Remaining service life prediction method based on adaptive random physical information neural network

Abstract

The embodiment of the invention provides a residual service life prediction method based on an adaptive random physical information neural network, and the method comprises the steps: obtaining equipment operation data; based on a feature extraction module, equipment degradation features are extracted from the equipment operation data, and the equipment degradation features comprise feature association information and time sequence dependence information; the feature extraction module is constructed on the basis of a time sequence context sensing Transform; based on the predictor, according to the equipment degradation characteristics, determining a residual service life prediction value; determining each order derivative of the equipment degradation characteristic; based on a self-adaptive random physical loss function, determining an input sample according to the residual service life prediction value and each order derivative of the equipment degradation characteristic; the adaptive random physical loss function is constructed based on a partial differential equation and a stochastic differential equation; and inputting the input sample into a prediction module to obtain the residual service life, and constructing the prediction module based on a meta-learning method. The technical scheme provided by the invention is used for solving the problems of few samples, black box effect and the like.

Description

Technical Field

[0001] This application relates to the field of point of interest recommendation technology, and in particular to a method for predicting remaining lifetime based on an adaptive random physical information neural network. Background Technology

[0002] Remaining useful life (RUL) prediction, as a core component of the Predictive Failure and Health Management (PHM) system, aims to provide a scientific basis for maintenance decisions by predicting the remaining operating time or cycle life of equipment.

[0003] The latest research in this field focuses on three major technological breakthroughs: enhancing the model's ability to extract time-series features, developing spatiotemporal fusion feature extraction mechanisms, and optimizing attention mechanism design. These advancements further strengthen the robustness, versatility, and ease of use of deep learning methods, enabling them to continue playing a crucial role in industrial big data environments.

[0004] However, existing methods have the following problems: 1. Existing methods primarily rely on statistical patterns in the data, neglecting the rich physical knowledge inherent in equipment operation. Models lacking physical information easily become "black boxes," their predictions unable to be explained by underlying physical principles. This leads to the model's decision-making process being detached from real-world physical mechanisms, making it difficult to verify the rationality of predictions from a causal perspective. It also weakens the model's generalization ability, especially in scenarios with data distribution shifts or requiring extrapolation, where the reliability of predictions significantly decreases. Furthermore, the model's over-reliance on training data leads to overfitting when data is limited, making it difficult to capture true physical laws.

[0005] 2. Scarcity of target domain degradation data for industrial equipment prediction. Due to complex operating conditions and safety and economic constraints, obtaining full lifecycle failure data is extremely difficult. Although there are sufficient training samples in the source domain, direct cross-domain migration will cause domain shift, leading to model performance degradation and reduced engineering reliability. Summary of the Invention

[0006] In view of the above-mentioned solutions, this application aims to propose a remaining useful life prediction method based on an adaptive random physical information neural network to solve at least one of the above problems.

[0007] In a first aspect, one or more embodiments of this application provide a method for predicting remaining useful life based on an adaptive stochastic physical information neural network, wherein the adaptive stochastic physical information neural network includes: a feature extraction module, a predictor, an adaptive stochastic physical loss function, and a prediction module, and the method includes: Acquire equipment operation data; Based on the feature extraction module, device degradation features are extracted from the device operation data. The device degradation features include feature association information and temporal dependency information. The feature extraction module is built based on a temporal context-aware Transformer. Based on the predictor, the remaining useful life is predicted according to the device degradation characteristics; Determine the derivatives of the degradation characteristics of the device; Based on the adaptive stochastic physical loss function, the input samples are determined according to the predicted remaining useful life and the derivatives of the equipment degradation characteristics; the adaptive stochastic physical loss function is constructed based on partial differential equations and stochastic differential equations. The input sample is input into the prediction module to obtain the remaining lifespan. The prediction module is constructed based on a meta-learning method.

[0008] Furthermore, the equipment degradation characteristics are extracted from the equipment operation data, including: The device operation data is converted into an input sequence; The input sequence is mapped to the embedding space through a linear transformation to obtain the embedding features; Determine the temporal dynamic features of the input sequence and encode the temporal dynamic features of the sequence; The embedded features and the encoded sequence time dynamic features are fused to obtain the fused features; Extract the device degradation features from the fusion features.

[0009] Furthermore, the feature extraction module includes multiple stacked Transformer encoding layers; The multi-layered stacked Transformer coding layers are used to extract the device degradation features; Each Transformer coding layer contains a multi-head self-attention layer and a feedforward network.

[0010] Furthermore, the method also includes: Based on the aforementioned partial differential equations, physical law constraints are constructed; Based on the aforementioned stochastic differential equation, a stochastic fluctuation constraint is constructed; Based on the random fluctuation constraints and the physical law constraints, a physical loss function is constructed; Based on the physical loss function and the preset data loss function, a dynamic weighted composite loss function is constructed; Based on the aforementioned dynamic weighted composite loss function, the adaptive stochastic physical loss function is constructed.

[0011] Furthermore, the dynamically weighted composite loss function is specifically as follows: ; Where α is the Wiener weight, L SDE Let L be the loss function corresponding to the stochastic differential equation. PDE L is the loss function corresponding to the partial differential equation. phy L is the loss function corresponding to the physical drive. Data This is the loss function for the data items.

[0012] Secondly, one or more embodiments of this application provide a device for predicting remaining useful life based on an adaptive random physical information neural network, comprising: The acquisition module is used to acquire device operating data; The feature extraction module is used to extract device degradation features from the device operation data. The device degradation features include feature association information and temporal dependency information. The feature extraction module is built based on a temporal context-aware Transformer. A predictor is used to determine a predicted remaining useful life based on the degradation characteristics of the device; A data processing module is used to determine the derivatives of the degradation characteristics of the device. An adaptive stochastic physical loss function module is used to determine input samples based on the predicted remaining useful life and the derivatives of the equipment degradation characteristics; the adaptive stochastic physical loss function is constructed based on partial differential equations and stochastic differential equations. A prediction module is used to obtain the remaining useful life, and the prediction module is constructed based on a meta-learning method.

[0013] Furthermore, the feature extraction module is used to convert the device operation data into an input sequence; map the input sequence to an embedding space through a linear transformation to obtain embedded features; determine the temporal dynamic features of the input sequence and encode the temporal dynamic features of the sequence; fuse the embedded features and the encoded temporal dynamic features of the sequence to obtain fused features; and extract the device degradation features from the fused features.

[0014] Furthermore, the feature extraction module includes multiple stacked Transformer encoding layers; The multi-layered stacked Transformer coding layers are used to extract the device degradation features; Each Transformer coding layer contains a multi-head self-attention layer and a feedforward network.

[0015] Furthermore, the adaptive stochastic physical loss function module is used to construct physical law constraints based on the partial differential equations; construct stochastic fluctuation constraints based on the stochastic differential equations; construct a physical loss function based on the stochastic fluctuation constraints and the physical law constraints; construct a dynamically weighted composite loss function based on the physical loss function and a preset data loss function; and construct the adaptive stochastic physical loss function based on the dynamically weighted composite loss function.

[0016] Thirdly, embodiments of this application provide a storage medium for storing computer-executable instructions, characterized in that, when the computer-executable instructions are executed, they implement the steps of the remaining useful life prediction method based on an adaptive random physical information neural network as described in the first aspect.

[0017] Compared with the prior art, this application can achieve at least the following technical effects: A feature extraction network built on a temporal context-aware Transformer mines feature associations and temporal dependencies to output device degradation features, enabling in-depth analysis of factors influencing RUL (Range Limit Indicator). These degradation features are input to the Predictor and directly mapped to RUL predictions via a fully connected network, calculating the data-driven loss. Simultaneously, automatic differentiation is used to obtain the derivatives of RUL with respect to time and hidden states, which are then concatenated and input into a partial differential equation (PDE). A Wiener process is introduced to construct a stochastic differential equation (SDE) to model fluctuations. The PDE and SDE losses constrain physical consistency, ensuring the prediction results can be explained by physical principles. Finally, a prediction module built using meta-learning methods determines the RUL, enabling rapid adaptation to a limited amount of target domain data during testing. This allows for accurate RUL prediction under new equipment or operating conditions, effectively addressing the challenge of RUL prediction with few samples and multiple operating conditions in industrial scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in one or more 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 some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a method for predicting remaining useful life based on an adaptive random physical information neural network, provided for one or more embodiments of this application; Figure 2 This is a schematic diagram of the structure of a device for predicting the remaining useful life based on an adaptive random physical information neural network, provided for one or more embodiments of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this application, the technical solutions in one or more embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on one or more embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this application.

[0021] This application provides a remaining lifetime prediction method based on an Adaptive Stochastic Physics-Informed Neural Network (AR-ASPINN, AdamW-Reptile based Adaptive Stochastic Physics-Informed Neural Network). The AR-ASPINN includes a feature extraction module, a predictor, an adaptive stochastic physics loss function, and a prediction module. As a deep learning architecture, the Physical Information Neural Network (PINN) integrates prior physical knowledge into the neural network, providing a promising solution to the challenge of insufficient physical information during model training. The core idea of ​​PINN is to introduce the partial differential equations describing the dynamic characteristics of the system as regularization terms or flexible constraints into the loss function of the neural network, thereby guiding the network to learn solutions that conform to physical laws.

[0022] Specific methods are as follows Figure 1 As shown, it includes the following steps: Step 1: Obtain equipment operation data.

[0023] In this embodiment, the monitoring data during equipment operation is first preprocessed, including data noise reduction, time-frequency domain feature evaluation and selection, and meta-tasks are constructed according to different devices. Equipment operation data includes: time series data, such as parameters reflecting real-time status, such as vibration amplitude, frequency (rotating machinery), temperature, and voltage / current (electronic equipment); fault records, such as historical data such as the number of faults, mean time between failures, and mean time to repair; material properties, such as the hardness and corrosion resistance of bearings; and environmental stresses, such as temperature cycling range, humidity, and mechanical vibration, which accelerate aging.

[0024] Step 2: Based on the feature extraction module, extract equipment degradation features from the equipment operation data.

[0025] In this embodiment, the device degradation feature includes feature association information and temporal dependency information. The temporal dependency information characterizes the association between the same feature at different time steps. That is, it represents the correlation between changes in a single feature over time; specifically, the feature value at a certain moment is affected by the feature values ​​at historical time steps, reflecting "continuity in the time dimension."

[0026] Feature association information: The relationship between different features at the same time step. That is, the mutual influence or coordinated change relationship between the operating features of multiple devices at the same time, reflecting the "relationship in the spatial dimension".

[0027] Both time-dependent information and feature-related information can reveal the physical laws governing the lifespan of equipment, laying the foundation for subsequent enhanced prediction results to conform to these physical laws.

[0028] In this embodiment, the feature extraction module is built based on Time Context-Aware Transformer (TCAT). Specifically, for high-dimensional, multi-source, and noisy monitoring time series of industrial machinery and equipment, TCAT maps the original input sequence X∈RB×T×F into a low-dimensional hidden state space H∈RB×D through feature embedding, temporal encoding, and self-attention mechanisms, providing reliable input for remaining lifetime prediction (RUL) and physical law mining. The overall module architecture integrates the physical priors of time series modeling with the expressive power of deep learning, achieving adaptive extraction of equipment degradation features.

[0029] Step 3: Based on the predictor, determine the predicted remaining useful life value according to the equipment degradation characteristics.

[0030] In the embodiments of this application, the predictor is directly mapped to the RUL predicted value via a fully connected network, laying the foundation for calculating the data-driven loss.

[0031] Step 4: Determine the derivatives of the equipment degradation characteristics.

[0032] In the embodiments of this application, the degradation of industrial equipment generally follows observable physical laws, such as Paris's law of mechanical wear (crack propagation rate is related to stress intensity factor) and Arrhenius equation of thermal aging (reaction rate is exponentially related to temperature). These laws can be abstracted into PDE (partial differential equation) form, that is, derivative form.

[0033] Step 5: Based on the adaptive stochastic physical loss function, determine the input samples according to the predicted remaining useful life and the derivatives of the equipment degradation characteristics.

[0034] In this embodiment, the adaptive stochastic physical loss function is constructed based on partial differential equations and stochastic differential equations. The adaptive stochastic physical loss function based on partial differential equations is used to ensure physical properties such as the monotonically decreasing RUL over time. The actual degradation process is affected by random factors such as environmental noise and sudden load changes, necessitating the introduction of stochastic differential equations to characterize the uncertainties.

[0035] Specifically, the PDE constraint mainly calculates the first derivative of RUL with respect to time t:

[0036] and the derivatives u with respect to the hidden state h. h ,u hh The derivative information is concatenated with the hidden state to form a feature sequence [h, u]. h ,u hh In AR-ASPINN, PDE constraints are implemented through an Attention LSTM Encoder (ALE). This module integrates a multi-head attention mechanism with an LSTM network to capture the long-range dependencies and temporal dynamics of degenerate features. When a feature sequence is input into the ALE, the multi-head attention mechanism first captures the correlations between different feature dimensions, generating attention-enhanced feature representations.

[0037] Where MultiHeadAttn is the multi-head attention mechanism function, and Featseq is the feature sequence. Subsequently, the temporal dependencies of the feature sequences are captured through an LSTM layer, outputting an encoded vector containing the temporal context:

[0038] Among them, Feat represents the features, and Attn represents the attention mechanism.

[0039] The encoded vector is mapped to the degradation rate function f through a two-layer fully connected network to construct the PDE residual:

[0040] By minimizing the loss function, the prediction is forced to conform to the degradation dynamics. The minimized loss function is:

[0041] The stochastic differential equation (SDE) takes the form of:

[0042] Where μ is the drift coefficient (average degradation rate), σ is the diffusion coefficient (fluctuation intensity), and dWt is the Wiener increment with mean 0 and variance dt. This equation describes the evolution of RUL with random perturbations superimposed on a deterministic trend.

[0043] During the training phase, noise dWt that conforms to the statistical characteristics of the Wiener process is generated, and a noise attenuation strategy is introduced. The noise coefficient (noise_scale) decreases with each training round, thus achieving a transition from "exploring random fluctuations" to "focusing on deterministic trends".

[0044]

[0045] Calculate the time derivative u of RUL t And compare it with the theoretical expression of SDE:

[0046] Here, Drift and Diffusion are learnable parameters, corresponding to the drift coefficient μ and diffusion coefficient σ, respectively.

[0047] To control noise introduction during training and testing, random noise is used during training to learn fluctuation characteristics, while only the drift term is retained during testing. This ensures the stability of the prediction.

[0048] Step 6: Input the input sample into the prediction module to obtain the remaining lifespan. The prediction module is built based on the meta-learning method.

[0049] In this embodiment, to enable the proposed AR-ASPINN to quickly adapt to new tasks with only a small number of samples, a meta-learning framework based on AdamW-Reptile (AR) is designed. This framework learns a set of optimized meta-parameters Φ through explicit regularization and first-order optimization strategies. This makes it an efficient starting point for adapting to the new RUL prediction task.

[0050] The meta-learning framework is specifically as follows: objective function The core objective of the AR framework is to find a set of meta-parameters Φ This allows the model to start from this initial point and perform tasks on T. p A small amount of training data After performing k steps of AdamW optimization (with an inner loop learning rate of α and an initial value of α), ), can be used in the corresponding verification data To reach the minimum expected loss:

[0051] Where L Total-pThis represents the total loss of the AR-ASPINN framework on task Tp.

[0052] Reptile employs a first-order optimization strategy to avoid the overhead of second-order derivatives. The AR algorithm proposed in this application is based on the Reptile mechanism and implements inner loop updates through the AdamW optimizer, eliminating the need for explicit computation of meta-gradients and implicitly optimizing the objective function.

[0053] 3.3.2 AWR Meta-Learning Algorithm The AR-ASPINN framework's adaptive regularization-based meta-parameter Φ update process comprises two core stages: meta-training and fast adaptation.

[0054] The goal of the meta-training phase is to learn an optimized set of meta-parameters Φ from a large number of source domain tasks. This gives it good generalization ability and rapid adaptation potential. Its meta-training phase iteratively executes the following process: 1. First, randomly sample a mini-batch containing B tasks from the source domain task distribution P(T). Complete batch sampling for the task; 2. Subsequently, for each task T in this batch p (p=1,…,B) Parallel execution of inner loop adaptation, first initializing task parameters. (Using the latest meta-parameters), and then using the learning rate α on the training data for task Tp. Perform k-step AdamW optimization updates to minimize the loss:

[0055] Where the gradient at step i is:

[0056] For the task parameter θ at the i-th The gradient at step 1, AdamW update involves an exponential moving average estimate of the first moment (momentum) m and the second moment v of the gradient, combined with an explicit weight decay mechanism:

[0057] Where β1, β2∈[0,1) are the exponential decay rates of the moment estimate. λ represents the numerical stability term; λ is the weight decay coefficient. The final parameters of task p after k steps of adaptation are obtained. .

[0058] 3. In the outer loop, based on the adaptation results of all tasks in the batch, calculate the update amount of the meta-parameter Φ. The update rule is:

[0059] Where η is the outer loop (meta) learning rate. This step causes the meta-parameter Φ to move along the initial point Φ. old After each task in the batch is adapted, the parameters are... The average vector direction shifts.

[0060] When the rapid adaptation phase faces a new target domain task, the final meta-parameters Φ obtained from meta-training are... This serves as the initial point for the new task model. It utilizes the very few supporting samples provided by the new task. From Φ By executing k′ steps of gradient updates similar to the inner loop of meta-training, model parameters adapted to the characteristics of the new task can be quickly obtained. This is used for subsequent RUL prediction.

[0061] In this embodiment of the application, the AR update direction is as follows:

[0062] It is to analyze the displacement of the intrinsic parameters of the task. The expected characteristics It is the result of the meta-parameter Φ being optimized by AdamW in k steps of task p, representing the displacement of Φ along the task loss Ltotal-p gradient iterated k times.

[0063] For simple SGD (k=1), the Taylor expansion approximates it as:

[0064]

[0065] For AdamW optimization (k>1), the update rule, which includes momentum, second moment, and weight decay, can still be approximately expanded as follows:

[0066]

[0067]

[0068] in, It's a mini-batch gradient. Let Φ be the Hessian matrix. It is the weight decay term of AdamW.

[0069] It can be decomposed into three types of optimization logic: the first is the average gradient term:

[0070] It originates from the iterative accumulation of the loss gradients of each task in the inner loop, driving the meta-parameters to move towards the initial point Φ for all tasks. old The average loss at each point evolves in the direction of minimizing, simulating the global optimization trend of joint training; Second, the weight decay term:

[0071] It relies on AdamW's explicit regularization mechanism to adapt parameters after task adaptation. It naturally carries parameter norm constraints, and indirectly guides the meta-parameters to compress their own norms through the difference with Φold, thereby suppressing overfitting and enhancing generalization ability; Thirdly, the gradient inner product term:

[0072] It captures the interaction characteristics of gradients from different batches of data within a task at the meta-parameters, and strengthens the consistency alignment of gradients within the task through iterative updates, enabling the meta-parameters to have the potential to quickly respond to gradient changes in new tasks and improve the adaptation efficiency in small sample scenarios.

[0073] Through Taylor expansion analysis, the expectation of dp can be approximated as:

[0074] In summary, the core advantage of the AR meta-learning framework proposed in this application lies in achieving more stable parameter norm control through an explicit weight decay mechanism, generating a simpler initial model to reduce dependence on specific tasks, and having a regularization mechanism equivalent to Gaussian priors under the Bayesian framework with a more rigorous theoretical foundation. At the same time, it can better balance data fitting and preservation of physical laws, enhance the synergy of physical constraints, and effectively suppress overfitting and improve the generalization stability of the model in small sample scenarios.

[0075] In this embodiment of the application, extracting device degradation features from the device operating data includes: The device operation data is converted into an input sequence; the input sequence is mapped to an embedding space through a linear transformation to obtain embedding features; the temporal dynamic features of the input sequence are determined and encoded; the embedding features and the encoded temporal dynamic features are fused to obtain fused features; and the device degradation features are extracted from the fused features.

[0076] The input sequence is arranged in chronological order, naturally preserving the temporal order of each feature, serving as the "original carrier" of temporal dependencies. Embedded features provide the "foundation for feature association," while temporal encoding enhances temporal dependencies. After fusion, temporal dependencies combine sequence structure and temporal dynamics, capturing dependencies over longer time steps. Feature associations combine high-dimensional feature associations with temporal synchronization, capturing "long-range feature associations across time steps." The feature extraction module comprises multiple stacked Transformer encoding layers; these layers are used to extract the device degradation features; each Transformer encoding layer includes a multi-head self-attention layer and a feedforward network.

[0077] In addition, this application employs two encoding methods. One is learnable temporal embedding, where the parameters are obtained by first randomly initializing them, and then jointly training and optimizing them end-to-end with other Transformer parameters using labeled industrial time-series data and task loss functions to ultimately adapt to the time patterns of a specific scenario. The other is sine and cosine encoding, where the parameters are obtained by predefining them according to mathematical formulas, and only the time step calculation or frequency scaling factor can be adjusted to adapt to the industrial time granularity. Even in semi-fixed expansion, only a small number of coefficients are fine-tuned, requiring no extensive training.

[0078] The specific procedures for extracting equipment degradation characteristics are as follows: The feature vector xt∈RF at each time step of the original input sequence X=[x1,x2,…,xT] is mapped to the embedding space through a linear transformation:

[0079] Among them W e ∈R F×D and b e ∈R D is the learnable parameter for feature embedding, B is the batch size, T is the sequence length, F is the feature dimension, and D is the embedding dimension.

[0080] This layer projects the high-dimensional original features into a unified low-dimensional space through a nonlinear mapping, reducing feature redundancy while preserving key information during device degradation. To capture the dynamic characteristics of time series data, the module introduces a temporal context encoding (Et) and fuses it with feature embeddings via broadcast addition.

[0081] There are two encoding methods: one is Learnable Time Embedding.

[0082] Where Wt is the embedding matrix with learnable temporal embeddings, and t∈{1,2,…,T} is the time index. This approach allows the model to adaptively learn temporal representations based on data characteristics, making it particularly suitable for capturing nonlinear temporal dependencies in complex device degradation processes.

[0083] Second is sine / cosine encoding:

[0084] Where t is the time step and i is the dimension index, the position encoding values ​​PE(t,i) of the sine and cosine encoding constitute the position encoding matrix Et. This fixed encoding method provides relative position information to the model through the periodicity of trigonometric functions, which helps to capture periodic patterns in equipment operation, such as the periodic impact characteristics of bearing failure.

[0085] Xtime features are extracted using stacked Transformer encoding layers, each layer containing multi-head attention and a feed-forward network.

[0086] The encoder structure for each layer is as follows:

[0087] The multi-head self-attention mechanism is defined as:

[0088] Among them W o Head The concatenation transformation matrix for the multi-head attention output is calculated for each head as follows:

[0089] The transformation matrix for the query, key, and value of the i-th head in multi-head attention is given. The feedforward network consists of two non-linear transformation layers.

[0090] W1, b1, W2, b2 are learnable parameters of the feedforward network (FFN). The Transformer output sequence H ∈ RB × T × D is subjected to time-dimensional average pooling and mapped to the target dimension through a linear layer.

[0091] Among them W o b oThe learnable parameters of the output layer, where C is the output dimension, can be set to RUL predicted value or number of health status categories depending on the task.

[0092] The proposed Temporal Context-Aware Transformer effectively captures temporal dependencies and feature associations in industrial time series by combining temporal encoding with a multi-head self-attention mechanism. It also improves model training stability through residual connections and layer normalization. Compared to traditional models, this module offers significant advantages in processing long-sequence data, better capturing long-term dependencies in equipment degradation processes.

[0093] Preferably, the loss function corresponding to the partial differential equation and the loss function corresponding to the stochastic differential equation are fused to balance long-term prediction conformity to physical laws and avoid overfitting and short-term noise. The fusion process is as follows: Based on partial differential equations, physical law constraints are constructed; based on stochastic differential equations, stochastic fluctuation constraints are constructed; based on stochastic fluctuation constraints and physical law constraints, a physical loss function is constructed; based on the physical loss function and a preset data loss function, a dynamic weighted composite loss function is constructed; based on the dynamic weighted composite loss function, an adaptive stochastic physical loss function is constructed.

[0094] Specifically, to balance deterministic patterns with random fluctuations, a dynamic weighted composite loss function is designed:

[0095] Where α is the Wiener weight, L SDE Let L be the loss function corresponding to the stochastic differential equation. PDE L is the loss function corresponding to the partial differential equation. phy L is the loss function corresponding to the physical drive. Data α is the loss function for the data items. The value of α is dynamically adjusted with each training round. In the early stage of training, α is set to a higher value to force the model to learn the random fluctuation patterns in the data and adapt to the personalized noise characteristics of different devices. In the later stage of training, α decays linearly to a minimum value to highlight the dominant role of PDE constraints, ensure that long-term predictions conform to physical laws, and avoid overfitting to short-term noise.

[0096] AR-ASPINN achieves an organic unity between "prior guidance of physical laws on model structure" and "data-driven learning of detailed features" by jointly embedding PDE and SDE, combined with an attention-based LSTM encoder (ALE) to model the temporal and dependency relationships of degenerate features. PDE ensures the physical rationality of predictions, while SDE captures the uncertainty of the degradation process. SDE allows the model to learn the core statistical features of the physical system at a lower cost in the early stages. This mechanism alleviates the initial learning difficulties caused by the strong constraints of PDE, laying the foundation for accurate physical modeling in later PDE stages, rather than simply adding noise, thus achieving a progressive learning from "random statistics" to "deterministic laws." Furthermore, ALE enhances feature representation capabilities through attention mechanisms and LSTM, and the dynamic weighting strategy further improves the model's robustness in small-sample, multi-condition scenarios.

[0097] This application provides a device for predicting remaining useful life based on an adaptive random physical information neural network, such as... Figure 2 As shown, it includes: Module 201 is used to acquire device operating data; Feature extraction module 202 is used to extract device degradation features from the device operation data. The device degradation features include feature association information and temporal dependency information. The feature extraction module is built based on temporal context-aware Transformer. Predictor 203 is used to determine a predicted value of remaining useful life based on the degradation characteristics of the device; Data processing module 204 is used to determine the derivatives of the degradation characteristics of the device. The adaptive stochastic physical loss function module 205 is used to determine the input samples based on the predicted remaining useful life and the derivatives of the equipment degradation characteristics; the adaptive stochastic physical loss function is constructed based on partial differential equations and stochastic differential equations. The prediction module 206 is used to obtain the remaining useful life, and the prediction module is constructed based on the meta-learning method.

[0098] In this embodiment, the feature extraction module is used to convert the device operation data into an input sequence; map the input sequence to an embedding space through a linear transformation to obtain embedded features; determine the temporal dynamic features of the input sequence and encode the temporal dynamic features of the sequence; fuse the embedded features and the encoded temporal dynamic features of the sequence to obtain fused features; and extract the device degradation features from the fused features.

[0099] In this embodiment, the feature extraction module includes multiple stacked Transformer coding layers; the multiple stacked Transformer coding layers are used to extract the device degradation features; each Transformer coding layer includes a multi-head self-attention layer and a feedforward network.

[0100] In this embodiment, the adaptive stochastic physical loss function module is used to construct physical law constraints based on the partial differential equation; construct stochastic fluctuation constraints based on the stochastic differential equation; construct a physical loss function based on the stochastic fluctuation constraints and the physical law constraints; construct a dynamically weighted composite loss function based on the physical loss function and a preset data loss function; and construct the adaptive stochastic physical loss function based on the dynamically weighted composite loss function.

[0101] This application provides a storage medium for storing computer-executable instructions, characterized in that, when executed, the computer-executable instructions implement the steps of the remaining useful life prediction method based on an adaptive random physical information neural network as described in any one of the embodiments.

[0102] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0103] In the 1930s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many improvements to the methodology today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that an improvement to the methodology cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0104] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0105] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0106] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, in implementing the embodiments of this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0107] Those skilled in the art will understand that one or more embodiments of this application can be provided as a method, system, or computer program product. Therefore, one or more embodiments of this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0112] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0113] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined in this application, computer-readable media does not include transient media, such as modulated data signals and carrier waves.

[0114] It should also be noted that 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 limitation, 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.

[0115] One or more embodiments of this application can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. One or more embodiments of this application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can reside in local and remote computer storage media, including storage devices.

[0116] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0117] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the remaining useful life based on an adaptive random physical information neural network, characterized in that, The adaptive stochastic physics information neural network comprises a feature extraction module, a predictor, an adaptive stochastic physics loss function and a prediction module, and the method comprises: obtaining equipment operation data; extracting equipment degradation features from the equipment operation data based on the feature extraction module, wherein the equipment degradation features comprise feature correlation information and time sequence dependence information; and the feature extraction module is constructed based on a time sequence context perception Transformer; determining a remaining useful life prediction value based on the equipment degradation features based on the predictor; determining derivatives of the equipment degradation features; determining an input sample based on the remaining useful life prediction value and the derivatives of the equipment degradation features based on an adaptive stochastic physics loss function; the adaptive stochastic physics loss function is constructed based on a partial differential equation and a stochastic differential equation; inputting the input sample into the prediction module to obtain a remaining useful life; and the prediction module is constructed based on a meta-learning method.

2. The method of claim 1, wherein the equipment degradation features are extracted from the equipment operation data by: converting the equipment operation data into an input sequence; mapping the input sequence to an embedding space through linear transformation to obtain embedding features; determining sequence time dynamic features of the input sequence and encoding the sequence time dynamic features; fusing the embedding features and the encoded sequence time dynamic features to obtain fused features; and extracting the equipment degradation features from the fused features.

3. The method of claim 2, wherein the feature extraction module comprises a plurality of stacked Transformer encoding layers; the plurality of stacked Transformer encoding layers are used to extract the equipment degradation features; each of the Transformer encoding layers comprises a multi-head self-attention layer and a feedforward network. The method further comprises: constructing a physical law constraint based on the partial differential equation; 4. The method of claim 1, wherein, constructing a stochastic volatility constraint based on the stochastic differential equation; constructing a physical loss function based on the stochastic volatility constraint and the physical law constraint; constructing a dynamic weighted composite loss function based on the physical loss function and a preset data loss function; and constructing the adaptive stochastic physics loss function based on the dynamic weighted composite loss function.

5. The method of claim 4, wherein the dynamic weighted composite loss function is specifically: comprising: an obtaining module configured to obtain equipment operation data; a feature extraction module configured to extract equipment degradation features from the equipment operation data, wherein the equipment degradation features comprise feature correlation information and time sequence dependence information; and the feature extraction module is constructed based on a time sequence context perception Transformer; ； where a is a Wiener weight, L SDE is the loss function corresponding to the stochastic differential equation, L PDE is the loss function corresponding to the partial differential equation, L phy is the loss function corresponding to the physical driver, L Data is the data item loss function.

6. An adaptive random physical information neural network-based residual useful life prediction device, characterized by, a predictor configured to determine a remaining useful life prediction value based on the equipment degradation features; a data processing module configured to determine derivatives of the equipment degradation features; and a prediction module configured to input an input sample into the prediction module to obtain a remaining useful life; and the prediction module is constructed based on a meta-learning method. ​ ​ An adaptive stochastic physical loss function module is configured to determine an input sample based on the remaining useful life prediction value and derivatives of the equipment degradation feature; the adaptive stochastic physical loss function is constructed based on a partial differential equation and a stochastic differential equation; A prediction module is configured to obtain a remaining useful life, and the prediction module is constructed based on a meta-learning method.

7. The apparatus of claim 6, wherein The feature extraction module is configured to convert the equipment operation data into an input sequence, map the input sequence to an embedding space through linear transformation to obtain embedding features, determine a time dynamic feature of the input sequence, and encode the sequence time dynamic feature; The embedding features and the encoded sequence time dynamic feature are fused to obtain a fusion feature, and the equipment degradation feature is extracted from the fusion feature.

8. The apparatus of claim 7, wherein The feature extraction module comprises a plurality of stacked Transformer encoding layers; The plurality of stacked Transformer encoding layers are configured to extract the equipment degradation feature; Each of the Transformer encoding layers comprises a multi-head self-attention layer and a feedforward network.

9. The apparatus of claim 6, wherein The adaptive stochastic physical loss function module is configured to construct a physical law constraint based on the partial differential equation, construct a stochastic fluctuation constraint based on the stochastic differential equation, construct a physical loss function based on the stochastic fluctuation constraint and the physical law constraint, construct a dynamic weighted composite loss function based on the physical loss function and a preset data loss function, and construct the adaptive stochastic physical loss function based on the dynamic weighted composite loss function. The computer executable instructions, when executed, implement the steps of the adaptive stochastic physical information neural network based remaining useful life prediction method of any one of claims 1-5.

10. A storage medium for storing computer-executable instructions, the computer-executable instructions comprising instructions for: ​

Images