Ship machinery bearing anti-noise type life prediction method fusing cross-modal attention mechanism

Abstract

The present application relates to the field of ship machinery bearing life prediction, and particularly relates to a ship machinery bearing anti-noise type life prediction method fusing a cross-modal attention mechanism. By constructing a mechanical bearing life prediction experiment platform, real-time simulation and collection of ship machinery bearing operation data are carried out, and multi-modal data are obtained covering normal operation, slight wear and severe failure conditions; a data preprocessing algorithm based on wavelet transform and dynamic threshold denoising is studied to suppress high-frequency noise in the marine environment; the collected multi-modal data are optimized and processed; and the processed multi-modal data are used for anti-noise life prediction of the ship machinery bearing.

Description

Technical Field

[0001] This invention relates to the field of marine mechanical bearing life prediction, and specifically to a noise-resistant life prediction method for marine mechanical bearings that integrates a cross-modal attention mechanism. Background Technology

[0002] Marine mechanical bearings, as core components of a ship's propulsion system, directly affect the ship's safety, reliability, and economy. In the complex marine environment, mechanical bearings endure harsh conditions such as high temperature, high humidity, salt spray corrosion, and multi-frequency vibrations, leading to accelerated failure rates and lifespan degradation. Traditional lifespan prediction models (such as shallow learning methods based on statistical regression or single sensor data) suffer from the following prominent problems: poor noise resistance: high-frequency noise in the complex marine environment (such as vibrations caused by wave surges and temperature fluctuations) severely interferes with bearing sensor data. Noise in the data can mask useful signal features, making it difficult for the model to accurately extract key information and reducing prediction accuracy.

[0003] Low interpretability: The "black box" nature of deep learning models makes it difficult for engineers to understand the physical mechanisms of the prediction results, resulting in insufficient interpretability. For example, it is impossible to locate key degradation features (such as bearing wear and lubrication failure), leading to a lack of basis for maintenance decisions.

[0004] Low efficiency of multimodal data fusion: The life decay of ship mechanical bearings involves multi-source heterogeneous data such as vibration, temperature and current. Existing methods are difficult to effectively integrate multimodal information, resulting in insufficient extraction and utilization of multi-scale features, which limits the generalization ability of the model. Summary of the Invention

[0005] The purpose of this invention is to propose a noise-resistant life prediction method for marine mechanical bearings that integrates cross-modal attention mechanisms, so as to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions.

[0006] This invention acquires multimodal data and introduces a cross-modal attention mechanism to fuse multi-source information, constructs a model, and designs an interpretability module, ultimately forming a closed-loop process for bearing life prediction. The interpretability module in the model includes quantifying noise in the multi-source information into bearing operating voltage and bearing operating angular velocity, and constructing two maximum normal operating ranges (positive and negative) using the two noise data. The normal operating state of the bearing at the operating moment is determined by the range, and the bearing life judgment is weighted according to abnormal changes in the bearing operating state.

[0007] A noise-resistant life prediction method for marine mechanical bearings integrating a cross-modal attention mechanism, the method comprising the following steps: S100: Data collection. By constructing a mechanical bearing life prediction experimental platform, the system simulates and collects real-time operating data of ship mechanical bearings, covering normal operation, minor wear, and severe failure conditions to obtain multimodal data. S200: Data preprocessing, researching data preprocessing algorithms based on wavelet transform and dynamic threshold denoising to suppress high-frequency noise in the marine environment; S300: Optimize the processing of the collected multimodal data; S400: Predicts the lifespan of marine mechanical bearings based on processed multimodal data to improve noise immunity.

[0008] Furthermore, in step S100, a multi-sensor network is deployed to collect mechanical bearing operation data in real time through the experimental platform, and at the same time, publicly available mechanical bearing life prediction datasets are retrieved and the datasets are used to establish a multimodal dataset.

[0009] Furthermore, in step S200, wavelet transform is used to perform preliminary noise reduction on the original signal, combined with dynamic thresholding; the algorithm further optimizes the data quality; the data is further normalized, and correlation analysis and PCA analysis are performed to establish a high-quality multimodal dataset.

[0010] Further, in step S300, the acquired multimodal data is classified to establish a high-quality multimodal dataset for definition. The bearing operation data is divided into bearing operating voltage and bearing operating angular velocity. The acquired noise data is defined as S_tll, and the noise data S_tll is defined as the influencing factors affecting the bearing operation data, including abnormal bearing operating voltage and abnormal bearing operating angular velocity. Abnormal bearing operating voltage is defined as S_tll_vol, and abnormal bearing operating angular velocity is defined as S_tll_ave. A noise influence value with an initial value of 0 is taken and compared with the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave. Let the initial value V0=0, and compare and calculate the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave within the period, Vp(vo)= Vp(av) = Where Vp(vo) and Vp(av) are the changes in operating voltage and rotational speed due to abnormal bearing operation, respectively, in the formula. This is expressed as the abnormal bearing operating voltage at time n within period T. This represents the abnormal bearing operating angular velocity at time n within a period T, where T is the period. This represents the average operating voltage of all abnormal bearings within period T. The maximum variable values ​​for normal bearing operation are calculated using Vp(vo) and Vp(av): Vps(Vp(vo)) = + *exp( ); Vps(Vp(av)) = + *exp( ); Vps is the maximum variable value during normal bearing operation, exp() is the logarithmic function with base e, min() is the minimum value function, and U is the average voltage of the bearing during normal operation. This represents the average angular velocity of the bearing during normal operation. This represents the normal operating voltage variation value for the bearing. The normal operating angular velocity fluctuation value of the bearing is used to construct two maximum normal operating ranges, positive and negative, based on two noise data. The normal operating state of the bearing at the time of operation is determined by the range, and the bearing life is judged by adding weight based on abnormal changes in the bearing operating state.

[0011] Preferably, the method in step S300 is based on a cross-modal attention mechanism. This mechanism integrates multi-source heterogeneous data related to vibration, temperature, and current, among others, concerning the lifespan degradation of marine machinery bearings, into bearing operating voltage and bearing operating angular velocity. When the bearing operating angular velocity is normal, the bearing operating voltage reflects the energy consumption of the bearing's operating state; when the bearing operating voltage is normal, the bearing operating angular velocity reflects the bearing wear. The attention mechanism retrieves the feedback from the multi-source heterogeneous data regarding the bearing operating voltage and bearing operating angular velocity. The weights of the corresponding multi-source heterogeneous data feedbacks and the bearing operating voltage and bearing operating angular velocity are normalized. A weighted average of the weights of the multi-source heterogeneous data is then applied to the corresponding Value to obtain the final attention output, which serves as the focus of the current task.

[0012] Furthermore, by adding weights to the bearing noise data, the bearing life prediction model is weighted accordingly. Sets [t1] and [t2] are constructed for the bearing operating range, where [t1] is the set of voltage values ​​within the maximum normal operating range and [t2] is the set of angular velocities within the maximum normal operating range. Data not in sets [t1] and [t2] are added to sets [t1`] and [t2`]. Set [t1`] is the set of abnormal bearing operating voltage values ​​and [t2`] is the set of abnormal bearing operating angular velocities. Sets [t1], [t2], [t1`, and [t2`] are arranged by time. The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated as P1 and P2. = , = The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated, H1 and H2. = , = The , , , , , , , For the k-th and (k+1)-th elements in the corresponding set, if the adjacent ratios P1, P2, H1, and H2 are positive and greater than or equal to 1, then the adjacent ratios are defined as positive weights; if the adjacent ratios P1, P2, H1, and H2 are positive and less than 1, then the adjacent ratios are defined as negative weights. The direction is determined by the positive and negative weights. In the data difference construction sequence Ck, the positive weight direction points from the k-th element to the (k+1)-th element, and the negative weight direction points from the (k+1)-th element to the k-th element. Using the directions of normal operating voltage and normal operating angular velocity as a reference, the adjacent ratios P1, P2, H1, and H2 are assigned directions, and the weight vectors of the adjacent ratios H1 and H2 are calculated. , , , in, , , , These are the weight vectors for adjacent ratios P1, P2, H1, and H2, respectively. , , , , respectively, are the vector values ​​of adjacent ratios P1, P2, H1, and H2 after assigning directions to adjacent ratios, and m and g are the sum of voltage and angular velocity values ​​obtained throughout the entire cycle.

[0013] Furthermore, in step S400, a noise data comparison is established to optimize the bearing life superposition comparison analysis of the prediction model; The network optimizes the comparison of noisy data by using positive and negative weight directions. The input to the neural network is the comparison of noisy data, and the network output is determined by both the positive and negative outputs. The calculation formula is as follows: ; ; ; ; Where x represents the input value, and the input value is the sequence Ck constructed from the data differences. These represent the directions of the positive and negative weights, respectively. This is the positive output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the forward output, respectively. The input layer weight matrix is ​​constructed using adjacent ratios P1, P2, H1, and H2. The reverse output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the reverse output, respectively. The input layer weights are obtained by comparing noise data when determining whether to use a forward or reverse output. , For input layer weights, , Here are the hidden layer weights, and the bias vector is... , and The sum, if obtained through a comparison at the second-to-last level, then , For input layer weights, , The hidden layer weights are given by the bias vector. , and The sum of; For the final output value, and These are the output layer weight matrix and bias vector, respectively, with symbols... For concatenation operations; Final output value The data is fed into a neural network model for training. After training is completed, a fast indicator X is output for the input data. The indicator X is used to judge whether the comparison data meets the qualified standard. Each indicator X needs to be trained through a separate neural network model to establish the mapping relationship between the input data and the indicator. In this neural network model, the number of neurons in the input layer is equal to the number of input data types, and the number of neurons in the output layer is equal to the number of output data types. The number of intermediate layers and the number of neurons in each layer are freely set. The configuration of the intermediate layers directly affects the accuracy of the neural network's approximate data. The decision to increase or decrease the number of intermediate layers or neurons is based on the training results. The method for calculating the number of intermediate layers and neurons is as follows: The intermediate layers in the neural network model are added layer by layer, and the output index X corresponding to the intermediate layer number is named. When the intermediate layer number is 1, the output index is... And so on, and based on the index X, a nonlinear regression function is constructed, f(X) = Let a be a constant value. Calculate the variance D of the independent variable of the function, D = Where L is the number of intermediate layers in the output. For the i-th index X in the output, the For the input data corresponding to the i-th output index X, input the output index X into the neural network model, increment the intermediate layer number by 1, and output index X. Update the weighting coefficients of neurons using gradient descent: (l) = - ; in, ( ) is used to find the partial derivative of the function, where P and Q are the system output error and the neuron weight increment, respectively, both being constant values. (l) represents the weighting coefficients of the neurons. (l) represents the updated neuron weighting coefficients. Indicates the neuron's learning rate, through the (l) Determine the number of neurons to be set in the intermediate layer, and combine the superposition analysis algorithm with the output value. The algorithm was subjected to deep learning, and comparative analysis data on the impact of noise data on bearing life was output, ensuring the quality of the comparative data.

[0014] The beneficial effects of this invention are as follows: by judging the impact of noise data on the normal operation of the bearing, a new weight ratio is assigned to the bearing life prediction model. Through this method, the multi-source impacts on the operation of ship bearings on the sea surface can be predicted more accurately for the life of ship bearings through a cross-modal attention mechanism, and the weight allocation of multi-source data is also more reasonable. Attached Figure Description

[0015] The above and other features of the present invention will become more apparent from the detailed description of the embodiments shown in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The diagram shows the overall flowchart of a noise-resistant life prediction method for marine mechanical bearings that integrates a cross-modal attention mechanism. Detailed Implementation

[0016] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0017] The vibration signals of rolling bearings in ship shafting systems are characterized by strong noise and non-stationarity, making it difficult to predict bearing life based on the obtained vibration signals. Furthermore, other high-frequency noises in the complex marine environment (such as vibrations caused by wave surges and temperature fluctuations) severely interfere with bearing sensor data. Noise in the data can mask useful signal features, making it difficult for models to accurately extract key information and reducing prediction accuracy. To address these problems, the following solutions are proposed: according to Figure 1 As shown, a noise-resistant life prediction method for marine mechanical bearings integrating a cross-modal attention mechanism is presented. The method includes the following steps: S100: Data collection. By constructing a mechanical bearing life prediction experimental platform, the system simulates and collects real-time operating data of ship mechanical bearings, covering normal operation, minor wear, and severe failure conditions to obtain multimodal data. S200: Data preprocessing, researching data preprocessing algorithms based on wavelet transform and dynamic threshold denoising to suppress high-frequency noise in the marine environment; S300: Optimize the processing of the collected multimodal data; S400: Predicts the lifespan of marine mechanical bearings based on processed multimodal data to improve noise immunity.

[0018] Furthermore, in step S100, a multi-sensor network is deployed to collect mechanical bearing operation data in real time through the experimental platform, and at the same time, publicly available mechanical bearing life prediction datasets are retrieved and the datasets are used to establish a multimodal dataset.

[0019] Furthermore, in step S200, wavelet transform is used to perform preliminary noise reduction on the original signal, combined with dynamic thresholding; the algorithm further optimizes the data quality; the data is further normalized, and correlation analysis and PCA analysis are performed to establish a high-quality multimodal dataset.

[0020] Further, in step S300, the acquired multimodal data is classified to establish a high-quality multimodal dataset for definition. The bearing operation data is divided into bearing operating voltage and bearing operating angular velocity. The acquired noise data is defined as S_tll, and the noise data S_tll is defined as the influencing factors affecting the bearing operation data, including abnormal bearing operating voltage and abnormal bearing operating angular velocity. Abnormal bearing operating voltage is defined as S_tll_vol, and abnormal bearing operating angular velocity is defined as S_tll_ave. A noise influence value with an initial value of 0 is taken and compared with the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave. Let the initial value V0=0, and compare and calculate the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave within the period, Vp(vo)= Vp(av) = Where Vp(vo) and Vp(av) are the changes in operating voltage and rotational speed due to abnormal bearing operation, respectively, in the formula. This is expressed as the abnormal bearing operating voltage at time n within period T. This represents the abnormal bearing operating angular velocity at time n within a period T, where T is the period. This represents the average operating voltage of all abnormal bearings within period T. The maximum variable values ​​for normal bearing operation are calculated using Vp(vo) and Vp(av): Vps(Vp(vo)) = + *exp( ); Vps(Vp(av)) = + *exp( ); Vps is the maximum variable value during normal bearing operation, exp() is the logarithmic function with base e, min() is the minimum value function, and U is the average voltage of the bearing during normal operation. This represents the average angular velocity of the bearing during normal operation. This represents the normal operating voltage variation value for the bearing. The normal operating angular velocity fluctuation value of the bearing is used to construct two maximum normal operating ranges, positive and negative, based on two noise data. The normal operating state of the bearing at the time of operation is determined by the range, and the bearing life is judged by adding weight based on abnormal changes in the bearing operating state.

[0021] Preferably, the above scheme can be replaced by the following scheme: Based on cross-modal attention, multimodal data information is effectively utilized through several different single-head cross-modal attention and learnable linear projection; the multimodal data is transformed into attention weight vectors Q for two modes: bearing operating voltage and bearing operating angular velocity. The vector Q includes K and V, where K is the attention weight vector for bearing operating voltage and V is the attention weight vector for bearing operating angular velocity. It is divided into H different subspaces. The temporal information is transformed from information mode to what can be regarded as image mode and then transferred to its corresponding image mode to be transformed into text mode. The cross-modal self-attention function is obtained by concatenating the outputs of all single-head cross-modal attention together and performing linear transformation through a fully connected layer to form a cross-modal attention function. The attention weights of bearing operating voltage and bearing operating angular velocity are calculated and re-allocated through the cross-modal attention function.

[0022] The beneficial effects of this step are as follows: by fusing external multimodal data with a cross-modal attention mechanism, the multimodal data is transformed into feedback on the voltage and angular velocity of the bearing during operation, while a new weighting ratio is applied to the bearing life prediction, making the data source for bearing life prediction more refined.

[0023] Furthermore, by adding weights to the bearing noise data, the bearing life prediction model is weighted accordingly. Sets [t1] and [t2] are constructed for the bearing operating range, where [t1] is the set of voltage values ​​within the maximum normal operating range and [t2] is the set of angular velocities within the maximum normal operating range. Data not in sets [t1] and [t2] are added to sets [t1`] and [t2`]. Set [t1`] is the set of abnormal bearing operating voltage values ​​and [t2`] is the set of abnormal bearing operating angular velocities. Sets [t1], [t2], [t1`, and [t2`] are arranged by time. The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated as P1 and P2. = , = The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated, H1 and H2. = , = The , , , , , , , For the k-th and (k+1)-th elements in the corresponding set, if the adjacent ratios P1, P2, H1, and H2 are positive and greater than or equal to 1, then the adjacent ratios are defined as positive weights; if the adjacent ratios P1, P2, H1, and H2 are positive and less than 1, then the adjacent ratios are defined as negative weights. The direction is determined by the positive and negative weights. In the data difference construction sequence Ck, the positive weight direction points from the k-th element to the (k+1)-th element, and the negative weight direction points from the (k+1)-th element to the k-th element. Using the directions of normal operating voltage and normal operating angular velocity as a reference, the adjacent ratios P1, P2, H1, and H2 are assigned directions, and the weight vectors of the adjacent ratios H1 and H2 are calculated. , , , in, , , , These are the weight vectors for adjacent ratios P1, P2, H1, and H2, respectively. , , , , respectively, are the vector values ​​of adjacent ratios P1, P2, H1, and H2 after assigning directions to adjacent ratios, and m and g are the sum of voltage and angular velocity values ​​obtained throughout the entire cycle.

[0024] Furthermore, in step S400, a noise data comparison is established to optimize the bearing life superposition comparison analysis of the prediction model; The network optimizes the comparison of noisy data by using positive and negative weight directions. The input to the neural network is the comparison of noisy data, and the network output is determined by both the positive and negative outputs. The calculation formula is as follows: ; ; ; ; Where x represents the input value, and the input value is the sequence Ck constructed from the data differences. These represent the directions of the positive and negative weights, respectively. This is the positive output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the forward output, respectively. The input layer weight matrix is ​​constructed using adjacent ratios P1, P2, H1, and H2. The reverse output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the reverse output, respectively. The input layer weights are obtained by comparing noise data when determining whether to use a forward or reverse output. , For input layer weights, , The hidden layer weights are given by the bias vector. , and The sum, if obtained through a comparison at the second-to-last level, then , For input layer weights, , The hidden layer weights are given by the bias vector. , and The sum of; For the final output value, and These are the output layer weight matrix and bias vector, respectively, with symbols... For concatenation operations; Final output value The data is fed into a neural network model for training. After training is completed, a fast indicator X is output for the input data. The indicator X is used to judge whether the comparison data meets the qualified standard. Each indicator X needs to be trained through a separate neural network model to establish the mapping relationship between the input data and the indicator. In this neural network model, the number of neurons in the input layer is equal to the number of input data types, and the number of neurons in the output layer is equal to the number of output data types. The number of intermediate layers and the number of neurons in each layer are freely set. The configuration of the intermediate layers directly affects the accuracy of the neural network's approximate data. The decision to increase or decrease the number of intermediate layers or neurons is based on the training results. The method for calculating the number of intermediate layers and neurons is as follows: The intermediate layers in the neural network model are added layer by layer, and the output index X corresponding to the intermediate layer number is named. When the intermediate layer number is 1, the output index is... And so on, and based on the index X, a nonlinear regression function is constructed, f(X) = Let a be a constant value. Calculate the variance D of the independent variable of the function, D = Where L is the number of intermediate layers in the output. For the i-th index X in the output, the For the input data corresponding to the i-th output index X, input the output index X into the neural network model, increment the intermediate layer number by 1, and output index X. Update the weighting coefficients of neurons using gradient descent: (l) = - ; in, ( ) is used to find the partial derivative of the function, where P and Q are the system output error and the neuron weight increment, respectively, both being constant values. (l) represents the weighting coefficients of the neurons. (l) represents the updated neuron weighting coefficients. Indicates the neuron's learning rate, through the (l) Determine the number of neurons to be set in the intermediate layer, and combine the superposition analysis algorithm with the output value. The algorithm was subjected to deep learning, and comparative analysis data on the impact of noise data on bearing life was output, ensuring the quality of the comparative data.

[0025] Preferably, model training and optimization Let θ be the training parameters of the model, then the mathematical expression of the model can be simplified to y=f(x,θ). However, θ is affected by the network architecture of the model, and the network's ability to extract data features is severely affected by hyperparameters such as convolution kernel size, learning rate, and batch size, so the model needs to be trained iteratively through a two-layer nested algorithm.

[0026] During training, the outer layer hyperparameters are optimized using the Hyperband algorithm. The mathematical expression for the optimization process is: f(x, )=Arg (x∈ ); The inner layer parameters are trained using the Adam algorithm, and the mathematical expression for the optimization process is as follows: =Arg (x∈ ) in, It is the hyperparameter space. yes A set of hyperparameters in ; For parameter space, Hyperparameters The corresponding model parameters, ; and These are the training set and the validation set, respectively. and These are the training loss and validation loss functions, respectively.

[0027] Preferably, a trained prediction model is used to identify the dynamic characteristics of bearing health status in relation to remaining service time, and the bearing health status is predicted based on the trend of historical health status scores, thus completing the prediction of bearing life.

[0028] By training the model through the above steps, the prediction model can identify the dynamic characteristics of bearing health status on usage time and predict the subsequent bearing health status based on the trend of historical health status scores, that is, the estimation of the bearing's usable duration.

[0029] Based on historical operating data of ship bearings, the maintenance cycle of ship bearings is obtained, and the current operating conditions of ship bearings are extracted to obtain life-related data. Based on the aforementioned lifespan correlation data and health status index H, where the lifespan correlation data is derived data obtained from the historical operating data and maintenance cycle of the ship bearings, and the health status index H is the current health status of the ship bearings output after neural network learning of the ship's health status using the above method, expressed by the formula: L = ; Calculate the remaining life of the ship's bearings to obtain the quantitative results of the remaining life; Where L represents the remaining lifetime. It is the theoretical maximum lifespan. It is the sensitivity adjustment coefficient for operating conditions. It is the maintenance cycle impact coefficient. is the weighting coefficient, e is the base of the natural logarithm, and H represents the health status index.

[0030] Although the invention has been described in considerable detail and particularly with regard to several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

Claims

1. A noise-resistant life prediction method for marine mechanical bearings integrating cross-modal attention mechanisms, characterized in that, The method includes the following steps: S100: Data collection. By constructing a mechanical bearing life prediction experimental platform, the system simulates and collects real-time operating data of ship mechanical bearings, covering normal operation, minor wear, and severe failure conditions to obtain multimodal data. S200: Data preprocessing, researching data preprocessing algorithms based on wavelet transform and dynamic threshold denoising to suppress high-frequency noise in the marine environment; S300: Optimize the processing of the collected multimodal data; S400: Predicts the lifespan of marine mechanical bearings using processed multimodal data.

2. The method for predicting the noise-resistant lifespan of marine mechanical bearings based on a cross-modal attention mechanism according to claim 1, characterized in that, In step S100, a multi-sensor network is deployed to collect mechanical bearing operation data in real time through the experimental platform. At the same time, publicly available datasets for predicting the lifespan of mechanical bearings are retrieved and the datasets are used to establish a multimodal dataset.

3. The method for predicting the noise-resistant lifespan of marine mechanical bearings based on a cross-modal attention mechanism according to claim 1, characterized in that, In step S200, wavelet transform is used to perform preliminary noise reduction on the original signal, combined with dynamic thresholding; the algorithm further optimizes the data quality; the data is further normalized, and correlation analysis and PCA analysis are performed to establish a high-quality multimodal dataset.

4. The method for predicting the noise-resistant lifespan of marine mechanical bearings based on a cross-modal attention mechanism according to claim 1, characterized in that, In step S300, the acquired multimodal data is classified to establish a high-quality multimodal dataset for definition. The bearing operation data is divided into bearing operating voltage and bearing operating angular velocity. The acquired noise data is defined as S_tll, and the noise data S_tll is defined as the influencing factors affecting the bearing operation data, including abnormal bearing operating voltage and abnormal bearing operating angular velocity. Abnormal bearing operating voltage is defined as S_tll_vol, and abnormal bearing operating angular velocity is defined as S_tll_ave. A noise influence value with an initial value of 0 is used to compare the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave. Let the initial value V0=0, and calculate the comparison between the abnormal bearing operating voltage S_tll_vol and the abnormal bearing operating angular velocity S_tll_ave within the period, Vp(vo) = Vp(av) = Where Vp(vo) and Vp(av) are the changes in operating voltage and rotational speed due to abnormal bearing operation, respectively, in the formula. This is expressed as the abnormal bearing operating voltage at time n within period T. This represents the abnormal bearing operating angular velocity at time n within a period T, where T is the period. This represents the average operating voltage of all abnormal bearings within period T. The maximum variable values ​​for normal bearing operation are calculated using Vp(vo) and Vp(av): Vps(Vp(vo))= + *exp( ); Vps (Vp (off)) = + *exp( ); Vps is the maximum variable value during normal bearing operation, exp() is the logarithmic function with base e, min() is the minimum value function, and U is the average voltage of the bearing during normal operation. This represents the average angular velocity of the bearing during normal operation. This represents the normal operating voltage variation value for the bearing. The normal operating angular velocity fluctuation value of the bearing is used to construct two maximum normal operating ranges, positive and negative, based on two noise data. The normal operating state of the bearing at the time of operation is determined by the range, and the bearing life is judged by adding weight based on abnormal changes in the bearing operating state.

5. The method for predicting the noise-resistant lifespan of marine mechanical bearings based on a cross-modal attention mechanism according to claim 4, characterized in that, By adding weights to the bearing noise data, the bearing life prediction model is weighted accordingly. Sets [t1] and [t2] are constructed for the bearing operating range, where [t1] is the set of voltage values ​​within the maximum normal operating range and [t2] is the set of angular velocities within the maximum normal operating range. Data not in sets [t1] and [t2] are added to sets [t1`] and [t2`]. Set [t1`] is the set of abnormal bearing operating voltage values ​​and [t2`] is the set of abnormal bearing operating angular velocities. Sets [t1], [t2], [t1`, and [t2`] are arranged by time. The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated as P1 and P2. = , = The adjacent differences of set [t1] and the adjacent ratios of adjacent differences of set [t2] are calculated, H1 and H2. = , = The , , , , , , , For the k-th and (k+1)-th elements in the corresponding set, if the adjacent ratios P1, P2, H1, and H2 are positive and greater than or equal to 1, then the adjacent ratios are defined as positive weights; if the adjacent ratios P1, P2, H1, and H2 are positive and less than 1, then the adjacent ratios are defined as negative weights. The direction is determined by the positive and negative weights. In the data difference construction sequence Ck, the positive weight direction points from the k-th element to the (k+1)-th element, and the negative weight direction points from the (k+1)-th element to the k-th element. Using the directions of normal operating voltage and normal operating angular velocity as a reference, the adjacent ratios P1, P2, H1, and H2 are assigned directions, and the weight vectors of the adjacent ratios H1 and H2 are calculated. , , , in, , , , These are the weight vectors for adjacent ratios P1, P2, H1, and H2, respectively. , , , , respectively, are the vector values ​​of adjacent ratios P1, P2, H1, and H2 after assigning directions to adjacent ratios, and m and g are the sum of voltage and angular velocity values ​​obtained throughout the entire cycle.

6. The method for predicting the noise-resistant lifespan of marine mechanical bearings based on a cross-modal attention mechanism according to claim 1, characterized in that, In step S400, a noise data comparison is established to optimize the bearing life superposition comparison analysis of the prediction model; The network optimizes the comparison of noisy data by using positive and negative weight directions. The input to the neural network is the comparison of noisy data, and the network output is determined by both the positive and negative outputs. The calculation formula is as follows: ； ； ； ； Where x represents the input value, and the input value is the sequence Ck constructed from the data differences. These represent the directions of the positive and negative weights, respectively. This is the positive output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the forward output, respectively. The input layer weight matrix is ​​constructed using adjacent ratios P1, P2, H1, and H2. The reverse output value of the network. , , These are the input layer weights, hidden layer weights, and bias vector for the reverse output, respectively. The input layer weights are obtained by comparing noise data when determining whether to use a forward or reverse output. , For input layer weights, , The hidden layer weights are given by the bias vector. , and The sum, if obtained through a comparison at the second-to-last level, then , For input layer weights, , The hidden layer weights are given by the bias vector. , and The sum of; For the final output value, and These are the output layer weight matrix and bias vector, respectively, with symbols... For concatenation operations; Final output value The data is fed into a neural network model for training. After training is completed, a fast indicator X is output for the input data. The indicator X is used to judge whether the comparison data meets the qualified standard. Each indicator X needs to be trained through a separate neural network model to establish the mapping relationship between the input data and the indicator. In this neural network model, the number of neurons in the input layer is equal to the number of input data types, and the number of neurons in the output layer is equal to the number of output data types. The number of intermediate layers and the number of neurons in each layer are freely set. The configuration of the intermediate layers directly affects the accuracy of the neural network's approximate data. The decision to increase or decrease the number of intermediate layers or neurons is based on the training results. The method for calculating the number of intermediate layers and neurons is as follows: The intermediate layers in the neural network model are added layer by layer, and the output index X corresponding to the intermediate layer number is named. When the intermediate layer number is 1, the output index is... And so on, and based on the index X, a nonlinear regression function is constructed, f(X) = Let a be a constant value. Calculate the variance D of the independent variable of the function, D = Where L is the number of intermediate layers in the output. For the i-th index X in the output, the For the input data corresponding to the i-th output index X, input the output index X into the neural network model, increment the intermediate layer number by 1, and output index X. Update the weighting coefficients of neurons using gradient descent: （l）=- ； in, ( ) is used to find the partial derivative of the function, where P and Q are the system output error and the neuron weight increment, respectively, both being constant values. (l) represents the weighting coefficients of the neurons. (l) represents the updated neuron weighting coefficients. Indicates the neuron's learning rate, through the (l) Determine the number of neurons to be set in the intermediate layer, and combine the superposition analysis algorithm with the output value. The algorithm was subjected to deep learning, and comparative analysis data on the impact of noise data on bearing life was output, ensuring the quality of the comparative data.

Images