Multi-agent-based enterprise production product defect and failure analysis method and system

Abstract

The invention discloses a multi-agent-based enterprise production product defect and failure analysis method and system. The analysis method comprises the following steps: collecting product quality data from a production line in real time; preprocessing the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate preprocessed data; a machine learning model is constructed and trained, the trained machine learning model is used to carry out mode recognition on the preprocessed data, and a potential mode of poor products is recognized; based on the identification mode, a prediction model is constructed and trained to predict product failure, and the trained prediction model is used to predict potential failure time and probability; a multi-objective optimization algorithm or a decision tree algorithm is used to generate a plurality of possible production adjustment schemes, the feasibility, cost effectiveness and potential influence of each scheme are evaluated, and decision suggestions are provided for users according to evaluation results. According to the invention, the accuracy and efficiency of product defect and failure analysis can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of production quality analysis, and in particular relates to a method and system for analyzing defects and failures of enterprise production products based on multiple intelligent agents. Background Art

[0002] Product defects and failures not only affect the market competitiveness of enterprises, but may also lead to huge economic losses and damage to brand image. Therefore, how to effectively analyze and prevent product defects and failures has become an urgent problem to be solved in the manufacturing industry. Traditional data collection methods mostly rely on manual operations, and have problems such as low efficiency, incomplete data, and poor real-time performance. Current quality analysis methods are mostly based on experience judgment, lack scientificity and systematicity, often stay on the surface, and it is difficult to accurately identify the deep-seated causes of product defects. The existing quality control process relies heavily on manual operations and lacks automation and intelligent support. For the prediction of product failure, existing technologies mostly use simple statistical analysis, lack comprehensive consideration of complex factors, and have limited prediction accuracy.

[0003] In order to solve the reliance of existing quality control processes on manual labor, the Chinese invention patent with the authorization announcement number CN113516285B discloses a method for analyzing and predicting product quality in production of an automated assembly and inspection production line. The solution collects quality data from each production station (process) in the production line, pre-processes the collected data to form a quality data set, then constructs a random forest, XGBoost and LSTM algorithm sub-model and uses the formed quality data set to train the algorithm model, and then uses the Stacking algorithm fusion technology for fusion and prediction. The prediction model is applied in actual production, and when quality problems occur, warnings are sent to staff through message push.

[0004] The above solution uses data acquisition technology to collect quality data, uses data preprocessing technology to process the collected data to obtain a data set, uses machine learning and neural networks to build an algorithm model, and uses algorithm fusion technology to fuse the model to complete the quality analysis and prediction of products in the production process of the automated assembly and inspection production line. It only analyzes the product quality in the production process and issues warnings, and cannot analyze quality parameters such as product failure time. In addition, there is a lack of real-time analysis and intelligent recommendations of dynamic data. After product defects and failures occur, there is a lack of an effective decision support system, which leads to slow response and improper handling of these problems by enterprises. Summary of the invention

[0005] The present invention provides a method and system for enterprise production product defect and failure analysis based on multi-agent, aiming to solve the problem of slow response in the prior art when analyzing product defects and failures in the production process based on static data, thereby improving the accuracy and efficiency of product defect and failure analysis and providing adjustment suggestions for production plans.

[0006] In order to solve the above technical problems, the present invention proposes a method for analyzing defective and failure products produced by enterprises based on multi-agents, comprising the following steps:

[0007] Collect product quality data from the production line in real time;

[0008] Preprocessing the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate preprocessed data;

[0009] Build and train a machine learning model, and use the trained machine learning model to perform pattern recognition on preprocessed data to identify potential patterns of product defects;

[0010] Based on the identified patterns, a prediction model is constructed and trained to predict product failures, and the trained prediction model is used to predict potential failure times and probabilities;

[0011] Use multi-objective optimization algorithms or decision tree algorithms to generate a variety of possible production adjustment plans, use Monte Carlo simulation to generate a variety of scenarios, apply constrained optimization algorithms to find feasible solutions while satisfying constraints, and combine expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision-making recommendations to users based on the evaluation results.

[0012] Preferably, the machine learning model is a support vector machine.

[0013] Preferably, the machine learning model is a neural network model.

[0014] Preferably, the prediction model is a time series analysis model, including an autoregressive integrated moving average model, a seasonal autoregressive integrated moving average model, Prophet and a long short-term memory network.

[0015] Preferably, the process of prediction by the time series analysis model is as follows:

[0016] Collect time series data related to product failure, process missing values ​​and outliers, perform data smoothing and denoising, and divide the processed data into training sets and test sets according to the set ratio;

[0017] Extract the time characteristics of the data in the training set and perform periodic analysis;

[0018] Select a specific time series analysis model, use the training set for training and the test set for evaluating the model performance, and save the model whose evaluation results reach the set value;

[0019] Use the saved model to make predictions for future points in time based on real-time data on the production line.

[0020] Preferably, the prediction model is a survival analysis model, including Kaplan-Meier estimator, Cox proportional hazard model, accelerated failure time model.

[0021] Preferably, the process of prediction by the survival analysis model is:

[0022] Collect product life cycle data, including product usage time and expiration time, process missing values ​​and outliers, and divide the processed data into training set and validation set according to the set ratio;

[0023] Select variables related to product failure and encode categorical variables;

[0024] Select a specific survival analysis model, use the training set for training, and estimate the survival function or risk function;

[0025] Use the validation set to evaluate model performance, calculate survival rates and hazard ratios, and save models that meet model assumptions;

[0026] Use saved models to predict the remaining life or failure probability of your product.

[0027] Preferably, the multi-objective optimization algorithm first defines multiple objective functions, including maximizing profits, minimizing costs, and maximizing customer satisfaction, and uses Pareto optimization, genetic algorithm or particle swarm optimization method to find the balance point of multiple objectives.

[0028] Preferably, the decision tree algorithm constructs a decision tree model based on historical data, uses information gain and Gini coefficient indicators to select the best splitting attribute, and avoids overfitting through pruning.

[0029] Another aspect of the present invention is to provide a system for analyzing defective and failure products produced by enterprises based on multi-agents, which implements the above-mentioned method for analyzing defective and failure products produced by enterprises based on multi-agents, including a data collection agent, a data processing agent, a pattern recognition agent, a fault prediction agent, and a decision support agent;

[0030] The data collection agent is used to collect product quality data from the production line in real time, including raw materials, semi-finished products, product appearance dimensions, performance parameters, equipment status and production environment data;

[0031] The data processing agent pre-processes the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate pre-processed data;

[0032] The pattern recognition agent uses a machine learning algorithm to perform pattern recognition on the data to identify potential patterns of product defects, and the machine learning algorithm includes a support vector machine and a neural network model;

[0033] The fault prediction agent predicts product failure based on the identified pattern and uses time series analysis or survival analysis methods to provide risk assessment. The models of the time series analysis include autoregressive integrated moving average model, seasonal autoregressive integrated moving average model, Prophet and long short-term memory network. The models of the survival analysis include Kaplan-Meier estimator, Cox proportional hazard model and accelerated failure time model.

[0034] The decision support agent uses a multi-objective optimization algorithm or a decision tree algorithm to generate a variety of possible production adjustment plans, uses Monte Carlo simulation to generate a variety of scenarios, applies a constrained optimization algorithm to find feasible solutions while satisfying constraints, and combines expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision recommendations to users based on the evaluation results.

[0035] Compared with the prior art, the present invention has the following technical effects:

[0036] 1. The analysis method proposed in the present invention uses automated data collection and processing to build a machine learning model for pattern recognition and prediction model to predict product failures, uses a multi-objective optimization algorithm or a decision tree algorithm to generate a variety of possible production adjustment plans, and provides decision-making suggestions to users based on the results of the plan evaluation. It can effectively improve the reaction speed when analyzing defects and failures in the product production process, and improve the efficiency, accuracy and intelligence level of product quality management.

[0037] 2. The analysis method proposed in the present invention can simplify the prediction process and improve the accuracy of product failure prediction by respectively constructing machine learning models for user pattern recognition, identifying potential patterns of product defects, and then constructing a prediction model to predict products with defective patterns.

[0038] 3. The analysis system proposed in the present invention includes data collection agents, data processing agents, pattern recognition agents, fault prediction agents and decision support agents. The agents share information through message transmission and event triggering mechanisms to ensure data consistency and analysis synergy, so as to achieve efficient operation of the entire system. The system dynamically adjusts the task allocation of each agent according to real-time data and analysis requirements to adapt to changes in the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the process of the analysis method of the present invention. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in combination with specific embodiments of the present application and with reference to the accompanying drawings.

[0041] Embodiment 1

[0042] This embodiment is a method for analyzing product defects and failures in enterprise production based on multi-agents. Figure 1 As shown, the following steps are included:

[0043] Collect product quality data from the production line in real time. The collected data includes raw materials, semi-finished products, product appearance dimensions and performance parameters, equipment status, environmental factors, etc. Specifically, data collection requires the execution of data source identification, real-time monitoring and data synchronization steps.

[0044] The data source identification identifies and determines the type and source of the data that needs to be collected, such as production line sensor data (such as operating parameters and status information of CNC machine tools, injection molding machines, etc.), environmental monitoring data (such as temperature, pressure, humidity, etc.), raw material inspection report data (such as production records in the ERP system, quality inspection records, etc.), etc.

[0045] The real-time monitoring monitors the status of the production line in real time and uses a variety of data collection methods to ensure real-time data collection and updating. Data collection can be carried out in the following ways: scheduled collection: set a fixed time interval to collect data regularly; event-driven collection: trigger data collection based on specific events (such as equipment failure, parameter abnormality); real-time collection: continuously collect data to ensure the real-time nature of the data.

[0046] The data synchronization synchronizes the collected data to ensure the consistency and integrity of the data. Reliable data transmission technology is used to ensure that the data is not lost or repeated during the transmission from the collection point to the central server or database.

[0047] The product quality data is preprocessed, including data cleaning, standardization, feature extraction and data fusion, to generate preprocessed data. Among them, data cleaning includes noise removal (filtering out obvious abnormal or noise data), filling missing values ​​(inferring and completing missing data), etc.

[0048] In a preferred embodiment of the present invention, pandas, numpy and scipy libraries are introduced by python to clean the collected data. For example, for the processing of missing values, fillna() method is used to fill missing values, wherein the filled value can be the average value of the data or the value specified manually. For duplicate rows in product quality data, deletion operation is taken. For abnormal values ​​(outliers), outliers are identified by calculating z-scores, and these abnormal data are filtered out, wherein the z-score represents the deviation between the data point and the average value of the column, and the unit is standard deviation. The larger the absolute value of the z-score, the farther the data point deviates from the average value. The data points greater than the set threshold value can be filtered out by setting the threshold value of the z-score. For example, in some embodiments of the present invention, the z-score greater than 3 is generally considered to be an outlier, so the threshold value of the z-score is set to 3, and the abnormal data points that deviate from the mean value by more than 3 times the standard deviation are retained using Boolean indexing. The rows with z-scores less than 3 are filtered out, and the data points greater than or equal to 3 are filtered out to improve data quality and reduce the impact of noise on subsequent analysis and modeling.

[0049] Data cleaning also converts the data types of specific data in product quality data. For example, the to_datetime() method of the pandas library is used to unify time values ​​into datetime format, which facilitates the use of the rich date and time operation functions provided by Pandas, such as time difference calculation, time series analysis, etc.; and ensures that all time data is in a standardized date and time format to reduce problems caused by inconsistent data formats.

[0050] Data cleaning also performs consistency checks on specific data in the product quality data. For the ratio of output to input resources, usually, the physical efficiency (output / input) cannot exceed 1. Exceeding 1 is usually impossible and usually indicates that there are errors in the data (for example, the output value is recorded incorrectly or the input value is inaccurate), so these data points are filtered out.

[0051] Feature extraction is used to extract useful features from raw data, such as statistical features, time domain features, etc. Specifically, the extraction of time domain features can be done by introducing pandas, numpy and scipy libraries in Python, and then using the rolling() method to use the sliding window calculation method to extract some statistical features from the data. These features include mean, standard deviation, maximum value and minimum value. These features can be used for time series analysis or machine learning modeling, and can also be used as input features for model training or further data analysis. In addition, the sliding window calculation can smooth the data, reduce noise, and capture local trends and fluctuations in the data.

[0052] You can also use the sliding window calculation method to extract two advanced statistical features of the data: kurtosis and skewness. Kurtosis describes the thickness or sharpness of the tail of the data distribution. High kurtosis means that the distribution has a heavy tail or peak, while low kurtosis means that the tail of the distribution is lighter. Skewness describes the symmetry of the data distribution. Positive skewness means that the data is skewed to the right (longer tail), and negative skewness means that the data is skewed to the left (shorter tail). These features can help describe the distribution shape of the data and can be used as input features for time series analysis or machine learning model training to improve prediction performance.

[0053] In addition, the frequency domain characteristics of the data can be extracted, the Fast Fourier Transform (FFT) of the input signal can be calculated and the corresponding frequency and FFT values ​​can be returned.

[0054] Standardization is to standardize the sensor data, including Min-Max scaling, Z-score standardization (standardized scaling), and robust scaling (Robust scaling). Min-Max scaling scales the data to the range of [0,1], which is suitable for data with a known exact range; Z-score standardization (standardized scaling) converts the data into a normal distribution with a mean of 0 and a standard deviation of 1, which is suitable for data close to a normal distribution; robust scaling (Robust scaling) uses the median and quartiles of the data for scaling, which is suitable for data with outliers. You can choose a suitable standardization method based on the distribution characteristics of the data. Standardization and scaling processing help reduce the scale differences between different features and improve the effectiveness of data analysis and machine learning models.

[0055] Data fusion fuses data from different data sources to generate a comprehensive data set for comprehensive analysis. In one embodiment of the present invention, data from three different sensors (temperature, pressure, vibration) are merged together and a comprehensive status indicator is calculated to reflect the overall state of the system. First, the erge_asof() function is used to align the data of different sensors to the same time axis by timestamp; then a comprehensive status indicator reflecting the overall state of the sensor data is calculated, and the value range of each sensor data is limited by clipping operations to prevent the influence of extreme values ​​on the comprehensive indicator. Then, each clipped sensor value is normalized, and finally the average value of the three normalized sensor data is calculated. The average value is used as the unified indicator after fusion. The unified indicator provides a unified measure to reflect the overall health of the system. The indicator can be used as the basis for subsequent analysis, monitoring or early warning.

[0056] Build and train a machine learning model, and use the trained machine learning model to perform pattern recognition on the preprocessed data to identify potential patterns of product defects.

[0057] In one embodiment of the present invention, the machine learning model is a support vector machine (SVM). The basic principle of SVM is to find an optimal hyperplane in a high-dimensional space to separate data points of different categories. In the scenario of product defect identification, its working principle is as follows:

[0058] SVM uses kernel functions to map the original feature space to a higher-dimensional space, making nonlinearly separable problems become linearly separable in the high-dimensional space. In the mapped space, SVM looks for a hyperplane that maximizes the class interval. This hyperplane is the decision boundary, which is used to distinguish good products from defective products. Data points close to the decision boundary are called support vectors. These points are the most critical to defining the boundary and determine the classification performance of SVM. In order to deal with noisy data in reality, SVM introduces the concept of soft margins, which allows some data points to be misclassified but minimizes errors.

[0059] In one embodiment of the present invention, the machine learning model is a neural network model. Neural networks, especially deep neural networks, learn complex nonlinear patterns by simulating the neuron connections of the human brain. In product defect recognition, its working principle is as follows:

[0060] Neural networks consist of input layers, hidden layers, and output layers. Each layer contains multiple neurons, which can capture different levels of abstract features in the data; through multiple layers of nonlinear transformations, neural networks can automatically learn hierarchical features in the data without manual feature engineering; each neuron uses a nonlinear activation function (such as ReLU, Sigmoid, etc.) to introduce nonlinearity and enhance the network's expressiveness; the network learns parameters through the back-propagation algorithm to minimize prediction errors; and uses dropout, L1 / L2 regularization and other techniques to prevent overfitting and improve the generalization ability of the model.

[0061] Based on the identified patterns, a prediction model is constructed and trained to predict product failures, and the trained prediction model is used to predict potential failure times and probabilities.

[0062] In one embodiment of the present invention, the prediction model is a time series analysis model, including an autoregressive integrated moving average model (ARIMA), a seasonal autoregressive integrated moving average model (SARIMA), Prophet and a long short-term memory network (LSTM).

[0063] Among them, the structure of the ARIMA model is:

[0064] ARIMA(p,d,q)

[0065] Where p is the number of autoregressive terms, d is the number of differences, and q is the number of moving average terms.

[0066] The structure of the SARIMA model is:

[0067] SARIMA(p,d,q)(P,D,Q)m

[0068] Where the additional (P,D,Q) represents the seasonal component of (p,d,q) and m represents the seasonal period.

[0069] The structure of the Prophet model is:

[0070] y(t)=g(t)+s(t)+h(t)+ε(t)

[0071] Where g(t) represents the trend term, s(t) represents the seasonal term, h(t) represents the holiday effect, and ε(t) represents the error term.

[0072] The process of prediction by the time series analysis model is as follows:

[0073] Collect time series data related to product failures, such as sensor readings, operation logs, environmental parameters, etc., process missing values ​​and outliers, perform data smoothing and denoising, and divide the processed data into training and test sets according to the set ratio;

[0074] Extract the time features of the data in the training set and perform periodic analysis to identify the periodic and seasonal components in the data. The time features include hours, days, months, quarters, years, etc.

[0075] Select a specific time series analysis model, use the training set for training and the test set to evaluate the model performance, and save the model whose evaluation results reach the set value; take the ARIMA model as an example of the model training process, first load the data and divide the data into a training set and a test set, then define the evaluation function, use the evaluation function to evaluate different parameters, and then use the best parameters evaluated to train the final model;

[0076] Evaluation performance indicators include MSE (Mean Squared Error), RMSE (Root Mean Squared Error), MAE (Mean Absolute Error) and other indicators

[0077] Use the saved model to make predictions for future points in time based on real-time data on the production line.

[0078] In one embodiment of the present invention, the prediction model is a survival analysis model, including Kaplan-Meier estimator, Cox proportional hazard model, accelerated failure time model (Accelerated Failure Time Model, AFT model).

[0079] Among them, the Kaplan-Meier estimator is a nonparametric method used to estimate the survival function.

[0080] The structure of the Cox proportional hazards model is as follows:

[0081] h(t|X)=h0(t)·exp(β1X1+β2X2+…+βpXp)

[0082] Where h(t|X) is the hazard function for a given covariate X, h0(t) is the baseline hazard function, and βi is the regression coefficient.

[0083] The structure of the accelerated failure time model is as follows:

[0084] log(T)=β0+β1X1+β2X2+…+βpXp+σε

[0085] Where T is the survival time, βi is the regression coefficient, σ is the scale parameter, and ε is the error term.

[0086] The prediction process of the survival analysis model is as follows:

[0087] Collect product life cycle data, including product usage time and expiration time, process missing values ​​and outliers, and divide the processed data into training set and validation set according to the set ratio;

[0088] Select variables related to product failure, such as usage environment, operating conditions, etc., and encode categorical variables;

[0089] Select a specific survival analysis model, use the training set for training, and estimate the survival function or hazard function; the model training process takes the Cox proportional hazard model as an example. First, use python to introduce pandas, numpy, lifelines, and sklearn libraries, load data and divide the data into training and validation sets, use the CoxPHFItter() function to create a model, use the fit() method to train the model, the score() method to evaluate the model, and the predict_partial_hazard() method to predict the risk. Finally, use the KaplanMeierFitter() function of the lifelines library to draw the survival curve;

[0090] Use the validation set to evaluate model performance, calculate survival rates and hazard ratios, and save models that meet model assumptions;

[0091] Use saved models to predict the remaining life or failure probability of your product.

[0092] Through time series analysis and survival analysis, product failure time and probability can be effectively predicted. Time series analysis is suitable for processing continuous time data, while survival analysis is suitable for processing time-to-event data. In practical applications, appropriate methods can be selected according to specific data characteristics, and corresponding model training and evaluation can be performed to improve prediction accuracy and reliability.

[0093] Use multi-objective optimization algorithms or decision tree algorithms to generate a variety of possible production adjustment plans, use Monte Carlo simulation to generate a variety of scenarios, apply constrained optimization algorithms to find feasible solutions while satisfying constraints, and combine expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision-making recommendations to users based on the evaluation results.

[0094] The multi-objective optimization algorithm first defines multiple objective functions, including maximizing profits, minimizing costs, and maximizing customer satisfaction, and uses Pareto optimization, genetic algorithms, or particle swarm optimization methods to find the balance point of multiple objectives.

[0095] The decision tree algorithm constructs a decision tree model based on historical data, uses information gain and Gini coefficient indicators to select the best splitting attribute, and avoids overfitting through pruning.

[0096] Generate multiple possible production adjustment plans based on forecast results and business constraints. Use Monte Carlo simulation to generate multiple scenarios, apply constrained optimization algorithms to find feasible solutions under the constraints, and combine expert knowledge to design heuristic rule generation solutions.

[0097] Evaluate the feasibility, cost-effectiveness and potential impact of each solution. Design a multi-dimensional evaluation index system, including technical feasibility, economic benefits, risks, etc. Use multi-criteria decision-making methods such as Analytic Hierarchy Process (AHP) for comprehensive evaluation, and apply sensitivity analysis to evaluate the robustness of the solution to changes in key parameters. Provide users with the best decision-making recommendations based on the evaluation results. Weighted summary of the evaluation results of multiple dimensions, combined with the decision maker's preferences, provide several top-ranked solutions, clearly present the advantages and disadvantages of each solution, and assist decision makers in making judgments.

[0098] The following is the implementation process of the decision suggestion in this embodiment:

[0099] In the example of using a multi-objective optimization algorithm, there are the following optimization goals: maximizing production efficiency, minimizing production costs, and maximizing order completion rate; a genetic algorithm is used to find the balance point of these three goals.

[0100] In the example of using the decision tree algorithm, the decision tree is used to predict raw material price fluctuations and changes in order demand. The characteristics include historical price trends, seasonal factors, market indicators, etc. The decision tree is constructed using the CART (Classification and Regression Trees) algorithm.

[0101] Based on the forecast results and business constraints, a variety of possible production adjustment plans are generated, such as using Monte Carlo simulation to generate scenarios for raw material prices and order demand in the next three months; applying the constrained optimization algorithm, the constraints considered include production capacity constraints, inventory space constraints, and minimum order quantity requirements; the generated plans may include: increasing the production volume of certain products and reducing the production volume of other products, adjusting the raw material procurement strategy such as early or delayed procurement, and temporarily increasing production shifts to increase overall production capacity; then evaluate each plan from the aspects of technical feasibility (considering equipment utilization, human resource requirements, etc.), economic benefits (calculating expected profits, return on investment, etc.), and risk assessment (considering market volatility risks, supply chain disruption risks, etc.); using the hierarchical analysis method for comprehensive evaluation, establishing a hierarchical structure: goal->criterion->plan, making pairwise comparisons, calculating weights, and finally calculating the comprehensive score; based on the evaluation results, providing recommendations to decision makers, such as presenting the top three plans and their detailed information, and providing each plan with information including expected output and financial benefits, required resource inputs, potential risks and countermeasures, and implementation schedule.

[0102] In some embodiments of the present invention, continuous optimization of decision-making methods is also included. Specifically, actual production data, including output, cost, order completion, etc., are first collected, and the decision tree model is updated regularly (such as monthly) to improve prediction accuracy; reinforcement learning algorithms are used to optimize the weight settings of multi-objective optimization; a knowledge base is established to record the results and lessons learned from each decision; and a comprehensive system evaluation and adjustment is conducted every quarter.

[0103] Embodiment 2

[0104] The present embodiment is a system for analyzing product defects and failures in enterprise production based on multiple agents. The system implements the method for analyzing product defects and failures in enterprise production based on multiple agents as described in the first embodiment, including a data acquisition agent, a data processing agent, a pattern recognition agent, a fault prediction agent, and a decision support agent.

[0105] The data acquisition agent is used to collect product quality data from the production line in real time, including raw materials, semi-finished products, product appearance dimensions, performance parameters, equipment status and production environment data.

[0106] The data processing intelligent agent pre-processes the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate pre-processed data.

[0107] The pattern recognition agent uses a machine learning algorithm to perform pattern recognition on data to identify potential patterns of product defects. The machine learning algorithm includes a support vector machine and a neural network model.

[0108] The fault prediction agent predicts product failure based on the identified patterns and uses time series analysis or survival analysis methods to provide risk assessment. The models of the time series analysis include autoregressive integrated moving average model, seasonal autoregressive integrated moving average model, Prophet and long short-term memory network. The models of the survival analysis include Kaplan-Meier estimator, Cox proportional hazard model, and accelerated failure time model.

[0109] The decision support agent uses a multi-objective optimization algorithm or a decision tree algorithm to generate a variety of possible production adjustment plans, uses Monte Carlo simulation to generate a variety of scenarios, applies a constrained optimization algorithm to find feasible solutions while satisfying constraints, and combines expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision recommendations to users based on the evaluation results.

[0110] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for analyzing defective and failure products of enterprise production based on multi-agent, characterized in that: The following steps are involved: Collect product quality data from the production line in real time; Preprocessing the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate preprocessed data; Build and train a machine learning model, and use the trained machine learning model to perform pattern recognition on preprocessed data to identify potential patterns of product defects; Based on the identified patterns, a prediction model is constructed and trained to predict product failures, and the trained prediction model is used to predict potential failure times and probabilities; Use multi-objective optimization algorithms or decision tree algorithms to generate a variety of possible production adjustment plans, use Monte Carlo simulation to generate a variety of scenarios, apply constrained optimization algorithms to find feasible solutions while satisfying constraints, and combine expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision-making recommendations to users based on the evaluation results.

2. The method for analyzing enterprise production product defects and failures based on multi-agents according to claim 1, characterized in that: The machine learning model is a support vector machine.

3. The method for analyzing enterprise production product defects and failures based on multi-agents according to claim 1, characterized in that: The machine learning model is a neural network model.

4. The method for analyzing enterprise production product defects and failures based on multi-agents according to claim 1, characterized in that: The prediction model is a time series analysis model, including an autoregressive integrated moving average model, a seasonal autoregressive integrated moving average model, Prophet and a long short-term memory network.

5. The method for enterprise production product defect and failure analysis based on multi-agent according to claim 4, characterized in that: The process of prediction by the time series analysis model is as follows: Collect time series data related to product failure, process missing values ​​and outliers, perform data smoothing and denoising, and divide the processed data into training sets and test sets according to the set ratio; Extract the time characteristics of the data in the training set and perform periodic analysis; Select a specific time series analysis model, use the training set for training and the test set for evaluating the model performance, and save the model whose evaluation results reach the set value; Use the saved model to make predictions for future points in time based on real-time data on the production line.

6. The method for enterprise production product defect and failure analysis based on multi-agent according to claim 1, characterized in that: The prediction model is a survival analysis model, including Kaplan-Meier estimator, Cox proportional hazard model, and accelerated failure time model.

7. The method for analyzing enterprise production product defects and failures based on multi-agents according to claim 6 is characterized in that: The prediction process of the survival analysis model is as follows: Collect product life cycle data, including product usage time and expiration time, process missing values ​​and outliers, and divide the processed data into training set and validation set according to the set ratio; Select variables related to product failure and encode categorical variables; Select a specific survival analysis model, use the training set for training, and estimate the survival function or risk function; Use the validation set to evaluate model performance, calculate survival rates and hazard ratios, and save models that meet model assumptions; Use saved models to predict the remaining life or failure probability of your product.

8. The method for enterprise production product defect and failure analysis based on multi-agent according to claim 1, characterized in that: The multi-objective optimization algorithm first defines multiple objective functions, including maximizing profits, minimizing costs, and maximizing customer satisfaction, and uses Pareto optimization, genetic algorithm or particle swarm optimization method to find the balance point of multiple objectives.

9. The method for enterprise production product defect and failure analysis based on multi-agent according to claim 1, characterized in that: The decision tree algorithm constructs a decision tree model based on historical data, uses information gain and Gini coefficient indicators to select the best splitting attribute, and avoids overfitting through pruning.

10. The enterprise production product defect and failure analysis system based on multi-agent is characterized by: The system implements the enterprise production product defect and failure analysis method based on multi-agent as described in any one of claims 1 to 9, including a data acquisition agent, a data processing agent, a pattern recognition agent, a fault prediction agent and a decision support agent; The data collection agent is used to collect product quality data from the production line in real time, including raw materials, semi-finished products, product appearance dimensions, performance parameters, equipment status and production environment data; The data processing agent pre-processes the product quality data, including data cleaning, standardization, feature extraction and data fusion, to generate pre-processed data; The pattern recognition agent uses a machine learning algorithm to perform pattern recognition on the data to identify potential patterns of product defects, and the machine learning algorithm includes a support vector machine and a neural network model; The fault prediction agent predicts product failure based on the identified pattern and uses time series analysis or survival analysis methods to provide risk assessment. The models of the time series analysis include autoregressive integrated moving average model, seasonal autoregressive integrated moving average model, Prophet and long short-term memory network. The models of the survival analysis include Kaplan-Meier estimator, Cox proportional hazard model and accelerated failure time model. The decision support agent uses a multi-objective optimization algorithm or a decision tree algorithm to generate a variety of possible production adjustment plans, uses Monte Carlo simulation to generate a variety of scenarios, applies a constrained optimization algorithm to find feasible solutions while satisfying constraints, and combines expert knowledge to design heuristic rules to generate plans, evaluate the feasibility, cost-effectiveness and potential impact of each plan, and provide decision recommendations to users based on the evaluation results.

Images