A device residual life prediction method, device, apparatus and storage medium

Abstract

The application discloses a device residual life prediction method and device, equipment and storage medium, relates to the technical field of device management, and comprises the following steps: obtaining operation data of a target industrial device to be predicted at present, and obtaining current operation data; inputting the current operation data into a residual life prediction model comprising a multiscale hollow convolution module, a sparse Transformer encoder, an adaptive gating unit and a residual life predictor, so as to extract transient degradation features caused by sudden physical impact in the device operation process based on the current operation data, obtain local transient impact features, and extract performance degradation trends changing with time to obtain global degradation trend features; based on the statistical characteristics of the operation data, the local transient impact features and the global degradation trend features are dynamically fused, and residual life prediction is performed based on the fused degradation features. The application can improve the accuracy of industrial equipment residual life prediction.

Description

Technical Field

[0001] This application relates to the field of equipment management technology, and in particular to a method, apparatus, device and storage medium for predicting the remaining life of equipment. Background Technology

[0002] Predicting the remaining useful life (RUL) of industrial equipment is a core aspect of achieving health management and failure prediction for industrial equipment. By analyzing the equipment's operational data throughout its entire lifecycle, the remaining operating time of the equipment can be predicted, thereby providing a basis for decision-making regarding preventive maintenance, spare parts management, and operational safety assurance.

[0003] Currently, the mainstream approach is to predict the remaining life of equipment based on deep learning models. However, this method has the following shortcomings in complex industrial scenarios, which limit the accuracy and generalization performance of the remaining life prediction: 1. Insufficient feature extraction capability: Current prediction models based on CNN (Convolutional Neural Network) or LSTM (Long Short-Term Memory) can extract local features or process time-series data, but CNN is limited by its receptive field and has difficulty capturing long-term degradation trends (such as slow wear) throughout the equipment's life cycle; while pure Transformer models can capture global dependencies, they easily ignore microscopic local features caused by instantaneous impacts (such as electric arcs, sudden vibrations), thus leading to a decrease in prediction accuracy. 2. Poor generalization ability across operating conditions: Current deep learning models usually assume that the training data (source domain, such as laboratory data) and test data (target domain, such as field data) have the same distribution. However, in actual industrial scenarios, equipment operating conditions (such as load, speed) are varied, resulting in significant differences in data distribution, which greatly reduces the prediction accuracy of the model. 3. Simple feature fusion mechanism: Current hybrid models mostly adopt serial structure or simple feature splicing, which fail to dynamically adjust the contribution weight of local and global features according to the different degradation stages of the device (such as early, middle and late stages). This results in low feature utilization efficiency and difficulty in accurately depicting the degradation trajectory of the entire life cycle, thereby limiting the model's representation ability and prediction robustness. Furthermore, the lack of adaptive fusion strategies leads to low prediction accuracy. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, apparatus, device, and storage medium for predicting the remaining life of equipment, which can improve the accuracy of predicting the remaining life of industrial equipment. The specific solution is as follows: In a first aspect, this application discloses a method for predicting the remaining useful life of equipment, including: Obtain the current operating data of the target industrial equipment to be predicted; The current operating data is input into the trained remaining lifetime prediction model. The remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets. The multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture. The remaining lifetime predictor consists of multiple fully connected layers. The historical dataset includes laboratory aging data and actual operating data of different industrial equipment. Using the multi-scale dilated convolution module and based on the current operating data, the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment are extracted to obtain local transient impact features. At the same time, using the sparse Transformer encoder and based on the current operating data, the performance degradation trend of the target industrial equipment over time is extracted to obtain global degradation trend features. The local transient impact features and the global degradation trend features are dynamically fused using the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features. The remaining life of the target industrial equipment is predicted using the remaining life predictor and based on the fused post-degradation features, resulting in the current prediction result.

[0005] Optionally, before obtaining the current operating data of the target industrial equipment to be predicted, the method further includes: Collect laboratory aging data from different industrial equipment in the source domain and actual operating data from different industrial equipment in the target domain; The laboratory aging data and the actual operating data are input into a generative adversarial network (GAN) containing a feature extractor, an adaptive gated recurrent unit, a gradient inversion layer, a domain discriminator, and a remaining lifetime predictor. The GAN is then adversarially trained based on the laboratory aging data and the actual operating data, using label smoothing and gradient penalty terms, to obtain the remaining lifetime prediction model. The feature extractor includes a multi-scale dilated convolution module and a sparse Transformer encoder. The multi-scale dilated convolution module contains multiple convolution kernels with different dilation rates, a batch normalization layer, a ReLU activation function, and a max pooling layer. The sparse Transformer encoder contains a multi-head sparse attention layer with residual connections, a feedforward neural network, and a layer normalization layer. The gradient inversion layer is located between the feature extractor and the domain discriminator.

[0006] Optionally, the training process of the generative adversarial network further includes: Using the feature extractor and based on the laboratory aging data and the actual operating data respectively, degradation features of different industrial equipment are extracted to obtain laboratory degradation features and actual degradation features; The adaptive gated loop unit dynamically weights and fuses the transient degradation features and performance degradation features in the laboratory degradation features, as well as the transient degradation features and performance degradation features in the actual degradation features, to obtain the first fused features and the second fused features. The first fused feature and the second fused feature are respectively input into the domain discriminator to perform domain classification on the input features; The gradient inversion layer enables the feature extractor to maximize the discrimination loss of the domain discriminator and enables the domain discriminator to minimize the classification loss of the domain discriminator.

[0007] Optionally, the step of dynamically fusing the local transient impact features and the global degradation trend features through the adaptive gating unit and based on the statistical characteristics of the current operating data to obtain the fused degradation features includes: Statistical calculations are performed on the current running data to obtain statistical results; Based on the statistical results and through the adaptive gating unit, corresponding fusion weights are dynamically generated for the local transient impact features and the global degradation trend features, respectively, to obtain the first fusion weight and the second fusion weight. Calculate the product of the local transient impact feature and the first fusion weight, and the product of the global degradation trend feature and the second fusion weight to obtain the first product and the second product, and calculate the sum of the first product and the second product to obtain the fused degradation feature.

[0008] Optionally, the step of performing statistical calculations on the current running data to obtain statistical results includes: The current running data is statistically calculated using a variety of preset statistical indicators to obtain multiple statistical results; the various statistical indicators include any one or more of variance, peak value, and kurtosis.

[0009] Optionally, the adaptive gating unit includes a multilayer perceptron composed of a multilayer fully connected network, an activation function layer, and a normalization layer; Accordingly, the adaptive gating unit dynamically generates corresponding fusion weights for the local transient impact features and the global degradation trend features based on the statistical results, respectively, to obtain the first fusion weight and the second fusion weight, including: Obtain the historical fusion weights from the previous moment, and concatenate multiple statistical results to obtain a statistical feature vector; The statistical feature vector and the historical fusion weights are concatenated to obtain a concatenated vector, and the concatenated vector is sequentially input into the multilayer perceptron and activation function layer in the adaptive gating unit to output the first score and the second score. The first score and the second score are input to the normalization layer in the adaptive gating unit to normalize the first score and the second score, thereby obtaining the first fusion weight corresponding to the local transient impact feature and the second fusion weight corresponding to the global degradation trend feature.

[0010] Optionally, inputting the current operating data into the trained remaining lifespan prediction model includes: The current running data is segmented according to a preset time window to obtain segmented running data; The segmented running data is preprocessed to obtain processed running data, and the processed running data is input into the trained remaining lifetime prediction model.

[0011] Secondly, this application discloses a device for predicting the remaining life of equipment, comprising: The acquisition module is used to acquire the current operating data of the target industrial equipment to be predicted, and obtain the current operating data. The input module is used to input the current operating data into the trained remaining lifetime prediction model. The remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets. The multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture. The remaining lifetime predictor consists of multiple fully connected layers. The historical dataset includes laboratory aging data and actual operating data of different industrial equipment. The feature extraction module is used to extract the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment through the multi-scale dilated convolution module and based on the current operating data, to obtain local transient impact features. At the same time, it uses the sparse Transformer encoder and based on the current operating data to extract the performance degradation trend of the target industrial equipment over time during operation, to obtain global degradation trend features. The dynamic fusion module is used to dynamically fuse the local transient impact features and the global degradation trend features through the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features; The prediction module is used to predict the remaining life of the target industrial equipment using the remaining life predictor and based on the fused post-degradation features, and obtain the current prediction result.

[0012] Thirdly, this application discloses an electronic device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the aforementioned device remaining life prediction method.

[0013] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for predicting the remaining lifespan of a device.

[0014] As can be seen, this application first obtains the current operating data of the target industrial equipment to be predicted, and then inputs the current operating data into the trained remaining lifetime prediction model. The remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using a historical dataset. The multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture. The remaining lifetime predictor consists of multiple fully connected layers. The historical dataset includes laboratory aging data and actual operating data of different industrial equipment. Then, the data is processed by the multi-scale dilated convolutional module and based on... The current operating data is used to extract transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment, resulting in local transient impact features. Simultaneously, the sparse Transformer encoder, based on the current operating data, extracts features of the performance degradation trend of the target industrial equipment over time, resulting in global degradation trend features. Then, the adaptive gating unit, based on the statistical characteristics of the current operating data, dynamically fuses the local transient impact features and the global degradation trend features to obtain fused degradation features. Finally, the remaining lifetime predictor, based on the fused degradation features, predicts the remaining lifetime of the target industrial equipment to obtain the current prediction result. This application predicts the remaining life of industrial equipment using a generative adversarial network (GAN) model. The model includes a multi-scale dilated convolutional module and a sparse Transformer encoder with a parallel architecture. These are used to extract features of transient degradation caused by sudden physical shocks and performance degradation trends over time during equipment operation, respectively, yielding local transient shock features and global degradation trend features. This application considers not only long-term degradation trends throughout the entire lifecycle (macro-degradation trends based on long-term dependencies, such as slow wear) but also micro-transient features caused by sudden physical shocks such as instantaneous impacts, arc discharges, and vibration mutations. This provides richer, multi-dimensional data for predicting equipment remaining life, thus improving the accuracy of remaining life prediction. Furthermore, this application dynamically fuses the synchronously captured local transient shock features and global degradation trend features based on the statistical characteristics of current operating data and through an adaptive gating unit in the model. Compared to traditional static fusion methods based on fixed weights, this allows adjustment of fusion weights according to the current equipment operating conditions, further improving the accuracy of remaining life prediction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of a method for predicting the remaining life of equipment disclosed in this application; Figure 2 This is a schematic diagram of a specific hybrid feature extraction and fusion process disclosed in this application; Figure 3 This is a schematic diagram of a specific adversarial training process disclosed in this application; Figure 4 This is a schematic diagram of a specific hybrid feature extraction and fusion process disclosed in this application; Figure 5 This is a flowchart of a specific method for predicting the remaining life of equipment disclosed in this application; Figure 6 This is a schematic diagram of the structure of a device for predicting the remaining life of equipment disclosed in this application; Figure 7 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application discloses a method for predicting the remaining life of equipment. See also Figure 1 As shown, the method includes: Step S11: Obtain the current operating data of the target industrial equipment to be predicted, and obtain the current operating data.

[0019] In this embodiment, the on-site operation data of the target industrial equipment to be predicted in the industrial scenario is first acquired. For example, multi-dimensional time series data (such as vibration acceleration, vibration velocity, vibration displacement, current, temperature, etc.) detected by various sensors are acquired to obtain the current operation data. Among them, industrial equipment includes, but is not limited to, electric motors, generators, water pumps, fans (compressors), air compressors, etc.

[0020] Step S12: Input the current operating data into the trained remaining lifetime prediction model; the remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets, wherein the multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture; the remaining lifetime predictor consists of multiple fully connected layers; the historical dataset includes laboratory aging data and actual operating data of different industrial equipment.

[0021] In this embodiment, after acquiring the operating data of the target industrial equipment to be predicted, the current operating data is input into a remaining lifetime prediction model trained on a generative adversarial network (GAN) consisting of a multi-scale dilated convolutional block, a sparse Transformer encoder, an adaptive gating unit (AGU), and a remaining lifetime predictor (composed of multiple fully connected layers) using historical datasets (including laboratory aging data and actual operating data of different industrial equipment). The GAN is then used to predict the remaining lifetime. The adaptive gating unit can adaptively learn and control the information flow between feature channels or dimensions, dynamically suppressing noise, enhancing effective features, and improving the model's ability to model complex data. It should be noted that the remaining lifetime prediction model can be deployed on an edge computing platform / device.

[0022] It should be noted that the remaining lifespan prediction model contains two branches: a CNN (Convolutional Neural Network) branch and a Transformer (a sequence model based on an attention mechanism) branch. The CNN branch includes a multi-scale dilated convolution module, and the Transformer branch includes a sparse Transformer encoder. The multi-scale dilated convolution module and the sparse Transformer encoder in both branches adopt a parallel architecture. By adopting a parallel architecture, rather than a serial or integrated approach, the two branches can extract features independently, avoiding information loss or unidirectional interference caused by serial structures, and achieving synchronous capture of microscopic transient shocks and macroscopic degradation trends. Furthermore, the multi-scale dilated convolution module is equipped with multiple convolution kernels with different dilation rates to expand the receptive field without increasing the number of parameters, and to extract microscopic transient features caused by sudden physical impacts such as instantaneous impacts, arc discharges, and vibration mutations (i.e., non-stationary local features caused by sudden physical impacts). The sparse Transformer encoder adopts an improved sparse Transformer attention mechanism to reduce computational complexity and capture the slow degradation trend of equipment performance over time (i.e., the long-term degradation trend in the entire life cycle of the equipment, such as slow wear).

[0023] It is understandable that transient shocks and arc discharges generated during the operation of industrial equipment will induce microscopic transient features in the signal. These features are characterized by rapid changes, short duration, and significant amplitude abrupt changes, making them difficult to capture effectively by conventional convolutional structures. However, by constructing local feature extraction branches through multi-scale dilated convolutions, the receptive field can be expanded without significantly increasing the number of network parameters, accurately extracting the high-frequency abrupt changes and subtle distortion features corresponding to the aforementioned transient shocks and arc discharges, thereby providing a reliable basis for remaining lifetime prediction.

[0024] Specifically, the training process of the remaining life prediction model may include: collecting laboratory aging data from different industrial equipment in the source domain and actual operating data from different industrial equipment in the target domain; inputting the laboratory aging data and the actual operating data into a system comprising a feature extractor, an adaptive gated recurrent unit (GRU, an improved recurrent neural network), a gradient inversion layer (GRL), a domain discriminator, and a remaining life predictor. In the generative adversarial network (GAN) of the Predictor, the GAN is adversarially trained based on the laboratory aging data and the actual operating data, using label smoothing and gradient penalty terms, to obtain the remaining lifetime prediction model. The feature extractor includes a multi-scale dilated convolution module and a sparse Transformer encoder. The multi-scale dilated convolution module contains multiple convolutional kernels with different dilation rates, batch normalization layers, ReLU activation functions, and max-pooling layers. The sparse Transformer encoder contains a multi-head sparse attention layer with residual connections, a feedforward neural network, and layer normalization layers. The gradient inversion layer is located between the feature extractor and the domain discriminator. In this embodiment, laboratory aging data (i.e., accelerated aging data of different equipment simulated in the laboratory) from different industrial equipment in the source domain and actual operating data (i.e., data detected by sensors in different industrial fields, such as current, temperature, vibration velocity, vibration displacement, etc.) from different industrial equipment in the target domain are first collected. Then, the collected laboratory aging data and actual operating data are input into a generative adversarial network (GAN) that includes a feature extractor, an adaptive gated recurrent unit, a gradient reversal layer (GRL), a domain discriminator, and a remaining lifetime predictor. Based on the laboratory aging data and actual operating data, and using label smoothing technology and gradient penalty terms, the GAN is adversarially trained to obtain a model for real-time remaining lifetime prediction. It should be noted that the Feature Extractor includes a multi-scale dilated convolution module and a sparse Transformer encoder with a parallel architecture. The multi-scale dilated convolution module contains multiple convolutional kernels with different dilation rates, a batch normalization (BN) layer, a ReLU activation function, and a max pooling layer. The sparse Transformer encoder contains a multi-head sparse attention layer with residual connections, a feedforward neural network, and a layer normalization layer. The gradient inversion layer is located between the Feature Extractor and the domain discriminator.The input of the adaptive gated loop unit can include not only the data features of the current window, but also the fusion weights from the previous moment can be introduced through a loop unit, so that the weight changes have temporal continuity and avoid drastic weight jitter caused by instantaneous noise.

[0025] Specifically, the CNN branch can be composed of multiple multi-scale dilated convolutional modules stacked together. Each multi-scale dilated convolutional module contains three parallel dilated convolutional layers with dilation rates of 1, 2, and 5, corresponding to different receptive fields to capture transient impact features at different scales. Furthermore, each multi-scale dilated convolutional module can be followed by a batch normalization layer and a ReLU activation function, then dimensionality reduction is performed through a max pooling layer, finally outputting the local degradation feature F_cnn (represented in vector form).

[0026] For the Transformer branch, an improved sparse Transformer encoder is employed. This encoder extracts global degradation features based on a sparse attention mechanism. This mechanism reduces the computational complexity from O(n²) to O(nlogn) by sparsifying the attention matrix (computing attention only near the diagonal and at random sampling points), while maintaining the ability to capture long-term dependencies. The sparse Transformer encoder includes a multi-head sparse attention layer, a feed-forward network, and layer normalization, and uses residual connections to finally output the global degradation feature F_trans (represented as a vector).

[0027] In this embodiment, the multi-scale dilated convolution module and sparse Transformer encoder in the feature extractor can extract features from the microscopic transient features (i.e., local degradation features, such as instantaneous impact, arc discharge, and sudden vibration) caused by sudden physical impacts during the operation of different industrial equipment, as well as the global degradation features of the equipment performance over time (such as slow wear). That is, features from the source domain and the target domain are extracted. The lightweight adaptive gating unit can then be used to extract features based on the input data (i.e., laboratory aging data and actual operating data). The model dynamically learns a set of fusion weights α and β (α+β=1) based on the statistical properties of the signal (such as variance, peak value, etc.). These weights are used to perform element-wise weighted fusion of the local degradation features F_cnn from the CNN branch output and the global degradation features F_trans from the Transformer branch output. The specific fusion formula is: F_fused=α⊙F_cnn+β⊙F_trans. Through this dynamic fusion mechanism, the model can focus more on local degradation features when the equipment experiences early and minor faults, and more on global degradation trends during stable operation. Furthermore, to address the inconsistency in data distribution between the source and target domains, as well as the data distribution drift caused by changes in operating conditions in industrial settings, this application introduces a gradient inversion layer (GRL) and a domain discriminator into the model. The domain discriminator can be a binary classifier used to determine whether the input features come from the source or target domain.

[0028] Furthermore, this application utilizes label smoothing techniques and gradient penalty terms to perform adversarial training on the generative adversarial network (GAN). Adversarial training deeply obfuscates features between the source and target domains, thereby avoiding overfitting and improving model performance (i.e., enhancing the predictive ability of remaining lifetime). The introduction of label smoothing and gradient penalty terms optimizes the adversarial training process, ensuring that the game between the feature extractor and the domain discriminator reaches a Nash equilibrium, thus stably generating domain-invariant features. Moreover, adversarial training through a gradient inversion layer (GRL) more effectively handles complex and non-linear data distribution shifts, truly achieving "zero-sample" or "few-sample" adaptive transfer from the laboratory to the field, and forcing the feature extractor to learn domain-invariant features.

[0029] Furthermore, the training process of the generative adversarial network specifically includes: using the feature extractor and based on the laboratory aging data and the actual operating data respectively, extracting degradation features of different industrial equipment to obtain laboratory degradation features and actual degradation features; using the adaptive gated recurrent unit, dynamically weighting and fusing the transient degradation features and performance degradation features in the laboratory degradation features and the transient degradation features and performance degradation features in the actual degradation features to obtain a first fused feature and a second fused feature; inputting the first fused feature and the second fused feature into the domain discriminator to classify the input features into a domain; and using a gradient inversion layer to maximize the discrimination loss of the domain discriminator by the feature extractor and minimize the classification loss of the domain discriminator by the domain discriminator. In this embodiment, during training, the degradation features (including local degradation features (i.e., transient degradation features) and global degradation features (i.e., performance degradation features) of different industrial equipment are first extracted using the multi-scale dilated convolution module and sparse Transformer encoder in the feature extractor, based on laboratory aging data from the source domain and actual operating data from the target domain. This yields the corresponding laboratory degradation features and actual degradation features. Then, an adaptive gated recurrent unit is used to extract the transient degradation features and performance degradation features from the laboratory degradation features, as well as the transient degradation features and performance degradation features from the actual degradation features. The performance degradation features are dynamically weighted and fused to obtain the corresponding first and second fused features. The two fused features are then input into the domain discriminator to classify the input features into the source domain (i.e., determine whether they belong to the source domain or the target domain). Finally, a gradient reversal layer is used to maximize the discriminative loss of the domain discriminator (even if the features extracted by the feature extractor cannot distinguish which domain the feature data comes from) and minimize the classification loss of the domain discriminator. Through the above adversarial game, the feature extractor can learn a feature representation that contains rich degradation information and is robust to domain changes.

[0030] In addition, during training, the prediction loss (which uses the true lifetime label of the source domain to calculate the prediction error, such as mean square error, to ensure that the extracted features can accurately reflect the remaining lifetime of the device) and the domain discrimination loss can be jointly optimized. After training is completed, the model parameters are frozen to obtain a high-precision model of remaining lifetime for real-time remaining lifetime prediction.

[0031] For details, see Figure 3As shown, features are extracted by the feature extractor G_f (parameter θf), and the extracted features are fused to obtain the fused feature F_fused. Then, the remaining lifetime predictor (composed of several fully connected layers, parameter θy) is used to predict the remaining lifetime, obtaining the predicted value y^o. The prediction loss L_y (e.g., mean-square error, MSE) can be calculated using the true label y of the source domain. The fused feature F_fused is then passed through a gradient inversion layer (GRL) and input into the domain discriminator G_d (parameter θd) to determine whether the input feature comes from the source domain (label 0) or the target domain (label 1). It should be noted that during forward propagation, the gradient inversion layer acts as an identity transformation: GRL(x) = x; while during backward propagation, the gradient inversion layer inverts the gradient from the domain discriminator before passing it to the feature extractor: dxdGRL = -λdxdL_d (where λ is the tradeoff coefficient and L_d is the domain discrimination loss).

[0032] The adversarial logic of adversarial training is as follows: the domain discriminator G_d strives to minimize the domain discriminative loss L_d (such as cross-entropy loss) to accurately distinguish the data source; while the feature extractor G_f, due to the gradient reversal layer, actually strives to maximize L_d, that is, to extract features that the domain discriminator cannot distinguish the source (i.e., domain-invariant features). The total loss is L = Ly + λL_d. By jointly optimizing the total loss, a model that is both accurate in prediction and robust to domain changes can be obtained.

[0033] Step S13: Using the multi-scale dilated convolution module and based on the current operating data, extract the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment to obtain local transient impact features. At the same time, using the sparse Transformer encoder and based on the current operating data, extract the performance degradation trend of the target industrial equipment over time to obtain global degradation trend features.

[0034] In this embodiment, see Figure 2 As shown, the current operating data can be processed by the multi-scale dilated convolution module (located in the CNN branch) in the trained remaining life prediction model, thereby extracting the transient degradation features caused by sudden physical shocks (such as instantaneous impacts, arc discharges, vibration mutations, etc.) during the operation of the current target industrial equipment, and obtaining local transient impact features. At the same time, the current operating data can be processed by the sparse Transformer encoder in another Transformer branch to extract the performance degradation trend of the target industrial equipment over time during operation, and obtain the global degradation trend features.

[0035] The CNN branch can use depthwise separable convolution kernels and dilated convolution kernels to reduce the number of parameters, while the Transformer branch uses a sparse attention mechanism to compute only the most relevant tokens, thereby reducing the computational load and ensuring real-time performance.

[0036] Step S14: The local transient impact features and the global degradation trend features are dynamically fused using the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features.

[0037] In this embodiment, see Figure 2 As shown, an adaptive gating unit can dynamically learn a set of fusion weights α and β (α+β=1) based on the statistical characteristics (such as variance and peak value) of the current operating data. Then, based on the fusion weights α and β, local transient impact features and global degradation trend features are dynamically fused to obtain the fused degradation features. It is evident that the adaptive gating unit can adjust the contribution of local and global features in real time according to the characteristics of the data itself, ensuring that the model maintains optimal feature representation at different degradation stages of the device.

[0038] Specifically, the step of dynamically fusing the local transient impact feature and the global degradation trend feature based on the statistical characteristics of the current operating data through the adaptive gating unit to obtain the fused degradation feature may include: performing statistical calculations on the current operating data to obtain statistical results; dynamically generating corresponding fusion weights for the local transient impact feature and the global degradation trend feature respectively through the adaptive gating unit and based on the statistical results to obtain a first fusion weight and a second fusion weight; calculating the product of the local transient impact feature and the first fusion weight, and the product of the global degradation trend feature and the second fusion weight, to obtain a first product and a second product, and calculating the sum of the first product and the second product to obtain the fused degradation feature. See also... Figure 4 As shown, statistical calculations can be performed on the current running data X_t according to any one or more of the preset statistical indicators, such as variance, peak-to-peak, and kurtosis. Then, through an adaptive gating unit (composed of two fully connected layers (FC) and a hyperbolic tangent function (Tanh)) and based on the statistical results, corresponding fusion weights are dynamically generated for local transient impact characteristics and global degradation trend characteristics, respectively, to obtain the first fusion weight α_t and the second fusion weight β_t (α_t + β_t = 1).

[0039] In one specific implementation, the adaptive gating unit includes a multilayer perceptron (MLP) composed of a multilayer fully connected network (FC), a Tanh activation function layer (located in the middle layer), and a normalization layer.

[0040] Accordingly, the step of dynamically generating corresponding fusion weights for the local transient impact feature and the global degradation trend feature based on the statistical results through the adaptive gating unit, to obtain the first fusion weight and the second fusion weight, may specifically include: obtaining the historical fusion weights of the previous moment and concatenating multiple statistical results to obtain a statistical feature vector; concatenating the statistical feature vector and the historical fusion weights to obtain a concatenated vector, and sequentially inputting the concatenated vector into the multilayer perceptron and activation function layer in the adaptive gating unit to output a first score and a second score; inputting the first score and the second score into the normalization layer in the adaptive gating unit to normalize the first score and the second score, to obtain the first fusion weight corresponding to the local transient impact feature and the second fusion weight corresponding to the global degradation trend feature. In this embodiment, see... Figure 4 As shown, the historical fusion weights α_t-1 from the previous time step (introduced through a loop connection) can be used to ensure the temporal continuity of the weights and avoid instantaneous noise interference. Then, multiple statistical results (such as variance, peak value, and kurtosis) are concatenated to obtain a statistical feature vector S_t. The statistical feature vector S_t and the historical fusion weights α_t-1 are then concatenated to obtain the concatenated vector [S_t, α_t-1]. Finally, the concatenated vector [S_t, α_t-1] is sequentially input into the multilayer perceptron and activation function in the adaptive gating unit. The adaptive gating unit first outputs two scores, a and b. These scores are then fed into a normalization layer (e.g., a Softmax layer) to normalize the scores (a and b) using Softmax, yielding the first fusion weight α_t corresponding to the local transient impact feature and the second fusion weight β_t corresponding to the global degradation trend feature F_trans. It is crucial to ensure that α_t + β_t = 1 (i.e., β_t = 1 - α_t) and that α_t and β_t ∈ (0, 1). Finally, the products of α_t and F_cnn, and β_t and F_trans, are calculated, and the sum of these products is used as the final fused degradation feature F_fused.

[0041] Step S15: Predict the remaining life of the target industrial equipment using the remaining life predictor and based on the fused degradation features to obtain the current prediction result.

[0042] In this embodiment, after feature fusion, the fused degradation features are input into the remaining lifetime predictor (composed of multiple fully connected layers) to predict the remaining lifetime of the current target industrial equipment based on the fused degradation features, thereby obtaining the current prediction result.

[0043] Furthermore, the usage status of the target industrial equipment can be determined based on the current prediction results. If the prediction results indicate that the remaining lifespan is less than the preset time, a corresponding alarm message will be generated and sent to the corresponding management terminal to remind the management personnel to conduct a manual inspection of the corresponding equipment as soon as possible and take corresponding measures, such as replacing equipment / components, to extend the equipment lifespan or ensure that the equipment continues to work.

[0044] As can be seen, this application embodiment predicts the remaining life of industrial equipment using a generative adversarial network-based model. This model includes a multi-scale dilated convolutional module and a sparse Transformer encoder with a parallel architecture, used to extract features of transient degradation characteristics caused by sudden physical shocks during equipment operation and performance degradation trends over time, respectively, resulting in local transient shock features and global degradation trend features. This application considers not only long-term degradation trends throughout the entire lifecycle (macro-degradation trends based on long-term dependencies, such as slow wear), but also micro-transient features caused by sudden physical shocks such as instantaneous impacts, arc discharges, and vibration mutations. This provides richer, multi-dimensional data for predicting the remaining life of equipment, thereby improving the accuracy of the prediction. Furthermore, this application embodiment dynamically fuses the synchronously captured local transient shock features and global degradation trend features based on the statistical characteristics of current operating data and through an adaptive gating unit in the model. Compared to the traditional static fusion method based on fixed weights, this method can adjust the fusion weights according to the current equipment operating conditions, further improving the accuracy of the remaining life prediction.

[0045] This application discloses a specific method for predicting the remaining life of equipment. (See also...) Figure 5 As shown, the method includes: Step S21: Obtain the current operating data of the target industrial equipment to be predicted, and obtain the current operating data.

[0046] Step S22: Divide the current running data according to a preset time window to obtain the divided running data.

[0047] In this embodiment, after obtaining the current operating data of the target industrial equipment to be predicted, the current operating data can be further sliced ​​according to a preset time window to obtain multiple batches of sliced ​​operating data.

[0048] Step S23: Perform preprocessing on the segmented running data to obtain processed running data, and input the processed running data into the trained remaining lifetime prediction model.

[0049] In this embodiment, the segmented running data of each batch are preprocessed sequentially, such as abnormal data removal and normalization, to obtain processed running data. Then, the processed running data is input into the trained remaining lifetime prediction model.

[0050] Step S24: Input the current operating data into the trained remaining lifetime prediction model; the remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets, wherein the multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture; the remaining lifetime predictor consists of multiple fully connected layers; the historical dataset includes laboratory aging data and actual operating data of different industrial equipment.

[0051] Step S25: Using the multi-scale dilated convolution module and based on the current operating data, extract the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment to obtain local transient impact features. At the same time, using the sparse Transformer encoder and based on the current operating data, extract the performance degradation trend of the target industrial equipment over time to obtain global degradation trend features.

[0052] Step S26: The local transient impact features and the global degradation trend features are dynamically fused using the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features.

[0053] Step S27: Predict the remaining life of the target industrial equipment using the remaining life predictor and based on the fused degradation features to obtain the current prediction result.

[0054] For more detailed processing procedures regarding steps S21, S24 to S27, please refer to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

[0055] As can be seen, in this embodiment, the current running data is segmented according to a preset time window to obtain segmented running data. Then, the segmented running data is preprocessed to obtain processed running data. The processed running data is then input into the trained remaining lifetime prediction model for remaining lifetime prediction. Through slicing and preprocessing operations, the accuracy and efficiency of equipment remaining lifetime prediction can be further improved.

[0056] Accordingly, this application also discloses a device for predicting the remaining life of equipment, see [link to relevant documentation]. Figure 6 As shown, the device includes: The acquisition module 11 is used to acquire the current operating data of the target industrial equipment to be predicted, and obtain the current operating data; Input module 12 is used to input the current operating data into the trained remaining lifetime prediction model; the remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets, wherein the multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture; the remaining lifetime predictor consists of multiple fully connected layers; the historical dataset includes laboratory aging data and actual operating data of different industrial equipment. The feature extraction module 13 is used to extract the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment through the multi-scale dilated convolution module and based on the current operating data, to obtain local transient impact features. At the same time, it uses the sparse Transformer encoder and based on the current operating data to extract the performance degradation trend of the target industrial equipment during operation over time, to obtain global degradation trend features. The dynamic fusion module 14 is used to dynamically fuse the local transient impact features and the global degradation trend features through the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features; The prediction module 15 is used to predict the remaining life of the target industrial equipment through the remaining life predictor and based on the fused degradation features, and obtain the current prediction result.

[0057] The specific workflow of each of the above modules can be found in the relevant content disclosed in the foregoing embodiments, and will not be repeated here.

[0058] Furthermore, embodiments of this application also disclose an electronic device, Figure 7This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0059] Figure 7 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the device remaining life prediction method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0060] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0061] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0062] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the device remaining life prediction method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0063] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed method for predicting the remaining lifespan of a device. The specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0064] Furthermore, embodiments of this application also disclose a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the device remaining lifetime prediction method disclosed above.

[0065] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0066] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0068] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only 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. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0069] The foregoing has provided a detailed description of a method, apparatus, device, and storage medium for predicting the remaining life of equipment. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the remaining life of equipment, characterized in that, include: Obtain the current operating data of the target industrial equipment to be predicted; The current operating data is input into the trained remaining lifetime prediction model. The remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets. The multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture. The remaining lifetime predictor consists of multiple fully connected layers. The historical dataset includes laboratory aging data and actual operating data of different industrial equipment. Using the multi-scale dilated convolution module and based on the current operating data, the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment are extracted to obtain local transient impact features. At the same time, using the sparse Transformer encoder and based on the current operating data, the performance degradation trend of the target industrial equipment over time is extracted to obtain global degradation trend features. The local transient impact features and the global degradation trend features are dynamically fused using the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features. The remaining life of the target industrial equipment is predicted using the remaining life predictor and based on the fused post-degradation features, resulting in the current prediction result.

2. The method for predicting the remaining life of equipment according to claim 1, characterized in that, Before obtaining the current operating data of the target industrial equipment to be predicted, the process further includes: Collect laboratory aging data from different industrial equipment in the source domain and actual operating data from different industrial equipment in the target domain; The laboratory aging data and the actual operating data are input into a generative adversarial network (GAN) containing a feature extractor, an adaptive gated recurrent unit, a gradient inversion layer, a domain discriminator, and a remaining lifetime predictor. The GAN is then adversarially trained based on the laboratory aging data and the actual operating data, using label smoothing and gradient penalty terms, to obtain the remaining lifetime prediction model. The feature extractor includes a multi-scale dilated convolution module and a sparse Transformer encoder. The multi-scale dilated convolution module contains multiple convolution kernels with different dilation rates, a batch normalization layer, a ReLU activation function, and a max pooling layer. The sparse Transformer encoder contains a multi-head sparse attention layer with residual connections, a feedforward neural network, and a layer normalization layer. The gradient inversion layer is located between the feature extractor and the domain discriminator.

3. The method for predicting the remaining life of equipment according to claim 2, characterized in that, The training process of the generative adversarial network also includes: Using the feature extractor and based on the laboratory aging data and the actual operating data respectively, degradation features of different industrial equipment are extracted to obtain laboratory degradation features and actual degradation features; The adaptive gated loop unit dynamically weights and fuses the transient degradation features and performance degradation features in the laboratory degradation features, as well as the transient degradation features and performance degradation features in the actual degradation features, to obtain the first fused features and the second fused features. The first fused feature and the second fused feature are respectively input into the domain discriminator to perform domain classification on the input features; The gradient inversion layer enables the feature extractor to maximize the discrimination loss of the domain discriminator and enables the domain discriminator to minimize the classification loss of the domain discriminator.

4. The method for predicting the remaining life of equipment according to claim 3, characterized in that, The step involves dynamically fusing the local transient impact features and the global degradation trend features using the adaptive gating unit and based on the statistical characteristics of the current operating data to obtain the fused degradation features, including: Statistical calculations are performed on the current running data to obtain statistical results; Based on the statistical results and through the adaptive gating unit, corresponding fusion weights are dynamically generated for the local transient impact features and the global degradation trend features, respectively, to obtain the first fusion weight and the second fusion weight. Calculate the product of the local transient impact feature and the first fusion weight, and the product of the global degradation trend feature and the second fusion weight to obtain the first product and the second product, and calculate the sum of the first product and the second product to obtain the fused degradation feature.

5. The method for predicting the remaining life of equipment according to claim 4, characterized in that, The statistical calculation of the current operating data to obtain statistical results includes: The current running data is statistically calculated using a variety of preset statistical indicators to obtain multiple statistical results; the various statistical indicators include any one or more of variance, peak value, and kurtosis.

6. The method for predicting the remaining life of equipment according to claim 4, characterized in that, The adaptive gating unit includes a multilayer perceptron composed of a multilayer fully connected network, an activation function layer, and a normalization layer. Accordingly, the adaptive gating unit dynamically generates corresponding fusion weights for the local transient impact features and the global degradation trend features based on the statistical results, respectively, to obtain the first fusion weight and the second fusion weight, including: Obtain the historical fusion weights from the previous moment, and concatenate multiple statistical results to obtain a statistical feature vector; The statistical feature vector and the historical fusion weights are concatenated to obtain a concatenated vector, and the concatenated vector is sequentially input into the multilayer perceptron and activation function layer in the adaptive gating unit to output the first score and the second score. The first score and the second score are input to the normalization layer in the adaptive gating unit to normalize the first score and the second score, thereby obtaining the first fusion weight corresponding to the local transient impact feature and the second fusion weight corresponding to the global degradation trend feature.

7. The method for predicting the remaining life of equipment according to any one of claims 1 to 6, characterized in that, The step of inputting the current operating data into the trained remaining lifespan prediction model includes: The current running data is segmented according to a preset time window to obtain segmented running data; The segmented running data is preprocessed to obtain processed running data, and the processed running data is input into the trained remaining lifetime prediction model.

8. A device for predicting the remaining life of equipment, characterized in that, include: The acquisition module is used to acquire the current operating data of the target industrial equipment to be predicted, and obtain the current operating data. The input module is used to input the current operating data into the trained remaining lifetime prediction model. The remaining lifetime prediction model is a model obtained by training a generative adversarial network containing a multi-scale dilated convolutional module, a sparse Transformer encoder, an adaptive gating unit, and a remaining lifetime predictor using historical datasets. The multi-scale dilated convolutional module and the sparse Transformer encoder adopt a parallel architecture. The remaining lifetime predictor consists of multiple fully connected layers. The historical dataset includes laboratory aging data and actual operating data of different industrial equipment. The feature extraction module is used to extract the transient degradation features caused by sudden physical impacts during the operation of the target industrial equipment through the multi-scale dilated convolution module and based on the current operating data, to obtain local transient impact features. At the same time, it uses the sparse Transformer encoder and based on the current operating data to extract the performance degradation trend of the target industrial equipment over time during operation, to obtain global degradation trend features. The dynamic fusion module is used to dynamically fuse the local transient impact features and the global degradation trend features through the adaptive gating unit and based on the statistical characteristics of the current running data to obtain the fused degradation features; The prediction module is used to predict the remaining life of the target industrial equipment using the remaining life predictor and based on the fused post-degradation features, and obtain the current prediction result.

9. An electronic device, characterized in that, It includes a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the device remaining life prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the device remaining life prediction method as described in any one of claims 1 to 7.

Images