Component aging multi-mode degradation trend prediction method

Abstract

The invention discloses a component aging multi-mode degradation trend prediction method, and belongs to the technical field of prediction and health management. The method comprises the following steps: firstly, constructing a priori causal graph containing a causal prohibition constraint based on physical knowledge; then, a dynamic causal graph is constructed and evolved by using a time sequence graph neural network during operation, and the graph is guided by priori knowledge and reflects the dynamic state of multi-modal data in real time; the core of the method is that when a strong correlation path learned by data driving in a dynamic graph conflicts with a causal prohibition constraint of a prior graph, a causal contradiction arbitration module is activated. The module does not directly deny data, but automatically introduces a new potential variable node. By intelligently arbitrating the contradiction between data and knowledge, the model can adaptively discover and model an unknown degradation mechanism, and the prediction accuracy and the system interpretability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of prediction and health management technology, specifically a method for predicting the multimodal degradation trend of component aging. Background Technology

[0002] In modern industrial sectors such as aerospace, rail transportation, and energy, the reliability and safety of critical systems and core components are paramount. To effectively avoid catastrophic failures, reduce maintenance costs, and improve equipment utilization, predictive health management (PHM) technology has become a focus of research and application. The core objective of this technology is to accurately predict the performance degradation trend and remaining useful life (RUL) of equipment by analyzing various monitoring data generated during operation.

[0003] Currently, the technical approaches to predicting degradation trends can be mainly categorized into two types: physical model-based methods and data-driven methods. Physical model-based methods establish analytical equations describing the degradation process by accurately mathematically modeling the failure mechanisms of components (such as fatigue accumulation, crack propagation, and material corrosion). The advantage of this type of method lies in the clear physical meaning and high interpretability of its conclusions, providing reliable predictions when the model closely matches the actual situation. However, its inherent limitations are also significant. The actual operating conditions of industrial equipment are extremely complex and variable, and its degradation is often the result of the coupling effects of multiple physical and chemical factors. Establishing accurate physical models for all potential failure modes is extremely difficult, if not impossible, in practice. Therefore, physical models often suffer from insufficient prediction accuracy due to oversimplification or inability to cover unknown degradation mechanisms.

[0004] Meanwhile, with the development of artificial intelligence and big data technologies, data-driven methods, especially those represented by deep learning, have demonstrated tremendous potential. These methods do not rely on profound mechanistic understanding; they can autonomously learn complex nonlinear mapping relationships directly from massive amounts of multimodal sensor data (such as temperature, vibration, pressure, and current), constructing predictive models from monitoring data to degradation states in an end-to-end manner. While data-driven models have advantages in handling complex coupling relationships, their limitations cannot be ignored. First, they typically require a large amount of data covering the entire lifecycle of the equipment, including fault labels, for training, and acquiring such high-quality data is costly and time-consuming. More importantly, these models are essentially "black boxes," lacking the guidance and constraints of physical laws in their learning process. Therefore, they are prone to learning spurious correlations or noise in the data, producing predictive results that violate basic physical principles, which is unacceptable in applications with extremely high safety requirements.

[0005] To overcome the limitations of the aforementioned single methods, the industry has begun to explore hybrid approaches that integrate physical knowledge with data-driven models. These methods typically use prior physical knowledge as a static constraint or regularization term to guide the training direction of the data-driven model or to verify the rationality of its output. However, existing fusion techniques face a deeper challenge: when empirical evidence continuously emerges in the data and creates sharp and persistent contradictions with the pre-established physical knowledge model, the strategy is usually to forcibly suppress the correlations revealed by the data, forcing it to "compromise" with the existing model. The fundamental flaw of this approach is that it ignores the valuable information that the contradiction itself may contain—that is, there may be a new or unknown degradation mechanism in the real world that is not covered by prior knowledge. Therefore, existing technologies generally lack an intelligent arbitration mechanism, unable to dynamically evolve and correct their causal understanding by generating new, reasonable physical hypotheses when data and knowledge conflict. This results in the model's inability to adapt to changing system dynamics, limiting its predictive accuracy and ability to detect unknown risks. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method, apparatus, and computer-readable storage medium for predicting multimodal degradation trends of component aging.

[0007] The first aspect of this invention provides a method for predicting the multimodal degradation trend of component aging, the method comprising:

[0008] First, a prior cause-effect graph is constructed based on the physical knowledge of the components, and multimodal time-series data of the components during operation is acquired synchronously. In one embodiment, the step of constructing the prior cause-effect graph may include:

[0009] Define observation variable nodes that correspond one-to-one with various sensor modes, as well as potential variable nodes that cannot be directly observed by sensors, defined based on physical mechanisms or expert experience.

[0010] Furthermore, a set of rules is defined to constrain causal relationships between nodes. This set of rules may include at least causal prohibition constraints, which explicitly define which variable nodes do not have a direct causal relationship according to physical laws or basic common sense.

[0011] Next, based on the acquired multimodal time-series data and guided by the prior causal graph, a dynamic causal graph is constructed and evolved. This dynamic causal graph is used to represent the dynamic correlations of the multimodal time-series data at different time steps in real time and dynamically. In one embodiment, this step can be implemented using a time-series graph neural network model. The adjacency matrix of this model, i.e., the matrix representing the graph's connection structure, is jointly determined by a weighted fusion of the fixed structure of the prior causal graph and a data-driven graph structure learned from real-time multimodal time-series data. This ensures that the evolution of the graph both follows a basic physical framework and flexibly reflects the real-time dynamics of the data.

[0012] Subsequently, one of the core innovations of this invention lies in monitoring the consistency between the data-driven discovery of associations and prior knowledge in the above process, that is, monitoring causal contradictions between the dynamic causal graph and the prior causal graph. When a causal contradiction is detected, an arbitration mechanism is initiated. Specifically, the step of monitoring causal contradictions can be implemented in the following way: for any node pair (v) prohibited by causal prohibition constraints in the prior causal graph... u ,v v The conflict score D at time step t can be calculated using the following formula. uv,t :

[0013] D uv,t =ReLU(a uv,t -θ conflict );

[0014] Among them, a uv,t For the dynamic causal graph from node v u to v v The connection weights, θ conflict A pre-set conflict determination threshold is used. When the calculated conflict score D... uv,t When the value is greater than 0, it can be determined that the system has detected a causal contradiction that requires arbitration.

[0015] When a causal contradiction is detected, it is arbitrated by introducing latent variables and reconstructing causal paths in the dynamic causal graph. In one embodiment, this arbitration step may specifically include: dynamically introducing a new latent variable node into the dynamic causal graph; then, replacing the original data-driven direct causal path that conflicts with the causal prohibition constraints of the prior causal graph with an indirect causal path through the new latent variable node. This approach does not simply negate data or knowledge, but intelligently resolves the conflict by proposing a deeper and more reasonable causal hypothesis, thereby enabling the model to discover unknown, potential systemic influencing factors.

[0016] Finally, based on the arbitrated dynamic causal graph, system degradation features characterizing the overall health of the components are aggregated, and the aging and degradation trends of the components are predicted accordingly. In one embodiment, this step can use a readout function to aggregate the hidden states of all nodes in the dynamic causal graph to generate a graph-level embedding vector as the system degradation feature. Then, the sequence of graph-level embedding vectors at consecutive time steps is input into a sequence prediction model (such as Transformer or LSTM) to generate predictions of future aging and degradation trends.

[0017] In some embodiments, the method of the present invention can optimize a composite loss function L. total Training is then performed. The composite loss function can be defined by the following formula:

[0018] L total =L pred +αL causal ;

[0019] Among them, L pred L is the prediction loss used to measure the difference between the predicted result and the true label value. causal The causal loss is used to penalize behaviors that violate the causal prohibition constraints in the prior causal graph. For example, this loss term can be calculated by summing or averaging the connection weights on all prohibited causal paths. α is a hyperparameter used to balance the prediction loss and the causal loss. In some embodiments, the composite loss function may further include a physical consistency loss L. phys This loss term is used to penalize cases where the model's predictions do not conform to the physical conservation laws (such as energy conservation) defined in the prior causal graph, thereby further improving the physical accuracy of the predictions.

[0020] A second aspect of the present invention provides a device for predicting the multimodal degradation trend of component aging, the device comprising:

[0021] A building block for constructing a prior cause-effect graph based on physical knowledge and acquiring multimodal time-series data during component operation;

[0022] An evolution module is used to construct and evolve a dynamic causal graph based on the multimodal time series data and guided by the prior causal graph. The dynamic causal graph is used to characterize the dynamic correlation of the multimodal time series data at different time steps.

[0023] An arbitration module is used to monitor causal contradictions between the dynamic causal graph and the prior causal graph; when a causal contradiction is detected, the causal contradiction is arbitrated by introducing latent variables into the dynamic causal graph and reconstructing the causal path.

[0024] A prediction module is used to aggregate system degradation features based on the arbitrated dynamic causal graph and predict the aging and degradation trends of components.

[0025] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect of the present invention.

[0026] This invention provides a method for predicting the multimodal degradation trend of component aging. It has the following beneficial effects:

[0027] 1. This invention does not simply merge data with static physical knowledge. Instead, through a core causal conflict arbitration module, when data and knowledge conflict, it does not forcibly suppress the data. Instead, it introduces latent variables and reconstructs the graph structure to find a more reasonable physical explanation for the conflict. This mechanism enables the model to dynamically adapt to the actual changes in the system, more profoundly approximating the real physical degradation process, thereby significantly improving the accuracy of prediction results and robustness to changes in operating conditions.

[0028] 2. Unlike traditional "black box" data-driven models, the entire prediction process of this invention is based on an explicit dynamic causal graph. In particular, when the arbitration module introduces new latent variables, this behavior can be understood as the system autonomously proposing a new, expert-verifiable physical hypothesis (e.g., discovering an unmonitored common cause). This makes the model's decision-making process no longer elusive but has a clear, traceable physical logic, enhancing user trust in the prediction results.

[0029] 3. The greatest advantage of this invention lies in its ability to learn from contradictions. When continuous observational data contradicts inherent prior causal prohibitions, the system treats it as an opportunity to discover new knowledge. By initiating arbitration and reconstructing the graph, the model can autonomously identify and model emerging or coupled degradation mechanisms that were not considered in the initial design phase. This adaptive capability is crucial for handling devices operating under complex and non-stationary conditions.

[0030] 4. The existence of a priori causal graphs provides strong guidance and constraints for the data-driven learning process, preventing the model from learning spurious associations without physical meaning. Furthermore, the causal conflict arbitration does not react to every data fluctuation, but is only triggered when the strength of data-driven associations consistently exceeds a preset threshold. This effectively filters out random noise interference, ensuring that the dynamic relationships learned by the model have a high degree of confidence.

[0031] 5. Traditional knowledge fusion methods heavily rely on a complete and accurate set of prior knowledge. If the knowledge base contains defects, the model's performance will be significantly reduced. The arbitration and graph reconstruction mechanism of this invention allows the system to start from an incomplete or even partially erroneous prior causal graph. During operation, the system can use data to "reverse-engineer" and supplement the original knowledge graph. This greatly reduces the stringent requirements on the initial knowledge model, making this method more versatile and practical in engineering practice. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the system functional modules according to an embodiment of the present invention;

[0033] Figure 2 This is a flowchart of a method according to an embodiment of the present invention;

[0034] Figure 3 A block diagram of the prediction device according to an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0036] refer to Figures 1 to 3 This invention provides a method for predicting the multimodal degradation trend of component aging, which can be applied to industrial equipment, especially for health status monitoring and remaining service life prediction of critical rotating parts or electronic components. This method is implemented through a system integrating data processing, model evolution, intelligent arbitration, and trend prediction.

[0037] The prediction system may include: a physical knowledge and data processing module, a dynamic graph evolution module, a causal contradiction arbitration module, and a degradation trend prediction module. These modules work together to achieve accurate modeling and prediction of the component aging process.

[0038] The physics knowledge and data processing module is responsible for providing structured prior knowledge and standardized input data for the entire prediction process. The output of this module directly serves the dynamic graph evolution module, providing it with initial topological guidance and real-time feature input.

[0039] The dynamic graph evolution module is the core engine of the entire system. It receives data and knowledge from other modules and constructs a graphical model that reflects the dynamic relationships between various physical quantities within the component. At each time step, this module performs state updates and structural evolution, and then passes the evolution results to the causal contradiction arbitration module and the degradation trend prediction module.

[0040] The causal conflict arbitration module, a key innovation of this invention, functions to monitor in real time the data-driven relationships learned by the dynamic graph evolution module and compare them with the prior knowledge provided by the physical knowledge and data processing module. When a conflict is detected, the module proactively intervenes in the graph structure of the dynamic graph evolution module, resolving the conflict by introducing new hypotheses. This is a closed-loop feedback control process that ensures the system's learning achieves a balance between following physical laws and respecting objective data.

[0041] The degradation trend prediction module receives and aggregates the system state information from the dynamic graph evolution module after arbitration and evolution, and uses this information to predict the future degradation trend of components, and finally outputs the prediction results for operation and maintenance personnel to refer to.

[0042] The method described in this invention can be implemented on one or more computing devices. The computing device can be a general-purpose computer, an industrial control computer, a server, or a dedicated embedded system. The device includes at least a processor, a memory, a communication interface, and a human-machine interface. The memory stores computer program instructions, and when the processor executes these instructions, it implements all or part of the steps of the method described in this invention. This implementation environment is well known to those skilled in the art and will not be described in detail here. The operation of the entire system coordinates the above-mentioned functional modules to form a complete, closed-loop workflow from data input to predicted output, achieving deep insight and forward-looking prediction of the component aging process.

[0043] In a specific embodiment of the present invention, the specific implementation of the physical knowledge and data processing module will be described in detail. The function of this module is to provide high-quality data input and structured knowledge guidance for subsequent modeling and prediction; the completeness and accuracy of its processing are the foundation of the effectiveness of the entire method.

[0044] First, this module acquires multimodal time-series data of the component's operation from an external data acquisition system. This data can originate from various sensors installed on or around the component, such as accelerometers for monitoring mechanical vibrations, temperature sensors for measuring thermal states, current or voltage sensors for reflecting electrical characteristics, and acoustic sensors for detecting abnormal noises. Assuming there are M different sensor modes, the acquired raw dataset can be represented as S = {S1, S2, ..., S...} M}, where each item Sm Both are time series that record the changes of the m-th mode over time.

[0045] After obtaining the raw data, a series of preprocessing operations are required to improve data quality. In one embodiment, a Gaussian filter or median filter can be used to smooth the signal and remove high-frequency noise interference. For any missing values ​​that may exist in the data, linear interpolation or cubic spline interpolation can be used to fill them in, ensuring the continuity of the data sequence. To eliminate the influence of different physical units and dimensions between different modal data, the data needs to be normalized. A specific normalization method is Z-Score standardization, the calculation formula of which is:

[0046]

[0047] Where, x m,t These are the raw data points of mode m at time t, x′ m,t These are the standardized data points, μ m and σ m These are the mean and standard deviation of mode m over a reference time period (e.g., normal operation phase), respectively. After preprocessing, a standardized dataset X = {X1, X2, ..., X...} is obtained, which can be directly used as model input. M}

[0048] Meanwhile, one of the core tasks of this module is to construct the prior causal graph G. p =(V p E p ,R p This process involves formalizing and structuring the accepted physical laws, engineering common sense, and expert diagnostic experience within the field.

[0049] The node set V of the graph p Defined as the union of two classes of variables: V p =V o ∪V l Among them, V o It is a set of observed variable nodes, each node corresponding one-to-one with one of the M preprocessed sensor modes. For example, the node "v_temp" corresponds to temperature sensor data. V l It is a set of latent variable nodes, which represent internal states or physical phenomena that cannot be directly measured by sensors but play a key role in component degradation. For example, for a rolling bearing, v_crack_density (crack density) or v_lubrication_degradation (lubrication degradation) can be defined as latent variable nodes. The introduction of these latent variables allows the model to reason at a level closer to the physical essence.

[0050] The edge set E of the graph p It is a set of directed edges used to represent prior causal relationships between variables. An edge originating from node v... i Pointing to v j The edge e ij This means that, according to physical mechanisms, v i The change is what causes v j One direct cause of the change. For example, an edge from "v_lubrication_degradation" to "v_temp" can be defined to represent the prior knowledge that lubrication failure will lead to an increase in temperature.

[0051] More importantly, the rule set R p The construction of this rule set imposes strong constraints on the learning of subsequent models. In one embodiment, this rule set contains at least the causal prohibition constraint R. forbid The constraint is a set of node pairs, where each node pair (v u ,v v Both clearly state that, based on first principles of physics or established engineering knowledge, the variable v u It is impossible to directly cause variable v v Changes in the frequency of an electrical system. For example, in most cases, current fluctuations in an electrical system do not directly cause changes in the vibration frequency of an independent mechanical structure, although they may be related through a common cause (such as load changes). Defining this relationship as a causal prohibition constraint can effectively prevent the model from constructing pseudo-causal or physically unexplained connections in subsequent learning simply because of the apparent correlation of the data. Furthermore, the rule set R... p It can also include other types of constraints, such as type constraints that define the direction of causal influence (positive or negative), or physical conservation constraints that describe the macroscopic conservation laws of the system (such as energy conservation).

[0052] In one specific embodiment of the present invention, the dynamic graph evolution module receives the output of the physics knowledge and data processing module, and its core task is to construct and evolve a dynamic causal graph. This graph aims to capture the complex and time-varying influence relationships between various physical quantities within a component in real time, serving as a bridge connecting static physics knowledge and dynamic observation data.

[0053] This module employs a specially designed Temporal Graph Neural Network (T-GNN) as the core of its technical implementation. At each time step t, the network maintains and updates a dynamic graph G. t =(V t A t ), where V t Let V be the set of nodes at this moment, and its initial state be the same as the node set V of the prior causal graph. p Same; A tLet a be the dynamic adjacency matrix at this moment, and let its element a be the adjacency matrix. ij,t This represents v at time node t. i For node v j The instantaneous impact intensity.

[0054] First, in order to integrate time-series data into the computational framework of graph neural networks, it is necessary to perform calculations for each observation node v. m ∈V o Extract its feature embedding at time step t. In one embodiment, this is achieved using an embedding function φ implemented by a one-dimensional convolutional neural network (1D-CNN). emb This is completed. The network operates on the corresponding modal data X. m Within a fixed-length time window W preceding time t, i.e., X m,[t-W+1:t] This is to capture the recent local patterns and dynamic features of the mode. The output z of the embedding function is... m,t =φ emb (X m,[t-W+1:t] ) is a fixed-dimensional vector, representing a node v. m The input features at time step t. For a latent variable node, its input features can be initialized as a zero vector or a learnable embedding vector.

[0055] Dynamic adjacency matrix A t The generation of the adjacency matrix A′ is crucial to this module, cleverly combining the stability of prior knowledge with the flexibility of data-driven approaches. Specifically, its generation involves two steps. The first step involves learning a fully data-driven adjacency matrix A′ from the data using a self-attention mechanism. data,t At time step t-1, all nodes hide the state {h}. k,t-1 Based on}, node v i For node v j Attention score ij,t The calculation is as follows:

[0056]

[0057] Among them, W Q and W K It is a learnable linear transformation matrix used to map the hidden state of a node to the query space and the key space; d k This is the dimension of the key vector, used to scale the dot product result. Subsequently, the score is normalized using the Softmax function to obtain the attention weights a′. ij,t These weights constitute the data-driven adjacency matrix A′ data,t Element.

[0058] The second step involves fusing data-driven discovery with prior knowledge to generate the dynamic adjacency matrix A used at time step t.t The calculation formula is as follows:

[0059] A t =σ(λA) p +(1-λ)A′ data,t )

[0060] In this formula, A p It is based on the prior cause-effect graph G p edge set E p The resulting binary adjacency matrix (1 if an edge exists, 0 otherwise). λ is a hyperparameter between 0 and 1, acting as "knowledge confidence," controlling the strength of prior knowledge's guidance on graph structure evolution. σ is the sigmoid activation function, ensuring that the final adjacency matrix A... t All elements a ij,t All of them fall within the (0,1) interval, which can be interpreted as the probability or strength of the connection.

[0061] The dynamic adjacency matrix A was obtained. t After considering the input features of each node, the temporal graph neural network evolves the node state through message passing and node update mechanisms. At time step t, each node v... j First, start from its neighbor set Message aggregation. A specific aggregation method is a weighted sum of the hidden states of neighboring nodes at the previous time step:

[0062]

[0063] Where a ij,t That is, the dynamic adjacency matrix A t The weights in the data are then determined. Subsequently, a gated recurrent unit (GRU) is used as the update function to update the aggregated message a. j,t The node's own input features z j,t And the hidden state h from the previous moment j,t-1 Combined, calculate the new hidden state h at the current time. j,t :

[0064] h j,t =GRU(h j,t-1 ,[z j,t ,a j,t ])

[0065] Here, [·,·] denotes the vector concatenation operation. By repeating this process at each time step, the dynamic graph evolution module can continuously track and model the evolving relationships between variables within the component, and its output is the node hidden state sequence {h j,t It contains a wealth of system state information that integrates physical knowledge and dynamic data.

[0066] In one specific embodiment of the present invention, the causal contradiction arbitration module constitutes the core innovation of this technical solution. The function of this module is not simply to choose between data and knowledge, but rather to act as an intelligent arbitrator and hypothesis generator when discrepancies arise, thereby deepening the system's understanding of component degradation mechanisms. Its specific implementation process includes two closely linked stages: contradiction detection and graph reconstruction.

[0067] This module operates in an online monitoring manner. It continuously receives the dynamic adjacency matrix A generated by the dynamic graph evolution module at each time step t. t And it is combined with the prior causal graph G constructed by the physics knowledge and data processing module. p The rule set R in p A comparison is performed. The focus of contradiction detection is on the causal prohibition constraint R. forbid Regarding the violation. Specifically, for rule set R forbid Any prohibited causal path existing in the process, i.e., node pair (v u ,v v This module checks the corresponding element 'a' in the dynamic adjacency matrix in real time. uv,t The value. When this value continuously or significantly exceeds a preset conflict determination threshold θ conflict At this point, the system considers the data to contain a strongly correlated signal that contradicts prior knowledge. To quantify the severity of this conflict and trigger subsequent arbitration mechanisms, the system calculates a conflict score D. uv,t The formula is as follows:

[0068] D uv,t =ReLU(a uv,t -θ conflict )

[0069] In this formula, a uv,t At time step t, from node v u to v v Data-driven learned connection weights; θ conflict It is a positive scalar value set empirically or by system calibration to distinguish between genuine conflicts and random fluctuations; ReLU(X) is the modified linear unit function, defined as max(0,x), which ensures that the conflict score is positive only when the connection weights actually exceed the threshold, otherwise it is zero. A conflict score D greater than zero uv,t This is a clear signal that the arbitration mechanism is being restructured in the startup diagram.

[0070] Once a significant causal contradiction, i.e. D, is detected... uv,tIf the value is greater than 0, the causal contradiction arbitration module will immediately perform a graph reconstruction operation. The core idea of ​​this operation is: not to deny the strong correlation revealed by the data, but to find a more logical and indirect explanation path. This process is executed precisely in steps.

[0071] The first step is to... in the current set of nodes V t Introduce and instantiate a new latent variable node v new This new node physically represents a previously unconsidered element that may lead to v. u and v v Intermediate physical phenomena or states that generate strong correlations.

[0072] The second step is to reconstruct the graph's topology. The system will refactor the original direct path v, which was supported by data but prohibited by knowledge. u →v v Make corrections. Specifically, this is done in the dynamic adjacency matrix A. t In the middle, element a uv,t The value is forced to zero or a very small number, thereby "severing" this direct connection that does not conform to prior knowledge.

[0073] The third step is to establish new indirect causal paths through latent variables. The system will then add two new edges, e, in the adjacency matrix. u,new and e new,v This assigns initial weights. This is equivalent to constructing a new two-step path v in the graph. u →v new →v v To make this hypothesis generation process more reasonable, the initial weights of the new edges can be related to the detected conflict score D. uv,t Related. For example, the new weight a u,new,t and a new,v,t It can be set to be compatible with D uv,t Monotonic correlation values, such as those mapped through an activation function σ, i.e., a u ,new,t=σ(w·D uv,t ), where w is an adjustable parameter.

[0074] Through the aforementioned detection and reconstruction loop, the method of this invention achieves a unique adaptive learning capability. It enables the model to learn from conflicts between data and knowledge, continuously refining its internal causal graph by proposing and validating new physical hypotheses, thereby maximizing the utilization of information contained in the data while ensuring physical consistency.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the multimodal degradation trend of component aging, characterized in that, Includes the following steps: A prior cause-effect graph is constructed based on physical knowledge, and multimodal time-series data during component operation is obtained; Based on multimodal time series data and guided by a prior causal graph, a dynamic causal graph is constructed and evolved. The dynamic causal graph is used to characterize the dynamic correlation of multimodal time series data at different time steps. Monitor causal contradictions between the dynamic causal graph and the prior causal graph; when a causal contradiction is detected, arbitrate the contradiction by introducing latent variables into the dynamic causal graph and reconstructing the causal path; Based on the arbitrated dynamic causal graph, the system degradation characteristics are aggregated and generated, and the aging and degradation trends of components are predicted.

2. The method according to claim 1, characterized in that, The steps for constructing the prior cause-effect graph include: Define the observed variable nodes corresponding to the sensor modes and the latent variable nodes defined by the physical mechanism; and, Define a set of rules to constrain the causal relationships between the nodes, the set of rules including at least a causal prohibition constraint to define that there is no direct causal relationship between the nodes.

3. The method according to claim 1, characterized in that, The steps for constructing and evolving the dynamic causal graph include: A time-series graph neural network model is adopted, and the hidden state of nodes in the graph is updated through a message passing mechanism; The adjacency matrix of the time-series graph neural network model is jointly determined by the structure of the prior causal graph and the data-driven graph structure learned based on the multimodal time-series data.

4. The method according to claim 1, characterized in that, The specific steps for monitoring causal contradictions are as follows: for the node pairs (v) prohibited by the causal prohibition constraint in the prior causal graph... u ,v v ), calculate its conflict score D according to the following formula. uv,t : D uv,t JReLU(a uv,t -θ conflict )4 Among them, a uv,t For node v in the dynamic causal graph u to v v The connection weights, θ conflict The conflict determination threshold is set when the conflict score D is reached. uv,t When the value is greater than 0, a causal contradiction is determined.

5. The method according to claim 1, characterized in that, The steps for arbitrating causal contradictions include: introducing a new latent variable node into the dynamic causal graph; and replacing a direct causal path generated by data but conflicting with the prior causal graph with an indirect causal path that passes through the new latent variable node.

6. The method according to claim 1, characterized in that, The method optimizes a composite loss function L. total The composite loss function is defined as follows for training: L total =L pred +αL causal ; Among them, L pred L is the prediction loss used to measure the difference between the predicted result and the actual value. causal α is a hyperparameter that balances the loss term, used to penalize causal losses for violating the causal prohibition constraints of the prior causal graph.

7. The method according to claim 6, characterized in that, The causal loss L causal The calculation method is as follows: sum or average the connection weights of all node pairs prohibited by the causal prohibition constraint in the prior causal graph.

8. The method according to claim 6, characterized in that, The composite loss function L total It also further includes a physical consistency loss L phys The physical consistency loss is used to penalize predictions that do not conform to the physical conservation constraints defined in the prior causal graph.

9. The method according to claim 1, characterized in that, The step of generating system degradation features by aggregation includes: using a readout function to aggregate the hidden states of all nodes in the dynamic causal graph to generate graph-level embedding vectors as the system degradation features.

10. The method according to claim 9, characterized in that, The step of predicting the aging and degradation trend of components includes: inputting the sequence of graph-level embedding vectors of continuous time steps into a sequence prediction model to generate a prediction result of future aging and degradation trends.

Images