Artificial intelligence-based automobile parts production control method and system

Abstract

The application relates to the technical field of automobile accessory production, and discloses an automobile accessory production control method and system based on artificial intelligence, which comprises the following steps: preprocessing equipment operation data, production environment data and part quality data collected by an Internet of Things device in real time to obtain standardized data after preprocessing; performing feature extraction and selection on the standardized data after preprocessing to obtain equipment operation data and production environment data features related to the part quality data; constructing a prediction model by combining a full connection neural network with a Pontryagin algorithm, training and optimizing the prediction model by using historical data features, and obtaining a trained prediction model. The application can effectively improve the response capability to complex time-varying factors in the automobile accessory production process, and reduce quality fluctuation and the waste rate.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of automobile accessory production, and in particular to an automobile accessory production control method and system based on artificial intelligence. BACKGROUND

[0002] Under the background of rapid development of the global automobile industry, automobile accessory production as a key link of automobile manufacturing directly affects the overall competitiveness of the automobile industry in terms of production efficiency and quality of automobile accessories. Traditional automobile accessory production control methods mainly rely on manual experience and preset fixed procedures for production scheduling, quality detection and equipment management. However, with the increasing diversification and customization of automobile products, the traditional production control method has exposed many drawbacks.

[0003] The traditional production control method is difficult to cope with complex production environment. The production environment factors and equipment operation data are closely related in time, and have a significant impact on the quality of automobile accessories.

[0004] In automobile accessory production, environmental factors such as temperature, humidity, vibration and dust change dynamically with time. These environmental factors and equipment operation data interact with each other, for example: the workshop temperature fluctuates with the outside air temperature and the heat generated by the equipment operation, and high temperature causes the motor temperature of the equipment to rise, which leads to a decrease in motor speed and an increase in energy consumption, thereby affecting the processing precision of the accessories. In addition, in the plum rain season, the humidity in the workshop will increase significantly, which is easy to cause the tool of the metal processing equipment to rust. With the passage of time, the tool wear is aggravated, and the cutting precision is reduced when producing automobile metal parts, and the size deviation of the parts is increased.

[0005] In addition, the vibration generated by the long-time operation of the equipment will cause the equipment mounting base to loosen and the parts to displace, and the running data collected by the sensor will also deviate.

[0006] The traditional production control method is difficult to capture the relationship between these time-varying environmental and equipment data in real time, resulting in unstable quality of the produced automobile accessories, high scrap rate and inability to meet the increasingly strict quality requirements of the automobile industry. Therefore, it is urgent to use artificial intelligence technology to realize precise production control.

[0007] Therefore, the application provides an automobile accessory production control method and system based on artificial intelligence. SUMMARY

[0008] The application provides an automobile accessory production control method and system based on artificial intelligence, which is used to improve the response capability to complex time-varying factors in the automobile accessory production process and reduce the quality fluctuation and scrap rate.

[0009] In a first aspect, the application provides an automobile accessory production control method based on artificial intelligence, which comprises:

[0010] Preprocess the device operation data, production environment data and component quality data collected in real time by the Internet of Things device to obtain standardized data after preprocessing;

[0011] Extract and select features from the standardized data after preprocessing to obtain device operation features and production environment features associated with component quality data;

[0012] Construct a prediction model by combining a fully connected neural network with a Pontryagin algorithm, and train and optimize the prediction model based on the device operation features and production environment features to obtain a trained dynamic quality prediction model;

[0013] Input the preprocessed data collected in real time into the dynamic quality prediction model for prediction to obtain a component quality prediction value for the current batch, and dynamically adjust the device operation parameters based on the prediction value through a feedback control mechanism to obtain a set of real-time optimized device operation parameters;

[0014] Perform multivariate global optimization on the set of real-time optimized device operation parameters and production environment data using a differential evolution algorithm to obtain an optimal production parameter combination, and apply the optimal production parameter combination to an actual production line to obtain a stable production process after parameter optimization;

[0015] Track and verify the stable production process after parameter optimization using a statistical process control method to obtain production stability improvement results and key quality indicator improvement data.

[0016] In combination with the first aspect, in a first implementation manner of the first aspect of the present application, the preprocessing of the device operation data, production environment data and component quality data collected in real time by the Internet of Things device comprises:

[0017] Clean the original data to obtain cleaned data, specifically including: for pulse noise in an automobile parts production device, smoothing the data using a sliding window mean filtering method, and eliminating data frames with a standard deviation exceeding a set threshold within the window;

[0018] Normalize the cleaned data, and normalize temperature, pressure and speed data according to device type to obtain standardized data;

[0019] Detect and replace outliers in the standardized data, identify outliers beyond ±3 times the standard deviation of the mean using a Z-score algorithm, and replace the outliers with the average value of the same device historical data in the same time period to obtain standardized data after preprocessing.

[0020] In a second implementation manner of the first aspect, the feature extraction and selection on the pre-processed standardized data comprises:

[0021] The time series data of the equipment operation data and the production environment data are segmented into a plurality of data frames of fixed time windows according to production batches, to obtain a segmented time series data set;

[0022] The equipment operation data in the segmented time series data set is extracted to obtain an equipment operation feature set, including a sliding average trend line slope, a speed difference sequence variance, an equipment operation time cumulative value, and a time domain statistic of an amplitude;

[0023] The production environment data in the segmented time series data set is extracted to obtain a production environment feature set, including extracting a frequency component whose energy proportion cumulative value reaches a preset threshold through fast Fourier transform on each data sequence, obtaining a dominant frequency component set, and taking a first principal component score as a dominant frequency fluctuation feature through principal component analysis for dimension reduction processing on a spectrum distribution of the dominant frequency component set;

[0024] A Gaussian mixture model is trained based on historical data of normal working conditions of the equipment, and a negative log-likelihood value of each data frame is calculated for the segmented time series data set by using the trained Gaussian mixture model, to obtain an anomaly score set;

[0025] The prediction importance of the equipment operation feature set, the production environment feature set, and the anomaly score set on the part quality index is evaluated through a random forest algorithm, and features with an importance score greater than a preset threshold are screened to obtain a key feature set as a model input.

[0026] In a third implementation manner of the first aspect, the prediction model is constructed through a fully connected neural network combined with a Pontryagin algorithm, comprising:

[0027] An initial prediction model is obtained by defining the equipment operation data as a state variable and the production environment data as a control variable, and constructing a fully connected neural network structure with an input layer dimension of N;

[0028] A dynamic system differential equation and a Hamilton function are constructed through a Pontryagin algorithm, and a companion equation and an extreme condition are solved to obtain an optimal control solution;

[0029] The optimal control solution is taken as an L2 regularization term, combined with a cross-entropy loss function of the neural network to obtain a total loss function;

[0030] The network weight and the control variable parameter are alternately optimized through a gradient descent method until the total loss function converges to a preset threshold, to obtain a trained dynamic quality prediction model.

[0031] With reference to the first aspect, in a fourth implementation form of the first aspect of the present disclosure, the dynamically adjusting the equipment operation parameters according to the predicted value through the feedback control mechanism comprises:

[0032] calculating a deviation value of the predicted value of the part quality output by the prediction model from the target quality index, the deviation value being a weighted sum of a root mean square error and an absolute error, to obtain a comprehensive deviation index;

[0033] inputting the comprehensive deviation index into a PID controller to generate a dynamic adjustment instruction for the equipment operation parameters, including an adjustment range of each parameter, to obtain a parameter adjustment instruction set;

[0034] downloading the parameter adjustment instruction set to a device controller through an industrial Internet of Things platform to control the equipment operation parameters within a preset safety threshold in real time, to obtain a real-time optimized equipment operation parameter set.

[0035] With reference to the first aspect, in a fifth implementation form of the first aspect of the present disclosure, the multi-variable global optimization of the real-time optimized equipment operation parameter set and the production environment data is performed using a differential evolution algorithm, comprising:

[0036] defining a variable range of the equipment operation parameters and the production environment parameters as a constraint condition, constructing a target function with the minimization of the quality prediction error and the energy consumption as a target, to obtain an optimization problem definition;

[0037] performing mutation and crossover operations on the optimization problem definition, generating a trial vector and performing binomial crossover with a parent vector, selecting individuals with better fitness to enter the next generation population, to obtain an iterative optimization population;

[0038] when the target function value of the iterative optimization population changes by less than a threshold value within a continuous preset number of iterations, outputting a current optimal solution, to obtain an optimal production parameter combination.

[0039] With reference to the first aspect, in a sixth implementation form of the first aspect of the present disclosure, the tracking and verification of the stable production process after the parameter optimization is performed using a statistical process control method, comprising:

[0040] collecting key quality index data of a continuous preset batch, drawing an X-bar-R control chart, to obtain a production process stability analysis result;

[0041] calculating a process capability index C pk pk of the production process stability analysis result, and determining the optimization effectiveness in combination with a defect rate improvement rate, if a preset threshold is not reached, retriggering a differential evolution algorithm optimization process, to obtain a final verification conclusion.

[0042] With reference to the first aspect, in a seventh implementation form of the first aspect, the method further comprises device maintenance optimization, specifically comprising:

[0043] The device health state evaluation result is combined with the device failure probability, the device influence on the production line, and the device importance weight, and is input into a maintenance priority scoring model to obtain a maintenance priority score.

[0044] The device health state evaluation result is combined with the device failure probability, the device influence on the production line, and the device importance weight, and is input into a maintenance priority scoring model to obtain a maintenance priority score.

[0045] Based on the maintenance priority score, the intelligent digital platform is used to dynamically adjust the maintenance plan to obtain an optimization effect of reducing the unplanned downtime.

[0046] With reference to the first aspect, in an eighth implementation form of the first aspect, the generation of the maintenance priority scoring model comprises:

[0047] The device failure probability is statistically obtained from historical data, and the failure probability is classified according to the device type to obtain a failure probability level.

[0048] The loss of device downtime on production efficiency, the maintenance cost, and the quality risk are quantified by using the analytic hierarchy process to obtain a device influence degree score.

[0049] According to the functional positioning of the device in the production process, the device importance weight is assigned to obtain a weight distribution result.

[0050] The failure probability level, the device influence degree score, and the weight distribution result are input into a weighted scoring model to generate a maintenance priority score.

[0051] By screening devices with a maintenance priority score higher than a preset threshold, a priority maintenance list is generated and an early warning is triggered to obtain an optimized maintenance execution scheme.

[0052] Secondly, the application provides an automobile parts production control system based on artificial intelligence, which comprises:

[0053] A data acquisition and preprocessing module is configured to preprocess device operation data, production environment data, and part quality data collected by an Internet of Things device in real time to obtain standardized data after preprocessing.

[0054] A feature extraction and selection module is configured to extract and select features from the standardized data after preprocessing to obtain device operation features and production environment features associated with the part quality data.

[0055] a dynamic quality prediction module, configured to construct a prediction model by combining a full connection neural network with a Pontryagin algorithm, and train and optimize the prediction model based on the equipment operation features and the production environment features to obtain a trained dynamic quality prediction model;

[0056] a real-time feedback control module, configured to input real-time collected pretreatment data into the dynamic quality prediction model for prediction to obtain a quality prediction value of the current batch of parts, and dynamically adjust equipment operation parameters based on the prediction value through a feedback control mechanism to obtain a set of real-time optimized equipment operation parameters;

[0057] a production optimization module, configured to perform multivariate global optimization on the set of real-time optimized equipment operation parameters and production environment data by using a differential evolution algorithm to obtain an optimal production parameter combination, and apply the optimal production parameter combination to an actual production line to obtain a stable production process after parameter optimization;

[0058] a production tracking and verification module, configured to track and verify the stable production process after parameter optimization by using a statistical process control method to obtain a production stability improvement result and key quality index improvement data.

[0059] Compared with the prior art, the application has at least the following advantages:

[0060] In the technical scheme provided in the application, the intelligent level of the production process is significantly improved through multi-link collaborative optimization, effectively solving the problems of quality fluctuation and high scrap rate faced by traditional methods in a complex time-varying environment, and specifically embodying in the precise prediction and dynamic control driven by data and the global optimization to improve stability.

[0061] In the application, equipment operation, production environment and part quality data are collected in real time based on Internet of Things equipment, data cleaning and standardization processing are performed through sliding window mean filtering and Z-score anomaly detection technology, time domain and frequency domain feature extraction and random forest algorithm are combined to filter key features, a dynamic quality prediction model is constructed by combining a full connection neural network and a Pontryagin algorithm, and precise prediction of quality indexes such as part size precision and surface roughness is realized.

[0062] In addition, the equipment parameters are adjusted in real time through a PID feedback control mechanism, an optimal production parameter combination is generated through global optimization by using a differential evolution algorithm, quality error and energy consumption are reduced, an LSTM neural network is further introduced to predict potential equipment faults and perform maintenance in advance, non-scheduled downtime is reduced, production cost is saved by reducing scrap rate, energy consumption and non-scheduled downtime, and product quality stability is improved.

[0063] In summary, the application converts production equipment operation data, production environment data and part quality data from independent collection to collaborative optimization, significantly improving part quality stability, production process capability and resource utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0065] Figure 1 A flowchart of the method for controlling automobile parts production based on artificial intelligence in the embodiments of the application;

[0066] Figure 2 A structural diagram of the system for controlling automobile parts production based on artificial intelligence in the embodiments of the application;

[0067] Figure 3 A flowchart of constructing a prediction model by a fully connected neural network combined with a Pontryagin algorithm in the embodiments of the application;

[0068] Figure 4 A flowchart of dynamically adjusting equipment operation parameters by a feedback control mechanism according to a prediction value in the embodiments of the application. DETAILED DESCRIPTION

[0069] The embodiments of the application provide a method and system for controlling automobile parts production based on artificial intelligence. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the application and the above drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variation thereof is intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0070] For the sake of understanding, the specific flow of the embodiments of the application will be described below. Please refer to Figure 1 One embodiment of the method for controlling automobile parts production based on artificial intelligence in the embodiments of the application includes the following steps:

[0071] Step S1, the device running data, production environment data and part quality data collected by the Internet of Things device in real time are preprocessed to obtain standardized data after preprocessing.

[0072] Specifically, the device running data includes machine tool speed, motor temperature, feed amount and other time series data reflecting the state of the device, which are collected by sensors deployed on the device at a frequency of 100ms, such as speed sensor, temperature sensor, displacement sensor.

[0073] The production environment data includes workshop temperature, humidity, vibration amplitude, dust concentration and other environmental data that affect the production process, which are collected by the environmental monitoring terminal deployed in the workshop at a frequency of 5s.

[0074] The part quality data includes size accuracy, surface roughness, material strength and other key data directly representing product quality, which are collected offline by three coordinate measuring instruments, roughness detectors, universal material testing machines and other devices after each batch of production is completed.

[0075] The purpose of preprocessing is to convert the collected raw data into a standardized form suitable for subsequent analysis and modeling.

[0076] Step S2, feature extraction and selection are performed on the standardized data after preprocessing to obtain device running features and production environment features associated with part quality data.

[0077] Specifically, the purpose of feature extraction is to extract indicators or attributes reflecting key information from raw data, such as extracting time domain features from device running data and extracting frequency domain features from production environment data.

[0078] Feature selection is to filter features that have a significant impact on part quality from the extracted features. Random forest algorithm is used to evaluate feature importance, and features with importance score higher than the preset threshold are selected. The purpose is to reduce model complexity, improve prediction efficiency and avoid noise interference.

[0079] Association with part quality data means that the selected features have actual connection with quality indicators. The contribution of features to quality prediction is verified by random forest algorithm, such as device vibration amplitude and part surface roughness, which are highly correlated, so the vibration feature is retained.

[0080] Device running features are dynamic parameters reflecting the running state of the device, including time domain features and abnormal score:

[0081] Among them, the time domain features include:

[0082] The slope of the moving average trend line is used to reflect the running speed change trend.

[0083] Speed difference sequence variance, used to reflect the severity of speed fluctuation.

[0084] Device running time cumulative value, used to count the effective working time of the device.

[0085] Mean and standard deviation of amplitude, used to reflect the intensity of vibration.

[0086] Abnormal score includes the negative log-likelihood value of the data frame calculated by the Gaussian mixture model, used to identify abnormal operating conditions.

[0087] Production environment features are parameters reflecting the impact of workshop environment on the production process, extracted through frequency domain analysis, including:

[0088] Frequency domain features:

[0089] Dominant frequency component, i.e. the frequency with the highest energy proportion extracted by fast Fourier transform.

[0090] First principal component score after principal component analysis dimension reduction, used to reflect the periodic fluctuation pattern of the environment.

[0091] Environmental parameters: temperature and humidity fluctuation characteristics, dust concentration trend.

[0092] More specifically, the time series data of device operation data and production environment data are divided into several data frames with fixed time windows, such as 30 minutes / window, with an overlap rate of 50%, and each data frame corresponds to a production sub-batch.

[0093] Time domain features are extracted from device operation data, including:

[0094] A 5-minute moving average is performed on the speed data within each data frame, a trend line is fitted and the slope is calculated;

[0095] The difference sequence of adjacent time speed is calculated, and its variance is calculated;

[0096] The total time of the device in the running state within the data frame is counted;

[0097] Time domain statistics of amplitude, including mean, standard deviation, maximum and minimum, reflecting the intensity distribution of the vibration signal, obtaining the device operation feature set.

[0098] Fast Fourier transform is performed on the production environment data, and the frequency component whose energy proportion accumulates to a preset threshold is extracted, obtaining the dominant frequency component set, and principal component analysis is performed on it, taking the first principal component score as the dominant frequency fluctuation feature, obtaining the production environment feature set.

[0099] Based on the historical data of the device under normal working conditions, a Gaussian mixture model is trained, and the negative log-likelihood value of each data frame is calculated. The larger the value, the more it deviates from the normal working condition, which is used as the abnormal score set.

[0100] The random forest algorithm is used to calculate the feature importance, and the features with importance scores greater than the preset threshold are selected as the key feature set as the model input.

[0101] Step S3, a prediction model is constructed by a fully connected neural network combined with a Pontryagin algorithm, and the prediction model is trained and optimized based on the device running features and production environment features, to obtain a trained dynamic quality prediction model.

[0102] Specifically, the fully connected neural network is used to learn the nonlinear mapping relationship between the input features and the output target. In this scenario, the input features are the device running features and the production environment features, and the output target is the part quality prediction value.

[0103] The input layer dimension of the fully connected neural network structure is the number of features.

[0104] The hidden layer of the fully connected neural network structure is 2 layers, with 128 and 64 neurons respectively, and the hidden layer activation function is LeakyReLU, and the output layer is a linear activation function.

[0105] The LeakyReLU activation function is an improvement based on the ReLU activation function. When the input is less than 0, it no longer outputs 0, but outputs a very small negative number, avoiding the "neuron death" problem of the ReLU function.

[0106] The Pontryagin algorithm is used to solve the optimal control strategy of the dynamic system by constructing the Hamilton function and the adjoint equation. The device running data is regarded as the system state variable, such as speed and temperature, and the production environment data is regarded as the control variable, such as temperature and humidity setting value. Combined with the state equation and the objective function, such as minimizing the quality prediction error, the optimal control solution is generated.

[0107] Taking the processing of automobile metal parts as an example:

[0108] The device running characteristics are used as input features, and the sliding average slope of the machine tool speed is set to 0.05 / s, and the standard deviation of the vibration amplitude is 2.3μm.

[0109] The features of the production environment are represented by the principal component score of the workshop temperature, which is assumed to be -0.7, reflecting the low temperature fluctuation.

[0110] The predicted surface roughness Ra value of the current batch of parts is 1.6μm, but the actual measured value is 1.5μm, indicating that there is a prediction error.

[0111] For this prediction error, the PID controller issues instructions to increase the coolant flow by 5% to reduce the machining temperature and reduce the fluctuation of surface roughness.

[0112] More specifically, a fully connected neural network with input layer dimension N is constructed, and a dynamic system differential equation and Hamilton function are constructed.

[0113] Solve the adjoint equation, and obtain the optimal control solution according to the extreme value condition.

[0114] The optimal control solution is combined with the cross-entropy loss function as an L2 regularization term to obtain the total loss function.

[0115] Based on the total loss function, the Adam optimizer is used to train the prediction model to obtain the trained dynamic quality prediction model.

[0116] The cross-entropy loss is used to measure the deviation of the predicted value from the true quality indicator, and the L2 regularization term is used to constrain the network weight and improve the model generalization ability.

[0117] The dynamic quality prediction model can dynamically update the prediction results according to the real-time acquisition of device running data and production environment data, such as re-predicting after each batch production. The prediction results are used to generate device parameter adjustment instructions, such as adjusting the speed through the PID controller, forming a "prediction-control-re-prediction" closed loop, and outputting the part quality prediction value, such as size deviation, surface roughness, material strength, etc.

[0118] Step S4, input the real-time collected preprocessed data into the dynamic quality prediction model for prediction to obtain the current batch of part quality prediction value, and dynamically adjust the device running parameters according to the prediction value through the feedback control mechanism to obtain the real-time optimized device running parameter set.

[0119] Specifically, the real-time collected preprocessed data is the device running data and production environment data obtained by the Internet of Things sensor in real time, after data cleaning, standardization and outlier replacement preprocessing steps, the processed data is input into the model in a structured format, such as time series data frame.

[0120] The feedback control mechanism generates device parameter adjustment instructions by taking the deviation between the predicted value and the target value as input. The PID controller generates the adjustment amount according to the current value, cumulative value and rate of change of the deviation. If the predicted surface roughness is higher than the target value, the PID controller calculates the required speed reduction or coolant flow increase.

[0121] Exemplarily, in the machining of automobile engine cylinder block:

[0122] Real-time data are machine speed 1200 RPM, tool temperature 85℃, workshop humidity 60%.

[0123] After denoising by pre-processing filtering, the speed is normalized to 0.8 and the humidity is normalized to 0.6.

[0124] By applying the dynamic quality prediction model, the deviation of the cylinder bore is predicted to be +0.02 mm. Compared with the preset target value of ±0.01 mm, there is a certain difference between the predicted value and the target value.

[0125] In order to solve this deviation, a PID controller is used for calculation, and it is decided to reduce the speed by 3% to 1164 RPM and increase the coolant flow by 10% to compensate for this deviation.

[0126] After parameter optimization and adjustment, the new parameter set is {speed: 1164 RPM, coolant flow: 110%}. The actual machining result shows that the bore deviation has been reduced to +0.008 mm.

[0127] Further, by calculating the deviation between the predicted value and the target quality index, a comprehensive deviation index E is obtained. After inputting the deviation index E into the PID controller, the controller parameters are adjusted by the Ziegler-Nichols method, thereby generating adjustment instructions for the device operating parameters.

[0128] Through the industrial Internet of Things platform, these instructions are issued to the device controller, and the real-time control parameters are within the safety threshold, thereby obtaining the real-time optimized device operating parameter set.

[0129] The optimized device operating parameter set includes device operating parameters such as speed, pressure, feed rate, cooling power, etc., and production environment parameters such as temperature and humidity set values, dust removal system frequency, etc.

[0130] Step S5, the differential evolution algorithm is used for multivariate global optimization on the real-time optimized device operating parameter set and production environment data, and the optimal production parameter combination is obtained. The optimal production parameter combination is applied to actual production to obtain stable production data after parameter optimization.

[0131] Specifically, the differential evolution algorithm searches for the optimal solution in the solution space by simulating the "mutation-crossover-selection" mechanism in biological evolution.

[0132] Multivariate global optimization avoids falling into local optimal solution by optimizing device parameters and environment parameters at the same time, and finds the globally optimal parameter combination. The optimal production parameter combination obtained minimizes the quality prediction error and minimizes energy consumption. The device parameters need to be within the safety threshold. The optimal production parameter combination includes a combination of device parameters and environment parameters.

[0133] A local optimal solution refers to an optimal solution found within a limited parameter range, but this optimal solution is not optimal in the entire possible parameter space. For example, adjusting the parameter combination of device rotation speed and environmental temperature may reduce the quality error of the current batch to a certain level, i.e., a local optimum, but there is another parameter combination, such as higher rotation speed combined with specific humidity control, which can make the error smaller and energy consumption lower, i.e., a global optimum.

[0134] Applying the optimal production parameter combination to actual production includes:

[0135] Sending the optimized parameters to the device controller through the industrial Internet of Things platform, adjusting the device operating parameters using the controller, and continuously collecting data from the sensors to ensure that the parameter execution meets the expectations.

[0136] The stable production data after parameter optimization includes stability performance, which is reflected in the following aspects: the fluctuation range of key quality indicators is reduced; the fluctuation of device operating parameters is reduced; and the impact of environmental interference on the production process is reduced.

[0137] And verify by drawing X-bar-R control chart, calculate process capability index C pk , ensure that C pk greater than or equal to 1.33.

[0138] Taking automobile gear machining as an example:

[0139] The input parameters include device parameters and environmental parameters, such as the initial rotation speed of the machine tool device set to 1200 RPM and the cooling liquid flow set to 80 L / min. In the environmental parameters, the initial temperature of the workshop is 25℃.

[0140] A series of test parameter combinations are generated through the differential evolution algorithm, such as adjusting the rotation speed to 1250 RPM, increasing the cooling liquid flow to 90 L / min, and reducing the workshop temperature to 22℃.

[0141] After the calculation of the target function, the results show that the quality error, i.e., the gear tooth spacing deviation, is reduced by 15%, and the energy consumption is reduced by 10%. Specifically, the gear tooth spacing deviation is reduced from ±0.03mm to ±0.01mm, the pass rate is increased from 90% to 98%, the device energy consumption is reduced by 8%, and the impact of workshop temperature and humidity fluctuations on the machining process is significantly alleviated.

[0142] At this time, the test parameter combination generated by the differential evolution algorithm is the optimal production parameter combination.

[0143] More specifically, the variable range of device operating parameters and production environment parameters is defined as a constraint condition, such as rotation speed, feed rate, and cooling pressure, and temperature and humidity.

[0144] The objective function is constructed, the dimension influence is eliminated by normalization processing, the mutation operation and the crossover operation are combined, the test vector is compared with the parent vector, and the individual with better fitness is selected into the next generation.

[0145] When the change amplitude of the objective function value is less than a preset threshold or reaches a maximum iteration number in continuous preset iteration rounds, output the current optimal solution as a parameter combination, and apply it to the production line.

[0146] Step S6, track and verify the stable production process after parameter optimization by the statistical process control method, and obtain the production stability improvement result and the key quality index improvement data.

[0147] Specifically, the statistical process control method monitors the stability of the process and identifies abnormal fluctuations by collecting and analyzing data in the production process, wherein whether the process is in a statistical control state is determined by an X-bar-R control chart, and the ability of the process to meet the quality specification is quantified by a process capability index C pk to ensure that the process is stable and controllable.

[0148] The tracking and verification include:

[0149] Continuously collect key quality data in the optimized production process, such as size measurement value of each batch and surface roughness, and collect data according to production batch or fixed time interval;

[0150] Draw an X-bar-R control chart to check whether the data points are within the control limit, such as ±3σ, and identify abnormal patterns, such as consecutive 7-point rise, calculate the C pk value after optimization, and determine whether the preset target is reached, such as whether the C pk value is greater than or equal to 1.33.

[0151] The quantitative indicators of the production stability improvement result include: control chart performance, process capability index C pk , fluctuation range, etc.

[0152] In the key quality index improvement data, the key quality index includes: accessory size precision, accessory surface quality and accessory performance index, etc.

[0153] The improvement data includes: defect rate, pass rate and customer feedback, etc.

[0154] Taking automobile brake disc processing as an example:

[0155] First, by differential evolution algorithm, the speed of the machine tool is increased from 1200 RPM to 1280 RPM, and the cooling liquid flow is adjusted from 80 L / min to 95 L / min;

[0156] Then, 20 batches of brake disc thickness data were continuously collected, each batch containing 50 products. By drawing an X-bar-R chart, it was found that all data points were within the control limit and showed no abnormal trend.

[0157] Through process capability analysis, the Cpk value was improved from 1.2 to 1.8, indicating that the process capability was significantly improved.

[0158] In terms of quality indicators, the thickness tolerance was reduced from ±0.05mm to ±0.02mm, the surface roughness Ra value was reduced from 1.8μm to 1.2μm, and the batch defect rate was reduced from 2.5% to 0.3%.

[0159] More specifically, the key quality indicator data of the specified batch was continuously collected, an X-bar-R control chart was drawn, the control limit was calculated, the control limit was ±3σ, and whether the data points exceeded the control limit or showed abnormal patterns was analyzed.

[0160] The process capability index C pk , combined with the defect rate improvement rate, if C pk is less than the preset threshold or the defect rate improvement rate is less than the preset threshold, the differential evolution algorithm optimization process is retriggered, otherwise the optimization is confirmed to be effective.

[0161] In a specific embodiment, the device running data, production environment data and component quality data collected in real time by the Internet of Things device are preprocessed, specifically including the following steps:

[0162] The original data is cleaned to obtain cleaned data, specifically including: for the pulse noise in the automobile accessory production equipment, the sliding window mean filtering method is used to smooth the data, and the data frames with standard deviation exceeding the set threshold in the window are removed;

[0163] The cleaned data is normalized, and the temperature, pressure and speed data are normalized according to the device type to obtain standardized data;

[0164] The standardized data is subjected to abnormal value detection and replacement, abnormal values exceeding ±3 times the standard deviation of the mean value are identified by Z-score algorithm, and the average value of the same device historical data in the same time period is used to replace the abnormal values to obtain preprocessed standardized data.

[0165] Specifically, the sliding window mean filtering method is a signal processing method that slides a fixed size window over the data sequence, calculates the average value of the data in the window, and replaces the data value at the center of the window with the average value to smooth the data and reduce noise interference.

[0166] The min-max normalization method is used to linearly transform data to a specific interval, usually [0, 1] or [-1, 1] etc.

[0167] Z-score algorithm is used to detect outliers in the data, which is used to convert the data into a standard normal distribution with mean 0 and standard deviation 1.

[0168] Further, the automobile parts production equipment generates pulse noise during operation, which interferes with the normal data characteristics. A sliding window mean filtering method is adopted, and the window size is set to 5 data points. The window is slid on the data sequence, and the standard deviation of the data in the window is calculated.

[0169] If the standard deviation of the data in a window exceeds three times the global standard deviation, it indicates that the data in the window may be affected by pulse noise. In this case, these data frames should be removed.

[0170] For example, when processing the time series data of the speed of a machine tool, this method can effectively remove abnormal high or low speed values caused by transient interference.

[0171] Different types of equipment data have different dimensions and value ranges. In order to facilitate subsequent processing and model learning, normalization is needed. According to the type of equipment, the minimum-maximum normalization method is used.

[0172] Taking temperature data as an example, it is normalized to the range of [0, 1] to unify the scale of data collected by different temperature sensors; pressure data is normalized to the range of [-1, 1], and speed data is normalized to the range of [0.2, 0.8].

[0173] Taking a certain equipment as an example, the temperature data range is between 20℃ and 50℃. After normalization, 20℃ will be converted to 0, and 50℃ will be converted to 1.

[0174] The Z-score algorithm is used to detect the standardized data, calculate the mean μ and standard deviation σ of the data, and those data points that exceed the mean ± 3 times the standard deviation are judged as outliers.

[0175] In order to ensure the continuity and reliability of the data, the average value of the same production shift in the historical data of the same equipment is used to replace it.

[0176] Taking a specific example, when a certain equipment has an abnormal value significantly higher than the normal range in the temperature data of a specific production shift, the historical average temperature of the same equipment in the same production shift can be used to replace the abnormal value.

[0177] In a specific embodiment, feature extraction and selection are performed on the preprocessed standardized data, which includes the following steps:

[0178] The time series data of the equipment operation data and the production environment data is divided into several data frames of fixed time windows according to production batches, to obtain a segmented time series data set;

[0179] Time domain features are extracted from the equipment operation data in the segmented time series data set, including sliding average trend line slope, speed difference sequence variance, equipment operation time cumulative value, and time domain statistics of amplitude, to obtain an equipment operation feature set;

[0180] Frequency domain features are extracted from the production environment data in the segmented time series data set, by performing fast Fourier transform on each data sequence, extracting frequency components whose energy proportion cumulative value reaches a preset threshold, selecting the top 10 frequency components, obtaining a dominant frequency component set, and performing dimensionality reduction processing on the frequency spectrum distribution of the dominant frequency component set by principal component analysis, taking the first principal component score as the dominant frequency fluctuation feature, to obtain a production environment feature set;

[0181] A Gaussian mixture model is trained based on historical data of normal equipment operating conditions, and the trained Gaussian mixture model is used to calculate the negative log-likelihood value of each data frame of the segmented time series data set, to obtain an anomaly score set;

[0182] The prediction importance of the equipment operation feature set, the production environment feature set, and the anomaly score set on the quality indicator of the part is evaluated by a random forest algorithm, and features with importance scores greater than a preset threshold are selected to obtain a key feature set as the model input.

[0183] Specifically, the continuous time series data is divided into multiple fixed-length segments according to certain rules, so that subsequent independent feature extraction and analysis can be performed on each segment.

[0184] The time series data of the equipment operation data and the production environment data is divided into 30-minute fixed time windows according to production batches, with a 15-minute overlap between adjacent windows. The purpose of this is to ensure the continuity and integrity of the data, while also extracting data features of different time periods.

[0185] For example, for the time series data of the machine tool speed, each 30-minute data frame has a 15-minute overlapping portion with the adjacent data frame.

[0186] When extracting time domain features from the equipment operation data:

[0187] The speed data in each data frame is subjected to 5-minute sliding average using the least squares method, and then a trend line is fitted, and the slope of the trend line is calculated. The slope reflects the change trend of the equipment operating speed, for example, a positive slope indicates that the equipment speed is rising, and a negative slope indicates that the equipment speed is falling;

[0188] The difference sequence of the rotation speed at adjacent time points is calculated, and then the variance of the difference sequence is calculated. The greater the variance, the more intense the speed fluctuation, reflecting the stability of the equipment operation.

[0189] The number of timestamps of the equipment in the running state within the statistical window is counted to obtain the total running time of the equipment in the time period.

[0190] The mean, standard deviation, kurtosis and skewness of the amplitude are calculated. The mean reflects the average level of the amplitude, the standard deviation reflects the fluctuation degree of the amplitude, the kurtosis reflects the peak degree of the amplitude distribution, and the skewness reflects the asymmetry degree of the amplitude distribution.

[0191] The production environment data is subjected to fast Fourier transform to convert the time domain signal into a frequency domain signal. The frequency components whose energy proportion accumulates to 90% are extracted to obtain a dominant frequency component set. The spectral distribution data of the dominant frequency components are input into principal component analysis. The spectral distribution is processed by dimension reduction through principal component analysis. The obtained principal components are sorted according to the variance contribution rate. The first three principal components are retained. The cumulative variance contribution rate of the first three principal components is required to be greater than or equal to 95%. The first three principal components can explain more than 95% of the changes in the original data, indicating that they almost cover all important periodic characteristics. The first principal component score is taken as the dominant frequency fluctuation feature. This feature reflects the periodic variation law of the environmental factors.

[0192] Based on the historical data of the normal working condition of the equipment, 1000 data frames with no fault record are selected to train the Gaussian mixture model. The EM algorithm is used to initialize the model parameters. The negative log-likelihood value of each data frame is calculated. If the value is greater than the mean + 2 times the standard deviation, the data frame is determined to be abnormal. The negative log-likelihood value is taken as the abnormal score.

[0193] The feature importance is evaluated by the random forest algorithm. The random forest algorithm calculates the feature importance of each feature in the decision tree. The features with an importance score greater than 1.2 times the average score are selected to obtain a key feature set as the model input.

[0194] It should be noted that the first principal component generated by principal component analysis is the direction with the largest data variance, which can explain the largest variability in the original spectral distribution. The first principal component is PC1. The first principal component score is the projection value of the spectrum of each data frame in the direction of the first principal component. The first principal component score quantifies the position of the sample in the main fluctuation mode, reflecting the intensity or regularity of the periodic variation of the production environment.

[0195] The first principal component score integrates the fluctuation characteristics of the dominant frequency components, for example:

[0196] If the workshop temperature presents periodic fluctuation, for example, the temperature rises once every two hours, the main frequency component may correspond to 0.5 Hz, and the period is two hours.

[0197] A higher first principal component score means that the periodic fluctuation is more significant in the current time period, while a lower score indicates that the fluctuation is more flat or there is an anomaly.

[0198] For example:

[0199] Suppose the spectral analysis of production environment data such as vibration signals shows that the dominant frequency components are 10Hz, 20Hz, and 30Hz, with amplitudes of 5, 3, and 2 respectively, and the total energy proportion reaches 90%.

[0200] Further, it is found that the amplitude changes of 10Hz and 20Hz are highly correlated, and they can be combined into PC1 through principal component analysis, for example: the first principal component = 0.7 x 10Hz amplitude + 0.3 x 20Hz amplitude.

[0201] If the 10Hz amplitude of a certain data frame is 6 and the 20Hz amplitude is 4, then the first principal component score is: 0.7 x 6 + 0.3 x 4 = 5.4

[0202] This first principal component score quantifies the intensity of the main vibration mode in this time period, which can be used as a production environment feature input into the prediction model.

[0203] In a specific embodiment, a prediction model is constructed by a fully connected neural network combined with Pontryagin algorithm, as shown in Figure 2 The prediction model construction process is shown in the figure, including the following steps:

[0204] A fully connected neural network structure with input layer dimension N is constructed, the device running data is defined as the state variable, and the production environment data is defined as the control variable, to obtain an initial prediction model;

[0205] The Pontryagin algorithm is used to construct the differential equation of the dynamic system and the Hamilton function, and the adjoint equation and the extremum condition are solved to obtain the optimal control solution;

[0206] The optimal control solution is used as an L2 regularization term, combined with the cross-entropy loss function of the neural network, to obtain the total loss function;

[0207] Based on the total loss function, the network weights and control variable parameters are alternately optimized by gradient descent method until the total loss function converges to a preset threshold, and a trained dynamic quality prediction model is obtained.

[0208] Specifically, the fully connected neural network is composed of an input layer, a hidden layer, and an output layer, which learns the features and patterns of data by continuously adjusting the weights between neurons.

[0209] Pontryagin algorithm solves the adjoint equation and the extremum condition to obtain the optimal control solution by constructing Hamilton function.

[0210] Further, a fully connected neural network with an input layer dimension of N is constructed to obtain an initial prediction model, where N = the number of equipment operation features + production environment features + abnormal score.

[0211] For example, if there are 10 equipment operation features, 5 production environment features, and 1 abnormal score, the input layer dimension N = 16.

[0212] A dynamic system differential equation is constructed by Pontryagin algorithm, such as:

[0213]

[0214] Where: A is the state matrix, B is the control matrix, x is the state variable, u is the control variable, and w is the Gaussian white noise.

[0215] The Hamilton function is constructed as:

[0216] H = (x - x d ) T Q(x - x d ) + u T Ru + λ T (Ax + Bu)

[0217] Where: x d is the target state, Q and R are weight matrices, and λ is the adjoint variable.

[0218] Solve the adjoint equation and the extremum condition to obtain the optimal control solution u * = -R -1 B T λ.

[0219] The optimal control solution is used as an L2 regularization term with a weight coefficient of 0.005, combined with the cross-entropy loss function of the neural network, to obtain the total loss function:

[0220] L = L ce + 0.005||u * || 2

[0221] Where: L ce is the cross-entropy loss function, and u * is the optimal control solution.

[0222] The L2 regularization term can prevent model overfitting and improve the generalization ability of the model.

[0223] The network weight and the control variable parameter are alternately optimized by the gradient descent method, the gradient descent method learning rate is 0.0005, the momentum is 0.9, the control variable is updated once every 100 batches, until the total loss function converges to a preset threshold, such as 1e-4, and the trained dynamic quality prediction model is obtained.

[0224] In a specific embodiment, the equipment operation parameters are dynamically adjusted according to the predicted value through a feedback control mechanism, such as Figure 3 The dynamic adjustment process is shown in the schematic diagram, which specifically includes the following steps:

[0225] The deviation value of the predicted value of the part quality predicted by the prediction model and the target quality index is calculated, the deviation value is the weighted sum of the root mean square error and the absolute error, and the comprehensive deviation index is obtained;

[0226] The comprehensive deviation index is input into the PID controller to generate a dynamic adjustment instruction for the equipment operation parameters, including the adjustment range of each parameter, and a parameter adjustment instruction set is obtained;

[0227] The parameter adjustment instruction set is issued to the equipment controller through the industrial Internet of Things platform to control the equipment operation parameters within the preset safety threshold in real time, and a real-time optimized equipment operation parameter set is obtained, and the industrial Internet of Things platform is preferably an MQTT protocol.

[0228] Specifically, the root mean square error is an index for measuring the difference between the predicted value and the true value, which calculates the square root of the average value of the square of the difference between the predicted value and the true value.

[0229] The absolute error is the average value of the absolute value of the difference between the predicted value and the true value.

[0230] The PID controller is a common feedback control algorithm, which is composed of three parts: proportion, integral and differential. By adjusting these three parameters, the output of the system can be as close to the target value as possible.

[0231] Further, by calculating the deviation value of the predicted value of the part quality predicted by the prediction model and the target quality index, the deviation value is the weighted sum of the root mean square error and the absolute error, and the weights are assumed to be 0.7 and 0.3, respectively, to obtain the comprehensive deviation index.

[0232] Specifically, if the predicted value is The true value is y, the root mean square error is The absolute error is The comprehensive deviation index E = 0.7 x RMSE + 0.3 x MAE.

[0233] The comprehensive deviation index is input into the PID controller, the proportion coefficient Kp = 0.5, the integral coefficient K i = 0.1, and the differential coefficient K d = 0.2.

[0234] The PID controller generates dynamic adjustment instructions for the device operating parameters according to the deviation value, including the adjustment range of each parameter, such as a speed adjustment range of ±100 rpm and a feed amount adjustment range of ±5 mm / min, to obtain a parameter adjustment instruction set.

[0235] The parameter adjustment instruction set is distributed to the device controller through an industrial Internet of Things platform, such as the MQTT protocol, and the device controller controls the device operating parameters in real time according to the instructions within the preset safety threshold, such as a speed not lower than 80% of the minimum rated speed, to obtain a real-time optimized device operating parameter set.

[0236] In a specific embodiment, a differential evolution algorithm is used for multivariate global optimization of the real-time optimized device operating parameter set and production environment data, specifically including the following steps:

[0237] The variable range of the device operating parameters and the production environment parameters is defined as a constraint condition, a target function is constructed with the quality prediction error and energy consumption minimization as the target, and an optimization problem definition is obtained.

[0238] The optimization problem definition is subjected to mutation and crossover operations, a trial vector is generated and binomially crossed with the parent vector, and individuals with better fitness are selected into the next generation population, to obtain an iterative optimization population.

[0239] When the target function value of the iterative optimization population changes by less than a threshold value within a continuous preset number of iterations, the current optimal solution is output, and an optimal production parameter combination is obtained.

[0240] Specifically, the differential evolution algorithm is a global optimization algorithm based on swarm intelligence, which constantly searches for the optimal solution by simulating mutation, crossover, and selection operations in the biological evolution process.

[0241] Mutation operation is an operation in the differential evolution algorithm, which generates a mutation vector by performing differential operation on individuals in the population.

[0242] Crossover operation is an operation that combines the mutation vector and the parent vector to generate a trial vector.

[0243] Exponential crossover is a crossover operation method that selects elements of the parent vector and the mutation vector according to an exponential distribution probability to generate a trial vector.

[0244] Binomial crossover is another crossover operation method that selects elements of the parent vector and the mutation vector with a certain probability to generate a trial vector.

[0245] Further, define the variable range of the equipment operating parameters and the production environment parameters as constraints, such as the rotation speed 600-1400 rpm, the temperature 18-28℃, and construct the objective function with the mass prediction error and the energy consumption minimization as the target:

[0246] Get the optimization problem definition, RMSE is the root mean square error.

[0247] Perform mutation and crossover operations on the optimization problem definition, use DE / rand / 2 strategy for mutation operation, mutation factor F=0.6, that is, randomly select three individuals in the population for difference operation to generate a mutation vector.

[0248] The crossover operation adopts exponential crossover, the crossover number is 3 times, generates a trial vector, and then performs binomial crossover on the trial vector and the parent vector, selects the individual with better fitness, i.e. smaller objective function value, to enter the next generation population, and gets the iterative optimization population.

[0249] When the objective function value of the iterative optimization population changes by less than 0.3% in continuous 30 iterations, it means that the algorithm has converged, and the current optimal solution is output, and the optimal production parameter combination is obtained.

[0250] In a specific embodiment, the stable production process after parameter optimization is tracked and verified by statistical process control method, which includes the following steps:

[0251] Collect the key quality index data of continuous preset batches, draw the X-bar-R control chart, and get the production process stability analysis result;

[0252] Calculate the process capability index C pk p of the production process stability analysis result, and determine the optimization effectiveness combined with the defect rate improvement rate, if the preset threshold is not reached, re-trigger the differential evolution algorithm optimization process to get the final verification conclusion.

[0253] Specifically, statistical process control is a quality management method based on statistical principles, which collects, analyzes and monitors data in the production process, discovers abnormal fluctuations in the process in time, and takes measures to adjust to ensure the stability of product quality.

[0254] X-bar-R control chart is a commonly used statistical process control tool, which consists of mean control chart and range control chart, mean control chart is used to monitor the mean value change of the process, and range control chart is used to monitor the fluctuation degree of the process.

[0255] Process capability index C pk is an index to measure the ability of the process to meet quality requirements, which considers the deviation of the mean value of the process from the specification center and the fluctuation degree of the process.

[0256] Further, the key quality indicator data of 15 consecutive batches, such as size accuracy and surface roughness, are collected, and the sample mean of each batch data is calculated and the sample range R.

[0257] According to the sample size, the control chart coefficient table is checked to obtain the A2 value, and the upper control limit is determined the lower control limit Draw the X-bar-R control chart to obtain the production process stability analysis result. Whether the production process is stable can be judged by observing whether the data points exceed the control limit or present abnormal patterns, such as continuous 7-point rise.

[0258] The process capability index C pk is calculated according to the production process stability analysis result, and the formula is:

[0259]

[0260] Among them:

[0261] C pk is the process capability index, which is used to measure the ability of the production process to meet the quality specifications;

[0262] μ is the sample mean of the quality indicator, which is used to reflect the process center position;

[0263] LSL is the lower limit of the specification, that is, the minimum allowable value of the quality indicator;

[0264] USL is the upper limit of the specification, that is, the maximum allowable value of the quality indicator;

[0265] σ is the sample standard deviation of the quality indicator, which is used to reflect the process fluctuation degree.

[0266] The optimization effectiveness is determined in combination with the defect rate improvement rate. If C pk is less than 1.0 or the defect rate improvement rate is less than 5%, it means that the production process capacity is insufficient or the optimization effect is not obvious, then the differential evolution algorithm optimization process is retriggered to obtain the final verification conclusion.

[0267] The defect rate improvement rate is calculated as follows:

[0268] In a specific embodiment, the method further includes equipment maintenance optimization, specifically including the following steps:

[0269] Real-time analysis of equipment operation data through LSTM neural network, prediction of potential equipment failure and generation of maintenance warning, and obtaining equipment health status evaluation result;

[0270] The device health state evaluation result is combined with the device failure probability, the influence degree of the device on the production line and the importance weight of the device, input into a maintenance priority score model, and a maintenance priority score is obtained.

[0271] Based on the maintenance priority score, the intelligent digital platform is used to dynamically adjust the maintenance plan, and an optimization effect of reducing the non-scheduled downtime is obtained.

[0272] The generation of the maintenance priority score model includes the following steps:

[0273] The device failure probability is statistically obtained from historical data, and the failure probability level is divided according to the type of the device, and the failure probability level is obtained.

[0274] The loss of device downtime on production efficiency, maintenance cost and quality risk are quantified by using the analytic hierarchy process, and the device influence degree score is obtained.

[0275] According to the functional positioning of the device in the production process, the importance weight of the device is given, and the weight distribution result is obtained.

[0276] The failure probability level, the device influence degree score and the weight distribution result are input into a weighted scoring model to generate a maintenance priority score.

[0277] By screening the devices with a maintenance priority score higher than a preset threshold, a priority maintenance list is generated and a warning is triggered, and an optimized maintenance execution scheme is obtained.

[0278] Specifically, the fault prediction and health assessment includes:

[0279] Through the LSTM neural network, the sensor data on the device, such as vibration, temperature, current signal, etc., is analyzed in real time, the potential failure of the device, such as bearing wear, motor overheating, etc., is predicted, and the device health state score is output, the score range is 0-10 points, 10 points for the worst, for example, continuous high-frequency vibration warning score is 8 points, the LSTM neural network can capture the time sequence dependence of the device operation data, for example, continuous 5 hours of high-frequency vibration may indicate bearing abnormality.

[0280] Further, multi-factor priority determination:

[0281] The following three types of parameters are combined to generate the maintenance priority score:

[0282] Failure probability score: based on historical failure rate statistics, the annual failure times are divided into levels and converted into 0-10 points, for example, 1 point for each occurrence of annual failure, and the highest score is 10 points.

[0283] Influence degree score:

[0284] Production efficiency loss score: divided by the amount of loss per hour, such as 1 point for every 100,000 yuan loss, up to 10 points;

[0285] Maintenance cost score: divided by the single maintenance cost, such as 1 point for every 10,000 yuan of maintenance cost, up to 10 points;

[0286] Quality risk score: divided by the percentage increase in defective rate, such as 1 point for every 1% increase in defective rate, up to 10 points.

[0287] Weight distribution: production efficiency loss weight is 0.6, maintenance cost weight is 0.3, and quality risk weight is 0.1.

[0288] Influence degree score calculation formula:

[0289] Influence degree score = (production efficiency loss score × 0.6) + (maintenance cost score × 0.3) + (quality risk score × 0.1)

[0290] The device importance weight distribution method is:

[0291] Core equipment, i.e. directly affecting the production capacity of equipment, is 0.8-1.0 points;

[0292] Auxiliary equipment, i.e. indirectly affecting the production capacity of equipment, is 0.3-0.7 points;

[0293] Redundant equipment, i.e. replaceable equipment, is 0.1-0.2 points.

[0294] Through the weighted score model, the maintenance priority score is generated.

[0295] Priority score = (health status score × 0.4) + (failure probability score × 0.3) + (influence degree score × 0.2) + (importance weight × 0.1).

[0296] Based on the priority score, the intelligent digital platform dynamically adjusts the maintenance plan according to the score:

[0297] If the priority score is greater than 8 points: immediately stop for maintenance, and automatically assign standby equipment;

[0298] If the priority score is 4-8 points: included in the weekly maintenance plan to avoid the production peak;

[0299] If the priority score is less than 4 points: continue to monitor and delay to handle during the production gap.

[0300] Taking the automobile engine cylinder body processing production line as an example:

[0301] Failure prediction: The vibration sensor of the CNC machine tool detects that the amplitude increases from 5 μm to 18 μm, and the LSTM model predicts that the main shaft bearing has a risk of failure within 48 hours, with a health score of 9.

[0302] The machine tool has an average of 4 failures per year, so the score is 4.

[0303] Production efficiency loss: 20 cylinder bodies will be lost for 1 hour of downtime each time, equivalent to a loss of 200,000 yuan, scoring 2.

[0304] Maintenance cost: The cost of replacing the main shaft is 50,000 yuan, scoring 5.

[0305] Quality risk: The defect rate increases from 1% to 5%, an increase of 4%, scoring 4.

[0306] The influence degree score is calculated by AHP:

[0307] (2 x 0.6) + (5 x 0.3) + (4 x 0.1) = 3.1

[0308] Importance weight: As a core processing equipment, its weight is 0.9.

[0309] The priority score is: (9 x 0.4) + (4 x 0.3) + (3.1 x 0.2) + (0.9 x 0.1) = 5.51

[0310] Since the priority score is 5.51, which is in the range of 5-7, it is included in the weekly maintenance plan to avoid maintenance during the production peak.

[0311] The above describes the method for controlling the production of automobile parts based on artificial intelligence in the embodiments of the application, and the following describes the system for controlling the production of automobile parts based on artificial intelligence in the embodiments of the application, please refer to Figure 4 An embodiment of the system for controlling the production of automobile parts based on artificial intelligence in the embodiments of the application includes:

[0312] A data acquisition and preprocessing module for preprocessing device operation data, production environment data and component quality data collected by Internet of Things devices in real time to obtain standardized data after preprocessing;

[0313] A feature extraction and selection module for extracting and selecting features from standardized data after preprocessing to obtain device operation features and production environment features associated with component quality data;

[0314] a dynamic quality prediction module, configured to construct a prediction model by combining a full connection neural network with a Pontryagin algorithm, and train and optimize the prediction model based on equipment operation characteristics and production environment characteristics to obtain a trained dynamic quality prediction model;

[0315] a real-time feedback control module, configured to input real-time collected pretreatment data into the dynamic quality prediction model for prediction to obtain a current batch of component quality prediction values, and dynamically adjust equipment operation parameters according to the prediction values through a feedback control mechanism to obtain a real-time optimized equipment operation parameter set;

[0316] a production optimization module, configured to perform multivariate global optimization on the real-time optimized equipment operation parameter set and production environment data by using a differential evolution algorithm to obtain an optimal production parameter combination, and apply the optimal production parameter combination to actual production to obtain a parameter-optimized stable production process;

[0317] a production tracking and verification module, configured to track and verify the parameter-optimized stable production process by using a statistical process control method to obtain production stability improvement results and key quality index improvement data.

[0318] The above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the same; even though the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced by equivalents; and such modifications or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for controlling the production of automotive parts based on artificial intelligence, characterized in that, The method includes: Preprocess the real-time equipment operation data, production environment data and component quality data collected by IoT devices to obtain preprocessed standardized data; Feature extraction and selection are performed on the preprocessed standardized data to obtain equipment operation characteristics and production environment characteristics associated with component quality data; A prediction model is constructed by combining a fully connected neural network with the Pontryagin algorithm, and the prediction model is trained and optimized based on the equipment operation characteristics and production environment characteristics to obtain a trained dynamic quality prediction model. The preprocessed data collected in real time is input into the dynamic quality prediction model for prediction, and the predicted quality value of the parts in the current batch is obtained. Based on the predicted value, the equipment operating parameters are dynamically adjusted through a feedback control mechanism to obtain a set of real-time optimized equipment operating parameters. The differential evolution algorithm is used to perform multivariate global optimization on the real-time optimized set of equipment operating parameters and production environment data to obtain the optimal combination of production parameters. The optimal combination of production parameters is then applied to the actual production line to obtain a stable production process after parameter optimization. The stable production process after parameter optimization is tracked and verified by statistical process control methods to obtain the results of improved production stability and data on the improvement of key quality indicators. The method of constructing a prediction model using a fully connected neural network combined with the Pontryagin algorithm includes: A fully connected neural network structure with an input layer dimension of N is constructed, and the equipment operation data is defined as the state variable and the production environment data as the control variable to obtain the initial prediction model; The dynamic system differential equations and Hamiltonian functions are constructed using the Pontryagin algorithm. The adjoint equations and extremum conditions are then solved to obtain the optimal control solution. The optimal control solution is used as an L2 regularization term and combined with the cross-entropy loss function of the neural network to obtain the total loss function; The network weights and control variable parameters are alternately optimized using gradient descent until the total loss function converges to a preset threshold, thus obtaining the trained dynamic quality prediction model.

2. The method according to claim 1, characterized in that, Preprocessing of real-time equipment operation data, production environment data, and component quality data collected by IoT devices includes: The original data is cleaned to obtain cleaned data. Specifically, for impulse noise in automotive parts production equipment, the sliding window mean filtering method is used to smooth the data, and data frames with a standard deviation exceeding a set threshold are removed from the window. The cleaned data is normalized by normalizing the temperature, pressure, and speed data according to the equipment type to obtain standardized data. Outlier detection and replacement are performed on the standardized data. Outliers exceeding the mean ± 3 times the standard deviation are identified using the Z-score algorithm, and the outliers are replaced with the average value of historical data from the same device over the same time period, resulting in preprocessed standardized data.

3. The method according to claim 1, characterized in that, The feature extraction and selection of the preprocessed standardized data includes: The time-series data of equipment operation data and production environment data are divided into several data frames with fixed time windows according to production batches to obtain segmented time-series datasets; Time-domain features are extracted from the equipment operation data in the segmented time-series dataset, including the slope of the moving average trend line, the variance of the velocity difference sequence, the cumulative value of equipment operation time, and the time-domain statistics of amplitude, to obtain the equipment operation feature set; Frequency domain features are extracted from the production environment data in the segmented time series dataset. By performing a fast Fourier transform on each data sequence, frequency components whose cumulative energy proportion reaches a preset threshold are extracted to obtain the dominant frequency component set. Principal component analysis is used to reduce the dimensionality of the spectral distribution of the dominant frequency component set, and the score of the first principal component is taken as the dominant frequency fluctuation feature to obtain the production environment feature set. A Gaussian mixture model is trained based on historical data of normal operating conditions of the equipment. The negative log-likelihood value of each data frame is calculated using the trained Gaussian mixture model on the segmented time series dataset to obtain an anomaly score set. The random forest algorithm is used to evaluate the predictive importance of the equipment operation feature set, production environment feature set, and anomaly score set to the component quality indicators. Features with importance scores greater than a preset threshold are selected to obtain a key feature set as model input.

4. The method according to claim 1, characterized in that, The step of dynamically adjusting the equipment operating parameters based on the predicted values ​​through a feedback control mechanism includes: The deviation between the predicted component quality value output by the prediction model and the target quality index is calculated. The deviation value is the weighted sum of the root mean square error and the absolute error, and a comprehensive deviation index is obtained. The comprehensive deviation index is input into the PID controller to generate dynamic adjustment instructions for the equipment operating parameters, including the adjustment range of each parameter, thus obtaining a parameter adjustment instruction set; The parameter adjustment instruction set is sent to the device controller through the industrial IoT platform, and the device operating parameters are controlled in real time within the preset safety threshold to obtain a set of real-time optimized device operating parameters.

5. The method according to claim 1, characterized in that, The process of using a differential evolution algorithm to perform multivariate global optimization on the real-time optimized set of equipment operating parameters and production environment data includes: Define the range of variables for equipment operating parameters and production environment parameters as constraints, construct an objective function with the goal of minimizing quality prediction error and energy consumption, and obtain the definition of the optimization problem; The optimization problem is defined by performing mutation and crossover operations to generate an experimental vector, which is then crossed with the parent vector using a binomial method. Individuals with better fitness are selected to enter the next generation of the population, resulting in an iteratively optimized population. When the objective function value of the iterative optimization population changes by less than a threshold within a preset number of iterations, the current optimal solution is output, and the optimal combination of production parameters is obtained.

6. The method according to claim 1, characterized in that, The process of tracking and verifying the stable production process after parameter optimization using statistical process control methods includes: Collect key quality index data for consecutive preset batches, draw X-bar-R control charts, and obtain the results of production process stability analysis; The process capability index C is calculated based on the stability analysis results of the production process. pk The effectiveness of the optimization is determined by combining the defect rate improvement rate. If the preset threshold is not reached, the differential evolution algorithm optimization process is re-triggered to obtain the final verification conclusion.

7. The method according to claim 1, characterized in that, The method also includes equipment maintenance optimization, specifically including: By analyzing equipment operation data in real time through LSTM neural networks, potential equipment failures can be predicted and maintenance warnings can be generated to obtain equipment health status assessment results. The equipment health status assessment results are combined with the probability of equipment failure, the degree of impact of the equipment on the production line, and the weight of equipment importance, and then input into the maintenance priority scoring model to obtain the maintenance priority score. Based on the maintenance priority score, the maintenance plan is dynamically adjusted using an intelligent digital platform, resulting in an optimization effect of reducing unplanned downtime.

8. The method according to claim 7, characterized in that, The generation of the maintenance priority scoring model includes: The probability of equipment failure is statistically analyzed by using historical data, and the failure probability level is divided according to the equipment type to obtain the failure probability level; The Analytic Hierarchy Process (AHP) was used to quantify the loss of production efficiency, maintenance costs, and quality risks caused by equipment downtime, and to obtain a score of the degree of equipment impact. Based on the functional positioning of the equipment in the production process, an importance weight is assigned to the equipment to obtain the weight allocation result; The failure probability level, equipment impact score, and weight allocation results are input into the weighted scoring model to generate a maintenance priority score. By selecting devices with maintenance priority scores higher than a preset threshold, a priority maintenance list is generated and an alert is triggered, resulting in an optimized maintenance execution plan.

9. An artificial intelligence-based automotive parts production control system, used to implement the method as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition and preprocessing module is used to preprocess the real-time equipment operation data, production environment data and component quality data collected by IoT devices to obtain preprocessed standardized data. The feature extraction and selection module is used to extract and select features from the preprocessed standardized data to obtain equipment operation features and production environment features associated with component quality data. The dynamic quality prediction module is used to construct a prediction model by combining a fully connected neural network with the Pontryagin algorithm, and to train and optimize the prediction model based on the equipment operation characteristics and production environment characteristics to obtain a trained dynamic quality prediction model. The real-time feedback control module is used to input the preprocessed data collected in real time into the dynamic quality prediction model for prediction, to obtain the predicted quality value of the parts in the current batch, and to dynamically adjust the equipment operating parameters according to the predicted value through the feedback control mechanism to obtain a set of real-time optimized equipment operating parameters. The production optimization module is used to perform multivariate global optimization on the real-time optimized set of equipment operating parameters and production environment data using the differential evolution algorithm to obtain the optimal combination of production parameters. The optimal combination of production parameters is then applied to the actual production line to obtain a stable production process after parameter optimization. The production tracking and verification module is used to track and verify the stable production process after parameter optimization through statistical process control methods, so as to obtain the results of improved production stability and data on key quality indicators.

Images