediction method mainly has the following defects:

[0006] Traditional statistical and physical model-based methods usually require assuming that the data follows a certain specific probability distribution (such as normal distribution, Weibull distribution, etc.), and rely on these distributions to establish a prediction model. However, in practical applications, the degradation process of equipment is often affected by various complex factors, including environmental conditions, load changes, and random disturbances, etc., which may cause dynamic changes in data distribution, making it difficult for fixed distribution assumptions to accurately fit the real life degradation process, and thus affecting the prediction accuracy.

[0007] In addition, existing methods still have certain deficiencies in feature extraction. For example, a single CNN model is mainly good at extracting local features, but due to its limited receptive field, it performs poorly in capturing long-term spatio-temporal dependencies, and may ignore some key global features. In addition, traditional feature extraction methods are prone to cause feature imbalance, and some key information may be weakened or lost, thereby affecting the overall performance of the prediction model, making the characterization of the device health state not comprehensive enough, and reducing the accuracy of the remaining life prediction. SUMMARY

[0008] The present application provides a device remaining life prediction method based on broadcast attention mechanism to solve the technical problem of low prediction accuracy of existing device remaining life prediction methods.

[0009] To solve the above technical problems, the present application provides the following technical solutions:

[0010] On the one hand, the present application provides a device remaining life prediction method based on broadcast attention mechanism, which comprises:

[0011] Collecting device operating state data;

[0012] Preprocessing the collected device operating state data; wherein the preprocessing includes data cleaning and denoising, data normalization and data enhancement;

[0013] Using the preprocessed device operating state data to train a pre-set device remaining life prediction model; wherein the device remaining life prediction model contains a broadcast attention mechanism;

[0014] Using the trained device remaining life prediction model to realize device remaining life prediction.

[0015] Further, the device operating state data includes temperature data, vibration data, current data and voltage data during device operation; the data format adopts multi-channel time series.

[0016] Further, the data cleaning and data denoising include:

[0017] missing data is filled using a sliding window mean;

[0018] an abnormal data point is removed using a One-Class SVM-based method;

[0019] data denoising is performed using wavelet transform filtering.

[0020] Further, the data augmentation includes:

[0021] training samples are generated using a fixed-length sliding window; and long time series are randomly cropped.

[0022] Further, the device remaining life prediction model includes a convolutional scalar long short-term memory neural network module, a broadcast attention module, a convolutional difference decoder module, a feature fusion module, and a fully connected layer.

[0023] The preprocessed device operating state data is input into the convolutional scalar long short-term memory neural network module to deeply mine the features between sequences and extract first feature information of the data; the first feature information is input into the broadcast attention module, and the output of the broadcast attention module is input into the convolutional difference decoder module to extract second feature information of the data; the feature fusion module fuses the first feature information and the second feature information to obtain fused feature information; and the fully connected layer realizes device remaining life prediction based on the fused feature information.

[0024] Further, the data processing process of the convolutional scalar long short-term memory neural network module includes:

[0025] First, the preprocessed device operating state data is processed by a convolutional operation to generate a feature map; then, one-dimensional maximum pooling is used to convert each kernel size of the feature map into a maximum number of single outputs; and then, for the output of the one-dimensional maximum pooling, a sLSTM network is used to capture deeper temporal and spatial feature information of the data.

[0026] Further, the broadcast attention module is composed of a 3x3 convolutional layer and a broadcast attention encoder, the activation function uses ReLU, and Batch Normalization is used for training optimization.

[0027] Further, the data processing process of the broadcast attention encoder includes:

[0028] First, three broadcast factors are initially set as m=3, n=5, k=7, respectively, corresponding to three different sizes of convolution kernels; then, the preset broadcast factor adaptive correction algorithm is used to dynamically adjust the numerical range of m, n, and k, so that the multi-scale receptive field can be flexibly adjusted when facing different input structures and scene distributions, thereby improving the adaptability and modeling ability of feature extraction;

[0029] The input of the broadcast attention encoder is first passed through an n x n convolution kernel in the main branch to capture feature information; the captured feature information is input into the multi-modal broadcast branch, which includes three branches of different scales, and the three branches of different scales use separable convolution operations; these branches are connected in series with convolution kernels of different sizes, and the scales corresponding to each branch are 1 x m→m x 1, 1 x n→n x 1, and 1 x k→k x 1, respectively;

[0030] The features from branches of different scales are element-wise added at the convergence node to fuse multi-scale information, and then a 1 x 1 convolution layer is used for dimension reduction or transformation in the channel dimension;

[0031] The output of the main branch is directly connected to the final 1 x 1 convolution layer using an addition operation, so that the output of the main branch is added to the output of the multi-modal broadcast branch;

[0032] The multiplication connection is used for feature weighting and selection, so that the input is directly multiplied by the final output element by element to realize the gating mechanism, control the transmission of feature information, enhance important features, and suppress irrelevant features.

[0033] Further, the process of dynamically adjusting the numerical range of m, n, and k by the broadcast factor adaptive correction algorithm includes:

[0034] In each training epoch, the average activation strength and gradient contribution of the output features of each scale branch in the multi-modal broadcast branch are counted;

[0035] The branches of different scales are sorted according to the gradient contribution, the kernel size of the branch with a response lower than the preset threshold is reduced, and the kernel size of the branch with a response higher than the preset threshold is enlarged to form adaptive scale evolution;

[0036] The exponential smoothing update formula is used to realize stable convergence and ensure the continuity and differentiability in the training process;

[0037] The corrected m, n, and k are used as a reference for the design of the convolution kernel in the next stage, or used in the search optimization stage.

[0038] Further, the convolution difference decoder module comprises four convolution difference decoders; the convolution difference decoder comprises a bilinear interpolation module and a 3*3 convolution module;

[0039] The data processing process of the convolution difference decoder module comprises:

[0040] The output of the broadcast attention encoder is bilinearly interpolated by the first convolution difference decoder to realize up-sampling, and a 3*3 convolution is used to fuse the up-sampled result with the input of the broadcast attention encoder as the output of the first convolution difference decoder; after the first convolution difference decoder, each convolution difference decoder respectively bilinearly interpolates the output of the broadcast attention encoder to realize up-sampling, and a 3*3 convolution is used to fuse the output of the broadcast attention encoder and the output of the previous convolution difference decoder; finally, point-wise convolution is used to fuse the features from the last three convolution difference decoders.

[0041] In another aspect, the present application also provides an electronic device comprising a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to realize the above method.

[0042] In another aspect, the present application also provides a computer readable storage medium, which stores at least one instruction, which is loaded and executed by the processor to realize the above method.

[0043] The technical scheme provided by the present application has at least the following beneficial effects:

[0044] Firstly, the present application adopts a convolution scalar long short-term memory network, combines 1D-CNN for local feature extraction, and models the spatial and temporal dependence relationship through sLSTM, thereby improving the long-term dependence modeling capability and prediction accuracy of the device life prediction; secondly, the present application introduces a broadcast factor adaptive correction algorithm, which can dynamically adjust different scale convolution kernel parameters (such as m, n, k) according to the feature response, to realize adaptive matching of the receptive field. This mechanism enhances the flexibility and robustness of the model in multi-scale modeling, making the network more stable and efficient when processing input features of different sizes, structures and complexities. Thirdly, the present application designs a broadcast attention encoder, which uses the broadcast factor adaptive correction algorithm, deep convolution, multi-scale strip convolution and 1*1 convolution to construct an efficient attention mechanism, improves the global feature extraction capability, realizes linear complexity calculation, balances information integrity and calculation efficiency, and combines a convolution interpolation decoder for feature reconstruction optimization, effectively improves the feature expression capability, ensures the stability and robustness of the prediction, and is easy to deploy in practice.

[0045] Through the above improvement, the application is superior in enhancing the accuracy, calculation efficiency and application adaptability of the remaining life prediction of the equipment, and provides efficient and reliable technical support for intelligent operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 is a flow chart of the device remaining life prediction method based on the broadcast attention mechanism provided by the embodiment of the present application;

[0048] Figure 2 is a schematic diagram of the convolutional scalar long short-term memory neural network module provided by the embodiment of the present application;

[0049] Figure 3 is a schematic diagram of the broadcast attention encoder provided by the embodiment of the present application;

[0050] Figure 4 is a schematic diagram of the broadcast attention module provided by the embodiment of the present application;

[0051] Figure 5 is a schematic diagram of the convolutional interpolation decoder module provided by the embodiment of the present application;

[0052] Figure 6 is a system block diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0054] First of all, it should be noted that in the embodiments of the present application, the words such as "exemplarily", "for example" are used to represent as an example, illustration or description. Any embodiment or design scheme described as "exemplary" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "exemplarily" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0055] First embodiment

[0056] In view of the deficiencies of existing equipment residual life prediction methods in terms of spatio-temporal feature extraction, feature fusion and computational complexity, the embodiment provides an equipment residual life prediction method based on a broadcast attention mechanism, aiming to build an efficient, accurate and deployable equipment residual life prediction method to improve the accuracy, computational efficiency and robustness of equipment residual life prediction, thereby reducing the risk of equipment failure, reducing maintenance costs and improving the intelligent level of equipment management, and thus better supporting intelligent operation and maintenance and predictive maintenance.

[0057] As shown in Figure 1 The implementation process of the method is as follows: first, the running state data of the equipment is collected by various sensors (temperature, vibration, current, voltage, etc.), and data cleaning, denoising, normalization and data enhancement are performed to improve the data quality and model generalization ability. Second, the CNN-sLSTM module is designed, which uses 1D-CNN to extract local features and combines sLSTM for spatio-temporal feature modeling to improve the learning ability of long-term dependency information. Third, the broadcast factor adaptive correction algorithm is designed, which dynamically adjusts the value range of the broadcast factors m, n and k by analyzing the weight distribution of different scale broadcast branches on the feature response, realizing the adaptive matching of the receptive field; then, the broadcast attention encoder is introduced, which realizes multi-scale feature extraction through deep convolution, multi-scale strip convolution and 1x1 convolution, enhances the fusion of local and global information, and reduces the computational complexity. Further, the convolution interpolation decoder is constructed, which uses bilinear interpolation and 3x3 convolution for upsampling and feature reconstruction to optimize the feature expression ability of the model, improve the information integration ability, prevent feature loss and improve the stability of the prediction. During the training process, the weighted mean square error (WMSE) is used to calculate the RUL error, and the training rounds are dynamically adjusted to obtain the optimal model.

[0058] The method can be implemented by an electronic device, and specifically, the execution process includes the following steps:

[0059] S1, collecting equipment running state data;

[0060] It should be noted that the equipment health monitoring data is usually collected by multiple sensors, such as temperature sensors to monitor the heating condition of the equipment, vibration sensors to detect mechanical structure abnormalities, current / voltage sensors to monitor the electrical state of the motor or electronic equipment, and the data format adopts multi-channel time series, i.e.: X = [X1, X2,..., Xn]T, where X1, X2,..., Xn represent the n-dimensional sensor features at different time points. t ], where X i represents the n-dimensional sensor feature at the i-th time point.

[0061] S2, preprocessing the collected equipment running state data; wherein the preprocessing includes data cleaning and denoising, data normalization and data enhancement;

[0062] Specifically, in the present embodiment, the data preprocessing process is as follows:

[0063] S21, data cleaning and denoising: missing data is filled using sliding window mean; abnormal data points are removed using a One-Class SVM-based method; denoising is performed using wavelet transform (Wavelet Transform) filtering.

[0064] S22, data normalization and standardization: due to the different dimensions of different sensors, Z-score standardization is used, which is expressed by the formula:

[0065]

[0066] S23, data augmentation and dataset generation: in order to increase the amount of training data and improve the robustness of the model, the following methods are used for data augmentation: training samples are generated using fixed-length sliding windows (such as 50 or 100 steps); long time series are randomly cropped to improve the generalization ability of the model.

[0067] S3, using the preprocessed equipment operating state data to train a pre-set equipment remaining life prediction model; wherein the equipment remaining life prediction model comprises a broadcast attention mechanism;

[0068] S4, using the trained equipment remaining life prediction model to realize equipment remaining life prediction;

[0069] It should be noted that in order to improve the accuracy of equipment remaining life prediction, the present embodiment proposes a model CNN-sLSTM composed of CNN and sLSTM to deeply mine the features between sequences, reduce data noise, and improve prediction accuracy. Specifically, the model extracts local time series features through 1D-CNN, and uses sLSTM to deeply model the spatiotemporal dependence relationship and enhance the learning ability of long-term dependence information. In addition, the present embodiment also designs a broadcast factor adaptive correction algorithm, a broadcast attention encoder, and a convolution interpolation decoder, uses multi-scale strip convolution and 1x1 convolution to construct an efficient attention mechanism, improves the global feature extraction capability while reducing the computational complexity, realizes linear complexity calculation, improves the real-time performance, and combines the convolution interpolation decoder for feature reconstruction optimization, strengthens the interaction and reconstruction between different features, avoids information loss, balances information integrity and computational efficiency, and improves the stability and robustness of the model.

[0070] Specifically, in the present embodiment, the construction process of the equipment remaining life prediction model is as follows:

[0071] ​Step 1, design a convolutional scalar long short-term memory neural network module (CNN-sLSTM): use 1D-CNN for time series feature extraction, h t = f(W·X t +b), where X t is the input time step data, W and b are the weights and biases of the CNN, and f(·) is the ReLU activation function. The CNN model can effectively reduce the number of features to prevent overfitting, while reducing the time and parameter complexity, as shown in Figure 2 The convolution operation first produces a feature map, and then the pooling layer extracts the most important features to calculate the local sufficient statistics. One-dimensional max pooling converts each kernel size of the input to the maximum number of single outputs to reduce or downsample the version of the input, followed by an sLSTM network to capture deeper temporal and spatial feature information of the data.

[0072] Step 2, design a broadcast factor adaptive tuning algorithm (BFAT): during network training, to further improve the model's adaptability to different tasks and input features, a broadcast factor adaptive tuning algorithm is introduced. This algorithm analyzes the weight distribution of different scale broadcast branches on feature responses, dynamically adjusts the numerical range of broadcast factors m, n, k, to improve the matching ability of the receptive field. Specifically:

[0073] 1) Within each training epoch, the average activation intensity and gradient contribution of the output features of each scale broadcast branch are counted;

[0074] 2) According to the contribution ranking, the response of the lower branch is reduced, and the kernel size of the higher branch is enlarged, forming an adaptive scale evolution;

[0075] 3) The adjustment process is realized by an exponential smoothing update formula to ensure stability and convergence, ensuring continuity and differentiability during training;

[0076] 4) The modified m, n, k will be used as a reference for the design of the next stage of convolution kernel, or further used in the search optimization stage.

[0077] Step 3, design a broadcast attention encoder (BAEncoder): to extract multi-scale features and fuse their respective feature domains, a broadcast attention encoder is designed, which adopts a pyramid structure, as shown in Figure 3As shown, it consists of four parts: the broadcast factor adaptive correction algorithm designed in step 2, deep convolution for aggregating local information, deep strip convolution for capturing multi-scale broadcast branches, and 1x1 convolution for modeling the relationship between different channels. Finally, a double residual connection is adopted, and the output of the 1x1 convolution is directly used as the attention weight to reweight the input of the BA. Specifically:

[0078] 1) Broadcast factor correction: Initially, three broadcast factors are set as m = 3, n = 5, and k = 7, respectively, corresponding to three different scale strip convolution kernels. The broadcast factor adaptive correction algorithm in step 2 is used to dynamically adjust the value range of m, n, and k to achieve flexible adjustment of multi-scale receptive fields when facing different input structures and scene distributions, thereby improving the adaptability and modeling ability of feature extraction.

[0079] 2) Main branch: The input features first pass through an n x n convolution kernel to capture feature information within a larger receptive field.

[0080] 3) Multi-modal broadcast branch:

[0081] a) Through three branches of different scales (1x m→m x 1, 1x n→n x 1, and 1x k→k x 1), separable convolution operations are used to reduce computational cost and improve feature expression ability.

[0082] b) These branches effectively decompose large-scale convolution kernels by concatenating different size convolution kernels, making the calculation more efficient while still retaining the characteristics of a larger receptive field.

[0083] 4) Feature fusion:

[0084] a) The features from different paths are element-wise added at the convergence node to fuse multi-scale information.

[0085] b) Then, a 1x1 convolution layer is used for dimension reduction or transformation in the channel dimension.

[0086] 5) Residual connection:

[0087] a) A jump connection (Residual Connection) is adopted for gradient propagation and information reuse. The output of the main branch is directly connected to the final 1x1 convolution using addition operation. This connection adds the output of the main branch to the result extracted by the multi-modal broadcast branch, thereby enhancing gradient flow and preserving input information, and alleviating the gradient vanishing problem during deep network training.

[0088] b) using a multiplicative connection to weight and select features, so that the input is directly multiplied element-wise with the final output, realizing a gating mechanism to control the transmission of feature information, enhance important features and suppress irrelevant features.

[0089] Step 4, constructing the broadcast attention module: using the broadcast attention encoder constructed in step 3, the broadcast attention module is composed of a 3x3 convolution layer and a broadcast attention encoder, as shown in Figure 4 The input data is processed in turn through the 3x3 convolution layer and the broadcast attention encoder. The activation function uses ReLU, and Batch Normalization (BN) is used for training optimization.

[0090] Step 5, designing a convolution interpolation decoder (CIDecoder): the CI module contains two parts: bilinear interpolation of the output features from the encoder in step 3 to realize upsampling, and using a 3x3 convolution to fuse information from the encoder and the previous CI module. Finally, pointwise convolution is used to fuse features from the last three CI modules, as shown in Figure 5 .

[0091] Step 6, feature fusion: the feature information obtained from the above steps is jointly input into the feature fusion module for weighted feature fusion.

[0092] Step 7, training and testing of the model. Given a training sample, the data is preprocessed by S2, and each sample data format uses a multi-channel time series. The preprocessed data is first input into the CNN-sLSTM module constructed in step 1 to preliminarily extract local features and long sequence modeling to extract spatiotemporal features; then the output features are input into the broadcast attention module constructed in step 4 to extract multi-scale features and fuse their respective feature domains; the output obtained is then input into the convolution feature decoder constructed in step 5 to optimize feature reconstruction and integrate multi-scale feature information; next, the spatiotemporal features obtained by the CNN-sLSTM module and the features obtained by the broadcast attention module and the convolution interpolation decoder are jointly input into the feature fusion module for weighted feature fusion; finally, the RUL error is calculated using the weighted mean square error (WMSE), and the error formula is as follows, and the test set is tested, and the training rounds are dynamically adjusted according to the stability of the error.

[0093]

[0094] In summary, the embodiment provides a device residual life prediction method based on a broadcast attention mechanism, aiming at the problem that the traditional time series modeling method is difficult to fully capture long and short term dependencies, resulting in limited prediction accuracy, the embodiment designs a model Conv-sLSTM composed of CNN and sLSTM, combines 1D-CNN for local feature extraction, and learns deep spatio-temporal features through sLSTM, effectively enhancing the modeling capability of long-term dependency information; In view of the problem that the fixed scale convolution has insufficient receptive field adaptability when processing diversified input, the embodiment proposes a broadcast factor adaptive correction algorithm, which can dynamically adjust the size of the convolution kernel according to the feature response and gradient contribution of different scale broadcast branches in the training process. This strategy enables the model to flexibly match the receptive field range, thereby more effectively modeling multi-scale feature information; In view of the problem that the feature extraction algorithm of the traditional framework exists down-sampling leading to information loss and high computational complexity, the embodiment proposes a broadcast attention encoder, which uses the broadcast factor adaptive correction algorithm, deep convolution, multi-scale bar convolution and 1x1 convolution to construct an attention mechanism, realizes efficient feature extraction, and balances information retention and computational efficiency, and further optimizes information reconstruction through a convolution interpolation decoder, improves the feature expression capability, and ensures the accuracy and stability of the prediction.

[0095] Based on the above, the method realizes efficient, accurate and robust device residual life prediction through multi-level feature extraction, fusion and optimization mechanism, effectively improves the accuracy, computational efficiency and application robustness of device life prediction, and provides efficient and reliable technical support for industrial equipment intelligent operation and maintenance and predictive maintenance.

[0096] Second embodiment

[0097] The embodiment provides an electronic device, such as Figure 6 As shown, the electronic device includes a processor and a memory; wherein the processor and the memory can be connected through a communication bus; the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment. In addition, the electronic device can also include a transceiver, and the processor and the transceiver can be connected through a communication bus, and the transceiver is used for communication with other devices.

[0098] Next, combined with Figure 6 The various components of the electronic device will be described in detail:

[0099] The processor is the control center of the electronic device. The electronic device can include multiple processors. Each of the processors can be a single-CPU or a multi-CPU. The processor can be one processor or a collective term of multiple processing elements. For example, the processor can be one or more central processing units (CPUs), other general purpose processors, application specific integrated circuits (ASICs), or one or more integrated circuits configured to implement one or more embodiments of the present application, such as one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. The general purpose processor can be a microprocessor or any conventional processor, or the like. The processor can perform various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0100] In a specific implementation, as an embodiment, the processor can include one or more CPUs, such as CPU0 and CPU1 shown in FIG. 1, of course, this is only an exemplary description. Figure 6

[0101] The memory is used to store software programs for implementing the solution of the present application, and is controlled by the processor to perform the implementation. The specific implementation can refer to the above-mentioned method embodiments, which will not be described here.

[0102] ​Optionally, the memory may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may be integrated with the processor or exist independently and accessed through the interface circuit ( Figure 6 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0103] The transceiver may include a receiver and a transmitter ( Figure 6 The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function. The transceiver can be integrated with the processor or exist independently and communicate with the electronic device through the interface circuit ( Figure 6 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0104] In addition, it should be noted that Figure 6 The structure of the electronic device shown in the figure does not constitute a limitation on the device. The actual device may include more or fewer components than shown, or may combine certain components, or arrange the components differently. In addition, the technical effects achieved by the electronic device when executing the method of the first embodiment can refer to the technical effects described in the first embodiment above, and therefore will not be repeated here.

[0105] Third embodiment

[0106] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device. The instructions stored therein can be loaded by a processor in a terminal to execute the method described above.

[0107] Moreover, it should be noted that the present application can be provided as a method, an apparatus, or a computer program product. Therefore, the embodiments of the present application can take the form of an entirely or partially hardware embodiment, an entirely or partially software embodiment, or an embodiment combining software and hardware aspects. Furthermore, when implemented in software, the embodiments of the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, a computer diskette, an optical storage medium, a magnetic storage medium, and a semiconductor memory device). The computer program product includes one or more computer instructions that when loaded and executed by a computer, cause the computer to carry out the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, such as from a website, a computer, a server, or a data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, or the like) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, data center, or the like, including one or more collections of available media. The available media can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0108] The embodiments of the present application are described with reference to flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate a device that implements the flowcharts and / or block diagrams. Figure 1 The flowcharts and / or block diagrams Figure 1 The apparatus that implements the functions specified in one or more flows and / or blocks.

[0109] These computer program instructions can also be stored in a computer-readable storage medium that can direct the computer or other programmable data processing terminal device to work in a specific manner, so that the instructions stored in the computer-readable storage medium produce a manufactured product that includes instruction apparatus, which implements the flowcharts and / or block diagrams. Figure 1 The flowcharts and / or block diagrams Figure 1the functions specified in the individual block or blocks. Such computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable devices to generate a computer-implemented process, thus the instructions executed on the computer or other programmable devices provide a process for implementing the functions specified in the flowchart block(s). Figure 1 the functions specified in the individual block or blocks. Such computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable devices to generate a computer-implemented process, thus the instructions executed on the computer or other programmable devices provide a process for implementing the functions specified in the flowchart block(s). Figure 1 the functions specified in the individual block or blocks. Such computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable devices to generate a computer-implemented process, thus the instructions executed on the computer or other programmable devices provide a process for implementing the functions specified in the flowchart block(s).

[0110] It should also be noted that, in the present document, the terms such as first and second, etc. are merely used to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or terminal device. Without more limitations, the element defined by the statement "including a…", does not exclude the presence of other identical elements in the process, method, article or terminal device including the element. In addition, the term "and / or" is merely a description of the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone, where A and B can be singular or plural. In addition, the character " / " in the present document generally represents an "or" relationship between the front and rear associated objects, but it can also represent an "and / or" relationship, which can be understood in the context before and after. "One or more" means one or more, and "multiple" means two or more. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can mean a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be single or multiple.

[0111] In addition, it can be understood that in various embodiments of the present application, the size of the sequence number of the above processes does not mean the order of execution, and the execution order of the processes should be determined by their functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0112] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware or in a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0113] In several embodiments provided by the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely schematic, for example, the division of functional modules / units is merely a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another device, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms. The units described as separate components can be or can not be physically separated, and the components displayed as units can be or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present, or two or more units can be integrated in one unit.

[0114] If the method is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0115] Finally, it should be noted that the above description is only the preferred embodiment of the application, it should be pointed out that although the preferred embodiment of the application has been described, for those skilled in the art, once the basic creative concept of the application is known, several improvements and refinements can be made without departing from the principles of the application, and these improvements and refinements should also be considered as the protection scope of the application. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the application.

Claims

1. A device remaining life prediction method based on broadcast attention mechanism, characterized in that: include: Collect equipment operating status data; Preprocess the collected equipment operation status data; preprocessing includes: data cleaning and data denoising, data normalization, and data enhancement; The pre-processed equipment operating status data is used to train a preset equipment remaining life prediction model; wherein the equipment remaining life prediction model includes a broadcast attention mechanism; The remaining life of the equipment is predicted using the trained equipment remaining life prediction model.

2. The device remaining life prediction method based on the broadcast attention mechanism according to claim 1, characterized in that: The equipment operation status data includes temperature data, vibration data, current data and voltage data when the equipment is running; the data format adopts a multi-channel time series.

3. The device remaining life prediction method based on the broadcast attention mechanism according to claim 1, characterized in that: The data cleaning and data denoising include: Use sliding window mean to fill missing data; A method based on One-Class SVM is used to eliminate abnormal data points; Wavelet transform filtering is used to denoise the data.

4. The device remaining life prediction method based on the broadcast attention mechanism according to claim 1, characterized in that: The data enhancement includes: Use a fixed-length sliding window to generate training samples; randomly crop time series whose length exceeds a preset value.

5. The device remaining life prediction method based on the broadcast attention mechanism according to claim 1, characterized in that: The equipment remaining life prediction model includes: a convolutional scalar long short-term memory neural network module, a broadcast attention module, a convolutional difference decoder module, a feature fusion module and a fully connected layer; The preprocessed equipment operation status data is input into the convolutional scalar long short-term memory neural network module to deeply mine the features between sequences and extract the first feature information of the data; the first feature information is input into the broadcast attention module, and the output of the broadcast attention module is input into the convolutional difference decoder module to extract the second feature information of the data; the feature fusion module fuses the first feature information and the second feature information to obtain fused feature information; the fully connected layer realizes the remaining life prediction of the equipment based on the fused feature information.

6. The device remaining life prediction method based on the broadcast attention mechanism according to claim 5, characterized in that: The data processing process of the convolutional scalar long short-term memory neural network module includes: First, the preprocessed device operating status data is processed by a convolution operation to generate a feature map; then, one-dimensional maximum pooling is used to convert each kernel size of the feature map into a single output with the maximum number; then, the sLSTM network is used to capture the deeper spatiotemporal feature information of the data based on the output of the one-dimensional maximum pooling.

7. The device remaining life prediction method based on the broadcast attention mechanism according to claim 5, characterized in that: The broadcast attention module consists of a 3×3 convolutional layer and a broadcast attention encoder. The activation function uses ReLU and Batch Normalization is used for training optimization.

8. The device remaining life prediction method based on the broadcast attention mechanism according to claim 7, characterized in that: The data processing of the broadcast attention encoder includes: First, the three broadcast factors are initially set as: m = 3, n = 5, and k = 7, corresponding to three strip convolution kernels of different scales. The three broadcast factors are adaptively adjusted during the training process. The input of the broadcast attention encoder first passes through an n×n convolution kernel in the main branch to capture feature information. The captured feature information is then input into the multimodal broadcast branch, which consists of three branches of different scales. Each of the three branches of different scales uses a separable convolution operation. These branches are connected in series with convolution kernels of different sizes, and the corresponding scales of each branch are: 1×m→m×1, 1×n→n×1, and 1×k→k×1. The features from branches of different scales are summed element-by-element at the aggregation node to fuse multi-scale information. After that, they pass through a 1×1 convolutional layer to reduce or transform the channel dimension. Skip connections are used for gradient propagation and information reuse. The output of the backbone branch is directly connected to the final 1×1 convolutional layer, and an addition operation is used to add the output of the backbone branch to the output of the multimodal broadcast branch. Multiplication connection is used for feature weighting and selection, so that the input is directly multiplied by the final output element by element, and a gating mechanism is implemented to control the transmission of feature information, enhance important features and suppress irrelevant features.

9. The device remaining life prediction method based on the broadcast attention mechanism according to claim 8, characterized in that: The process of adaptive dynamic adjustment of the broadcast factor during training includes: In each training epoch, the average activation intensity and gradient contribution of the branch output features at each scale in the multimodal broadcast branch are counted; The branches of different scales are sorted according to their gradient contributions. The convolution kernel size is reduced for branches whose responses are lower than a preset threshold, and the kernel size is enlarged for branches whose responses are higher than the preset threshold, thus forming an adaptive scale evolution. Stable convergence is achieved through exponential smoothing update formula, ensuring continuity and differentiability during training; The modified m, n, and k are used as a reference for the design of the convolution kernel in the next stage, or used in the search and optimization stage.

10. The device remaining life prediction method based on the broadcast attention mechanism according to claim 7, characterized in that: The convolution difference decoder module includes four convolution difference decoders; the convolution difference decoder includes a bilinear interpolation module and a 3×3 convolution module; The data processing process of the convolution difference decoder module includes: The output of the broadcast attention encoder is bilinearly interpolated by the first convolutional difference decoder to achieve upsampling, and the upsampling result is fused with the input of the broadcast attention encoder using 3×3 convolution as the output of the first convolutional difference decoder; after the first convolutional difference decoder, each convolutional difference decoder performs bilinear interpolation on the output of the broadcast attention encoder to achieve upsampling, and uses 3×3 convolution to fuse the output of the broadcast attention encoder and the output of the previous convolutional difference decoder; finally, point-by-point convolution is used to fuse the features from the last three convolutional difference decoders.

Images