Remaining life prediction method, device and storage medium

Abstract

The present invention discloses a method, device, and storage medium for predicting remaining life, wherein the method comprises: obtaining raw data from a device; performing unsupervised clustering training on the raw data to obtain pseudo-labels corresponding to decay stages; establishing a remaining life prediction model based on adaptive perception of decay stages of spatiotemporal dual-cell states, and performing auxiliary self-supervised course training on the remaining life prediction model based on the pseudo-labels; collecting real-time data from the device, inputting the real-time data into the trained remaining life prediction model for prediction, and outputting the predicted value of the remaining life of the device and the corresponding stage value. The present invention obtains decay stage information in an unsupervised manner and enhances the model's ability to learn decay states through self-supervised auxiliary learning, thereby enabling the model to better integrate the concept of decay stages into the remaining life prediction task, further improving the model's prediction effect. The present invention can be widely applied to the field of equipment remaining life prediction.

Description

Technical Field

[0001] The present invention relates to the field of equipment remaining life prediction, and in particular to a remaining life prediction method, device and storage medium. Background Art

[0002] Prognostic and Health Management (PHM) refers to the technology that uses large amounts of condition monitoring data and information to evaluate and manage the health status of equipment with the help of statistical algorithms or models. It can predict potential failures in advance, and combine various equipment information to provide maintenance decisions, achieve situation-based maintenance, thereby improving the safety of the production process and reducing maintenance costs. Remaining Useful Life (RUL) prediction technology is one of the key technologies of PHM. It predicts the RUL of a system or component in advance by analyzing the operating data monitored by sensors or establishing a suitable degradation model. The data-driven approach provides an important means of RUL prediction. Through data acquisition, data preprocessing, feature engineering, model building, model training and prediction, an effective RUL prediction model is established to process the monitoring data to achieve RUL prediction.

[0003] In actual production RUL prediction, how to use data-driven methods to accurately predict the life of equipment components still faces many challenges. First, under different working conditions, the life decay behavior of equipment components (such as bearings) is different. Even under the same working conditions, due to the statistical variance of the factory status of the components, the initial differences between individuals lead to different changes in the decay process. Therefore, the equipment life is deeply affected by the environment. Second, in the remaining life prediction task, the data collected in different sampling segments have time correlation. That is, in addition to the temporal relationship of the data of a single continuous sampling segment, the data between different sampling segments with a temporal sequence also have a certain temporal correlation. For example, the data of different decay states have a relationship between the occurrence of the previous and the next. Third, the monitoring data often has the problem of data imbalance, especially the obvious imbalance between healthy and unhealthy. It is necessary to find a way to balance the monitoring data in a healthy state with the data in an unhealthy and severely decayed state. This requires the introduction of stage information with dual temporal and spatial features. Summary of the Invention

[0004] In order to at least solve one of the technical problems existing in the prior art to a certain extent, the present invention aims to provide a remaining life prediction method, device and storage medium.

[0005] The technical solution adopted in the present invention is:

[0006] A method for predicting remaining life includes the following steps:

[0007] Get the raw data of the device;

[0008] Performing unsupervised clustering training on the original data to obtain pseudo labels corresponding to the attenuation stage;

[0009] Establish a remaining life prediction model based on adaptive perception of the spatiotemporal dual-cell state decay stage, and perform auxiliary self-supervised course training on the remaining life prediction model based on pseudo labels;

[0010] Collect real-time data of the equipment, input the real-time data into the trained remaining life prediction model for prediction, and output the predicted value of the remaining life of the equipment and the corresponding stage value.

[0011] Furthermore, performing unsupervised clustering training on the original data to obtain pseudo labels corresponding to the attenuation stage includes:

[0012] Use a deep model to convert raw data into high-dimensional features, cluster the high-dimensional features, and obtain pseudo labels based on each clustering result;

[0013] Feed the pseudo labels back to the deep model and update the parameters to finally obtain a deep clustering network;

[0014] The deep clustering network is used to identify the decay stage corresponding to the original data.

[0015] Furthermore, the remaining life prediction model includes a multi-layer convolutional embedded recurrent neural network and a prediction layer;

[0016] The original data is sampled in a sliding window manner to obtain window data as the input of the remaining life prediction model;

[0017] After the window data passes through the recurrent neural network and prediction layer, the remaining life prediction value is output;

[0018] Among them, each layer of the recurrent neural network consists of two parts: temporal state transfer and spatial state transfer. The temporal state propagates horizontally with the time steps, while the spatial state zigzags between adjacent time steps in the vertical direction.

[0019] Furthermore, the remaining life prediction model adopts a convolutional recurrent neural network, a convolutionally embedded long short-term memory network with dual spatiotemporal cell states;

[0020] The long short-term memory network can fully utilize the ability of the convolutional network to extract data spatial features and the ability of the recurrent neural network to extract temporal features, combining the advantages of the two, and can extract features from original signals with spatiotemporal features, ultimately establishing a more reliable mapping relationship for the remaining life of the equipment.

[0021] Furthermore, the remaining life prediction model uses a long short-term memory network as the backbone, and replaces the original point product operations of the input-hidden and hidden-hidden transfer functions of the long short-term memory network with convolution operations;

[0022] The spatial state is propagated vertically across different network layers, carrying state information for spatial feature extraction at each layer. The temporal state is propagated horizontally across different time steps in chronological order, carrying long-term dependency state information at each time step. Furthermore, the hidden state is used to achieve the transmission of long-term and short-term states.

[0023] Based on the dual-feature state transfer of space and time, a dynamic and static sub-cell differential mechanism is introduced. In the dynamic sub-cell, the difference between the current output of the previous layer and the output of the previous time step is used as input. After calculation, it is output to the static sub-cell for calculation, and the invariant features that contribute more to the predicted remaining life value in space and time are obtained.

[0024] Finally, the health factors extracted from the spatial and temporal features are regressed through a fully connected regression network to obtain a fitted equipment remaining useful life curve.

[0025] The spatial state of the next time step can be obtained by upsampling and accumulating the spatial state of the last layer of the previous time step layer by layer, as shown in the following formula:

[0026]

[0027] in, It is the spatial state input of the first layer. Upsamplinq refers to the upsampling algorithm, which generally uses linear interpolation. refers to the spatial state of the i-th layer.

[0028] Furthermore, the auxiliary self-supervised course training of the remaining life prediction model based on the pseudo labels includes:

[0029] By designing a new deep model loss function, based on the basis of auxiliary task learning to improve the performance of the main task, the equipment health stage classification task is learned simultaneously with the remaining life prediction task; over time, from the healthy state to the final entering the severe decay period, the learning focus is gradually shifted from the stage classification task to the remaining life prediction task, so as to better serve the main task of remaining life prediction.

[0030] Furthermore, the loss function is expressed as follows:

[0031]

[0032] Where λ is a time-varying parameter representing the weight of auxiliary learning, which is calculated using the following formula:

[0033] λ(i)=max(ζ, λ(0)-∈·i)

[0034] Where ζ is the maximum value of λ, ∈ is the decay rate factor;

[0035] The expressions of mean square error and cross entropy loss function are as follows:

[0036]

[0037]

[0038] Where, is the true value of the remaining life, is the predicted value of remaining life, n is the total number of samples, y i is the i-th true value, is the i-th predicted value, P(y ij ) is the distribution of the true category label, which is a one-hot vector with a value of 0 or 1. is the distribution of the predicted category, which is a vector whose values ​​are between [0,1], p is the total number of samples, and q is the total number of categories.

[0039] Furthermore, the auxiliary self-supervised course training of the remaining life prediction model based on the pseudo labels includes:

[0040] According to the pseudo-labels obtained in the unsupervised stage, solving the dual tasks of main and auxiliary tasks strengthens the model's ability to perceive the stages, thereby forcing the temporal and spatial states to follow the stage changes and play a role; among them, the temporal and spatial states influence each other and are constrained by the stage information.

[0041] Another technical solution adopted in the present invention is:

[0042] A remaining life prediction device, comprising:

[0043] at least one processor;

[0044] at least one memory for storing at least one program;

[0045] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0046] Another technical solution adopted in the present invention is:

[0047] A computer-readable storage medium stores a program executable by a processor, wherein the program executable by the processor is used to perform the method described above when executed by the processor.

[0048] The beneficial effects of the present invention are: the present invention obtains attenuation stage information in an unsupervised manner, and enhances the model's ability to learn the attenuation state in a self-supervised assisted learning manner, thereby enabling the model to better integrate the concept of the attenuation stage into the remaining life prediction task, further improving the model's prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following introduction is made to the drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0050] Figure 1 Schematic diagram of the original vibration signal of the bearing 1_1 in the horizontal direction according to an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of the original vibration signal of the bearing 1_1 in the vertical direction according to an embodiment of the present invention;

[0052] Figure 3 1 is a schematic diagram of the phase division of the original vibration signal in an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of clustering effect in an embodiment of the present invention;

[0054] Figure 5 2 is a graph showing a decrease in the loss value of the loss function according to an embodiment of the present invention;

[0055] Figure 6 1 is a graph showing the predicted remaining life of bearing 1_3 in working condition 1 according to an embodiment of the present invention;

[0056] Figure 7 1 is a flow chart of a method for predicting remaining useful life according to an embodiment of the present invention;

[0057] Figure 8 Schematic diagram of the structure of an adaptive remaining life prediction model for spatiotemporal dual-cell decay state perception according to an embodiment of the present invention;

[0058] Figure 9 This is a structural flow chart of a remaining life prediction method according to an embodiment of the present invention;

[0059] Figure 10Schematic diagram of batch data organization for training a stateful cyclic model in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention. The step numbers in the following embodiments are provided for ease of explanation only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0061] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0062] In the description of the present invention, "several" means one or more, "many" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, while "above," "below," and "within" are understood to include the number itself. The use of "first" and "second" in the description is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0063] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0064] Based on the existing problems, the purpose of the present invention is to provide a remaining life prediction method based on adaptive perception of spatiotemporal dual-cell states. According to the raw data generated in real time by the sensors of production equipment in actual industrial production, a "virtual" health factor is constructed by directly modeling the raw data of such equipment without relying on any preprocessing steps that require expert prior experience. A recurrent network prediction model with embedded convolution operations in which time and space states coexist is adopted, combined with the spatiotemporal dual-cell state and sub-cell state differential mechanism, and state-assisted self-supervised curriculum learning mechanism to perform model training with state, perform adaptive perception of the equipment attenuation state, and realize accurate prediction of equipment life in actual production.

[0065] like Figure 7and Figure 9 As shown, this embodiment provides a method for predicting remaining life, comprising the following steps:

[0066] S1. Obtain the original data of the device.

[0067] Among them, step S1 also includes a data preprocessing step: based on the type of industrial production equipment and deployed sensors, it can be determined what type of data is collected to facilitate the storage and integration of the data, and convert it into a usable form that can be used as system input, including but not limited to eliminating outlier data and processing null value data.

[0068] S2. Perform unsupervised clustering training on the original data to obtain pseudo labels corresponding to the attenuation stage.

[0069] Unsupervised stage clustering analysis is performed based on the original sensor data or preprocessing features in production, that is, deep clustering feature learning is automatically performed according to the algorithm without any human-labeled information, and a deep model with data feature recognition and clustering capabilities is obtained.

[0070] As an optional implementation, based on the raw data generated in real time by sensors of industrial production equipment, the raw data of such equipment is directly modeled, a deep clustering model and the corresponding traditional clustering algorithm are initialized, initial features are obtained through the deep clustering model, clustering is performed using the traditional clustering algorithm, and feedback is then given to the deep model until the convergence condition of the traditional clustering algorithm is reached, i.e., the algorithm is stopped. This algorithm can obtain a model that can adaptively determine which category (stage) the raw data belongs to based on high-level abstract features. These high-level features are the final features obtained by the deep convolutional model. They are distinguishable in a high-dimensional extended space, and have clear boundaries between clusters, so they can be distinguished into different categories, i.e., clusters of different stages.

[0071] Because the health degradation stage of a device is generally unknown, statistical analysis and other methods are traditionally used to determine whether the device is in a state of severe degradation. For example, the three-standard-deviation method is used to determine whether the device is in a state of severe degradation. When the standard deviation of sensor data (such as vibration data) reaches three times or more of the cumulative standard deviation, the device is considered to be in a state of severe degradation. However, this method cannot effectively distinguish between normal degradation and healthy states, or between normal degradation and severe degradation. Moreover, different devices under different operating conditions will have certain differences. In addition, due to the influence of environmental factors, simple statistical methods often cannot accurately indicate the health degradation state of the device. Therefore, an effective algorithm is needed to accurately classify or cluster the device's sensor data. Unsupervised deep clustering learning has been a hot topic of research in recent years. It can effectively perform deep clustering analysis on unclustered data without any manually annotated label information, automatically grouping similar data into the same category, which is of great help in predictive analysis of deep models. This method is an unsupervised clustering algorithm based on deep networks. It mainly maps the original input data into latent space features by the deep model, and then performs clustering analysis on this feature based on high-dimensional distance measurement. The results are then fed back to the deep network for parameter update based on backpropagation.

[0072] S3. Establish a remaining life prediction model based on adaptive perception of the spatiotemporal dual-cell state decay stage, and perform auxiliary self-supervised course training on the remaining life prediction model based on pseudo labels.

[0073] S31. Remaining life prediction modeling based on deep learning:

[0074] See also Figure 8 , the data modeling step of this embodiment will adopt a convolutional recurrent neural network, and a long short-term memory network with dual spatiotemporal cell states embedded in the convolution. This network can make full use of the ability of the convolutional network to extract data spatial features and the ability of the recurrent neural network to extract temporal features, combining the advantages of the two in a native way, and can more effectively and accurately extract features from the original signal with spatiotemporal features, and finally establish a more reliable mapping relationship for the remaining life of the equipment. This network mainly uses the long short-term memory network as the backbone, and replaces the original point multiplication operation of the input-hidden and hidden-hidden transfer functions with convolution operations. Based on the above-mentioned convolution embedded long short-term memory network, the present invention introduces dual spatiotemporal cell states, and state adaptive perception is obtained by differentiating the dynamic and static spatial sub-cells. For the convolution operation, the most core operations are the convolution layer and the respective back-propagation weight updates, which can be expressed as follows:

[0075] Convolution operation:

[0076]

[0077] Convolution back propagation:

[0078]

[0079] Fully connected layer:

[0080] X l =σ(W l X l-1 +b l )

[0081] Back propagation of the fully connected layer:

[0082]

[0083] The long short-term memory network formula for convolutional embedding is:

[0084] Forget Gate:

[0085] f t =σ(W f *[x,h t-1 ]+b f )

[0086] Input Gate:

[0087] i t =σ(W f *[x,h t-1 ]+b i )

[0088] Output Gate:

[0089] o t =σ(W f *[x,h t-1 ]+b o )

[0090] Cell status:

[0091]

[0092]

[0093] Hidden state:

[0094] h t =o t ⊙tanh(c t )

[0095] The long short-term memory network with spatiotemporal dual-state adaptive perception has the following calculation formula for the temporal state part:

[0096] Output candidates:

[0097]

[0098] Input Gate:

[0099]

[0100] Dynamic substates:

[0101]

[0102] Static substate:

[0103]

[0104] Time series cell body state output:

[0105]

[0106] in, represents the dynamic sub-cell state, represents the static sub-cell state, Indicates the hidden state of the upper layer at the current moment,

[0107] The long short-term memory network with spatiotemporal dual-state adaptive perception, the calculation formula of the spatial state part is as follows:

[0108] Spatial output candidates:

[0109]

[0110] Spatial Input Gate:

[0111]

[0112] Spatial forget gate:

[0113]

[0114] Spatial cell body state output:

[0115]

[0116]

[0117] The total cell body output of the model is calculated as follows:

[0118] Overall output gate:

[0119]

[0120] Overall hidden state output:

[0121]

[0122] The calculation formula for the static sub-cell state is as follows:

[0123] Output candidates:

[0124]

[0125] Input Gate:

[0126]

[0127] Forget Gate:

[0128]

[0129] Static cell state:

[0130]

[0131] Output Gate:

[0132]

[0133] Static hidden state:

[0134]

[0135] The calculation formula of dynamic subcells is as follows:

[0136] Dynamic output candidates:

[0137]

[0138] Input Gate:

[0139]

[0140] Forget Gate:

[0141]

[0142] Dynamic cell state:

[0143]

[0144] Output Gate:

[0145]

[0146] Dynamic hidden state:

[0147]

[0148] Where * represents the convolution operation, ⊙ represents the Hadamard product, σ represents the sigmoid activation function, tanh represents the tanh activation function, and softmax represents the softmax activation function.

[0149] As an optional implementation, the specific steps are as follows: First, preprocessed raw data is used as network input, organized either as multidimensional vector data points or as a two-dimensional matrix of multidimensional vectors. Then, the network structure incorporates both spatial and temporal state transfer. Spatial states propagate vertically across different network layers, carrying state information extracted from spatial features at each layer. Temporal states propagate horizontally across different time steps in chronological order, carrying long-term temporal dependency information at each time step. Furthermore, in conjunction with hidden states, both long-term and short-term state transfer is enabled. Building on this dual-feature state transfer, a dynamic and static subcell differential mechanism is introduced. In dynamic subcells, the difference between the current output of the previous layer and the output of the previous time step is used as input. After gate activation, the difference is then output to static subcells for further computation. This mechanism identifies invariant features, focusing on which spatial and temporal states contribute most to the predicted remaining lifespan, thereby enabling preferential selection of effective features. Finally, the health factors extracted from the spatial and temporal features are regressed through a fully connected regression network to obtain a fitted equipment remaining useful life (RUL) curve.

[0150] See also Figure 10 , where the original data or data features x∈R M×N , M is the data dimension, N is the data length, The total number of samples is sampled in a sliding window format to form a data window as the model input. This data window passes through the proposed multi-layer convolutional embedding convolutional recurrent neural network, and finally passes through a prediction layer to output the predicted remaining life. Each network layer consists of two components: temporal state transition and spatial state transition, corresponding to state transitions in the horizontal and vertical directions. The temporal state propagates horizontally with each time step, while the spatial state zigzags vertically between adjacent time steps. The spatial state output of the previous time step serves as the input for the state of the next time step. Furthermore, these two components, respectively, contain an input gate, a forget gate, and an output gate. Each component uses the current moment input and the current hidden layer input as input parameters for state transitions. Both components participate in the calculation of gates and output values, and the final output is obtained by combining the two state outputs. The simultaneous introduction of spatial and temporal cell states allows for both vertical spatial feature extraction and temporal feature extraction based on temporal state, ensuring that state information in either direction is not neglected, ensuring that the extracted data features better reflect the device's true state. At the same time, in the time series modeling of the remaining life prediction task, the concept of cells in cells was introduced, and two sub-cell states were introduced in the overall cell state calculation, namely the dynamic sub-cell state (COC-D) and the static sub-cell state (COC-S). For the current network layer, the temporal differential calculation of the spatially related hidden state output of the previous layer was performed to obtain spatial features that are unrelated to the dynamic noise in the temporal sequence, further reducing the impact of external interference on the prediction results, and obtaining more robust remaining life prediction results.

[0151] S32. Model training based on assisted self-supervised learning:

[0152] A self-supervised curriculum learning mechanism is employed to learn both the primary and secondary tasks, with the secondary task assisting the primary task. By designing a novel deep model loss function, based on the assumption that learning the auxiliary task improves the performance of the primary task, the device health stage classification task is learned simultaneously with the RLL prediction task. Over time, as the device transitions from a healthy state to a period of severe degradation, the learning focus gradually shifts from the stage classification task to the RLL prediction task, thereby better serving the primary RLL prediction task. This step primarily utilizes the unsupervised stage clustering information from step S2 and a two-state flow mechanism model to model the stage perception of the device's RLL. The clustering model obtained in step S2 provides pseudo-stage labels for this step, strengthening the model's stage perception by solving a dual-task problem. This forces the temporal and spatial states to follow the stage changes. The temporal and spatial states influence each other and are constrained by the stage information. This interaction leads to further improvements in the RLL prediction accuracy.

[0153] The new depth model loss function designed in this embodiment has the following specific formula:

[0154] The overall learning loss function is as follows:

[0155]

[0156] λ is a time-varying parameter that represents the weight of auxiliary learning and is calculated as follows:

[0157] λ(i)=max(ζ,λ(0)-∈·i)

[0158] where ζ is the maximum value of λ and ∈ is the decay rate factor.

[0159] The specific basic mean square error and cross entropy loss functions are shown in the following formula:

[0160]

[0161]

[0162] At the beginning of each round of model training, the secondary tasks are learned with greater weight, while the primary tasks are learned with less weight. Over time, the learning of the secondary tasks is gradually reduced while the learning of the primary tasks is strengthened until the end of the round. Throughout the training process, as the device progresses from stable operation to normal status, then to severe degradation, and finally to failure, the focus of learning gradually shifts from stage prediction to remaining life prediction. Alternatively, the same learning focus shifts from early in the training round to later in the training round. Through auxiliary self-supervised training methods, the model's state memory can obtain more information about the device's health degradation state transitions, allowing for dynamic and adaptive data feature extraction based on the device's current stage during online prediction, resulting in more accurate predictions.

[0163] S4. Collect real-time data of the equipment, input the real-time data into the trained remaining life prediction model for prediction, and output the predicted value of the remaining life of the equipment and the corresponding stage value.

[0164] After training the prediction model, real-time data from the device is collected, input into the model for prediction, and the prediction results are output.

[0165] The above method is explained in detail below with reference to the accompanying drawings and specific embodiments.

[0166] This embodiment provides a remaining life prediction method based on spatiotemporal dual-cell state adaptive perception. 1) Obtain the original bearing monitoring data and perform corresponding data preprocessing; 2) Perform unsupervised clustering training on the obtained original data or data features to obtain the corresponding initial stage pseudo-labels; 3) Use the existing data and pseudo-labels to assist in self-supervised course training of the model; 4) Train the spatiotemporal dual-cell state model to obtain spatiotemporal features based on the equipment attenuation state; 5) Use the features to model the remaining life of the equipment through the RUL prediction layer to obtain the final mapping relationship model; 6) Perform real-time data prediction online. The specific steps are as follows:

[0167] Step 1. Data preparation.

[0168] After collecting relevant sensor data and performing certain preprocessing operations, the original sensor data is obtained. This embodiment uses the PRONOSTIA bearing data set of the PHM2012 competition as an example. The data set is collected by interval sampling, with 0.1 second sampling every 10 seconds and a sampling frequency of 25.6kHz, that is, 2560 data points are sampled each time. The data set contains data of three working conditions. Each working condition data contains vibration signals in the vertical and horizontal directions, as well as temperature signals. Since the vibration signal contains more attenuation information, the vibration signal is generally used as the input data of the model. Therefore, the size of each sample data is 2560x2. Among them, the horizontal vibration signal of working condition 1 is as follows Figure 1 As shown, the vertical vibration signal is as follows Figure 2 shown.

[0169] Step 2. Unsupervised stage clustering.

[0170] In this step, the original signal is converted into high-dimensional features using a deep model, and then clustering is performed using a traditional clustering algorithm based on the features. Each clustering result is fed back to the deep model as a "pseudo-label" for parameter update. When the convergence condition of the clustering algorithm is met, the unsupervised clustering process ends. Figure 3 As shown, Figure 3 The figure shows the stage division of the original vibration signal. Specifically, the 2560x2 data obtained in step 1 is input into several one-dimensional convolutional layer networks. After each convolutional layer, it will pass through a pooling layer with a kernel size of 2. The last pooling layer is global pooling, and the data length is finally compressed to a certain size. Here, the number of convolutional layers is 4, and the number of convolution kernels is 16, 32, 64, and 128, respectively, with a size of 10. The final output feature size is 1x128. After obtaining the features, traditional clustering algorithms such as K-Means are used to cluster high-dimensional vectors, and then the single clustering results are fed back to the deep network. The clustering results use the One-Hot encoding mechanism, and the conversion between features and clustering results requires a linear transformation layer. Specifically, the fully connected layer can be used as a linear mapping layer to connect the clustering results and the deep network. Here, the fully connected input is 128 and the output is 3, where 3 is the number of stages that a device may go through in its entire life cycle based on human experience. Ultimately, a deep clustering network that can distinguish raw high-frequency data can be obtained and automatically determine which stage the raw data belongs to. The clustering results are visualized as follows: Figure 4 As shown in the figure, the features that could not be mapped into distinct clusters by the deep model can achieve a better clustering effect after training.

[0171] Step 3. Training of the spatiotemporal dual-cell state model.

[0172] This step is to train a model that combines stage pseudo-labels and a dual-cell state mechanism. The model training adopts a self-supervised auxiliary training mechanism to learn auxiliary tasks and main tasks separately. Specifically, a composite loss function is used. The loss function is obtained by adding two terms, which correspond to the output of the RUL prediction network and the output of the stage prediction respectively. A weight factor is added to gradually shift the focus of learning from the auxiliary task to the main task, namely the remaining life (RUL) prediction, over time. Specifically, the weight factor is multiplied by the RUL prediction loss term, starting from a small value and increasing in a certain way over time (linear, nonlinear), and finally reaching the maximum value. Figure 5 As shown in Figure 2, it is a graph of the loss reduction results after 200 rounds of training.

[0173] Step 4. Online prediction.

[0174] This step mainly uses the trained model to accept unknown data outside the training set to calculate the corresponding remaining life value of the equipment. Specifically, the unknown real-time data needs to be preprocessed first, and then input into the spatiotemporal dual-cell state model to obtain the predicted value of its remaining life and the corresponding stage value. Figure 6 The following are the curve diagram and stage prediction result diagram of the remaining life prediction results. At this point, all steps of this algorithm are completed and the final remaining life prediction results are obtained.

[0175] In summary, this embodiment has the following advantages and beneficial effects compared to the prior art:

[0176] (1) The present invention is based on a dual decay information capture framework. This framework captures the characteristic states of equipment monitoring data in both time and space dimensions. It supplements the original time state with the spatial state and captures the invariant state features in the form of sub-cell state differences. This allows the combination of the spatiotemporal dual cell states to describe the equipment decay process in a more refined manner, focusing more on the static invariant features in the decay state, thereby obtaining accurate remaining life prediction results.

[0177] (2) The present invention uses a state-based prediction framework to predict the remaining life of the equipment, and strengthens it with an auxiliary self-supervised curriculum learning mechanism. This framework can better reflect the state changes in the equipment life decay process than the general framework that assumes that samples are independent. That is, on a global scale, different decay moments have a temporal state flow relationship, rather than allowing the time series cell state to exist only within a single input window. This method not only highlights the global temporal order relationship, but also allows the cyclic model to natively capture these states and adaptively utilize the past state to affect the feature extraction of the current moment.

[0178] (3) The present invention obtains the decay stage information in an unsupervised manner and enhances the model's ability to learn the decay state in a self-supervised assisted learning manner, that is, it has dual learning objectives of remaining life prediction and decay state prediction, so that the model can better integrate the concept of decay stage into the remaining life prediction task, further improving the latter's prediction effect.

[0179] This embodiment further provides a remaining life prediction device, comprising:

[0180] at least one processor;

[0181] at least one memory for storing at least one program;

[0182] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 9 The method shown.

[0183] A remaining life prediction device of this embodiment can execute a remaining life prediction method provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0184] The present application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs Figure 9 The method shown.

[0185] This embodiment also provides a storage medium storing instructions or programs that can execute a remaining life prediction method provided by an embodiment of the method of the present invention. When the instructions or program are run, any combination of implementation steps of the method embodiment can be executed, and the corresponding functions and beneficial effects of the method can be obtained.

[0186] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0187] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0188] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0189] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0190] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0191] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0192] In the above description of this specification, reference to the terms "one embodiment / example," "another embodiment / example," or "certain embodiments / examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0193] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0194] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for predicting remaining service life, characterized in that: The following steps are involved: Get the raw data of the device; Performing unsupervised clustering training on the original data to obtain pseudo labels corresponding to the attenuation stage; Establish a remaining life prediction model based on adaptive perception of the spatiotemporal dual-cell state decay stage, and perform auxiliary self-supervised course training on the remaining life prediction model based on pseudo labels; Collect real-time data of the equipment, input the real-time data into the trained remaining life prediction model for prediction, and output the predicted value of the remaining life of the equipment and the corresponding stage value; The remaining life prediction model adopts the long short-term memory network as the backbone, and replaces the original point product operation of the input-hidden and hidden-hidden transfer functions of the long short-term memory network with a convolution operation; The spatial state is propagated vertically across different network layers, carrying state information for spatial feature extraction at each layer. The temporal state is propagated horizontally across different time steps in chronological order, carrying long-term dependency state information at each time step. Furthermore, the hidden state is used to achieve the transmission of long-term and short-term states. Based on the dual-feature state transfer of space and time, a dynamic and static sub-cell differential mechanism is introduced. In the dynamic sub-cell, the difference between the current output of the previous layer and the output of the previous time step is used as input. After calculation, it is output to the static sub-cell for calculation, and the invariant features that contribute more to the predicted remaining life value in space and time are obtained. Finally, the health factors extracted from the spatial and temporal features are regressed through a fully connected regression network to obtain a fitted equipment remaining service life curve.

2. A method for predicting remaining service life according to claim 1, characterized in that: The performing unsupervised clustering training on the original data to obtain pseudo labels corresponding to the attenuation stage includes: Use a deep model to convert raw data into high-dimensional features, cluster the high-dimensional features, and obtain pseudo labels based on each clustering result; Feed the pseudo labels back to the deep model and update the parameters to finally obtain a deep clustering network; The deep clustering network is used to identify the decay stage corresponding to the original data.

3. The method for predicting remaining service life according to claim 1, wherein: The remaining life prediction model includes a multi-layer convolutional embedded recurrent neural network and a prediction layer; The original data is sampled in a sliding window manner to obtain window data as the input of the remaining life prediction model; After the window data passes through the recurrent neural network and prediction layer, the remaining life prediction value is output; Among them, each layer of the recurrent neural network consists of two parts: temporal state transfer and spatial state transfer. The temporal state propagates horizontally with the time steps, while the spatial state zigzags between adjacent time steps in the vertical direction.

4. The method for predicting remaining service life according to claim 1, wherein: The remaining life prediction model adopts a convolutional recurrent neural network and a long short-term memory network with dual spatiotemporal cell states embedded in the convolution; The long short-term memory network makes full use of the ability of the convolutional network to extract data spatial features and the ability of the recurrent neural network to extract temporal features, and can extract features from original signals with spatiotemporal features, and ultimately establish a mapping relationship for the remaining life of the equipment.

5. A method for predicting remaining service life according to claim 4, characterized in that: The spatial state of the next time step can be obtained by upsampling and accumulating the spatial state of the last layer of the previous time step layer by layer, as shown in the following formula: in, It is the spatial state input of the first layer. Upsampling refers to the upsampling algorithm. Refers to the spatial state of the lth layer.

6. The method for predicting remaining service life according to claim 1, wherein: The auxiliary self-supervised course training of the remaining life prediction model based on pseudo labels includes: By designing a deep model loss function and improving the performance of the main task based on auxiliary task learning, the equipment health stage classification task is learned simultaneously with the remaining life prediction task. Over time, from a healthy state to the final stage of severe degradation, the learning focus gradually shifts from the stage classification task to the remaining life prediction task, so as to better serve the main task of remaining life prediction.

7. The method for predicting remaining service life according to claim 1, wherein: The expression of the loss function is as follows: Where λ is a time-varying parameter representing the weight of auxiliary learning and is calculated using the following formula: λ(i)=max(ζ,λ(0)-∈·i) Where ζ is the maximum value of λ, ∈ is the decay rate factor; The expressions of mean square error and cross entropy loss functions are as follows: Where, is the true value of the remaining life, is the predicted value of remaining life, n is the total number of samples, y i is the true value of the i-th point, is the predicted value of the i-th point, P(y i ) is the distribution of true category labels, is the distribution of predicted categories, j represents the jth category, p is the total number of samples, and q is the total number of categories.

8. The method for predicting remaining service life according to claim 1, wherein: The auxiliary self-supervised course training of the remaining life prediction model based on pseudo labels includes: Based on the pseudo-labels obtained in the unsupervised stage, the model's ability to perceive the stage is strengthened in the form of a dual task, namely the main task and the auxiliary task, thereby forcing the temporal and spatial states to follow the stage changes and play a role; among them, the temporal and spatial states influence each other and are constrained by the stage information.

9. A remaining life prediction device, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to perform the method according to any one of claims 1 to 8 when executed by the processor.

Images