Multi-modal data modeling method and device for rotating part degradation prediction

Abstract

The invention discloses a multi-modal data modeling method and a multi-modal data modeling device for rotating part degradation prediction, which are used for solving the technical problem of poor model prediction accuracy caused by the existing rotating part degradation prediction modeling method. The method comprises the following steps: acquiring to-be-trained initial multi-modal signal data, and preprocessing the to-be-trained initial multi-modal signal data based on a preset signal processing algorithm to generate to-be-trained target multi-modal signal data; adopting a preset distribution function to calculate a degradation index according to the to-be-trained target multi-mode signal data; and performing model training on the initial mechanical rotating part degradation prediction model by adopting the to-be-trained target multi-mode signal data and the degradation index, and determining a target mechanical rotating part degradation prediction model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multimodal data modeling method and device for predicting degradation of rotating parts. Background Art

[0002] Mechanical rotating parts are key components of complex discrete manufacturing equipment, and their safe and stable operation is crucial. However, with the long-term operation of complex equipment, mechanical parts will inevitably experience degradation such as wear and fatigue. If there is no real-time monitoring and timely intervention, it will seriously affect the reliability and service life of the equipment. Therefore, how to effectively monitor the health status of key components and perform maintenance or replacement in advance has become the key to improving production efficiency and reducing unplanned downtime.

[0003] In recent years, with the development of sensing technology, the Internet of Things, big data analysis, and artificial intelligence technology, the condition monitoring and fault diagnosis of mechanical systems have shifted from the traditional periodic inspection mode to condition-based predictive maintenance. The core of this mode is to accurately predict the degradation process of mechanical components so that maintenance and remedial measures can be taken before the components fail completely, thereby effectively avoiding unexpected downtime caused by sudden failures and ensuring the safe and stable operation of equipment.

[0004] Most of the existing rotating component degradation prediction modeling methods use data from a single sensor to complete the model training. However, in practical applications, the degradation process of mechanical components is often the result of the combined action of multiple factors. A single type of sensor data is difficult to fully capture the degradation trend of components under complex working conditions, resulting in poor accuracy of model prediction. Summary of the invention

[0005] The present invention provides a multimodal data modeling method and device for rotating component degradation prediction, which are used to solve the technical problem that the existing rotating component degradation prediction modeling method leads to poor accuracy of model prediction.

[0006] A first aspect of the present invention provides a multimodal data modeling method for predicting degradation of a rotating component, comprising:

[0007] Acquire initial multimodal signal data to be trained, and preprocess the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained;

[0008] Calculating a degradation index based on the target multimodal signal data to be trained using a preset distribution function;

[0009] The target multimodal signal data to be trained and the degradation index are used to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model.

[0010] Optionally, the preset signal processing algorithm includes an improved empirical mode decomposition algorithm and a wavelet packet algorithm; the preprocessing of the initial multimodal signal data to be trained based on the preset signal processing algorithm to generate target multimodal signal data to be trained includes:

[0011] Adding white noise to the initial multimodal signal data to be trained to generate noisy signal data to be trained;

[0012] Decomposing the noisy signal data to be trained by using an improved empirical mode decomposition algorithm to output a plurality of multimodal sub-signal data;

[0013] Using a wavelet packet algorithm to perform noise reduction on each of the multimodal sub-signal data to generate a plurality of multimodal noise-reduced sub-signal data;

[0014] The plurality of multimodal denoising sub-signal data are reconstructed based on a preset adaptive threshold, and target multimodal signal data to be trained is output.

[0015] Optionally, the preset distribution function includes an improved gamma distribution probability density function and a cumulative distribution function; the step of using the preset distribution function to calculate the degradation index according to the target multimodal signal data to be trained includes:

[0016] Determining a gamma probability density distribution according to the continuous degradation amount in the target multimodal signal data to be trained using an improved gamma distribution probability density function;

[0017] Substituting the continuous degradation amount in the target multimodal signal data to be trained into the cumulative distribution function, and outputting the joint failure risk probability;

[0018] Based on the gamma probability density distribution, determining the current moment entropy value corresponding to the target multimodal signal data to be trained;

[0019] Performing a difference operation on the current moment entropy value and the historical moment entropy value corresponding to the target multimodal signal data to be trained, to determine the difference between the entropy values ​​corresponding to the target multimodal signal data to be trained;

[0020] A degradation index is calculated according to the difference between the entropy values ​​and the joint failure risk probability.

[0021] Optionally, the adopting the target multimodal signal data to be trained and the degradation index to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model includes:

[0022] Inputting the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction, and generating a predicted degradation result of the mechanical rotating component to be trained;

[0023] Taking the degradation index as the real degradation result of the mechanical rotating component, and determining the model gradient according to the real degradation result of the mechanical rotating component and the predicted degradation result of the mechanical rotating component to be trained;

[0024] Using the model gradient to update the model parameters of the initial mechanical rotating component degradation prediction model to determine an intermediate mechanical rotating component degradation prediction model;

[0025] Determine a target loss value by using the intermediate mechanical rotating component degradation prediction model and a preset root mean square error loss function according to the target multimodal signal data to be trained and the actual degradation result of the mechanical rotating component;

[0026] Determining whether the target loss value converges;

[0027] If so, the intermediate mechanical rotating component degradation prediction model is used as the trained target mechanical rotating component degradation prediction model.

[0028] Optionally, the initial mechanical rotating component degradation prediction model includes an embedding layer, a plurality of Mamba blocks and a linear output layer; the step of inputting the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction to generate a predicted degradation result of the mechanical rotating component to be trained includes:

[0029] Mapping the target multimodal signal data to be trained by using a time vector algorithm through an embedding layer to generate multimodal mapping signal data;

[0030] Using a plurality of Mamba blocks to perform efficient sequence processing on the multimodal mapping signal data to generate target degradation features;

[0031] The target degradation feature is used as the input of the linear output layer, and the predicted degradation result of the mechanical rotating part to be trained is output.

[0032] Optionally, the Mamba block includes a first linear layer, a second linear layer, a third linear layer, a SiLU activation function layer, a dimensional channel attention module, and a state space module; the data processing process of the Mamba block includes:

[0033] Performing linear transformation on the input multimodal signal data input to the Mamba block through a first linear layer and a second linear layer respectively, to generate a first multimodal transformation signal feature and a second multimodal transformation signal feature;

[0034] Using a SiLU activation function layer to perform nonlinear mapping on the first multimodal transformation signal feature, and outputting a multimodal nonlinear signal feature;

[0035] Capturing cross-dimensional channel information of the second multimodal transformation signal feature through a dimensional channel attention module to generate a multimodal attention signal feature;

[0036] multiplying the multimodal attention signal feature and the second multimodal transformation signal feature to output a first multimodal multiplied signal feature;

[0037] Using a state space module to extract features of the first multimodal multiplication signal to generate signal degradation features;

[0038] multiplying the signal degradation feature and the multimodal nonlinear signal feature to output a second multimodal multiplied signal feature;

[0039] A third linear layer is used to perform a linear transformation on the second multimodal multiplication signal feature to generate an output signal feature.

[0040] Optionally, it also includes:

[0041] When receiving the initial multimodal signal data to be detected, preprocessing the initial multimodal signal data to be detected based on a preset signal processing algorithm to generate target multimodal signal data to be detected;

[0042] The target mechanical rotating component degradation prediction model is used to predict the target multimodal signal data to be detected, and a target mechanical rotating component prediction degradation result is generated.

[0043] A second aspect of the present invention provides a multimodal data modeling device for predicting degradation of a rotating component, comprising:

[0044] An acquisition module, used for acquiring initial multimodal signal data to be trained, and preprocessing the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained;

[0045] A calculation module, used to calculate a degradation index according to the target multimodal signal data to be trained by using a preset distribution function;

[0046] The training module is used to perform model training on the initial mechanical rotating component degradation prediction model using the target multimodal signal data to be trained and the degradation index to determine the target mechanical rotating component degradation prediction model.

[0047] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the multimodal data modeling method for predicting degradation of rotating components as described in any one of the above items.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the multimodal data modeling method for predicting degradation of rotating components as described in any one of the above items.

[0049] It can be seen from the above technical solutions that the present invention has the following advantages:

[0050] The above technical scheme of the present invention provides a multimodal data modeling method for rotating component degradation prediction. First, initial multimodal signal data to be trained is obtained, and based on a preset signal processing algorithm, the initial multimodal signal data to be trained is preprocessed to generate target multimodal signal data to be trained; then, a preset distribution function is used to calculate the degradation index according to the target multimodal signal data to be trained; finally, the target multimodal signal data to be trained and the degradation index are used to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model; based on the above scheme, the present invention uses the preprocessed target multimodal signal data to be trained, and combines the degradation index output by the preset distribution function to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model. The process can overcome the limitations of traditional prediction methods based on single-modal data, provide more comprehensive degradation information for predictive modeling, and thus improve the accuracy of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 A flowchart of the steps of a multimodal data modeling method for predicting degradation of rotating components provided in the first embodiment of the present invention;

[0053] Figure 2 This is an example diagram of a framework for multimodal data modeling for rotating component degradation prediction provided in the first embodiment of the present invention;

[0054] Figure 3A schematic diagram of the execution process of the multimodal data modeling method for rotating component degradation prediction provided in the second embodiment of the present invention;

[0055] Figure 4 A flowchart of the steps of model prediction of a target mechanical rotating component degradation prediction model provided in the third embodiment of the present invention;

[0056] Figure 5 This is a structural block diagram of a multimodal data modeling device for rotating component degradation prediction provided in Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0057] The embodiments of the present invention provide a multimodal data modeling method and device for rotating component degradation prediction, which are used to solve the technical problem that the existing rotating component degradation prediction modeling method leads to poor model prediction accuracy.

[0058] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] See also Figure 1 , Figure 1 A flowchart of the steps of a multimodal data modeling method for rotating component degradation prediction provided in Embodiment 1 of the present invention.

[0060] The present invention provides a multimodal data modeling method for predicting degradation of rotating parts, comprising:

[0061] Step 101: Acquire initial multimodal signal data to be trained, and pre-process the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained.

[0062] The preset signal processing algorithms include improved empirical mode decomposition algorithm and wavelet packet algorithm.

[0063] The initial multimodal signal data to be trained is the multimodal signal data used for model training. The multimodal signal data includes various signal data throughout the life cycle of the component, such as temperature, acoustic emission, vibration, etc.

[0064] It should be noted that, under complex working conditions, sensor signals such as vibration or acoustic emission (multimodal signal data) will collect a large amount of noise signals due to their high-frequency sampling characteristics. In order to improve the signal-to-noise ratio of the vibration signal, it is necessary to perform denoising preprocessing; among them, the improved empirical mode decomposition (EEMD, Ensemble Empirical Mode Decomposition) combined with wavelet threshold denoising algorithm is a method commonly used in the field of signal processing, which can effectively remove the noise components in the signal and avoid the problem of transition denoising. Therefore, the present invention adopts EEMD combined with wavelet threshold denoising method to perform denoising preprocessing on the signal.

[0065] Specifically, based on the preset signal processing algorithm, the process of preprocessing the initial multimodal signal data to be trained to generate the target multimodal signal data to be trained can be achieved by executing the following sub-steps S11 to S14:

[0066] Step S11, adding white noise to the initial multimodal signal data to be trained to generate noisy signal data to be trained;

[0067] It should be noted that the auxiliary white noise signal is added to the original signal x(t) (i.e., the initial multimodal signal data to be trained) to obtain the noisy signal data to be trained X n (t), where the noisy signal data X to be trained n (t) can be expressed as:

[0068] ;

[0069] in, is the noisy signal data to be trained; is the residual component; N is the order of IMF decomposed; is the nth order IMF component.

[0070] Step S12, using an improved empirical mode decomposition algorithm to decompose the noisy signal data to be trained, and outputting a plurality of multi-modal sub-signal data;

[0071] It should be noted that EEMD is used to decompose the noisy signal (the noisy signal data to be trained) to obtain the decomposition result (i.e. multiple multimodal sub-signal data).

[0072] Step S13, using a wavelet packet algorithm to perform noise reduction on each multimodal sub-signal data to generate a plurality of multimodal noise-reduced sub-signal data;

[0073] Step S14: reconstruct the plurality of multimodal denoised sub-signal data based on a preset adaptive threshold, and output the target multimodal signal data to be trained.

[0074] It should be noted that the wavelet packet algorithm is used to reduce the noise of the noisy signal (multimodal denoising sub-signal data). In order to prevent the loss of useful information in the data, the noise standard deviation estimation is introduced to design the adaptive threshold (preset adaptive threshold), and finally the final denoised signal (multimodal signal data to be trained) is reconstructed; the adaptive threshold calculation is shown in the following formula:

[0075] ;

[0076] in, To preset the adaptive threshold; is the threshold factor, set to 0.5; is the noise standard deviation estimate; N is the length of the signal; MAD is the median of the absolute value of the difference between the detail coefficient of the wavelet transform and its median; 0.6745 represents the ratio of the median absolute deviation to the standard deviation of the normal distribution.

[0077] In this embodiment, initial multimodal signal data to be trained is acquired, and based on a preset signal processing algorithm, the initial multimodal signal data to be trained is preprocessed to generate target multimodal signal data to be trained.

[0078] Step 102: Calculate the degradation index based on the target multimodal signal data to be trained using a preset distribution function.

[0079] The preset distribution functions include the modified gamma distribution probability density function and the cumulative distribution function.

[0080] It should be noted that in the degradation index construction stage, the existing health index construction method only uses a single sensor signal, which is not enough to fully reflect the overall health status of the system. The present invention uses multi-sensor information to obtain more comprehensive degradation characteristics of components, and combines gamma distribution and Bayesian inference to propose a comprehensive degradation index construction based on the combination of failure risk probability and entropy value difference.

[0081] Further, step 102 may include the following sub-steps:

[0082] S21, using an improved gamma distribution probability density function to determine a gamma probability density distribution according to the continuous degradation amount in the target multimodal signal data to be trained;

[0083] It should be noted that, considering that the degradation process of mechanical parts of complex equipment without repair actions is continuous and monotonic, and the gamma process is also a non-negative, strictly monotonic random process, with monotonic, stable and independent Gaussian increments, it can well describe the degradation process of parts caused by wear and the like. Therefore, the present invention assumes that the continuous degradation process of key components obeys the gamma process, then the continuous degradation amount of the component at any time obeys the gamma distribution, and the continuous degradation amount is obtained by non-negative transformation of the original data (for example, calculating its effective value, etc.). In addition, by introducing time-varying shape parameters and initial degradation offsets, the gamma distribution probability density function that describes the degradation process of the component is obtained, that is, the improved gamma distribution probability density function, which is specifically:

[0084] ;

[0085] in, is the gamma probability density distribution; is the continuous degradation amount; is a gamma distribution; is the gamma function; is the scale parameter; is the shape parameter that changes with time t; s is the observation data input; is the offset term, indicating the initial value at which degradation begins; is the indicator function, ensuring that s is Within the range.

[0086] It is worth mentioning that The value increases with time, indicating that the cumulative effect in the degradation process is obvious, that is, as the usage time increases, the degree of degradation of the component increases significantly. The size indicates the significance of the degradation. The smaller the value, the faster the degradation rate, and the larger the value, the slower the rate. It can help calibrate the starting point of the degradation model to make it closer to the actual situation.

[0087] S22, substituting the continuous degradation amount in the target multimodal signal data to be trained into the cumulative distribution function, and outputting the joint failure risk probability;

[0088] It should be noted that in order to dynamically adjust the degradation index and evaluate the uncertainty of the degradation state of the component, Bayesian inference is introduced to estimate the parameters of the degradation data distribution, and the prior knowledge is combined with the observed data to infer the posterior distribution. The posterior distribution is defined as:

[0089] ;

[0090] in, is the observation data input at time t-1, that is, the observation data input at the past time; is the first update parameter of the posterior distribution; is the second update parameter of the posterior distribution; is the posterior distribution; For the observed value The gamma distribution obtained by posterior inference is as follows.

[0091] Furthermore, the likelihood function is defined as:

[0092] ;

[0093] in, is the likelihood function, which means that given the known observed data In the case of The joint probability distribution of is the observation sample at time i; For a given parameter Next, the observed value Gamma distribution probability density function.

[0094] Furthermore, Bayesian inference is a parameter estimation based on Bayes' theorem, which combines prior knowledge and likelihood function to update the posterior distribution of parameters. The posterior probability distribution is calculated as follows:

[0095] ;

[0096] in, is the posterior probability distribution; is the prior distribution; is the likelihood function; For parameters , and ; s is the observed sample.

[0097] After the above Bayesian inference degradation data distribution model construction process, a degradation data distribution model based on multiple variables can be obtained. In order to achieve a comprehensive understanding and evaluation of component degradation, the present invention constructs a multivariate joint distribution model. Assuming that vibration and temperature are two variables, the joint distribution model is defined as:

[0098] ;

[0099] in, is the joint distribution model; is the edge distribution function of vibration; is the marginal distribution function of temperature; v is the observed value of vibration degradation; t is the observed value of temperature degradation; C is the Gaussian Couple function.

[0100] Furthermore, the Gaussian Couple function can be expressed as:

[0101] ;

[0102] in, is the cumulative distribution function of the standard normal distribution; is the inverse cumulative function of the standard normal distribution; is the cumulative distribution function of the bivariate standard normal distribution of the correlation coefficient R; is the first standardized variable, and its value range is [0,1]; is the second standardized variable, and its value range is [0,1].

[0103] Next, the degradation index calculation is performed, including the failure risk probability calculation and the entropy value difference calculation. For the degradation amount of a certain variable, a threshold is given , the probability of exceeding this threshold :

[0104] ;

[0105] in, is the marginal distribution function of the gamma distribution; is the probability that is less than or equal to the threshold; is the component failure risk probability calculated based on a certain variable distribution at time t.

[0106] Based on the above foundation, assuming that the target multimodal signal data to be trained includes two variables, vibration and temperature, the calculation process of the corresponding joint failure risk probability can be expressed as:

[0107] ;

[0108] in, is the joint failure risk probability; is the threshold value of vibration degradation; is the threshold value of temperature degradation; V is the vibration degradation; T is the temperature degradation; and are the cumulative distribution functions of vibration and temperature degradation respectively; C is the Gaussian Couple function.

[0109] It is worth mentioning that the cumulative distribution function can refer to the specific expression of the existing cumulative distribution function, and the present invention will not go into details.

[0110] S23, determining the current moment entropy value corresponding to the target multimodal signal data to be trained based on the gamma probability density distribution;

[0111] It should be noted that when evaluating the operating status of a component over time during its entire life cycle, the instability of the component's operating status can be reflected by calculating the difference in the entropy values ​​of the distribution in the two time periods before and after. is the probability density function of the component state at time, then the entropy value at this time, that is, based on the gamma probability density distribution, determines the current entropy value corresponding to the target multimodal signal data to be trained ; The process can be expressed as:

[0112] ;

[0113] in, is the entropy value at the current moment; is the gamma probability density distribution.

[0114] S24, performing a difference operation on the current moment entropy value and the historical moment entropy value corresponding to the target multimodal signal data to be trained, and determining the difference in entropy values ​​corresponding to the target multimodal signal data to be trained;

[0115] The historical moment entropy value is the entropy value of the previous moment corresponding to the target multimodal signal data to be trained.

[0116] It should be noted that, assuming and Calculate the change in entropy value during the time period It can be expressed as the difference between the entropy values ​​of the two time periods before and after, that is, the calculation process of the difference in entropy values ​​corresponding to the multimodal signal data of the target to be trained can be expressed as:

[0117] ;

[0118] in, is the difference in entropy values ​​corresponding to the target multimodal signal data to be trained; is the current moment entropy value corresponding to the target multimodal signal data to be trained; is the historical moment entropy value corresponding to the target multimodal signal data to be trained.

[0119] S25. Calculate the degradation index based on the difference in entropy values ​​and the joint failure risk probability.

[0120] It should be noted that when constructing the degradation index, the difference in entropy values ​​can be used to measure the degree of uncertainty change during the system degradation process. Combining the difference in entropy values ​​with reliability can form a more comprehensive degradation index. Assuming that the target multimodal signal data to be trained includes two variables, vibration and temperature, the calculation process of the corresponding degradation index can be expressed as:

[0121] ;

[0122] Among them, DI is the degradation index; is the weight factor of failure risk probability; is the joint failure risk probability at time t; is the weight factor of the entropy difference; is the difference in entropy values ​​corresponding to the vibration variables; is the difference in entropy values ​​corresponding to the temperature variables.

[0123] In this embodiment, a preset distribution function is used to calculate the degradation index based on the target multimodal signal data to be trained.

[0124] Step 103: Use the target multimodal signal data to be trained and the degradation index to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model.

[0125] Specifically, step 103 may include the following sub-steps S31-S36:

[0126] Step S31, inputting the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction, and generating a predicted degradation result of the mechanical rotating component to be trained;

[0127] It should be noted that in the degradation trend prediction modeling stage. In view of the modeling problem of long-distance dependencies, the present invention adopts the Mamba prediction network based on the state space model to predict the degradation trend of components. The state space model is a model that uses ordinary differential equations to describe the evolution of states over time, which is suitable for time series modeling tasks. Based on the state space model, the Mamba model introduces the HiPPO matrix to efficiently represent the accumulated historical information in a limited storage space, and adopts a selection mechanism to filter out irrelevant information and reset the state, thereby enhancing the traditional state space model to achieve fast reasoning and long-distance dependency modeling. The present invention improves on the basis of the Mamba model and proposes an improved Mamba degradation prediction model based on attention fusion, that is, an initial mechanical rotating component degradation prediction model, which includes three parts, namely an embedding layer, N Mamba blocks (Mamba blocks) and a linear output layer.

[0128] Further, step S31 may include the following sub-steps:

[0129] Step S311, mapping the target multimodal signal data to be trained using a time vector algorithm through an embedding layer to generate multimodal mapping signal data;

[0130] It should be noted that the embedding layer uses the Time2vec algorithm (time vector algorithm) to map the time series data. The basic idea is to jointly learn the representation of time steps through a linear layer and a periodic activation function. The algorithm maps the time step to a vector space of a specific dimension, which can capture the periodic and non-periodic patterns in the time series data. In the context of time changes, similar features can still remain stable. The implementation formula is as follows:

[0131] ;

[0132] in, Mapping signal data for multimodality; is the target multimodal signal data to be trained; is the weight; is the bias; F is the periodic activation function; k is the dimension of time2vec; i is the dimension index.

[0133] Furthermore, assuming that Batchsize is set to B, the length of the time series data is L, and the number of input features is D, the input time series data is , after the embedding layer of time2vec, the output is .

[0134] Step S312: using multiple Mamba blocks to perform efficient sequence processing on the multimodal mapping signal data to generate target degradation features;

[0135] It should be noted that multiple Mamba blocks are in a cascade relationship. After the multimodal mapping signal data is processed by each Mamba block in turn (that is, the output of the previous Mamba block is the input of the next Mamba block), the feature data output by the last Mamba block is used as the target degradation feature.

[0136] Optionally, the Mamba block includes a first linear layer, a second linear layer, a third linear layer, a SiLU activation function layer, a dimensional channel attention module, and a state space module; the data processing process of the Mamba block includes:

[0137] S1, performing linear transformation on the input multimodal signal data input to the Mamba block through a first linear layer and a second linear layer respectively, to generate a first multimodal transformation signal feature and a second multimodal transformation signal feature;

[0138] S2, using a SiLU activation function layer to perform nonlinear mapping on the first multimodal transformation signal feature, and outputting a multimodal nonlinear signal feature;

[0139] S3, capturing cross-dimensional channel information of the second multimodal transformation signal feature through a dimensional channel attention module to generate a multimodal attention signal feature;

[0140] S4, multiplying the multimodal attention signal feature and the second multimodal transformation signal feature, and outputting a first multimodal multiplication signal feature;

[0141] S5. Using a state space module to extract features of the first multimodal multiplication signal to generate signal degradation features;

[0142] S6, multiplying the signal degradation feature and the multimodal nonlinear signal feature, and outputting a second multimodal multiplied signal feature;

[0143] S7. Use the third linear layer to perform linear transformation on the second multimodal multiplication signal feature to generate an output signal feature.

[0144] The input multimodal signal data is the signal data input to the Mamba block. It can be understood that the input multimodal signal data can correspond to any signal data input to the Mamba block for data processing during the model training or prediction process;

[0145] The first multimodal transformation signal feature, the second multimodal transformation signal feature, the multimodal nonlinear signal feature, the multimodal attention signal feature, the first multimodal multiplication signal feature, the signal degradation feature, and the second multimodal multiplication signal feature are all intermediate feature data generated in the Mamba block;

[0146] The output signal feature is the feature data output by the Mamba block. It can be understood that it can correspond to any feature data output after the Mamba block performs data processing during model training or prediction.

[0147] It should be noted that the multimodal mapping signal data enters N Mamba modules (Mamba blocks) for processing. The Mamba module mainly includes linear layers (first linear layer, second linear layer, third linear layer), dimensional channel attention module and SSM module (State Space Model Module). The data processing process of the Mamba module can be expressed as:

[0148] ;

[0149] in, The output signal feature is the output signal feature output by the last Mamba block as the target degradation feature; is the linear transformation layer (linear layer); It is a composite activation function based on Sigmoid function and Linear function; It is a dimensional channel attention module, which has good cross-dimensional channel information capture capabilities; For SSM module.

[0150] Furthermore, the processing process of the first multimodal multiplication signal feature can be expressed as:

[0151] ;

[0152] in, is a second multi-modal transformation signal feature; is the first multi-modal multiplication signal feature; is the Sigmoid activation function; GAP is the global average pooling operation; C1D is the 1×1 convolution operation; To multiply.

[0153] Furthermore, the SSM module maps the input signal to the corresponding output through state representation. This step is used to describe the evolution of the component degradation state over time and to extract degradation features. The main implementation formula is:

[0154] ;

[0155] in, is the state vector at time k; is the first discretization parameter of the SSM module; is the state vector at time k-1; is the second discretization parameter of the SSM module; is the input vector of the SSM module at time k; is the output vector of the SSM module at time k; is the third discretization parameter of the SSM module; It is the fourth discretization parameter of the SSM module.

[0156] It is worth mentioning that Mamba also introduces a selective scanning mechanism to enhance the interaction between sequences, so that the model can filter out noise information irrelevant to the time series task, while selectively propagating or forgetting relevant information of the current input.

[0157] Step S313: Use the target degradation feature as the input of the linear output layer to output the predicted degradation result of the mechanical rotating component to be trained.

[0158] Furthermore, we finally enter the linear output layer, which includes two linear layers and an activation function. The specific calculation is as follows:

[0159] ;

[0160] in, is the output predicted degradation value, i.e., the predicted degradation result of the mechanical rotating component to be trained; is the Sigmoid activation function; is the second weight matrix of the linear layer; is the first weight matrix of the linear layer; is the target degradation feature; is the first bias vector of the linear layer; is the second bias vector of the linear layer.

[0161] Step S32, taking the degradation index as the actual degradation result of the mechanical rotating component, and determining the model gradient according to the actual degradation result of the mechanical rotating component and the predicted degradation result of the mechanical rotating component to be trained;

[0162] It should be noted that the degradation index is taken as the real degradation result (true value) of the mechanical rotating part, and is substituted into the root mean square error (RMSE) loss function together with the predicted degradation result of the mechanical rotating part to be trained and derived to determine the model gradient.

[0163] Step S33, using the model gradient to update the model parameters of the initial mechanical rotating component degradation prediction model to determine the intermediate mechanical rotating component degradation prediction model;

[0164] Step S34, using the intermediate mechanical rotating component degradation prediction model and the preset root mean square error loss function to determine the target loss value according to the target multimodal signal data to be trained and the actual degradation result of the mechanical rotating component;

[0165] It should be noted that the degradation index is taken as the actual degradation result (true value) of the mechanical rotating part, and is substituted into the root mean square error (RMSE) loss function together with the predicted degradation result of the mechanical rotating part to be trained to calculate the target loss value.

[0166] Step S35, judging whether the target loss value converges;

[0167] Step S36: If yes, the intermediate mechanical rotating component degradation prediction model is used as the trained target mechanical rotating component degradation prediction model.

[0168] It should be noted that if the target loss value has not converged, the intermediate mechanical rotating component degradation prediction model is used as the new initial mechanical rotating component degradation prediction model, and the execution step S31 is jumped until the target loss value converges, and the intermediate mechanical rotating component degradation prediction model determined when the target loss value converges is used as the trained target mechanical rotating component degradation prediction model.

[0169] In this embodiment, the target multimodal signal data to be trained and the degradation index are used to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model.

[0170] For example, see Figure 2The framework of multimodal data modeling for rotating component degradation prediction proposed in the present invention is divided into three steps: multimodal data acquisition and preprocessing, degradation index construction, and degradation trend prediction modeling. First, in the multimodal data acquisition and preprocessing stage, multi-source sensors are used to comprehensively monitor complex equipment mechanical components, and various signal data (i.e., initial multimodal signal data) such as temperature and vibration are collected throughout the life cycle of the components. In order to improve the data quality, the original signal is denoised by using an improved empirical mode decomposition combined with a wavelet threshold denoising algorithm (i.e., a preset signal processing algorithm), thereby obtaining denoised multi-source data (i.e., target multimodal signal data). Secondly, in the degradation index construction stage, considering the non-negativity and strict monotonicity exhibited by mechanical components in degradation processes such as wear, and the gamma distribution is a non-negative random variable distribution, the present invention selects the gamma distribution as a probability model to describe the degradation process. In order to better evaluate the uncertainty of the degradation state of the component and realize the dynamic adjustment of the degradation index, the present invention introduces Bayesian inference for parameter estimation, and derives the posterior distribution in the degradation process by combining prior knowledge and observed data. On this basis, a multivariate joint distribution model is constructed in combination with the distribution of multi-source data, and a failure risk probability calculation formula is established in combination with the cumulative distribution function of the gamma distribution, and a comprehensive degradation index is constructed in combination with the change in entropy value. Finally, in the degradation prediction modeling stage, the present invention constructs an improved Mamba degradation prediction model (i.e., a mechanical rotating component degradation prediction model) based on the fusion of attention mechanisms, which combines the advantages of both the state space model and the dimensional attention mechanism, effectively extracts degradation features that evolve over time, and realizes accurate prediction of the degradation trend of mechanical components.

[0171] As a comparison of technical effects, we can refer to existing technologies. Currently, most degradation prediction modeling methods for mechanical components rely on data from a single sensor. However, in practical applications, the degradation process of mechanical components is often the result of multiple factors, and a single type of sensor data is difficult to fully capture the degradation trend of components under complex working conditions.

[0172] In order to make up for the limitation of insufficient information from a single sensor, the present invention proposes a multimodal data modeling method for rotating component degradation prediction to achieve accurate prediction of the degradation evolution trend of mechanical components. Specifically, a gamma distribution with time-varying shape parameters and initial degradation offset is used to describe the degradation process of mechanical components caused by wear and other reasons. At the same time, the Bayesian inference method is used to estimate the parameters of the data distribution in the degradation process, so as to achieve dynamic adjustment of the degradation index and uncertainty evaluation under the degradation state of the component. Then, the calculation formula of the comprehensive degradation index is designed in combination with the failure risk probability and the entropy value difference. The index can evaluate the health status of mechanical components from multiple dimensions and is a more comprehensive and accurate comprehensive degradation index. Finally, the trained Mamba degradation prediction model (target mechanical rotating component degradation prediction model) based on the fusion of the attention mechanism is used to predict the degradation trend of the component. Combined with the above content, and then through the steps of the proposed system, the accurate prediction of the degradation trend of mechanical components can be effectively achieved in practical applications.

[0173] In summary, the present invention comprehensively monitors the operating status of mechanical parts of complex equipment through multi-source sensors (such as vibration, temperature, acoustic emission, etc.), collects multi-source signal data of the whole life cycle of the parts, and the data reflects the degradation trend process of mechanical parts from different angles, providing more comprehensive degradation information for subsequent degradation index construction and predictive modeling, thereby overcoming the limitations of traditional prediction methods based on single-modal data. The present invention also innovatively proposes to use the gamma distribution that introduces time-varying shape parameters and initial degradation offsets to describe the degradation process of mechanical parts caused by wear and other reasons, so that the model can more flexibly adapt to different degradation forms of different mechanical parts. At the same time, the Bayesian inference method is used to estimate the parameters of data distribution in the degradation process, realize the dynamic adjustment of degradation indicators and uncertainty assessment under the degradation state of parts. At the same time, the present invention designs a joint distribution of multivariate variables, and calculates the failure risk probability of mechanical parts at a specific time point by setting a threshold value according to the cumulative distribution function of the gamma distribution; at the same time, the difference between the entropy values ​​of the data distribution in two adjacent time periods is calculated to measure the degree of increase in the instability of the parts over time. Combining the failure risk probability and the entropy value difference, the weight factor of the failure risk probability and the weight factor of the entropy value difference that can be adjusted as needed are introduced to construct a more comprehensive and accurate comprehensive degradation index of mechanical components. In addition, the present invention innovatively proposes an improved Mamba degradation prediction model based on the fusion of dimensional attention mechanism, which includes an embedding layer, a Mamba block, and an output layer. The embedding layer uses the Time2vec algorithm to map the time series data to capture the periodic and non-periodic features in the time series data. The state space model and the dimensional channel attention module are innovatively fused in the Mamba block, where the state space model can effectively describe the process of state evolution over time, and is used to extract degradation features that change over time. The dimensional channel attention mechanism has excellent cross-dimensional information capture capabilities and can efficiently fuse multi-source data. The model can effectively extract the degradation features of mechanical components throughout their life cycle and achieve high-precision prediction of component degradation trends.

[0174] In an embodiment of the present invention, the present invention provides a multimodal data modeling method for rotating component degradation prediction. First, initial multimodal signal data to be trained is obtained, and based on a preset signal processing algorithm, the initial multimodal signal data to be trained is preprocessed to generate target multimodal signal data to be trained; then, a preset distribution function is used to calculate the degradation index according to the target multimodal signal data to be trained; finally, the target multimodal signal data to be trained and the degradation index are used to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model; based on the above scheme, the present invention uses the preprocessed target multimodal signal data to be trained, and combines the degradation index output by the preset distribution function to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model. The process can overcome the limitations of traditional prediction methods based on single-modal data, provide more comprehensive degradation information for predictive modeling, and thus improve the accuracy of model prediction.

[0175] For better explanation, refer to Figure 3 , showing a schematic diagram of the execution process of the multimodal data modeling method for rotating component degradation prediction provided by the second embodiment of the present invention. It should be pointed out that this embodiment only briefly describes the general process of the multimodal data modeling method for rotating component degradation prediction, and the specific implementation process of each step can be understood by referring to the relevant content in the aforementioned embodiment, which will not be described here. It can be understood that the present invention is not limited to this.

[0176] (1) Collect degradation data of mechanical components: Use multi-source sensors to collect degradation data of mechanical components throughout their life cycle, such as vibration, acoustic emission, temperature, etc., to ensure data integrity and availability.

[0177] (2) Data preprocessing: The signal to be denoised is denoised based on the improved empirical mode decomposition combined with the wavelet threshold algorithm, and then normalized or standardized preprocessing is performed; the data is divided into multiple samples in chronological order using the sliding window partitioning method, and then divided into training set, validation set and test set.

[0178] (3) Degradation index construction: The preprocessed multimodal data is used to construct preliminary degradation indicators. For example, for the vibration signal, its effective value can be taken as the degradation indicator as the input of the degradation index construction stage. Then, the above-mentioned gamma distribution, Bayesian inference, failure risk probability calculation, entropy value calculation and degradation index calculation processes are combined to realize the extraction of degradation indicators.

[0179] (4) Model training: The preprocessed training set is used as the input data for model training, and the degradation index constructed above is used as the label. Figure 1In the improved Mamba degradation prediction model based on attention mechanism fusion shown in the figure, the parameters are optimized by regression loss back propagation during the training process to obtain the optimal network model parameters. The loss function adopts the root mean square error (RMSE) loss function, which can give greater weight to larger prediction errors during the gradient back propagation process, prompting the model to pay more attention to reducing these larger errors and improve the prediction accuracy and convergence speed of the model.

[0180] (5) Model evaluation and parameter adjustment: After completing model training, use the validation set to evaluate the model, calculate indicators such as RMSE and MAE on the validation set, and observe the comparison between the model's prediction performance on the validation set and the true label. Adjust model hyperparameters such as learning rate and number of hidden layer units based on the prediction effect of the validation set, and save the model with the best performance during the training process.

[0181] (6) Model testing and application: Input the degradation data in the test set or actual task into the best performing model, output the degradation value of the component, and realize the degradation trend prediction task of mechanical components.

[0182] In an embodiment of the present invention, the present invention describes the degradation process of mechanical parts caused by wear and other reasons by using a gamma distribution that introduces time-varying shape parameters and initial degradation offsets, and uses the Bayesian inference method to estimate the parameters of the data distribution in the degradation process, thereby realizing dynamic adjustment of the degradation index and uncertainty assessment under the degradation state of the parts. Then, a calculation formula for the comprehensive degradation index is designed in combination with the failure risk probability and the entropy value difference. This index can evaluate the health status of mechanical parts from multiple dimensions and is a more comprehensive and accurate comprehensive degradation index. Finally, the Mamba degradation prediction model based on the fusion of the attention mechanism is used to predict the degradation trend of the parts. Combined with the above content, and then through the steps of the proposed system, it is finally possible to effectively realize the accurate prediction of the degradation trend of mechanical parts in practical applications.

[0183] See also Figure 4 , Figure 4 This is a flow chart of the steps of model prediction of the target mechanical rotating component degradation prediction model provided in the third embodiment of the present invention.

[0184] For better explanation, refer to Figure 4 , shows a flowchart of the steps of model prediction of the target mechanical rotating component degradation prediction model provided by the third embodiment of the present invention. The process may include the following steps:

[0185] Step 401: When receiving the initial multimodal signal data to be detected, preprocessing the initial multimodal signal data to be detected based on a preset signal processing algorithm to generate target multimodal signal data to be detected;

[0186] Step 402: Use the target mechanical rotating component degradation prediction model to predict the target multimodal signal data to generate a target mechanical rotating component prediction degradation result.

[0187] In an embodiment of the present invention, the present invention designs a Mamba degradation prediction model (mechanical rotating component degradation prediction model) based on attention mechanism fusion by combining the time evolution feature extraction advantages of the state-space model and the selective fusion advantages of the attention mechanism, thereby achieving accurate prediction of the component degradation process.

[0188] See also Figure 5 , Figure 5 This is a structural block diagram of a multimodal data modeling device for rotating component degradation prediction provided in Embodiment 4 of the present invention.

[0189] The present invention provides a multimodal data modeling device for predicting degradation of rotating parts, comprising:

[0190] The acquisition module 501 is used to acquire the initial multimodal signal data to be trained, and pre-process the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained;

[0191] A calculation module 502, used to calculate a degradation index according to target multimodal signal data to be trained using a preset distribution function;

[0192] The training module 503 is used to perform model training on the initial mechanical rotating component degradation prediction model using the target multimodal signal data to be trained and the degradation index, so as to determine the target mechanical rotating component degradation prediction model.

[0193] Furthermore, the preset signal processing algorithm includes an improved empirical mode decomposition algorithm and a wavelet packet algorithm; the acquisition module 501 is specifically used for:

[0194] Adding white noise to the initial multimodal signal data to be trained to generate noisy signal data to be trained;

[0195] An improved empirical mode decomposition algorithm is used to decompose the noisy signal data to be trained, and multiple multi-modal sub-signal data are output;

[0196] Using a wavelet packet algorithm to reduce the noise of each multimodal sub-signal data to generate multiple multimodal noise-reduced sub-signal data;

[0197] Based on the preset adaptive threshold, multiple multimodal denoising sub-signal data are reconstructed to output the target multimodal signal data to be trained.

[0198] Further, the preset distribution function includes an improved gamma distribution probability density function and a cumulative distribution function; the calculation module 502 is specifically used for:

[0199] An improved gamma distribution probability density function is used to determine the gamma probability density distribution according to the continuous degradation amount in the target multimodal signal data to be trained;

[0200] Substituting the continuous degradation amount in the target multimodal signal data to be trained into the cumulative distribution function, and outputting the joint failure risk probability;

[0201] Based on the gamma probability density distribution, determine the current moment entropy value corresponding to the target multimodal signal data to be trained;

[0202] Performing a difference operation on the current moment entropy value and the historical moment entropy value corresponding to the target multimodal signal data to be trained, and determining the difference in entropy values ​​corresponding to the target multimodal signal data to be trained;

[0203] The degradation index is calculated based on the difference in entropy values ​​and the joint failure risk probability.

[0204] Furthermore, the training module 503 includes:

[0205] The first submodule is used to input the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction, and generate a predicted degradation result of the mechanical rotating component to be trained;

[0206] The second submodule is used to use the degradation index as the real degradation result of the mechanical rotating component, and determine the model gradient according to the real degradation result of the mechanical rotating component and the predicted degradation result of the mechanical rotating component to be trained;

[0207] The third submodule is used to update the model parameters of the initial mechanical rotating component degradation prediction model using the model gradient to determine the intermediate mechanical rotating component degradation prediction model;

[0208] The fourth submodule is used to determine the target loss value according to the target multimodal signal data to be trained and the actual degradation result of the mechanical rotating component by using the intermediate mechanical rotating component degradation prediction model and the preset root mean square error loss function;

[0209] The fifth submodule is used to determine whether the target loss value converges;

[0210] The sixth submodule is used for, if yes, using the intermediate mechanical rotating component degradation prediction model as the trained target mechanical rotating component degradation prediction model.

[0211] Furthermore, the initial mechanical rotating component degradation prediction model includes an embedding layer, a plurality of Mamba blocks and a linear output layer; a first submodule, specifically used for:

[0212] The time vector algorithm is used to map the multimodal signal data of the training target through the embedding layer to generate multimodal mapping signal data;

[0213] Multiple Mamba blocks are used to perform efficient sequence processing on multimodal mapping signal data to generate target degradation features;

[0214] The target degradation features are used as the input of the linear output layer, and the predicted degradation results of the mechanical rotating parts to be trained are output.

[0215] Optionally, the Mamba block includes a first linear layer, a second linear layer, a third linear layer, a SiLU activation function layer, a dimensional channel attention module, and a state space module; the data processing process of the Mamba block includes:

[0216] Performing linear transformation on the input multimodal signal data input to the Mamba block through a first linear layer and a second linear layer respectively, to generate a first multimodal transformation signal feature and a second multimodal transformation signal feature;

[0217] A SiLU activation function layer is used to perform nonlinear mapping on the first multimodal transformation signal feature, and a multimodal nonlinear signal feature is output;

[0218] Capturing cross-dimensional channel information of the second multimodal transformation signal feature through a dimensional channel attention module to generate a multimodal attention signal feature;

[0219] Multiplying the multimodal attention signal feature and the second multimodal transformation signal feature, and outputting a first multimodal multiplied signal feature;

[0220] Using a state space module to extract features of the first multimodal multiplication signal to generate signal degradation features;

[0221] Multiplying the signal degradation feature and the multimodal nonlinear signal feature to output a second multimodal multiplied signal feature;

[0222] The third linear layer is used to perform a linear transformation on the second multimodal multiplication signal features to generate output signal features.

[0223] In an optional device embodiment, it also includes:

[0224] The first module is used for, when receiving the initial multimodal signal data to be detected, preprocessing the initial multimodal signal data to be detected based on a preset signal processing algorithm to generate target multimodal signal data to be detected;

[0225] The second module is used to use the target mechanical rotating component degradation prediction model to predict the multi-modal signal data of the target to be detected, and generate the target mechanical rotating component prediction degradation result.

[0226] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices, modules and sub-modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0227] An embodiment of the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the multimodal data modeling method for predicting degradation of rotating components as described in any of the above embodiments.

[0228] An embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of a multimodal data modeling method for predicting degradation of rotating components as described in any of the above embodiments are implemented.

[0229] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0230] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0231] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal data modeling method for rotating component degradation prediction, characterized in that: include: Acquire initial multimodal signal data to be trained, and preprocess the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained; Calculating a degradation index based on the target multimodal signal data to be trained using a preset distribution function; The target multimodal signal data to be trained and the degradation index are used to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model.

2. The multimodal data modeling method for rotating component degradation prediction according to claim 1, characterized in that: The preset signal processing algorithm includes an improved empirical mode decomposition algorithm and a wavelet packet algorithm; based on the preset signal processing algorithm, the initial multimodal signal data to be trained is preprocessed to generate target multimodal signal data to be trained, including: Adding white noise to the initial multimodal signal data to be trained to generate noisy signal data to be trained; Decomposing the noisy signal data to be trained by using an improved empirical mode decomposition algorithm to output a plurality of multimodal sub-signal data; Using a wavelet packet algorithm to perform noise reduction on each of the multimodal sub-signal data to generate a plurality of multimodal noise-reduced sub-signal data; The plurality of multimodal denoising sub-signal data are reconstructed based on a preset adaptive threshold, and target multimodal signal data to be trained is output.

3. The multimodal data modeling method for rotating component degradation prediction according to claim 1, characterized in that: The preset distribution function includes an improved gamma distribution probability density function and a cumulative distribution function; the use of the preset distribution function to calculate the degradation index according to the target multimodal signal data to be trained includes: Determining a gamma probability density distribution according to the continuous degradation amount in the target multimodal signal data to be trained using an improved gamma distribution probability density function; Substituting the continuous degradation amount in the target multimodal signal data to be trained into the cumulative distribution function, and outputting the joint failure risk probability; Based on the gamma probability density distribution, determining the current moment entropy value corresponding to the target multimodal signal data to be trained; Performing a difference operation on the current moment entropy value and the historical moment entropy value corresponding to the target multimodal signal data to be trained, to determine the difference between the entropy values ​​corresponding to the target multimodal signal data to be trained; A degradation index is calculated according to the difference between the entropy values ​​and the joint failure risk probability.

4. The multimodal data modeling method for rotating component degradation prediction according to claim 1, characterized in that: The adopting the target multimodal signal data to be trained and the degradation index to perform model training on the initial mechanical rotating component degradation prediction model to determine the target mechanical rotating component degradation prediction model includes: Inputting the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction, and generating a predicted degradation result of the mechanical rotating component to be trained; Taking the degradation index as the real degradation result of the mechanical rotating component, and determining the model gradient according to the real degradation result of the mechanical rotating component and the predicted degradation result of the mechanical rotating component to be trained; Using the model gradient to update the model parameters of the initial mechanical rotating component degradation prediction model to determine an intermediate mechanical rotating component degradation prediction model; Determine a target loss value by using the intermediate mechanical rotating component degradation prediction model and a preset root mean square error loss function according to the target multimodal signal data to be trained and the actual degradation result of the mechanical rotating component; Determining whether the target loss value converges; If so, the intermediate mechanical rotating component degradation prediction model is used as the trained target mechanical rotating component degradation prediction model.

5. The multimodal data modeling method for rotating component degradation prediction according to claim 4, characterized in that: The initial mechanical rotating component degradation prediction model includes an embedding layer, a plurality of Mamba blocks and a linear output layer; the step of inputting the target multimodal signal data to be trained into the initial mechanical rotating component degradation prediction model for prediction to generate a predicted degradation result of the mechanical rotating component to be trained includes: Mapping the target multimodal signal data to be trained by using a time vector algorithm through an embedding layer to generate multimodal mapping signal data; Using a plurality of Mamba blocks to perform efficient sequence processing on the multimodal mapping signal data to generate target degradation features; The target degradation feature is used as the input of the linear output layer, and the predicted degradation result of the mechanical rotating part to be trained is output.

6. The multimodal data modeling method for rotating component degradation prediction according to claim 5, characterized in that: The Mamba block includes a first linear layer, a second linear layer, a third linear layer, a SiLU activation function layer, a dimensional channel attention module, and a state space module; the data processing process of the Mamba block includes: Performing linear transformation on the input multimodal signal data input to the Mamba block through a first linear layer and a second linear layer respectively, to generate a first multimodal transformation signal feature and a second multimodal transformation signal feature; Using a SiLU activation function layer to perform nonlinear mapping on the first multimodal transformation signal feature, and outputting a multimodal nonlinear signal feature; Capturing cross-dimensional channel information of the second multimodal transformation signal feature through a dimensional channel attention module to generate a multimodal attention signal feature; multiplying the multimodal attention signal feature and the second multimodal transformation signal feature to output a first multimodal multiplied signal feature; Using a state space module to extract features of the first multimodal multiplication signal to generate signal degradation features; multiplying the signal degradation feature and the multimodal nonlinear signal feature to output a second multimodal multiplied signal feature; A third linear layer is used to perform a linear transformation on the second multimodal multiplication signal feature to generate an output signal feature.

7. The multimodal data modeling method for rotating component degradation prediction according to claim 1, characterized in that: Also includes: When receiving the initial multimodal signal data to be detected, preprocessing the initial multimodal signal data to be detected based on a preset signal processing algorithm to generate target multimodal signal data to be detected; The target mechanical rotating component degradation prediction model is used to predict the target multimodal signal data to be detected, and a target mechanical rotating component prediction degradation result is generated.

8. A multimodal data modeling device for predicting degradation of rotating parts, characterized in that: include: An acquisition module, used for acquiring initial multimodal signal data to be trained, and preprocessing the initial multimodal signal data to be trained based on a preset signal processing algorithm to generate target multimodal signal data to be trained; A calculation module, used to calculate a degradation index according to the target multimodal signal data to be trained by using a preset distribution function; The training module is used to perform model training on the initial mechanical rotating component degradation prediction model using the target multimodal signal data to be trained and the degradation index to determine the target mechanical rotating component degradation prediction model.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of the multimodal data modeling method for rotating component degradation prediction as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the multimodal data modeling method for rotating component degradation prediction according to any one of claims 1 to 7 is implemented.

Images