A Method for Predicting Remaining Mechanical Service Based on Implicit Dynamics Hybrid Drive

Abstract

This invention relates to the field of mechanical life prediction and discloses a method for predicting the remaining service life of machinery based on implicit dynamics hybrid drive. The method includes the following steps: acquiring vibration signals of the machinery under various operating conditions and constructing a training set; extracting evolutionary feature vectors through a mapping evolution model and outputting preliminary prediction results; constructing an implicit dynamics model based on the evolutionary feature vectors and the preliminary prediction results; calculating a joint loss function composed of the data feature loss of the mapping evolution model and the dynamic loss of the implicit dynamics model, and updating the model parameters in the mapping evolution model and the implicit dynamics model through backpropagation; and using the trained mapping evolution model to predict the remaining service life of the tested rotating machinery to obtain the remaining service life of the tested sample. This invention utilizes the vibration data of rotating machinery operation to automatically learn and mine the implicit dynamic features hidden in the degradation process, achieving accurate prediction of the remaining service life of rotating machinery.

Description

Technical Field

[0001] This invention relates to the field of mechanical life prediction technology, and more specifically to a method for predicting the remaining service life of machinery based on implicit dynamic hybrid drive. Background Technology

[0002] Rotating machinery (such as bearings, gears, and shafts) is widely used in industrial systems such as aerospace equipment, construction machinery, and energy equipment. Its health status directly determines the safety, operational reliability, and lifecycle management benefits of the equipment. Accurately predicting the remaining service life of rotating machinery allows for the early detection of potential failure risks, enabling the development of proactive maintenance strategies. This effectively avoids equipment downtime or accidents caused by sudden failures in rotating machinery, ensuring the stable and reliable completion of predetermined tasks. In recent years, thanks to advancements in sensor technology and monitoring methods, artificial intelligence-based methods for predicting the remaining service life of rotating machinery have gradually gained industrial application. These methods typically collect vibration signals from rotating machinery during operation and use convolutional neural networks, deep belief networks, and recurrent neural networks to automatically learn signal features, thereby predicting the remaining service life of components. This has become a crucial support for improving equipment reliability in current engineering practice.

[0003] However, most existing Remaining Useful Life (RUL) prediction methods based on monitoring data only model the statistical distribution or trend of observed signals, superficially fitting the changes in RUL within a data-driven feature space. They fail to effectively integrate the dynamic evolution laws of rotating components during operation, resulting in model prediction accuracy being significantly affected by data distribution and changes in operating conditions. This leads to poor generalization ability in real-world engineering environments, making long-term reliable life prediction difficult. A few methods attempt to establish explicit dynamic models, relying on prior assumptions and physical parameter settings to impose dynamic consistency constraints on the degradation process of rotating machinery. However, due to the highly nonlinear nature of the degradation mechanism of rotating machinery and the complex interactions of multiple operating conditions and variables, explicit dynamic modeling struggles to fully capture the nonlinear coupling relationship between changes in health status and multi-source features. This often results in insufficient extraction of dynamic features, missing key degradation information, and large model parameter uncertainties, making it difficult to adapt to actual operating conditions and hindering effective application in practice. Summary of the Invention

[0004] In view of this, the present invention provides a method for predicting the remaining service life of machinery based on implicit dynamics hybrid drive. This method utilizes vibration data from the operation of rotating machinery to automatically learn and mine implicit dynamic characteristics hidden in the degradation process, thereby achieving accurate prediction of the remaining service life of rotating machinery.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting the remaining service life of machinery based on implicit dynamics hybrid drive includes the following steps:

[0007] Vibration signals of machinery under various working conditions are acquired and a training set is constructed.

[0008] The training set is input into a pre-built mapping evolution model, the evolution feature vector is extracted through the mapping evolution model, and preliminary prediction results are output.

[0009] An implicit dynamic model is constructed based on the evolutionary feature vector and the preliminary prediction results.

[0010] Calculate the joint loss function consisting of the data feature loss of the mapping evolution model and the dynamic loss of the implicit dynamic model, and update the model parameters in the mapping evolution model and the implicit dynamic model through backpropagation.

[0011] The remaining service life of the rotating machinery under test is obtained by predicting the life of the test sample using the trained mapping evolution model.

[0012] Preferably, the steps of constructing the training set include:

[0013] The vibration signal is sliced ​​to obtain multiple signal samples. After modeling domain scale alignment of each signal sample, the signal samples are organized into tensor form. A remaining lifetime label is added to each signal sample.

[0014] Preferably, the step of constructing the training set further includes:

[0015] Add operating condition labels to each signal sample, distinguish the input data according to the operating condition labels, and divide the data into training set and test set so that the training set and test set cover all operating conditions.

[0016] Preferably, the mapping evolution model includes a feature projection module and a prediction module; the feature projection module, which is stacked in multiple layers, extracts features from the signal sample to obtain the evolution feature vector; the prediction module predicts the remaining lifetime based on the evolution feature vector.

[0017] Preferably, the data feature loss is:

[0018]

[0019] in, The data feature loss is defined as N, where N is the total number of samples and i is the sample variable. The preliminary prediction results are as follows. This is the remaining lifetime label for the current sample.

[0020] Preferably, the implicit dynamic model is:

[0021]

[0022] in, Let be the partial derivative of the initial RUL prediction result for sample i with respect to time t. For the preliminary RUL prediction results corresponding to sample i, apply to the features The partial derivatives, For the implicit nonlinear dynamic operator to be learned, Its parameter set.

[0023] Preferably, the implicit nonlinear dynamic operator for:

[0024]

[0025] in, The coefficient parameters to be learned. Let k be the order of the polynomial, and k be the order variable.

[0026] Preferably, the kinetic loss is:

[0027]

[0028] in, This refers to the dynamic loss.

[0029] Preferably, the joint loss function is:

[0030]

[0031] in, For the aforementioned joint loss, For the data feature loss, For the aforementioned dynamic loss, This is the weighting hyperparameter for dynamic loss.

[0032] Preferably, during the backpropagation, the dynamic loss hyperparameter is also updated in the following manner:

[0033]

[0034] Where t represents the current training round, and Let represent the dynamic loss weight hyperparameters for round t+1 and round t, respectively; and These represent the data feature loss and dynamic loss for round t, respectively. This is a hyperparameter.

[0035] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for predicting the remaining service life of machinery based on implicit dynamics hybrid drive, which has the following beneficial effects:

[0036] 1) A constraint mechanism based on dynamic characteristics is proposed. The model constrains the training process by constructing a dynamic loss and incorporating it into a joint loss function. This enables in-depth mining of the implicit dynamic degradation and evolution laws during actual equipment operation, embedding the dynamic characteristics represented in the data into the lifetime prediction model to achieve physical consistency constraints on the degradation process. This mechanism, in synergy with observation constraints based on data characteristics, not only improves the prediction accuracy of the model under diverse data distributions and complex operating conditions, but also significantly enhances the model's robustness to external operating condition disturbances and data noise, making the lifetime prediction results more stable and reliable.

[0037] 2) An implicit dynamics hybrid driving method was introduced, which directly maps the preliminary life prediction results to the implicit dynamics feature space, establishing a hybrid learning mechanism that combines data feature constraints and dynamic constraints. This mechanism can effectively integrate data features and degradation dynamics information, significantly enhancing the model's adaptability and generalization ability to diverse degradation modes of rotating machinery. It meets the high-precision life prediction requirements of practical application scenarios such as complex working conditions and variable loads, providing solid technical support for equipment health management and proactive maintenance. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of a method for predicting the life of rotating machinery based on implicit dynamics hybrid drive, provided in an embodiment of the present invention.

[0040] Figure 2 This is a structural diagram of the prediction model for implicit dynamics hybrid drive in an embodiment of the present invention.

[0041] Figure 3 This is a schematic diagram of the experimental results verifying the effectiveness of the present invention in an embodiment.

[0042] Figure 4 This is a schematic diagram comparing the performance of the present invention with existing deep learning methods in an embodiment of the invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] like Figure 1 and Figure 2 This invention discloses a method for predicting the remaining service life of machinery based on implicit dynamics hybrid drive, comprising the following steps:

[0045] S1: Acquire vibration signals of machinery under various operating conditions and construct a training set. Examples include rotating machinery such as bearings, gears, and shafts.

[0046] S2: Input the training set into the pre-built mapping evolution model, extract the evolution feature vector through the mapping evolution model, and output the preliminary prediction results.

[0047] S3: Construct an implicit dynamic model based on the evolutionary eigenvectors and preliminary prediction results.

[0048] S4: Calculate the joint loss function consisting of the data feature loss of the mapping evolution model and the dynamic loss of the implicit dynamic model, and update the model parameters in the mapping evolution model and the implicit dynamic model through backpropagation.

[0049] S5: The remaining service life of the rotating machinery under test is obtained by predicting the life of the test sample using a trained mapping evolution model.

[0050] In this embodiment, a hybrid driving approach combining data-driven and implicit dynamics is used to capture complex state changes and constrain the model with dynamic laws. This approach can accurately extract life-related features, avoid irrelevant interference, and ensure that the model conforms to the inherent evolution laws of machinery, thereby improving the accuracy and reliability of predicting the remaining service life of rotating machinery.

[0051] In this embodiment, S1 mainly includes data acquisition and preprocessing steps. First, continuous vibration signals are acquired from the target equipment under various actual working conditions. Accelerometers are placed at key parts of the rotating machinery to obtain vibration signal data under different speeds and loads. The vibration signal data used is shown in Table 1. The acquired signals are one-dimensional time series, and the length of each signal segment is set to... The sampling frequency is The duration of each collection is The original signal is denoted as ,in .

[0052] The acquired raw vibration signals are sliced ​​along the time axis using a sliding window method, and each sample signal is represented as follows:

[0053]

[0054] in, For the first Vibration signal segments of a sample For each time step, This represents the total number of samples.

[0055] To ensure training stability and comparability between different samples, scale alignment of the modeling domain is performed on all acquired signal segments. Specifically, for each signal segment... The mean-variance scaling method is used for processing:

[0056]

[0057] in, and These are the mean and standard deviation of all data in the training set, respectively. This is the scale-aligned input.

[0058] Each preprocessed sample is organized into a uniform tensor form, suitable for batch training. Let the batch size be... If the time step is l, then each batch can be represented as a three-dimensional tensor. The operating condition categories can be used to distinguish between the training and test sets. For example, data containing three operating conditions, with five sets for each condition, can be randomly seeded to select four sets for each of the three operating conditions (a total of twelve sets) for the training set, and one set for each of the remaining three operating conditions (a total of three sets) for the test set. The model can learn the behavior of all operating conditions, achieving joint modeling under different operating conditions.

[0059] Furthermore, for each signal sample, its remaining lifespan tag is recorded. This refers to the number of time steps remaining from the current acquisition time to the failure termination point. The input and label pair can then be denoted as... .

[0060] In this embodiment, S2 is the evolutionary mapping construction stage, which involves aligning the scale of the vibration signal segments. The input is fed into the evolutionary mapping model, and based on the residual minimization criterion within the evolutionary mapping domain, the parameters of all mapping subsets are adaptively tuned to achieve the initial mapping of the remaining lifetime. To enhance the evolutionary expressive power and fully reflect the complex evolutionary information in the signal, the model structure employs a combination of multi-layer fully connected mapping and nonlinear activation functions to achieve high-dimensional transformation and compression of signal evolutionary features.

[0061] Specifically, for each input sample First, the original signal segment is feature-projected through several linear transformation layers (i.e., fully connected layers). Taking the first... Taking layer feature mapping as an example, its output is expressed as follows:

[0062]

[0063] in, , and The first Layer weights and bias parameters, This is a non-linear activation function (such as ReLU). Through multi-layer stacking, deep feature vectors related to evolution are extracted step by step. ,Right now , This represents the number of feature extraction layers.

[0064] Subsequently, the final evolutionary characteristics will be represented. The input is fed into the RUL prediction mapping layer to achieve a direct mapping from the signal to the remaining lifetime. The prediction layer adopts a multi-layer fully connected structure, and its output is denoted as:

[0065]

[0066] in, For lifetime prediction mapping function, Its parameter set, For the first The RUL estimate for each sample, i.e., the prediction result.

[0067] In this embodiment, S3 is the implicit dynamics modeling stage. Within the dynamics feature space, a degradation dynamics modeling method based on implicit partial differential equations (PDEs) is used to achieve mechanism constraints on the degradation process of rotating machinery.

[0068] Specifically, let the preliminary RUL output sequence predicted by the evolutionary mapping model be... to regard it as time The implicit function, combined with the feature vector The following implicit dynamic relationship expression is established:

[0069]

[0070] in, Let RUL be the partial derivative with respect to time. Let RUL be the partial derivative with respect to the feature. For the implicit nonlinear dynamic operator to be learned, Its parameter set.

[0071] This expression demonstrates that the dynamic evolution of RUL is driven by the current health state, characteristic change trends, and implicit dynamic relationships, reflecting the complex degradation mechanisms under actual operating conditions. Nonlinear dynamic operators Based on the sample size and physical laws, a polynomial approximation is used as follows:

[0072]

[0073] in, The coefficient parameters to be learned. Let be the order of the polynomial.

[0074] In this embodiment, S4 mainly involves the automatic differentiation mechanism and loss function design. The automatic differentiation mechanism is fully utilized to calculate the partial derivatives of the implicit dynamics model output with respect to time and features, ensuring that dynamic consistency constraints are efficiently and accurately reflected in the model optimization process.

[0075] The implicit dynamic relationship expression is written as a PDE residual term:

[0076]

[0077] Specifically, the model predicts the remaining lifetime sequence as follows: The corresponding feature vector is Using automatic differentiation techniques, it is possible to efficiently obtain... Regarding time partial derivatives and partial derivatives with respect to features Then, substituting the above partial derivatives directly into the calculation of the residuals of the implicit dynamic equation, we obtain the dynamic loss term:

[0078]

[0079] This process eliminates the need for explicitly writing complex differentiation formulas, adapts to various model structures, and helps improve the method's versatility and engineering operability. It is then directly integrated into the joint loss function as a model dynamics consistency constraint.

[0080] In terms of loss function design, this invention adopts a hybrid optimization strategy driven by data feature loss and dynamic loss. By jointly minimizing the data feature loss and dynamic loss during the training process of the evolutionary mapping model, the model can automatically satisfy the dynamic consistency constraint while fitting the observed data, thereby improving the generalization ability and physical interpretability of RUL prediction.

[0081] The overall loss function is designed as follows:

[0082]

[0083] in, The data feature loss term measures the error between the model's predicted RUL and the true label, and is defined using mean squared error (MSE):

[0084]

[0085] This is a hyperparameter for the dynamic loss weights, used to balance the contribution ratios of data feature loss and dynamic loss.

[0086] To further improve the adaptability of model training and the stability of prediction results, this invention uses a dynamic adjustment mechanism for the dynamic loss weights. Specifically, during model training, Instead of using fixed constants, it dynamically and adaptively updates the loss terms based on their changing trends and relative magnitudes. The basic idea is: in each training iteration... Then, according to the current round and The value will automatically adjust the next round. The update formula is expressed as:

[0087]

[0088] in, To prevent small constants with a denominator of zero (such as...) ), This is the initial setting value at t=0, which can be set based on engineering experience. The minimum value is at Magnitude Can be set to or . Its value is much smaller than the normal value of the loss item, so we first use the normal value, such as... If the training process results in an explosion / non-convergence, the training level should be increased by two stages (e.g., This mechanism enables the dynamic loss term to gradually strengthen the constraint effect in the early stage of training, while gradually balancing the fit of observed data and dynamic consistency in the later stage of model training. This allows the overall optimization process to converge adaptively, effectively avoiding a single loss from dominating model training and ensuring the coordination and unity of physical consistency and prediction accuracy.

[0089] During model training, the joint loss function is optimized through backpropagation, and all parameters (including feature mapping layer parameters, RUL prediction layer parameters, and nonlinear dynamic operator parameters) are synchronously and iteratively updated under the goal of minimizing the loss. Automatic differentiation is employed throughout the training process, seamlessly supporting efficient gradient calculation of residuals from complex partial differential equations, effectively ensuring compliance with dynamic constraints and convergence stability during the implicit dynamic model training process. Upon completion of training, the optimal lifetime prediction model is finally obtained.

[0090] like Figure 3 and Figure 4 To further illustrate this point, an accelerated degradation experiment of a rolling bearing is used as an example. The data collected during the experiment, from operation to failure, is used to evaluate the effectiveness of the method of this invention. LDK UER204 bearings were used in the experimental data. Figure 3 The two parts represent the experimental results corresponding to two repeated experiments, which can show that the invention's prediction of the remaining service life is close to the true value.

[0091] The accelerated degradation experiment applied radial force, perpendicular to the direction of gravity, to the housing of the bearing under test using a hydraulic loading system. Vibration signals from the bearing were acquired by vertical and horizontal accelerometers at a sampling frequency of 25.6 kHz, recording 2560 data points every 0.1 seconds, with a data acquisition interval of 1.28 seconds every 60 seconds. Three different operating conditions were simulated by setting different rotational speeds (2100 rpm, 2250 rpm, and 2400 rpm) and radial forces (12 kN, 11 kN, and 10 kN). Fifteen bearings were tested. The test results are as follows: Figure 4 , Figure 4 (a), (b), and (c) in the figure correspond to the three working conditions mentioned above, respectively; (d) is the experimental result that combines the three working conditions.

[0092] The experiment compares the method of this invention with existing learning methods:

[0093] Method A constructs a model structure similar to that of the present invention, but does not introduce eigenvectors in the implicit dynamics modeling stage. Only the initial output sequence of RUL predicted by the evolutionary mapping model is used. Considered time The implicit function is used to establish the implicit dynamic relationship expression as follows: .

[0094] Method B retains the introduced feature vector However, instead of using the dynamic adjustment mechanism of dynamic loss weights in the step of constructing the loss function, a fixed constant is used instead, and the dynamic adaptive update is no longer based on the changing trend and relative size of each loss term.

[0095] Method C uses only dynamic loss when constructing the loss function, considering only dynamic feature constraints and excluding data feature constraints.

[0096] Method D uses only data feature loss for end-to-end mapping RUL when constructing the loss function, without dynamic consistency constraints.

[0097] Method E constructs a model structure similar to that of the present invention, but no implicit operators are used in the dynamic modeling stage. Instead of using a constraint, an exponential decay model was chosen. The explicit dynamic equations serve as physical consistency constraints.

[0098] Total loss:

[0099]

[0100] in, This represents the consistency loss of the decay function model. It is an adjustable hyperparameter.

[0101] In the experimental results, this invention uses the root mean square error (RMSE) and the symmetric mean absolute percentage error (sMAPE) to evaluate the performance of the proposed model. The calculation formulas are as follows:

[0102]

[0103]

[0104] Where n represents the amount of data, and These represent the predicted value and the actual value, respectively.

[0105] The experimental setup is shown in Table 1, and the learning scenarios for the constructed implicit dynamics hybrid model are shown in Table 2. The experimental results are summarized in Table 3.

[0106] Table 1. Bearing operating condition configurations in the examples

[0107]

[0108] Table 2. Learning scenarios constructed in the examples

[0109]

[0110] Table 3. Performance evaluation results of the method of the present invention

[0111]

[0112] Figure 4 In the figure, (a), (b) and (c) represent the RMSE results under different working conditions, and (d) is the SMAPE result that integrates the three working conditions. The experimental results show that the present invention has good health assessment performance in the learning process, and the method of the present invention has obvious advantages over other methods.

[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A mechanical residual useful life prediction method based on implicit dynamics hybrid driving, characterized in that, The method comprises the following steps: obtaining vibration signals of a mechanical device under various working conditions and constructing a training set; inputting the training set into a pre-constructed mapping evolution model, extracting an evolution feature vector through the mapping evolution model, and outputting a preliminary prediction result; constructing an implicit dynamics model according to the evolution feature vector and the preliminary prediction result; the implicit dynamics model is: ； wherein, is the partial derivative of the RUL preliminary prediction result corresponding to sample i with respect to time t, is the partial derivative of the RUL preliminary prediction result corresponding to sample i with respect to feature , is the implicit nonlinear dynamics operator to be learned, is its parameter set; the implicit nonlinear dynamics operator is: ； wherein, is a coefficient parameter to be learned, is a polynomial order, k is an order variable; calculating a joint loss function composed of a data feature loss of the mapping evolution model and a dynamics loss of the implicit dynamics model, and updating model parameters in the mapping evolution model and the implicit dynamics model through back propagation; performing life prediction on a to-be-tested sample through the trained mapping evolution model to obtain a remaining useful life of the to-be-tested rotating mechanical device.

2. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 1, wherein, The step of constructing the training set comprises: slicing the vibration signals to obtain a plurality of signal samples; aligning model domain scales of the signal samples and arranging the signal samples into a tensor form; adding a remaining life label to each signal sample.

3. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 2, characterized in that, The step of constructing the training set further comprises: adding a working condition label to each signal sample, distinguishing input data according to the working condition label, and dividing the training set and the test set, so that the training set and the test set cover all working conditions.

4. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 2, characterized in that, The mapping evolution model comprises a feature projection module and a prediction module; the feature projection module is stacked in multiple layers to extract features of the signal samples, and the evolution feature vector is obtained; the prediction module predicts the remaining life according to the evolution feature vector.

5. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 1 or 4, characterized in that, The data feature loss is: ； wherein, is the data feature loss, N is the total number of samples, i is the sample variable, is the preliminary prediction result, is the remaining life label corresponding to the current sample.

6. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 1, wherein, The dynamics loss is: ； wherein, is the kinetic loss.

7. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 1, characterized in that, The joint loss function is: ； wherein, is the joint loss, is the data feature loss, is the dynamics loss, is a dynamics loss weight hyperparameter.

8. The mechanical residual useful life prediction method based on implicit dynamics hybrid driving according to claim 7, characterized in that, When the back propagation is performed, the dynamics loss hyperparameter is also updated, and the updating mode is: ； where t denotes the current training epoch, and denote the dynamics loss weight hyperparameters for the (t+1)th and tth epochs, respectively; and are the data feature loss and dynamics loss for the tth epoch, respectively, is a hyperparameter.

Images