CMC residual life prediction method based on TCN network and multi-feature fusion

Abstract

The invention discloses a TCN network and multi-feature fusion-based CMC residual life prediction method, belongs to the technical field of material nondestructive testing and life prediction, and constructs a TCN network-based CMC residual life prediction model in combination with acoustic emission signal clustering analysis and fatigue hysteresis behavior features. By introducing features such as a clustering analysis result and a hysteresis loop area, the model can more comprehensively capture multi-scale features of CMC fatigue damage, and the capability of distinguishing different damage modes is enhanced. According to the method, contributions of different damage modes to the residual life can be quantified more accurately, the model can automatically learn the influence weight of each damage mode, so that the prediction precision of the CMC residual life is remarkably improved, in addition, the universality and reliability of the model are further improved due to the advantages of the TCN network in processing long sequence data, and the prediction precision of the CMC residual life is improved. And effective prediction of the residual life of the CMC is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-destructive testing and life prediction of materials, and relates to a method for predicting the remaining life of ceramic matrix composites (CMC) based on acoustic emission (AE) clustering and hysteresis loop characteristics and a TCN network. Background Technique

[0002] CMC has excellent properties such as high specific strength, high specific modulus, and high temperature resistance, and is an ideal material for hot-end components of aeroengines. However, in actual use, CMC inevitably bears fatigue loads, and its failure mechanism is complex, involving multi-damage coupling such as matrix cracking, interface wear, and fiber fracture. Under long-term cyclic fatigue loads, the life dispersion of CMC is significant, mainly due to the non-uniformity of its microstructure, component content, and randomly distributed defects. Traditional physical or empirical models usually rely on pre-measured ultimate strength and loading modes, ignoring the uncertainties in damage evolution and accumulation, and it is difficult to provide accurate life prediction. Therefore, it is of great significance to explore a more general method to accurately describe the damage mechanical behavior of CMC under long-term loads.

[0003] Acoustic emission (AE) technology is an important means in the field of structural health monitoring and has been widely used in damage detection and identification of composite materials. In recent years, AE-based data-driven models have gradually become a research hotspot for remaining useful life (RUL) prediction due to their excellent ability to process large amounts of data and map complex non-linear relationships. However, the existing methods have the following deficiencies:

[0004] 1) Insufficient time series dependence: Traditional artificial neural networks (ANN) do not consider the dependence in time series data and cannot effectively capture the cumulative damage characteristics during the fatigue process of CMC. Although the LSTM network performs well in sequence modeling, its processing ability for long sequence data is limited, and the training time is relatively long (Rohman M N, Hidayat M IP, Purniawan A. Prediction of composite fatigue life under variable amplitude loading using artificial neural network trained by genetic algorithm[C] / / AIP Conference Proceedings. AIP Publishing, 2018, 1945(1).).

[0005] 2) Insufficiencies in acoustic emission signal processing: Traditional research usually conducts clustering analysis on acoustic emission signals to obtain the evolution laws of different damage types. However, existing research mostly takes the overall acoustic emission signal as the input and rarely considers the contribution differences of different damage types to the remaining life (Hao J, Rupp M, Lomov S V, et al. Remaining useful life prediction of flax fibre biocomposites under creep load by acoustic emission and deep learning[J]. Composites Part A: Applied Science and Manufacturing, 2025, 188: 108572.).

[0006] 3) Insufficient consideration of hysteresis behavior: Existing RUL prediction methods are mostly based on damage detection signals, and the cumulative characteristics of CMC fatigue damage are also significantly reflected in hysteresis behavior. For example, some research has developed a fatigue life prediction method for CMC structures using hysteresis loops as damage variables, but existing models consider this less, resulting in limitations in the generality of the models (Li J, Cao X, Chen R, et al. Prediction of remaining fatigue life of metal specimens using data-driven method based on acoustic emission signal[J]. Applied Acoustics, 2023, 211: 109571).

[0007] In recent years, the Temporal Convolutional Network (TCN) has shown better performance than LSTM and GRU in sequence modeling tasks through dilated causal convolution and residual block structures, especially having significant advantages in processing long sequence data. Therefore, based on the TCN network, combining the results of AE signal clustering analysis and fatigue hysteresis behavior to construct a CMC remaining life prediction model to improve the life prediction accuracy of CMC under long-term fatigue service loads has important value. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a CMC remaining life prediction method based on the TCN network and multi-feature fusion to effectively predict the remaining life of CMC under long-term service loads in view of the above existing deficiencies.

[0009] To achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0010] CMC Remaining Life Prediction Method Based on TCN Network and Multi-Feature Fusion, including the following steps:

[0011] Step 1: Conduct CMC fatigue tests and acoustic emission tests to obtain the hysteresis behavior curves and acoustic emission characteristic parameters at different cycle numbers.

[0012] Step 2: Conduct clustering analysis on the acoustic emission signals to identify the evolution law of different component damages with the cycle number.

[0013] Step 3: Extract features from the hysteresis behavior, calculate the area of the hysteresis loop, analyze the variation law of the cumulative area of the hysteresis loop with the cycle number, and quantify the influence of the hysteresis behavior on the fatigue damage accumulation.

[0014] Step 4: Convert the acoustic emission signals related to time into acoustic emission signals at different cycle numbers. At the same time, merge the cumulative areas of the hysteresis loops at different cycle numbers as the input for the training of the TCN network model, and the output is the remaining fatigue life of CMC; at different cycle numbers, the acoustic emission characteristic parameters, hysteresis characteristics, and remaining cycle numbers constitute the training data set.

[0015] Step 5: Normalize the input and output data sets.

[0016] Step 6: Build a TCN network model and set the training parameters.

[0017] Step 7: Use the constructed data set to train the TCN network model, and use the trained TCN network model to predict the remaining life of CMC, and evaluate the accuracy and reliability of the model through error analysis.

[0018] To optimize the above technical solution, the specific measures taken also include:

[0019] The specific process of Step 1 is as follows: Conduct CMC fatigue tests under different stresses, measure the stress and strain during the tests to obtain the hysteresis behavior curves at different cycle numbers; at the same time, conduct acoustic emission tests during the fatigue tests and record the acoustic emission characteristic parameters, and the acoustic emission characteristic parameters include peak frequency, amplitude, and count.

[0020] In Step 2, the specific method for clustering analysis of acoustic emission signals is as follows: Select the peak frequency and combine it with the amplitude and count respectively to identify different meso-damage forms. First, comprehensively consider the elbow method and the silhouette coefficient method to determine the optimal number of clusters and the CMC meso-damage forms, and determine the number of clusters for CMC acoustic emission signal clustering. The CMC damage form is related to its meso-composite structure, which is composed of fibers, matrix, and interface. Under fatigue load, the corresponding damage forms are fiber fracture, interface debonding and slip, and matrix cracking. In summary, set the number of clusters to 3, and use K-means for clustering analysis. Analyze the acoustic emission characteristic signals of the peak frequency-energy and peak frequency-count combinations, and classify the signal with the highest peak frequency to the fiber, followed by the matrix and the interface. Subsequently, respectively count the energy and count acoustic emission signals corresponding to the meso-damage of the fiber, matrix, and interface, obtain their cumulative change rules with the number of cycles, and obtain the acoustic emission component damage identification result.

[0021] In Step 5, maximum normalization is used to process the input and output data sets.

[0022] The TCN network model combines causal convolution and dilated convolution. Causal convolution ensures that the output of each time stamp depends only on historical inputs, avoiding leakage of future information; dilated convolution exponentially expands the receptive field, enabling the network to capture long-term dependencies. The residual block of the TCN network includes a dilated causal convolution layer, a weight normalization layer, a ReLU activation function, a spatial Dropout layer, and a 1×1 convolution layer. The feature transfer and gradient optimization processes are as follows:

[0023] 1) The input signal sequentially passes through the first dilated causal convolution layer to extract multi-scale temporal features;

[0024] 2) Through weight normalization, ReLU activation, and the spatial Dropout layer, standardization, non-linear transformation, and regularization are completed;

[0025] 3) The second dilated causal convolution layer repeats Step 2) to form a two-layer convolution process;

[0026] 4) After the original input is aligned with the channels through the 1×1 convolution layer, it performs residual addition with the output of Step 3) to construct a gradient direct connection structure;

[0027] 5) The output that fuses local features and global dependencies is stacked through multiple layers to achieve deep feature fusion.

[0028] The network adopts a zero-padding strategy to maintain temporal causality, the residual connection alleviates the vanishing gradient, the dilated convolution expands the receptive field, and each component collaborates to optimize, ensuring the stable convergence of the TCN network model.

[0029] In step 6, the training parameters include the learning rate of the TCN network model, batch size, number of iterations, weight decay parameter, Dropout rate, and dilation coefficient.

[0030] In step 7, the specific method for training the TCN network model using the constructed dataset is as follows: First, the dataset is divided into a training set and a test set. The training set is used for training the TCN network model, and the test set is used for verification. During the training process, the TCN network model calculates the predicted values through forward propagation and measures the deviation between the predicted values and the true values using a loss function. Subsequently, the network parameters are updated using the backpropagation algorithm.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] (1) By performing clustering analysis, the present invention classifies acoustic emission signals into three mesoscopic damage modes: fiber fracture, interface debonding and slip, and matrix cracking, and uses the classification results as model inputs, rather than directly using the overall acoustic emission signal characteristics. This method can more accurately quantify the contributions of different damage modes to the remaining life, enabling the model to automatically learn the influence weights of each damage mode, thereby significantly improving the prediction accuracy of the remaining life of CMC.

[0033] (2) The present invention combines acoustic emission signal clustering analysis and fatigue hysteresis behavior characteristics to construct a CMC remaining life prediction model based on the TCN network. By introducing features such as clustering analysis results and hysteresis loop area, the model can more comprehensively capture the multi-scale characteristics of CMC fatigue damage, enhancing the ability to distinguish different damage modes. In addition, the advantages of the TCN network in processing long sequence data further improve the generality and reliability of the model, achieving an effective prediction of the remaining life of CMC. Description of the Drawings

[0034] Figure 1 is the peak frequency-energy combination clustering result diagram;

[0035] Figure 2 is the peak frequency-count combination clustering result diagram;

[0036] Figure 3 is the diagram of the change of acoustic emission cumulative energy with the number of cycles

[0037] Figure 4 is the diagram of the change of acoustic emission cumulative count with the number of cycles;

[0038] Figure 5 is the diagram of the change of the cumulative area of the hysteresis loop with the number of cycles;

[0039] Figure 6 is the visualization diagram of the TCN network model;

[0040] Figure 7It is the convergence graph of the training error of the TCN network model;

[0041] Figure 8 It is the graph of the predicted remaining life of the typical CMC. Detailed implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments provided in the present application without creative efforts fall within the scope of protection of the present application.

[0043] Obviously, the accompanying drawings in the following description are only some examples or embodiments of the present application. For those of ordinary skill in the art, without creative efforts, the present application can also be applied to other similar scenarios based on these drawings. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in the present application, some design, manufacturing or production changes based on the technical content disclosed in the present application are only conventional technical means and should not be understood as the content disclosed in the present application being insufficient.

[0044] Referring to "embodiment" in the present application means that the specific features, structures or characteristics described in connection with the embodiment can be included in at least one embodiment of the present application. The phrase appears in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments without conflict.

[0045] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the ordinary meanings understood by those with ordinary skills in the technical field to which this application belongs. The words such as "a", "an", "one kind", "the" and the like involved in this application do not indicate a limitation in quantity and may represent singular or plural. The terms "include", "comprise", "have" and any variations thereof involved in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product or device that includes a series of steps or units (units) is not limited to the listed steps or units, but may further include unlisted steps or units, or may further include other steps or units inherent to these processes, methods, products or devices. The words such as "connect", "be connected", "couple" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" / "several" involved in this application means two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the front and back associated objects. The terms "first", "second", "third" and the like involved in this application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0046] Step 1: Conduct CMC fatigue tests and acoustic emission tests to obtain the hysteresis behavior curves and acoustic emission characteristic parameters at different cycle numbers, providing a data basis for model construction. Specifically, conduct CMC fatigue failure tests under different stresses, measure the stress and strain during the tests to obtain the hysteresis curves at different cycle numbers; meanwhile, conduct acoustic emission tests during the fatigue tests and record key acoustic emission characteristic parameters such as peak frequency, amplitude, and count.

[0047] In this example, the relevant parameters for the fatigue experiment are: the peak fatigue stress is 180 MPa, the environmental temperature is 800 °C, and the material is unidirectional CMC.

[0048] Step 2: Conduct clustering analysis on the acoustic emission signals to identify the evolution laws of different component damages such as matrix cracking, interface wear, and fiber fracture with the cycle number. Since different mesoscopic damages have different influences on the remaining life of CMC, clustering analysis is used to classify the acoustic emission signals, and the classification results are used as the input of the TCN network model instead of directly using the overall signal characteristics of the acoustic emission. In this way, the TCN network model can automatically learn the influence weights of different damage modes on the remaining life, thereby improving the prediction accuracy.

[0049] In specific implementation, different combinations of characteristic parameters are selected to perform clustering analysis on acoustic emission signals. Since there are significant differences in the frequency characteristics generated by damage signals of different components, the peak frequency is combined with other acoustic emission statistical characteristics to identify different mesoscopic damage forms. First, considering the elbow method and the silhouette coefficient method comprehensively to determine the optimal number of clusters, as well as the common mesoscopic damage forms of CMC, the number of clusters for CMC acoustic emission signal clustering is determined. Generally, the damage form of CMC is mainly related to its mesoscopic composition structure, which is mainly composed of fibers, matrix, and interface. The corresponding damage forms under fatigue load are mainly three main mesoscopic damage forms: fiber fracture, interface debonding and slip, and matrix cracking. In summary, the number of clustering categories is set to 3. K-means is used for clustering analysis to analyze the acoustic emission characteristic signals of the peak frequency-energy and peak frequency-count combinations. The signal with the highest peak frequency is attributed to the fiber, followed by the matrix and the interface. As Figure 1 shown in the results of the frequency-energy combination clustering analysis, it can be seen that the damage frequency ranges of the fiber, matrix, and interface are 0 - 200k, 200 - 540k, and 540 - 1114k respectively, Figure 2 The results of the frequency-count combination clustering analysis are shown in . The peak frequency division range is similar to that of the frequency-energy combination, indicating that the distinction degree of different component damages is relatively large, and different damages can be well identified and distinguished.

[0050] Subsequently, the energy and count acoustic emission signals corresponding to the mesoscopic damages of the fiber, matrix, and interface are respectively statistically analyzed to obtain their cumulative change laws with the number of cycles. As Figure 3 and Figure 4 show the laws of the cumulative energy and cumulative count of different components changing with the number of cycles respectively. It can be seen that during the entire fatigue cycle, the damage signal of the matrix is relatively strong, while the damage signal of the fiber is the weakest.

[0051] Step 3: Extract the characteristics of the hysteresis behavior, calculate the area of the hysteresis loop, analyze the change law of the cumulative area of the hysteresis loop with the number of cycles, and quantify the influence of the hysteresis behavior on the cumulative fatigue damage. The hysteresis behavior of CMC can comprehensively reflect damage forms such as fiber failure, interface slip, and matrix cracking. Therefore, performing cumulative statistics on it can describe the cumulative damage evolution process of CMC during fatigue, and it is an important parameter for measuring the fatigue damage behavior of unidirectional CMC. When evaluating the remaining life of CMC, the cumulative area of the hysteresis loop is a key input.

[0052] As Figure 5 shown, it is the change law of the cumulative area of the hysteresis loop with the number of cycles. It has significant monotonicity, which can reflect the process of gradual damage accumulation and the trend of gradual reduction of the remaining life.

[0053] Step 4: Construct the remaining life prediction dataset. Based on the correspondence between time and the number of cycles, convert the acoustic emission signals related to time into acoustic emission signals at different numbers of cycles. Meanwhile, merge the cumulative areas of the hysteresis loops at different numbers of cycles as the input. The output is the remaining fatigue life of CMC, that is, subtract the life used in the current number of cycles from the total life of CMC to obtain the change in the remaining fatigue life of CMC. At different numbers of cycles, the acoustic emission characteristic parameters, hysteresis characteristics, and the remaining number of cycles constitute the training dataset. In this example, under a peak stress of 180 MPa, the fatigue life is 16,560 N.

[0054] Step 5: Perform data normalization to eliminate the dimensional difference and improve the training efficiency and prediction accuracy of the model. Since there is a large difference in the order of magnitude between different input and output variables, use maximum normalization to eliminate the influence of different magnitudes, laying a foundation for accurately predicting the remaining life of CMC.

[0055] Step 6: Build a TCN network model and set the training parameters. The core feature of the TCN (Temporal Convolutional Network) network lies in the combination of causal convolution and dilated convolution. Causal convolution ensures that the output at each timestamp depends only on the historical input, avoiding interference from future information, and is suitable for time series prediction tasks; dilated convolution exponentially expands the receptive field, enabling the network to capture long-term dependencies without significantly increasing the network depth or the number of parameters. The key component of TCN is the residual block, which usually includes dilated causal convolutional layers, weight normalization layers, ReLU activation functions, spatial Dropout layers, and an optional 1x1 convolutional layer (used to adjust the dimensions of the residual input and output). By stacking multiple residual blocks, TCN can flexibly adapt to different sequence learning tasks.

[0056] When training the TCN network model, key parameters need to be set: the initial learning rate and dynamically adjust it in combination with a learning rate scheduler; the batch size to balance the training speed and memory occupancy; the number of iterations and combine an early stopping strategy to prevent overfitting; it is recommended to use Adam as the optimizer and set the weight decay parameter to regularize the model; the Dropout rate is used to prevent overfitting; the dilation coefficient grows exponentially to capture dependencies at different time scales. By reasonably setting the parameters and tuning them in combination with cross-validation or grid search, the model performance can be effectively improved.

[0057] In this example, the TCN network model is as Figure 6 shown, and some model parameters are: the initial learning rate is 0.01, the maximum number of training times is 300, the optimizer is Adam; L1 and L2 regularization are adopted.

[0058] Step 7: Train and validate the model using the constructed dataset. First, divide the data into a training set and a test set. The training set is used for model training, and the test set is used for validation. During the training process, the model calculates the predicted values through forward propagation and measures the deviation between the predicted values and the true values using a loss function. Subsequently, the network parameters are updated using the backpropagation algorithm. After training is completed, evaluate the model performance using the test set, quantify the prediction accuracy by calculating the error metrics, and plot the comparison curve between the predicted values and the true values. Finally, through error analysis and result comparison, verify the effectiveness and reliability of the model in the CMC fatigue remaining life prediction task.

[0059] In this example, the training set is 80% of the data, and the proportion of the validation set is 20%. The curves of the errors on the training set and the validation set versus the number of iterations are as Figure 7 shown, and it can be seen that the errors converge quickly. Finally, over the entire life time, the predicted remaining life is compared with the test results as Figure 8 shown, and it can be seen that the overall trends of the predictions are consistent, indicating that the remaining life prediction method proposed in this patent is feasible.

[0060] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the described embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent substitution methods and are all included in the protection scope of the present invention.

Claims

1. The CMC remaining life prediction method based on TCN network and multi-feature fusion is characterized by: The following steps are involved: Step 1: Carry out CMC fatigue test and acoustic emission test to obtain hysteresis behavior curves and acoustic emission characteristic parameters under different cycle numbers. Step 2: cluster analysis is performed on the acoustic emission signals to identify the evolution of damage of different components with the number of cycles; Step 3: Extract the characteristics of hysteresis behavior, calculate the hysteresis loop area, analyze the variation of the cumulative area of ​​hysteresis loop with the number of cycles, and quantify the influence of hysteresis behavior on fatigue damage accumulation; Step 4: Convert the time-related acoustic emission signal into acoustic emission signals at different cycle numbers. At the same time, merge the cumulative areas of the hysteresis loops at different cycle numbers as the input for TCN network model training, and output the CMC remaining fatigue life. Under different numbers of cycles, acoustic emission characteristic parameters, hysteresis characteristics and remaining number of cycles constitute the training data set; Step 5: Normalize the input and output data sets; Step 6: Build the TCN network model and set the training parameters; Step 7: Use the constructed data set to train the TCN network model, and use the trained TCN network model to predict the remaining life of CMC, and evaluate the accuracy and reliability of the model through error analysis.

2. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: The specific process of step 1 is: carry out CMC fatigue tests under different stresses, measure the stress and strain during the test to obtain the hysteresis behavior curves under different cycle numbers; at the same time, carry out acoustic emission tests during the fatigue test and record the acoustic emission characteristic parameters, which include peak frequency, amplitude and count.

3. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: In step 2, the specific method for clustering analysis of acoustic emission signals is as follows: the peak frequency is selected and combined with the amplitude and count respectively to identify different mesoscopic damage forms. First, the elbow method and the silhouette coefficient method are comprehensively considered to determine the optimal number of clusters and the CMC mesoscopic damage form, and the number of CMC acoustic emission signal clusters is determined. The CMC damage form is related to its mesoscopic composition structure. The mesoscopic composition structure consists of fiber, matrix and interface. The corresponding damage forms under fatigue load are fiber fracture, interface debonding and slippage, and matrix cracking. In summary, the cluster type is set to 3, and K-means is used for cluster analysis. The acoustic emission characteristic signals of the peak frequency-energy and peak frequency-count combinations are analyzed, and the signal with the highest peak frequency is attributed to the fiber, followed by the matrix and the interface. Subsequently, the energy and count acoustic emission signals corresponding to the mesoscopic damage of the fiber, matrix and interface are respectively counted to obtain their cumulative change law with the number of cycles, and the acoustic emission component damage identification result is obtained.

4. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: In step 5, the input and output data sets are processed using maximum normalization.

5. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: The TCN network model combines causal convolution and dilated convolution. Causal convolution ensures that the output of each timestamp depends only on historical input to avoid future information leakage; dilated convolution enables the network to capture long-term dependencies by exponentially expanding the receptive field. The residual block of the TCN network model includes an expanded causal convolution layer, a weight normalization layer, a ReLU activation function, a spatial Dropout layer, and a 1×1 convolution layer. The feature transfer and gradient optimization process is as follows: 1) The input signal passes through the first dilated causal convolutional layer in turn to extract multi-scale temporal features; 2) After weight normalization, ReLU activation and spatial Dropout layer, standardization, nonlinear transformation and regularization are completed; 3) The second dilated causal convolutional layer repeats step 2) to form a double-layer convolution process; 4) After the original input is aligned through the 1×1 convolutional layer channels, the residual is added to the output of step 3) to construct a gradient pass-through structure; 5) Fuse local features with global dependent outputs, stack multiple layers, and achieve deep feature fusion; The network uses a zero-padding strategy to maintain temporal causality, residual connections to alleviate gradient vanishing, dilated convolutions to expand the receptive field, and coordinated optimization of various components to ensure stable convergence of the TCN network model.

6. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: In step 6, the training parameters include the TCN network model learning rate, batch size, number of iterations, weight decay parameter, Dropout rate, and expansion coefficient.

7. The CMC remaining life prediction method based on TCN network and multi-feature fusion according to claim 1 is characterized in that: In step 7, the specific method of using the constructed data set to train the TCN network model is as follows: first, the data set is divided into a training set and a test set. The training set is used for TCN network model training, and the test set is used for verification. During the training process, the TCN network model calculates the predicted value through forward propagation, and measures the deviation between the predicted value and the true value through the loss function, and then uses the back propagation algorithm to update the network parameters.

Images