Firearm reliability prediction and evaluation method

Abstract

The invention discloses a firearm reliability prediction and evaluation method, and belongs to the technical field of data evaluation. The method comprises the following steps: firstly, collecting and preprocessing multi-dimensional reliability data of a full life cycle of a firearm through a big data technology; secondly, constructing a'feature extraction-correlation analysis-prediction output 'three-layer progressive historical equipment reliability level model, and mining reliability change and distribution rules; and finally, based on the model, realizing reliability prediction and evaluation of the in-research and in-production equipment, and identifying reliability weak links through multi-dimensional correlation analysis to form a complete closed loop of'data-model-analysis-application-optimization '.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data evaluation, and particularly relates to a firearm reliability prediction and evaluation method. BACKGROUND

[0002] As a core light weapon equipment of national defense, the reliability of a firearm directly determines the combat effectiveness, use safety and life cycle cost. With the continuous improvement of the performance requirements of modern war on equipment, the firearm technology develops towards high precision, high firing rate and complex environment adaptability, and the factors affecting the reliability of the firearm present multi-source, non-linear and dynamic characteristics. The traditional firearm reliability evaluation method has been difficult to meet the development needs of modern equipment, and mainly has the following deficiencies: 1) On the data level, the reliability data of the firearm runs through the whole life cycle, covers material performance, structure parameters and simulation failure data in the design stage, machining precision, assembly process and quality inspection data in the production stage, environmental parameters (temperature, humidity, dust), operation frequency and failure records in the use stage, and repair frequency, replacement parts and failure cause analysis in the repair stage. These data have the characteristics of large scale, various types and strong correlation, but the traditional technology has two big bottlenecks: one is the lack of efficient data integration capability, which cannot associate the data scattered in the design, production and use links into an organic whole; the other is the lack of deep mining capability, which is difficult to capture the potential law of the multi-factor synergistic influence on reliability in the data, resulting in that the data value is not fully utilized.

[0003] 2) On the evaluation demand level, there are three core contradictions in the current firearm equipment system: one is the contradiction between the idle historical data and the prediction demand of new equipment, the traditional method cannot utilize the massive historical data to deduce the reliability change law, and it is difficult to predict the reliability of the design scheme of the in-process equipment and the production quality of the in-process equipment; the second is the contradiction between local evaluation and overall reliability demand, the traditional method mainly focuses on a single link or local system, and cannot reflect the overall reliability level of the firearm; the third is the contradiction between single factor analysis and multi-factor weak link positioning demand, the traditional method can only identify the direct correlation between a single structure parameter and failure, and cannot locate the system-level weak link of the multi-factor synergistic action of “material + process + environment”, resulting in that the firearm design optimization, production quality control and use maintenance strategy lack scientific basis, and the equipment reliability is restricted.

[0004] Patent CN202310832994.2 discloses a firearm firing ignition reliability control method. The method realizes firearm firing ignition reliability control through the process of "obtaining key parameters → constructing sample database → fitting relationship model → determining target structure parameters". The specific steps include: determining the key influence parameters of firearm firing ignition reliability through failure mode and effects analysis (FMEA), based on typical fault tree and fault tree qualitative analysis, including lock gap, primer shell thickness, firing pin protrusion, firing platform head diameter, firing pin head diameter, primer loading depth, powder height, and transfer hole diameter; then, based on the key influence parameters, the control variable analysis of firearm firing ignition reliability is carried out, under the premise of unchanged other parameters and same firing environment, the normal distribution mean and standard deviation of single key parameter or at least two key parameters are adjusted respectively, the parameter tolerance changes according to the preset step, the firing ignition reliability corresponding to each tolerance range is recorded, and single influence parameter sample database and multiple influence parameter sample database are formed; using polynomial distribution fitting model, support vector machine model or artificial neural network proxy model, data fitting is carried out on the above two sample databases respectively, the mathematical relationship model of "single parameter-reliability" and "multi-parameter coupling-reliability" is established, and the firearm firing ignition reliability influence relationship model is obtained; the expected value of firearm firing ignition reliability is obtained, and the target firearm firing ignition structure parameters are determined based on the reliability influence relationship model and the expected value of firearm firing ignition reliability, so as to control the reliability of firearm firing ignition through the target firearm firing ignition structure parameters.

[0005] Based on the core logic and implementation details of the above existing technical solutions, combined with the actual engineering requirements of firearm reliability, patent CN202310832994.2 has the following limitations: 1) Data dimension is single, which cannot reflect the overall reliability of the firearm. The patent only takes the local structure parameters of the firing ignition system as input variables, without considering the material data in the design stage, the processing technology data in the production stage, the maintenance history in the use stage, and the failure data of other systems of the firearm. Because the coverage of the input variables is limited, the reasoning conclusion is only applicable to the firing ignition link and cannot be extended to the overall reliability evaluation of the firearm, resulting in one-sided evaluation results. 2) The evaluation target is lagging behind and lacks prediction ability. The patent takes the firing ignition failure control of active equipment as the target, and can only make post-correction to the local failure of the served equipment through parameter optimization. Since the historical equipment reliability model is not constructed, the reliability change rule cannot be derived using a large amount of historical data, and the rule cannot be migrated to the reliability prediction of the in-research and in-production equipment, which cannot meet the forward-looking needs of equipment research and production. 3) The identification of weak links is limited and cannot support the optimization of the whole system. Limited by the data dimension and the evaluation range, the patent can only identify the correlation between the local structure parameters of the firing ignition system and the failure, and cannot locate the weak links affecting the overall reliability of the firearm through multi-dimensional data correlation analysis, such as the synergistic effect of material fatigue performance, processing error, and high-temperature environment, which makes it difficult to guide the design optimization and production quality improvement of the whole system of the firearm. 4) The data value is not fully excavated, and the evaluation accuracy is limited. The patent does not use big data technology, and cannot integrate large-scale and heterogeneous data generated in the whole life cycle of the firearm. It only relies on a small amount of structured parameters for causal reasoning. This leads to the fact that the potential multi-factor synergistic effect in the data is not excavated, the sample size and variable richness of the causal relationship analysis are insufficient, and finally the accuracy and comprehensiveness of the evaluation results are limited. SUMMARY

[0006] To solve the problems of single data dimension, lagging evaluation, and inability to predict the reliability of in-research and in-production equipment and locate multi-factor synergistic weak links in the prior art, the present application provides a firearm reliability prediction and evaluation method. The method integrates multi-dimensional reliability data of light weapons in the whole life cycle through big data technology, constructs a historical equipment reliability level model, accurately obtains the reliability change and distribution rule, locates the reliability weak links, and realizes efficient and comprehensive analysis and evaluation of the reliability level of in-research and in-production equipment.

[0007] The present application discloses a firearm reliability prediction and evaluation method, which comprises: Step S1, around the firearm whole life cycle reliability related data, a structured multi-dimensional reliability database is constructed by building multi-source data acquisition channels and standardized preprocessing procedures; Step S2, based on the processed whole life cycle multi-dimensional reliability data, a three-layer progressive historical equipment reliability level model of feature extraction-correlation analysis-prediction output is constructed, the model of the firearm reliability is constructed by quantifying key features, strengthening the correlation between features and reliability, and optimizing the prediction algorithm. Step S3, based on the historical reliability prediction model, input the multi-dimensional running data of the gun under different working conditions and environments, and extract the reliability change law and distribution characteristics through the double-layer mining strategy of trend fitting + distribution verification; Step S4, through the Apriori association rule mining algorithm, the feature combination-failure sample data collected in the whole life cycle of the gun are analyzed, and the weak links affecting reliability are identified based on the core logic of double threshold judgment + structured list output; Step S5, for the two scenes of in-service equipment design verification and in-production equipment quality control, combined with the historical reliability prediction model and quantitative threshold judgment, the equipment reliability prediction and dynamic control are realized; Step S6, by periodically supplementing new data, repeating the preprocessing process, and retraining the model with gradient descent algorithm, the historical reliability prediction model is ensured to improve with data accumulation.

[0008] Step S1 specifically includes: S11, data collection; Through big data technology, a multi-source data collection channel is built to collect multi-dimensional reliability data in the whole life cycle of the gun, covering reliability data in different dimensions of design, production, use and failure; wherein: Design data: obtain material performance and structure parameters through CAD design software interface; obtain simulation failure data through simulation platform; Production data: connect production line sensors and MES system to collect machining precision, assembly process and quality inspection pass rate data; Usage data: collect environmental parameters and operation data through sensors; Failure data: connect equipment maintenance system to collect failure records and maintenance history; S12, data preprocessing; Use big data cleaning technology to remove outliers, integrate scattered data based on the unique code of the gun, perform standardized conversion, and form a structured multi-dimensional reliability database; Data cleaning: remove sensor failure and manual recording errors using 3σ criterion, and supplement missing data by interpolation method; Data integration: associate design, production, use and failure data based on the unique code of the gun to form a single gun life cycle data chain; Data standardization: convert unstructured data into structured features, unify physical units, and form a structured multi-dimensional reliability database.

[0009] Step S2 specifically includes: S21, feature extraction layer: quantize multi-dimensional data features; For the four dimensions of gun design, production, use and failure, the data are transformed into quantifiable and highly correlated structured features through feature engineering. The definition, calculation logic and physical meaning of key features in each dimension are clarified. Design dimension features: Material fatigue life : where k is the total number of stress cycles that the material undergoes during the use of the firearm, is the actual cycle number of the i-th stress cycle, is the limit fatigue life of the material at the stress level corresponding to the i-th stress cycle, measured through material fatigue testing; Structural strength margin : where is the maximum stress value of the firearm structure calculated by finite element simulation software such as ABAQUS, is the material allowable stress determined according to the material standard, which is the ratio of material yield strength and design safety factor; Production dimension features: Machining error standard deviation : where n is the sampling number of parts in the same batch, is the measured size of the j-th part, is the arithmetic mean of the measured size of the parts in this batch; Assembly process stability : where USL / LSL is the upper and lower limit of the assembly parameter, is the standard deviation of the assembly parameter collected by the production line sensor; Use dimension features: Environmental severity coefficient : where T, RH, are the actual service environment temperature, relative humidity and dust concentration, , , are the weights of temperature, humidity and dust concentration, calculated by AHP method, satisfying T0, RH0 are the standard environment temperature and standard relative humidity specified by the design standard, T max, RH max 、 The maximum ambient temperature, maximum relative humidity, and maximum sand concentration allowed by design, respectively.

[0010] Cumulative usage intensity : wherein, is the cumulative number of shots of the firearm, is the cumulative number of shots of the firearm, is the environmental severity coefficient calculated above; Failure dimension characteristics: Historical failure frequency : wherein, is the total number of failures of the same type of historical firearm; M is the number of historical samples; is the cumulative usage intensity of the mth sample; Effective feature set screening: significant screening + redundancy elimination two-step method is used to screen the effective feature set, and finally form the effective feature set as the input variable for subsequent association analysis and prediction modeling.

[0011] Step S2 specifically includes: S22, association analysis layer: quantifying the correlation strength between features and reliability; To identify key features that have a significant impact on firearm reliability, the effective feature set X obtained by screening is used as the input, and the mean time between failures MTBF is used as the output Y. The correlation strength between each feature and reliability is quantified by random forest and gradient boosting tree fusion algorithm; Random forest feature importance calculation: A multi-decision tree ensemble model is constructed using the random forest algorithm, and the importance of a single feature is calculated based on the reduction of node impurity Gini coefficient: Feature importance formula: wherein, K is the total number of decision trees in the random forest, T k is the kth decision tree, and node is the node in the decision tree that uses feature x i to split; is the Gini coefficient change before and after the node splits; Gini coefficient change formula: Wherein, Gini (parent) is the Gini coefficient of the parent node, child is the child node after the parent node is split, nchild is the sample quantity of the child node, nparent is the sample quantity of the parent node, and Gini (child) is the Gini coefficient of the child node. Gini coefficient basic formula: , Wherein, C is the number of reliability level categories, is the proportion of samples belonging to the cth reliability level in a certain node; GBDT gradient boosting tree feature weight iterative optimization: The gradient boosting tree GBDT algorithm is adopted to iteratively update the feature weight by minimizing the prediction error, so as to realize secondary quantification of the correlation strength between the feature and reliability; Loss function definition: The mean square error MSE is used as the loss function to measure the deviation between the model prediction value and the true MTBF value, and the formula is: Wherein, t is the iteration number; N is the total number of samples; Y i is the true MTBF value of the ith sample, F t-1 (X i ) is the model prediction value of the previous t-1 iterations, and ht(X i ) is the prediction value of the newly added base learner in the tth iteration; Feature weight update rule: Based on the negative gradient direction of the loss function, the weight coefficient of each feature is updated, and in each iteration, the weight i of the feature x t is adjusted to minimize the loss function L i ; The greater the absolute value of the finally output feature weight, the stronger the correlation between the feature and the reliability of the firearm; Key feature set screening: The feature importance output by the random forest and the absolute value of the feature weight output by the GBDT are normalized respectively and mapped to the [0, 1] interval; Calculate the comprehensive correlation strength of each feature: Score(x i )=0.5×Norm(Importance(x i ))+0.5×Norm(|ω i |), wherein Norm( ) is a normalization function; Sort the comprehensive correlation strength from large to small, and screen the top-ranked features to form a key feature set , which is used as the core input of the subsequent prediction model.

[0012] Step S2 specifically comprises: S23, a prediction model layer: constructing a reliability prediction model; Based on the key feature set , an improved gradient boosting decision tree (GBDT) prediction model is constructed; Model structure design: The improved GBDT prediction model adopts a multi-level feature mapping and nonlinear transformation architecture, introduces an activation function and weight optimization mechanism of a deep neural network, and the model expression is specifically: Wherein: is the reliability prediction output value, representing the quantitative prediction result of the reliability of the target object; W1, W2, and W3 are the feature mapping weight matrices of the first to third levels of the model, respectively, for realizing linear transformation of features; b1, b2, and b3 are the bias vectors of the first to third levels of the model, respectively, for adjusting the offset of feature transformation and improving the generalization ability of the model. / Both of them adopt ReLU activation function, whose expression is , which is used to introduce nonlinear feature mapping, enhance the fitting ability of the model to complex nonlinear relationship, and avoid the problem of gradient disappearance; Sigmoid activation function is adopted, whose expression is , which is used to constrain the prediction output value in the interval (0, 1], so as to ensure that the output result is consistent with the physical meaning of reliability index.

[0013] Model training: Design a targeted training strategy, including data set division, loss function selection, optimizer configuration and training termination condition; Data set division: the sample set containing the historical reliability data of the target object and the corresponding key feature set is randomly divided into training set, validation set and test set according to the proportion; among them, the training set is used for model parameter learning, the validation set is used for adjusting hyperparameters in the iteration process to avoid overfitting, and the test set is used for final evaluation of the generalization performance of the model; Loss function selection: mean absolute percentage error (MAPE) is used as the loss function of model training, whose expression is: Wherein, N is the number of samples, Y i is the actual reliability value of the i-th sample, is the predicted reliability value of the i-th sample by the model, and MAPE is selected as the loss function; Optimizer and training termination condition: the Adam optimizer is used to iteratively optimize the model weight matrices W1, W2, W3 and bias vectors b1, b2, b3 to minimize the MAPE loss function; the training process continues to iterate until the following two convergence conditions are met simultaneously: The MAPE value on the test set is less than a first threshold value; and the determination coefficient R2 on the test set is greater than a second threshold value, where the R2 expression is: , is the average value of the actual reliability value of the sample; R2 is used to measure the explanation ability of the model to data variation.

[0014] Step S3 specifically includes: S31, reliability change law mining, based on the attenuation modeling of the number of shots; For the same type of gun due to mechanical wear, component fatigue, and barrel ablation, a quantitative correlation model of MTBF average failure-free time and core use parameters is constructed by exponential fitting; Input the historical operation data of the same type of gun, including the number of shots , the time series t = [t1, t2,..., t n ] is used to correspond to the measured value Y = [Y1, Y2,..., Y n ] of MTBF under the use state and auxiliary operation parameters; An exponential decay function is used to fit the number of shots-MTBF data to construct a mathematical model of MTBF change with the number of shots, which can reflect the dynamic influence of use intensity on reliability, and the expression is: Where: Y0 is the initial MTBF value of the gun, which is determined by fitting the performance test data and initial operation data when the gun is manufactured; k is the attenuation coefficient; N shot is the core independent variable affecting reliability decay; Based on the above quantitative model, the key threshold law of gun reliability decay is extracted, which is directly used as a quantitative basis for preventive maintenance of the gun, that is, when the cumulative number of shots of the gun approaches the threshold value, the barrel, firing mechanism, and recoil spring components are repaired or replaced in advance.

[0015] Step S3 specifically includes: S32, reliability distribution law mining, statistical modeling based on environmental conditions; A normal distribution fitting is used to construct an MTBF distribution model under multiple environments, and statistical tests are used to verify the significance of environmental factors to realize quantitative characterization of the influence of environment on reliability; Environmental variable screening and data grouping: Screening the environmental factors that have significant impact on the reliability of firearms through correlation analysis, and selecting temperature as the key variable in priority; dividing the historical reliability data into two typical sample sets according to the environmental temperature: Normal temperature group: environmental temperature is 20℃±5℃, containing the measured value Y of MTBF under this environment norm =[Y norm1 ,Y norm2 ,..,Y normm ] and the corresponding temperature record; High temperature group: environmental temperature > 40℃, containing the measured value Y of MTBF under this environment high =[Y high1 ,Y high2 ,...,Y highp ] and the corresponding temperature record; Distribution model fitting and effectiveness verification: Distribution type test: K-S test is used to verify the normal distribution characteristics of the two groups of MTBF data. When the test statistic D < Dα, it is determined that the data obeys the normal distribution; Distribution parameter solving: the normal distribution parameters of the two groups of data are solved by maximum likelihood estimation method, and the normal distribution model of MTBF under different environments is obtained: Under normal temperature environment: MTBF obeys normal distribution N(μ1,σ12), where the mean μ1 represents the average reliability level of firearms under normal temperature, and the variance σ12 represents the fluctuation degree of reliability under normal temperature; Under high temperature environment: MTBF obeys normal distribution N(μ2,σ22), where the mean μ2 represents the average reliability level of firearms under high temperature, and the variance σ22 represents the fluctuation degree of reliability under high temperature.

[0016] Step S4 specifically includes: Using Apriori association rule mining algorithm, the feature combination-failure sample data collected in the whole life cycle of the product is analyzed; Setting double threshold as the weak link determination standard, including support threshold and confidence threshold; outputting the results to generate a structured reliability weak link list, which contains the following core information: weak factor, associated component, failure probability, improvement scheme and verification requirement.

[0017] Step S5 specifically includes: S51, reliability prediction and evaluation of in-development equipment; The design parameters and simulation data of in-development equipment are input into the historical reliability prediction model, wherein the design parameters need to include the fatigue life of new materials and the structure size; the simulation data covers the stress-strain simulation results and vibration impact simulation data under typical working conditions; The historical model outputs a theoretical reliability prediction value of the equipment under development based on input data through a multi-feature weighting algorithm , collects the actual MTBF value Y during the service period in , calculates the average value as a comparison benchmark If , it is determined that the reliability of the equipment under development does not meet the requirements of the active benchmark, and a design optimization feedback report is immediately generated. The feedback report clearly quantifies the optimization indicators and feeds them back to the design department to optimize parameters S52, reliability prediction evaluation of in-production equipment A real-time data collection system is built for the production line, with a collection frequency of ≥1 time / minute. The collected data includes the processing parameters of the current production batch, the assembly process parameters, and the performance parameters of the raw material batch. The data is uploaded to the cloud database in real time with a delay of ≤10s The real-time collected standardized data is input into the historical reliability prediction model, and the reliability prediction value of the equipment after production of this batch is output The reliability standard threshold of this type of equipment is retrieved , and a real-time comparison is made. If The system immediately sends a process adjustment instruction to the production line control system and suspends the subsequent production of this batch until the data is re-collected after process adjustment ≥ .

[0018] Step S6 specifically includes: S61, evaluation result output specification A structured and visual reliability evaluation report is generated, which includes the following core modules, and all data needs to be accompanied by calculation basis and error range Reliability score module: the calculation formula is , and the score level is marked Trend analysis module: with time as the horizontal axis and reliability prediction value as the vertical axis, a reliability change curve of the equipment under development at different design stages and the in-production equipment at different batches is generated, and the reliability improvement amplitude of key nodes is marked Weak link and improvement module: integrate the reliability prediction results of the equipment under development and in-production and the weak link identification results, list the weak link list and provide implementable improvement suggestions S62, model iteration optimization Establish a closed-loop model iteration mechanism to ensure that the prediction accuracy continuously improves with data accumulation Periodically supplement the actual operation data of new equipment to the multi-dimensional database, including fault records of in-production equipment after service, test data of equipment under development, and actual reliability verification data after production line process adjustment ​The data quality and feature integrity are ensured through the repeated processes of outlier rejection, data standardization and feature re-extraction. Based on the updated database, the historical reliability prediction model is retrained by using the gradient descent algorithm to adjust the weight coefficients of each feature in the model. After the training is completed, the validation set is selected. If the model prediction error does not exceed the error threshold, it is determined that the model optimization is qualified and is put into use. Otherwise, the feature extraction logic needs to be rechecked and retrained until the error requirement is met. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application. Those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0020] Figure 1 is a schematic diagram of a firearm reliability prediction and evaluation process provided by an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creating any creative labor are within the scope of protection of the present application.

[0022] The present application integrates big data technology and multi-dimensional modeling to realize a firearm reliability prediction and evaluation method for integrating, mining rules, positioning weak links and forward-looking evaluation of firearms in the whole life cycle. The steps include: collecting and preprocessing multi-dimensional reliability data of firearms in the whole life cycle by using big data technology; building a three-layer progressive historical equipment reliability level model of "feature extraction-correlation analysis-prediction output"; mining the reliability change and distribution rules of firearms based on the model; identifying the reliability weak links through multi-dimensional correlation analysis; using the model to predict and evaluate the reliability of in-research and in-production equipment; and iteratively optimizing the model combined with actual data of new equipment to improve the evaluation accuracy.

[0023] Related term explanation: Reliability: refers to the ability of a firearm to complete a specified function under specified conditions and within a specified time. Common indicators include mean time between failures, failure rate, reliability, etc.

[0024] Multi-dimensional reliability data: refers to the data collected from different stages of the whole life cycle of firearms, including design, production, testing, use, and maintenance, from different angles such as structural parameters, process parameters, environmental parameters, performance parameters, and failure parameters.

[0025] Big data technology: refers to a collection of technologies for processing massive, multi-source heterogeneous data, including data collection (sensors, database interface, file parsing), preprocessing (abnormal value cleaning, multi-source data integration, standardization conversion), data mining (association rules, random forest, gradient boosting tree), etc.

[0026] Historical equipment reliability level model: refers to a mathematical model constructed based on the whole life cycle reliability data of the listed or retired firearms through machine learning and statistical analysis, which can quantify the correlation between various influencing factors and the core indicators of firearm reliability, and realize reliability prediction and rule analysis.

[0027] Reliability weak link: refers to the key factors or components that have a significant impact on the overall reliability of firearms and have a high failure probability, which is obtained through multi-dimensional data correlation analysis.

[0028] The present application first collects and preprocesses the multi-dimensional reliability data of the whole life cycle of firearms through big data technology; secondly, constructs a three-layer progressive historical equipment reliability level model of "feature extraction-correlation analysis-prediction output", and mines the reliability change and distribution rule; finally, based on the model, the reliability prediction and evaluation of the in-research and in-production equipment are realized, and the reliability weak link is identified through multi-dimensional correlation analysis, forming a complete closed loop of "data-model-analysis-application-optimization".

[0029] The specific process is shown in Figure 1 .

[0030] S1, multi-dimensional reliability data collection and preprocessing.

[0031] S11, data collection; Through big data technology, a multi-source data collection channel is built to collect multi-dimensional reliability data in the whole life cycle of firearms, including reliability data of different dimensions such as design, production, use, and failure: Design data: obtain material performance and structural parameters through CAD design software interface, and obtain simulation failure data through simulation platform; Production data: connect production line sensors and MES system to collect machining precision, assembly process, and quality inspection pass rate data; Use data: collect environmental parameters and operation data (firing times, firing rate) through sensors; Failure data: interface with equipment maintenance system, collect failure records (firing times, use environment), failure records (failure location, mode, time of occurrence), maintenance history (number of repairs, replacement parts, average repair time).

[0032] S12, data preprocessing; Use big data cleaning technology to remove outliers, associate dispersed data based on "unique gun code" through data integration, and then perform standardized conversion to form a structured multi-dimensional reliability database.

[0033] Data cleaning: remove abnormal data such as sensor failure and manual recording error using 3σ criterion, and supplement missing data by interpolation method; Data integration: based on "unique gun code", associate design, production, use and failure data to form single gun life cycle data chain; Data standardization: convert unstructured data into structured features, unify physical units, and form a structured multi-dimensional reliability database.

[0034] S2, multi-dimensional historical equipment reliability level model construction Based on the pre-processed multi-dimensional reliability data of the gun throughout its life cycle, a three-layer progressive historical equipment reliability level model of "feature extraction-correlation analysis-prediction output" is constructed. First, the data features are quantified from four dimensions and the effective feature set is selected. Then, the correlation strength between the features and reliability is quantified and the key feature set is determined through fusion algorithm. Finally, based on the key feature set, an improved GBDT prediction model is constructed. Through the design of model structure and training strategy, the reliability of the gun is predicted.

[0035] S21, feature extraction layer: quantification of multi-dimensional data features For the four-dimensional data of gun design, production, use and failure, the data is transformed into quantifiable and highly correlated structured features through feature engineering. The definition, calculation logic and physical meaning of key features in each dimension are as follows: Design dimension features: Design dimension features focus on the inherent properties of the gun design stage, quantifying the basic influence of material performance and structural safety on reliability. The core features include: Material fatigue life : represents the ability of the material used in the core components of the gun to resist fatigue failure under cyclic stress, which is a key indicator of long-term reliability of the gun.

[0036] Based on the Miner linear cumulative damage theory, considering the multi-stress cycle working condition of the material in actual use, the material fatigue life , where k is the total number of stress cycles the material undergoes during the use of the gun, Actual cycle number of the ith stress cycle, Limit fatigue life of the material under the stress level corresponding to the ith stress cycle, measured by fatigue test.

[0037] Structural strength margin : Quantify the strength redundancy of the core structure of the firearm at the design stage, reflecting the safety capability of the structure against extreme working conditions.

[0038] Based on finite element simulation and material mechanics theory, combined with the safety factor requirements specified in the firearm design standard, the structural strength margin , wherein The maximum stress value of the firearm structure calculated by finite element simulation software such as ABAQUS, The material allowable stress determined according to the firearm material standard, which is the ratio of material yield strength and design safety factor.

[0039] 2) Production dimension characteristics: Production dimension characteristics focus on the process stability and machining accuracy of the firearm manufacturing process, quantifying the impact of production links on reliability, and the core characteristics include: Machining error standard deviation : Characterize the dispersion degree of the machining dimensions of the same batch of firearm parts such as the firing pin and spring, reflect the consistency of the machining accuracy of the production line, the smaller the dispersion degree, the higher the interchangeability and assembly reliability of the parts.

[0040] Based on the statistical standard deviation formula, combined with statistical analysis of the measured data of the production line, the machining error standard deviation , wherein n is the sampling number of the same batch of parts, The measured size of the jth part, The arithmetic mean of the measured size of the parts in this batch; Assembly process stability (Process capability index): Quantify the process capability of the firearm assembly process such as bolt tightening and component matching, determine whether the assembly parameters are stable within the design allowed tolerance range, and is a core indicator for evaluating production consistency and reliability.

[0041] Based on the process capability analysis theory, combined with the parameter tolerance requirements in the firearm assembly process document, , wherein USL / LSL is the upper and lower limit of the assembly parameter, The standard deviation of the assembly parameter collected by the production line sensor.

[0042] 3) Use dimension characteristics: Use dimension characteristics focus on the environmental load and use strength of the firearm service process, and the core is to quantify the dynamic impact of external working conditions on reliability, and the core characteristics include: Environmental severity coefficient (Weighted sum): Comprehensive quantification of the influence degree of environmental factors such as temperature, humidity, sand dust on the reliability of firearms, the larger the coefficient, the stronger the damage of the environment to the firearm.

[0043] Based on the analytic hierarchy process (AHP), the weights of each environmental factor are determined, and the normalized calculation is carried out combined with the difference between the actual service environment and the standard environment: Among them, T, RH, respectively, the temperature, relative humidity, sand dust concentration of the actual service environment, , , respectively, the weight of temperature, humidity, sand dust concentration, calculated by AHP method, meet , T0, RH0 respectively, the standard environment temperature, standard relative humidity specified by the design standard, T max , RH max , respectively, the maximum environmental temperature, maximum relative humidity, maximum sand dust concentration allowed by the design.

[0044] Cumulative use intensity : Comprehensive quantification of the use frequency, use intensity and environmental influence of firearms, representing the cumulative damage degree of firearms in the service process, and is the core bridge between use process and failure risk.

[0045] Based on the synergistic mechanism of "use frequency x use intensity x environmental influence", the shooting parameters and environmental coefficients are integrated: Among them is the cumulative number of firearm shots, is the cumulative number of firearm shots, is the environmental severity coefficient calculated above.

[0046] Failure dimension characteristics: Failure dimension characteristics focus on historical failure data of firearms, the core is to quantify the frequency and intensity of failure, and provide direct failure basis for reliability modeling, and its core characteristics are: Historical failure frequency : Represents the failure probability of similar firearms under the same use intensity, reflecting the reliability level of historical equipment, and is a key indicator for correlating historical data and new equipment prediction.

[0047] Based on the failure statistics theory, the historical use and failure records of similar firearms are combined: Among them Total failure number of the same type of historical gun; M is the number of historical samples; Cumulative use intensity of the mth sample.

[0048] 5) Effective feature set screening To avoid feature redundancy and dimension disaster, ensure the simplicity and prediction accuracy of the model, the "significance screening + redundancy elimination" two-step method is used to screen the effective feature set: Significance screening (analysis of variance ANOVA): taking the core indicator of gun reliability (mean time between failures MTBF, denoted as Y) as the dependent variable, and the above extracted dimensional features as the independent variable, single factor analysis of variance is carried out. Calculate the F statistic of each feature, retain features (where is the F distribution critical value at a significance level of 0.05 and degrees of freedom (1, n-2)), and eliminate features that have no significant effect on reliability.

[0049] Redundancy elimination (mutual information entropy): calculate the mutual information entropy I(X, Y) between the remaining features after significance screening (representing the degree of linear or nonlinear correlation between two features), if , it is determined that there is strong redundancy between the two features, and the feature that has less effect on reliability (determined by the feature importance ranking in the correlation analysis layer) is eliminated.

[0050] Finally form an effective feature set , as the input variable of subsequent correlation analysis and prediction modeling.

[0051] S22, correlation analysis layer: quantifying the correlation strength between features and reliability To identify key features that have a significant impact on gun reliability, take the effective feature set X screened as input and the mean time between failures MTBF (denoted as Y) as output, and quantify the correlation strength between each feature and reliability through random forest and gradient boosting tree (GBDT) fusion algorithm: Random forest feature importance calculation: A random forest algorithm is used to build a multi-decision tree ensemble model, and the importance of a single feature is calculated based on the reduction of node impurity (Gini coefficient). The formula is defined as follows: Feature importance formula: Where K is the total number of decision trees in the random forest, T k is the kth decision tree, node is the node in the decision tree that uses feature x i to split; is the change in Gini coefficient before and after the node splits.

[0052] Gini coefficient change formula: Where Gini (parent) is the Gini coefficient of the parent node, child is the child node after the parent node is split, nchild is the sample size of the child node, nparent is the sample size of the parent node, and Gini (child) is the Gini coefficient of the child node.

[0053] Gini coefficient basic formula: , Where C is the number of reliability level categories, is the proportion of samples in a certain node that belong to the cth reliability level.

[0054] GBDT (Gradient Boosting Tree) feature weight iterative optimization: The Gradient Boosting Tree (GBDT) algorithm is used to iteratively update the feature weights by minimizing the prediction error, achieving secondary quantification of the strength of the association between features and reliability.

[0055] Loss function definition: The mean square error (MSE) is used as the loss function to measure the deviation between the model prediction value and the true MTBF value, and the formula is: Where t is the number of iterations; N is the total number of samples; Y i is the true MTBF value of the ith sample, F t-1 (X i ) is the model prediction value of the previous t-1 iterations, h t (X i ) is the prediction value of the newly added base learner in the tth iteration.

[0056] Feature weight update rule: Based on the negative gradient direction (i.e., the residual direction) of the loss function, the weight coefficients of each feature are updated. In each iteration, the weight i of feature x is adjusted to minimize the loss function L t . The larger the absolute value of the final output feature weight, the stronger the association between the feature and the reliability of the firearm.

[0057] 3) Key feature set screening To ensure the accuracy and simplicity of the subsequent prediction model, the quantification results of the above two algorithms are combined.

[0058] The "feature importance" output by the random forest and the "feature weight absolute value" output by the GBDT are normalized (mapped to the [0, 1] interval) respectively; Calculate the comprehensive correlation strength of each feature: Score(x i )=0.5×Norm(Importance(x i ))+0.5×Norm(|ω i |), where Norm() is a normalization function; Sort the comprehensive correlation strength from large to small, and select the top 30% of features to form a key feature set , as the core input of the subsequent prediction model.

[0059] S23, prediction model layer: build a reliability prediction model To achieve high-precision prediction of the reliability of the target object, based on the key feature set , an improved gradient boosting decision tree (GBDT) prediction model is constructed, and its core technical scheme is as follows: 1) Model structure design The improved GBDT prediction model adopts a multi-level feature mapping and nonlinear transformation architecture, which introduces the activation function of a deep neural network and a weight optimization mechanism to solve the technical problem of insufficient prediction accuracy of traditional GBDT models under complex feature mapping. The model expression is as follows: Where: is the reliability prediction output value, representing the quantitative prediction result of the reliability of the target object; W1, W2, and W3 are the feature mapping weight matrices of the first to third levels of the model, respectively, used to realize linear transformation of features; b1, b2, and b3 are the bias vectors of the first to third levels of the model, respectively, used to adjust the offset of feature transformation and improve the generalization ability of the model; / Both use the ReLU activation function, whose expression is ( ), used to introduce nonlinear feature mapping to enhance the model's fitting ability for complex nonlinear relationships while avoiding the problem of gradient vanishing; The Sigmoid activation function is used, whose expression is , used to constrain the prediction output value in the interval (0, 1], ensuring that the output result is consistent with the physical meaning of the reliability index.

[0060] 2) Model training method To ensure that the model reaches the required prediction accuracy and stability for engineering applications, a targeted training strategy is designed, including data set division, loss function selection, optimizer configuration, and training termination conditions.

[0061] a. Data set division: The sample set containing the target object's historical reliability data and corresponding key feature set is randomly divided into training set, validation set, and test set in the ratio of 7:1.5:1.5; among them, the training set is used for model parameter learning, the validation set is used to adjust hyperparameters in the iteration process to avoid overfitting, and the test set is used for final evaluation of the model's generalization performance.

[0062] b. Loss function selection: The mean absolute percentage error (MAPE) is used as the loss function for model training, and its expression is: where N is the number of samples, Y i is the actual reliability value of the i-th sample, is the predicted reliability value of the i-th sample by the model. By choosing MAPE as the loss function, the relative error between the predicted value and the actual value can be effectively quantified, which is more in line with the evaluation requirements of reliability prediction accuracy in engineering scenarios.

[0063] c. Optimizer and training termination conditions: The Adam optimizer is used to iteratively optimize the model weight matrix W1, W2, W3 and bias vector b1, b2, b3 to minimize the MAPE loss function; the training process continues to iterate until the following two convergence conditions are met simultaneously: - The MAPE value on the test set is less than 5%; - The determination coefficient R2 on the test set is greater than 0.9, where R2 is expressed as: where is the average value of the actual reliability value of the sample; R2 is used to measure the model's ability to explain data variation, and R2>0.9 indicates that the model can explain more than 90% of the actual data variation, ensuring the credibility of the prediction results.

[0064] S3, Reliability change and distribution law mining Based on the historical reliability prediction model, a "trend fitting-distribution verification" double-layer mining strategy is adopted. First, a quantitative correlation model between MTBF and firing times is constructed through exponential fitting to mine the decay law of gun reliability with use and determine the maintenance threshold. Then, the key environmental factor grouping data is screened, and the K-S test and parameter solving are used to construct the MTBF normal distribution model under multiple environments to mine the influence law of environment on reliability, providing a basis for evaluation and design.

[0065] S31, Reliability change law mining (based on the number of shots to model the decay) For the same type of gun due to mechanical wear, component fatigue, bore ablation and other use loss, the quantitative correlation model of MTBF (mean time between failures) and core use parameters is constructed by exponential fitting, which breaks through the limitation of traditional reliance on experience to judge the decay trend and realizes the quantitative expression of the decay law.

[0066] 1) Input the historical operation data of the same type of gun, including the number of shots , the time series t=[t1,t2,...,t n ] corresponding to the measured value Y=[Y1,Y2,...,Y n ] of MTBF under the use state, and auxiliary operation parameters (such as barrel temperature, firing force, bullet type, etc.).

[0067] 2) Exponential decay function is used to fit the "number of shots-MTBF" data, and the mathematical model of MTBF changing with the number of shots is constructed, which can reflect the dynamic influence of use intensity on reliability, and the expression is: Among them: Y0 is the initial MTBF value of the gun (the reliability index when the number of shots N shot =0), which is determined by the performance test data and initial operation data when the gun is out of the factory; k is the decay coefficient (positive real number), which reflects the reliability decay rate, the larger the k value, the more significant the influence of the number of shots on MTBF, and the parameter is obtained by least squares fitting of the historical "number of shots-MTBF" data, which ensures that the fitting degree R2 of the model and the actual decay trend is greater than 0.95; N shot is the core independent variable that affects the reliability decay.

[0068] Based on the above quantitative model, the key threshold law of gun reliability decay is extracted, which can be directly used as the quantitative basis for preventive maintenance of the gun, that is, when the cumulative number of shots of the gun approaches the threshold value, the barrel, firing mechanism, recoil spring and other core components prone to wear should be repaired or replaced in advance to avoid the risk of failure in combat or training due to sudden reliability drop, while avoiding the waste of resources caused by premature maintenance.

[0069] S32, Reliability distribution law mining (based on statistical modeling of environmental conditions) In view of the problem that different environmental stresses (temperature, humidity, sand dust, etc.) have no quantitative support on the influence on the reliability of the gun, the MTBF distribution model under multiple environments is constructed by normal distribution fitting, the influence significance of environmental factors is verified by combining statistical test, and the quantitative characterization of the influence of environment on reliability is realized.

[0070] 1) Environment variable screening and data grouping: The environmental factors that have significant influence on the reliability of the gun are screened through correlation analysis, and temperature is preferentially selected as the key variable; the historical reliability data is divided into two groups of typical sample sets according to the environmental temperature: Normal temperature group: the environmental temperature is 20℃±5℃, and the MTBF measured values Y norm =[Y norm1 ,Y norm2 ,..,Y normm ] and the corresponding temperature records under this environment are contained; High temperature group: the environmental temperature is greater than 40℃, and the MTBF measured values Y high =[Y high1 ,Y high2 ,...,Y highp ] and the corresponding temperature records under this environment are contained.

[0071] 2) Distribution model fitting and effectiveness verification Distribution type test: K-S test (Kolmogorov-Smirnov test) is used to verify the normal distribution characteristics of the two groups of MTBF data, when the test statistic D is less than Dα (α=0.05 is the significance level), it is determined that the data obeys the normal distribution; Distribution parameter solving: the normal distribution parameters (mean μ, variance σ2) of the two groups of data are solved by maximum likelihood estimation method, and the normal distribution model of MTBF under different environments is obtained: Under normal temperature environment: MTBF obeys the normal distribution N(μ1,σ12), wherein the mean μ1 represents the average reliability level of the gun under normal temperature, and the variance σ12 represents the fluctuation degree of reliability under normal temperature; Under high temperature environment: MTBF obeys the normal distribution N(μ2,σ22), wherein the mean μ2 represents the average reliability level of the gun under high temperature, and the variance σ22 represents the fluctuation degree of reliability under high temperature.

[0072] The mean values of MTBF under high temperature environment and normal temperature condition can be derived from the two groups of distribution models. Since the high temperature environment has a negative influence on the reliability of the gun, therefore, the reliability design target under high temperature environment can be set based on the law during the gun research and development stage, and the whole environmental application ability of the product is improved.

[0073] S4, identification of weak links of reliability Adopt Apriori association rule mining algorithm to conduct association analysis on the "feature combination-failure" sample data collected in the whole product life cycle: Set double thresholds as the weak link determination standard, including the support threshold and the confidence threshold. If the support of the feature combination (xa, xb,..., xc) and the failure mode F is ≥80% and the confidence is ≥75%, it is determined that the factor corresponding to the combination is a weak link, wherein the support is defined as the number of samples in the sample set in which the "feature combination and failure mode F appear simultaneously", and the proportion of the total number of samples; the confidence is defined as the number of samples in the sample set in which "the feature combination appears, and the failure mode F occurs subsequently", and the proportion of the number of samples in which only the feature combination appears.

[0074] Output the results to generate a structured list of reliability weak links, which includes the following core information: Weak factor: clearly identify the specific parameters corresponding to the weak link; Associated components: label the specific product components to which the weak factor belongs; Failure probability: failure probability based on confidence conversion; Improvement scheme: propose practical technical improvement direction and clearly quantify the indicators; Verification requirement: supplement the verification method after improvement.

[0075] S5, Reliability prediction and evaluation of in-research and in-production equipment By inputting the core data of the in-research or in-production stage into the historical reliability prediction model, and comparing with the benchmark threshold or standard threshold, the forward-looking judgment of equipment reliability is realized.

[0076] S51, Reliability prediction and evaluation of in-research equipment Input the design parameters and simulation data of the in-research equipment into the historical reliability prediction model, wherein the design parameters need to include key performance parameters such as new material fatigue life and structure size parameters; simulation data covers stress-strain simulation results, vibration impact simulation data, etc. under typical working conditions.

[0077] Based on the input data, the historical model outputs the theoretical reliability prediction value of the in-research equipment through a multi-feature weighting algorithm Select 3-5 types of the same type of active equipment, collect the actual value Yin of MTBF in the service period of nearly 12 months, and calculate the average value as the comparison benchmark; If , it is determined that the reliability of the in-research equipment does not meet the requirements of the active benchmark, and a design optimization feedback report is immediately generated, which clearly quantifies the optimization indicators and feeds back the optimization parameters to the design department.

[0078] S52, Reliability prediction and evaluation of in-production equipment Build a real-time data acquisition system for the production line, with a collection frequency of ≥1 time / minute. The collected data should include the current production batch's processing parameters, assembly process parameters, and raw material batch performance parameters. The data should be uploaded to the cloud database in real time, with a delay of ≤10 seconds.

[0079] Input the real-time collected standardized data into the historical reliability prediction model, and output the reliability prediction value of the batch of equipment after production ; retrieve the reliability standard threshold of the type of equipment (determined by the equipment technical specification), and perform real-time comparison; if , the system immediately sends a process adjustment instruction to the production line control system and suspends the subsequent production of the batch until the data is re-collected and verified after process adjustment .

[0080] S6, evaluation result output and model iteration optimization S61, evaluation result output specification Generate a structured and visual reliability evaluation report. The report should include the following core modules, and all data should be accompanied by calculation basis and error range (≤3%): Reliability score module: the scoring calculation formula is , and the score level is marked (e.g., ≥90 points for excellent, 70-89 points for qualified, and <70 points for unqualified); Trend analysis module: take time as the horizontal axis and reliability prediction value as the vertical axis to generate the reliability change curve of the equipment under research at different design stages and the equipment in production at different batches, and mark the reliability improvement amplitude at key nodes (such as after design parameter optimization and process adjustment); Weak link and improvement module: integrate the reliability prediction results and weak link identification results of the equipment under research and in production, list the weak link list (including weak factors, related components, and impact degree), and provide practical improvement suggestions.

[0081] S62, model iteration optimization process Establish a closed-loop model iteration mechanism to ensure that the prediction accuracy continuously improves with data accumulation. The specific process is as follows: 1) Supplement the actual operation data of new equipment to the multi-dimensional database regularly. The data types include fault records of equipment in production, test data of equipment under research, and actual reliability verification data after production line process adjustment; 2) Repeat the "outlier removal → data standardization → feature re-extraction" process for the supplemented data to ensure data quality and feature integrity; ​​3) Based on the updated database, the historical reliability prediction model is retrained using the gradient descent algorithm to adjust the weight coefficients of each feature in the model. After training, 20% of the new supplementary data is selected as the validation set. If the model prediction error is ≤5%, the model optimization is qualified and put into use. If the error is >5%, the feature extraction logic needs to be rechecked and trained again until the error requirement is met.

[0082] The technical effects brought by the technical scheme of the present application include: 1) The present application integrates design, production, use and failure data, and the input variables cover the whole life cycle of the firearm. The overall reliability of the firearm can be reflected from the system level, avoiding the misjudgment of local qualification but overall failure.

[0083] 2) The present application can directly output the reliability prediction value by inputting the design data of the research equipment and the production data of the in-production equipment into the model, which can predict the reliability level of the research and in-production equipment in advance and provide the basis for early optimization of research and production.

[0084] 3) The present application can locate the weak link of the system level caused by the superposition of multiple factors such as materials, processes and environment by analyzing the synergistic correlation between feature combinations and failures through the Apriori algorithm, which can guide the design optimization and production process improvement of the whole system.

[0085] 4) The present application integrates large-scale heterogeneous data by using big data technology, and mines the potential correlation of data through algorithms such as random forest and GBDT. The sample size of model training is more sufficient, and the variables are more abundant, which solves the problem of low data value utilization rate and improves the evaluation accuracy. The technical improvements brought by the technical scheme of the present application include: 1) Breakthrough the local data limitation of existing technology, take the whole life cycle data of the firearm as the whole analysis object, realize the correlation and value mining of multi-source data through big data technology, and lay the data foundation for comprehensive evaluation.

[0086] 2) Build a historical equipment reliability model, transfer the law of historical data to the reliability prediction of research and in-production equipment, realize the evaluation closed loop from history to future, and solve the core pain point that traditional technology cannot predict the reliability of new equipment.

[0087] 3) Not limited to the causal relationship between single-dimensional parameters and failures, through the correlation analysis of multi-dimensional feature combinations, the weak link of multi-factor superposition is located, which makes the identification result more practical and meaningful.

[0088] 4) The actual data of new equipment is continuously fed back to the database to realize dynamic optimization of the model, ensuring that the prediction accuracy is continuously improved with data accumulation, meeting the long-term evaluation demand.

[0089] The various embodiments described in the specification are intended to be exemplary only and the same are not to be taken in a limiting sense. Unless otherwise noted, documents and published material described in this specification are hereby incorporated by reference. Various modifications and changes can be made thereto by those skilled in the art which freely pro- vided that such modifications and changes are within the scope of the application as claimed. It is therefore desired that the present application be considered only as exemplary and not limiting.

[0090] The above detailed description merely illustrates exemplary embodiments of the application, and is not intended to limit the scope of the application. The embodiments described above are merely given as examples and are not intended to limit the scope of the application. The scope of the application is limited only by the claims.

[0091] All the optional technical solutions described above can be combined to form optional embodiments of the application, which will not be described one by one here.

[0092] The above is only a preferred embodiment of the application, and is not intended to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A firearm reliability prediction assessment method, characterized by, The method comprises: Step S1, around the gun full life cycle reliability related data, through the construction of multi-source data acquisition channel and standardized pretreatment process, structured multi-dimensional reliability database is constructed; Step S2, on the basis of the processed full life cycle multi-dimensional reliability data, a three-layer progressive history equipment reliability level model of feature extraction-correlation analysis-prediction output is constructed, the model of gun reliability is realized by quantifying key features, strengthening the correlation between features and reliability, and optimizing the prediction algorithm; Step S3, based on the historical reliability prediction model, input the multi-dimensional running data of the gun under different working conditions and environment, through the double-layer mining strategy of trend fitting+distribution verification, the reliability change rule and distribution characteristics are extracted; Step S4, through the Apriori association rule mining algorithm, the feature combination-failure sample data collected in the life cycle of the gun is analyzed, and the weak links affecting reliability are identified by taking double threshold judgment+structured list output as the core logic; Step S5, for the two scenes of in-service equipment design verification and in-production equipment quality control, combined with the historical reliability prediction model and quantitative threshold judgment, the equipment reliability prediction and dynamic control are realized; Step S6, by periodically supplementing new data, repeating the pretreatment process, and retraining the model by gradient descent algorithm, the accuracy of the historical reliability prediction model is improved with the accumulation of data.

2. The method of claim 1, wherein, Step S1 specifically comprises: S11, data acquisition; Multi-source data acquisition channel is built through big data technology, and multi-dimensional reliability data in the full life cycle of the gun is collected, including design, production, use and failure reliability data in different dimensions; wherein: Design data: material performance and structure parameters are obtained through CAD design software interface; simulation failure data are obtained through simulation platform; Production data: connect production line sensors and MES system to collect machining precision, assembly process and quality inspection pass rate data; Use data: environmental parameters and operation data are collected through sensors; Failure data: connect equipment maintenance system to collect failure records and maintenance history; S12, data preprocessing; Abnormal values are removed by big data cleaning technology, scattered data are associated based on the unique code of the gun through data integration, standardized conversion is carried out, and structured multi-dimensional reliability database is formed; Data cleaning: 3σ criterion is used to remove sensor failure and manual recording error, and interpolation method is used to supplement missing data; Data integration: design, production, use and failure data are associated based on the unique code of the gun to form single gun full life cycle data chain; Data standardization: unstructured data is converted into structured features, physical units are unified, and structured multi-dimensional reliability database is formed.

3. The firearm reliability prediction and evaluation method of claim 2, wherein, Step S2 specifically comprises: S21, feature extraction layer: quantifying multi-dimensional data features; For the four-dimensional data of gun design, production, use and failure, the data is converted into quantifiable and highly correlated structured features through feature engineering, and the definition, calculation logic and physical meaning of key features in each dimension are defined; Design dimension features: Material fatigue life : wherein k is the total number of stress cycles that the material is subjected to during use of the firearm, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the ith stress cycle, N is the actual number of cycles for the Structural strength margin : wherein, is the maximum stress value of the firearm structure calculated by finite element simulation software such as ABAQUS, is the material allowable stress determined according to the firearm material standard, that is, the ratio of the material yield strength and the design safety factor; Production dimension features: Process error standard deviation : Wherein, n is the sampling number of the same batch of parts, is the measured size of the jth part, is the arithmetic mean of the measured size of the batch of parts; Assembly process stability : wherein USL / LSL are upper / lower limits of the assembly parameter, is the standard deviation of the assembly parameter collected by the line sensors; Use dimension features: environmental severity coefficient : Wherein, T, RH, respectively are the temperature, relative humidity, sand concentration of the actual service environment, , , respectively are the weight of temperature, humidity, sand concentration, which are calculated by AHP method, and satisfy , T0, RH0 respectively are the standard environmental temperature, standard relative humidity specified by the design standard, max , RH max , respectively are the maximum environmental temperature, maximum relative humidity, maximum sand concentration allowed by the design. cumulative use intensity : wherein, is the cumulative number of shots fired by the firearm, is the cumulative number of shots fired by the firearm, is the environmental severity factor calculated above. Failure dimension features: Historical failure frequency : wherein, M is the total number of failures of the same type of historical firearm; and is the cumulative usage intensity of the mth sample. Effective feature set screening: the effective feature set is screened by the two-step method of significance screening + redundancy elimination, and finally forms the effective feature set as the input variable of subsequent correlation analysis and prediction modeling. , as the input variable of subsequent correlation analysis and prediction modeling.

4. The firearm reliability prediction and evaluation method of claim 3, wherein, Step S2 specifically comprises: S22, correlation analysis layer: quantify the correlation strength between features and reliability; To identify the key features that have a significant impact on the reliability of firearms, the effective feature set X obtained by screening is taken as the input, and the mean time between failures MTBF is taken as the output Y. The correlation strength between each feature and reliability is quantified by random forest and gradient boosting tree fusion algorithm; Random forest feature importance calculation: A random forest algorithm is used to construct a multi-decision tree ensemble model. The importance of a single feature is calculated based on the reduction of node impurity Gini coefficient: Feature importance formula: where K is the total number of decision trees in the random forest, T k is the kth decision tree, node is the node in the decision tree that uses feature x i to split; is the change in Gini coefficient before and after the node splits; Gini coefficient change formula: Where Gini(parent) is the Gini coefficient of the parent node, child is the child node after the parent node is split, nchild is the number of samples of the child node, nparent is the number of samples of the parent node, and Gini(child) is the Gini coefficient of the child node. Gini coefficient basic formula: ， wherein C is the number of categories of reliability levels, is the proportion of samples in a certain node that belong to the cth category of reliability levels; GBDT gradient boosting tree feature weight iterative optimization: The gradient boosting tree GBDT algorithm is used to iteratively update the feature weight by minimizing the prediction error, realizing the secondary quantification of the correlation strength between the feature and the reliability. Loss function definition: The mean square error MSE is taken as the loss function to measure the deviation between the model prediction value and the true MTBF value, and the formula is: where t is the iteration number; N is the total number of samples; Y i is the true MTBF value of the i-th sample, F t-1 (X i ) is the model prediction value of the previous t-1 rounds of iterations, ht(X i ) is the prediction value of the newly added base learner in the t-th round of iterations; Feature weight update rule: The weight coefficients of each feature are updated based on the negative gradient direction of the loss function. In each iteration, the weight of the feature x i is adjusted to minimize the loss function L t ; the greater the absolute value of the final output feature weight, the stronger the correlation between the feature and the reliability of the firearm. ​ Key feature set screening: The absolute values of the feature importance output by the random forest and the feature weight output by the GBDT are normalized and mapped to the [0, 1] interval. Calculate the comprehensive correlation strength of each feature: Score(x i )=0.5 x Norm(Importance(x i ))+0.5 x Norm(|ω i |), where Norm() is a normalization function; Sort the integrated correlation strength from large to small, filter the top-ranked features to form a key feature set as the core input of the subsequent prediction model.

5. The firearm reliability prediction and evaluation method of claim 4, wherein, Step S2 specifically includes: S23, prediction model layer: build a reliability prediction model; Based on a set of key features An improved gradient boosting decision tree (GBDT) prediction model is constructed. Model structure design: The improved GBDT prediction model adopts a multi-level feature mapping and nonlinear transformation architecture, introduces the activation function and weight optimization mechanism of deep neural networks, and the model expression is as follows: wherein: is the reliability prediction output value, representing the quantitative prediction result of the reliability of the target object; W1, W2, and W3 are respectively the feature mapping weight matrices of the first to third levels of the model, used to realize linear transformation of the features; b1, b2, and b3 are respectively the bias vectors of the first to third levels of the model, used to adjust the offset of the feature transformation and improve the generalization ability of the model; / Both adopt the ReLU activation function, and its expression is , used to introduce nonlinear feature mapping, enhance the fitting ability of the model to complex nonlinear relationships, and avoid the problem of gradient disappearance; adopt the Sigmoid activation function, and its expression is , used to constrain the prediction output value in the interval (0, 1], ensuring that the output result is consistent with the physical meaning of the reliability index. Model training: Design targeted training strategies, including data set division, loss function selection, optimizer configuration, and training termination conditions. Dataset partitioning: This involves dividing the dataset into historical reliability data of the target object and corresponding key feature sets. The sample set is randomly divided into training set, validation set and test set according to the proportion; the training set is used for model parameter learning, the validation set is used to adjust hyperparameters during the iteration process to avoid overfitting, and the test set is used to finally evaluate the generalization performance of the model. Loss function selection: The mean absolute percentage error MAPE is taken as the loss function for model training, and its expression is: wherein N is the number of samples, Y i is the actual reliability value of the i-th sample, is the predicted reliability value of the i-th sample by the model, and MAPE is selected as the loss function. Optimizer and training termination condition: The Adam optimizer is used to iteratively optimize the model weight matrix W1, W2, W3 and bias vector b1, b2, b3 to minimize the MAPE loss function. The training process continues to iterate until the following two convergence conditions are met simultaneously: The MAPE value on the test set is less than the first threshold value; and the determination coefficient R2 on the test set is greater than the second threshold value, where R2 expression is: ， R2is the average of the actual reliability values for the samples; R2is used to measure the model's ability to explain the variation in the data.

6. The firearm reliability prediction and evaluation method of claim 5, wherein, Step S3 specifically includes: S31, reliability change law mining, based on the decay modeling of the number of shots; For the dynamic decay of the reliability of the same type of firearm due to mechanical wear, component fatigue, and barrel ablation, an exponential fitting is used to construct a quantitative correlation model between MTBF (mean time between failures) and core usage parameters. Input historical running data of the same model of firearm, including a sequence of shooting times , using a time sequence t = [t1, t2,..., t n ] corresponding to the measured values of MTBF Y = [Y1, Y2,..., Y n ] in the state of use and auxiliary running parameters; An exponential decay function is used to fit the number of shots-MTBF data to construct a mathematical model of the change of MTBF with the number of shots. This model can reflect the dynamic influence of usage intensity on reliability, and the expression is: Wherein: Y0 is the initial MTBF value of the firearm, determined by fitting the performance test data and initial operation data when the firearm is shipped; k is the attenuation coefficient; N shot is the core independent variable affecting reliability attenuation; Based on the above quantitative model, the key threshold law of gun reliability degradation is extracted, which is directly used as the quantitative basis for preventive maintenance of guns. When the cumulative number of gun shots approaches the threshold, the gun barrel, firing mechanism and recoil spring components are repaired or replaced in advance.

7. The firearm reliability prediction evaluation method of claim 6, wherein, Step S3 specifically includes: S32, reliability distribution law mining, statistical modeling based on environmental conditions; Construct the MTBF distribution model under multiple environments through normal distribution fitting, and verify the significance of the influence of environmental factors through statistical test to realize the quantitative characterization of the influence of environment on reliability; Environment variable screening and data grouping: Through correlation analysis, select the environmental factors that significantly affect the reliability of the gun, and preferentially select temperature as the key variable; divide the historical reliability data into two typical sample sets according to the environmental temperature: Normal temperature group: the ambient temperature is 20℃±5℃, including the measured value Y of MTBF under the ambient temperature norm =[Y norm1 ,Y norm2 ,..,Y normm ] and the corresponding temperature record; High temperature group: ambient temperature > 40°C, comprising the measured value Y of MTBF at this ambient temperature high = [Y high1 , Y high2 ,..., Y highp ] and the corresponding temperature record; Distribution model fitting and effectiveness verification: Distribution type test: use K-S test to verify the normal distribution characteristics of the two groups of MTBF data. When the test statistic D < Dα, it is determined that the data follows the normal distribution; Distribution parameter solving: solve the normal distribution parameters of the two groups of data through the maximum likelihood estimation method to obtain the MTBF normal distribution model under different environments: Under normal temperature environment: MTBF follows the normal distribution N(μ1,σ12), where the mean μ1 represents the average reliability level of the gun under normal temperature, and the variance σ12 represents the fluctuation degree of reliability under normal temperature; Under high temperature environment: MTBF follows the normal distribution N(μ2,σ22), where the mean μ2 represents the average reliability level of the gun under high temperature, and the variance σ22 represents the fluctuation degree of reliability under high temperature.

8. The firearm reliability prediction evaluation method of claim 7, wherein, Step S4 specifically includes: Use the Apriori association rule mining algorithm to perform association analysis on the feature combination-failure sample data collected in the whole life cycle of the product; Set double thresholds as the weak link determination standard, including support threshold and confidence threshold; output the results to generate a structured reliability weak link list, which includes the following core information: weak factors, associated components, failure probability, improvement scheme, and verification requirements.

9. The firearm reliability prediction evaluation method of claim 8, wherein, Step S5 specifically includes: S51, reliability prediction and evaluation of in-development equipment; Input the design parameters and simulation data of in-development equipment into the historical reliability prediction model, where the design parameters need to include the fatigue life of new materials and structural size; the simulation data covers stress-strain simulation results and vibration impact simulation data under typical working conditions; The historical model outputs a theoretical reliability prediction value of the equipment under research based on input data through a multi-feature weighting algorithm , collect the actual value Y of MTBF during the service period in , calculate the average value as a comparison reference If , it is determined that the reliability of the development equipment does not meet the active standard requirement, a design optimization feedback report is immediately generated, the feedback report clearly quantifies the optimization indicators, and the optimization parameters are fed back to the design department; S52, reliability prediction and evaluation of in-production equipment; Build a real-time data acquisition system for the production line, with a collection frequency of ≥1 times / minute. The collected data need to include the processing process parameters, assembly process parameters, and raw material batch performance parameters of the current production batch. The data need to be uploaded to the cloud database in real time with a delay of ≤10s; The real-time collected standardized data is input into a historical reliability prediction model, and a reliability prediction value of the batch of equipment after production is output ; the reliability standard threshold of the type of equipment is called , and real-time comparison is performed If the reliability prediction value is less than the reliability standard threshold, the system immediately sends a process adjustment instruction to the production line control system and suspends the subsequent production of the batch until the data after process adjustment is collected and verified again ≥ .​ 10. The firearm reliability prediction evaluation method of claim 9, wherein, Step S6 specifically includes: S61, evaluation result output specification; Generate a structured and visual reliability evaluation report, which includes the following core modules, and all data need to be accompanied by calculation basis and error range; Reliability score module: the calculation formula is and mark the score level; Trend Analysis Module: Reliability Predictions with time as the horizontal axis Using the vertical axis, generate reliability change curves for different design stages of equipment under development and different batches of equipment in production, and mark the reliability improvement rate at key nodes; Weak link and improvement module: integrate the reliability prediction results of in-development and in-production equipment and the weak link identification results, list the weak link list and provide practical improvement suggestions; S62, model iteration optimization; Establish a closed-loop model iteration mechanism to ensure that the prediction accuracy continues to improve as data accumulates; Periodically supplement the actual operation data of new equipment to the multidimensional database, including fault records of in-service equipment, test data of in-development equipment prototypes, and actual reliability verification data after production line process adjustments; Repeat the process of outlier removal, data standardization, and feature re-extraction on the supplemented data to ensure data quality and feature integrity; Based on the updated database, retrain the historical reliability prediction model using the gradient descent algorithm to adjust the weight coefficients of each feature in the model. After training, select the validation set. If the model prediction error does not exceed the error threshold, the model optimization is considered qualified and put into use. Otherwise, the feature extraction logic needs to be rechecked and retrained until the error requirement is met.

Images