Method for predicting service life of rubber V belt by fusing multi-dimensional features and double attention

Abstract

The invention discloses a rubber V belt residual life prediction method fusing multi-dimensional features and double attention. The method comprises the steps that multi-source time sequence data such as slip ratio, tension and temperature are collected and preprocessed; extracting local coupling features of multi-sensor signals at the same moment by using one-dimensional convolution, fusing the local coupling features with original features, and capturing long-term time sequence dependence through a long-short-term memory network; introducing a damage increment inversion module based on rubber fatigue physics, and mapping a running state into a physical quantity sequence such as accumulated damage degree; and focusing key stages and states by using a double attention mechanism aiming at the fusion feature sequence, and outputting a life value by a regression layer. According to the method, the problems of poor interpretability and insufficient small sample generalization in the prior art are solved through fusion of a physical mechanism and data driving, and the prediction precision, the generalization ability and the engineering credibility are improved.

Description

Technical Field

[0001] This invention relates to the field of rubber V-belt life prediction technology, and more specifically to a method for predicting the life of rubber V-belts that integrates multi-dimensional features and dual attention. Background Technology

[0002] Accurately predicting the remaining life of rubber V-belts is the core of achieving predictive maintenance of transmission systems. The degradation process of V-belts is a typical spatiotemporal coupling problem: in the time dimension, its performance degradation has a long-term dependence; in the characteristic dimension, multi-source sensor signals (such as slip rate, tension, and temperature) have complex local coupling relationships at the same time.

[0003] Existing prediction methods mostly employ single models, such as LSTM and CNN. LSTM excels at capturing long-term dependencies in time series but cannot effectively extract local coupling patterns between multiple features at the same time point; simply concatenating multiple features as input ignores this fine-grained interaction information. CNN can extract local features, but traditional CNNs struggle to model long-range time dependencies when processing time series signals. Furthermore, during V-belt degradation, the importance of different sensor channels and different time points for lifetime prediction changes dynamically, and existing methods lack adaptive attention mechanisms to this dynamic importance. More importantly, existing data-driven methods are essentially "black box" models; their prediction process is completely disconnected from the physical mechanisms of fatigue damage in rubber materials, resulting in poor generalization ability under small sample conditions, prediction results lacking physical basis, and difficulty in gaining trust in engineering practice.

[0004] Therefore, there is an urgent need for a dedicated prediction model that can simultaneously model local coupling and long-term temporal dependence between features, and can incorporate physical mechanisms into the model to achieve interpretable and highly robust prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting the lifespan of rubber V-belts that integrates multi-dimensional features and dual attention mechanisms. By integrating multi-dimensional features and dual attention mechanisms, and combining physical inversion with data-driven models, this method solves the problems that a single model cannot take into account local coupling features and long-term temporal dependencies, lacks dynamic attention mechanisms, and has poor physical interpretability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention, comprising the following steps: Step 1: Data acquisition and preprocessing. Real-time acquisition of multi-source time-series data during the operation of the rubber V-belt. The multi-source time-series data includes at least slip rate, tension, and temperature. The multi-source time-series data is then cleaned and standardized. Step 2: Construct sample sequences. Use the sliding window method to segment the preprocessed multi-source time series data to construct a supervised learning sample set with time steps. Then, each sample in the sample set is input into a one-dimensional convolutional neural network to independently convolve the feature dimension at each time step, explicitly extracting the local coupling features between signals from different sensors at the same time. Then, the local features extracted by the CNN are concatenated with the original input features, and adaptive weights are generated through a learnable fully connected layer to achieve weighted fusion and form an enhanced feature representation. Step 3: Perform temporal dependency modeling, input the fused feature sequence into the long short-term memory network, and capture its long-term temporal evolution pattern; Step 4: Physical inversion and feature reconstruction of rubber belt damage increment. Based on the feature vector sequence representing the operating state of the equipment output by the LSTM network in Step 3, as well as the preset geometric parameters of the rubber V-belt structure and the fatigue parameters of the rubber material, the cumulative damage time series sequence inside the rubber material is calculated in real time through an embedded simplified physical model. The cumulative damage sequence is fused with the LSTM output feature sequence to generate a physical-data fusion feature sequence. Step 5: The physical-data fusion feature sequence output in Step 4 is subjected to temporal attention weighting and channel attention weighting in sequence to focus on key damage time periods and physical state signals that are crucial for life prediction. Step six involves integrating the feature sequences refined through dual attention, outputting the remaining life prediction value of the rubber V-belt through a fully connected regression layer, and performing end-to-end joint training using the Huber loss function and the Adam optimizer.

[0007] As a further improvement to the present invention, the specific steps of data acquisition and preprocessing in step one are as follows: Step 1: Data cleaning and alignment. First, the raw data is inspected. For spike noise caused by transient sensor interference, it is identified using the 3-Sigma criterion and replaced using linear interpolation of the preceding and following data. For missing values ​​at fewer than 5 consecutive time points, linear interpolation is used to fill in the missing values. Steps one and two, data standardization: To eliminate the influence of dimensions, Z-score standardization is performed on the three feature sequences of slip rate, tension, and temperature after cleaning. That is, for each feature, its mean on the training set is calculated. ) and standard deviation ( Then, transform all the data (including the training and test sets): ; Step 13, Constructing supervised learning samples: The complete, standardized multidimensional time series data is divided into continuous samples using the sliding window method, and then all samples are divided into training set, validation set and test set.

[0008] As a further improvement of the present invention, the one-dimensional convolutional neural network used in step two is specifically as follows: a two-layer one-dimensional convolutional neural network; The first layer consists of 64 convolutional kernels of size (1,3) with ReLU activation. When the time dimension is 1 and the feature dimension is 3, the kernel performs convolution operations across the three features of slip rate, tension, and temperature at each independent time step, explicitly extracting instantaneous coupling patterns such as "high slip rate accompanied by high temperature". The second layer consists of 32 (1,3) convolutional kernels, which further abstract the features. The CNN output features are weighted and fused with the original input to form an enhanced feature sequence. .

[0009] As a further improvement to the present invention, the specific method for modeling the temporal dependency relationship in step three is as follows: The fused feature sequences are input into a two-layer stacked long short-term memory network. The optimal configuration determined by grid search is as follows: the number of hidden units in each LSTM layer is set to 128. Dropout layers and batch normalization layers are introduced between LSTM layers to capture the long-term dynamic law of V-belt performance degradation. The output of all time steps is used as the input of subsequent steps.

[0010] As a further improvement to the present invention, the specific steps of the physical inversion and feature reconstruction of the rubber strip damage increment in step four are as follows: Step 41 involves joint encoding of data abstraction and physical entities, including temporal state characteristics and an inherent physical parameter set P. The inherent physical parameter set P contains preset inherent parameters of the rubber V-belt, determined based on a type B 1499mm V-belt, including: Geometric parameters: wedge angle Nodal height Working corner According to the test bench settings Material fatigue parameters: SN curve coefficients were determined by standard fatigue testing on the same batch of materials. The exponent m = 5.8, and these parameters are in vector form. Input provides an invariant physical benchmark for the learning process; Step 42: Establish a physical mapping network, the structure of which includes: Input layer: The abstract state h_t at time t is concatenated with the physical parameter vector P to form a 133-dimensional joint input; Hidden layer: A fully connected layer with 128 neurons, using the ReLU activation function, responsible for non-linear fusion and feature transformation of the joint input; Output layer: A single linear output neuron, a direct scalar. This refers to the equivalent alternating stress amplitude, with a learning target range of approximately 8 MPa to 25 MPa, corresponding to stress levels from normal to failure. This process is formally represented as: ,in These are the trainable parameters of PMN.

[0011] Step 43: Damage calculation and feature reconstruction to generate an interpretable sequence of physical states.

[0012] As a further improvement of the present invention, the working mechanism of the physical mapping network is as follows: Physics-guided adaptive mapping: PMN learns the mapping function f from "abstract state + physical parameters" to "equivalent stress"; End-to-end gradient-level fusion: The total loss function of the physical mapping network is Where L_physics is the physical consistency loss, and the constraints are... Approximating the reference values ​​provided by finite element simulation ensures that the learning does not deviate from the laws of mechanics; the gradient of L_task, which is the lifetime prediction Huber loss, will propagate back through the entire network.

[0013] As a further improvement to the present invention, the damage calculation and feature reconstruction in step four-three are specifically as follows: Cumulative damage: ; Feature reconstruction: converting physical state sequences The abstract state sequence H of LSTM is concatenated with it to form the physical-data deep fusion feature sequence F_fused, with dimensions [B, 64, 131].

[0014] As a further improvement of the present invention, the specific method for performing temporal attention weighting and channel attention weighting in step five is as follows: Channel attention implementation: Global average pooling is performed on the input feature sequence along the channel dimension to obtain a channel description vector. Then, this vector is input into a two-layer feedforward neural network, and the weight coefficient α of each channel is output using the Sigmoid activation function. Finally, the weight coefficient α is multiplied with the original features channel by channel. Temporal attention implementation: For the feature vector of each time step in the input feature sequence, it is mapped to a scalar score through a fully connected layer. The scalar scores of all time steps are input into the Softmax function for normalization to obtain the importance weight β of each time step. The weight β is then multiplied by the feature vector of the corresponding time step. Synergistic effect and verification results: The weights α and β are applied together to the feature sequence to achieve dynamic control of "focusing on key physical states during critical damage stages".

[0015] The beneficial effects of this invention are as follows: by introducing a differentiable physical inversion module and a channel-temporal dual attention mechanism, the following technical effects are achieved: High prediction accuracy and strong generalization ability: In the rubber V-belt life prediction test, the method of this invention achieved a high-precision prediction with a coefficient of determination (R²) of 0.962 and a mean absolute percentage error (MAPE) of 2.8% (see the comparative experimental data in Specific Implementation S4). Its root mean square error (RMSE) of 8.7 is significantly lower than that of traditional time series models (ARIMA, RMSE 28.4) and mainstream machine learning models (XGBoost, RMSE 22.1). This complete set of comparative experiments (see Table 1 in Section S4) demonstrates the excellent fitting and generalization ability of the model. This effect is fundamentally due to the unique physical inversion design of this invention (see Section S2.c of Specific Implementation). By constructing a differentiable physical computation graph through a trainable physical mapping network (PMN), the model learns a more physically consistent "data → physical state → life" mapping, rather than a purely data-driven "black box" mapping. This reasoning mechanism based on the physical state space is the fundamental reason why the model can maintain high accuracy with small samples and has stronger generalization potential for unexperienced conditions.

[0016] A dual-attention synergy mechanism for fusion-physical-data features is proposed: an innovative cascaded structure of temporal attention and channel attention is designed specifically for processing the fused physical-data feature sequence. Temporal attention enables the model to focus on critical degradation stages that accelerate damage; channel attention automatically distinguishes and reinforces the importance of key physical quantity channels such as "cumulative damage degree". Ablation experiments demonstrate a synergistic effect between the two; removing either channel attention or temporal attention leads to a sharp increase in RMSE (approximately 80% and 63%, respectively), indicating that this dual design is necessary and not obvious for achieving dynamic and accurate prediction that "focuses on key states during critical time periods".

[0017] A deep and reliable fusion of physical mechanisms and data-driven approaches was achieved: by introducing a trainable and differentiable Physical Mapping Network (PMN), data-driven feature learning and the physical calculation of rubber fatigue damage (based on the SN curve and Miner's linear cumulative damage rule) are constructed in a unified, gradient-differentiable computational graph. This allows the model to learn the mapping relationship of "running data → physical damage state → remaining life," rather than an uninterpretable "black box" mapping. The PMN uses physical consistency auxiliary loss for constraint, ensuring that its output conforms to basic mechanical laws. This fusion mechanism fundamentally improves the model's generalization ability and the reliability of prediction results under small sample conditions. Ablation experiments show that after removing this physical inversion module, the model's root mean square error (RMSE) increased from 8.7 to 15.8, and the coefficient of determination (R²) decreased from 0.962 to 0.89, demonstrating the core contribution of this module. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the overall system architecture as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-dimensional feature fusion module (1D-CNN) structure described in an embodiment of the present invention; Figure 3 This is a diagram illustrating the internal working mechanism of the physical inversion module described in this embodiment of the invention. Figure 4 This is a detailed structural diagram of the dual attention mechanism module described in an embodiment of the present invention; Figure 5 This is a schematic diagram of the sliding window method and sequence construction described in an embodiment of the present invention; Figure 6 This is a comparison chart of the prediction results described in the embodiments of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the embodiments shown in the accompanying drawings.

[0020] Reference Figure 1 As shown, the rubber V-belt lifetime prediction method integrating multi-dimensional features and dual attention in this embodiment includes the following steps: S1. Data Preparation and Preprocessing The data used in this embodiment comes from a rubber V-belt accelerated life testing rig. Multiple sensors were arranged on the rig, and time-series data of the following three key parameters were simultaneously acquired at a sampling frequency of 1Hz: Slip ratio: Calculated by measuring the rotational speeds of the driving and driven wheels using an encoder.

[0021] Tension force: Measured by a force sensor installed on the tensioning mechanism.

[0022] Temperature: The surface temperature of the V-belt is measured non-contactly using an infrared temperature sensor.

[0023] Each V-belt records its complete lifecycle data from the start of operation until failure, and marks the remaining lifetime (RUL) corresponding to its failure point as 0. The RUL label is generated by reverse calculation from the initial lifetime using a linear or exponential degradation model.

[0024] Data preprocessing and sequence construction (corresponding) Figure 5 The steps are as follows: a) Data cleaning and alignment: First, the raw data is inspected. For spike noise caused by instantaneous sensor interference, it is identified using the 3-Sigma criterion and replaced using linear interpolation of the preceding and following data. For missing values ​​at fewer than 5 consecutive time points, linear interpolation is used to fill in the missing values.

[0025] b) Data Standardization: To eliminate the influence of dimensions, Z-score standardization was performed on the three feature sequences of slip rate, tension, and temperature after cleaning. That is, for each feature, its mean on the training set was calculated ( ) and standard deviation ( Then, transform all the data (including the training and test sets): .

[0026] c) Constructing Supervised Learning Samples: The complete, standardized multidimensional time-series data is segmented into continuous samples using a sliding window method. In this embodiment, after hyperparameter optimization, the window length is set to 64 time steps, and the sliding step size is set to 1 time step. The 64×3 data matrix within each window serves as the input feature (X) of a sample. The true RUL value at a certain time point after the end of the window is used as the prediction target (Y) for that sample.

[0027] S2. Batch organization of supervised learning sample sets After completing the sample construction described in S1, all samples are divided into training, validation, and test sets. During model training, a DataLoader is used to randomly organize the training set samples into multiple batches. Each batch contains B samples (B=64 in this embodiment), forming an input tensor X_batch with dimension [B, 64, 3] and a corresponding label tensor Y_batch with dimension [B], which serve as the basic data unit for one forward and backward propagation of the model.

[0028] S3. Constructing and training deep prediction models that integrate physical mechanisms a) Multi-dimensional feature fusion module (corresponding to) Figure 2 ) Input Sample (B represents the batch size). Using a two-layer one-dimensional convolutional neural network (1D-CNN): The first layer consists of 64 convolutional kernels of size (1,3) with ReLU activation. The kernel has a length of 1 in the time dimension and a length of 3 in the feature dimension, which means that at each independent time step, it performs convolution operations across the three features of slip rate, tension, and temperature, explicitly extracting instantaneous coupling patterns such as "high slip rate accompanied by high temperature".

[0029] The second layer consists of 32 (1,3) convolutional kernels, which further abstract the features.

[0030] The CNN output features are weighted and summed with the original input (projected to the same dimension by a linear layer) to form an enhanced feature sequence. .

[0031] b) Temporal Dependency Modeling Module The fused feature sequences were input into a two-layer stacked Long Short-Term Memory (LSTM) network. The optimal configuration, determined through grid search, was set to 128 hidden units per LSTM layer. To optimize training and prevent overfitting, Dropout layers (with a dropout rate of 0.2) and BatchNorm layers were introduced between LSTM layers. This module is responsible for capturing the long-term dynamics of V-belt performance degradation, and its output at all time steps serves as the input to subsequent modules. The optimizer was Adam, with an initial learning rate of 0.001, and training was performed using the Mini-Batch method. This configuration (64-step window, 128 hidden units) achieved the lowest root mean square error (RMSE of 2.83 hours) on the validation set, and its final performance on the independent test set is shown in Table 1 of Section S4 below.

[0032] c) Damage increment inversion module based on rubber fatigue physics (corresponding to) Figure 3 ): The goal of this module is not simply to call a fixed formula, but to build a learnable and optimizable "physical law simulator". This simulator (PMN) can receive high-dimensional abstract states extracted by a data-driven network and, referring to the inherent physical nature of the rubber V-belt, deduce the internal damage state that conforms to the laws of fatigue mechanics.

[0033] 1. Module Input: Joint encoding of data abstraction and physical entities Temporal state characteristics: This is a 128-dimensional abstract vector obtained by the LSTM network after deep encoding of current and historical multi-source sensor signals (slip rate, tension, temperature). It represents the model's "understanding" of the current device's operational health status, but it does not possess direct physical dimensions. Its dimensions are [B, 64, 128], where B is the batch size, 64 is the time step, and 128 is the number of LSTM hidden units (determined through hyperparameter optimization). This sequence encodes the complex temporal coupling patterns of multi-source sensor signals.

[0034] The inherent physical parameter set P: Preset inherent parameters of the rubber V-belt, determined based on a type B 1499mm V-belt, including: Geometric parameters: wedge angle Nodal height Working corner According to the test bench settings.

[0035] Material fatigue parameters: SN curve coefficients were determined by standard fatigue testing on the same batch of materials. The exponent m = 5.8 (these are typical values, and the specific values ​​may vary depending on the formula). These parameters are in vector form. Input provides an invariant physical benchmark for the learning process.

[0036] 2. Specialized design and working mechanism of Physical Mapping Networks (PMNs): PMN is a differentiable, trainable feedforward neural network designed to surpass general regression networks and become a dedicated physics engine for rubber fatigue.

[0037] Network structure: Input layer: The abstract state h_t (128-dimensional) at time t is concatenated with the physical parameter vector P (5-dimensional) to form a 133-dimensional joint input. This concatenation operation is the first step in achieving deep coupling between data and physics, forcing the network to understand data features in the context of physical priors.

[0038] Hidden layer: A fully connected layer with 128 neurons, using the ReLU activation function, responsible for nonlinear fusion and feature transformation of the joint input.

[0039] Output layer: A single linear output neuron, a direct scalar. This refers to the equivalent alternating stress amplitude. Its learning target (reference stress σ_ref) ranges from approximately 8 MPa to 25 MPa, corresponding to stress levels from normal to failure. This process is formally represented as: ,in These are the trainable parameters of PMN.

[0040] Specialized working mechanism: Physics-guided adaptive mapping: PMN learns the mapping function f from "abstract state + physical parameters" to "equivalent stress". This is compared with fixed mechanical formulas (such as...). Unlike rigidity calculations, PMN achieves adaptive approximation through its trainable weights. It can implicitly learn from data complex coupling effects that are difficult to describe by traditional formulas (e.g., how temperature softening of rubber modulus affects actual stress at the same slip ratio), thus combining consistency with physical laws with adaptability to complex working conditions.

[0041] End-to-end gradient-level fusion: PMN is not an isolated module. Its total loss function is... Among them, the L_physics (physical consistency loss) constraint By approximating reference values ​​provided by finite element simulations, the learning process is ensured to adhere to the laws of mechanics (addressing the credibility issue); the gradient of L_task (lifetime prediction Huber loss) propagates back through the entire network. This means that the parameter updates of PMN not only fit the stress but also directly contribute to improving the final lifetime prediction accuracy. This design achieves closed-loop collaborative optimization of "feature extraction → physical inversion → lifetime prediction," representing a deep unification of mechanism and data at the optimization level (addressing the accuracy issue).

[0042] 3. Damage calculation and feature reconstruction: generating interpretable physical state sequences. Based on PMN output Differentiable calculations are performed strictly following fatigue damage theory: Single-cycle damage increment: Where N(t) is the theoretical failure cycle number under the current stress.

[0043] Cumulative damage: This sequence monotonically increases from 0 (new) to 1 (failed), and is related to the remaining lifetime. Core state quantities with clear physical relationships.

[0044] Feature reconstruction: converting physical state sequences The (3D) state sequence is concatenated with the LSTM abstract state sequence H (128D) to form the physical-data deep fusion feature sequence F_fused, with dimensions [B, 64, 131]. This essentially aligns and enhances the "data-driven understanding state" with the "physical law derivation state", providing a highly discriminative input for the subsequent attention mechanism.

[0045] 4. Comparison of the effects brought by this module The technical approach implemented in this module is fundamentally different from existing V-belt life prediction methods.

[0046] Regarding the depth of utilization of physical knowledge, existing methods typically place physical formulas outside the model for independent calculation, with the calculation results serving only as static feature inputs; or they rely entirely on data-driven approaches, becoming unreliable "black boxes." This invention innovatively embeds physical parameters and laws into the computational graph within the model in the form of an embedded, trainable Physical Mapping Network (PMN), transforming physical knowledge from an "external constraint" into an "internal driving engine" capable of participating in the entire process of forward inference and backward propagation.

[0047] In modeling complex operating conditions, existing methods are limited by simplistic assumptions based on fixed formulas or the limitations of pure data models in generalizing to unknown patterns, making it difficult to accurately characterize the nonlinear effects of dynamic coupling of multiple factors such as slip rate, temperature, and tension. In stark contrast, the PMN of this invention can adaptively learn the mapping relationship between complex operating conditions and equivalent stress from data. Its model itself possesses the reasoning ability to understand and adapt to operating conditions it has not experienced, thereby achieving a qualitative improvement in robustness under varying operating conditions.

[0048] d) Dual attention mechanism module This step uses channel attention and temporal attention mechanisms to adaptively and finely control the physical-data fusion feature sequence.

[0049] (1) Channel Attention Implementation: The input feature sequence is subjected to global average pooling along the channel dimension to obtain a channel description vector. This vector is then input into a two-layer feedforward neural network (the first layer is a dimensionality reduction layer, and the second layer is a recovery layer), and the Sigmoid activation function is used to output the weight coefficient α (0 < α < 1) for each channel. Finally, the weight coefficient α is multiplied by the original features channel by channel. This mechanism enables the model to automatically distinguish and enhance the contribution of different sensor channels, such as slip rate and temperature, to the final lifetime prediction.

[0050] (2) Temporal Attention Implementation: For each time step in the input feature sequence, the feature vector is mapped to a scalar score through a fully connected layer. The scalar scores of all time steps are input into the Softmax function for normalization to obtain the importance weight β (0 < β < 1) for each time step. The weight β is multiplied by the feature vector of the corresponding time step. This mechanism enables the model to automatically focus on critical time periods where damage accumulation occurs abruptly or enters a phase of accelerated degradation, rather than simply averaging all historical information.

[0051] (3) Synergistic effect and verification results: The weights α and β are applied together to the feature sequence to achieve dynamic control over "focusing on key physical states during critical damage stages". To verify the necessity and superiority of this design, an ablation experiment was conducted: when only the channel attention mechanism was removed, the root mean square error (RMSE) of the model prediction increased by about 80% compared to the complete model; when only the temporal attention mechanism was removed, the RMSE increased by about 63%. This proves that the synergistic design of dual attention is indispensable and effective in improving prediction accuracy.

[0052] (4) Comparative Experimental Results: The model employing this dual attention mechanism significantly outperformed the baseline models that did not use attention or used single attention on the test set. Specifically, its prediction results showed a coefficient of determination (R²) of 0.962, a mean absolute percentage error (MAPE) of 2.8%, and a root mean square error (RMSE) of 8.7. In comparison, the traditional ARIMA model had an RMSE of 28.4 and the XGBoost model had an RMSE of 22.1 on the same test set.

[0053] e) Regression Output Layer The feature sequence, which has been weighted by dual attention, is converted into a fixed-length feature vector by global average pooling. Finally, a scalar, namely the predicted remaining lifetime (RUL) value, is output through a fully connected layer.

[0054] The specific implementation of data sample construction is as follows: A sliding window method is used to segment the standardized multi-source time series data. Based on the hyperparameter grid search results, to balance prediction accuracy and computational efficiency, this embodiment sets the window length (time step) to 64 and the sliding step size to 1. The 64 time steps of data within each window serve as the input features of a sample, and the actual remaining lifetime value at a certain time point after the end of the window is used as the prediction target for that sample, thereby constructing a supervised learning sample set.

[0055] S4. Model Training Details and Prediction Performance Verification a) Model training: The Adam optimizer (initial learning rate 0.001) was used, with batches X_batch and Y_batch as units, to minimize the total loss L_total, and training was performed for 100 epochs. The batch size B was set to 64.

[0056] b) Validation of Prediction Performance: To verify the effectiveness of the method, we validated it on a test set containing 100 V-band full lifecycle data points. We compared this method with ARIMA, XGBoost, a single LSTM model, a single CNN model, and the closest existing technology—a degradation model based on the Wiener process.

[0057] The results are as follows Figure 6 As shown, the predicted curve (solid line) of this invention best matches the actual degradation curve (dotted line). The quantitative evaluation indicators are shown in the table below: Model R² MAPE RMSE ARIMA 0.712 12.5% 28.4 Wiener process model 0.870 9.2% 18.5 XGBoost 0.805 8.7% 22.1 Single LSTM 0.853 6.1% 18.9 Single CNN 0.821 7.5% 20.5 This invention 0.962 2.8% 8.7 The results show that the present invention significantly outperforms all the comparison models in terms of R², MAPE and RMSE, achieving high-precision and high-stability remaining lifetime prediction.

[0058] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention, characterized in that: Includes the following steps: Step 1: Data acquisition and preprocessing. Real-time acquisition of multi-source time-series data during the operation of the rubber V-belt. The multi-source time-series data includes at least slip rate, tension and temperature. The multi-source time-series data is cleaned and standardized preprocessed. Step 2: Construct sample sequences. Use the sliding window method to segment the preprocessed multi-source time series data to construct a supervised learning sample set with time steps. Then, each sample in the sample set is input into a one-dimensional convolutional neural network to independently convolve the feature dimension at each time step, explicitly extracting the local coupling features between signals from different sensors at the same time. Then, the local features extracted by the CNN are concatenated with the original input features, and adaptive weights are generated through a learnable fully connected layer to achieve weighted fusion and form an enhanced feature representation. Step 3: Perform temporal dependency modeling, input the fused feature sequence into the long short-term memory network, and capture its long-term temporal evolution pattern; Step 4: Physical inversion and feature reconstruction of rubber belt damage increment. Based on the feature vector sequence representing the operating state of the equipment output by the LSTM network in Step 3, as well as the preset geometric parameters of the rubber V-belt structure and the fatigue parameters of the rubber material, the cumulative damage time series sequence inside the rubber material is calculated in real time through an embedded simplified physical model. The cumulative damage sequence is fused with the LSTM output feature sequence to generate a physical-data fusion feature sequence. Step 5: The physical-data fusion feature sequence output in Step 4 is subjected to temporal attention weighting and channel attention weighting in sequence to focus on key damage time periods and physical state signals that are crucial for life prediction. Step six involves integrating the feature sequences refined through dual attention, outputting the remaining life prediction value of the rubber V-belt through a fully connected regression layer, and performing end-to-end joint training using the Huber loss function and the Adam optimizer.

2. The rubber V-belt life prediction method integrating multi-dimensional features and dual attention as described in claim 1, characterized in that: The specific steps for data acquisition and preprocessing in step one are as follows: Step 1: Data cleaning and alignment. First, the raw data is inspected. For spike noise caused by transient sensor interference, it is identified using the 3-Sigma criterion and replaced using linear interpolation of the preceding and following data. For missing values ​​at fewer than 5 consecutive time points, linear interpolation is used to fill in the missing values. Steps one and two, data standardization: To eliminate the influence of dimensions, Z-score standardization is performed on the three feature sequences of slip rate, tension, and temperature after cleaning. That is, for each feature, its mean on the training set is calculated. ) and standard deviation ( Then, transform all the data (including the training and test sets): ; Step 13, Constructing supervised learning samples: The complete, standardized multidimensional time series data is divided into continuous samples using the sliding window method, and then all samples are divided into training set, validation set and test set.

3. The method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention as described in claim 1 or 2, characterized in that: The one-dimensional convolutional neural network used in step two is specifically as follows: it is a two-layer one-dimensional convolutional neural network; The first layer consists of 64 convolutional kernels of size (1,3) with ReLU activation. When the time dimension is 1 and the feature dimension is 3, the kernel performs convolution operations across the three features of slip rate, tension, and temperature at each independent time step, explicitly extracting instantaneous coupling patterns such as "high slip rate accompanied by high temperature". The second layer consists of 32 (1,3) convolutional kernels, which further abstract the features. The CNN output features are weighted and fused with the original input to form an enhanced feature sequence. .

4. The rubber V-belt life prediction method integrating multi-dimensional features and dual attention as described in claim 3, characterized in that: The specific method for modeling temporal dependencies in step three is as follows: The fused feature sequences are input into a two-layer stacked long short-term memory network. The optimal configuration determined by grid search is as follows: the number of hidden units in each LSTM layer is set to 128. Dropout layers and batch normalization layers are introduced between LSTM layers to capture the long-term dynamic law of V-belt performance degradation. The output of all time steps is used as the input of subsequent steps.

5. The method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention as described in claim 1 or 2, characterized in that: The specific steps of the physical inversion and feature reconstruction of the rubber strip damage increment in step four are as follows: Step 41 involves joint encoding of data abstraction and physical entities, including temporal state characteristics and an inherent physical parameter set P. The inherent physical parameter set P contains preset inherent parameters of the rubber V-belt, determined based on a type B 1499mm V-belt, including: Geometric parameters: wedge angle Nodal height Working corner According to the test bench settings Material fatigue parameters: SN curve coefficients were determined by standard fatigue testing on the same batch of materials. The exponent m = 5.8, and these parameters are in vector form. Input provides an invariant physical benchmark for the learning process; Step 42: Establish a physical mapping network, the structure of which includes: Input layer: The abstract state h_t at time t is concatenated with the physical parameter vector P to form a 133-dimensional joint input; Hidden layer: A fully connected layer with 128 neurons, using the ReLU activation function, responsible for non-linear fusion and feature transformation of the joint input; Output layer: A single linear output neuron, a direct scalar. This refers to the equivalent alternating stress amplitude, with a learning target range of approximately 8 MPa to 25 MPa, corresponding to stress levels from normal to failure. This process is formally represented as: ,in These are the trainable parameters of PMN.

6. Step 43: Damage calculation and feature reconstruction to generate an interpretable sequence of physical states.

7. The method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention as described in claim 5, characterized in that: The working mechanism of the physical mapping network is as follows: Physics-guided adaptive mapping: PMN learns the mapping function f from "abstract state + physical parameters" to "equivalent stress"; End-to-end gradient-level fusion: The total loss function of the physical mapping network is Where L_physics is the physical consistency loss, and the constraints are... Approximating the reference values ​​provided by finite element simulation ensures that the learning does not deviate from the laws of mechanics; the gradient of L_task, which is the lifetime prediction Huber loss, will propagate back through the entire network.

8. The method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention as described in claim 6, characterized in that: The damage calculation and feature reconstruction in step four-three are as follows: Cumulative damage: ; Feature reconstruction: converting physical state sequences The sequence is concatenated with the abstract state sequence H of LSTM to form the physical-data deep fusion feature sequence F_fused, with dimensions [B, 64, 131].

9. The method for predicting the lifespan of rubber V-belts by integrating multi-dimensional features and dual attention as described in claim 1 or 2, characterized in that: The specific methods for performing temporal attention weighting and channel attention weighting in step five are as follows: Channel attention implementation: Global average pooling is performed on the input feature sequence along the channel dimension to obtain a channel description vector. Then, this vector is input into a two-layer feedforward neural network, and the weight coefficient α of each channel is output using the Sigmoid activation function. Finally, the weight coefficient α is multiplied with the original features channel by channel. Temporal attention implementation: For the feature vector of each time step in the input feature sequence, it is mapped to a scalar score through a fully connected layer. The scalar scores of all time steps are input into the Softmax function for normalization to obtain the importance weight β of each time step. The weight β is then multiplied by the feature vector of the corresponding time step. Synergistic effect and verification results: By applying weights α and β together to the feature sequence, dynamic control of "focusing on key physical states during critical damage stages" can be achieved.

Images