A method, device, and medium for determining a remaining life prediction model

Abstract

This invention discloses a method, apparatus, device, and medium for determining a remaining lifetime prediction model. The method includes: acquiring historical full-life vibration data of a bearing; determining the predicted start time of the bearing entering a degradation state and a training sample set based on the signal amplitude of the historical full-life vibration data; training and optimizing a pre-constructed kernel adaptive filter model based on sample data in the training sample set and a global convex loss function until the training termination condition is met, thus obtaining the trained remaining lifetime prediction model. The pre-constructed kernel adaptive filter model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained. This method accurately locates the bearing degradation start point, extracts high-quality samples, and uses a global convex loss function for optimization, avoiding local optima; combined with fixed-dimensional explicit feature mapping, it achieves nonlinear high-precision prediction and efficient computation, significantly improving the accuracy, stability, and real-time performance of remaining lifetime prediction.

Description

Technical Field

[0001] This invention relates to the field of equipment health management technology, and in particular to a method, apparatus, equipment and medium for determining a remaining useful life prediction model. Background Technology

[0002] Bearings are critical support components in the main shaft and rotary transmission parts of sewing equipment (such as industrial sewing machines, overlock machines, and embroidery machines). Their operating condition directly affects the stability of the equipment, sewing quality, and production continuity. Currently, the maintenance of sewing equipment bearings mainly relies on periodic maintenance, post-event repairs, or simple condition monitoring, and there is a lack of solutions for predicting remaining service life.

[0003] In the research on the remaining service prediction of general rolling bearings, existing methods are mainly divided into two categories: model-based methods (such as fatigue loss models and Wiener process models) have the advantages of clear mechanisms and good interpretability, but usually rely on bearing degradation mechanisms, statistical assumptions, or prior parameters; data-driven methods (such as support vector machines, neural networks, and long short-term memory networks) can learn degradation patterns from historical bearing operating data, but generally require a large number of historical samples or full-life data to support them. Kernel Adaptive Filtering (KAF) is a lightweight online learning method that can, to some extent, compensate for the shortcomings of the above two types of methods, providing a reference technical foundation for the online remaining service prediction of sewing equipment bearings.

[0004] In traditional KAF online learning, the size of the kernel dictionary grows linearly with the number of samples, leading to a continuous increase in storage overhead and computational complexity, which poses challenges for actual equipment deployment. During the operation of sewing equipment, there are start-up and shutdown shocks, mechanical vibrations, load changes and abnormal disturbances. Bearing status data is easily affected by non-Gaussian noise and outliers, affecting the stability of the remaining service life prediction results. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for determining a remaining useful life prediction model, in order to improve the accuracy of the remaining useful life prediction model.

[0006] According to a first aspect of the present invention, a method for determining a remaining useful life prediction model is provided, comprising:

[0007] Acquire historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data;

[0008] Based on the sample data in the training sample set and combined with the global convex loss function, the pre-constructed kernel adaptive filtering model is trained and optimized until the training termination condition is met, and the trained remaining lifetime prediction model is obtained. The pre-constructed kernel adaptive filtering model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained.

[0009] According to a second aspect of the present invention, an apparatus for determining a remaining useful life prediction model is provided, comprising:

[0010] The sample determination module is used to acquire the historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data.

[0011] The model determination module is used to train and optimize the pre-constructed kernel adaptive filtering model based on the sample data in the training sample set and the global convex loss function until the training termination condition is met, so as to obtain the trained remaining lifetime prediction model. The pre-constructed kernel adaptive filtering model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer and a weight parameter vector to be trained.

[0012] According to a third aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0013] At least one processor; and

[0014] A memory communicatively connected to the at least one processor; wherein,

[0015] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method for determining the remaining lifetime prediction model according to any embodiment of the present invention.

[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the method for determining the remaining lifetime prediction model according to any embodiment of the present invention.

[0017] According to a fifth aspect of the present invention, embodiments of the present invention also provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements a method for determining a remaining lifetime prediction model according to any embodiment of the present invention.

[0018] The technical solution of this invention acquires historical full-life vibration data of the bearing, determines the predicted start time of the bearing entering a degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data, and then trains and optimizes a pre-constructed kernel adaptive filter model based on the sample data in the training sample set and a global convex loss function until the training termination condition is met, resulting in a trained remaining life prediction model. The pre-constructed kernel adaptive filter model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained. This approach accurately locates the bearing degradation start point, extracts high-quality samples, and uses a global convex loss function for optimization, avoiding local optima. Combined with fixed-dimensional explicit feature mapping, it achieves nonlinear high-precision prediction and efficient computation, significantly improving the accuracy, stability, and real-time performance of remaining life prediction.

[0019] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a method for determining a remaining useful life prediction model according to Embodiment 1 of the present invention;

[0022] Figure 2 This is a flowchart of a method for determining a remaining useful life prediction model according to Embodiment 2 of the present invention;

[0023] Figure 3 This is an example diagram of a sewing equipment for determining a remaining life prediction model according to Embodiment 2 of the present invention;

[0024] Figure 4 This is an example diagram illustrating the effect of a method for determining a remaining life prediction model according to Embodiment 2 of the present invention;

[0025] Figure 5 This is a schematic diagram of the structure of a device for determining a remaining lifetime prediction model according to Embodiment 3 of the present invention;

[0026] Figure 6 This is a schematic diagram of the structure of an electronic device that implements an embodiment of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] Example 1

[0030] Figure 1 This is a flowchart illustrating a method for determining a remaining life prediction model according to Embodiment 1 of the present invention. This embodiment is applicable to model training for predicting the bearing life in sewing equipment. The method can be executed by a device for determining the remaining life prediction model, which can be implemented in hardware and / or software and can be configured in an electronic device. Figure 1 As shown, the method includes:

[0031] S110. Obtain the historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data.

[0032] In this embodiment, the historical full-life vibration data of the bearing refers to the raw vibration signal data collected by vibration sensors throughout the entire life cycle of the bearing, from its initial use to complete failure, serving as the raw data source for model training. The signal amplitude can be understood as the maximum amplitude value in the vibration signal sampling segment, used to construct health indicators that reflect the degree of bearing degradation. The prediction start time can be understood as the critical moment when the bearing transitions from the healthy stage to the degradation stage. The training sample set refers to the training dataset constructed using health indicator sequence segments before the prediction start time as input features and subsequent health indicator values ​​at the corresponding time as labels.

[0033] Specifically, the processor can acquire the complete vibration signal of the bearing from its intact state to failure, extract the maximum amplitude from each sample segment to construct a health indicator sequence, select a benchmark for the healthy stage from the health indicator sequence, and calculate the mean and variance to obtain the health judgment interval. The prediction start time is determined based on the health indicator values ​​and the health judgment interval. This time marks the transition of the bearing from the healthy stage to the degradation stage and is the origin time for initiating remaining life prediction. A sample set is constructed for training the model using the health indicator sequence before the prediction start time as input and the corresponding health indicator values ​​after the start time as labels.

[0034] S120. Based on the sample data in the training sample set and combined with the global convex loss function, the pre-constructed kernel adaptive filtering model is trained and optimized until the training termination condition is met, and the trained remaining lifetime prediction model is obtained.

[0035] In this embodiment, the global convex loss function can be understood as the weighted least squares form transformed from the original exponential robust loss function through the auxiliary variable method (semi-quadratic optimization). This form is a globally convex function, ensuring that the optimization process converges to the global optimum. The pre-built kernel adaptive filtering model can be understood as a prediction model pre-built based on the kernel adaptive filtering framework. The fixed-dimensional explicit feature mapping layer can be understood as a mapping layer constructed using a random Fourier feature method based on a generalized Gaussian distribution, mapping the input data into a fixed-dimensional (M-dimensional) explicit feature vector, so that the model structure no longer grows with the number of samples, replacing the kernel dictionary of the traditional KAF. The linear combination layer can be understood as a layer that performs an inner product operation between the mapped explicit feature vector and the weight parameter vector, realizing the calculation from features to predicted output. The training termination condition can be understood as the judgment condition for terminating iterative training, including the convergence of the weight parameter vector (the change in continuous iterations is less than a threshold) or reaching the preset maximum number of iterations.

[0036] Specifically, the pre-built kernel adaptive filtering model comprises three components: ① Fixed-dimensional explicit feature mapping layer: This layer maps the input health indicators to fixed-dimensional (M-dimensional) explicit feature vectors using a random Fourier feature method based on a generalized Gaussian distribution, replacing the kernel dictionary that grows with the number of samples in traditional KAF, thus making the model structure controllable; ② Linear combination layer: This layer performs an inner product operation between the explicit feature vector and the weight parameter vector to obtain the predicted value; ③ Weight parameter vector to be trained: These are the parameters to be learned by the model. During training, the processor sequentially passes the input data of the training samples through the mapping layer to obtain explicit feature vectors, then calculates the predicted value using the linear combination layer and the current weight parameter vector, and calculates the error between the predicted value and the label. The optimization objective uses a global convex loss function (derived from the original exponential robust loss function through the auxiliary variable method), and iteratively updates the weight parameter vector using the conjugate direction method combined with an exponential forgetting factor. The model is iterated repeatedly by traversing the training sample set until the training termination condition is met (weight convergence or the maximum number of iterations is reached), resulting in a trained model. This model can be directly used for online prediction of the remaining service life of bearings.

[0037] The technical solution of this invention acquires historical full-life vibration data of the bearing, determines the predicted start time of the bearing entering a degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data, and then trains and optimizes a pre-constructed kernel adaptive filter model based on the sample data in the training sample set and a global convex loss function until the training termination condition is met, resulting in a trained remaining life prediction model. The pre-constructed kernel adaptive filter model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained. This approach accurately locates the bearing degradation start point, extracts high-quality samples, and uses a global convex loss function for optimization, avoiding local optima. Combined with fixed-dimensional explicit feature mapping, it achieves nonlinear high-precision prediction and efficient computation, significantly improving the accuracy, stability, and real-time performance of remaining life prediction.

[0038] As a first optional embodiment of this embodiment, based on the above embodiment, the steps for constructing the global convex loss function can be refined as follows:

[0039] By introducing auxiliary weight variables, the original exponential robust loss function is rewritten as the sum of the product of the auxiliary weight variable and the square of the error, plus the dual function of the auxiliary weight variable, resulting in the rewritten expression. Here, the error is the difference between the predicted health indicator sequence and the label. Based on the rewritten expression, the optimal value function for the auxiliary weight variable is determined with the error as the independent variable, ensuring that the auxiliary weight variable reaches its extreme value under any error, and that the auxiliary weight variable is negatively correlated with the square of the error. Substituting the optimal value function of the auxiliary weight variable back into the rewritten expression yields a globally convex function in the form of weighted least squares.

[0040] In this embodiment, the auxiliary weight variable, introduced through semi-quadratic optimization, is denoted as and is used to rewrite the original non-convex loss function into a weighted least squares form. Each sample corresponds to an auxiliary weight, which is negatively correlated with the square of the error. The original exponential robust loss function refers to a robust loss function constructed in the exponential form of the square of the error. It has a suppressive effect on non-Gaussian noise but is not globally convex. The rewritten expression can be understood as the intermediate form of the original loss function after introducing the auxiliary weight variable. The dual function refers to the dual function term corresponding to the auxiliary weight variable in the semi-quadratic optimization. The error refers to the difference between the predicted health indicator sequence and the label of the current sample, which can be denoted as . The optimal value function can be understood as the optimal value function of the auxiliary weight variable with respect to the error. The weighted least squares form can be understood as the final loss function form obtained by substituting the optimal value function back into the final form.

[0041] Specifically, while the original exponential robust loss function exhibits good robustness to non-Gaussian noise, its mathematical form contains a nested exponential structure, making it not a globally convex function. This means that directly optimizing the original loss function may lead to local optima instead of finding the global optimum, impacting the accuracy of the prediction model. By introducing an additional auxiliary variable, the original loss function is rewritten into a new form. The rewritten expression consists of two terms: the first is the product of the auxiliary weight variable and the square of the error, and the second is the dual function of the auxiliary weight variable. For any given error value, by appropriately choosing the value of the auxiliary weight variable, the rewritten expression can be made identical to the original loss function. In other words, the two expressions before and after the rewrite are mathematically equivalent, and their optimal solutions are exactly the same. The auxiliary weight variable is negatively correlated with the square of the error: when the sample error is large (meaning the sample may be an outlier or noise), the corresponding auxiliary weight variable takes a smaller value; when the sample error is small (meaning the prediction of the sample is more accurate), the corresponding auxiliary weight variable takes a larger value. By automatically adjusting the weight of each sample in the loss function using auxiliary weight variables, the impact of outliers on model updates is automatically suppressed, while the contribution of normal samples to model updates is preserved. The second step has already obtained the optimal value rule for the auxiliary weight variables with respect to the error; that is, for any given error, the optimal auxiliary weight value can be calculated. The third step substitutes this rule into the rewritten expression obtained in the first step, eliminating the auxiliary weight variables and making the loss function a function that only contains the error. The final loss function obtained after back-substitution has the form of weighted least squares, where the squared error of each sample is multiplied by a weight associated with that sample's error, and then summed over all samples. Unlike ordinary least squares (where all samples have the same weight), here each sample has a different weight.

[0042] For example, the original exponential robust loss function is:

[0043]

[0044] in, Sensitivity coefficient For kernel width, For the first The error between the actual value and the predicted value at any given time. This represents the total number of training samples.

[0045] Since the original exponential robust loss function has nested exponents and is not a globally convex function, a semi-quadratic optimization is used to transform the original exponential robust loss function into a weighted least squares form:

[0046]

[0047] in, Let be the coefficients of the auxiliary weight variables for the i-th sample. , For the weight vector, Let be the auxiliary weight variable for the i-th sample. Let be the expected output (label) for the i-th sample. For the i-th input The displayed feature vector.

[0048] As a second optional embodiment of this first embodiment, based on the above embodiment, it further includes:

[0049] Vibration signals during bearing operation are collected in real time. Based on the vibration signals, a real-time health indicator sequence is determined. According to the statistical distribution of the health stages and the health judgment interval in the real-time health indicator sequence, the predicted start time of the bearing entering the degradation state is determined. The real-time health indicator sequence before the predicted start time is input into the trained remaining service life prediction model, and the predicted health indicator sequence for the prediction period is output. The predicted values ​​of each health indicator in the predicted health indicator sequence are compared with the preset failure threshold to determine the failure time corresponding to the predicted health indicator value that reaches or exceeds the preset failure threshold. Based on the difference between the failure time and the predicted start time, the remaining service life of the bearing is determined.

[0050] In this embodiment, the vibration signal refers to the raw vibration waveform data collected by the vibration sensor during the actual operation of the bearing. The real-time health indicator sequence can be understood as extracting the maximum amplitude from the vibration signal segment at each sampling moment during operation as the health indicator value. The health indicator prediction sequence can be understood as a sequence composed of predicted health indicator values ​​for subsequent prediction periods. The preset failure threshold can be understood as a pre-set critical value for determining bearing failure. The failure time is the first moment when the predicted health indicator value in the health indicator prediction sequence reaches / exceeds the preset failure threshold. The remaining service life can be understood as the time span from the current moment to the bearing's expected aging time.

[0051] Specifically, during the actual operation of the bearing, vibration sensors installed on the bearing continuously collect vibration signals. The system processes the vibration signal segment at each sampling moment, extracting the maximum amplitude from the segment as a degradation feature; this value is the health indicator value at that moment. As time progresses, the health indicator values ​​at each moment sequentially form a real-time health indicator sequence. When the processor detects that n consecutive health indicator values ​​exceed the health judgment interval determined during the training phase, it determines that the bearing has entered the degradation phase from the healthy phase, and determines the last moment of the consecutive exceedances as the prediction start moment, officially starting the remaining life prediction from that moment. The processor can input the real-time health indicator sequence before the current moment into the trained prediction model. Through the fixed-dimensional display feature mapping layer in the trained prediction model, the input data is mapped into a fixed-dimensional display feature vector. The display feature vector is then combined with the trained weight parameter vector for weighted summation, outputting the predicted health indicator sequence for the prediction period. The processor can compare the predicted value of each health indicator in the health indicator prediction sequence with the preset failure threshold, and determine the first moment when the preset failure threshold is reached or exceeded as the failure moment. This means that the bearing has lost its normal working ability when the failure moment is reached. The processor can use the difference between the failure moment and the prediction start moment as the remaining service life of the bearing.

[0052] For example, the preset failure threshold is a health indicator value of 0.5 m / s. 2 For example, data is sampled every hour for 200ms. The prediction time t0=350 hours is the prediction start time. Based on historical data from 1 to 349 hours, the prediction sequence of health indicators for the next 200 hours is determined. The predicted value of the health indicator determined at the 425th hour is 0.546. The value of the health indicator at the subsequent time all exceeds 0.546. Therefore, the preset failure threshold is exceeded for the first time at the 425th hour. Therefore, the remaining lifespan is 75 hours obtained by subtracting 350 from 425.

[0053] Example 2

[0054] Figure 2This is a flowchart illustrating a method for determining a remaining lifetime prediction model according to Embodiment 2 of the present invention. This embodiment is a further refinement of the above embodiment. Figure 2 As shown, the method includes:

[0055] S201. Obtain the historical full-life vibration data of the bearing, and determine the health index sequence based on the maximum amplitude of each sampling segment in the historical full-life vibration data.

[0056] In this embodiment, historical full-life vibration data can be understood as the raw vibration signal data collected by vibration sensors throughout the entire life cycle of the bearing, from its initial use to complete failure. A sampling segment can be understood as several small segments of continuously acquired vibration signals divided over time. The maximum amplitude refers to the largest amplitude value within the vibration signal sampling segment.

[0057] Specifically, the processor can acquire vibration signals continuously collected by the vibration sensor throughout the entire bearing's lifespan, and divide the collected raw vibration signals into multiple sampling segments according to time. For each sampling segment, the maximum amplitude value is extracted, which is the health indicator at that moment. The health indicators at all moments are arranged in chronological order to form a health indicator sequence.

[0058] For example, if each sample segment is S, then the maximum amplitude can be expressed as: The corresponding health indicator sequence HI can be represented as ,in, Represented as The first in A vibration signal sampling segment.

[0059] For example, a specific example can be used to demonstrate the location of the sensor. Figure 3 An example diagram of a sewing equipment is provided for a method of determining a remaining life prediction model according to Embodiment 2 of the present invention, as shown in the figure. Figure 3 As shown, sewing equipment (such as industrial sewing machines) includes the sewing machine spindle, bearing (rolling element) bearing housing, sewing machine frame, and vertical vibration sensor (S1) and horizontal vibration sensor (S2) installed on the bearing. Either direction can be selected. The vibration signal of the bearing is collected through the sensor during the entire operation process.

[0060] S202. Use the data of the first preset proportion in the health indicator sequence as the health benchmark to determine the health judgment interval.

[0061] In this embodiment, the pre-preset ratio can be understood as the proportion divided into healthy stages, for example, 70%. The health judgment interval can be understood as the numerical range used to determine whether the bearing deviates from a healthy state.

[0062] Specifically, the data from the earlier stages of the health indicator sequence correspond to the phase when the bearing is in a normal healthy state. A predetermined proportion of the data (e.g., 70%) is taken as the sample for the healthy phase. The mean (reflecting the central position of the healthy phase) and variance (reflecting the fluctuation range of the healthy phase) of these data are calculated. Based on this, a health judgment interval (usually the mean ± k times the standard deviation) is constructed. This interval represents the normal fluctuation range of the bearing's health indicators in a healthy state. As long as the health indicator values ​​are within this interval, the bearing can be considered to be in a healthy state.

[0063] For example, the preset proportion is 70%, the multiple k is 3, and the average value is calculated using the data of the top 70% of the HI sequence as the health benchmark. and variance To obtain the health assessment range At a certain moment, continuous When a HI value exceeds this range, the bearing is determined to have entered a degraded state from a healthy state, and the prediction of remaining service life is initiated from this moment.

[0064] S203. When a preset number of health indicator values ​​exceed the health judgment range, the last moment of the consecutive exceedance time is determined as the prediction start time.

[0065] In this embodiment, the preset quantity can be understood as a value that is continuously exceeded, so as to avoid misjudgment caused by a single fluctuation.

[0066] Specifically, as the bearing's operating time increases, the health indicator values ​​gradually increase and begin to deviate from the healthy stage. The processor can continuously monitor whether the health indicator values ​​at each moment exceed the health judgment range. To avoid misjudgment due to a single fluctuation, a preset number of consecutive exceedances (n) is used as the judgment condition. Only when n consecutive health indicator values ​​exceed the range is the bearing determined to have entered the degradation stage, and the last moment of the consecutive exceedances is taken as the prediction start moment.

[0067] S204. Using the health indicator sequence fragments before the prediction start time as the input features of the training samples, and using the health indicator values ​​after the corresponding time as the labels of the training samples, construct a training sample set.

[0068] In this embodiment, the label can be understood as the part of the training sample that serves as the correct answer, that is, the actual health indicator value at the corresponding time after the input sequence.

[0069] Specifically, for the health indicator sequence before the prediction start time, the sliding window method can be used to construct training samples. A fixed-length window (e.g., 7 points) is set, the continuous health indicator values ​​within the window are used as input features, and the health indicator values ​​of the next few time points (e.g., x time points, which can be set according to needs) immediately after the window are used as labels. Then the window slides forward one time point, and the above construction process is repeated to form multiple training samples, thus obtaining a training sample set.

[0070] S205. Based on the fixed-dimensional explicit feature mapping layer, the sample data in the training sample set is converted into a fixed-dimensional explicit feature vector.

[0071] In this embodiment, the explicit feature vector can be understood as a vector obtained through an explicit calculation formula.

[0072] Specifically, during training, the input data for each training sample (i.e., the health indicator sequence segment before the prediction start time, typically 7 HI values) is first processed through a fixed-dimensional explicit feature mapping layer. This mapping layer uses the Generalized Gaussian Distribution-based Stochastic Fourier Feature Method (GGD-RFF) to project the low-dimensional input data onto a fixed-dimensional feature space, obtaining a fixed-dimensional explicit feature vector (dimension M, where M is a preset fixed value).

[0073] Furthermore, based on the above embodiments, the step of converting sample data in the training sample set into fixed-dimensional explicit feature vectors based on a fixed-dimensional explicit feature mapping layer can be refined as follows:

[0074] The fixed-dimensional explicit feature mapping layer includes an explicit feature mapping function. The mapping dimension of the explicit feature mapping function is a preset fixed value. The explicit feature mapping function is constructed by sampling random frequency vectors from the generalized Gaussian distribution and random phases from the uniform distribution based on Bochner's theorem and the feature function of the generalized Gaussian distribution. The sample data in the training sample set is projected onto the fixed-dimensional feature space through the explicit feature mapping function to obtain the fixed-dimensional explicit feature vector.

[0075] In this embodiment, the explicit feature mapping function can be understood as a mapping function with a clear mathematical expression, used to project input data from the original space to the feature space. The mapping dimension can be understood as a set dimension value that determines the length of the feature vector and the model size. Bochner's theorem can be understood as the core theorem in harmonic analysis, guaranteeing that a positive definite kernel function can be represented by its Fourier transform. The characteristic function of the generalized Gaussian distribution refers to the characteristic function corresponding to the probability density function of the generalized Gaussian distribution, used to perform stochastic Fourier eigenvalue decomposition on the kernel function. The random frequency vector can be understood as the cosine term used to construct the explicit feature mapping function, controlling the frequency characteristics of the mapping function. The random phase can be understood as a random value sampled from a uniform distribution.

[0076] Specifically, the fixed-dimensional explicit feature mapping layer is the first processing stage in the model, and its core is the explicit feature mapping function. Input data is transformed from the original space to the feature space through this function. The mapping layer acts as a container, holding the mapping function and structural parameters such as the mapping dimension, together forming a complete feature mapping module. The explicit feature mapping function requires a specified dimension M during construction, which determines the length of the feature vector. M is preset during the model design phase and remains unchanged during training and prediction. Based on the feature function of the generalized Gaussian distribution, the mapping function is constructed according to Bochner's theorem. Bochner's theorem guarantees that the mapping function constructed using random Fourier features can approximate the inner product calculation of the kernel function. M random frequency vectors are sampled from the generalized Gaussian distribution, each with the same dimension as the input data. The generalized Gaussian distribution includes a shape parameter, which can be adjusted to adapt to the non-Gaussian features of the bearing degradation signal. M random phase values ​​are sampled from a uniform distribution. The phase value ensures that the mapping function satisfies translation invariance, guaranteeing insensitivity to translation of the input data. After the mapping function is constructed, during the training phase, the input data of each training sample is substituted into the function for calculation. After the input data is processed by the function, a numerical vector of length M is obtained, which is the explicit feature vector of the sample.

[0077] For example, input data Lifetime tag data is .

[0078] The characteristic function of the generalized Gaussian distribution is:

[0079]

[0080] in, Indicates shape parameters.

[0081] The feature mapping function is shown as follows:

[0082]

[0083] in, To display the feature vector, , From respectively Evenly distributed The random frequency vector and random phase obtained by sampling, For mapping dimensions.

[0084] S206. Based on the linear combination layer and the current weight parameter vector, a weighted sum is performed to obtain the health indicator prediction sequence of the current sample.

[0085] Specifically, the processor can feed the obtained fixed-dimensional explicit feature vector into the linear combination layer, calculate with the current weight parameter vector, and determine the health indicator prediction sequence with predicted values ​​for K future time moments.

[0086] S207. Determine the current loss value based on the health indicator prediction sequence, training sample labels, and the global convex loss function.

[0087] In this embodiment, the current loss value can be understood as the value calculated by substituting the prediction error of the current sample into the global convex loss function, which is used to drive the update of the weight parameters.

[0088] Specifically, the processor can compare the output health indicator prediction sequence with the labels of the training samples (i.e., the real health indicator values ​​at K future time points) to obtain K prediction errors. Then, it substitutes these K prediction errors into the global convex loss function to calculate K loss values ​​and then calculates the average value to obtain the current loss value.

[0089] S208. Iteratively update the current weight parameter vector using the conjugate direction method.

[0090] In this embodiment, the conjugate direction method can be understood as an iterative optimization algorithm that updates the weight parameters by searching for the optimal step size along the conjugate direction.

[0091] Specifically, based on the conjugate direction method, a search direction is determined in each iteration (this direction is conjugate with all previous search directions about the weighted correlation matrix), and then the optimal step size is determined along this direction. The weight parameter vector is moved from the current position to the new position, so that the objective function value gradually decreases until convergence.

[0092] Furthermore, based on the above embodiments, the steps for iteratively updating the current weight parameter vector using the conjugate direction method can be refined as follows:

[0093] Based on the current weight parameter vector, explicit feature vector, and corresponding auxiliary weight variables, determine the residual vector for the current iteration; based on the residual vector for the current iteration and the direction vector of the previous iteration, determine the conjugate direction vector for the current iteration; based on the conjugate direction vector and the weighted correlation matrix, determine the optimal step size for the current iteration, the weighted correlation matrix being calculated based on the explicit feature vectors of each sample and the corresponding auxiliary weight variables; update the weight parameter vector by adding the product of the optimal step size and the conjugate direction vector to the current weight parameter vector.

[0094] In this embodiment, the residual vector represents the gradient direction of the deviation between the current predicted value and the target value under the current weights. The direction vector can be understood as the search direction used to update the weight parameters in the conjugate direction method. The conjugate direction vector can be understood as direction vectors that satisfy the conjugate condition, ensuring that the search directions in each iteration are orthogonal to each other, thus guaranteeing algorithm convergence. The optimal step size can be understood as the step size value that maximizes the decrease in the objective function along the current conjugate direction. The weighted correlation matrix can be understood as a matrix calculated from the explicit feature vectors of each sample and the corresponding auxiliary weight variables.

[0095] Specifically, the processor first multiplies the explicit feature vectors of all training samples with their corresponding auxiliary weight variables, then compares this with the current weight parameter vector to obtain a residual vector reflecting the overall direction and degree of bias. The processor can then combine the current residual vector with the direction vector from the previous iteration, retaining the valuable parts of the previous iteration's direction while introducing a new direction indicated by the current residual, resulting in a conjugate direction vector. Based on these two factors, the optimal step size is calculated—that is, the step size along the current direction that maximizes the model's performance improvement in that direction. Based on the current weight parameter vector, the processor moves the optimal step size along the conjugate direction vector to a new position, which serves as the starting weight for the next iteration. The updated weight parameter vector, compared to the previous one, is closer to the optimal solution in the sense of the objective function (global convex loss function). By repeatedly executing this process, moving the optimal step size in the new conjugate direction each time, the weight parameter vector gradually approaches the global optimal solution that minimizes the objective function.

[0096] For example, the weighted least squares form can be further rewritten as a positive definite semi-quadratic form. , Represented as a weighted correlation matrix, This is represented as a weighted correlation cross-correlation vector. A conjugate gradient method combined with the exponential decay method is employed. and The iterative updates are represented as follows:

[0097] ,

[0098] in, For symbolic abbreviation, , It is a forgetting factor.

[0099] Weight vector It can be updated to:

[0100] Combination and The update equation can be used to obtain the update of the residual vector:

[0101] in, Let be the residual vector at time i+1. This represents the optimal step size for the i-th iteration. Let be the conjugate direction vector of the i-th iteration. This represents the displayed feature vector of the (i+1)th input sample.

[0102] The conjugate direction vector can be derived as follows: The step size of the conjugate gradient method is The conjugate parameter of the i-th iteration is It can be calculated using the following formula: ,

[0103] S209. Determine whether the iteration termination condition has been met. If yes, terminate the iteration; otherwise, continue the iteration update.

[0104] In this embodiment, the iteration termination condition can be understood as the judgment condition for terminating the iteration, such as the convergence of the weight parameters or the reaching of the maximum number of iterations.

[0105] Specifically, after each iteration, the processor can determine whether the training termination condition is met. If so, the iteration is terminated; otherwise, data is input for iterative updates.

[0106] The technical solution of this invention reduces computational and storage burdens, making it suitable for online deployment of sewing equipment. Traditional KAF methods grow linearly with the number of samples during online learning. This invention uses a stochastic Fourier feature method based on a generalized Gaussian distribution to transform the implicit kernel mapping of KAF into a fixed-dimensional explicit feature mapping, preventing the model structure from continuously expanding with online samples and effectively controlling computational and storage overhead. It enhances the stability of prediction results in complex noise environments: Sewing equipment operation is subject to interference such as start-up and shutdown shocks, mechanical vibrations, and fabric load changes, and health indicator sequences may exhibit non-Gaussian characteristics. By introducing a generalized Gaussian distribution and a robust KRSL loss function, the model's adaptability to non-Gaussian degradation features, shock components, and outliers is enhanced, reducing the impact of abnormal samples on model updates. It provides quantitative basis for bearing maintenance decisions in sewing equipment: The timing of bearing repair or replacement can be determined based on prediction results, avoiding interruptions to sewing production caused by sudden failures. Compared to regular maintenance or post-incident repairs, this significantly reduces maintenance costs and improves equipment utilization.

[0107] For example, training termination conditions typically include the following two scenarios: If the change in the weight parameter vector is less than a preset threshold over a preset number of iterations, it indicates that the model has converged to near the optimal solution, and further iterations will not significantly improve prediction performance; training should then be terminated. Alternatively, if the current iteration count reaches a preset maximum iteration count, training is forcibly terminated regardless of whether convergence is complete, to avoid unlimited consumption of computational resources.

[0108] For example, the algorithm proposed in this invention can be simulated and verified on a chaotic time series dataset, and compared with KAF algorithms such as LMS, KLMS, KRGMC, KRLS, and FrKMPE. The comparison results can be presented in the form of tables and graphs. Figure 4 This is an example diagram illustrating the effect of a method for determining a remaining lifetime prediction model provided in Embodiment 2 of the present invention. Figure 4 As shown, the evaluation metrics include mean squared error (MSE, in dB) and runtime (seconds). The table below shows the results of different KAFs on chaotic time series prediction. The table shows that the MSE of this invention is -31.53 dB, which is better than all the compared algorithms, verifying its prediction accuracy and stability. Furthermore, the runtime is only 0.0740 seconds, far lower than other KAF variants (such as FrKMPE, which is 10.449 seconds), making it easy to deploy on embedded controllers of sewing equipment.

[0109] Table 1 Comparison of Prediction Results

[0110]

[0111] Example 3

[0112] Figure 5 This is a schematic diagram of a device for determining a remaining lifetime prediction model according to Embodiment 3 of the present invention. Figure 5 As shown, the device includes:

[0113] The sample determination module 51 is used to acquire the historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data.

[0114] The model determination module 52 is used to train and optimize the pre-constructed kernel adaptive filtering model based on the sample data in the training sample set and the global convex loss function until the training termination condition is met, so as to obtain the trained remaining lifetime prediction model. The pre-constructed kernel adaptive filtering model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer and a weight parameter vector to be trained.

[0115] The technical solution of this invention acquires historical full-life vibration data of the bearing, determines the predicted start time of the bearing entering a degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data, and then trains and optimizes a pre-constructed kernel adaptive filter model based on the sample data in the training sample set and a global convex loss function until the training termination condition is met, resulting in a trained remaining life prediction model. The pre-constructed kernel adaptive filter model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained. This approach accurately locates the bearing degradation start point, extracts high-quality samples, and uses a global convex loss function for optimization, avoiding local optima. Combined with fixed-dimensional explicit feature mapping, it achieves nonlinear high-precision prediction and efficient computation, significantly improving the accuracy, stability, and real-time performance of remaining life prediction.

[0116] Furthermore, the sample determination module 51 is specifically used for:

[0117] Based on the maximum amplitude of each sampling segment in the historical life-cycle vibration data, a sequence of health indicators is determined;

[0118] The health assessment interval is determined by using the data of the first preset proportion in the health indicator sequence as the health benchmark.

[0119] When a preset number of health indicator values ​​exceed the health judgment range, the last moment of the consecutive exceedance time is determined as the prediction start time.

[0120] The training sample set is constructed by using health indicator sequence fragments before the predicted start time as input features of the training samples and health indicator values ​​after the corresponding time as labels of the training samples.

[0121] Furthermore, the model determination module 52 includes:

[0122] The first determining unit is used to convert the sample data in the training sample set into a fixed-dimensional explicit feature vector based on the fixed-dimensional explicit feature mapping layer.

[0123] The second determining unit is used to perform a weighted summation based on the linear combination layer and the current weight parameter vector to obtain the health indicator prediction sequence of the current sample.

[0124] The third determining unit is used to determine the current loss value based on the health indicator prediction sequence, training sample labels, and the global convex loss function.

[0125] The fourth determining unit is used to iteratively update the current weight parameter vector using the conjugate direction method;

[0126] The fifth determining unit is used to determine whether the iteration termination condition has been met. If so, the iteration is terminated; otherwise, the iteration continues.

[0127] Specifically, the first determining unit is used for:

[0128] The fixed-dimensional explicit feature mapping layer includes an explicit feature mapping function; the mapping dimension of the explicit feature mapping function is a preset fixed value, and the explicit feature mapping function is constructed by sampling random frequency vectors from the generalized Gaussian distribution and random phases from the uniform distribution based on Bochner's theorem and the feature function of the generalized Gaussian distribution.

[0129] The sample data in the training sample set is projected onto a fixed-dimensional feature space through the explicit feature mapping function to obtain a fixed-dimensional explicit feature vector.

[0130] Specifically, the fourth determining unit is used for:

[0131] The residual vector for the current iteration is determined based on the current weight parameter vector, the explicit feature vector, and the corresponding auxiliary weight variables.

[0132] Based on the residual vector of the current iteration and the direction vector of the previous iteration, determine the conjugate direction vector of the current iteration;

[0133] The optimal step size for the current iteration is determined based on the conjugate direction vector and the weighted correlation matrix, wherein the weighted correlation matrix is ​​calculated based on the explicit feature vectors of each sample and the corresponding auxiliary weight variables.

[0134] The weight parameter vector is updated by adding the product of the optimal step size and the conjugate direction vector to the current weight parameter vector.

[0135] Optionally, the device may also include a function building module.

[0136] The function construction module is specifically used for:

[0137] By introducing an auxiliary weight variable, the original exponential robust loss function is rewritten as the sum of the product of the auxiliary weight variable and the square of the error, plus the dual function of the auxiliary weight variable, resulting in the rewritten expression, where the error is the difference between the health indicator prediction sequence of the sample and the label.

[0138] Based on the rewritten expression, the optimal value function of the auxiliary weight variable is determined with the error as the independent variable, so that the auxiliary weight variable obtains an extreme value under any error, and the auxiliary weight variable is negatively correlated with the square of the error;

[0139] Substituting the optimal value function of the auxiliary weight variable back into the rewritten expression yields a global convex function in the form of weighted least squares.

[0140] Optionally, the device may also include an actual prediction module.

[0141] The actual prediction module is specifically used for:

[0142] Vibration signals during bearing operation are collected in real time, and a real-time health indicator sequence is determined based on the vibration signals.

[0143] Based on the statistical distribution of the health stages in the real-time health indicator sequence and the health judgment interval, the predicted start time of the bearing entering the degradation state is determined.

[0144] The real-time health indicator sequence before the prediction start time is input into the trained remaining lifespan prediction model, and the health indicator prediction sequence for the prediction period is output.

[0145] The predicted values ​​of each health indicator in the health indicator prediction sequence are compared with a preset failure threshold to determine the failure time corresponding to the predicted health indicator value that reaches or exceeds the preset failure threshold.

[0146] The remaining service life of the bearing is determined based on the difference between the failure time and the predicted start time.

[0147] The apparatus for determining the remaining lifetime prediction model provided in this embodiment of the invention can execute the method for determining the remaining lifetime prediction model provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0148] Example 4

[0149] Figure 6 A schematic diagram of an electronic device 40 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0150] like Figure 6As shown, the electronic device 40 includes at least one processor 41 and a memory, such as a read-only memory (ROM) 42 or a random access memory (RAM) 43, communicatively connected to the at least one processor 41. The memory stores computer programs executable by the at least one processor. The processor 41 can perform various appropriate actions and processes based on the computer program stored in the ROM 42 or loaded from storage unit 48 into the RAM 43. The RAM 43 may also store various programs and data required for the operation of the electronic device 40. The processor 41, ROM 42, and RAM 43 are interconnected via a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0151] Multiple components in electronic device 40 are connected to I / O interface 45, including: input unit 46, such as keyboard, mouse, etc.; output unit 47, such as various types of monitors, speakers, etc.; storage unit 48, such as disk, optical disk, etc.; and communication unit 49, such as network card, modem, wireless transceiver, etc. Communication unit 49 allows electronic device 40 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0152] Processor 41 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 41 performs the various methods and processes described above, such as the method for determining the remaining lifetime prediction model.

[0153] In some embodiments, the method for determining the remaining lifetime prediction model may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 48. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 40 via ROM 42 and / or communication unit 49. When the computer program is loaded into RAM 43 and executed by processor 41, one or more steps of the method for determining the remaining lifetime prediction model described above may be performed. Alternatively, in other embodiments, processor 41 may be configured to perform the method for determining the remaining lifetime prediction model by any other suitable means (e.g., by means of firmware).

[0154] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0155] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0156] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0157] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0158] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0159] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0160] In one embodiment, the present invention further includes a computer program product, which includes a computer program that, when executed by a processor, implements the method for determining the remaining lifetime prediction model of any embodiment of the present invention.

[0161] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0162] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0163] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for determining a remaining useful life prediction model, characterized in that, include: Acquire historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data; Based on the sample data in the training sample set and combined with the global convex loss function, the pre-constructed kernel adaptive filtering model is trained and optimized until the training termination condition is met, and the trained remaining lifetime prediction model is obtained. The pre-constructed kernel adaptive filtering model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer, and a weight parameter vector to be trained.

2. The method according to claim 1, characterized in that, The step of determining the predicted start time and training sample set for the bearing to enter a degradation state based on the signal amplitude of the historical full-life vibration data includes: Based on the maximum amplitude of each sampling segment in the historical life-cycle vibration data, a sequence of health indicators is determined; The health assessment interval is determined by using the data of the first preset proportion in the health indicator sequence as the health benchmark. When a preset number of health indicator values ​​exceed the health judgment range, the last moment of the consecutive exceedance time is determined as the prediction start time. The training sample set is constructed by using health indicator sequence fragments before the predicted start time as input features of the training samples and health indicator values ​​after the corresponding time as labels of the training samples.

3. The method according to claim 1, characterized in that, The process of training and optimizing a pre-constructed kernel adaptive filtering model based on sample data from the training sample set and a global convex loss function until the training termination condition is met, resulting in a trained remaining lifetime prediction model, includes: Based on the fixed-dimensional explicit feature mapping layer, the sample data in the training sample set is converted into a fixed-dimensional explicit feature vector. The health indicator prediction sequence of the current sample is obtained by weighted summation based on the linear combination layer and the current weight parameter vector. The current loss value is determined based on the predicted health indicator sequence, training sample labels, and the global convex loss function. The current weight parameter vector is iteratively updated using the conjugate direction method; Determine if the iteration termination condition has been met; if so, terminate the iteration; otherwise, continue the iteration update.

4. The method according to claim 3, characterized in that, The step of converting sample data in the training sample set into fixed-dimensional explicit feature vectors based on the fixed-dimensional explicit feature mapping layer includes: The fixed-dimensional explicit feature mapping layer includes an explicit feature mapping function; the mapping dimension of the explicit feature mapping function is a preset fixed value, and the explicit feature mapping function is constructed by sampling random frequency vectors from the generalized Gaussian distribution and random phases from the uniform distribution based on Bochner's theorem and the feature function of the generalized Gaussian distribution. The sample data in the training sample set is projected onto a fixed-dimensional feature space through the explicit feature mapping function to obtain a fixed-dimensional explicit feature vector.

5. The method according to claim 3, characterized in that, The iterative update of the current weight parameter vector using the conjugate direction method includes: The residual vector for the current iteration is determined based on the current weight parameter vector, the explicit feature vector, and the corresponding auxiliary weight variables. Based on the residual vector of the current iteration and the direction vector of the previous iteration, determine the conjugate direction vector of the current iteration; The optimal step size for the current iteration is determined based on the conjugate direction vector and the weighted correlation matrix, wherein the weighted correlation matrix is ​​calculated based on the explicit feature vectors of each sample and the corresponding auxiliary weight variables. The weight parameter vector is updated by adding the product of the optimal step size and the conjugate direction vector to the current weight parameter vector.

6. The method according to claim 1, characterized in that, The steps for constructing the global convex loss function are as follows: By introducing an auxiliary weight variable, the original exponential robust loss function is rewritten as the sum of the product of the auxiliary weight variable and the square of the error, plus the dual function of the auxiliary weight variable, resulting in the rewritten expression, where the error is the difference between the health indicator prediction sequence of the sample and the label. Based on the rewritten expression, the optimal value function of the auxiliary weight variable is determined with the error as the independent variable, so that the auxiliary weight variable obtains an extreme value under any error, and the auxiliary weight variable is negatively correlated with the square of the error; Substituting the optimal value function of the auxiliary weight variable back into the rewritten expression yields a global convex function in the form of weighted least squares.

7. The method according to claim 1, characterized in that, Also includes: Vibration signals during bearing operation are collected in real time, and a real-time health indicator sequence is determined based on the vibration signals. Based on the statistical distribution of the health stages in the real-time health indicator sequence and the health judgment interval, the predicted start time of the bearing entering the degradation state is determined. The real-time health indicator sequence before the prediction start time is input into the trained remaining lifespan prediction model, and the health indicator prediction sequence for the prediction period is output. The predicted values ​​of each health indicator in the health indicator prediction sequence are compared with a preset failure threshold to determine the failure time corresponding to the predicted health indicator value that reaches or exceeds the preset failure threshold. The remaining service life of the bearing is determined based on the difference between the failure time and the predicted start time.

8. A device for determining a remaining useful life prediction model, characterized in that, include: The sample determination module is used to acquire the historical full-life vibration data of the bearing, and determine the predicted start time of the bearing entering the degradation state and the training sample set based on the signal amplitude of the historical full-life vibration data. The model determination module is used to train and optimize the pre-constructed kernel adaptive filtering model based on the sample data in the training sample set and the global convex loss function until the training termination condition is met, so as to obtain the trained remaining lifetime prediction model. The pre-constructed kernel adaptive filtering model includes a fixed-dimensional explicit feature mapping layer, a linear combination layer and a weight parameter vector to be trained.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method for determining the remaining lifetime prediction model according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for determining the remaining lifetime prediction model according to any one of claims 1-7.

Images