A method for predicting the residual life of a roller screw pair

Abstract

The application relates to a roller screw pair residual life prediction method. The method comprises the following steps: acquiring a radial vibration signal in the operation process of a control rod drive mechanism roller screw pair, and pre-processing the radial vibration signal to obtain a plurality of one-dimensional vibration sub-samples; performing time-frequency conversion processing on each one-dimensional vibration sub-sample to generate a two-dimensional time-frequency distribution graph representing the time-frequency energy distribution of the signal, thereby obtaining a time-frequency graph stream; moving the time-frequency graph stream according to a preset time window, and performing time sequence embedding to obtain a target time-frequency distribution sequence; inputting the target time-frequency distribution sequence into a pre-trained residual life estimation model based on a dynamic time convolution network to obtain a residual life prediction value; the residual life estimation model comprises a dynamic convolution layer, an attention mechanism and a full connection layer; and the dynamic convolution layer adopts a dynamic activation function. The method can realize stable and accurate estimation of the residual life of the control rod drive mechanism roller screw pair.

Description

Technical Field

[0001] This invention belongs to the field of nuclear power plant equipment maintenance technology, and in particular relates to a method for predicting the remaining life of a roller screw pair. Background Technology

[0002] With the development of technologies in the fields of safe operation and maintenance of nuclear power plants and health management of rotating machinery, data-driven remaining life prediction technology has emerged. This technology relies on a large amount of historical monitoring data to build equipment degradation models, extract time-domain, frequency-domain, or time-frequency multi-dimensional features of vibration signals, and then trains deep learning networks to achieve life regression prediction, providing important support for the formulation of preventive maintenance plans for equipment.

[0003] In traditional technologies, for the life management of critical transmission components in nuclear reactors, such as the control rod drive mechanism roller screw pair, the vibration monitoring data of the roller screw pair during operation is first collected by sensors. Then, conventional time-domain statistics or frequency-domain spectral features such as root mean square value and peak value are extracted. Subsequently, a life prediction framework is built using mainstream time-series prediction models such as recurrent neural networks and convolutional neural networks to assess the wear state of the roller screw pair and avoid risks such as mechanism jamming and reactive control failure.

[0004] However, the above methods are subject to dual limitations of nuclear safety regulations and equipment operating conditions in data acquisition. High-frequency, multi-channel continuous monitoring is difficult to achieve. The cost of data acquisition throughout the entire life cycle is high and the signal-to-noise ratio of early degradation signals is extremely low, forming a typical small sample dilemma. Traditional generative data augmentation methods are prone to pattern collapse, and the distribution of synthetic samples does not match that of real data. In addition, conventional features cannot capture weak early degradation information, and the ability to represent non-stationary signals is insufficient. Mainstream prediction models have defects such as gradient vanishing or fixed receptive fields, making it difficult to model the long-term dependence between early features and final failure. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for predicting the remaining life of roller screw pairs that can achieve accurate and stable prediction of the remaining life of roller screw pairs under small sample conditions by combining generative data augmentation and adaptive feature extraction to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for predicting the remaining life of a roller screw pair, including:

[0007] The radial vibration signal of the roller screw pair of the control rod drive mechanism is acquired during operation, and the radial vibration signal is preprocessed to obtain multiple one-dimensional vibration sub-samples.

[0008] The time-frequency transformation of each one-dimensional vibration sub-sample is performed to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, thus obtaining a time-frequency map stream;

[0009] The time-frequency graph stream is embedded into the time-series by moving the window according to the preset time period to obtain the target time-frequency distribution sequence.

[0010] The target time-frequency distribution sequence is input into a pre-trained lifetime estimation model based on a dynamic temporal convolutional network to obtain the lifetime prediction value. The lifetime estimation model includes dynamic convolutional layers, attention mechanisms, and fully connected layers. The dynamic convolutional layers use dynamic activation functions.

[0011] In one embodiment, the radial vibration signal is preprocessed to obtain multiple one-dimensional vibration sub-samples, including:

[0012] The radial vibration signal is sliced ​​using a sliding window with no overlap or a specified overlap rate according to a preset number of sampling points, resulting in multiple one-dimensional sub-samples.

[0013] The amplitude of each one-dimensional subsample is normalized by using maximum-minimum normalization to obtain normalized vibration samples;

[0014] The normalized vibration samples are denoised, and outliers are removed from the denoised samples to obtain compensated vibration samples.

[0015] The sampling time information of the compensated vibration samples is labeled to form a time-stamped one-dimensional vibration sub-sample.

[0016] In one embodiment, time-frequency transformation processing is performed on each one-dimensional vibration sub-sample to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, resulting in a time-frequency map stream, including:

[0017] Based on the Morlet wavelet basis function, continuous wavelet transform is performed on each one-dimensional vibration sub-sample by scaling and translation parameters to obtain the scale-time shift coefficient matrix.

[0018] The scale-time shift coefficient matrix is ​​sequentially subjected to amplitude mapping, logarithmic compression, and grayscale processing, and then resampled at a preset resolution to obtain a set of time-frequency graphs.

[0019] Batch normalization and edge clipping are performed on the time-frequency graph set to obtain the time-frequency graph stream.

[0020] In one embodiment, the remaining lifetime estimation model based on a dynamic temporal convolutional network is trained using the following method:

[0021] Based on the time-frequency graph flow of the control rod drive mechanism life test platform, a pre-trained U-Net-based diffusion model is applied for data augmentation to generate a synthetic time-frequency graph with the same distribution as the time-frequency graph flow. The time-frequency graph flow and the synthetic time-frequency graph are then merged to obtain a time-frequency graph data augmentation set.

[0022] Each time-frequency image sample in the time-frequency image data augmentation set is labeled with its remaining useful life, and the remaining useful life labels are normalized to obtain a labeled augmented sample set; the labeled augmented sample set includes a training set and a validation set;

[0023] Each time-frequency image sample in the training set is converted into a feature vector, and each feature vector is embedded into a time series input of a preset time length based on a time-shifting window to obtain the sequence input set and its corresponding predicted target label.

[0024] The remaining lifetime estimation model based on the dynamic temporal convolutional network is trained and its hyperparameters are optimized using a sequence input set and a labeled augmented sample set.

[0025] In one embodiment, the remaining lifetime estimation model based on a dynamic temporal convolutional network is trained and its hyperparameters optimized using a sequence input set and a labeled augmented sample set to obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network, including:

[0026] The key structural hyperparameters of the remaining lifetime estimation model based on dynamic temporal convolutional networks are optimized by grid search to obtain multiple combinations of hyperparameters that can be traversed within a predefined search range; the key structural hyperparameters include the number of residual blocks and the size of the time window.

[0027] For each hyperparameter combination, the remaining lifetime estimation model based on the dynamic temporal convolutional network is trained according to the sequence input set and its corresponding predicted target label. The root mean square error and mean absolute error of the remaining lifetime prediction value of the trained dynamic temporal convolutional network-based remaining lifetime estimation model and the normalized remaining lifetime label are calculated using the validation set to obtain the performance evaluation index.

[0028] Choose the hyperparameter combination that minimizes the performance evaluation index as the final model configuration;

[0029] The remaining lifetime estimation model based on the dynamic temporal convolutional network was trained using all the training data of the final model configuration and labeled augmented sample set, and the loss was calculated based on the remaining lifetime prediction and the normalized remaining lifetime label.

[0030] The model parameters of the remaining lifetime estimation model based on the dynamic temporal convolutional network are updated iteratively based on the loss until convergence, thus obtaining the remaining lifetime estimation model based on the dynamic temporal convolutional network.

[0031] In one embodiment, the remaining lifetime estimation model based on a dynamic temporal convolutional network obtains the remaining lifetime prediction value through the following method:

[0032] Based on hyperparameter combination, causal dilated convolution of residual blocks is used to extract multi-scale temporal features from the sequence input set, and the hidden representation of the time step corresponding to each time shift window is obtained.

[0033] Global average pooling is performed on each channel based on the hidden representation, and the result of global average pooling is input into the hyperfunction network to obtain the segmented activation function parameter vector; the hyperfunction network includes two fully connected layers and intermediate ReLU and normalization layers;

[0034] Construct a parameterized piecewise linear activation function based on the piecewise activation function parameter vector, and obtain the activation features corresponding to the hidden representation based on the piecewise linear activation function;

[0035] The activation features are added to the hidden representation through residual connections to obtain the hidden state sequence, and the hidden state sequence is then time-channel weighted through an attention mechanism to obtain the weighted feature representation.

[0036] Based on a fully connected layer, the weighted feature representation is mapped to a normalized remaining lifetime prediction.

[0037] In one embodiment, the hidden state sequence is time-channel weighted using an attention mechanism to obtain a weighted feature representation, including:

[0038] The attention score vector is calculated by linearly transforming the hidden state sequence.

[0039] The attention score vector is subjected to Softmax normalization to obtain the attention weight sequence;

[0040] The hidden state sequence is weighted and summed based on attention weights to obtain a weighted feature representation.

[0041] The aforementioned method for predicting the remaining life of the roller screw pair acquires the radial vibration signal of the roller screw pair and preprocesses it into one-dimensional vibration sub-samples. Then, it performs time-frequency transformation processing on each sub-sample to generate a two-dimensional time-frequency distribution map that can simultaneously characterize the time-frequency energy distribution, thus overcoming the difficulty of feature extraction for non-stationary signals. The obtained time-frequency map stream is embedded in time sequence according to a preset time-shifting window to construct a target time-frequency distribution sequence with time context information. This sequence is input into a pre-trained remaining life estimation model composed of a dynamic convolutional layer with a dynamic activation function, an attention mechanism, and a fully connected layer. The dynamic activation function adaptively captures the dynamic features of different life stages, the attention mechanism focuses on key time step information, and the dynamic temporal convolutional structure models long-term dependencies, thereby achieving stable and accurate estimation of the remaining life of the roller screw pair of the control rod drive mechanism under small sample conditions. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the method for predicting the remaining life of the roller screw pair according to the present invention.

[0044] Figure 2 This is a schematic diagram of the training process of the remaining lifetime estimation model based on dynamic temporal convolutional network of the present invention.

[0045] Figure 3 This is a step-by-step flowchart for step S204. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] In one embodiment, such as Figure 1 As shown, a method for predicting the remaining life of a roller screw pair is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0048] S101. Obtain the radial vibration signal during the operation of the control rod drive mechanism roller screw pair, and preprocess the radial vibration signal to obtain multiple one-dimensional vibration sub-samples.

[0049] Indicatively, when acquiring radial vibration signals, the acquisition scheme is determined based on the operating mechanism and wear characteristics of the control rod drive mechanism (CRDM). Specifically, during CRDM operation, the rotor drives the rollers to move axially up and down on the lead screw assembly. The frictional force mainly acts in the horizontal direction. Therefore, the radial vibration signal contains the richest information on the wear and degradation of the roller-lead screw pair. Axial and tangential signals can be used as auxiliary references, so radial vibration signals are prioritized for acquisition. The acquisition equipment is a radiation-resistant, high-temperature and high-pressure resistant nuclear-grade piezoelectric accelerometer, such as the INV982X general-purpose piezoelectric accelerometer. When deploying the sensor, it must be fixed at a 90° angle to the axial and tangential directions of the CRDM housing to ensure that the signal acquisition direction is accurately aligned with the direction of frictional force, while avoiding data distortion caused by sensor displacement due to equipment vibration. Furthermore, the sampling system complies with the requirements of nuclear safety regulations regarding sampling rate and signal-to-noise ratio. The sampling rate must satisfy the Nyquist sampling theorem to ensure that it can capture the weak wear vibration components during the low-speed operation of the roller screw pair. Optionally, the sampling rate is ≥20kHz and the signal-to-noise ratio is ≥60dB to suppress electromagnetic interference and background noise in high-temperature, high-pressure, and high-radiation environments.

[0050] Furthermore, the collected radial vibration signals undergo preprocessing, consisting of two core steps: amplitude normalization and sliding window slicing. Amplitude normalization aims to eliminate dimensional differences in vibration amplitude under different operating conditions, preventing gradient vanishing or exploding problems caused by excessive amplitude spans during subsequent model training. Sliding window slicing aims to convert continuous long-sequence vibration signals into multiple independent one-dimensional vibration sub-samples, avoiding information redundancy or loss of local features due to excessive signal length. Optionally, a non-overlapping sliding window method is used for slicing. The window length needs to be determined based on the operating cycle and wear characteristic cycle of the roller screw pair, and must cover the vibration signal corresponding to at least one complete rotation cycle of the roller screw pair to ensure that each sub-sample can reflect the complete local wear state of the roller screw pair within that time period. For example, if the rotation cycle of the roller screw pair is 0.1s and the sampling rate is 20kHz, then each cycle contains 2000 sampling points. In this case, to balance computational efficiency and feature integrity, the sliding window length can be set to 1024 sampling points, ensuring that each sub-sample contains vibration features of more than half a cycle, avoiding information breakage between samples. During the slicing process, starting from the initial position of the vibration signal, the slide window is moved by a distance of one slide window length each time until the entire signal sequence is processed. The resulting multiple one-dimensional vibration sub-samples need to be numbered in order of acquisition time to provide temporal dimension correlation support for subsequent time-frequency transformation and time series modeling, ensuring that the timestamp of each sub-sample corresponds one-to-one with the actual running time of the roller screw pair.

[0051] S102. Perform time-frequency transformation processing on each one-dimensional vibration sub-sample to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, thus obtaining a time-frequency map stream.

[0052] Since the vibration signal of the CRDM roller screw pair is a typical non-stationary signal, its frequency characteristics change dynamically with time. Simply using time-domain statistics such as root mean square value and peak value, or frequency-domain spectral analysis, cannot fully characterize the time-frequency evolution of the wear process from early weak degradation to late severe wear. Illustratively, continuous wavelet transform (CWT) is used for time-frequency transformation processing. This transform can decompose the signal simultaneously in the time and frequency domains, accurately capturing the frequency variation over time and providing richer time-frequency joint information for wear feature extraction. Specifically, the wavelet function type is determined; the Morlet wavelet function can be selected, as it possesses good time-frequency locality and can effectively focus on specific frequency components in the signal, making it particularly suitable for capturing the vibration frequency shift and energy changes caused by roller screw pair wear. The expression for the Morlet wavelet function is: .in, Pi is a constant, with a value of approximately 3.1416. The center frequency of the Morlet wavelet is determined based on the design operating speed of the CRDM roller screw pair. For example, if the design speed of the roller screw pair is 60 r / min, its characteristic vibration frequency is approximately 1 Hz. It can be set to 5-10 to enhance the sensitivity to this characteristic frequency and harmonic components; It is a time variable, and its unit is consistent with the sampling time of the vibration signal; It is a natural constant with a value of approximately 2.7183.

[0053] Furthermore, the continuous wavelet transform is... .in, These are wavelet transform coefficients, and their absolute values ​​characterize the signal at different scales. ,Location The larger the coefficient of the time-frequency energy density at a given time-frequency point, the stronger the vibration energy at that point, which usually corresponds to the characteristic vibration caused by the wear of the roller screw pair. The scaling parameter determines the width of the wavelet function. Inversely proportional to frequency, The larger the value, the wider the wavelet function, corresponding to the low-frequency components of the analyzed signal, such as the low-frequency vibrations during the early slight wear of a ball screw pair. The smaller the value, the narrower the wavelet function, corresponding to the high-frequency components of the analyzed signal, such as the high-frequency impact vibration of a roller screw pair under severe late-stage wear. In actual calculations, the value needs to be set according to the frequency range of the vibration signal. The range of values ​​is determined to ensure that all characteristic frequencies of the roller screw pair are covered from normal operation to failure. The position parameter determines the center time position of the wavelet function. The value of is completely consistent with the time series of the one-dimensional vibration subsample, ensuring complete coverage of the time domain range of the entire subsample and omitting vibration information at any time point; Given a one-dimensional vibrational subsample as input, This is the time variable for the subsample, and its value range corresponds to the sampling duration of the subsample; wavelet function The complex conjugation operation ensures the energy conservation of wavelet transform coefficients, avoiding feature distortion caused by energy loss during the transform process; integral sign Represents the time variable Integrating over the entire range of values, since the signal has a finite length, the integration range can be reduced to the time span of a subsample to improve computational efficiency.

[0054] Through the continuous wavelet transform described above, each one-dimensional vibrational subsample is converted into a two-dimensional time-frequency distribution map, with the horizontal axis representing the time-corresponding location parameters. The vertical axis represents frequency. , by scale parameter The conversion yields, where The characteristic frequencies of the Morlet wavelet are given by... The calculations show that the grayscale value of each pixel in the image corresponds to a wavelet transform coefficient. The absolute value of the grayscale value indicates that the higher the grayscale value, the stronger the energy at that time-frequency point. The two-dimensional time-frequency distribution maps corresponding to all one-dimensional vibration sub-samples are arranged sequentially according to the time order of the sub-samples to form a time-frequency graph stream. This time-frequency graph stream completely preserves the evolution of the frequency characteristics of the vibration signal over time. For example, in the early wear stage of the roller screw pair, the energy in the low-frequency region of the time-frequency graph slowly increases; after entering the middle wear stage, energy concentration points begin to appear in the mid-frequency region; in the late wear stage, obvious impact energy peaks appear in the high-frequency region.

[0055] S103. Move the window according to the preset time and embed the time-frequency graph stream into the time sequence to obtain the target time-frequency distribution sequence.

[0056] Each time-frequency distribution map in the time-frequency graph stream only reflects the local wear state of the roller screw pair at a certain moment. The remaining life prediction of the CRDM roller screw pair is essentially based on the temporal evolution trend of the wear process, requiring the integration of continuous time-frequency features over a period to construct a complete wear trend. For example, by using a preset time-shifting window to perform temporal embedding on the time-frequency graph stream, the discrete time-frequency distribution maps are transformed into a target time-frequency distribution sequence with temporal correlation, providing structured input for the dynamic temporal convolutional network to model long-sequence dependencies. Roller screw pair wear is a continuous and gradual process, and the wear states at adjacent moments are strongly correlated. The time-shifting window needs to cover a sufficient number of continuous time-frequency distribution maps to include a complete wear trend segment. Simultaneously, the size of the time-shifting window needs to balance computational efficiency and feature completeness; an excessively large window increases model computational complexity, while an excessively small window cannot capture long-term temporal dependencies. Let the time-shifting window size be... That is, each target time-frequency distribution sequence contains Zhang's continuous time-frequency distribution plot, with a window movement step size of [missing information]. That is, the number of time-frequency distribution maps that the window moves along the time axis of the time-frequency graph each time, usually the step size. Set as 50%-80% to balance data utilization and computational efficiency.

[0057] Specifically, the total length of the time-frequency graph stream is determined. The total number of time-frequency distribution maps contained in the time-frequency map stream is calculated by extracting maps from the first to the last time-frequency distribution map, starting from the beginning of the stream. The time-frequency distribution plots form the first target time-frequency distribution sequence. Subsequently, the window moves along the time axis. Step, extract the first To the Zhang time-frequency distribution plots are used to form the second target time-frequency distribution sequence; this process is repeated until the end of the window exceeds the total length of the time-frequency plot stream. The final number of target time-frequency distribution sequences is ,in This indicates a floor operation, ensuring that the last sequence still contains the floor function. Zhang's continuous time-frequency distribution plot, if the window ends are insufficient Zhang can be supplemented by filling in the preceding time-frequency distribution map. In practical applications, it is necessary to ensure the length of the time-frequency map stream. satisfy To avoid feature distortion caused by padding, each target's time-frequency distribution sequence must strictly follow the chronological order, which can be represented as follows: ,in ( ) represents the first digit in the sequence. Zhang's time-frequency distribution map, and The corresponding time is earlier than , Earlier This process is repeated to ensure that the time-frequency characteristics within the sequence are arranged in chronological order of the wear time of the roller screw pair. This sequence not only preserves the time-frequency energy characteristics of each time-frequency distribution map but also constructs a dynamic evolution relationship of the wear state through temporal arrangement, as shown in the sequence. to The time-frequency characteristics of the changes can reflect The trend of the roller screw pair from slight wear to moderate wear, or from moderate wear to severe wear, within a time segment.

[0058] S104. Input the target time-frequency distribution sequence into the pre-trained remaining lifetime estimation model based on a dynamic temporal convolutional network to obtain the remaining lifetime prediction value. The remaining lifetime estimation model includes a dynamic convolutional layer, an attention mechanism, and a fully connected layer. The dynamic convolutional layer uses a dynamic activation function.

[0059] The pre-trained remaining lifetime estimation model uses a Dynamic Temporal Convolutional Network (DTCN) as its core architecture. Through the synergistic effect of dynamic convolutional layers, attention mechanisms, and fully connected layers, it extracts, weights, and regresses wear features from the target time-frequency distribution sequence. The dynamic convolutional layer is the core feature extraction module of the model. Based on an improvement to the causal dilated convolution of the Temporal Convolutional Network (TCN), a Dynamic Reactivation Function (DReLU) is introduced to replace the traditional ReLU, adapting to the dynamic changes in time-frequency features during the wear process of the roller screw pair. The role of the causal dilated convolution is to capture the long-term temporal dependence of the target time-frequency distribution sequence; its convolution operation formula is as follows: .in, For convolution operation in the output sequence The result of each time step represents the feature output of that time step; The first convolution kernel There are weight parameters, and the kernel size is . ; For the input target time-frequency distribution sequence at the th The feature vector for each time step is obtained by flattening the pixel values ​​of the corresponding time-frequency distribution map, and the dimension is consistent with the number of pixels in a single time-frequency distribution map. The expansion factor is set using an exponentially increasing method, such as the expansion factor for the first-level residual block. , second floor , No. layer By increasing Without increasing the number of kernel parameters, the receptive field of the convolution kernel can be significantly expanded, enabling the network to capture the feature associations between early and late time steps in the target time-frequency distribution sequence. This solves the problem of limited receptive field and inability to model long sequence dependencies in traditional convolution kernels; summation symbol This represents the summation of the products of the convolution kernel weights and their corresponding input features, achieving linear combination of features and extraction of local associations. The Dynamic Activation Function (DReLU) is used to perform a non-linear transformation on the output features of the causal dilated convolution; its expression is... ,in, The input feature vector of the dynamic convolutional layer is the output of the causal dilated convolution. ; This is a dynamic slope parameter used to adjust the slope of the linear transformation of the activation function; This is a dynamic offset parameter used to adjust the activation threshold; This means taking the larger of the input value and 0, ensuring that the output is non-negative. and It is not a fixed value, but is dynamically calculated by the hyperfunction. For example, the hyperfunction performs global average pooling on the feature map output by the causal dilated convolution to obtain the feature mean vector. , Dimensions and number of channels in the feature map Consistent, this vector reflects the global energy distribution of the feature map; Input two fully connected layers, the output dimension of the first layer is The activation function is ReLU, which achieves non-linear transformation and dimensionality expansion of the features; the output dimension of the second layer is... Then through the Sigmoid function Normalize the output values ​​to [0,1] to obtain the normalized vector. ;Will Split into two dimensions subvectors and , respectively as and The initial residual is introduced by setting a preset initial value. and ,pass , The final dynamic parameters are calculated. When the input feature is an early, weak wear signal, Enlarge Decrease, make Enlarge Reduces and enhances the activation ability of weak features; when the input feature is a late-stage strong wear signal, Reduce Increase, make Reduce It enhances and suppresses noise interference, thereby adaptively matching the feature extraction requirements of different wear stages.

[0060] Attention mechanisms are used to weight the temporal features output by dynamic convolutional layers, highlighting the feature contributions of key time steps. The weight calculation formula is as follows: .in, For the target time-frequency distribution sequence, the first... The attention weights for each time step feature, the sum of the weights for all time steps satisfying... This ensures the normalization of weights; Here is the weight matrix for the attention mechanism, with dimension 1. , The dimension of the output feature of the dynamic convolutional layer is the product of the number of pixels and the number of channels after flattening a single time-frequency distribution map; For the dynamic convolutional layer in the 1st The output feature vector at each time step has a dimension of . It contains all the feature information of the time-frequency plot at that time step; Let be the bias vector of the attention mechanism, with dimension . This is used to adjust the baseline value for weight calculation; The natural constant is denominator For the normalization term of the Softmax function, ensure The value is in the range [0,1]. Using this formula, the model will automatically assign higher values ​​to time steps containing key wear features. Assigning lower values ​​to time steps with flat features Then, the feature vector of each time step With corresponding weights Multiply and sum to obtain the weighted global feature vector. , dimension This vector encapsulates the key temporal features of the entire target time-frequency distribution sequence.

[0061] The fully connected layer is responsible for processing the global feature vector. The mapping to remaining lifetime predictions is divided into two layers. The first layer is a feature compression layer, which... Vie Compress to Dimension, the activation function is ReLU, and the expression is: ,in This is the weight matrix of the first fully connected layer. For bias vectors, The second layer is the prediction output layer, which performs a nonlinear transformation. Vie Mapped to 1D output, i.e., normalized remaining lifetime label The expression is ,in This is the weight matrix of the second fully connected layer. For bias vectors, The value range is [0,1], corresponding to the roller screw pair starting from the initial state. until the end of life The lifetime status. Finally, label inverse normalization is used to... Convert to actual remaining life value The inverse normalization formula is: ,in For the total design life of the CRDM roller screw pair, for example, if Hours, model output ,but The remaining life of the current roller screw pair is 4800 hours.

[0062] In the aforementioned method for predicting the remaining life of the roller screw pair, multiple one-dimensional vibration sub-samples are obtained by preprocessing the radial vibration signal. This eliminates the differences in vibration amplitude dimensions caused by different operating conditions, avoids gradient shift caused by inconsistent data scales during model training, and suppresses electromagnetic interference and background noise under high temperature, high pressure, and high radiation environments, while preserving the true vibration characteristics corresponding to the wear of the roller screw pair. Dividing the continuous signal into independent one-dimensional sub-samples ensures that each sub-sample covers the complete local wear cycle, avoiding information redundancy caused by excessively long individual samples and feature loss caused by insufficient sampling points. The one-dimensional vibration sub-samples are transformed into a two-dimensional time-frequency distribution map representing the time-frequency energy distribution through time-frequency transformation. The time-frequency distribution map can intuitively present the dynamic changes of signal frequency over time. This joint time-frequency feature can accurately capture early weak degradation information that is difficult to identify by traditional methods, providing rich representational dimensions for the model to extract full-life-cycle wear characteristics. By embedding the time-frequency graph stream into a time sequence through a preset time-shifting window, a target time-frequency distribution sequence is formed. This integrates discrete single time-frequency graphs into feature segments with temporal correlation, ensuring that the model can capture the correlation between early weak degradation features and late failure. The remaining lifetime estimation model based on a dynamic temporal convolutional network replaces the traditional static ReLU with a dynamic activation function. It can dynamically adjust the activation response according to the characteristics of different wear stages of the roller screw pair, avoiding the "one-size-fits-all" treatment of dynamic features throughout the entire lifespan by the static activation function and improving the adaptability of feature extraction. The attention mechanism can automatically identify and weight key time step features that contribute highly to the lifetime prediction, reducing the interference of redundant information on the model and focusing on the core degradation patterns. The causal dilated convolutional structure based on the temporal convolutional network can expand the receptive field of the convolutional kernel while avoiding the gradient vanishing / exploding problem, effectively modeling the long sequence dependency of the CRDM roller screw pair throughout its entire lifespan, and is more adaptable to the characteristics of the roller screw pair of "low-speed operation, slow wear, and long life cycle", significantly improving the accuracy and stability of the remaining lifetime prediction.

[0063] In one embodiment, the radial vibration signal is preprocessed to obtain multiple one-dimensional vibration sub-samples, including:

[0064] S11. The radial vibration signal is sliced ​​according to the preset number of sampling points in a sliding window with no overlap or a specified overlap rate to obtain multiple one-dimensional sub-samples.

[0065] The determination of the preset number of sampling points needs to be combined with the operating characteristics of the control rod drive mechanism's roller screw pair and the signal analysis requirements. This ensures that each one-dimensional sub-sample can completely contain the vibration characteristics corresponding to at least one rotation cycle of the roller screw pair or one control rod lifting action, avoiding the loss of local wear information due to too few sampling points or information redundancy due to too many. The sliding window operates in two modes: non-overlapping and specified overlap rate. In non-overlapping mode, the distance the window slides each time is equal to the preset number of sampling points, and adjacent one-dimensional sub-samples have no overlapping area on the time axis. This is suitable for scenarios with smooth changes in vibration signal characteristics and sufficient data, reducing computational redundancy. In specified overlap rate mode, the window sliding distance is less than the preset number of sampling points. The overlap rate, i.e., the ratio of the number of sampling points in the overlapping part of adjacent samples to the preset number of sampling points, can be set to 30%-70%. This preserves the temporal correlation between adjacent samples and avoids feature breakage caused by window jumps. It is suitable for the early wear stage of the roller screw pair, improving the completeness of subsequent feature extraction through feature complementarity in the overlapping area.

[0066] To illustrate, the specific execution flow of the slicing process is as follows: taking the starting sampling point of the radial vibration signal as the starting point of the window, a signal segment with a length equal to the preset number of sampling points is extracted as the first one-dimensional sub-sample. The window sliding distance is calculated according to the selected overlap rate, i.e., sliding distance = preset number of sampling points × (1 - overlap rate). The window is shifted along the time axis by the corresponding distance, and a new signal segment is extracted as the next one-dimensional sub-sample. This process is repeated until the end of the window exceeds the total sampling length of the radial vibration signal. The resulting multiple one-dimensional sub-samples need to be numbered sequentially according to the extraction time to ensure that their timing relationship is consistent with the actual running time of the roller screw pair.

[0067] S12. Max-min normalization is used to normalize the amplitude of each one-dimensional subsample to obtain normalized vibration samples.

[0068] Max-min normalization eliminates the difference in amplitude dimensions between different one-dimensional subsamples, thus avoiding prediction bias caused by fluctuations in the operating conditions of the roller screw pair. ,in, This represents the original vibration amplitude data in a one-dimensional subsample. Its value directly reflects the vibration intensity of the roller screw pair within the corresponding time period of that subsample. The more severe the wear in a time period, the higher the vibration intensity. The larger the fluctuation range and peak value, the better; This represents the minimum value of all original vibration amplitudes in the one-dimensional subsample. The maximum value of all original vibration amplitudes in the one-dimensional subsample is required to be calculated by traversing all sampling points of a single subsample to ensure that the full amplitude range of the subsample is covered. The normalized vibration amplitude is strictly constrained within the [0,1] interval, ensuring a unified comparison benchmark for vibration amplitudes of different one-dimensional subsamples. Specifically, max-min normalization is performed separately for each one-dimensional subsample to ensure that the amplitude variation range of each subsample is fully mapped to the [0,1] interval, thus fully preserving its internal vibration characteristic details.

[0069] S13. Denoise the normalized vibration samples and remove outliers from the denoised samples to obtain compensated vibration samples.

[0070] Normalized vibration samples may still contain two types of interference information: one is environmental noise, such as electromagnetic interference under high temperature and high pressure, and vibration coupling noise from other components of the equipment; the other is outliers, such as amplitude changes caused by instantaneous sensor failures and sampling point jumps caused by data transmission errors. Their influence needs to be eliminated through noise reduction and outlier removal. Indicatively, wavelet thresholding is the preferred method for denoising. Specifically, a wavelet basis matching the vibration signal characteristics, such as the Db4 or Sym8 wavelet, is selected. The normalized vibration sample undergoes N-level wavelet decomposition to obtain wavelet coefficients at different scales. Low-frequency coefficients correspond to the effective features of the signal, while high-frequency coefficients correspond to noise. A threshold is determined based on the signal energy distribution, such as using the Birgé-Massart threshold selection criterion. The threshold size is adaptively determined by minimizing the risk function. High-frequency wavelet coefficients are thresholded: coefficients below the threshold are set to zero, and coefficients above the threshold are shrunk according to a preset rule, retaining the high-frequency coefficients corresponding to the effective features. The processed low-frequency and high-frequency coefficients are then reconstructed using wavelets to obtain the denoised vibration sample. This sample effectively retains the characteristic vibration components corresponding to the wear of the roller screw pair while suppressing environmental noise interference. Indicatively, outlier removal uses the 3σ criterion, whose core logic is to identify extreme values ​​deviating from the main body of the signal based on the normal distribution characteristics. ,in, This represents the amplitude of a specific sampling point in the denoised vibration sample. The mean of the amplitudes of all sampling points of the denoised vibration sample reflects the overall vibration level of the sample. The standard deviation of the amplitude of all sampling points of the vibration sample after denoising reflects the dispersion of the sample amplitude; The threshold for outlier determination is defined as follows: when a sampling point satisfies the above inequality, it is determined to be an outlier. For identified outliers, interpolation compensation should be used instead of direct deletion to avoid signal timing breaks due to missing sampling points. The resulting sample after compensation is the compensated vibration sample, whose data integrity and reliability are significantly improved.

[0071] S14. Mark the sampling time information of the compensated vibration sample to form a time-stamped one-dimensional vibration sub-sample.

[0072] The illustrative time information mainly includes the start and end times of the sampling. The start time is the actual running time corresponding to the first sampling point of the compensated vibration sample, and the end time is the actual running time corresponding to the last sampling point. Both must be determined based on the synchronization time reference of the radial vibration signal acquisition system. For example, if a compensated vibration sample contains signals from the 10,000th to the 11,023rd sampling point, and the sampling rate of the acquisition system is... And the actual time of the first sampling point is The starting sampling time of this sample is The sampling ended at the time when .

[0073] In one embodiment, time-frequency transformation processing is performed on each one-dimensional vibration sub-sample to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, resulting in a time-frequency map stream, including:

[0074] S21. Based on the Morlet wavelet basis function, continuous wavelet transform is performed on each one-dimensional vibration sub-sample by scaling and translation parameters to obtain the scale-time shift coefficient matrix.

[0075] Indicatively, the Morlet wavelet basis function, due to its excellent time-frequency locality, can accurately capture the frequency shift and energy change caused by wear of the roller screw pair in a one-dimensional vibrating subsample, i.e. ,in, The Morlet wavelet basis functions are scaled and translated, and are the core analytical unit of the continuous wavelet transform; The scaling parameter determines the width of the wavelet basis functions. The larger the value, the wider the wavelet, corresponding to the analysis of the low-frequency components of the one-dimensional vibrational subsample. The smaller the value, the narrower the wavelet, which corresponds to the analysis of high-frequency components; The translation parameter determines the center time position of the wavelet basis function. The value of is perfectly aligned with the time axis of the one-dimensional vibration subsample, ensuring complete coverage of the vibration signal throughout the entire time span of the sample; It is a time variable, and the unit is consistent with the sampling time of the one-dimensional vibration sub-sample; This is the energy normalization factor, used to ensure the energy conservation of wavelet basis functions under different scaling parameters and avoid energy loss due to variations in scaling parameters. The change causes an energy deviation in the transformation coefficients; Here is the Morlet wavelet mother function, and its expression is: ,in The central angular frequency of the generating function is used to enhance sensitivity to specific characteristic frequencies. Pi is a constant. It is a natural constant. The term is a complex exponential term used to provide information in the frequency dimension; its imaginary and real parts correspond to the phase and amplitude changes of the signal, respectively. This is a correction term used to ensure the energy normalization of the wavelet function and avoid energy deviation in the transform coefficients due to changes in the center frequency. The Gaussian envelope term controls the temporal locality of the wavelet function, ensuring it exhibits a Gaussian distribution along the time axis. This allows analysis only of signals within a local time range, avoiding interference from signals over longer time intervals. The continuous wavelet transform is... By traversing the preset set of scaling parameters ,like , The scale number is determined based on the sample frequency range and the set of translation parameters. ,like , The number of time locations is consistent with the number of sample points, and all are calculated. corresponding , Arranged along the scale-time position dimension, a scale-time shift coefficient matrix is ​​formed, with each row of the matrix corresponding to a different scaling parameter. That is, different frequencies correspond to different translation parameters. That is, at different times, the matrix element values ​​are wavelet transform coefficients. .

[0076] S22. Perform amplitude mapping, logarithmic compression and grayscale processing on the scale-time shift coefficient matrix in sequence, and resample according to the preset resolution to obtain a time-frequency graph set.

[0077] Due to wavelet transform coefficients Since the coefficients are complex numbers containing both amplitude and phase information, and the time-frequency plot needs to reflect energy distribution, the absolute values ​​of the coefficients are used to achieve amplitude mapping. This value directly represents the energy intensity at the corresponding time-frequency point, and after mapping, a real-valued energy coefficient matrix is ​​obtained. Logarithmic compression is used to enhance the visibility of weak features in the energy coefficient matrix. Because the vibration energy corresponding to early wear of the roller screw pair is weak, it is easily masked by the later strong energy features. Through logarithmic transformation, the strong energy range can be compressed and the weak energy range stretched, i.e. ,in, This is the energy value after compression; This is the gain coefficient, used to adjust the degree of compression. The larger the value, the more obvious the stretching effect of the weak energy feature. It can be determined according to the overall energy range of the coefficient matrix to ensure that the dynamic range of the compressed energy value is adapted to the subsequent grayscale processing. It is the natural logarithm function. This is used to avoid the problem of meaningless logarithms caused by input values ​​of zero.

[0078] Grayscale processing is the process of converting compressed energy values... Convert to image grayscale values ​​to form a two-dimensional time-frequency distribution map, using linear mapping, i.e. ,in, The value is a grayscale value, ranging from 0 to 255, where 0 corresponds to black and 255 corresponds to white. The minimum energy value in the energy coefficient matrix. The maximum energy value is obtained by traversing the matrix; The total number of gray levels. The energy values ​​are normalized to the [0,1] range and then mapped to the 0-255 grayscale range, so that the differences in energy distribution are presented intuitively in the form of grayscale depth.

[0079] Furthermore, the time-frequency image size corresponding to all one-dimensional vibration sub-samples is unified by resampling at a preset resolution. The resampling adopts bilinear interpolation, based on the preset image width. and height Pixel interpolation is performed on the grayscale matrix to ensure that each time-frequency image obtained has the same resolution. The results obtained after processing all one-dimensional vibrational subsamples as described above. The resolution time-frequency graphs are compiled to form a time-frequency graph collection.

[0080] S23. Perform batch normalization and edge clipping on the time-frequency graph set to obtain the time-frequency graph stream.

[0081] Indicatively, batch normalization aims to eliminate differences in grayscale distribution between different time-frequency maps in a time-frequency map set, avoiding the impact of grayscale shifts caused by different sample collection periods and wear stages on model feature learning. Its calculation is performed on the grayscale values ​​of all pixels in the time-frequency map set. ,in, These are the grayscale values ​​after batch normalization. This represents the original grayscale value of a pixel in the time-frequency image set. The mean grayscale value of all pixels in the time-frequency image set is calculated by iterating through each pixel of all time-frequency images in the set. The variance of the gray values ​​of all pixels reflects the degree of dispersion of the gray value distribution; This is a minimum value used to avoid calculation errors caused by zero variance; the process standardizes the gray values ​​of all time-frequency maps to a distribution with a mean of 0 and a variance of 1, ensuring that the gray features of different time-frequency maps have a unified comparison benchmark.

[0082] Edge clipping is used to remove invalid noise regions at the edges of time-frequency maps. Because continuous wavelet transform is prone to boundary effects at the edges of the time and frequency axes, the grayscale values ​​at the edges of the time-frequency maps have no actual physical meaning and need to be removed through clipping. The clipping range is determined based on the energy distribution of the time-frequency maps. Specifically, the energy distribution of grayscale values ​​in each time-frequency map is statistically analyzed, and the effective region with an energy percentage exceeding 95% of the total energy is calculated. The horizontal and vertical boundary coordinates of the effective region are then determined. Edge clipping is performed on each time-frequency map in the set according to the unified effective region boundaries, removing invalid pixels outside the boundaries. The time-frequency maps after batch normalization and edge clipping are arranged sequentially according to the sampling time order of their corresponding one-dimensional vibration subsamples, forming a continuous time-frequency map stream. The time order of adjacent time-frequency maps in the time-frequency map stream is completely consistent with the actual operating time order of the roller screw pair, fully preserving the dynamic evolution law of the time-frequency characteristics during the wear process of the roller screw pair, providing structured input data for subsequent time-series embedding and remaining life prediction.

[0083] In one embodiment, such as Figure 2 As shown, the remaining lifetime estimation model based on a dynamic temporal convolutional network is trained using the following method:

[0084] S201. Based on the time-frequency graph stream of the control rod drive mechanism life test platform, apply a pre-trained U-Net-based diffusion model for data augmentation to generate a synthetic time-frequency graph with the same distribution as the time-frequency graph stream. Merge the time-frequency graph stream and the synthetic time-frequency graph to obtain the time-frequency graph data augmentation set.

[0085] The time-frequency graph stream collected by the control rod drive mechanism life test platform forms the original basis for data augmentation. This stream contains the time-frequency characteristics of different wear stages during the roller screw pair's journey from normal operation to failure. Further, synthetic samples are generated using a diffusion model to supplement the data. The illustrative, pre-trained U-Net-based diffusion model consists of a forward diffusion process and a backward diffusion process, with U-Net used as the noise estimator for backward diffusion. The forward diffusion process involves progressively adding Gaussian noise to the original time-frequency graph, gradually transforming the original data from a complex distribution to a simple Gaussian noise distribution. ,in, For positive diffusion Time-frequency plot data of the steps; For the first Time-frequency data of the steps, hour This is the original time-frequency diagram; For the first The weighting parameter of each step decreases as the number of diffusion steps increases, and is used to control the proportion of original data retained. For the first Standard Gaussian noise is added step by step to simulate data perturbation; and They are respectively and The weighting coefficients ensure data energy conservation during the diffusion process. The back-diffusion process starts with pure Gaussian noise and gradually removes noise to restore the original data distribution. In this process, the U-Net network is responsible for predicting the noise at each step, and its input is... With diffusion steps The encoded information is output as prediction noise. ,Right now ,in For cumulative weighting, For the first The variance of the steps Standard Gaussian noise, after After a back-diffusion step, a synthetic time-frequency map consistent with the original time-frequency map distribution is finally generated. The U-Net architecture is used as the noise prediction network. The U-Net contains a downsampling encoder and an upsampling decoder path, with skip connections fusing multi-scale features. The encoder consists of multiple convolutional blocks, each containing two convolutional layers followed by ReLU activation and max-pooling layers, progressively reducing spatial resolution and increasing the number of channels. The decoder symmetrically consists of multiple upsampling blocks. Each block first increases resolution through upsampling or deconvolution, then concatenates with the feature maps of the corresponding layers in the encoder, followed by convolution and ReLU. The final layer uses convolution and sigmoid activation to output predicted noise. During training, the original image set is used... Randomly selected images Random sampling time step Generate noisy images ,Will and Input U-Net and calculate the mean square error between its output and the actual added noise. Optimize network parameters through backpropagation that minimizes error loss. .

[0086] All time-frequency maps in the original time-frequency map stream are merged with the synthetic time-frequency map generated by the diffusion model along the time dimension to form a time-frequency map data augmentation set. This augmentation set not only retains the real wear characteristics of the original data, but also makes up for the lack of small samples through synthetic samples, providing sufficient data support for subsequent model training.

[0087] S202. Label each time-frequency image sample in the time-frequency image data augmentation set with the remaining lifespan label, and normalize the remaining lifespan label to obtain a labeled augmented sample set; the labeled augmented sample set includes a training set and a validation set.

[0088] As an illustration, the remaining service life labeling should be based on the time sequence relationship of the entire life cycle of the roller screw pair to determine the total number of time periods of the entire life cycle of the roller screw pair in the control rod drive mechanism. This refers to the total number of time segments corresponding to all time-frequency image samples in the time-frequency image data augmentation set. Each time-frequency image sample corresponds to one time segment. The remaining service life label for the initial operation stage of the roller screw pair, i.e., the first time segment, is set to 1, representing the longest remaining service life. The end of the service life stage, i.e., the [missing information], is [missing information]. The label 0 for the first time segment represents a remaining lifetime of 0, and the label for the middle time segment is 0. The labels for each time segment are calculated according to a linear decreasing rule. ,in, For the first The remaining lifetime label for each time segment corresponding to a time-frequency image sample; The total number of time periods throughout the entire life cycle is determined by the total number of samples in the time-frequency graph data augmentation set; The time segment number corresponding to the current time-frequency image sample The serial numbers are arranged in chronological order of the running time of the roller screw pair to ensure that the labels are negatively correlated with the degree of wear. The smaller the size, the less wear and tear; the larger the label. The larger the label, the more wear and tear it will have, and the smaller the label should be.

[0089] S203. Convert each time-frequency image sample in the training set into a feature vector, and embed each feature vector into a time series input with a preset time length based on a time-shifting window to obtain the sequence input set and its corresponding predicted target label.

[0090] Furthermore, the pixel information of the two-dimensional time-frequency image is converted into a one-dimensional vector. For example, let the resolution of each time-frequency image sample be... , The number of pixels in the horizontal time dimension. The number of pixels in the vertical frequency dimension is used to flatten the time-frequency graph row by row or column by column. The grayscale values ​​of each pixel are arranged sequentially to form a dimension of . One-dimensional feature vector Each element of the feature vector corresponds to a pixel grayscale value in the time-frequency map, completely preserving the time-frequency energy distribution information of the time-frequency map. The preset time-shift window size is... That is, each sequence input contains A series of consecutive feature vectors, starting from the first feature vector in the training set sequence. Begin, extract Form the first time series input Then, the window is moved one step along the time axis of the feature vector sequence to ensure temporal continuity between sequences, and the window is truncated. Forming the second time series input Repeat this process until the end of the window exceeds the total length of the feature vector sequence in the training set, finally obtaining the sequence input set. , The number of sequence inputs. Each sequence input... , The sequence number is set as the last feature vector of the sequence. Corresponding remaining service life label , i.e., target tag .

[0091] S204. The remaining lifetime estimation model based on the dynamic temporal convolutional network is trained and its hyperparameters are optimized using the sequence input set and the labeled expanded sample set to obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network.

[0092] During the model training phase, each sequence input from the sequence input set is used as an example. As input to the model, the corresponding predicted target label The parameters are learned by minimizing the difference between the model's predicted values ​​and the true labels, as the expected output of the model. The training process uses mean squared error (MSE) as the loss function. ,in, This represents the loss value used for model training. The number of sequences in the input set; For the model to input sequence The remaining useful life prediction; For sequence input The corresponding true predicted target label; Sum of the squares of prediction errors for all sequences, divided by Obtain the average error and loss value. The smaller the value, the higher the model's prediction accuracy. Model parameter optimization employs the Adam optimizer, which iteratively updates the parameters of the dynamically timed convolutional network (VTCNN) through backpropagation. This includes the kernel weights of the dynamically convolutional layers, the hyperfunction parameters of the dynamic activation functions, the weight matrix of the attention mechanism, and the weights and biases of the fully connected layers. The Adam optimizer adaptively adjusts the learning rate to balance model training speed and convergence stability, avoiding gradient vanishing or exploding problems. Hyperparameter optimization needs to be performed in conjunction with a validation set that has been expanded using labeled samples. The main hyperparameters to be optimized include the number of residual blocks in the VTCNN. Time-shifting window size kernel size The grid search method is used to traverse the preset hyperparameter combinations. For each hyperparameter combination, the model is trained using the training set, and the root mean square error (RMSE) or mean absolute error (MAE) of the model is calculated using the validation set. The hyperparameter combination with the best performance on the validation set is selected as the final model parameters.

[0093] Repeat the above training and hyperparameter optimization process until the model loss value converges on the training set (i.e., the loss value no longer decreases significantly with the number of iterations) and the error on the validation set is stable. Finally, the trained model for estimating the remaining lifetime based on a dynamic temporal convolutional network is obtained.

[0094] In one embodiment, the remaining lifetime estimation model based on a dynamic temporal convolutional network is trained and its hyperparameters optimized using a sequence input set and a labeled augmented sample set to obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network, including:

[0095] S31. Perform grid search optimization on the key structural hyperparameters of the remaining lifetime estimation model based on dynamic temporal convolutional networks to obtain multiple combinations of hyperparameters that can be traversed within a predefined search range; the key structural hyperparameters include the number of residual blocks and the size of the time window.

[0096] The number of residual blocks and the size of the time window in the key structural hyperparameters directly determine the ability of the dynamic temporal convolutional network to extract temporal features and its modeling efficiency. The number of residual blocks affects the feature extraction depth of the model; too few blocks make it difficult to capture the deep dynamic features of the wear of the ball screw pair, while too many blocks can easily lead to overfitting and gradient vanishing. The size of the time window determines the temporal length of a single input to the model; too small a window cannot cover the complete wear trend segment, while too large a window increases computational complexity and may introduce redundant noise. Specifically, the search range for the number of residual blocks needs to be determined in conjunction with model training resources and feature complexity, and is usually set to a continuous set of integers, which can balance feature extraction depth and computational cost. The search range for the size of the time window needs to match the feature period of the wear of the ball screw pair to ensure coverage of temporal information from short-term local wear to medium-term trend evolution. For example, each value within the search range for the number of residual blocks can be combined with each value within the search range for the size of the time window to generate all possible hyperparameter combinations, such as residual block number = 3, time window size = 20; residual block number = 4, time window size = 15, etc. Each combination represents an independent set of model structure configurations.

[0097] S32. For each hyperparameter combination, the remaining lifetime estimation model based on the dynamic temporal convolutional network is trained according to the sequence input set and its corresponding predicted target label. The root mean square error and mean absolute error of the remaining lifetime prediction value of the trained dynamic temporal convolutional network-based remaining lifetime estimation model and the normalized remaining lifetime label are calculated using the validation set to obtain the performance evaluation index.

[0098] For each hyperparameter combination, model training requires using the sequence input set as training data and the corresponding predicted target labels as supervision signals. Each time-series input from the sequence input set is sequentially fed into a dynamic temporal convolutional network. The network extracts temporal features through dynamic convolutional layers, weights key information through an attention mechanism, and outputs predicted values ​​through fully connected layers. Then, guided by the difference between the predicted values ​​and the predicted target labels, an optimizer adjusts the network parameters. Furthermore, after training, the model performance is evaluated using a validation set, i.e., by calculating the MAE (Maximum Averaging Effect). ,in, The mean absolute error is the average error; the smaller the value, the smaller the average deviation between the model's predicted value and the true label. To determine the number of samples in the validation set; For the model on the validation set The remaining lifetime prediction value is input from a sequence, and the value range is [0,1]. For the verification set The normalized remaining useful life true label corresponding to each sequence input, and... Consistent dimensions; The absolute prediction error for a single sample is summed and divided by . The average error is obtained. RMSE is... ,in, It is the root mean square error. Compared with MAE, it is more sensitive to larger errors and can highlight the prediction bias of the model for the severe wear stage or the early wear stage. The squared prediction error of a single sample can amplify the impact of larger errors. Summing these errors, averaging them, and then taking the square root yields an index reflecting the overall error dispersion. The MAE and RMSE corresponding to each hyperparameter combination are used as performance evaluation indicators, providing a quantitative basis for subsequent selection of the optimal configuration.

[0099] S33. Select the hyperparameter combination that minimizes the performance evaluation index as the final model configuration.

[0100] As an illustration, RMSE is compared first, because the sensitivity of this index to large errors can ensure the accuracy of the model's predictions during the critical wear stage of the ball screw pair, and avoid maintenance decision errors caused by large errors. If the RMSE of multiple hyperparameter combinations are similar, MAE is further compared, and the combination with smaller MAE is selected to ensure that the average deviation of the model's overall prediction is minimized.

[0101] S34. Using the final model configuration and all training data of the expanded sample set with labels, train the remaining lifetime estimation model based on the dynamic temporal convolutional network, and calculate the loss based on the remaining lifetime prediction and the normalized remaining lifetime label.

[0102] Furthermore, utilizing all training data can cover more wear scenarios of the roller screw pair, avoiding insufficient model learning due to insufficient data. Batch training mode should be adopted during training, reducing loss through iterative updates of model parameters: after each batch training, the optimizer calculates the gradient based on the loss value, and backpropagates to update the convolutional kernel weights of the dynamic convolutional layer, the hyperfunction parameters of the dynamic activation function, the weight matrix of the attention mechanism, and the weights and biases of the fully connected layer, ensuring that the model gradually learns the wear characteristics and patterns in the training data.

[0103] S35. Iteratively update the model parameters of the remaining lifetime estimation model based on the dynamic temporal convolutional network based on the loss until convergence, and obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network.

[0104] After each batch of training, the optimizer adjusts its parameters based on the gradient direction of the loss value. The adjustment magnitude is controlled by the learning rate, focusing on updating the weights of the dilated convolutional kernels of the dynamic convolutional layers, the weight matrix of the attention mechanism, and the hyperfunction parameters of the dynamic activation function. The model is considered converged when the training set loss value changes less than a preset threshold for multiple consecutive iterations, and the RMSE and MAE of the validation set no longer decrease with iteration. At this point, the model parameters have sufficiently learned the features and patterns in the training data, and further iterations could easily lead to overfitting. The converged model should have all its parameters saved to form a directly deployable residual lifetime estimation model based on a dynamic temporal convolutional network.

[0105] In one embodiment, such as Figure 3 As shown, the remaining lifetime estimation model based on dynamic temporal convolutional networks obtains the predicted remaining lifetime value through the following method:

[0106] S301. Based on hyperparameter combination, the causal dilated convolution of residual blocks is used to extract multi-scale temporal features from the sequence input set, and the hidden representation of the time step corresponding to each time shift window is obtained.

[0107] The number of residual blocks and the size of the time window are determined in the hyperparameter combination. The causal dilated convolution within the residual block is the core of multi-scale temporal feature extraction. It covers wear features at different time scales through different dilation factors, ensuring the capture of complete information from short-term local vibrations to long-term trend evolution. The sequence input corresponding to each time shift window is processed by the causal dilated convolution of all residual blocks, and the output is the hidden representation of each time step. The hidden representations are arranged in the order of time steps to form a temporal feature sequence containing multi-scale wear features, which provides the basis for subsequent activation function processing and attention weighting.

[0108] S302. Perform global average pooling on each channel according to the hidden representation, and input the global average pooling result into the hyperfunction network to obtain the segmented activation function parameter vector; the hyperfunction network includes two fully connected layers and intermediate ReLU and normalization layers.

[0109] Global average pooling is performed along the channel dimension of the hidden representation. Its purpose is to compress temporal information, obtain the global feature mean for each channel, and avoid interference from local noise. ,in, For the first The global average pooling result of each channel, that is, the mean of the hidden representation of all time steps of that channel; To hide the total number of time steps, the size should be the same as the time movement window size; For the first The time step, the first The hidden representation value of each channel reflects the feature strength of that channel at that time step; The average coefficient is used to ensure the results reflect the global characteristics of all channels. This applies to all channels. Forming the global mean vector , The dimensions and number of channels are the same, and the input is a hyperfunction network. The first layer of the hyperfunction network is a fully connected layer. Mapped to a high-dimensional feature space, with the output dimension set to... ( (where the number of channels is used), followed by a nonlinear transformation introduced through the ReLU activation function. ,in This is the weight matrix of the first fully connected layer. For bias vectors, This is the output feature vector of the first layer. Then, a normalization layer is applied using the Sigmoid function to... Normalization to The interval, the formula is To avoid unstable parameter updates due to excessively large eigenvalues, a second fully connected layer is finally applied. Mapped to a segmented activation function parameter vector , dimension It includes the activation function slope and offset parameters for each channel, and the formula is as follows: ,in This is the weight matrix for the second fully connected layer. For bias vectors, These are the core parameters for constructing the activation function later.

[0110] S303. Construct a parameterized piecewise linear activation function based on the piecewise activation function parameter vector, and obtain the activation features corresponding to the hidden representation based on the piecewise linear activation function.

[0111] Segmented activation function parameter vector The slope parameter is split into two sub-vectors based on the channel. With offset parameter Each channel corresponds to a set ( ), used to construct the piecewise linear activation function for this channel. ,in, This refers to a parameterized piecewise linear activation function, also known as a dynamic activation function. For the first Time step, first Hidden representation of channels; For the first The activation slope of the channel controls the intensity of the linear transformation of the feature. The larger the value, the more pronounced the amplification effect on the channel characteristics; For the first Channel activation offset, adjusting feature activation threshold. The smaller the value, the easier it is to activate weak features; Ensure the activation output is non-negative, meeting the non-negativity requirement of neural network feature extraction. This involves creating a hidden representation for each time step and each channel. Input the corresponding channel The function yields the activated feature values. ,all Arranged according to time step and channel order to form activation features The activation feature can adaptively highlight key wear features in different channels and at different time steps.

[0112] S304. The activation features are added to the hidden representation through residual connections to obtain the hidden state sequence. The hidden state sequence is then time-channel weighted through an attention mechanism to obtain the weighted feature representation.

[0113] Schematic, residual connection ,in, For the first The hidden state vector at each time step integrates information from activation features and the original hidden representation; For the first The activation feature vector at each time step; For the first The original hidden representation vectors at each time step are directly added together via residual connections to alleviate the gradient vanishing problem in deep networks, while preserving the temporal correlation of the original features. Arranged according to time steps to form a hidden state sequence .

[0114] The attention mechanism weights the hidden state sequence along two dimensions: time and channel. Specifically, the time dimension weights the importance weight of each time step. ,in For the first Weight of time steps This is the attention weight matrix. For bias vectors, The exponent term of the softmax function is used to highlight the weights of important time steps; the importance weight of each channel is calculated using a channel-weighted approach. ,in Here is the channel attention weight matrix. For bias vectors, For the first The global mean of the channel. The weighted feature representation is calculated by multiplying the time and channel weights, i.e. ,in, The final weighted feature representation vector integrates key features from both the time and channel dimensions. For the first Time step, first Joint weight of channels, The elements of the hidden state sequence are weighted to highlight the feature information that contributes most to the remaining lifetime prediction.

[0115] S305. Based on the fully connected layer, the weighted feature representation is mapped to the normalized remaining lifetime prediction value.

[0116] Schematic illustration: The fully connected layer is responsible for representing the high-dimensional weighted features. Mapped to one-dimensional normalized predicted values, the first layer will Compressed to a low-dimensional feature space, the formula is: ,in This is the weight matrix of the first fully connected layer. For bias vectors, The ReLU activation function introduces non-linearity into the first layer's output feature vector, enhancing the model's expressive power; the second layer maps low-dimensional features to normalized predicted values. ,in, The normalized remaining lifetime prediction value ranges from [0,1]. The closer to 1, the longer the remaining lifespan; the closer to 0, the shorter the remaining lifespan. This is the weight matrix for the second fully connected layer. As the bias vector, this layer transforms low-dimensional features into a single predicted value through a linear mapping, achieving a regression mapping from features to remaining lifetime.

[0117] In one embodiment, the hidden state sequence is time-channel weighted using an attention mechanism to obtain a weighted feature representation, including:

[0118] S41. Calculate the attention score vector by transforming the hidden state sequence into a linear transformation.

[0119] Indicative, the hidden state sequence is as follows ,in Representing the The hidden state vector at each time step has a dimension of . The core of the linear transformation is to map the high-dimensional hidden state vector into a single scalar attention score, quantifying the importance of each time step for remaining lifetime prediction. ,in, For the first The attention score for each time step is a scalar value. The higher the score, the richer the wear-related key information contained in the hidden state of that time step. The weight matrix for the linear transformation has dimensions of . , used to The hidden state vector is compressed into 1 dimension, and its element values ​​are optimized through model training to learn the contribution of different feature dimensions to the score. For the first The hidden state vector at each time step contains the temporal and channel features of that time step; This is a bias term for the linear transformation, used to adjust the baseline value of the score and avoid score distortion caused by the overall shift of the hidden state; These are learnable weight vectors. The process iterates through all time steps of the hidden state sequence. to The calculation yields the result for each time step. These scores are arranged in time step order to form an attention score vector. This vector fully reflects the preliminary assessment results of the importance of each time step in the hidden state sequence.

[0120] S42. Perform Softmax normalization on the attention score vector to obtain the attention weight sequence.

[0121] The purpose of Softmax normalization is to convert the scores in the attention score vector into probabilistic weights, ensuring that the sum of the weights at all time steps is 1. ,in, For the first The attention weights at each time step, with values ​​ranging from [value range missing]. The larger the weight, the greater the contribution of the hidden state at that time step to the remaining lifetime prediction. For the first The exponential operation of the attention score at each time step is used to convert the score into a non-negative value, while amplifying the differences in scores across different time steps. The higher the score, the larger the exponential value, and the more prominent the weighting. The denominator... The sum of attention score indices across all time steps is used to normalize the weights, ensuring... This satisfies the probability distribution characteristics. The attention score vector... Each of them Perform the above calculations to obtain the corresponding time step. The attention weights are arranged in chronological order to form an attention weight sequence. This sequence precisely quantifies the relative importance of each time step in the hidden state sequence; for example, the time step in the accelerated wear stage of the ball screw pair will receive higher importance. The time step in the early, slight wear stage is lower. .

[0122] S43. The hidden state sequence is weighted and summed based on the attention weights to obtain the weighted feature representation.

[0123] The core of weighted summation is to selectively fuse features at each time step in the hidden state sequence using an attention weight sequence, highlighting key wear features at high-weight time steps and suppressing redundant or noisy information at low-weight time steps. ,in, For the final weighted feature representation, the dimensions and hidden state vectors Consistent with this, the vector condenses the key information in the hidden state sequence and is the core input for subsequent fully connected layers to map the remaining lifetime prediction value; For the first The attention weights at each time step are used as weighting coefficients for the hidden state at that time step, controlling their contribution to the final feature representation. For the first The hidden state vector at each time step, containing the wear characteristics at that time step; summation symbol. This represents the accumulation of the weight-hidden state products across all time steps, achieving selective fusion of features from multiple time steps. Through this operation, the weighted feature representation... It can focus on the most critical features for predicting remaining lifetime in the hidden state sequence, such as the impact features of late wear and the frequency shift features of mid-term wear, while filtering out meaningless early steady vibration features or noise interference, providing high-quality feature input for subsequent accurate prediction.

[0124] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0125] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting the remaining life of a roller screw pair, characterized in that, The method includes: The radial vibration signal of the roller screw pair of the control rod drive mechanism is acquired during operation, and the radial vibration signal is preprocessed to obtain multiple one-dimensional vibration sub-samples. The time-frequency transformation process is performed on each of the one-dimensional vibration sub-samples to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, thus obtaining a time-frequency map stream; The time-frequency graph stream is embedded in a time sequence according to a preset time window to obtain the target time-frequency distribution sequence. The target time-frequency distribution sequence is input into a pre-trained remaining lifetime estimation model based on a dynamic temporal convolutional network to obtain a remaining lifetime prediction value. The remaining lifetime estimation model includes a dynamic convolutional layer, an attention mechanism, and a fully connected layer. The dynamic convolutional layer uses a dynamic activation function.

2. The method according to claim 1, characterized in that, The radial vibration signal is preprocessed to obtain multiple one-dimensional vibration sub-samples, including: The radial vibration signal is sliced ​​using a sliding window with a preset number of sampling points and either no overlap or a specified overlap rate to obtain multiple one-dimensional sub-samples. The amplitude of each one-dimensional subsample is normalized by using maximum-minimum normalization to obtain normalized vibration samples; The normalized vibration samples are denoised, and outliers are removed from the denoised samples to obtain compensated vibration samples. The sampling time information of the compensated vibration sample is labeled to form the time-stamped one-dimensional vibration sub-sample.

3. The method according to claim 1, characterized in that, The step of performing time-frequency transformation processing on each of the one-dimensional vibration sub-samples to generate a two-dimensional time-frequency distribution map characterizing the time-frequency energy distribution of the signal, resulting in a time-frequency map stream, includes: Based on the Morlet wavelet basis function, continuous wavelet transform is performed on each of the one-dimensional vibration sub-samples by scaling and translation parameters to obtain the scale-time shift coefficient matrix. The scale-time shift coefficient matrix is ​​sequentially subjected to amplitude mapping, logarithmic compression, and grayscale processing, and then resampled at a preset resolution to obtain a set of time-frequency graphs. Batch normalization and edge clipping are performed on the time-frequency graph set to obtain the time-frequency graph stream.

4. The method according to claim 1, characterized in that, The remaining lifetime estimation model based on dynamic temporal convolutional networks is trained using the following method: Based on the time-frequency graph stream of the control rod drive mechanism life test platform, a pre-trained U-Net-based diffusion model is applied for data augmentation to generate a synthetic time-frequency graph with the same distribution as the time-frequency graph stream. The time-frequency graph stream and the synthetic time-frequency graph are then merged to obtain a time-frequency graph data augmentation set. Each time-frequency image sample in the time-frequency image data augmentation set is labeled with its remaining useful life, and the remaining useful life labels are normalized to obtain a labeled augmentation sample set; the labeled augmentation sample set includes a training set and a validation set; Each time-frequency image sample in the training set is converted into a feature vector, and each feature vector is embedded into a time series input of a preset time length based on a time-shifting window to obtain a set of sequence inputs and their corresponding predicted target labels. The remaining lifetime estimation model based on the dynamic temporal convolutional network is trained and its hyperparameters optimized using the sequence input set and the labeled augmented sample set to obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network.

5. The method according to claim 4, characterized in that, The step of training and hyperparameter optimization of the remaining lifetime estimation model based on the dynamic temporal convolutional network using the sequence input set and the labeled augmented sample set to obtain the remaining lifetime estimation model based on the dynamic temporal convolutional network includes: The key structural hyperparameters of the remaining lifetime estimation model based on the dynamic temporal convolutional network are optimized by grid search to obtain multiple combinations of hyperparameters traversed within a predefined search range; the key structural hyperparameters include the number of residual blocks and the size of the time window; For each of the hyperparameter combinations, the remaining lifetime estimation model based on the dynamic temporal convolutional network is trained according to the sequence input set and its corresponding predicted target label, and the root mean square error and mean absolute error of the remaining lifetime prediction value of the trained remaining lifetime estimation model based on the dynamic temporal convolutional network and the normalized remaining lifetime label are calculated using the validation set to obtain the performance evaluation index. Select the hyperparameter combination that minimizes the performance evaluation index as the final model configuration; The remaining lifetime estimation model based on the dynamic temporal convolutional network is trained using all training data from the final model configuration and the labeled augmented sample set, and the loss is calculated based on the remaining lifetime prediction and the normalized remaining lifetime label. The model parameters of the remaining lifetime estimation model based on the dynamic temporal convolutional network are iteratively updated based on the loss until convergence, thus obtaining the remaining lifetime estimation model based on the dynamic temporal convolutional network.

6. The method according to claim 5, characterized in that, The remaining lifetime estimation model based on dynamic temporal convolutional networks obtains the predicted remaining lifetime value through the following method: Based on the hyperparameter combination, the causal dilated convolution of the residual block is used to extract multi-scale temporal features from the sequence input set, and the hidden representation of the time step corresponding to each time shift window is obtained. Global average pooling is performed on each channel according to the hidden representation, and the result of global average pooling is input into the hyperfunction network to obtain the segmented activation function parameter vector; the hyperfunction network includes two fully connected layers and intermediate ReLU and normalization layers; A parameterized piecewise linear activation function is constructed based on the piecewise activation function parameter vector, and the activation features corresponding to the hidden representation are obtained based on the piecewise linear activation function. The activation features are added to the hidden representation through residual connections to obtain a hidden state sequence, and the hidden state sequence is then time-channel weighted through an attention mechanism to obtain a weighted feature representation. Based on the fully connected layer, the weighted feature representation is mapped to a normalized remaining lifetime prediction value.

7. The method according to claim 6, characterized in that, The step of applying time-channel weighting to the hidden state sequence through an attention mechanism to obtain a weighted feature representation includes: The attention score vector is calculated by linearly transforming the hidden state sequence. The attention score vector is subjected to Softmax normalization to obtain the attention weight sequence; The weighted feature representation is obtained by summing the hidden state sequence based on the attention weights.

Images