Intelligent control system and method for sintering temperature of powder metallurgy blank

Abstract

The application discloses a kind of powder metallurgy blank sintering temperature intelligent control system and method, belong to powder metallurgy technical field, including data acquisition unit, data processing center, data storage unit, temperature regulating execution mechanism, data display unit and data communication unit.The temperature prediction model constructed by data processing center is composed of three core modules: topological perception multi-scale feature extraction module, chaotic dynamics time series dependence identification module and multi-source information fusion dynamic attention module, topological analysis preserves the complex correlation between parameters, chaotic theory captures the nonlinear dynamic characteristics of temperature, multi-dimensional dynamic attention mechanism pays attention to key information, realizes the high-precision prediction and control of sintering temperature, the system has self-adaptive learning ability, and can automatically adjust the control strategy according to different materials and process conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of powder metallurgy, in particular to a powder metallurgy blank sintering temperature intelligent control system and method, which is particularly suitable for accurately controlling the temperature of a powder metallurgy blank during sintering. BACKGROUND

[0002] Powder metallurgy is a process technology for manufacturing parts by shaping and sintering metal or alloy powders. In the powder metallurgy production process, accurate control of sintering temperature has a decisive influence on product quality. Traditional sintering temperature control mainly relies on PID controllers, which adjust heating power by setting the deviation between actual temperature and actual temperature, but this method is difficult to cope with the complex nonlinear characteristics and multivariable coupling problems in the sintering process.

[0003] With the development of Industry 4.0, intelligent manufacturing technology has been gradually applied to traditional manufacturing fields. Currently, some research has applied neural networks, fuzzy logic and other methods to sintering temperature control, but these methods still have the following shortcomings: first, they lack effective integration capabilities for multi-source heterogeneous data in the sintering process; second, they are difficult to capture complex time-dependent relationships in temperature changes; third, they cannot adapt to temperature control requirements under different materials and different process conditions. Therefore, it is of great practical significance to develop an intelligent system that can accurately control the sintering temperature of a powder metallurgy blank. SUMMARY

[0004] The purpose of the present application is to provide a powder metallurgy blank sintering temperature intelligent control system and method, which aims to solve the problems of insufficient sintering temperature control accuracy, poor adaptability, and low energy utilization in the prior art.

[0005] The present application proposes a powder metallurgy blank sintering temperature intelligent control system, which comprises:

[0006] A data acquisition unit is used to acquire multi-source heterogeneous data inside the reaction furnace in real time, the multi-source heterogeneous data including gas temperature value, gas flow value, reaction furnace electric heating data, reaction furnace temperature adjustment mechanism action value, blank temperature value, reaction furnace internal temperature value, and gas concentration value;

[0007] A data processing center is connected to the data acquisition unit, used to preprocess and screen the multi-source heterogeneous data, and to construct a temperature prediction model based on the preprocessed data, the temperature prediction model including a topological perception multi-scale feature extraction module, a chaotic dynamics time-dependent identification module, and a multi-source information fusion dynamic attention module;

[0008] A data storage unit is connected to the data processing center, used to store the multi-source heterogeneous data and training sample data of the temperature prediction model;

[0009] a temperature adjustment actuator connected with the data processing center, configured to adjust the temperature of the reaction furnace according to the output result of the temperature prediction model;

[0010] a data display unit connected with the data processing center, configured to visually display the temperature prediction result and the control process data;

[0011] a data communication unit configured to realize data interaction among the data acquisition unit, the data processing center, the data storage unit, the temperature adjustment actuator and the data display unit.

[0012] Preferably, the temperature adjustment actuator comprises a heating ring, a furnace door temperature detector and a temperature adjustment controller, the temperature adjustment controller receives the heating data of the heating ring and the in-furnace temperature data of the reaction furnace respectively, and performs temperature regulation according to the received data, the temperature regulation comprises calculation and application of a first temperature correction amount and a second temperature correction amount, and the second temperature correction amount takes effect when the first temperature correction amount approaches zero.

[0013] Preferably, the topologically-aware multi-scale feature extraction module comprises:

[0014] a feature space construction unit configured to construct the multi-source heterogeneous data into a high-dimensional feature space;

[0015] a topological relationship analysis unit configured to calculate the topological relationship strength between different parameters in the high-dimensional feature space;

[0016] a multi-scale feature extraction unit configured to hierarchically decompose and extract features based on the topological relationship strength;

[0017] a feature dimension reduction and retention unit configured to reduce the feature dimension while retaining key topological structure information.

[0018] Preferably, the chaotic dynamics time series dependence identification module comprises:

[0019] a time series data segmentation unit configured to divide the temperature time series data into three time windows of short-term, medium-term and long-term;

[0020] a phase space mapping unit configured to map the time series data of each time window to a phase space;

[0021] a system stability analysis unit configured to calculate system invariants in the phase space and evaluate system stability;

[0022] a turning point identification unit configured to identify key turning points and steady-state regions of temperature change;

[0023] A time-dependent network construction unit is configured to establish a dependent network representing temperature variation rules at different time scales.

[0024] As a preferred, the multi-source information fusion dynamic attention module comprises:

[0025] An information representation unification unit is configured to convert data of different sources and forms into a unified representation form.

[0026] A multi-level fusion unit is configured to fuse information at three levels of parameter level, parameter group level and global level.

[0027] A dynamic attention calculation unit is configured to calculate attention weights in time dimension, feature dimension and space dimension based on the current system state.

[0028] An uncertainty quantification unit is configured to construct a probabilistic graph network to quantify the uncertainty of the prediction result and generate a confidence interval.

[0029] As a preferred, the data processing center further comprises:

[0030] A data preprocessing module is configured to perform data cleaning, data conversion and data normalization on the multi-source heterogeneous data.

[0031] An abnormal data processing and screening module is configured to detect and remove abnormal data.

[0032] A training data extraction module is configured to extract data related to temperature regulation from the preprocessed data as a training data set.

[0033] A training data sample construction module is configured to construct a training sample comprising input data and output data.

[0034] A data model training module is configured to train the temperature prediction model based on the training sample.

[0035] A data model testing and evaluation module is configured to test and evaluate the temperature prediction model to obtain an optimal model.

[0036] As a preferred, the data preprocessing module comprises rules for data cleaning.

[0037] For missing values, the missing values are filled using the average value or median value of the type of missing data according to the type of missing object.

[0038] For abnormal samples, the detected abnormal samples are deleted.

[0039] For abnormal values, the abnormal values are estimated using the numbers within the range of three-quarters quantile to one-quarter quantile in the group where the object is located.

[0040] As preferred, the abnormal data processing and screening module detects the abnormal data by constructing an isolated forest model, including:

[0041] Preliminary screening of abnormal data is performed using all feature data after dimensionless processing;

[0042] The data after preliminary screening is clustered;

[0043] An isolated forest model is constructed according to the clustering results;

[0044] The isolated forest model is used to further screen abnormal data, and non-abnormal data is obtained as a training data set.

[0045] As preferred, the input data vector constructed by the training data sample construction module includes the reaction furnace internal temperature value, the blank temperature value, the heating ring temperature value, the input furnace gas temperature value, the input furnace gas concentration value, the gas flow value and the information whether there is artificial intervention, and the output data includes the reaction furnace door temperature value, the output furnace gas temperature value and the output furnace gas concentration value.

[0046] A control method using the powder metallurgy blank sintering temperature intelligent control system, comprising the following steps:

[0047] (1) The data acquisition unit acquires real-time multi-source heterogeneous data inside the reaction furnace, and transmits the multi-source heterogeneous data to the data processing center;

[0048] (2) The data processing center pre-processes the multi-source heterogeneous data, including data cleaning, data conversion and data normalization;

[0049] (3) The data processing center uses an abnormal data processing and screening module to detect and screen the pre-processed data, and obtains non-abnormal data as a training data set;

[0050] (4) The data processing center extracts data related to temperature regulation from the training data set, and constructs a training sample containing input data vectors and output data;

[0051] (5) The data processing center trains a temperature prediction model using the training sample, and the temperature prediction model includes a topological perception multi-scale feature extraction module, a chaotic dynamics time series dependence identification module and a multi-source information fusion dynamic attention module;

[0052] (6) The topological perception multi-scale feature extraction module extracts multi-scale features that retain topological relationships from the multi-source heterogeneous data;

[0053] (7) The chaotic dynamics time sequence dependence identification module analyzes the nonlinear dynamic characteristics of the temperature time sequence data and identifies the key turning points of temperature change;

[0054] (8) The multi-source information fusion dynamic attention module integrates features from different sources and focuses on key information through a multi-dimensional attention mechanism;

[0055] (9) The data processing center predicts the future temperature under the current state based on the temperature prediction model;

[0056] (10) The temperature adjustment actuator calculates the temperature correction amount and correction curve based on the prediction result;

[0057] (11) The temperature adjustment actuator adjusts the reaction furnace temperature according to the temperature correction amount and correction curve;

[0058] (12) The data display unit displays the prediction result and control process data in real time.

[0059] The present application fuses topology, chaos theory and deep learning technology, and constructs a deep neural network temperature prediction model based on multi-dimensional tensor analysis, realizing high-precision prediction and control of sintering temperature. The system has the following beneficial effects:

[0060] 1. Significantly improve temperature control accuracy, reduce control error from ±5-8℃ of traditional system to ±2℃ or less, reduce temperature fluctuation amplitude by about 65%, improve product quality consistency;

[0061] 2. By accurately identifying the temperature change rule and key turning point in the sintering process, the system can predict and adjust the control strategy in advance, reduce the phenomenon of temperature over-regulation and under-regulation, and improve the control response speed by about 40%;

[0062] 3. The optimized temperature control strategy significantly reduces energy consumption, sintering process energy consumption reduces by about 22%, gas consumption reduces by about 18%, improves production economy;

[0063] 4. The system has self-adaptive learning ability, can automatically adjust the control strategy according to different materials and different process conditions, significantly improves the universality and adaptability of the system;

[0064] 5. Through effective integration and utilization of multi-source heterogeneous data in the sintering process, the system provides comprehensive process monitoring and quality traceability capability, which helps process improvement and quality control. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is the overall structure framework diagram of the powder metallurgy blank sintering temperature intelligent control system of the present application;

[0066] Figure 2 Structure diagram of functional modules of the data processing center of the present application;

[0067] Figure 3 Structure diagram of the topological perception multi-scale feature extraction module of the present application;

[0068] Figure 4 Structure diagram of the chaotic dynamics time series dependence identification module of the present application;

[0069] Figure 5 Structure diagram of the multi-source information fusion dynamic attention module of the present application;

[0070] Figure 6 Structure diagram of the temperature regulation actuator of the present application;

[0071] Figure 7 Flow chart of the control method of the present application. DETAILED DESCRIPTION

[0072] The present application will be further described in detail below with reference to the accompanying drawings and specific examples. Those skilled in the art should understand that these examples are only used to illustrate the present application and do not limit the scope of the present application. Figure 1 - Figure 7 , and specific examples. Those skilled in the art should understand that these examples are only used to illustrate the present application and do not limit the scope of the present application.

[0073] Referring to Figure 1 The present application provides a powder metallurgy blank sintering temperature intelligent control system, which comprises a data acquisition unit 1, a data processing center 2, a data storage unit 3, a temperature regulation actuator 4, a data display unit 5 and a data communication unit 6.

[0074] The data acquisition unit 1 is used to acquire multi-source heterogeneous data in the reaction furnace in real time, wherein the multi-source heterogeneous data includes gas temperature value, gas flow value, reaction furnace electric heating data, action value of the reaction furnace temperature regulation mechanism, blank temperature value, reaction furnace temperature value, gas concentration value and other key parameters. Preferably, these data are collected in the form of discrete time series, and the sampling frequency is set to 5 times per second, which can fully capture the temperature change characteristics in the sintering process, while not producing too much redundant data. In actual application, the sampling frequency can be adjusted according to different sintering processes. For the rapid heating stage, it can be increased to 10 times per second; and for the constant temperature stage, it can be reduced to 2 times per second, to realize the optimal allocation of resources.

[0075] The data processing center 2 is connected with the data acquisition unit 1 through a high-speed data bus. The data processing center 2 is mainly responsible for pre-processing and screening of the multi-source heterogeneous data, and constructing a temperature prediction model based on the pre-processed data. The temperature prediction model includes three key modules: a topological perception multi-scale feature extraction module 21, a chaotic dynamics time series dependence identification module 22, and a multi-source information fusion dynamic attention module 23. The three modules cooperate with each other to realize accurate prediction of the sintering temperature change.

[0076] The data storage unit 3 is connected with the data processing center 2, and is used for storing multi-source heterogeneous data and training sample data of the temperature prediction model. In addition, the unit also stores historical sintering data, model parameters and system configuration information, providing data support for model training and system optimization. In actual sintering production line application, a typical powder metallurgy sintering system can generate about 20GB of raw data per day, and after preliminary compression and screening, about 5GB needs to be stored for a long time, so the data storage unit 3 adopts a layered storage strategy: hot data (recent data) is stored in high-speed storage medium, and cold data (historical data) is stored in large-capacity storage medium, which not only guarantees system performance, but also takes into account cost-effectiveness.

[0077] The temperature regulation execution mechanism 4 is connected with the data processing center 2 through an industrial control network, and is used for adjusting the temperature of the reaction furnace according to the output result of the temperature prediction model. In a typical powder metallurgy sintering application, the response time of temperature regulation is not more than 2 seconds, and the control accuracy can reach within ±2℃, meeting the process requirements of most powder metallurgy products.

[0078] The data display unit 5 is connected with the data processing center 2 through the data communication unit 6, and is used for visualizing and displaying the temperature prediction results and control process data. The unit provides an intuitive graphical interface, which is convenient for operators to monitor and analyze the sintering process. In actual application, the operator can real-time view the temperature curve, the comparison between the predicted value and the actual value, the change trend of key parameters and other information through the data display unit 5, providing decision support for process adjustment and quality control.

[0079] The data communication unit 6 is the connecting link of the whole system, and is used for realizing data interaction between the data acquisition unit 1, the data processing center 2, the data storage unit 3, the temperature regulation execution mechanism 4 and the data display unit 5. The unit supports multiple industrial communication protocols such as Modbus, Profibus, OPCUA, etc., ensuring seamless connection and efficient communication between the parts of the system. The data transmission rate is usually above 10Mbps, meeting the needs of real-time control of the system.

[0080] Reference Figure 6The embodiment details the structure and working principle of the temperature adjustment actuator 4. The temperature adjustment actuator 4 includes a heating coil 41, a furnace door temperature detector 42, and a temperature adjustment controller 43. The temperature adjustment controller 43 receives the heating data of the heating coil 41 and the in-furnace temperature data of the reaction furnace, and performs temperature regulation based on the received data.

[0081] The temperature regulation process includes two key steps: calculation and application of the first temperature correction and the second temperature correction. In the powder metallurgy sintering process, both accuracy and stability need to be considered in temperature control. The system first calculates the first temperature correction based on the current temperature error and the trend of change, and applies it to temperature adjustment; then, the system continues to monitor the temperature change, calculates the second temperature correction, and when the first temperature correction approaches zero (usually defined as an absolute value less than 0.5℃), the second temperature correction takes effect, further fine-tuning the temperature.

[0082] In a typical iron-based powder compact sintering process, the sintering temperature needs to be raised from room temperature to 1120℃ and maintained for 2 hours before slow cooling. In the heating stage, the first temperature correction mainly addresses large-scale temperature adjustment, with a correction amount of up to 10-15℃; while in the constant temperature stage, the second temperature correction plays a major role, with a correction amount usually between 0.5-2℃, ensuring that the temperature fluctuates around the target value. This two-stage regulation strategy significantly improves the accuracy and stability of temperature control, making the sintering process more controllable and the product quality more consistent.

[0083] Referring to Figure 3 The topologically-aware multi-scale feature extraction module 21 includes a feature space construction unit 211, a topological relationship analysis unit 212, a multi-scale feature extraction unit 213, and a feature dimension reduction and preservation unit 214.

[0084] The feature space construction unit 211 is used to construct multi-source heterogeneous data into a high-dimensional feature space. In specific implementation, this unit first normalizes different types of data to eliminate dimensional differences, and then maps them to a unified high-dimensional feature space. For example, for temperature data, flow data, and gas concentration data, different normalization methods are used:

[0085] 1. Temperature data uses maximum-minimum value normalization:

[0086] ,

[0087] where, is the original temperature value, is the normalized temperature value, and are the minimum and maximum values of the temperature, respectively. In copper-based powder metallurgy compact sintering, Usually set to room temperature 25 °C, Set to the highest sintering temperature 850 °C; while for iron-based powder metallurgy blanks, Also room temperature 25 °C, May be as high as 1200 °C.

[0088] 2. Flow data is standardized using Z-score:

[0089] ,

[0090] Where, is the original flow value, is the standardized flow value, is the mean value of the flow, is the standard deviation of the flow. In practical applications, the mean value of the gas flow and the standard deviation are usually determined based on production experience, for example, the of the hydrogen flow is usually 5 L / min, about 0.5 L / min; while the of the nitrogen flow can be as high as 20 L / min, about 2 L / min.

[0091] The topological relationship analysis unit 212 is used to calculate the topological relationship strength between different parameters in the high-dimensional feature space. This unit uses a neighborhood information preservation matrix to record the topological relationship between parameters, and the matrix element represents the association strength between parameter i and parameter j. The association strength is calculated by:

[0092] ,

[0093] Where, is the association strength between parameter and parameter , represents the Euclidean distance between parameter and , and are local scale parameters used to adjust the decay rate of the association strength. In practical applications, for parameters that are physically closely related, such as furnace temperature and furnace door temperature, and are usually set small (about 0.1-0.2), so that only very close values have high association and for indirectly related parameters, such as temperature and gas concentration, and are set larger (about 0.5-0.8), allowing more relaxed association determination.

[0094] The multi-scale feature extraction unit 213 is used for hierarchical decomposition and extraction of features based on topological relationship strength. In powder metallurgy green body sintering temperature control, this unit realizes three-layer feature extraction:

[0095] 1. Parameter-level single feature extraction: Extract the characteristics of each parameter itself, such as temperature change rate, fluctuation amplitude, etc. For example, temperature change rate is calculated as follows:

[0096] ,

[0097] where, is the temperature value at time t, is the temperature value at time t, is the time interval, usually set to 1 second.

[0098] Parameter combination feature extraction: Extract the interaction features of related parameter groups. For example, temperature-flow coupling feature

[0099] ,

[0100] where, is the normalized temperature value, is the normalized flow value, , and are weight coefficients, determined by data analysis, usually in the early stage of sintering, and adjusted to in the later stage of sintering. This dynamic adjustment reflects the changes in the importance of various parameters in different stages.

[0101] 3. System-level global feature extraction: Extract macro features reflecting the state of the entire sintering system, such as system energy state :

[0102] ,

[0103] where, is the normalized value of the th parameter, is the weight coefficient, is the total number of parameters. The weight of each parameter is determined according to its influence on the system state, for example, the weight of the temperature parameter is usually 0.4-0.5, the weight of the flow parameter is 0.2-0.3, and the weight of the concentration parameter is 0.2-0.25.

[0104] ​The feature dimension reduction and retention unit 214 is used to retain key topological structure information while reducing the feature dimension. In a large sintering furnace control system, the original features can be as many as dozens or even hundreds of dimensions, and direct use of these features will result in excessively high computational complexity, so dimension reduction processing is needed. This unit uses the local linear embedding (LLE) algorithm, which can maintain the local geometric relationship between data points. For a sintering control system containing more than 50 original features, the number of features can be reduced to 15-20 through dimension reduction, while retaining more than 90% of the information, greatly improving the computational efficiency.

[0105] Referring to Figure 4 , the chaotic dynamics time series dependence identification module 22 includes a time series data segmentation unit 221, a phase space mapping unit 222, a system stability analysis unit 223, a turning point identification unit 224, and a time series dependence network construction unit 225.

[0106] The time series data segmentation unit 221 is used to divide the temperature time series data into short-term, medium-term, and long-term three time windows. In powder metallurgy sintering process control, temperature changes have the characteristics of multiple time scales, so multi-scale analysis is needed. The short-term window is set to 30 seconds, mainly capturing rapid temperature fluctuations such as temperature fluctuations caused by heating element switching; the medium-term window is set to 5 minutes, used to identify temperature change trends such as acceleration or deceleration in the heating stage; the long-term window is set to 30 minutes, used to analyze overall sintering cycle characteristics such as preheating, constant temperature, and cooling stages. This multi-scale time window design enables the system to fully capture temperature change characteristics at different time scales.

[0107] The phase space mapping unit 222 is used to map the time series data of each time window to the phase space. This unit uses the time delay coordinate method to construct the phase space, and the specific method is:

[0108] ,

[0109] where, is the phase space vector at time t, is the temperature value at time t, is the time delay parameter, is the embedding dimension. In copper-based powder compact sintering temperature control, is usually set to 2 seconds (corresponding to 10 sampling points), is set to 4; while for more complex tungsten-based powder compact sintering, may need to be increased to 3-4 seconds, to 5-6, to capture more complex dynamic characteristics. These parameters are determined by optimizing the mutual information function and the false nearest neighbor method, which can maximize the dynamic characteristics of the system.

[0110] The system stability analysis unit 223 is used to calculate the system invariants in the phase space and assess the system stability. The key invariants calculated by this unit include:

[0111] 1. Maximum Lyapunov Exponent (MLE):

[0112] ,

[0113] where, MLE is the maximum Lyapunov exponent, represents the separation distance of two initial phase close trajectories in the phase space at time t, represents the norm of the separation distance, represents the norm of the initial separation distance. In the powder metallurgy sintering process, MLE>0 indicates that the system has chaotic characteristics, and the temperature change is difficult to predict in the long term, usually appearing in the rapid heating stage at the beginning of sintering; MLE<0 indicates that the system is stable, and the temperature change is predictable, usually appearing in the constant temperature stage. In practical applications, for the case of MLE>0.01, the prediction time domain of the system is shortened, and the control frequency is increased; while for the case of MLE<0.005, the prediction time domain can be extended, and the control frequency can be reduced, optimizing system resources.

[0114] 2. Correlation Dimension:

[0115] ,

[0116] where, D is the correlation dimension, is the correlation integral, representing the ratio of points in the phase space with a distance less than is the distance threshold. The higher the D, the higher the complexity of the system. In the iron-based powder sintering, the D of the preheating stage is usually between 2.5-3.5, indicating that the system has high complexity; while the D of the constant temperature stage drops to 1.5-2.0, indicating that the system tends to be stable.

[0117] The turning point identification unit 224 is used to identify the key turning points and stable regions of temperature change. This unit identifies the inflection points, intersection points and stable regions of the trajectory by analyzing the geometric characteristics of the phase space trajectory. In the powder metallurgy sintering process, the turning points usually correspond to the changes of process stages, such as the end point of preheating, the start point of constant temperature, etc. Identifying these key points helps to optimize the control strategy and improve the temperature control accuracy. For example, when transitioning from the preheating stage to the constant temperature stage, the system will adjust the control parameters 3-5 minutes in advance to slow down the heating rate and avoid temperature overshoot. This strategy reduces the temperature overshoot amplitude from 5-8℃ in traditional control to 1-2℃. ​​​

[0118] The time-dependent network construction unit 225 is used to establish a dependency network representing the regularity of temperature changes at different time scales. This network is a directed and weighted graph, where nodes represent states at different time points, and edge weights represent state transition probabilities. By analyzing the structural properties of this network, the system can identify typical patterns and regularities of temperature changes, providing a basis for prediction and control. In copper-based powder sintering, the system identifies four typical temperature change patterns: rapid heating, slow heating, constant temperature fluctuation, and slow cooling, and adopts different control strategies for each pattern, significantly improving control effectiveness.

[0119] Referring to Figure 5 , the dynamic attention module 23 of multi-source information fusion includes an information representation unification unit 231, a multi-level fusion unit 232, a dynamic attention calculation unit 233, and an uncertainty quantification unit 234.

[0120] The information representation unification unit 231 is used to convert data of different sources and forms into a unified representation. In the powder metallurgy sintering temperature control system, data sources are diverse, including temperature sensors, flow meters, gas concentration analyzers, etc. The data formats and units output by these devices are different. This unit realizes the standardized representation of data, ensuring that data from different sources can be processed and analyzed under the same framework. Specifically, for each type of data , it is converted to a standard representation :

[0121] ,

[0122] where is the standardized data representation, is the original data, is the conversion function for data type , including standardization, encoding, and feature extraction operations. For example, for temperature data, includes sliding window averaging, trend extraction, etc.; for gas concentration data, includes gradient calculation, volatility analysis, etc. This unified representation enables the system to effectively integrate information from different sources.

[0123] The multi-level fusion unit 232 is used to fuse information at three levels: parameter level, parameter group level, and global level. This unit implements a hierarchical information fusion strategy, gradually building a complete system representation from a single parameter to the entire system:

[0124] 1. Parameter-level fusion: merging information of the same parameter at different time points, such as calculating the average value, rate of change, and volatility of temperature in the past 5 minutes;

[0125] 2. Parameter group fusion: integrate information of related parameters, such as temperature-flow group, temperature-concentration group, etc. For temperature-flow group, the fusion function can be expressed as:

[0126] ,

[0127] where, is the fusion result, is the temperature feature, is the flow feature, is the temperature-flow interaction feature, , and are the corresponding weight coefficients. In practical applications, usually in the range of 0.4-0.6, in the range of 0.2-0.3, in the range of 0.2-0.3, these weights are dynamically adjusted with the sintering stage.

[0128] 3. Global fusion: integrate information of all parameter groups to form a system-wide state representation. The global representation is calculated as follows:

[0129] ,

[0130] where, is the system-wide state representation, is the fusion result of the th parameter group, is the corresponding weight coefficient, is the total number of parameter groups. is determined by a dynamic attention mechanism, reflecting the importance of different parameter groups in the current state.

[0131] The dynamic attention calculation unit 233 is used to calculate the attention weights of the time dimension, feature dimension and spatial dimension based on the current system state. In powder metallurgy sintering temperature control, the importance of data at different time points, different parameters and different locations is not the same, and the focus needs to be dynamically adjusted. The attention weight is calculated as follows:

[0132] ,

[0133] where, is the attention weight vector of the time dimension, is the weight matrix, is the current hidden state, is the current context vector, and softmax is the normalization function to ensure the sum of all weights is 1. Similarly, the feature dimension attention And spatial dimension attention are also calculated by corresponding weight matrices.

[0134] In practical applications, the attention weights reflect the importance of different information. For example, in the initial rapid heating stage of aluminum-based powder sintering, the attention weights of gas flow and concentration are higher (0.4-0.6) because these parameters directly affect the heating rate; in the middle constant temperature stage, the attention weight of temperature change rate rises (0.5-0.7) because stable temperature becomes more critical; in the late slow cooling stage, the attention weight of the billet temperature dominates (usually higher than 0.6) because it is directly related to product quality.

[0135] The uncertainty quantification unit 234 is used to construct a probabilistic graph network to quantify the uncertainty of the prediction results and generate a confidence interval. In a complex sintering process, there is uncertainty in temperature prediction, and understanding this uncertainty is crucial for control decisions. The uncertainty representation adopts a Bayesian neural network method to obtain the prediction distribution through Monte Carlo sampling:

[0136] ,

[0137] where, is the probability distribution of the prediction result under a given input , is the th sample of the neural network parameters, is the number of samples. In practical applications, is usually set to 100, and in cases where computing resources are limited, it can be reduced to 50, and a good uncertainty estimate can still be obtained. The confidence interval of the prediction is obtained from the prediction distribution by the quantile method:

[0138]

[0139] ,

[0140] where, is the 95% confidence interval, and are the 2.5% and 97.5% quantiles of the prediction distribution, respectively. In practical sintering control, the system will dynamically adjust the control strategy according to the width of the confidence interval: when the confidence interval is narrow (such as within ±3°C), a more aggressive control strategy is adopted; when the confidence interval is wide (such as more than ±5°C), a more conservative control strategy is adopted to avoid temperature fluctuations caused by excessive adjustments.

[0141] Referring to Figure 2 ​​​​The data processing center 2 further comprises a data preprocessing module 24, an abnormal data processing and screening module 25, a training data extraction module 26, a training data sample construction module 27, a data model training module 28, and a data model testing and evaluation module 29.

[0142] The data preprocessing module 24 is used for data cleaning, data conversion and data normalization of multi-source heterogeneous data. In the powder metallurgy sintering temperature control system, the original data often has problems such as noise, missing and inconsistency, which need to be preprocessed to improve the data quality. The data cleaning process removes data anomalies caused by sensor failure, electrical interference and other factors; data conversion converts different formats of original data into a unified format for subsequent processing; data normalization eliminates the scale difference between different data through dimensionless processing to improve the effect of model training.

[0143] The abnormal data processing and screening module 25 is used for detecting and removing abnormal data. In the sintering process control, factors such as sensor failure and human interference may cause data anomalies, which will affect the accuracy of model training and prediction. This module identifies abnormal points in the data through statistical methods and machine learning techniques to improve data quality.

[0144] The training data extraction module 26 is used to extract data related to temperature regulation from preprocessed data as a training data set. In powder metallurgy sintering temperature control, not all parameters are directly related to temperature prediction, so it is necessary to selectively extract relevant parameters to build a targeted training data set. For example, for stainless steel powder compact sintering, furnace temperature, gas flow and oxygen concentration are key parameters that need to be extracted; while for aluminum-based powder compact, additional attention is needed for hydrogen concentration and heating power.

[0145] The training data sample construction module 27 is used to construct training samples including input data and output data. In practical applications, this module organizes input parameters and output parameters into training samples according to time relationship, ensuring that the model can learn the mapping relationship between input and output.

[0146] The data model training module 28 is used to train the temperature prediction model based on the training samples. This module implements the training process of the model, including parameter optimization, loss function calculation, etc. In model training, a batch training strategy is adopted, each batch contains 64-128 samples, and the learning rate is dynamically adjusted during training to improve training efficiency and model performance.

[0147] The data model testing and evaluation module 29 is used to test and evaluate the temperature prediction model, so as to obtain the optimal model. The module comprehensively evaluates the model performance through various evaluation indicators such as mean absolute error (MAE), root mean square error (RMSE) and correlation coefficient (R²) and the like. In actual application, an excellent temperature prediction model usually has an MAE of less than 3°C, an RMSE of less than 4°C and an R² of greater than 0.95 on the test set, meeting the requirements of sintering temperature control.

[0148] In the powder metallurgy sintering temperature control system, data quality is crucial to model performance. The present application adopts three different cleaning strategies to handle different types of data problems:

[0149] For missing values, according to the type of missing object, the missing data average value or median value of the type is used to fill in the missing values. In actual application, temperature data is usually smooth, and missing values can be filled with the average value of the previous and next 10 time points (about 2 seconds of data); while for the gas flow data with greater volatility, the median value of the same type data within the last 30 minutes is used for filling, which can avoid the influence of extreme values and provide more reasonable estimation.

[0150] For abnormal samples, the detected abnormal samples are deleted. In the powder metallurgy sintering process, abnormal samples usually show sudden jumps in temperature or gas parameters, which are often caused by sensor failure or external interference, rather than real process changes. The system identifies abnormal samples by setting reasonable thresholds, such as temperature change rate exceeding 50°C / s (much higher than the maximum change rate of 10°C / s in normal sintering process), gas flow instantaneous change exceeding 200%, and the like. These thresholds are determined based on a large amount of production data analysis and expert experience, and can effectively identify abnormal situations.

[0151] For outliers, the number within the range of three-quarters to one-quarter of the object's group is used to estimate the outliers. This method is based on the box plot principle of statistics, which can effectively handle outliers in the data while maintaining the overall distribution characteristics of the data. In the powder metallurgy sintering temperature control, the correction error of the quartile range method for temperature data is usually controlled within ±3°C, meeting the requirements of the system for data quality.

[0152] Example 8

[0153] This embodiment details the working principle of the abnormal data processing and screening module 25. The module uses the Isolation Forest algorithm for anomaly detection, which is a unsupervised learning method particularly suitable for handling outliers in high-dimensional data.

[0154] The specific process of abnormal data detection is as follows:

[0155] First, all the feature data after dimensionless processing is used to preliminarily screen abnormal data. The dimensionless processing adopts the maximum-minimum value normalization and Z-score standardization method to ensure that different features are compared in the same scale. The preliminary screening adopts the 3-sigma rule, that is, the data points deviating from the mean value by more than 3 standard deviations are marked as potential anomalies. For powder metallurgy sintering data, this step can usually screen out about 5%-10% of potential abnormal points, greatly reducing the data amount for subsequent processing.

[0156] Second, clustering analysis is performed on the data after preliminary screening. In the powder metallurgy sintering process, the data characteristics of different stages are significantly different, so it is necessary to detect anomalies in stages. The clustering analysis adopts the K-means algorithm to divide the data into 3-5 categories, representing different working stages or states. For example, for standard iron-based powder blank sintering, the data is usually divided into four categories: preheating stage (room temperature to 600°C), rapid heating stage (600-1000°C), constant temperature stage (1000-1120°C) and cooling stage.

[0157] Then, an isolation forest model is constructed according to the clustering results. The core idea of the isolation forest algorithm is that abnormal points are more likely to be isolated, so the difficulty of isolating data points can be measured by constructing a random tree to detect anomalies. For each cluster, the system trains a dedicated isolation forest model to improve the accuracy of anomaly detection. The key parameters of the isolation forest model are set as follows:

[0158] Number of trees: 100-200, enough to ensure detection effect, and will not cause excessive computational burden

[0159] Subsample size: 60% of the original data, balancing sampling representativeness and randomness;

[0160] Maximum tree depth: usually set to (subsample size) to prevent overfitting;

[0161] Finally, the isolation forest model is used to further screen abnormal data, and the non-abnormal data is obtained as the training data set. The abnormality judgment threshold is usually set to the 95th percentile of the model output score, which means that about 5% of the data will be marked as abnormal and excluded from the training set. In the powder metallurgy sintering process, this proportion is the balance point verified by a large number of experiments, which can effectively remove abnormal data while retaining enough training samples.

[0162] This embodiment details the specific implementation of the training data sample construction module 27. In the powder metallurgy blank sintering temperature control system, the design of the training sample directly affects the learning effect and prediction performance of the model.

[0163] The input data vector constructed by the training data sample construction module 27 includes the reaction furnace internal temperature value, the blank temperature value, the heating ring temperature value, the input furnace internal gas temperature value, the input furnace internal gas concentration value, the gas flow value, and the information whether there is artificial intervention. These parameters comprehensively reflect the state of the sintering process, providing rich information input for the prediction model.

[0164] The output data includes the reaction furnace door temperature value, the output furnace internal gas temperature value, and the output furnace internal gas concentration value. These parameters are key indicators for evaluating the sintering effect and are also target parameters that the control system needs to predict and regulate.

[0165] When constructing the training sample, the system needs to consider the time lag relationship between the input parameters and the output parameters. In copper-based powder sintering, this time lag is usually 30-45 seconds; while in iron-based powder sintering, due to larger heat capacity, the time lag can reach 60-90 seconds. In order to capture this time lag relationship, the sample construction uses a sliding window method: using the input data at time t to predict the output data at time t+Δt, where Δt is the time delay determined according to the process characteristics.

[0166] In addition, in order to improve the generalization ability of the model, the training data sample also uses data enhancement techniques such as adding small amplitude random noise (usually ±2% of the original value), adjusting the sampling window size, etc. These techniques effectively improve the model's adaptability to different working conditions, allowing it to maintain good prediction performance when facing production fluctuations.

[0167] The embodiment details the control method of using the powder metallurgy blank sintering temperature intelligent control system of the present application, which includes the following steps:

[0168] (1) The data acquisition unit 1 acquires real-time multi-source heterogeneous data inside the reaction furnace and transmits the multi-source heterogeneous data to the data processing center 2. In actual application, the data acquisition frequency is set to 5Hz, that is, 5 times of data acquisition per second. This sampling frequency is determined based on the temperature change characteristics of the powder metallurgy sintering process: in the rapid heating stage, the temperature change rate can reach 5-10℃ / min, and the 5Hz sampling frequency can capture about 0.02℃ temperature change, meeting the fine control requirements; while in the constant temperature stage, the temperature change rate decreases to 0.5-1℃ / min, and the same sampling frequency provides more abundant data, which helps to identify small fluctuations and trend changes. Data transmission uses industrial Ethernet protocol, communication bandwidth is 10Mbps, transmission delay is less than 10ms, ensuring the real-time performance of data transmission.

[0169] (2) The data processing center 2 preprocesses the multi-source heterogeneous data, including data cleaning, data conversion, and data normalization. Data cleaning handles noise, missing values, and outliers; data conversion converts raw data of different formats into a unified format; and data normalization eliminates scale differences between different data through dimensionless processing. In the powder metallurgy compact sintering control system, the preprocessing link can improve the data quality by about 25% to 30%, laying the foundation for subsequent analysis and modeling.

[0170] (3) The data processing center 2 uses the abnormal data processing and screening module 25 to detect and screen the preprocessed data for abnormalities, obtaining non-abnormal data as the training data set. As previously mentioned, this module uses the Isolation Forest algorithm for anomaly detection, which can identify and exclude about 5% of abnormal data points, significantly improving the quality and representativeness of the data set.

[0171] (4) The data processing center 2 extracts temperature regulation-related data from the training data set and constructs training samples containing input data vectors and output data. In actual applications, the system dynamically adjusts the feature selection strategy for different sintering materials and processes. For example, for powder materials with high carbon content, the system will increase the weight of gas concentration features; for materials prone to oxidation, more attention will be paid to features such as gas flow and temperature change rate.

[0172] (5) The data processing center 2 uses the training samples to train the temperature prediction model, which includes the topologically aware multi-scale feature extraction module 21, the chaotic dynamics time series dependency identification module 22, and the multi-source information fusion dynamic attention module 23. Model training uses the batch gradient descent method, with a batch size of 64, an initial learning rate of 0.001, and a cosine annealing learning rate scheduling strategy, i.e., the learning rate gradually decays from the initial value to a minimum value (usually 1 / 10 of the initial value) according to the cosine function, and then returns to the initial value, repeating the cycle. This strategy can effectively avoid local optima and improve model performance. Training a typical iron-based powder compact sintering temperature prediction model requires about 5000 samples and takes about 30 minutes (on a standard industrial server).

[0173] (6) The topologically aware multi-scale feature extraction module 21 extracts multi-scale features that preserve topological relationships from multi-source heterogeneous data. As previously mentioned, this module preserves the internal structure of the data by constructing a topological relationship network between features, improving the expressiveness of the features. In actual applications, the system dynamically adjusts the level and depth of feature extraction for sintering processes of different complexities. For example, for simple copper-based powder sintering, two-level feature extraction (parameter level and parameter group level) is sufficient; for complex tungsten-cobalt alloy powder sintering, full three-level feature extraction is needed to fully capture the system characteristics.

[0174] (7) The chaotic dynamics time series dependence identification module 22 analyzes the nonlinear dynamic characteristics of the temperature time series data and identifies the key turning points of temperature changes. Through phase space reconstruction and system invariant analysis, this module reveals the internal laws of temperature changes and improves the accuracy of prediction. For different sintering stages, the system adjusts the size and overlap of the analysis window. For example, in the rapid heating stage, the window size is set to 20-30 seconds, and the overlap is 50%; while in the constant temperature stage, the window size can be increased to 60-90 seconds, and the overlap is reduced to 30%. This dynamic adjustment strategy balances the calculation efficiency and analysis accuracy.

[0175] (8) The multi-source information fusion dynamic attention module 23 integrates features from different sources and focuses on key information through a multi-dimensional attention mechanism. As mentioned earlier, this module can automatically adjust the attention to different information and focus on the most relevant parameters and time points according to the current system state. In the powder metallurgy sintering process, the dynamic attention mechanism can adapt to the control requirements of different stages and improve the prediction and control performance.

[0176] (9) The data processing center 2 predicts the future temperature under the current state based on the temperature prediction model. In the standard sintering process, the prediction time range is set to the next 5-10 minutes, and the prediction step is 30 seconds, which achieves a good balance between prediction accuracy and computational complexity. The system not only predicts the temperature value, but also predicts the temperature change trend and possible fluctuation range, providing comprehensive information for control decisions.

[0177] (10) The temperature regulation actuator 4 calculates the temperature correction amount and correction curve based on the prediction results. The calculation of the correction amount considers the current temperature error, temperature change trend, and predicted future temperature change, realizing a feedforward-feedback combined control strategy. This strategy is more forward-looking than traditional pure feedback control, which can respond to temperature changes in advance and reduce control lag.

[0178] (11) The temperature regulation actuator 4 adjusts the reaction furnace temperature according to the temperature correction amount and correction curve. In actual application, temperature regulation adopts a smooth change strategy to avoid drastic action of equipment and large fluctuations in temperature. For example, for the adjustment of heating power, the system limits the maximum adjustment amplitude to no more than 15% of the current power to protect the heating element and reduce temperature overshoot.

[0179] (12) The data display unit 5 displays the prediction results and control process data in real time. The display content includes the current temperature, predicted temperature, control error, key parameter change trend, etc., presented in various chart forms such as line charts, heat maps, and dashboards. These visual information helps operators understand the system state comprehensively, discover potential problems in time, and optimize control strategies.

[0180] The above steps form a complete closed-loop control process, which realizes accurate prediction and control of the sintering temperature of powder metallurgy blanks through deep learning technology and advanced data processing methods. In actual production, this method has been successfully applied to the sintering processes of various powder metallurgy products, such as copper-based self-lubricating bearings, stainless steel filters, and tungsten-cobalt hard alloy cutters, significantly improving product quality and production efficiency. Compared with traditional PID control, this method improves temperature control accuracy by about 60%, reduces energy consumption by about 20%, and increases product pass rate by about 15%, demonstrating significant technical and economic value.

[0181] The above-described embodiments only express specific implementations of the present application, which are described in detail and specifically, but should not be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, which are within the protection scope of the present application.

Claims

1. A powder metallurgy green body sintering temperature intelligent control system, characterized in that, The application relates to a temperature prediction system for a reaction furnace. The system comprises: a data acquisition unit for acquiring multi-source heterogeneous data inside the reaction furnace in real time, wherein the multi-source heterogeneous data comprises gas temperature values, gas flow values, reaction furnace electric heating data, action values of a reaction furnace temperature adjusting mechanism, blank body temperature values, reaction furnace internal temperature values and gas concentration values; a data processing center connected with the data acquisition unit, for pre-processing and screening the multi-source heterogeneous data, and constructing a temperature prediction model based on the pre-processed data, wherein the temperature prediction model comprises a topological perception multi-scale feature extraction module, a chaotic dynamics time series dependence identification module and a multi-source information fusion dynamic attention module; The topological perception multi-scale feature extraction module comprises: a feature space construction unit for constructing the multi-source heterogeneous data into a high-dimensional feature space; a topological relationship analysis unit for calculating the topological relationship strength between different parameters in the high-dimensional feature space; a multi-scale feature extraction unit for hierarchically decomposing and extracting features based on the topological relationship strength; a feature dimension reduction and retention unit for reducing the feature dimension while retaining key topological structure information; a data storage unit connected with the data processing center, for storing the multi-source heterogeneous data and training sample data of the temperature prediction model; a temperature adjusting execution mechanism connected with the data processing center, for adjusting the temperature of the reaction furnace according to the output result of the temperature prediction model; a data display unit connected with the data processing center, for visually displaying the temperature prediction result and control process data; a data communication unit for realizing data interaction between the data acquisition unit, the data processing center, the data storage unit, the temperature adjusting execution mechanism and the data display unit; The chaotic dynamics time series dependence identification module comprises: a time series data segmentation unit for dividing temperature time series data into three time windows of short-term, medium-term and long-term; a phase space mapping unit for mapping time series data of each time window to a phase space; a system stability analysis unit for calculating system invariants in the phase space and evaluating system stability; a turning point identification unit for identifying key turning points and steady-state regions of temperature change; a time series dependence network construction unit for establishing a dependence network representing temperature change rules of different time scales; The multi-source information fusion dynamic attention module comprises: an information representation unification unit for converting data of different sources and forms into a unified representation form; a multi-level fusion unit for fusing information at three levels of parameter level, parameter group level and global level; a dynamic attention calculation unit for calculating attention weights of time dimension, feature dimension and space dimension based on the current system state; an uncertainty quantification unit for constructing a probability graph network to quantify the uncertainty of the prediction result and generating a confidence interval.

2. The powder metallurgy green body sintering temperature intelligent control system according to claim 1, characterized in that, The temperature adjusting actuator comprises a heating coil, a furnace door temperature detector, and a temperature adjusting controller, the temperature adjusting controller receives heating data of the heating coil and furnace internal temperature data of the reaction furnace, and performs temperature regulation according to the received data, the temperature regulation comprises calculation and application of a first temperature correction amount and a second temperature correction amount, and the second temperature correction amount takes effect when the first temperature correction amount approaches zero.

3. The intelligent sintering temperature control system for powder metallurgy green body according to claim 1, characterized in that, The data processing center further comprises: a data preprocessing module for data cleaning, data conversion and data normalization of the multi-source heterogeneous data; an abnormal data processing and screening module for detecting and removing abnormal data; a training data extraction module for extracting temperature adjustment related data from the preprocessed data as a training data set; a training data sample construction module for constructing a training sample comprising input data and output data; a data model training module for training the temperature prediction model based on the training sample; a data model testing and evaluation module for testing and evaluating the temperature prediction model to obtain an optimal model.

4. The powder metallurgy green body sintering temperature intelligent control system according to claim 3, characterized in that, The data preprocessing module has the following rules for data cleaning: for missing values, the missing values are filled according to the type of the missing object, using the average value or median value of the type of missing data; for abnormal samples, the detected abnormal samples are deleted; for abnormal values, the abnormal values are estimated using the numbers within the range of three-quarters to one-quarter quantile of the group in which the object is located.

5. The intelligent sintering temperature control system for powder metallurgy green body according to claim 3, wherein, The abnormal data processing and screening module detects abnormal data by constructing an isolated forest model, comprising: using all feature data after dimensionless processing for preliminary screening of abnormal data; clustering the data after preliminary screening; constructing an isolated forest model according to the clustering results; further screening abnormal data using the isolated forest model to obtain non-abnormal data as a training data set.

6. The intelligent sintering temperature control system for powder metallurgy green body according to claim 3, wherein, The input data vector constructed by the training data sample construction module includes reaction furnace internal temperature value, billet temperature value, heating coil temperature value, input furnace gas temperature value, input furnace gas concentration value, gas flow value and information whether there is artificial intervention, and the output data includes reaction furnace door temperature value, output furnace gas temperature value and output furnace gas concentration value.

7. A control method using the sintering temperature intelligent control system of any one of claims 1-6, characterized in that, The method comprises the following steps: (1) a data acquisition unit acquires multi-source heterogeneous data inside a reaction furnace in real time, and transmits the multi-source heterogeneous data to a data processing center; (2) the data processing center preprocesses the multi-source heterogeneous data, including data cleaning, data conversion and data normalization; (3) the data processing center uses an abnormal data processing and screening module to detect and screen abnormal data after preprocessing, to obtain non-abnormal data as a training data set; (4) the data processing center extracts temperature adjustment related data from the training data set, and constructs a training sample comprising input data vector and output data; (5) The data processing center uses the training samples to train a temperature prediction model, which includes a topology-aware multi-scale feature extraction module, a chaotic dynamics time series dependency identification module, and a multi-source information fusion dynamic attention module; (6) The topology-aware multi-scale feature extraction module extracts multi-scale features that retain topological relationships from the multi-source heterogeneous data; (7) The chaotic dynamics time series dependency identification module analyzes the nonlinear dynamic characteristics of the temperature time series data and identifies key turning points of temperature changes; (8) The multi-source information fusion dynamic attention module integrates features from different sources and focuses on key information through a multi-dimensional attention mechanism; (9) The data processing center predicts future temperatures under the current state based on the temperature prediction model; (10) The temperature regulation execution mechanism calculates the temperature correction amount and correction curve based on the prediction results; (11) The temperature regulation execution mechanism adjusts the reaction furnace temperature according to the temperature correction amount and correction curve; (12) The data display unit displays the prediction results and control process data in real time.

Images