Method and device for predicting service life of barrel based on gradient enhancement and quantile recursive network

Abstract

The embodiment of the invention provides a barrel service life prediction method and device based on gradient enhancement and a quantile recursive network, and the method comprises the steps: collecting the multi-source sensor data of the whole life cycle of a barrel, and carrying out the preprocessing, and forming a standardized data set; constructing an XGBoost model, performing feature importance analysis on the standardized data set through the XGBoost model, screening key features based on an analysis result to obtain a data set after feature selection, and performing optimization training on the XGBoost model through gradient lifting and regularization methods; dividing the data set after feature selection into a training set and a test set; in combination with a quantile regression method, constructing a life prediction model based on a quantile recurrent neural network, training the life prediction model by using the training set, and training and optimizing model parameters through a bifurcated sequence to obtain an optimized barrel life prediction model; and inputting a test set into the model, outputting a residual service life prediction result containing a prediction value and a confidence interval, and evaluating the prediction performance.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the technical field of artillery barrel life prediction and related technical fields, and in particular, to a barrel life prediction method and device based on gradient enhancement and quantile recursive network. Background Art

[0002] As a core component of artillery, the barrel's lifespan directly determines its service life. When barrel damage reaches a certain level, the artillery will cease to function properly. Therefore, accurate barrel lifespan prediction is crucial for artillery maintenance and combat effectiveness.

[0003] Scholars from various countries have conducted research on barrel life prediction from different perspectives. The main factors affecting barrel life include chemical erosion of the inner barrel surface metal by gunpowder gas and projectiles, mechanical wear, and high-temperature ablation. Through analysis of the ablation mechanism, based on the assumption that the radial expansion of the barrel is caused by melt erosion on the inner surface of the barrel, and applying the semi-infinite object assumption and the theoretical model of the inner surface melt layer, a quantitative calculation method for the thickness of the melt layer on the inner wall of the barrel was derived. However, the mechanism of barrel life termination involves the interaction of multiple factors such as high temperature and high pressure, chemical erosion, and mechanical wear, which is very complex. Traditional theoretical models based on ablation mechanisms have low prediction accuracy due to the difficulty in accurately constructing mathematical models, and cannot meet actual combat needs.

[0004] In recent years, machine learning-based lifespan prediction methods have gained increasing attention. By analyzing small sample sizes of barrel degradation data, machine learning methods have been shown to effectively predict barrel lifespan. However, the small sample sizes, non-intervals, and nonlinearity of barrel degradation data place high demands on model robustness, generalizability, and uncertainty quantification. Existing machine learning models still have shortcomings in addressing these issues, making them difficult to meet the actual operational needs of military forces.

[0005] Therefore, there is an urgent need for an efficient, accurate and quantifiable uncertainty barrel life prediction method to break through the limitations of traditional theoretical models and existing machine learning methods and provide strong support for barrel maintenance and assurance. Summary of the Invention

[0006] The embodiments described herein provide a method and device for predicting barrel life based on gradient enhancement and quantile recursive networks, which solve the problems existing in the prior art.

[0007] According to a first aspect of the present disclosure, a barrel life prediction method based on gradient boosting and quantile recursive network is provided, comprising:

[0008] collecting multi-source sensor data over the entire life cycle of the barrel, and preprocessing the multi-source sensor data to form a standardized data set;

[0009] Constructing an XGBoost model, performing feature importance analysis on the standardized dataset using the XGBoost model, screening key features based on the analysis results to obtain a feature-selected dataset, and optimizing and training the XGBoost model using gradient boosting and regularization methods;

[0010] Dividing the feature-selected dataset into a training set and a test set;

[0011] Combining the quantile regression method, a life prediction model is constructed based on a quantile recursive neural network, the life prediction model is trained using the training set, and the model parameters are optimized through bifurcation sequence training to obtain an optimized barrel life prediction model;

[0012] The test set is input into the optimized barrel life prediction model, and a remaining service life prediction result including a predicted value and a confidence interval is output, and the prediction performance is evaluated.

[0013] In some embodiments of the present disclosure, the step of preprocessing the multi-source sensor data includes:

[0014] Performing time sequence alignment processing on the multi-source sensor data;

[0015] Outlier detection and correction based on statistical distribution;

[0016] Frequency domain filtering to eliminate high-frequency noise;

[0017] Select the data normalization method according to the working condition type, including: linear normalization for a single working condition, and clustering followed by group normalization for multiple working conditions to ensure that the data are in the same dimension.

[0018] In some embodiments of the present disclosure, after performing feature importance analysis on the standardized dataset using the XGBoost model and screening key features based on the analysis results, the following steps are included:

[0019] Eliminate redundant features and retain high-importance features;

[0020] The step of optimizing the XGBoost model by gradient boosting and regularization method comprises:

[0021] Residual fitting is used to improve accuracy, the second-order approximation of the loss function is used to accelerate training, the regularization term is used to control the complexity of the model, and pruning processing based on information gain is adopted, and the pruning criterion is that the information gain is non-negative.

[0022] In some embodiments of the present disclosure, before the step of performing feature importance analysis on the standardized dataset using the XGBoost model, the method further includes:

[0023] The XGBoost model is constructed based on the regression decision tree of the CART algorithm, and its node splitting criterion is minimization of the squared error.

[0024] In some embodiments of the present disclosure, the step of building an XGBoost model includes:

[0025] Based on the objective function minimization criterion, the optimal feature split point is determined through iterative search, wherein the objective function includes: a square error loss term, a regularization penalty term, and a second-order Taylor expansion approximation term;

[0026] Recursively partitioning the data set to obtain the regression tree decision tree until at least one of the following termination conditions is met: the node sample size is lower than a preset threshold, the error reduction is less than a set value, and the maximum tree depth limit is reached;

[0027] The XGBoost model is obtained by combining multiple regression decision trees.

[0028] In some embodiments of the present disclosure, the quantile recurrent neural network includes:

[0029] The encoder module uses a recurrent neural network to extract temporal features;

[0030] The decoder module includes a global feature decoder and a local digit decoder, wherein:

[0031] The global feature decoder captures long-term dependencies and outputs quantile predictions;

[0032] The local quantile decoder responds to the periodicity and mutation of the data and outputs quantile predictions.

[0033] In some embodiments of the present disclosure, the step of bifurcated sequence training includes:

[0034] Convert the artillery life cycle monitoring data into time series feature representation;

[0035] Dynamically determine multiple prediction starting points on time series data based on preset time interval rules; synchronously execute at each prediction starting point: context feature extraction, multi-scale feature decoding, and uncertainty quantification prediction;

[0036] The prediction error loss and physical constraints are integrated to construct a multi-objective optimization function.

[0037] In some embodiments of the present disclosure, after the step of evaluating the prediction performance, the method further includes:

[0038] Deploy the trained model to actual applications, collect sensor data in real time, and predict the remaining life of the barrel;

[0039] Formulate a barrel maintenance plan based on the prediction results to ensure the normal use of the artillery.

[0040] According to a second aspect of the present disclosure, a barrel life prediction device based on gradient boosting and quantile recursive network is provided, comprising:

[0041] a preprocessing module for collecting multi-source sensor data throughout the life cycle of the barrel and preprocessing the multi-source sensor data to form a standardized data set;

[0042] A feature selection module is used to build an XGBoost model, perform feature importance analysis on the standardized dataset using the XGBoost model, screen key features based on the analysis results, obtain a feature-selected dataset, and optimize and train the XGBoost model using gradient boosting and regularization methods;

[0043] A partitioning module, used to divide the dataset after feature selection into a training set and a test set;

[0044] a training module for constructing a life prediction model based on a quantile recursive neural network in combination with a quantile regression method, training the life prediction model using the training set, and optimizing model parameters through bifurcation sequence training to obtain an optimized barrel life prediction model;

[0045] The prediction module is used to input the test set into the optimized barrel life prediction model, output a remaining service life prediction result including a predicted value and a confidence interval, and evaluate the prediction performance.

[0046] In some embodiments of the present disclosure, the preprocessing module is specifically configured to:

[0047] Performing time sequence alignment processing on the multi-source sensor data;

[0048] Outlier detection and correction based on statistical distribution;

[0049] Frequency domain filtering to eliminate high-frequency noise;

[0050] Select the data normalization method according to the working condition type, including: linear normalization for a single working condition, and clustering followed by group normalization for multiple working conditions to ensure that the data are in the same dimension.

[0051] In some embodiments of the present disclosure, the feature selection module is further configured to:

[0052] Eliminate redundant features and retain high-importance features;

[0053] The step of optimizing the XGBoost model by gradient boosting and regularization method comprises:

[0054] Residual fitting is used to improve accuracy, the second-order approximation of the loss function is used to accelerate training, the regularization term is used to control the complexity of the model, and pruning processing based on information gain is adopted, and the pruning criterion is that the information gain is non-negative.

[0055] In some embodiments of the present disclosure, the feature selection module is further configured to:

[0056] The XGBoost model is constructed based on the regression decision tree of the CART algorithm, and its node splitting criterion is minimization of the squared error.

[0057] In some embodiments of the present disclosure, the feature selection module is specifically configured to:

[0058] Based on the objective function minimization criterion, the optimal feature split point is determined through iterative search, wherein the objective function includes: a square error loss term, a regularization penalty term, and a second-order Taylor expansion approximation term;

[0059] Recursively partitioning the data set to obtain the regression tree decision tree until at least one of the following termination conditions is met: the node sample size is lower than a preset threshold, the error reduction is less than a set value, and the maximum tree depth limit is reached;

[0060] The XGBoost model is obtained by combining multiple regression decision trees.

[0061] In some embodiments of the present disclosure, the quantile recurrent neural network includes:

[0062] The encoder module uses a recurrent neural network to extract temporal features;

[0063] The decoder module includes a global feature decoder and a local digit decoder, wherein:

[0064] The global feature decoder captures long-term dependencies and outputs quantile predictions;

[0065] The local quantile decoder responds to the periodicity and mutation of the data and outputs quantile predictions.

[0066] In some embodiments of the present disclosure, the training module is specifically used to:

[0067] Convert the artillery life cycle monitoring data into time series feature representation;

[0068] Dynamically determine multiple prediction starting points on time series data based on preset time interval rules;

[0069] Executed synchronously at each prediction starting point, including: context feature extraction, multi-scale feature decoding, and uncertainty quantification prediction;

[0070] The prediction error loss and physical constraints are integrated to construct a multi-objective optimization function.

[0071] In some embodiments of the present disclosure, further comprising:

[0072] The deployment module is used to deploy the trained model into actual applications, collect sensor data in real time, and predict the remaining life of the barrel;

[0073] The formulation module is used to formulate a barrel maintenance plan based on the prediction results to ensure the normal use of the artillery.

[0074] According to a third aspect of the present disclosure, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method in any one of the above embodiments are implemented.

[0075] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method in any one of the above embodiments are implemented.

[0076] The disclosed embodiments provide a barrel life prediction method and device based on gradient boosting and quantile recurrent network. The method includes steps such as data preprocessing, CART model construction, XGBoost model training, feature selection, MQ-RNN model training, and RUL prediction. By integrating the high-precision feature selection capability of XGBoost and the time series prediction capability of MQ-RNN, an efficient, accurate, and uncertainty-quantifiable prediction model is constructed. XGBoost and MQ-RNN are combined for barrel life prediction, which significantly improves the accuracy and reliability of barrel life prediction. At the same time, the uncertainty of the prediction results can be quantified, providing strong support for barrel maintenance and assurance, thereby addressing the shortcomings of existing technologies and improving prediction accuracy and practicality.

[0077] The above description is only an overview of the technical solutions of the embodiments of the present application. In order to more clearly understand the technical means of the embodiments of the present application, they can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiments of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly described below. It should be noted that the drawings described below only relate to some embodiments of the present disclosure and are not intended to limit the present disclosure.

[0079] Figure 1 This is a flow chart of a barrel life prediction method based on gradient enhancement and quantile recursive network provided by an embodiment of the present disclosure;

[0080] Figure 2 is a schematic diagram of the basic architecture of CART provided by an embodiment of the present disclosure;

[0081] Figure 3 is a schematic diagram of the structure of a quantile recurrent neural network provided by an embodiment of the present disclosure;

[0082] Figure 4 Schematic diagram of the structure of the LSTM model provided by the embodiment of the present disclosure;

[0083] Figure 5 Schematic diagram of the structure of the MLP hidden layer provided by the embodiment of the present disclosure;

[0084] Figure 6 Schematic diagram of a barrel life prediction device based on gradient enhancement and quantile recursive network provided by an embodiment of the present disclosure;

[0085] Figure 7 It is a structural diagram of a computer device provided by an embodiment of the present disclosure.

[0086] In the drawings, reference numerals having the same last two digits correspond to the same elements. It should be noted that the elements in the drawings are schematic and not drawn to scale. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work also fall within the scope of protection of the present disclosure.

[0088] References to "embodiments" herein mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase "embodiment" in various places in the specification does not necessarily refer to the same embodiment, nor does it necessarily refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0089] The term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists, A and B exist, and B exists. Additionally, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0090] Furthermore, in all embodiments of the present disclosure, terms such as “first” and “second” are used only to distinguish one component (or a portion of a component) from another component (or another portion of a component).

[0091] In the description of this application, unless otherwise specified, "plurality" means more than two (including two), and similarly, "multiple groups" means more than two (including two).

[0092] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0093] As a core component of artillery, the barrel's lifespan directly impacts its operational effectiveness. Traditional theoretical models based on ablation mechanisms are difficult to accurately construct due to the complex nature of the barrel's end-of-life mechanism, resulting in low prediction accuracy. In recent years, machine learning methods have been increasingly applied to barrel life prediction, but they still face challenges with processing small samples, non-intervals, and nonlinear data. Existing models lack robustness, generalizability, and uncertainty quantification, failing to meet actual operational requirements.

[0094] Based on the problems existing in the existing technology, Figure 1 This is a flow chart of a barrel life prediction method based on gradient enhancement and quantile recursive network provided by an embodiment of the present disclosure, such as Figure 1 As shown in the figure, the specific process of the barrel life prediction method based on gradient enhancement and quantile recurrent network includes:

[0095] S110 , in the data preparation stage, collecting multi-source sensor data of the entire life cycle of the barrel, and pre-processing the multi-source sensor data to form a standardized data set.

[0096] In one possible implementation, the multi-source sensor data of the barrel throughout its life cycle includes one or more of temperature, vibration, chamber pressure, bore wear, shot count, projectile attitude, material composition, microstructure, dimensional accuracy, residual stress, static pressure testing, dynamic fatigue, initial velocity and trajectory, cracks and defects, corrosion, coating degradation, ambient temperature and humidity, geographic positioning, and usage history.

[0097] The step of preprocessing the multi-source sensor data includes:

[0098] Performing time alignment processing on the multi-source sensor data to ensure time synchronization of the multi-source sensor data;

[0099] Outlier detection and correction based on statistical distribution;

[0100] Frequency domain filtering to eliminate high-frequency noise;

[0101] The data normalization method is selected according to the working condition type, including: linear Min-Max normalization for a single working condition, and k-means clustering followed by group normalization for multiple working conditions to ensure that the data are in the same dimension.

[0102] Specifically, normalization is to limit the preprocessed data to a fixed range, such as [0, 1] or [-1, 1], to avoid a series of troubles caused by single sample data.

[0103] Operating conditions refer to the specific operating states or environmental conditions experienced by a device or system during operation. For artillery barrels, operating conditions can be understood as the state of the gun under different operating conditions, such as varying firing rates, charge loads, ambient temperatures, and humidity. These conditions directly impact barrel wear and lifespan.

[0104] A single operating condition refers to a situation where a device or system is consistently operating under the same or similar operating conditions and environmental conditions. For a gun barrel, a single operating condition means that all ejection data and bore wear data are collected under the same firing frequency, charge, ambient temperature, and other conditions. For example, a gun can fire continuously at a fixed firing frequency (e.g., 1 round per minute) and the same charge, with all data collected at the same ambient temperature (e.g., 25°C). In this case, the data is considered a single operating condition.

[0105] Multiple operating conditions refer to the different operating states or environmental conditions experienced by a device or system during operation. For a gun barrel, multiple operating conditions mean that ejection data and bore wear data are collected under different conditions, such as firing rates, charge loads, and ambient temperatures. For example, a certain type of gun fires at different firing rates (e.g., 1 round per minute, 2 rounds per minute), with different charge loads (e.g., full charge, half charge), and at different ambient temperatures (e.g., 10°C, 25°C, 40°C). In this case, the data covers multiple operating conditions.

[0106] Optionally, it is determined whether the normalized data is a single working condition. If the normalized data is a single working condition, the Min-Max normalization method is directly used to normalize the original data to the range of [0, 1] by performing a linear transformation on the original data, that is,

[0107]

[0108] Where x is the original measurement value of a sensor (such as chamber pressure, temperature) at a certain moment (for example, the chamber pressure peak value x = 400 MPa during a certain firing), x max and x min They are the maximum and minimum values of the sensor in the sample data set, max is the maximum value and min is the minimum value. For example: temperature sensor: x = 300 ° C (the surface temperature of the rifling of a certain shot), vibration sensor: x = 5.2 m / s 2 (RMS value of vibration during a certain firing), chamber pressure sensor: x = 450 MPa (peak pressure during a certain firing).

[0109] Optionally, if the standardized data involves multiple working conditions, the data are first grouped by a clustering algorithm, and then each group of data is normalized to ensure that the model can accurately capture the characteristics under different working conditions, that is,

[0110]

[0111] Where i is a positive integer, k is the number of clusters, and the value of k is usually consistent with the number of types of working conditions. i,k is the data collected by the i-th sensor under the k-th working condition, m i,k and ν i,k are the variances of the average values of the original measurements of the i-th sensor under the k-th working condition.

[0112] Optionally, the clustering algorithm may be a hybrid entropy weight clustering algorithm, including:

[0113] The feature weight is calculated based on the index entropy value, and the standardized data is clustered to generate the working condition grouping label. The formula of the index entropy value is:

[0114] I i =∑(p ij lnp ij ) (3)

[0115] Among them, I i is the entropy value of the ith indicator, which is used to measure the discrete degree or uncertainty of the data under this indicator. ij is the standardized probability value (or proportion) of the jth sample under the i-th indicator, lnp ij For p ij The natural logarithm of , ∑ is the summation symbol, and the sum is taken for all samples j = 1, 2, …, n to calculate the entropy contribution of all samples under this indicator.

[0116] The formula for weight is:

[0117]

[0118] Among them, w i is the weight of the i-th indicator, I i is the entropy value of the i-th indicator, ∑ k (1-I k ) is 1 k The information utility values of all indicators k are summed up to normalize the weights to ensure ∑ i w i =1.

[0119] The improved k-means algorithm is used to iteratively correct the cluster centers to ensure the accuracy of the clustering results.

[0120] After the hybrid entropy weight clustering step, the working condition grouping labels are generated.

[0121] For different working condition data, Min-Max normalization or clustering followed by normalization is used to limit the data to an appropriate range, avoid the problem of single data, and optimize the data preprocessing process.

[0122] S130, constructing an XGBoost model, performing feature importance analysis on the standardized data set by the XGBoost model, screening key features based on the analysis results, obtaining a feature-selected data set, and optimizing and training the XGBoost model by gradient boosting and regularization methods;

[0123] XGBoost (Extreme Gradient Boosting): A decision tree-based ensemble learning algorithm, it integrates multiple CART trees. It improves accuracy by continuously constructing new barrel bore wear regression models based on the residual between the previous tree's prediction and the actual value. A second-order Taylor expansion of the loss function improves training efficiency, and the tree complexity is incorporated into the objective function to avoid overfitting, thereby improving the model's predictive performance.

[0124] CART (Classification and Regression Tree): A typical binary tree model used to solve classification and regression problems. For barrel life prediction, the number of ejections is divided into several parts, each corresponding to a corresponding amount of barrel bore wear. The division process is based on the greedy principle, with split points selected based on minimizing the squared error. Ultimately, a barrel bore wear regression decision tree is constructed.

[0125] In some embodiments, before step S120, the process further includes:

[0126] The XGBoost model is constructed based on the regression decision tree of the CART algorithm, and its node splitting criterion is minimization of the squared error.

[0127] Optionally, the step of constructing an XGBoost model based on the regression tree decision of the CART algorithm includes:

[0128] Based on the objective function minimization criterion, the optimal feature split point is determined through iterative search, wherein the objective function includes: a square error loss term, a regularization penalty term, and a second-order Taylor expansion approximation term;

[0129] The data set is recursively partitioned to obtain the regression tree decision tree until at least one of the following termination conditions is met: the node sample size is lower than a preset threshold (the node sample size threshold is 0.5%-1% of the total sample size), the error reduction is less than a set value (the error reduction setting value is 0.01-0.05), and the maximum tree depth limit is reached (the maximum tree depth is 3-10 layers);

[0130] The XGBoost model is obtained by combining multiple regression decision trees.

[0131] In a specific implementation process, after performing feature importance analysis on the standardized dataset using the XGBoost model and screening key features based on the analysis results, the following steps are included:

[0132] Eliminate redundant features and retain high-importance features.

[0133] The CART model is constructed to solve the problems of small sample, non-interval, and nonlinearity of barrel degradation data. CART is a typical binary tree structure. It divides the ejection number into several parts, so that each part corresponds to a barrel inner diameter wear value. The basic structure of CART is as follows: Figure 2 As shown, Figure 2 In this example, X represents the data feature, and c represents the output. For the ejection volume, we assign it to a unit based on the feature and obtain the corresponding barrel inner diameter wear value.

[0134] Given a barrel test training dataset:

[0135] D={(x1,y1),(x2,y2),…,(x n ,y n )} (5)

[0136] Among them, x is the ejection amount, y is the corresponding barrel inner diameter wear value, and n is the amount of barrel test training data.

[0137] The partitioning of the ejection count values is based on the greedy principle. Each partitioning step lists all ejection count values in the current dataset one by one. Based on the square error minimization criterion, the point where the error between the predicted and actual barrel diameter wear values is minimized is selected as the split point. Assuming that is the split point, the barrel diameter wear prediction can be divided into two regions:

[0138] R1(s)={x|x≤s} (6)

[0139] R2(s)={x|x>s} (7)

[0140] Among them, x is the ejection amount, s is the number of ejections in the training set, and R1 and R2 are two data regions divided by the ejection number threshold.

[0141] The expression of the sum of square errors of the two regions after division is:

[0142]

[0143] Where i is a positive integer, x is the ejection amount, y is the corresponding barrel inner diameter wear value, n is the amount of barrel test training data, R1 and R2 are two data areas divided by the ejection number threshold, c1 and c2 are the average wear values of areas R1 and R2 respectively, x is the average wear value of areas R1 and R2 respectively, and i is the specific value of the artillery ejection number, y i is the corresponding ejection number x i The inner diameter wear value of the barrel.

[0144] After solving Equation (11) to obtain the optimal cut-off point s for the number of ejections, the ejection input is divided into two intervals. The above division process is then repeated for the two intervals until the set threshold is met. Finally, a barrel inner diameter wear regression decision tree is constructed.

[0145] In the specific implementation process, the XGBoost model is used to calculate feature importance, remove redundant features, and improve data quality;

[0146] XGBoost is used for feature selection. It is an ensemble learning algorithm based on decision trees, designed for speed and performance. XGBoost aims to achieve the effect of a strong classifier by combining various weak classifiers. By analyzing multidimensional features, XGBoost can determine the importance of features and filter input data based on their importance.

[0147] For a dataset D with n examples and m features, i ,y i )},(|D|=n,x i ∈R m ,y i ∈R), the tree ensemble model uses K additional functions to predict the output

[0148]

[0149] in, For sample x i The predicted value of φ(x i ) is the prediction function, x i is the feature vector of the i-th sample, f k (xi ) is the kth basis function, For the accumulation of infinite numbers, F={f(x)=ω q(x)},(q:R m →T,ω∈R T ) is the space of regression trees. q is the structure of each tree that maps the dataset to the corresponding leaf weights. T is the number of leaves in the tree, i is a positive integer, x is the ejection volume, y is the corresponding barrel bore wear value, and n is the amount of barrel test training data.

[0150] The model is optimized using the following objective function

[0151]

[0152] is the loss function for a single sample, is the model’s predicted value for the i-th sample, y i is the true label or target value of the i-th sample, is the regularization term, γ and λ are the regularization coefficients.

[0153] For each iteration of the model, the objective function of the previous iteration is optimized. The prediction value of the tth iteration is

[0154]

[0155] Among them, The cumulative prediction value of the i-th sample after the t-th iteration (or the t-th tree), To represent the historical prediction value of the i-th sample after the previous iteration (t-1), f t (x i ) is the base learner (such as decision tree) added in the tth iteration for the sample x i The predicted output, x i is the feature vector of the i-th sample.

[0156] The objective function L can be transformed into

[0157]

[0158] Among them, y i is the corresponding ejection number x i The barrel inner diameter wear value, represents the historical prediction value of the i-th sample after the previous iteration (tt-1 times), f t (x i ) is the base learner (such as decision tree) added in the tth iteration for the sample x i The predicted output, x i is the feature vector of the i-th sample, is the regularization term.

[0159] In order to minimize the objective function, f in formula (14) t (x i ) performs a second-order Taylor expansion and deletes the constant term,

[0160]

[0161] in, and y i is the corresponding ejection number x i The barrel inner diameter wear value, represents the historical prediction value of the i-th sample after the previous iteration (tt-1 times), f t (x i ) is the base learner (such as decision tree) added in the tth iteration for the sample x i The predicted output, x i is the feature vector of the i-th sample, is the regularization term.

[0162] Let I j ={i|q(x i )=j} is the set of leaf j. Substitute into formula (14), then

[0163]

[0164] in, and y i is the corresponding ejection number x i The barrel inner diameter wear value, represents the historical prediction value of the i-th sample after the previous iteration (tt-1 times), f t (x i ) is the base learner (such as decision tree) added in the tth iteration for the sample x i The predicted output, x i is the feature vector of the i-th sample, γ and λ are regularization coefficients, and T is the number of leaves in the tree.

[0165] Minimize Equation (16) to obtain feature importance,

[0166]

[0167] Among them, I j ={i|q(x i )=j} is the set of leaf j, λ is the regularization coefficient.

[0168] Will Substitute into formula (17), then

[0169]

[0170] Among them, I j ={i|q(x i )=j} is the set of leaf j, λ is the regularization coefficient and T is the number of leaves in the tree.

[0171] Using XGBoost for feature selection can effectively analyze the importance of multi-dimensional features, remove redundant features, and improve data quality.

[0172] In a specific implementation process, the step of optimizing the XGBoost model by gradient boosting and regularization method includes:

[0173] Residual fitting is used to improve accuracy, the second-order approximation of the loss function is used to accelerate training, the regularization term is used to control the complexity of the model, and pruning is performed based on information gain. The pruning criterion is that the information gain is non-negative, specifically including:

[0174] Improve accuracy by fitting the residual between the previous tree's predicted value and the actual value;

[0175] Perform a second-order Taylor expansion on the loss function to speed up training;

[0176] Add model complexity regularization terms to the objective function to avoid overfitting;

[0177] Post-pruning is performed based on the information gain criterion. When the information gain is less than zero, the splitting is stopped, and the hyperparameters of the tree are set to limit the complexity of the decision tree. The hyperparameters of the tree include but are not limited to the maximum depth and the minimum number of samples of leaf nodes.

[0178] Information Gain (Gain): When building a decision tree based on XGBoost, a metric used to determine whether to split a node. It is defined as the loss before the split minus the loss after the split. A positive Gain indicates that the greater the loss before the split, the more likely the split is worthwhile. When Gain < 0, the split is stopped. Hyperparameters such as the maximum tree depth and the minimum number of leaf nodes are also used to limit the complexity of the decision tree and prevent overfitting.

[0179] XGBoost is the integrated result of multiple CART trees. It improves accuracy by continuously constructing new barrel inner diameter wear regression decisions to fit the residual between the predicted value and the actual value of the previous tree, and uses the second-order Taylor expansion of the loss function to improve training efficiency. At the same time, the complexity of the tree is added to the objective function to avoid overfitting.

[0180] Assume that the mathematical relationship between the expected number of ejections and the amount of wear on the barrel inner diameter is:

[0181]

[0182] F={f(x)=w q(x)} (20)

[0183] q:R→{1,2,…,T},w∈R T (twenty one)

[0184] in, is the predicted value of barrel inner diameter wear, F is the decision tree structure set, K is the number of decision trees, q is the tree structure of sample mapping to leaf nodes, T is the number of leaf nodes, w is the real number score of leaf nodes, R is a real number set, indicating that the input feature space is a real number domain (single variable), R T It is a T-dimensional real vector space (each tree leaf weight vector), which stores the leaf weight of each tree. For sample x i The predicted value of (such as remaining life), f k is the kth base model (such as a decision tree), K is the total number of base models (the complexity of the ensemble model), n is the number of samples (i.e., the number of instances in the dataset), which is a positive integer, q(x) is the partition function that maps samples to integer indices, i.e., the leaf node allocation function of the decision tree, and w_(q(x)) is the leaf node weight (prediction value).

[0185] Based on the principle of minimizing the objective function (Obj), the model parameters are solved to achieve the optimal state. In order to ensure that the prediction model is infinitely close to the actual value while avoiding the risk of overfitting, the model complexity factor is added to the objective function. The objective function can be written as:

[0186]

[0187] in, is the error function term; is the model complexity function term, is the predicted value of barrel inner diameter wear of the i-th sample, f k is the kth base model, K is the number of decision trees, and i is a positive integer.

[0188] The essence of the XGBoost model is to add a new decision function to the model while keeping the original model unchanged, so that the objective function is as close to 0 as possible. The specific function construction process is as follows:

[0189]

[0190]

[0191] in, is the predicted value of the t-th model; f t (x i ) is the new function added for the tth time. At this time, the objective function can be expressed as:

[0192]

[0193] in, For sample x i The predicted value of is the predicted value of the t-1th model, Ω(f t ) is the complexity penalty term (regularization term) of the tree, f t (x i ) is the new function added for the tth time.

[0194] In order to find the parameter value that makes the objective function the lowest, the original objective function is transformed through the second-order Taylor expansion to obtain the approximate objective function:

[0195]

[0196] in, For sample x i The predicted value of is the predicted value of the t-1th model, g i is the loss function l about The first partial derivative of h i is the loss function l about The second-order partial derivative of Ω(f t ) is the complexity penalty term (regularization term) of the tree, f t (x i ) is the new function added for the tth time.

[0197] Since the structure of t-1 trees has been determined, the constant term can be removed and the approximate objective function can be expressed as:

[0198]

[0199] Among them, g i is the loss function l about The first partial derivative of h i is the loss function l about The second-order partial derivative of Ω(f t ) is the complexity penalty term (regularization term) of the tree, f t (x i ) is the new function added for the tth time.

[0200] The complex terms of the decision tree can be expressed as:

[0201]

[0202] Among them, γT is the L1 regularization term, is the L2 regularization term, and λ is the weight decay coefficient (the strength parameter of the L2 regularization term), which is used to control the square and penalty terms of the leaf node weights to prevent the model from becoming complicated due to excessive weights, thereby improving the generalization ability.

[0203] Then the approximate objective function can be expressed as:

[0204]

[0205] in, G is the sum of the second-order partial derivatives of leaf node j for all samples belonging to node j, which is a constant and is recorded as G j . It is the cumulative sum of the second-order partial derivatives of leaf node j for all samples belonging to node j, which is a constant and is recorded as H j , w j is the output value of the jth leaf node, that is, the leaf node value when the model predicts, I j is the sample set belonging to the jth leaf node, γT is the L1 regularization term, is the L2 regularization term, λ is the weight attenuation coefficient, and T is the number of leaf nodes.

[0206] Derivate the objective function to obtain the weight of each leaf node of the decision tree And the optimal objective function value:

[0207]

[0208] Among them, G j Right now is the cumulative sum of the second-order partial derivatives of leaf node j for all samples belonging to node j, which is a constant, H j Right now is the cumulative sum of the second-order partial derivatives of leaf node j for all samples belonging to node j, is a constant, λ is the weight attenuation coefficient, and γT is the L1 regularization term.

[0209] By repeatedly recursively calling the construction process of the barrel inner diameter wear regression decision tree, a large number of regression decision tree models are obtained. The optimal tree structure is found based on the self-scaling function and integrated into the existing decision model to finally generate the barrel life prediction model with the best fitting effect.

[0210] Optionally, the process of generating the data feature filtering model further includes:

[0211] Model pruning is performed based on information gain (Gain), and splitting stops when Gain is less than 0;

[0212] Set hyperparameters such as the maximum depth of the tree and the minimum number of samples for leaf nodes to limit the complexity of the decision tree.

[0213] To ensure that the decision tree built based on XGBoost does not continue to split, which would lead to overfitting of the model, the Gain model is pruned. For the current node and the corresponding dataset, the loss before splitting is:

[0214]

[0215] Among them, G is the first-order gradient (sum of total gradients) of the parent node before splitting, H is the second-order gradient of the child node after splitting, γ is to prevent complex models from overfitting discrete data points, and λ is the weight decay coefficient.

[0216] The sum of the losses of the left and right subtrees split at a certain split point is the loss after splitting:

[0217]

[0218] Among them, G L is the first-order gradient (sum of gradients) of the left child node after splitting, G R is the first-order gradient (sum of gradients) of the right child node after splitting, H L is the second-order gradient of the left child node after the split, H R is the second-order gradient of the right child node after splitting, γ is to prevent complex models from overfitting discrete data points, and λ is the weight decay coefficient.

[0219] Define information gain as the loss before splitting minus the loss after splitting. Gain can be expressed as:

[0220]

[0221] Among them, G L is the first-order gradient (sum of gradients) of the left child node after splitting, G R is the first-order gradient (sum of gradients) of the right child node after the split, H is the second-order gradient of the child node after the split, and H L is the second-order gradient of the left child node after the split, H R is the second-order gradient of the right child node after splitting, γ is to prevent complex models from overfitting discrete data points, and λ is the weight decay coefficient.

[0222] If Gain is positive and the value is large, it means that the loss before the split is greater than the loss after the split, which means the split is more worthwhile. Otherwise, the split is less worthwhile. Traverse all current leaf nodes and select the split point with the maximum Gain. When Gain < 0, stop splitting. At the same time, set hyperparameters such as the maximum tree depth and the minimum number of leaf node samples to limit the complexity of the decision tree.

[0223] XGBoost is based on CART integration, which improves accuracy by fitting residuals and accelerates training by second-order Taylor expansion to prevent overfitting.

[0224] S130, dividing the feature-selected data set into a training set and a test set;

[0225] In a specific embodiment, the training set is used to train the model, enabling it to learn patterns and features in the data. The test set is used to evaluate the model's performance and verify its predictive ability on unseen data. By dividing the training set into two test sets, we can avoid model overfitting (i.e., the model performs well on the training set but poorly on new data) and ensure the model's generalization ability.

[0226] Optional and commonly used data partitioning methods include:

[0227] Random Partition: The dataset is randomly shuffled and then divided into training and test sets proportionally. This method is suitable for evenly distributed data. For example, if the dataset contains 1,000 data sets, 800 sets are randomly selected as the training set, and the remaining 200 sets are used as the test set.

[0228] Time Series Partitioning: For time series data (such as barrel life data), the training and test sets are divided chronologically. Typically, the earlier data is used as the training set, and the later data is used as the test set. This method better simulates real-world application scenarios. For example, if the dataset contains data from January 2020 to December 2023, the data from January 2020 to December 2022 is used as the training set, and the data from January 2023 to December 2023 is used as the test set.

[0229] Stratified sampling: When data is imbalanced, the training and test sets are divided according to the class ratio to ensure that each class is represented in both the training and test sets. For example, if the dataset contains different operating conditions (such as high temperature, normal temperature, and low temperature), the training and test sets are divided according to the operating condition ratio.

[0230] By properly dividing the data, we can ensure that the model learns effective rules during training and verify its generalization ability during testing.

[0231] S140, in a model training phase (offline phase), combining a quantile regression method with a quantile recurrent neural network to construct a life prediction model, using the training set to train the life prediction model, and optimizing model parameters through bifurcation sequence training to obtain an optimized barrel life prediction model;

[0232] In a specific embodiment, the training set is input into the barrel life prediction model and trained.

[0233] Optionally, the bifurcation sequence training adopts a dynamic bifurcation point generation strategy.

[0234] The quantile recurrent neural network (MQ-RNN) adopts an encoder-decoder structure, such as Figure 3 Shown, including:

[0235] The encoder module uses a recurrent neural network to extract temporal features and generate hidden states;

[0236] The decoder module includes a global feature decoder and a local digit decoder, wherein:

[0237] The decoder includes a global MLP and a local MLP, which are used to capture global information and local features respectively;

[0238] The global feature decoder captures long-term dependencies and outputs quantile predictions;

[0239] The local quantile decoder responds to the periodicity and mutation of the data and outputs a quantile prediction. Optionally, the quantile prediction range covers the [0.1, 0.5, 0.9] quantiles.

[0240] Quantile regression is used to train the model by minimizing the quantile loss function to quantify the uncertainty of the prediction results.

[0241] The quantile regression achieves quantile prediction through global MLP and local MLP.

[0242] This algorithm uses LSTM to encode all historical information into hidden states, and designs two MLPs (global and local) as decoders. It can handle large-scale time series regression problems. By minimizing the quantile loss function and employing a bifurcated sequence training scheme, it reduces data requirements and training time compared to traditional moving window approaches, making the model more stable.

[0243] MQ-RNN can handle large-scale time series regression problems.

[0244]

[0245] Among them, y iis the i-th time series to be predicted, are the time covariates available in the history, is the future time covariate, It is a time-invariant feature.

[0246] Specifically, the input information of the encoder is and y :t,i , after LSTM encoding, the output hidden state h t In the decoder, a global MLP and a local MLP are used. The input of the global MLP is the hidden state h t and future input The output is d t+N , N=1,2,...,n and d s , where the output of the global MLP can be expressed as

[0247]

[0248] Among them, d t+N , N=1,2,…n is the information captured by n points in the future, d s It is independent of time and is used to capture global information. g It is to combine historical state and feature input to generate future observations or control instructions, h t is the hidden state and It is future input.

[0249] The local MLP can respond well to the periodicity of the data and some mutations of the data, and is used for each specific horizontal plane. The input of the local MLP is the future input and the output of the global MLP, d t+N , N=1,2,...,n and d s The output of the local MLP is all the quantiles required for a specific future time step

[0250]

[0251] Among them, M l (·) is a local MLP, q(·) represents each Q quantile. Quantile regression fits different quantiles of the target value by giving different sites, d t+N The information captured for the next n points, d s Used to capture global information, is the future time covariate.

[0252] Train the prediction model by minimizing the quantile loss function

[0253]

[0254] Where y is the actual value, is the predicted value, q is the quantile coefficient, (·) + =max(0,·). The MQ-RNN prediction model is trained by minimizing the total quantile loss, i.e.

[0255]

[0256] Where t is the time, q is the task or variable, and N is the prediction step size.

[0257] MQ-RNN differs from other Seq2Seq architecture models in that it uses a forking-sequences training scheme to improve its performance. Many time series prediction models use a moving window scheme to split a time series into subsequences, called split sequences, but this requires extensive data. In the MQ-RNN prediction method, this is unnecessary due to the forking-sequences training scheme. The forking-sequences training scheme treats each time series of arbitrary length as a single sample, eliminating the need for large amounts of data in experiments and significantly reducing training time. Furthermore, the forking-sequences training scheme utilizes all available information at once, making the model more stable.

[0258] Quantile regression: A regression analysis method that fits different quantiles of a target value by giving different loci (quantiles). In the MQ-RNN model, training the prediction model by minimizing the quantile loss function can more comprehensively reflect the distribution characteristics of the data and quantify the uncertainty of the prediction results.

[0259] In a specific embodiment, a bifurcated sequence training strategy is adopted to train a RUL prediction model based on LSTM encoding and dual MLP decoding structure, wherein,

[0260] LSTM (Long Short-Term Memory) is the core temporal feature extraction module. Its unique gating mechanism effectively solves the gradient problem of traditional RNNs. Its innovative gating structure consists of three key components: the forget gate, the input gate, and the output gate. The forget gate controls the retention of historical information; the input gate regulates the input of new information; and the output gate manages the output of the current state.

[0261] These three gating units work together through trainable weight parameters to effectively capture long-term dependencies, fundamentally solve the gradient explosion / vanishing problem, and efficiently model complex temporal patterns.

[0262] The structure of LSTM is as follows Figure 4 shown.

[0263] The gate control unit is mainly composed of three logical control units: forget gate, input gate and output gate. These three logical control units control the three stages of LSTM by setting weights. The formula of forget gate is:

[0264] f t =σ(W f ·[ t-1 ,x t ]+b f ) (42)

[0265] The formula for the input gate is:

[0266] i t =σ(W i ·[h t-1 ,x t ]+b i ) (43)

[0267] The formula for the output gate is:

[0268] o t =σ(W o ·[h t-1 ,x t ]+b o ) (44)

[0269] Among them, x t is the input data, h t-1 is the hidden state at the last moment, W f 、W i and W o is the weight matrix in each gating unit, b f 、b i and b o is the deviation, t is the time, f t It is the forget gate, i t is the input gate, o t is the output gate.

[0270] The update of the cell state in LSTM can be described as:

[0271]

[0272] Among them C t is the cell state at time t, C t-1 is the cell state at time t-1, is the candidate vector to be added to the cell state, W c represents the weight matrix of the tanh activation function layer, b cis the bias, t is the time, c is the standard symbolic convention in LSTM, representing the calculation related to the candidate cell state, and f t is the forget gate, h t-1 is the hidden state at the last moment, x t is the input data. Finally, the output gate and tanh function layer are used to generate the unit network output h t , which can be expressed as

[0273] h t =o t *tanh(C t ) (47)

[0274] Among them, t is the time, o t is the output gate, C t is the cell state at time t.

[0275] Multilayer Perceptron (MLP) is a basic neural network architecture that uses a fully connected structure, including input layer, hidden layer and output layer. The hidden layer can have many layers. The simplest MLP has only one hidden layer, such as Figure 5 As shown in the figure, each layer is fully connected: the input layer receives data, the intermediate layers perform computations, and the output layer outputs the results. In this paper's MQ-RNN prediction model, the MLP uses the ReLU activation function. Compared to the Sigmoid and Tanh activation functions, this effectively alleviates the vanishing gradient problem, improves computational efficiency by over 40%, and accelerates model convergence. Furthermore, the ReLU activation function only has linear relationships, making both forward and backward propagation faster than the Sigmoid and Tanh activation functions.

[0276] The formula for the ReLU activation function is:

[0277] ReLU(x)=max(x,0) (48)

[0278] Here, x is the input value, such as the eigenvalue output by the convolutional layer, which may physically represent sensor signal values such as voltage or current. ReLU maps negative values to zero (suppression) and leaves positive values intact (activation). It is often used to transmit key features and is suitable for processing highly noisy sensor data.

[0279] In a specific embodiment, the step of bifurcated sequence training adopts a method of minimizing a quantile loss function, including:

[0280] Convert the artillery life cycle monitoring data into time series feature representation;

[0281] Dynamically determine multiple prediction starting points on time series data based on preset time interval rules;

[0282] Executed synchronously at each prediction starting point, including: context feature extraction, multi-scale feature decoding, and uncertainty quantification prediction;

[0283] The multi-objective optimization function is constructed by integrating the prediction error loss and physical constraints. The complete time series is retained without slicing.

[0284] Start the decoding process at any time point to generate multiple parallel prediction output branches;

[0285] Improving prediction stability through full-sequence context modeling.

[0286] In a specific implementation, a bifurcated sequence training strategy is adopted, i.e., minimizing the quantile loss function to train the model, which can better fit different quantiles of the target value, improve the accuracy and reliability of the prediction, and form a significant technical difference from the traditional moving window method.

[0287] Traditional method: A sequence of length T needs to be cut into K subsequences (K = T-(input + output window)), generating redundant data

[0288] Bifurcation method: Maintaining sequence integrity and achieving end-to-end learning through intelligent decoding.

[0289] Technical principle:

[0290] 1. Sequence length preservation: retain the complete time step (t=1→T) of a single time series sample without slicing to avoid information loss.

[0291] 2. Parallel decoding mechanism: Start the decoding process at any time point t (t>L_min, L_min = minimum training length) and start multiple parallel prediction output branches.

[0292] 3. Spatiotemporal correlation learning: Leveraging the inherent periodicity and trend characteristics of time series, full-sequence context modeling is used to improve prediction stability and enhance prediction robustness.

[0293] The specific implementation steps are as follows:

[0294] Step 1: Sequence preprocessing

[0295] Input: gun full cycle degradation sequence {x1,x2,...,x T}(T is the total number of ejections);

[0296] Output: corresponding to the RUL sequence {y t0 ,y t0+1 ,...,y T}(t0 is the initial observation point);

[0297] Encoder initialization:

[0298] h t =LSTM(x t ,h t-1 ),t∈[1,t0] (49)

[0299] Among them, h t is the hidden state of the LSTM unit at time step t, which contains the sequence information at the current moment, x t represents the input data at time step t (such as sensor readings, word embeddings, trajectory coordinates, etc.), h t-1 is the hidden state of the previous time step t-1. LSTM passes historical information through this parameter. t∈[1,t0] is the running time range of the encoder, from the initial time t=1 to t0. t0 is the predefined time window length of the encoder, which is used to initialize the model state. LSTM(·) is the operation process of the LSTM unit. According to the current input x t and the previous state h t-1 Update current status h t .

[0300] The input sequence is processed by LSTM in the time window [1, t0] to generate the hidden state h t Serves as the initial context for subsequent tasks.

[0301] Step 2: Generate bifurcation point

[0302] Define the set of fork start times:

[0303]

[0304] Among them, τ is the safety time interval (typical value 3-5 time steps), Δ is the preset prediction step size (related to the artillery maintenance cycle), is the set of all possible bifurcation start times, t is the specific time step of the candidate bifurcation point, t0 is the end time of encoder initialization, and T is the total time step of the sequence.

[0305] Starting from t0+τ and ending at T-Δ, a suitable time point is selected as the bifurcation starting position, which may be used for multimodal prediction or dynamic path planning.

[0306] Step 3: Parallel decoding process

[0307] For each bifurcation point

[0308] 1. Context Extraction:

[0309]

[0310] Among them, W c is a learnable parameter matrix used to linearly transform the hidden state, is the bifurcation point t p The context vector at , aggregates the information of the historical hidden state, From time 1 to t p All hidden state sequences, h k is the hidden state (LSTM output) of the kth time step, t p is the time step of the current bifurcation point.

[0311] 2. Global MLP decoding: obtaining temporal features:

[0312]

[0313] in, is the bifurcation point t p The global time series features at , used for subsequent prediction, is the bifurcation point t p The hidden state of is the context vector, [; 1 is the vector concatenation operation, MLP g It is a multi-layer perceptron (MLP) used for feature fusion.

[0314] 3. Quantile prediction:

[0315]

[0316] in, At the bifurcation point t p At , the predicted value corresponding to quantile q, q is the target quantile (such as 0.1, 0.5, 0.9), indicating different percentiles of the predicted distribution. It is an independent multi-layer perceptron with dedicated parameters for each quantile q.

[0317] Capture uncertainty (e.g., 10%, 50%, 90% quantiles) by predicting different quantiles through independent MLPs.

[0318] Step 4: Loss function calculation:

[0319] Use quantile loss weighted superposition:

[0320]

[0321] in, is the total loss function, is the bifurcation point t p The true value of At the bifurcation point t pAt, the predicted value corresponding to the quantile q, ρ q (·) is the quantile loss function, and Q is the quantile set.

[0322] The quantile loss function is:

[0323]

[0324] Where, ∈ is the prediction error q is the current quantile, which adjusts the weight of positive and negative errors.

[0325] Quantile loss ρ q Assign different weights to over / underestimation errors to optimize the prediction distribution.

[0326] Example: Visual illustration of the forking training process:

[0327] Input sequence: [x1 x2 x3 x4 x5 x6 x7 x8 x9 x 10 ];

[0328] Bifurcation point generation: t=5,τ=2→effective bifurcation point t p =5,6,7, must satisfy t p ≥t0+τ and t p ≤T-Δ;

[0329] Predicted Path:

[0330] ⑤──→MLP decoding→

[0331] ⑥──→MLP decoding→

[0332] ⑦──→MLP decoding→

[0333] Each bifurcation point t p Generate a prediction sequence for the next Δ steps (e.g., t p =5 predictions arrive ).

[0334] Loss calculation: Sum the quantile losses of the true values y5 to y8.

[0335] Fork t p =5, predicting the next 3 steps (i.e. ).

[0336] The true value is y5~y8.

[0337] Quantile set Q = {0.1, 0.5, 0.9}.

[0338] Calculate the quantile loss at each time step

[0339] For each time step t∈{5,6,7,8} and quantile q∈Q, compute:

[0340] t=5,q=0.1:

[0341] ∈=10-8=2≥0→ρ 0.1 =0.1×2=0.2

[0342] t=5,q=0.9:

[0343] ∈=10-12=-2<0→ρ 0.9 =(0.9-1)×(-2)=0.2

[0344] t=6,q=0.5:

[0345] ∈=15-15=2≥0→ρ 0.5 =0

[0346] Step 2: Accumulate the losses of all bifurcation points and quantiles

[0347] Total loss calculation:

[0348] Bifurcation point t p =5 covers t=5,6,7,8.

[0349] For each t and q, calculate ρ p And sum:

[0350]

[0351] MQ-RNN uses a bifurcated sequence training scheme. Compared to traditional moving window schemes, this eliminates the need for extensive data expansion, reduces training time, and makes the model more stable. By training the model by minimizing the quantile loss function, it can better fit different quantiles of the target value, improving the accuracy and reliability of predictions.

[0352] S150. In the model prediction stage (online stage), the test set is input into the optimized barrel life prediction model, a remaining service life prediction result including a predicted value and a confidence interval is output, and the prediction performance is evaluated.

[0353] In a specific implementation, barrel life is defined as the total time from commissioning until the gun fails due to accumulated damage. It essentially reflects the gun's overall service life. This lifespan is primarily governed by the following factors: chemical corrosion from propellant gases, mechanical friction loss from projectiles, and thermal ablation at high temperatures.

[0354] Remaining useful life (RUL) represents the usable life of a barrel in its current state. The barrel life prediction model constructed based on this embodiment can accurately estimate the RUL value, providing data support for maintenance decisions, enabling preventive maintenance planning, scientifically regulating usage intensity, and continuously ensuring combat effectiveness.

[0355] The remaining service life is predicted by the constructed barrel life prediction model, which provides a basis for the maintenance of the barrel, so that the maintenance plan can be arranged in advance to ensure the normal use of the barrel.

[0356] In some embodiments, after step S150, the method further includes:

[0357] S160, deploying the trained barrel life prediction model to actual applications, collecting sensor data in real time and predicting the remaining barrel life;

[0358] S170. Formulate a barrel maintenance plan based on the prediction results to ensure normal use of the artillery.

[0359] The disclosed embodiments provide a barrel life prediction method based on gradient boosting and quantile recurrent networks, including data preprocessing, CART model construction, XGBoost model training, feature selection, MQ-RNN model training, and RUL prediction. By integrating XGBoost's high-precision feature selection capabilities with MQ-RNN's time series prediction capabilities, an efficient, accurate, and uncertainty-quantified prediction model is constructed, significantly improving the accuracy and reliability of barrel life prediction. It can also quantify the uncertainty of the prediction results, providing strong support for barrel maintenance.

[0360] On the basis of the above embodiments, the present disclosure also provides a barrel life prediction device based on gradient enhancement and quantile recursive network, such as Figure 6 As shown, the barrel life prediction device based on gradient enhancement and quantile recursive network includes:

[0361] An acquisition module 610 is configured to acquire multi-source sensor data throughout the barrel's life cycle and pre-process the multi-source sensor data to form a standardized data set;

[0362] A feature selection module 620 is configured to construct an XGBoost model, perform feature importance analysis on the standardized dataset using the XGBoost model, screen key features based on the analysis results, obtain a feature-selected dataset, and optimize and train the XGBoost model using gradient boosting and regularization methods;

[0363] A partitioning module 630 is configured to partition the feature-selected dataset into a training set and a test set;

[0364] a training module 640 for constructing a life prediction model based on a quantile recurrent neural network in combination with a quantile regression method, training the life prediction model using the training set, and optimizing model parameters through bifurcation sequence training to obtain an optimized barrel life prediction model;

[0365] The prediction module 650 is configured to input the test set into the optimized barrel life prediction model, output a remaining service life prediction result including a predicted value and a confidence interval, and evaluate the prediction performance.

[0366] In a specific embodiment, the preprocessing module is specifically used to:

[0367] Performing time sequence alignment processing on the multi-source sensor data;

[0368] Outlier detection and correction based on statistical distribution;

[0369] Frequency domain filtering to eliminate high-frequency noise;

[0370] Select the data normalization method according to the working condition type, including: linear normalization for a single working condition, and clustering followed by group normalization for multiple working conditions to ensure that the data are in the same dimension.

[0371] In a specific embodiment, the feature selection module is further used to:

[0372] Eliminate redundant features and retain high-importance features;

[0373] The step of optimizing the XGBoost model by gradient boosting and regularization method comprises:

[0374] Residual fitting is used to improve accuracy, the second-order approximation of the loss function is used to accelerate training, the regularization term is used to control the complexity of the model, and pruning processing based on information gain is adopted, and the pruning criterion is that the information gain is non-negative.

[0375] In a specific embodiment, the feature selection module is further used to:

[0376] The XGBoost model is constructed based on the regression decision tree of the CART algorithm, and its node splitting criterion is minimization of the squared error.

[0377] In a specific embodiment, the feature selection module is specifically used to:

[0378] Based on the objective function minimization criterion, the optimal feature split point is determined through iterative search, wherein the objective function includes: a square error loss term, a regularization penalty term, and a second-order Taylor expansion approximation term;

[0379] Recursively partitioning the data set to obtain the regression tree decision tree until at least one of the following termination conditions is met: the node sample size is lower than a preset threshold, the error reduction is less than a set value, and the maximum tree depth limit is reached;

[0380] The XGBoost model is obtained by combining multiple regression decision trees.

[0381] In a specific embodiment, the quantile recurrent neural network includes:

[0382] The encoder module uses a recurrent neural network to extract temporal features;

[0383] The decoder module includes a global feature decoder and a local digit decoder, wherein:

[0384] The global feature decoder captures long-term dependencies and outputs quantile predictions;

[0385] The local quantile decoder responds to the periodicity and mutation of the data and outputs quantile predictions.

[0386] In a specific embodiment, the training module is specifically used to:

[0387] Convert the artillery life cycle monitoring data into time series feature representation;

[0388] Dynamically determine multiple prediction starting points on time series data based on preset time interval rules;

[0389] Executed synchronously at each prediction starting point, including: context feature extraction, multi-scale feature decoding, and uncertainty quantification prediction;

[0390] The prediction error loss and physical constraints are integrated to construct a multi-objective optimization function.

[0391] In a specific embodiment, it also includes:

[0392] The deployment module is used to deploy the trained model into actual applications, collect sensor data in real time, and predict the remaining life of the barrel;

[0393] The formulation module is used to formulate a barrel maintenance plan based on the prediction results to ensure the normal use of the artillery.

[0394] The disclosed embodiments provide a barrel life prediction device based on gradient boosting and quantile recurrent networks. By integrating the high-precision feature selection capabilities of XGBoost with the time series prediction capabilities of MQ-RNN, an efficient, accurate, and uncertainty-quantified prediction model is constructed. This significantly improves the accuracy and reliability of barrel life prediction, while also quantifying the uncertainty of the prediction results, providing strong support for barrel maintenance.

[0395] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0396] The following is an example to illustrate:

[0397] Predicting the Remaining Life of a Self-Propelled Gun Barrel: Over long-term use, the barrel of a self-propelled gun degrades due to high-temperature, high-pressure gas erosion and projectile friction. To accurately predict the remaining useful life (RUL) of the barrel, this example analyzes barrel lifecycle monitoring data and develops a prediction model based on XGBoost and MQ-RNN.

[0398] 1. Goal

[0399] The remaining number of shots in the gun barrel is predicted using chamber pressure-vibration time series data, with an error within ±5% and a 90% confidence interval.

[0400] 2. Data Preparation

[0401] Data source: 500 sets of barrel life cycle data collected on a 1:100 scale test bench.

[0402] Sampling frequency: 10kHz.

[0403] Data channels: temperature (4 channels), vibration (6 channels), chamber pressure (3 channels), strain (5 channels).

[0404] Data preprocessing:

[0405] Timing alignment: Dynamic Time Warping (DTW) algorithm is used to align multi-source signals;

[0406] Abnormal processing: 3σ criterion is used to eliminate abnormal points, and linear interpolation is used to fill missing values;

[0407] Filtering and noise reduction: Butterworth low-pass filter (cut-off frequency 1kHz);

[0408] Working condition classification: k-means clustering identifies three typical working conditions (full charge / half charge / special environment).

[0409] 3. Feature Selection

[0410] XGBoost parameter settings:

[0411] Tree depth: 6

[0412] Learning rate: 0.1

[0413] Regularization coefficient: λ = 1, γ = 0.5

[0414] 15 key features were selected (such as chamber pressure rise rate, third-order vibration mode energy, etc.) 4. Model construction

[0415] XGBoost Model:

[0416] Build 50 CART regression trees

[0417] Objective function:

[0418] Use early stopping strategy (the error of the validation set does not decrease for 5 consecutive rounds)

[0419] MQ-RNN model:

[0420] Encoder:

[0421] Bidirectional LSTM (128 hidden units)

[0422] Input window: 20 time steps

[0423] Decoder:

[0424] Global MLP: 3-layer fully connected (256-128-64)

[0425] Local MLP: Independent branches handle [0.1, 0.5, 0.9] quantiles

[0426] Split Training:

[0427] Safety interval Δs = 4 time steps

[0428] Prediction step Δp = 10 (corresponding to 100 live-fire shootings)

[0429] 5. Model Training

[0430] Training parameters:

[0431] Training set: 400 groups (80%)

[0432] Test set: 100 groups (20%)

[0433] Optimizer: Adam (lr=0.001)

[0434] Batch size: 32

[0435] Loss function:

[0436] 6. Prediction results

[0437] Performance indicators:

[0438] MAE: 2.3 shots (<3%)

[0439] MSE: 8.2 shots 2

[0440] Interval coverage: 91.2% (target 90%)

[0441] Typical output:

[0442] Predicted remaining life: 85 ± 4 shots (90% confidence interval)

[0443] Recommended maintenance time: After the 80th shot

[0444] 7. Results and Discussion

[0445] Model performance: The model has high prediction accuracy on the test set, with an MSE of 8.3 shots and a MAE of 2.3 shots, which meets the design error requirements.

[0446] Uncertainty quantification: The quantile regression results provide confidence intervals for the predicted values, enhancing the credibility of the prediction results.

[0447] Practical application value: This model can provide a scientific basis for artillery barrel maintenance and help troops formulate reasonable maintenance plans and extend the service life of the barrel.

[0448] This example combines the strengths of XGBoost and MQ-RNN to construct an efficient and accurate barrel life prediction model. Experimental results demonstrate that the model can effectively predict the remaining number of rounds in a barrel and quantify the uncertainty of the prediction results, demonstrating its significant practical application value.

[0449] The present application also provides a computer device. Figure 7 , Figure 7 This is a basic structural block diagram of the computer device in this embodiment.

[0450] The computer device includes a memory 710 and a processor 720 that are interconnected and communicate with each other via a system bus. It should be noted that the figure only shows a computer device with components 710-720, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0451] Computer devices can be desktop computers, laptops, PDAs, cloud servers, etc. Computer devices can interact with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0452] The memory 710 includes at least one type of readable storage medium, including non-volatile memory or volatile memory, such as flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, magnetic disk, optical disk, etc. The RAM may include static RAM or dynamic RAM. In some embodiments, the memory 710 may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory 710 may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, or a Flash Card equipped on the computer device. Of course, the memory 710 may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory 710 is generally used to store the operating system and various application software installed on the computer device, such as the program code of the above-mentioned method. In addition, the memory 710 may also be used to temporarily store various types of data that have been output or are about to be output.

[0453] The processor 720 is generally used to perform the overall operation of the computer device. In this embodiment, the memory 710 is used to store program code or instructions, which include computer operating instructions. The processor 720 is used to execute the program code or instructions stored in the memory 710 or process data, such as the program code for running the above method.

[0454] In this article, a bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. This bus system can be divided into address buses, data buses, and control buses. For ease of illustration, the figure uses only one thick line, but this does not mean that there is only one bus or only one type of bus.

[0455] Another embodiment of the present application further provides a computer-readable medium, which may be a computer-readable signal medium or a computer-readable medium. A processor in a computer reads the computer-readable program code stored in the computer-readable medium, enabling the processor to execute the functional actions specified in each step or combination of steps in the above method, and to generate a device that implements the functional actions specified in each block or combination of blocks in the block diagram.

[0456] Computer-readable media include but are not limited to electronic, magnetic, optical, electromagnetic, infrared memory or semiconductor systems, devices or apparatuses, or any appropriate combination of the foregoing, the memory is used to store program codes or instructions, the program codes include computer operating instructions, and the processor is used to execute the program codes or instructions of the above-mentioned methods stored in the memory.

[0457] For the definitions of memory and processor, please refer to the description of the aforementioned computer device embodiment and will not be repeated here.

[0458] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0459] Each functional unit or module in each embodiment of the present application may be integrated into a processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The above-mentioned integrated units may be implemented in the form of hardware or software functional units.

[0460] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0461] Unless the context clearly indicates otherwise, as used herein and in the appended claims, the singular includes the plural, and vice versa. Thus, when referring to the singular, the plural of the corresponding term is generally included. Similarly, the words "include" and "comprising" are to be interpreted as inclusive rather than exclusive. Likewise, the terms "include" and "or" should be interpreted as inclusive unless such interpretation is expressly prohibited herein. Where the term "example" is used herein, particularly when it follows a group of terms, the "example" is merely exemplary and illustrative and should not be considered exclusive or comprehensive.

[0462] Further aspects and scope of adaptability become apparent from the description provided herein. It should be understood that various aspects of the present application can be implemented individually or in combination with one or more other aspects. It should also be understood that the description and specific embodiments herein are intended to be illustrative only and are not intended to limit the scope of the present application.

[0463] Several embodiments of the present disclosure have been described in detail above, but it is obvious that those skilled in the art can make various modifications and variations to the embodiments of the present disclosure without departing from the spirit and scope of the present disclosure. The scope of protection of the present disclosure is defined by the appended claims.

Claims

1. A barrel life prediction method based on gradient boosting and quantile recursive network, characterized in that: include: collecting multi-source sensor data over the entire life cycle of the barrel, and preprocessing the multi-source sensor data to form a standardized data set; Constructing an XGBoost model, performing feature importance analysis on the standardized dataset using the XGBoost model, screening key features based on the analysis results to obtain a feature-selected dataset, and optimizing and training the XGBoost model using gradient boosting and regularization methods; Dividing the feature-selected dataset into a training set and a test set; Combining the quantile regression method, a life prediction model is constructed based on a quantile recursive neural network, the life prediction model is trained using the training set, and the model parameters are optimized through bifurcation sequence training to obtain an optimized barrel life prediction model; The test set is input into the optimized barrel life prediction model, and a remaining service life prediction result including a predicted value and a confidence interval is output, and the prediction performance is evaluated.

2. The method according to claim 1, characterized in that The step of preprocessing the multi-source sensor data includes: Performing time sequence alignment processing on the multi-source sensor data; Outlier detection and correction based on statistical distribution; Frequency domain filtering to eliminate high-frequency noise; Select the data normalization method according to the working condition type, including: linear normalization for a single working condition and clustering followed by group normalization for multiple working conditions.

3. The method according to claim 1, characterized in that After performing feature importance analysis on the standardized dataset using the XGBoost model and screening key features based on the analysis results, the method further includes: Eliminate redundant features and retain high-importance features; The step of optimizing the XGBoost model by gradient boosting and regularization method comprises: Residual fitting is used to improve accuracy, the second-order approximation of the loss function is used to accelerate training, the regularization term is used to control the complexity of the model, and pruning processing based on information gain is adopted, and the pruning criterion is that the information gain is non-negative.

4. The method according to claim 1, wherein Before the step of performing feature importance analysis on the standardized dataset by the XGBoost model, the method further includes: The XGBoost model is constructed based on the regression decision tree of the CART algorithm, and its node splitting criterion is minimization of the squared error.

5. The method according to claim 4, characterized in that The steps of building the XGBoost model include: Based on the objective function minimization criterion, the optimal feature split point is determined through iterative search, wherein the objective function includes: a square error loss term, a regularization penalty term, and a second-order Taylor expansion approximation term; The data set is recursively partitioned to obtain a regression tree decision tree until at least one of the following termination conditions is met: the node sample size is lower than the preset threshold, the error decrease is less than the set value, and the maximum tree depth limit is reached; The XGBoost model is obtained by combining multiple regression decision trees.

6. The method according to claim 1, characterized in that The quantile recurrent neural network includes: The encoder module uses a recurrent neural network to extract temporal features; The decoder module includes a global feature decoder and a local digit decoder, wherein: The global feature decoder captures long-term dependencies and outputs quantile predictions; The local quantile decoder responds to the periodicity and mutation of the data and outputs quantile predictions.

7. The method according to claim 1, characterized in that The steps of the bifurcation sequence training include: Convert the artillery life cycle monitoring data into time series feature representation; Dynamically determine multiple prediction starting points on time series data based on preset time interval rules; Executed synchronously at each prediction starting point, including: context feature extraction, multi-scale feature decoding, and uncertainty quantification prediction; The prediction error loss and physical constraints are integrated to construct a multi-objective optimization function.

8. A barrel life prediction device based on gradient enhancement and quantile recursive network, characterized in that: include: a preprocessing module for collecting multi-source sensor data throughout the life cycle of the barrel and preprocessing the multi-source sensor data to form a standardized data set; A feature selection module is used to build an XGBoost model, perform feature importance analysis on the standardized dataset using the XGBoost model, screen key features based on the analysis results, obtain a feature-selected dataset, and optimize and train the XGBoost model using gradient boosting and regularization methods; A partitioning module, used to divide the dataset after feature selection into a training set and a test set; a training module for constructing a life prediction model based on a quantile recursive neural network in combination with a quantile regression method, training the life prediction model using the training set, and optimizing model parameters through bifurcation sequence training to obtain an optimized barrel life prediction model; The prediction module is used to input the test set into the optimized barrel life prediction model, output a remaining service life prediction result including a predicted value and a confidence interval, and evaluate the prediction performance.

9. A computer device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images