Method for diagnosing faults and predicting service life of robot joint module phase perception

Abstract

The application relates to a robot joint module stage perception fault diagnosis and life prediction method, which comprises the following steps: converting original vibration signals of a robot joint module in a whole life cycle into semantic health representations describing health state evolution; constructing a degradation reasoning model based on a large language model by taking the semantic health representations as input and taking structured prior knowledge containing degradation stages, degradation trends and physical boundary constraints of the robot joint module as output; constructing a predictor based on a deep neural network by taking the structured prior knowledge as input and taking degradation stage recognition and residual life prediction as output; in the training process, based on the structured prior knowledge, dynamically adjusting training loss weights of different degradation stages and constructing a physical heuristic loss, fusing training loss and physical information constraint loss into a total loss, and jointly training and optimizing model parameters. The application enhances prediction accuracy, physical consistency and robustness under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of machine learning-based mechanical key component condition monitoring and health management technology, and in particular to a method for fault diagnosis and life prediction of robot joint module stage perception. Background Technology

[0002] As a core component of industrial machinery, the operational status of robot joint modules directly determines the reliability and safety of the entire equipment system. The high-precision transmission and dynamic response capabilities of robot joint modules are fundamental to compliant robot assembly. Unplanned downtime failures in these components can not only lead to substantial economic losses but also potentially trigger serious production safety accidents. Therefore, conducting research on the remaining life prediction and fault severity diagnosis of robot joint modules is a core task in establishing a preventative maintenance and health management system for equipment.

[0003] With the development of industrial big data technology, existing methods for predicting the remaining lifespan and diagnosing the fault level of robot joint modules mainly focus on the application of deep learning models (such as CNN and Transformer). These methods aim to extract degradation features from massive vibration signals by constructing complex nonlinear mapping structures. However, existing data-driven methods still face the following bottlenecks in practical applications:

[0004] First, there is a lack of stage-specific adaptability. The degradation process of robot joint modules typically exhibits distinct stage-specific heterogeneity, such as slow degradation in the early stages of operation and exponentially accelerated degradation in the later stages of failure. Existing deep learning models mostly use static training objective functions, making it difficult to dynamically perceive the different degradation stages during model training, leading to decreased prediction accuracy during critical rapid degradation periods.

[0005] Second, physical consistency is difficult to guarantee. Purely data-driven neural network models are considered "black box" models, and their predictions often rely solely on signal pattern matching, lacking the constraints of physical laws. In the absence of explicit degradation physical equations, the lifetime predictions output by the model may violate basic physical evolution logic, reducing the reliability of the prediction results.

[0006] Third, high-level knowledge integration is difficult. Although domain experts possess profound prior knowledge and reasoning logic regarding robot joint module degradation, there is still a lack of efficient technical means to effectively integrate this high-level, structured degradation reasoning knowledge into the underlying signal processing and end-to-end neural predictors. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for fault diagnosis and life prediction based on stage perception of robot joint modules, thereby solving the problems of limited prediction accuracy, physical consistency, and robustness of existing technologies under complex working conditions.

[0008] The technical solution adopted in this invention is as follows:

[0009] This invention provides a method for fault diagnosis and lifespan prediction based on stage perception of a robot joint module, comprising:

[0010] The original vibration signals of the robot joint module throughout its entire life cycle are collected and transformed into a semantic health representation describing the evolution of its health status after signal processing.

[0011] A degradation inference model based on a large language model is constructed, which takes the semantic health representation as input and generates structured prior knowledge containing the degradation stage, degradation trend and physical boundary constraints of the robot joint module.

[0012] A predictor based on a deep neural network is constructed, which takes the structured prior knowledge as input and generates degradation stage identification and remaining life prediction results.

[0013] The predictor structure includes a GRU encoder, a degradation stage classification head, and a RUL regression head; the GRU encoder extracts the temporal degradation features of the vibration signal to obtain a degradation embedding; the degradation stage classification head maps the degradation embedding to a degradation stage category; the RUL regression head predicts the remaining lifetime based on the degradation embedding;

[0014] The predictor is equipped with a stage-aware adaptive learning module, which uses the degradation stage in the structured prior knowledge to dynamically adjust the weights of the training loss function of the predictor in different degradation stages, so as to balance the stage classification loss and RUL regression loss in the training loss; the predictor is equipped with a physical heuristic constraint regularization module, which constructs a physical heuristic loss based on the structured prior knowledge.

[0015] The training loss and the physical information constraint loss are weighted and fused to form the total loss. The model parameters are then jointly trained and optimized using the backpropagation algorithm to form the predictor.

[0016] The physical heuristic loss consists of three constraints related to degradation characteristics, including:

[0017]

[0018] In the formula: For time points t Physical heuristic loss; The strength of each constraint; monotonicity loss constraint , and They are time points respectively t and t -1 is the predicted remaining lifetime value; degradation rate constraint , , Upper bound of degradation rate; degradation acceleration constraint , , For time points t -2 is the predicted remaining lifetime value. Indicates the stage of degeneration. The expected rate of degradation.

[0019] The preferred technical solution is:

[0020] The calculation of the training loss function weights includes:

[0021]

[0022] In the formula, For time points t The training loss function weights are the i-th Degradation probability at each degradation stage Expectations Output by the degradation stage classification head; These are the preset stage base weights; K Total number of degradation stages.

[0023] The degradation stage The expected degradation acceleration level is defined as follows:

[0024]

[0025] Wherein, early, middle, and late represent the early, middle, and late degradation stages, respectively, i.e., the total number of degradation stages. K= 3, k =1,2,3 correspond to early, middle, and late, respectively.

[0026] when K= 3. The preset stage base weights are:

[0027]

[0028] in, This is to characterize the fact that as the degree of degradation increases, the remaining life regression analysis receives a higher weight.

[0029] The construction of the semantic health representation includes:

[0030] Time-frequency analysis was performed on the acquired raw vibration signal segments to extract a set of degradation-sensitive characteristics from the time-frequency representation;

[0031] In the original vibration signal segment t Multiple degradation-sensitive features are extracted at time to form a multi-dimensional time-frequency feature vector, which is then normalized.

[0032] Principal component analysis was used to fuse the multidimensional time-frequency feature vectors into a single health indicator. ;right Performing a first-order difference, we obtain , , They represent time respectively , Health indicator values ​​at the location, Represent the rate of change of health indicators; construct a series of health indicators. , ;

[0033] A set of statistical features is extracted from the vibration signal segment to provide supplementary degradation information, time. Statistical eigenvectors at [location] RMS, Kurtosis, Skewness, Crest, and Entropy represent root mean square, kurtosis, skewness, peak factor, and signal entropy, respectively.

[0034] Determine the running condition vector , , , They represent time respectively Speed, load level, and temperature at the location;

[0035] A semantic health representation is constructed based on the health indicator sequence, statistical feature vector, and operating condition vector. .

[0036] The degradation-sensitive characteristics include time-frequency energy, spectral centroid, spectral bandwidth, and high-frequency energy ratio.

[0037] The construction of the degradation reasoning model includes:

[0038] Construct a prompt template, convert the semantic health representation into a natural language description, input it into a pre-trained large language model, and output the parsed structured prior knowledge. , The terms indicate the inferred degradation stage, with early, middle, and late representing the early, middle, and late degradation stages, respectively. Indicating a degradation trend, stable, moderate, and accelerating represent stable degradation trends, moderate degradation trends, and accelerating degradation trends, respectively. This represents the inferred constraints related to the degradation characteristics, which are embedded as soft regularization terms into the learning objective of the loss function.

[0039] The upper bound of the degradation rate Adaptive determination based on degradation trends inferred from large language models:

[0040] .

[0041] The GRU encoder includes an input projection layer, a temporal coding layer, and a global degenerate embedding layer. The input projection layer is the input of the GRU encoder. The temporal coding layer includes multiple GRU units. The global degenerate embedding layer takes the hidden state of the last GRU unit at the last time step as the degenerate embedding vector.

[0042] The RUL regression head performs regression based on the degenerate embedding vector and outputs a predicted remaining lifetime value:

[0043]

[0044] In the formula, the remaining lifetime prediction output value during training is... It is constrained to be non-negative to ensure physical rationality; Indicates the activation function of the hidden layer; , This represents the weights and biases of the RUL regression head output layer; Represents a degenerate embedding vector. , This represents the weights and biases of the hidden layer in the RUL regression head;

[0045] The degradation stage classification head includes a fully connected layer with a ReLU activation function and a softmax output layer.

[0046] The technical solution of the present invention can achieve at least some of the following beneficial effects:

[0047] This invention leverages the capabilities of large language models in logical reasoning and cross-modal knowledge association. Utilizing their high-level knowledge reasoning abilities, it designs an SHR-driven degradation reasoning approach to guide the knowledge cascading behavior from raw vibration signals to high-level structured prior knowledge, and performs multi-source information fusion during model training. A stage-aware adaptive learning strategy and physically heuristic constraint regularization are employed to generate physically consistent deep embeddings. This suppresses prediction biases and non-physical fluctuations in the heterogeneous degradation stage of robot joint modules under purely data-driven models, enabling the model to further learn implicit degradation physical characteristics and achieve accurate predictions of the remaining lifespan and fault characteristics of robot joint modules.

[0048] Through numerical examples, this invention combines degradation inference of a large language model, stage-aware adaptive dynamic adjustment of the predictor's training loss function weights at different stages, and physical constraint regularization to robustly model complex degradation patterns, thereby enhancing the model's prediction accuracy, physical consistency, and robustness under complex conditions. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention.

[0050] Figure 2 A schematic diagram of the predictor structure with added physical information constraint loss for an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram illustrating the prediction results of the remaining lifespan of a robot joint module using the method of an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram illustrating the classification results (fault severity classification) of the degradation stage of a robot joint module according to the method of this embodiment of the invention. Detailed Implementation

[0053] The specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0054] See Figure 1 This embodiment of a method for fault diagnosis and lifespan prediction of a robot joint module stage perception includes the following steps:

[0055] S1. Raw vibration signals from the robot joint module throughout its entire lifecycle are collected using a PCB piezoelectric accelerometer. After signal processing, these signals are converted into a semantic health representation (SHR) describing the evolution of the health state. A preferred method includes:

[0056] S11. Preferably, short-time Fourier transform (STFT) is used to perform time-frequency analysis on the acquired original vibration signal segment to obtain a local spectral representation, and a set of degradation-sensitive characteristics are extracted from the time-frequency representation, including but not limited to the characteristics listed in Table 1.

[0057] Table 1. Characteristics including degradation susceptibility.

[0058]

[0059] S12. In the original vibration signal segment t Multiple degradation-sensitive features are extracted at time to form a multi-dimensional time-frequency feature vector, which is then normalized.

[0060]

[0061] In the formula, the multidimensional time-frequency feature vector , Indicates the first Each feature vector in time The original value, , Indicates the first The minimum and maximum values ​​of each feature vector in the entire time series. This represents the normalized eigenvalues, ranging from [0,1].

[0062] S13. Principal component analysis is used to fuse the multidimensional time-frequency feature vectors into a single health indicator. :

[0063]

[0064] In the formula, Covariance matrix eigenvector matrix; Based on the results obtained from principal component analysis, principal components are the preferred choice. As ;

[0065] S14. To Performing a first-order difference, we obtain , , They represent time respectively , Health indicator values ​​at the location, Representing the rate of change of health indicators, reflecting the speed of deterioration; constructing a health indicator series. , ;

[0066] S15. Extract a set of statistical features from the vibration signal segment to provide supplementary degradation information, time. Statistical eigenvectors at [location] RMS, Kurtosis, Skewness, Crest, and Entropy represent root mean square, kurtosis, skewness, peak factor, and signal entropy, respectively.

[0067] S16. To clearly represent operating conditions and distinguish between changes caused by degradation and fluctuations caused by operating conditions, determine the operating condition vector. , , , They represent time respectively Speed, load level, and temperature at the location;

[0068] S17. Construct a semantic health representation based on the health indicator sequence, statistical feature vector, and operating condition vector. This representation captures the degeneracy process in a compact and semantically clear form, making it suitable for reasoning based on large language models.

[0069] S2. Construct a degradation inference model based on a Large Language Model (LLM), which takes the semantic health representation as input and generates structured prior knowledge containing the degradation stages, degradation trends, and physical boundary constraints of the robot joint module. Preferably, the construction of the degradation inference model includes:

[0070] Construct a prompt template, convert the semantic health representation into a natural language description, input it into a pre-trained large language model, and output the parsed structured prior knowledge. , The terms "early," "middle," and "late" represent the inferred degradation stages, respectively. Indicating a degradation trend, stable, moderate, and accelerating represent stable degradation trends, moderate degradation trends, and accelerating degradation trends, respectively. This indicates the inferred constraints related to the degradation characteristics, which will be embedded as a soft regularization term into the learning objective of the loss function. The preferred structure of the prompt template is shown in Table 2.

[0071] Table 2 Prompt Templates

[0072]

[0073] S3. Construct a predictor based on a deep neural network, which takes the structured prior knowledge as input and generates degradation stage identification and remaining life prediction results.

[0074] The predictor structure includes a GRU (Gated Recurrent Unit) encoder, a degradation stage classification head, and a RUL regression head. The GRU encoder extracts the temporal degradation features of the vibration signal to obtain a degradation embedding. The degradation stage classification head maps the degradation embedding to a degradation stage category. The RUL regression head predicts the remaining lifetime based on the degradation embedding.

[0075] The predictor is equipped with a stage-aware adaptive learning module, which uses the degradation stage of the structured prior knowledge to dynamically adjust the weights of the training loss function of the predictor at different degradation stages, so as to balance the stage classification loss and RUL regression loss in the training loss and achieve targeted learning for degradation heterogeneity.

[0076] Wherein, the training loss , Indicates the stage classification loss, Representing the RUL regression loss; preferably, the calculation of the weights of the training loss function includes:

[0077]

[0078] In the formula, For time points t The training loss function weights are the i-th Degradation probability at each degradation stage Expectations Output by the degradation stage classification head; K Total number of degradation stages in this embodiment K= 3, k =1,2,3 correspond to early, middle, and late, respectively; Preset stage base weights:

[0079]

[0080] in, This is to characterize the fact that as the degree of degradation increases, the remaining life regression analysis receives a higher weight.

[0081] The predictor includes a physical heuristic constraint regularization module, which constructs a physical heuristic loss based on the degradation rules contained in the structured prior knowledge. It consists of three constraints related to degradation characteristics, including:

[0082]

[0083] In the formula: The physical heuristic loss at time point t; The strength of each constraint;

[0084] As a monotonic loss constraint, it is calculated over all adjacent time steps within each input sequence and averaged over the batch: ,in and These are the predicted remaining lifetime values ​​at time points t and t-1, respectively.

[0085] To constrain the degradation rate, a first-order time difference approximation is used for calculation: , , The upper bound of the degradation rate can be adaptively determined based on the degradation trend inferred from a large language model:

[0086]

[0087] To approximate the degenerate acceleration constraint, a second-order finite difference method is used: , , For time points t -2 is the predicted remaining lifetime value. Indicates the stage of degeneration. The expected rate of degradation acceleration is preferably set as follows:

[0088]

[0089] The training loss and the physical information constraint loss are weighted and fused to form the total loss. The model parameters are jointly trained and optimized using the backpropagation algorithm to form the predictor.

[0090] The monotonicity loss constraint in this embodiment considers the monotonic trend of the theoretically decreasing lifespan of machine parts over time, while the degradation rate constraint and degradation acceleration constraint consider the rate of degradation, making the trend more realistic. Therefore, the physical heuristic loss constructed by combining these three constraints can make the predictor's prediction results more consistent with physical meaning.

[0091] The predictor structure in this embodiment, which incorporates physical information constraint loss, is described in [reference needed]. Figure 2 As shown. The predictor structure includes an input layer, a hidden layer, and an output layer. The predictor outputs... That is, the predicted degradation stage category and the remaining lifetime value.

[0092] Specifically, the GRU encoder includes an input projection layer, a temporal coding layer, and a global degenerate embedding layer. The input projection layer is the input to the GRU encoder, the temporal coding layer includes multiple GRU units, and the global degenerate embedding layer takes the hidden state of the last GRU unit at the last time step as the degenerate embedding vector. The specific architecture includes:

[0093]

[0094] in, Represents the ReLU activation function. Indicates time Input at the location, This represents the weight matrix of the input projection layer. This represents the bias term of the input projection layer. This represents the initial hidden state after projection;

[0095]

[0096] Among them, multi-layer GRU units process projection features. Indicates the first Layer GRU unit in time The hidden state, Indicates the first Layer in time The output, Indicates the first Layer in time The hidden state, Indicates the first Layer GRU unit; This indicates the number of GRU layers, and the hidden state at the final time step is taken as a degenerate embedding. , For the final degenerate embedding vector, For the last layer of GRU at the last time step The hidden state.

[0097] As a preferred embodiment, the degradation stage classification head includes a fully connected layer with a ReLU activation function and a softmax output layer. The stage probability vector is calculated as follows:

[0098]

[0099] in, , This represents the weights and biases of the hidden layer in the classification head during the stage. , This represents the weights and biases of the output layer of the stage classification head. This represents the activation function of the hidden layer. This indicates that the output will be converted into a probability distribution. Corresponding to the early, middle, and late degradation stages respectively, the predicted degradation stage can be obtained through the following methods:

[0100]

[0101] in, Indicates the first The probability of degradation in each prediction stage. This indicates that the category with the highest probability is selected.

[0102] Specifically, the RUL regression head performs regression based on the degenerate embedding vector to output a predicted remaining lifetime value:

[0103]

[0104] In the formula, the remaining lifetime prediction output value during training is... It is constrained to be non-negative to ensure physical rationality; Indicates the activation function of the hidden layer; , This represents the weights and biases of the RUL regression head output layer; Represents a degenerate embedding vector. , This represents the weights and biases of the hidden layer in the RUL regression head.

[0105] The feasibility and effectiveness of the method in this embodiment are verified below with specific examples.

[0106] The robot joint module and related data acquisition data used in this example are shown in Table 3:

[0107] Table 3 Specific experimental parameters

[0108]

[0109] The fault diagnosis and life prediction method for robot joint module stage perception in this example includes:

[0110] Step 1: Using the parameters shown in Table 3, vibration signals of the robot joint module were collected throughout its entire life cycle using a PCB piezoelectric accelerometer. Based on this, the root mean square (RMS), kurtosis, and centroid frequency, and other time-frequency domain features of each sample were extracted and converted into SHR that can describe the degradation state, thus constructing an experimental dataset containing both signal and semantic features.

[0111] The second step is to divide the dataset into a training set and a test set. Based on this, the semantic health representations from the training set are input into a pre-trained LLM to build an LLM inference engine. This engine utilizes the LLM's logical reasoning capabilities to output advanced guidance information with structured degradation knowledge, including the current degradation stage and degradation trend.

[0112] Step 3: Using a stage-aware adaptive module, dynamic training weights are assigned to vibration signals at different time stamps based on the degradation stage information output by the LLM, thus achieving the cascading of stage information. In the later stages of degradation, the weight coefficients of the loss function are automatically increased, thereby obtaining a fusion representation that can sensitively capture the characteristics of severe degradation.

[0113] Step 4: Repeat steps 2 and 3, inputting the vibration signal sequence of the entire life cycle into the neural network prediction model in batches, and using the GRU encoder to deepen the feature extraction layers to improve the model's training and learning ability for nonlinear degradation trajectories.

[0114] During the model's training process, feature vectors incorporating high-level degradation knowledge are input into the physical constraint regularization module. A physical heuristic regularization module is constructed using monotonicity constraints, degradation rate constraints, and acceleration constraints. The initial RUL predictions output by the neural network are subjected to physical consistency checks. A penalty function is used to suppress non-physical rebounds or abnormal fluctuations in the predictions over time. The physical constraints are implemented by calculating the first and second differences of the predictions, forcing the model to follow the physical laws of fatigue damage accumulation. The output of the physical constraint module is weighted and fused with the task prediction error (MSE) to calculate the global multi-objective loss function.

[0115] An early stopping strategy is employed during training. The Adam optimizer is used to train, optimize, and update the parameters of the overall model during backpropagation until the model converges on the training set. The network model with optimal parameters is saved, and a health assessment is performed on the robot joint modules in the test set. The remaining life prediction curve and degradation stage classification are output, where the degradation stage category represents the degree of failure (e.g., primary failure, moderate failure, severe failure), thereby realizing fault diagnosis.

[0116] The trained model is deployed to predict the remaining lifetime and classify the degradation stage of real-time data, thereby achieving a complete closed loop from signal processing and knowledge reasoning to predictive decision-making.

[0117] The remaining lifetime prediction results and degradation stage classification results obtained in this example are as follows: Figure 3 and Figure 4 As shown. By Figure 3 It can be seen that the RUL comparison chart of the predicted value and the actual value shows a close match throughout the entire life cycle; from Figure 4It can be seen that the fault severity characteristics of the robot joint module are clearly classified in the early, middle and late stages. This indicates that the combination of degradation inference based on a large language model, stage-aware adaptive dynamic adjustment of the predictor's training loss function weights at different stages, and physical constraint regularization can robustly model complex degradation patterns.

[0118] In summary, this invention leverages the capabilities of large language models in logical reasoning and cross-modal knowledge association, utilizing their high-level knowledge reasoning abilities to design an SHR-driven degradation reasoning approach. This guides the knowledge cascading behavior from raw vibration signals to high-level structured prior knowledge, and incorporates multi-source information fusion during model training. Furthermore, a stage-aware adaptive learning strategy and physically heuristic constraint regularization are employed to generate physically consistent deep embeddings. This suppresses prediction biases and non-physical fluctuations in the heterogeneous degradation stage of robot joint modules by purely data-driven models, enabling the model to further learn implicit degradation physical characteristics. Consequently, this achieves accurate prediction of the remaining lifespan and fault characteristics of robot joint modules.

[0119] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the lifespan of a robot joint module based on stage perception, characterized in that, include: The original vibration signals of the robot joint module throughout its entire life cycle are collected and transformed into a semantic health representation describing the evolution of its health status after signal processing. A degradation inference model based on a large language model is constructed, which takes the semantic health representation as input and generates structured prior knowledge containing the degradation stage, degradation trend and physical boundary constraints of the robot joint module. A predictor based on a deep neural network is constructed, which takes the structured prior knowledge as input and generates degradation stage identification and remaining life prediction results. The predictor structure includes a GRU encoder, a degradation stage classification head, and a RUL regression head; the GRU encoder extracts the temporal degradation features of the vibration signal to obtain a degradation embedding; the degradation stage classification head maps the degradation embedding to a degradation stage category; the RUL regression head predicts the remaining lifetime based on the degradation embedding; The predictor is equipped with a stage-aware adaptive learning module, which dynamically adjusts the weights of the training loss function of the predictor at different degradation stages based on the degradation stages in the structured prior knowledge, so as to balance the stage classification loss and RUL regression loss in the training loss; the predictor is equipped with a physical heuristic constraint regularization module, which constructs a physical heuristic loss based on the structured prior knowledge. The training loss and the physical heuristic loss are weighted and fused to form the total loss. The model parameters are then jointly trained and optimized using the backpropagation algorithm to form the predictor. The physical heuristic loss consists of three constraints related to degradation characteristics, including: ， In the formula: For time points t Physical heuristic loss; The strength of each constraint; monotonicity loss constraint , and Predicted remaining lifetime values ​​at time points t and t-1, respectively; degradation rate constraint. , , Upper bound of degradation rate; degradation acceleration constraint , , The predicted remaining lifetime value at time point t-2. Indicates the degeneration stage The expected rate of degradation.

2. The method according to claim 1, characterized in that, The calculation of the training loss function weights includes: ， In the formula, For time points t The training loss function weights are the first... Degradation probability at each degradation stage Expectations Output by the degradation stage classification head; These are the preset stage base weights; K This represents the total number of degradation stages.

3. The method according to claim 2, characterized in that, The degradation stage The expected degradation acceleration level is defined as follows: ， Here, early, middle, and late represent the early, middle, and late stages of degradation, respectively.

4. The method according to claim 3, characterized in that, when K= 3. The preset stage base weights are: ， in, This is to characterize the higher weighting of remaining lifetime regression analysis as the degree of degradation increases; k =1,2,3 correspond to early, middle, and late, respectively.

5. The method according to claim 1, characterized in that, The construction of the semantic health representation includes: Time-frequency analysis was performed on the acquired raw vibration signal segments to extract a set of degradation-sensitive characteristics from the time-frequency representation; In the original vibration signal segment t Multiple degradation-sensitive features are extracted at time to form a multi-dimensional time-frequency feature vector, which is then normalized. Principal component analysis was used to fuse the multidimensional time-frequency feature vectors into a single health indicator. ;right Performing a first-order difference, we obtain , , They represent time respectively , Health indicator values ​​at the location, Represent the rate of change of health indicators; construct a series of health indicators. , ; A set of statistical features is extracted from the vibration signal segment to provide supplementary degradation information, time. Statistical eigenvectors at [location] RMS, Kurtosis, Skewness, Crest, and Entropy represent root mean square, kurtosis, skewness, peak factor, and signal entropy, respectively. Determine the running condition vector , , , They represent time respectively Speed, load level, and temperature at the location; A semantic health representation is constructed based on the health indicator sequence, statistical feature vector, and operating condition vector. .

6. The method according to claim 5, characterized in that, The degradation-sensitive characteristics include time-frequency energy, spectral centroid, spectral bandwidth, and high-frequency energy ratio.

7. The method according to claim 1, characterized in that, The construction of the degradation reasoning model includes: Construct a prompt template, convert the semantic health representation into a natural language description, input it into a pre-trained large language model, and output the parsed structured prior knowledge. , The terms indicate the inferred degradation stage, with early, middle, and late representing the early, middle, and late degradation stages, respectively. Indicating a degradation trend, stable, moderate, and accelerating represent stable degradation trends, moderate degradation trends, and accelerating degradation trends, respectively. This represents the inferred constraints related to the degradation characteristics, which are embedded as soft regularization terms into the learning objective of the loss function.

8. The method according to claim 7, characterized in that, The upper bound of the degradation rate Adaptive determination based on degradation trends inferred from large language models: 。 9. The method according to claim 1, characterized in that, The GRU encoder includes an input projection layer, a temporal coding layer, and a global degenerate embedding layer. The input projection layer is the input of the GRU encoder. The temporal coding layer includes multiple GRU units. The global degenerate embedding layer takes the hidden state of the last GRU unit at the last time step as the degenerate embedding vector.

10. The method according to claim 9, characterized in that, The RUL regression head performs regression based on the degenerate embedding vector and outputs a predicted remaining lifetime value: ， In the formula, the remaining lifetime prediction output value during training is... It is constrained to be non-negative to ensure physical rationality; Indicates the activation function of the hidden layer; , This represents the weights and biases of the RUL regression head output layer; Represents a degenerate embedding vector. , This represents the weights and biases of the hidden layer in the RUL regression head; The degradation stage classification head includes a fully connected layer with a ReLU activation function and a softmax output layer.

Images