Diamond particle solidification sintering furnace intelligent control system

Abstract

The application discloses a kind of diamond diamond particle solidification sintering furnace intelligent control system, belong to industrial furnace control technical field, this system aims at solving the problems of existing sintering control experience, process monitoring is not comprehensive, it is difficult to realize real-time optimization and product quality consistency is difficult to guarantee, system includes sensor module, multispectral imaging device, data preprocessing module, liquid neural network module, attention fusion module, user interaction module, thermal structure coupling module and real-time feedback control module, system is collected multimodal data by sensor and multispectral camera, after pre-processing, utilize liquid neural network to process timing dynamic, and by multi-head attention mechanism fusion temperature, humidity, spectrum, mechanical force and other multidimensional features, combine user input and internal state results of thermal structure coupling analysis, real-time feedback control module adopts closed-loop strategy to calculate optimal firing parameter and generate control signal, drive actuating mechanism to carry out adaptive adjustment.

Description

Technical Field

[0001] This invention relates to the field of industrial furnace control technology, specifically an intelligent control system for a diamond particle solidification and sintering furnace. Background Technology

[0002] Diamond granules, due to their superior properties such as high hardness and high wear resistance, are widely used in the manufacture of various tools, abrasives, and functional components. Sintering is a crucial process in the preparation of high-performance diamond composite materials. This process is complex, involving multiple physicochemical changes under high temperature and pressure. Product quality (such as density, hardness, internal stress, and microstructure uniformity) is extremely sensitive to process parameters such as temperature, pressure, holding time, and heating / cooling rates during sintering. Traditional sintering furnace control often relies on operators' experience to set fixed process curves or uses simple feedback control based on classical control theories such as PID (proportional-integral-derivative). This makes it difficult to fully adapt to batch fluctuations in raw materials, differences in furnace loading methods, and the complex nonlinear and time-varying dynamic characteristics during sintering. This leads to problems such as inconsistent product quality, insufficient yield, and difficulty in optimizing production efficiency and energy consumption. Therefore, developing a system capable of real-time monitoring of the sintering state and intelligent, adaptive control is of great significance.

[0003] To improve the control of the sintering process, some existing technologies were explored, such as:

[0004] Chinese invention patent CN103697686B discloses a diamond broaching block sintering furnace and its sintering process. The system uses multiple temperature and pressure sensors to collect data, which is then processed by a PID controller. In particular, it attempts to use the ideal gas law to convert pressure values ​​into indirect temperature values. By comparing the measured temperature values ​​with the set values, it calculates control quantities to adjust the heating circuit. The purpose of this design is to more accurately adjust the furnace temperature and ensure temperature uniformity by using multi-point measurements and combining pressure information.

[0005] Chinese invention patent CN113084718B discloses a forming and sintering process for a metal-bonded diamond grinding head. The focus of this patent is to optimize the pre-sintering preparation process (segmented forming and interlocking assembly) and to set up a detailed and fixed vacuum sintering procedure (stage heating, holding, cooling and gas pressure control) to solve the cracking and porosity problems that are prone to occur after sintering of specific products (grinding heads with large length-to-diameter ratio), aiming to improve yield and reduce costs.

[0006] The above designs, by employing PID control combined with multi-point temperature and pressure measurement, or by optimizing pretreatment processes and setting fixed sintering programs, have improved the temperature control accuracy of the sintering process or solved the molding problems of specific products to a certain extent. However, they still have certain limitations, such as: a single perception dimension, a lack of in-depth insight into the state of the material itself, and an inability to understand in real time the changes in the physicochemical state of the diamond grains in the furnace, such as the uniformity of composition, internal stress and strain distribution, microstructure evolution, and degree of densification, which are key information that directly affect the quality of the final product; relatively simple control strategies, and limited adaptability and control accuracy for complex processes with strong nonlinearity, multivariable coupling, time delay, and time-varying characteristics in sintering; a lack of real-time adaptive adjustment capabilities; insufficient process understanding and model integration; and a lack of a mechanism to combine deep process understanding such as materials science and physical field simulation with real-time control. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent control system for a diamond particle solidification sintering furnace to solve the aforementioned problems. This system achieves multi-dimensional real-time perception of the state of diamond particles by integrating sensing technologies such as multispectral imaging. It uses artificial intelligence models such as liquid neural networks and multi-head attention to process complex time-series and multimodal data and capture process dynamics. It integrates a thermal-structure coupled physical model to analyze the internal state. Finally, it achieves precise, adaptive, and closed-loop optimization adjustment of sintering parameters through a real-time feedback control module, thereby significantly improving the product quality, consistency, yield, and production efficiency of diamond particle sintering.

[0008] The objective of this invention is achieved through the following technical solution: an intelligent control system for a diamond particle solidification sintering furnace, comprising:

[0009] The sensor module collects data through sensors and a multispectral camera to form multimodal raw data;

[0010] The data preprocessing module performs unified formatting, noise reduction, and time-series alignment on the multimodal raw data, and outputs standardized data.

[0011] The liquid neural network module uses the principle of continuous-time dynamic modeling.

[0012] The system processes standardized data, generates hidden state data by simulating the changes in neuron state over time, and quickly solves the state change equation using Taylor expansion and matrix exponential approximation methods to output the hidden state data.

[0013] The attention fusion module processes the latent state data. This module has at least four parallel attention heads: the first head performs a linear transformation on the temperature data to generate query, key, and value vectors, and extracts temperature features using an attention mechanism; the second head performs a linear transformation on the humidity data to generate query, key, and value vectors, and extracts humidity features using an attention mechanism; the third head processes the spectral features extracted from the image that reflect material composition, uniformity, defects, and humidity; and the fourth head processes the mechanical force data. The outputs of all heads are fused through concatenation and linear mapping to form unified multimodal information.

[0014] The user interaction module receives preset material ratios and pressing parameters input by the operator through the operating terminal, reads reference process data from the database, and outputs user-preset parameters.

[0015] The thermal structure coupling module receives real-time temperature and image data from the site. After extracting the appearance changes of diamond grains through image analysis, it uses a preset physical model and computer simulation to map the external temperature and appearance changes into internal stress, strain and temperature distribution, and outputs the internal structural state analysis results.

[0016] The real-time feedback control module receives multimodal information, user-preset parameters, and structural state analysis results. It uses a closed-loop control strategy to calculate the optimal firing parameters, converts the calculation results into control signals, and feeds them back to the sintering furnace actuator in real time to achieve automatic adjustment and stable control.

[0017] The sensors in the sensor module collect data on temperature, humidity, pressure, and mechanical force, while the multispectral camera is fixedly installed inside the sintering furnace and aimed at the diamond particles to be sintered.

[0018] The data preprocessing module also provides some of the preprocessed data to the user interaction module for comparison with the user input information.

[0019] The liquid neural network module consists of a data input unit, a continuous state modeling unit, and a closed-form solution unit. The continuous state modeling unit is responsible for simulating the evolution of the neuron's state over time, while the closed-form solution unit is responsible for quickly approximating the solution of the state change equation.

[0020] Each parallel attention head in the attention fusion module uses an independent linear transformation unit to generate its query, key, and value vectors respectively. The fusion method of the outputs of each head is to concatenate the output vectors of each head and then generate unified multimodal information through a linear mapping layer.

[0021] The user interaction module is also configured to automatically compare the actual material composition, uniformity, defects and moisture content information detected and analyzed by a multispectral camera with the preset material ratio input by the operator, and issue an alarm signal through the operation terminal when the deviation between the two exceeds a preset threshold.

[0022] The thermal structure coupling module includes a data conversion unit, a physical modeling unit, and a simulation analysis unit. The data conversion unit is responsible for converting real-time temperature data and extracted appearance change features into an input format suitable for the physical model. The physical modeling unit performs mapping based on a preset coupling model, and the simulation analysis unit is responsible for calculating the internal state.

[0023] The preset physical model and computer simulation used in the thermal structure coupling module can dynamically and accurately reflect the distribution of internal stress, strain and temperature of diamond grains at different stages of the sintering process, providing decision support for the real-time feedback control module.

[0024] The real-time feedback control module includes a decision calculation unit and a control signal generation unit. The decision calculation unit is responsible for executing the closed-loop control strategy to calculate the optimal firing parameters, and the control signal generation unit is responsible for converting the calculated parameters into control signals compatible with the sintering furnace actuator and implementing adjustment logic including anomaly compensation.

[0025] It also includes an online learning module, which is configured to receive real-time feedback data and control status data generated during system operation, and adaptively update the network weights and dynamic parameters in the liquid neural network module and attention fusion module to continuously optimize the accuracy and response speed of the control model to adapt to different sintering conditions.

[0026] The beneficial effects of this invention are:

[0027] 1. By integrating various traditional sensors (monitoring temperature, pressure, humidity, mechanical force, etc.) with multispectral imaging technology, this system can acquire multi-dimensional, high-resolution real-time process information. It can not only monitor macroscopic process parameters, but also precisely capture the subtle changes in material composition, distribution uniformity, surface defects, humidity, spatial temperature distribution, appearance geometric evolution, and optical features related to crystal structure of diamond grains during the sintering process. Based on the physical principle of thermal structure coupling modeling and simulation analysis capabilities, it can map and correlate easily measurable external appearance data (such as surface temperature and appearance changes) with key internal states (such as internal stress field, strain distribution, and internal temperature gradient) that are difficult to measure directly. This allows the system to focus on the internal physicochemical changes during the sintering process, deepening the real-time understanding of complex sintering mechanisms.

[0028] 2. Employing a liquid neural network, it can effectively capture and simulate the complex dynamic characteristics of the sintering process itself, including strong nonlinearity, time-varying continuity, and long-term dependence. It utilizes closed-form solution approximation to handle the neural network's state evolution equations, ensuring that the complex dynamic model can operate at speeds that meet real-time control requirements. It integrates an attention mechanism with multiple parallel processing paths, intelligently fusing multimodal information flows from different sensors, multispectral image analysis, physical models, and user input. Based on the current sintering stage and state, it dynamically evaluates and weights the importance of different information sources, ensuring that the most critical and relevant features play a leading role in system decision-making at the current moment, avoiding information redundancy or the submersion of critical information.

[0029] 3. A real-time feedback closed-loop control loop was constructed, which can continuously compare the comprehensive evaluation results reflecting the actual sintering state with the preset target or dynamically optimized trajectory. The control decision module can calculate the optimal control parameter adjustment amount (such as heating / cooling rate, holding time, target pressure, pressurization rate, etc.) under the current operating conditions based on the control strategy. The control commands are converted into operation signals for the actuators such as the heating system and pressure system of the sintering furnace in real time, realizing precise, continuous or high-frequency periodic dynamic adjustment of the sintering process. It can automatically compensate for the slight differences between batches of raw materials, fluctuations in environmental conditions, and unexpected disturbances that may occur during the process, so that the sintering process always stays on the optimal or near-optimal path.

[0030] 4. Through process optimization, the quality consistency of the final sintered diamond product can be improved, resulting in a more uniform internal structure, higher density, ideal hardness, and fewer internal defects. Automated material detection and verification before sintering can effectively identify unqualified raw materials and avoid ineffective production. At the same time, the adaptive adjustment capability in the process reduces scrap caused by process fluctuations, thereby improving the yield and material utilization rate. By controlling energy input and process time, redundant heating or excessively long heat preservation is avoided, optimizing energy consumption, shortening the production cycle, and reducing production costs.

[0031] 5. It improves the automation level of the sintering process, reduces the dependence on operator experience and the burden of manual monitoring and parameter adjustment. The dynamic neural network and attention module can continuously adjust and optimize themselves using real-time feedback data during system operation, enabling the system to adapt to slowly changing working conditions (such as equipment wear, seasonal environmental changes, and fine-tuning of raw material supply) for a long time, continuously improve control accuracy and response characteristics, maintain optimal performance, and reduce the frequency and cost of offline model maintenance and retraining. Attached Figure Description

[0032] Figure 1 This is a system interaction diagram of the present invention;

[0033] Figure 2 This is a system architecture diagram of the present invention;

[0034] Figure 3 This is a timing diagram for the present invention.

[0035] Explanation of the labels in the diagram

[0036] 1. Sensor module; 2. Data preprocessing module; 3. Liquid neural network module; 4. Attention fusion module; 5. User interaction module; 6. Thermal structure coupling module; 7. Real-time feedback control module; 8. Online learning module; 101. Multimodal raw data; 102. Standardized data; 103. Latent state data; 104. Multimodal information; 105. User preset parameters; 106. Structural state analysis results; 107. Control signal. Detailed Implementation

[0037] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0039] Example 1:

[0040] like Figures 1 to 3 As shown in the figure, this embodiment discloses an intelligent control system for a diamond particle solidification sintering furnace, which realizes real-time acquisition, preprocessing, time-series modeling, information fusion and closed-loop control of multimodal data.

[0041] The system first collects various operating data inside the sintering furnace through sensor module 1. Sensor module 1 includes multiple field sensors for collecting temperature, humidity, pressure and mechanical force data. It also includes a multispectral camera, which is fixedly installed inside the sintering furnace. Its field of view is precisely aligned with the diamond grains to be sintered to collect image data covering the visible light and near-infrared bands. All the collected data is integrated by sensor module 1 to form multimodal raw data 101, and then output to data preprocessing module 2.

[0042] After receiving the multimodal raw data 101 from the sensor module 1, the data preprocessing module 2 first performs a unified format conversion on the data from different sources; secondly, it uses filtering and other algorithms to reduce noise in the raw data; finally, it performs precise time-series alignment of the data based on the timestamps of each data point. After processing, the standardized data 102 output by the data preprocessing module 2 not only has a regular data format and a high signal-to-noise ratio, but also meets the requirements of the liquid neural network module 3 for the continuity and temporality of the input data. In addition, some of the preprocessed data will be transmitted to the user interaction module 5 for process parameter comparison.

[0043] The liquid neural network module 3 is used for time-series dynamic modeling of the standardized data 102. This module consists of a data input unit, a continuous state modeling unit, and a closed-form solution unit. The data input unit receives the standardized data 102. The continuous state modeling unit simulates the process of neuron state change over time based on the principle of continuous-time dynamic modeling, thereby capturing the complex dynamics in the sintering process. The closed-form solution unit uses Taylor expansion and matrix exponential approximation methods to quickly approximate the solution of the differential equation describing the state change. The hidden state data 103 output by the liquid neural network module 3 not only records the historical information of the process, but also reflects the current dynamics in real time, providing high-quality feature data for subsequent processing.

[0044] The dynamic modeling formula for liquid neural network module 3 is as follows:

[0045] Dynamic coupling time constant and gain

[0046] In traditional liquid neural networks, the time constant τ is a single learnable parameter. Here, we introduce a dual dynamic coupling where the time constant τ(t) and the gain γ(t) both depend on the input x(t):

[0047]

[0048] The learnable weight matrix; It is a learnable scalar;

[0049] State evolution differential equation: Add a quadratic coupling term θ(t)h(t)⊙h(t) to enhance nonlinear expressive power:

[0050]

[0051] Nonlinear activation; These are dynamic quadratic coupling coefficients; ⊙ represents element-wise multiplication;

[0052] Closed-form discrete update: Let the discrete step size be Δt. Using the closed-form approximation of the dynamic coupling time constant, the state update formula can be obtained:

[0053]

[0054] (Available) approximate)

[0055] and Step size weights for primary and secondary coupling, respectively.

[0056] This formula integrates three contribution terms: first-order attenuation, adaptive gain drive, and secondary coupling, achieving efficient and expressive dynamic modeling.

[0057] The latent state data 103 is passed to the attention fusion module 4 for intelligent feature fusion. This module is designed with at least four parallel attention heads:

[0058] The first attention head performs a linear transformation on the temperature data to generate query, key, and value vectors, and then uses a scaled dot product attention mechanism to extract temperature features.

[0059] The second attention head performs a linear transformation on the humidity data to generate query, key, and value vectors, and uses a scaling dot product attention mechanism to extract humidity features.

[0060] The third focus is to extract spectral features from image data that reflect material composition, uniformity, defects, and humidity.

[0061] The fourth attention head processes mechanical force data. The output vectors of each attention head are concatenated and then fused through a linear mapping layer to form unified multimodal information 104. This multimodal information 104 comprehensively reflects the important characteristics of the four channels of temperature, humidity, spectrum and mechanical force, and becomes the key basis for subsequent decision-making.

[0062] The multi-head fusion formula for Attention Fusion Module 4:

[0063] Head-wise vector generation: For each head i∈{1,2,3,4}, extract different subspace features from the hidden state hn to generate query, key, and value vectors:

[0064]

[0065] Each group Learnable parameters

[0066] Scaling Similarity and Adaptive Weights: Unlike traditional softmax, we first normalize and scale the similarity, then introduce dynamic head weights λi.

[0067]

[0068] Prevent division by zero; A head-specific weight vector, used to adjust the weights based on the current state. Adaptively adjust the importance of each head;

[0069] Weighted output, the output of each head combines scaled similarity with dynamic weights:

[0070]

[0071] , , Indicates first to Perform softmax, then left multiply .

[0072] Concatenation and Remapping: Finally, the outputs of each head are concatenated, and a unified multimodal information Mn is obtained through the remapping layer.

[0073]

[0074] For learnable linear mapping parameters, This indicates vector concatenation;

[0075] User interaction module 5 serves as the interface between the system and the operator. It receives preset material ratios and pressing parameters input by the operator through an operating terminal (such as a touch screen or computer interface). At the same time, it can read reference process data from the internal database. After integrating various information, this module outputs the user-preset parameter 105 and transmits it to the real-time feedback control module 7 for subsequent decision-making.

[0076] The thermal structure coupling module 6 is used to map measurable external data to the internal state of diamond grains. The thermal structure coupling module 6 first receives real-time acquired temperature data and image data from a multispectral camera; then, it extracts the appearance change characteristics of diamond grains during the sintering process through image analysis; finally, using a preset thermal structure coupling model based on physical principles and computer simulation methods, it maps external temperature and appearance changes to internal stress, strain, and temperature distribution. The internal structural state analysis results 106 output by this module provide important physical basis for control decisions.

[0077] The dynamics formula for thermal structure coupling is as follows: Let the state vector

[0078]

[0079] For strain tensor; For stress tensor; It is a temperature scalar;

[0080] The entire thermal-structure coupling system can then be written as:

[0081]

[0082] The explanations for each item are as follows:

[0083] Elastic-thermal coupling (strain rate)

[0084] Stress evolution

[0085] Temperature evolution

[0086]

[0087]

[0088] Stress relaxation coefficient

[0089] Basic elastic stiffness

[0090] thermal softening coefficient

[0091] Basic thermal conductivity

[0092] Strain-dependent thermal conductivity amplification factor

[0093] Mechanical dissipation coupling coefficient

[0094] Reference temperature

[0095] The real-time feedback control module 7 is the decision-making and execution center of the system. This module includes a decision calculation unit and a control signal generation unit. The decision calculation unit receives multimodal information 104 from the attention fusion module 4, user preset parameters 105 from the user interaction module 5, and internal structural state analysis results 106 from the thermal structure coupling module 6. It calculates the optimal firing parameters using a closed-loop control strategy (such as model predictive control, PID control, etc.). The control signal generation unit converts the optimal parameters into control signals 107, which are fed back to the actuators of the sintering furnace (such as heaters and pressure systems) in real time, thereby achieving automatic adjustment and stable control.

[0096] Work process

[0097] After the system starts up, it performs a self-test. The operator inputs the preset material ratio and pressing parameters through the user interaction module 5. The user interaction module 5 generates the user preset parameter 105 from this information and reads the reference process data at the same time.

[0098] After the raw materials are put in, sensor module 1 continuously collects on-site sensor data and multispectral camera images to form multimodal raw data 101.

[0099] The multimodal raw data 101 is formatted, denoised, and time-series aligned by the data preprocessing module 2, and then outputs standardized data 102; some data is also transmitted to the user interaction module 5 for parameter comparison.

[0100] Standardized data 102 is input into the liquid neural network module 3, and hidden state data 103 is generated in real time through the synergistic effect of the data input unit, the continuous state modeling unit, and the closed-form solution unit.

[0101] The latent state data 103 is fed into the attention fusion module 4. The four attention heads of the attention fusion module 4 extract features from four dimensions: temperature, humidity, image spectrum and mechanical force, and then perform weighted fusion to finally form unified multimodal information 104.

[0102] Meanwhile, the thermal structure coupling module 6 receives real-time temperature data and image data. After extracting appearance change features through image analysis, it maps the data into internal stress, strain and temperature distribution through a preset physical model and computer simulation, and outputs the internal structure state analysis results 106.

[0103] The real-time feedback control module 7 summarizes the multimodal information 104, user preset parameters 105, and internal structural state analysis results 106, and calculates the optimal firing parameters using a closed-loop control strategy. These parameters are converted into control signals 107 by the control signal generation unit and sent to the sintering furnace actuator in real time to adjust the heater, pressure system, etc.

[0104] The system continuously performs data acquisition, preprocessing, modeling, feature fusion, state mapping, and decision control to ensure that it responds to dynamic changes on site at any time until the sintering process is completed.

[0105] The sensor module 1 collects temperature, humidity, pressure, mechanical force and multispectral image data, enabling the system to acquire multimodal information inside the sintering furnace in a comprehensive and real-time manner. The raw multimodal data 101 can truly reflect the operating status inside the furnace.

[0106] The data preprocessing module 2 performs uniform formatting, noise reduction, and temporal alignment on the multimodal raw data 101. The output standardized data 102 ensures data continuity and temporal accuracy, providing high-quality input for the liquid neural network module 3, thereby improving the reliability of the hidden state data 103.

[0107] The liquid neural network module 3 utilizes continuous-time dynamic modeling and closed-form solution technology, which can capture the complex dynamics of the sintering process in real time with low memory usage and high-speed response, laying a solid foundation for subsequent feature fusion. The output hidden state data 103 reflects the process history and current dynamics.

[0108] The multi-head attention mechanism built into the attention fusion module 4 can independently weight temperature, humidity, image spectrum and mechanical force data, and finally form a unified multimodal information 104 through concatenation and linear mapping. This enables the system to intelligently and dynamically adjust the information weights of each channel, thereby achieving a more accurate state description.

[0109] The real-time feedback control module 7 integrates the multimodal information 104 output by the attention fusion module 4, the user preset parameters 105 of the user interaction module 5, and the internal structural state analysis results 106 of the thermal structure coupling module 6. It then uses a closed-loop control strategy to calculate the optimal firing parameters and converts the results into control signals 107 to provide precise instructions to the sintering furnace actuators. This achieves automatic adjustment and anomaly compensation, ensuring process stability and product quality consistency.

[0110] This embodiment achieves integrated control from data acquisition, preprocessing, state modeling, feature fusion to closed-loop control. The systematic automatic control can reduce human intervention and improve production efficiency and process stability.

[0111] In summary, Example 1, through the collaborative work of sensor module 1, data preprocessing module 2, liquid neural network module 3, attention fusion module 4, user interaction module 5, thermal structure coupling module 6, and real-time feedback control module 7, achieves real-time acquisition, processing, modeling, fusion, and closed-loop control of multimodal data within the sintering furnace. This significantly improves the monitoring accuracy and control effect of the sintering process, and has a marked improvement effect on product quality and process stability.

[0112] Example 2:

[0113] like Figures 1 to 3 As shown, this embodiment adopts the core intelligent control system architecture in Embodiment 1, namely sensor module 1, data preprocessing module 2, liquid neural network module 3, attention fusion module 4, user interaction module 5, thermal structure coupling module 6, and real-time feedback control module 7. At the same time, the functions of user interaction module 5 and thermal structure coupling module 6 are further enhanced and refined.

[0114] In this embodiment, the user interaction module 5 not only receives the preset material ratio and pressing parameters input by the operator through the operating terminal to form user preset parameters 105, but also has automatic material verification and alarm functions. Its internal configuration includes a comparison logic unit for the following operations:

[0115] Receive partial preprocessed data (e.g., initially detected material property data) from the output of data preprocessing module 2.

[0116] The system receives actual material information, including data on the material composition, distribution uniformity, surface defects, and moisture content of diamond particles, which is collected by the multispectral camera through the sensor module 1 and preprocessed by the data preprocessing module 2.

[0117] It automatically compares the actual material information with the material ratio preset by the operator and judges the deviation based on the preset safety or process threshold.

[0118] If the comparison results show a deviation exceeding the threshold, an alarm function will be triggered, and the operator will be notified through the display interface or voice prompts to confirm or handle the situation.

[0119] At the same time, the verified user-preset parameter 105 is output to the real-time feedback control module 7.

[0120] In this embodiment, the thermal structure coupling module 6 is further refined into three cooperating sub-units based on embodiment 1 to enhance the ability to analyze the internal physical state during sintering:

[0121] The data conversion unit receives real-time temperature data collected from the sensor module 1 and diamond grain appearance change characteristics (such as size shrinkage rate, surface roughness change, etc.) extracted by image analysis, and converts these data into a standardized data format that matches the physical modeling requirements.

[0122] The physical modeling unit has a built-in preset model based on the principle of thermo-structure coupling. This model describes the interaction between diamond particles, grain boundary migration, deformation diffusion, and interaction with the binder under high temperature and high pressure. The unit uses the standardized data 102 output by the data conversion unit as the model input to calculate the stress distribution, strain field, and temperature field inside the diamond particles.

[0123] The simulation analysis unit uses the preliminary results obtained by mapping the physical modeling unit to perform numerical calculations and simulations through the finite element or finite difference method, and obtains more detailed and accurate internal structural state analysis results 106. These results can dynamically reflect the changes in the internal state at each stage of the sintering process.

[0124] Apart from the enhancements mentioned above, the structure and function of sensor module 1, data preprocessing module 2, liquid neural network module 3, attention fusion module 4, and real-time feedback control module 7 are consistent with those of embodiment 1. However, they can now make full use of the richer and more accurate information provided by the enhanced user interaction module 5 and thermal structure coupling module 6 to form higher quality multimodal information 104 and internal structural state analysis results 106, thereby supporting more precise control decisions.

[0125] Work process

[0126] The working process of this embodiment is based on that of embodiment 1, but adds material verification and enhanced physical state analysis steps. The specific steps are as follows:

[0127] During the initialization phase, after the system self-check is completed, the operator inputs the preset material ratio and pressing parameters into the operation terminal through the user interaction module 5, forming user preset parameters 105.

[0128] Meanwhile, the multispectral camera acquires initial material images, which are then processed by the data preprocessing module 2 to generate standardized data 102. The information about material properties in the preprocessed data is then transmitted to the user interaction module 5.

[0129] The comparison logic unit in the user interaction module 5 automatically compares the material ratio preset by the operator with the actual detected material composition, uniformity, defects, and moisture content. If the deviation between the detection result and the preset information exceeds the safety threshold, the alarm function is automatically triggered, and the operator is prompted to confirm or adjust through the display interface or voice prompts to ensure that the input raw materials meet the requirements.

[0130] The data acquisition, preprocessing, liquid neural network processing, and attention fusion processes are carried out according to the workflow of Example 1 to generate unified multimodal information 104.

[0131] The thermal-structure coupling module 6 simultaneously receives real-time temperature data and image data from the site. The data conversion unit converts this data into a format suitable for physical modeling. The physical modeling unit performs state mapping based on the preset thermal-structure coupling model. The simulation analysis unit calculates the detailed distribution of internal stress, strain, and temperature field, and outputs the internal structure state analysis results 106.

[0132] The dynamic simulation results of the thermal structure coupling module 6 can reflect the changes in the internal state of the sintering process in real time during different stages such as heating, holding, pressurizing, holding pressure, and cooling.

[0133] The real-time feedback control module 7 receives multimodal information 104 from the attention fusion module 4, user preset parameters 105 from the user interaction module 5, and internal structural state analysis results 106 from the thermal structure coupling module 6.

[0134] The control module calculates the optimal firing parameters for the current or next stage through a closed-loop control strategy and converts them into control signal 107.

[0135] The control signal 107 is fed back to the sintering furnace actuators (including heater controller, pressure controller, etc.) in real time, thereby realizing automatic adjustment and stable control of the sintering process.

[0136] The system continuously performs data acquisition, state modeling, physical analysis, and control decision-making in a loop until the sintering process ends.

[0137] With the enhanced user interaction module 5, the system automatically compares the preset material ratio input by the operator with the actual material detection data before sintering. If the deviation exceeds the standard, an alarm signal is issued in time, thereby identifying raw material abnormalities in advance and reducing production risks and quality decline caused by unqualified raw materials.

[0138] The thermal-structure coupling module 6 is further subdivided into a data conversion unit, a physical modeling unit, and a simulation analysis unit. The high-precision thermal-structure coupling model, combined with computer simulation technology, dynamically reflects the internal stress, strain, and temperature distribution of diamond grains at different sintering stages, fundamentally revealing the physical evolution mechanism during the sintering process and providing a more scientific basis for control decisions.

[0139] The real-time feedback control module 7 utilizes enhanced internal structure state analysis results 106 and multi-dimensional fusion multimodal information 104 to perform more accurate closed-loop calculations of the control strategy, thereby rapidly responding to changes in raw materials and the environment during the sintering process, effectively suppressing process anomalies, and improving the consistency and quality stability of the finished product.

[0140] This embodiment achieves automated integration of the entire process from material verification and in-depth internal state analysis to control decision-making, reducing the need for manual intervention, improving production efficiency and system safety, and providing rich analytical basis for process monitoring and fault diagnosis.

[0141] In summary, by enhancing the functions of the user interaction module 5 and the thermal structure coupling module 6, Example 2 not only improves the raw material quality control and early warning capabilities, but also provides more accurate information for real-time feedback control through in-depth physical state analysis. This enables the entire sintering process to achieve higher precision and stability of automatic closed-loop control, resulting in a significant improvement in product quality and process reliability.

[0142] Example 3:

[0143] like Figures 1 to 3 As shown, this embodiment further integrates the online learning module 8 on the basis of the core intelligent control system of embodiment 1 or embodiment 2, so that the system has the ability to continuously adapt and dynamically optimize.

[0144] Sensor module 1 collects temperature, humidity, pressure and mechanical force data from field sensors, and simultaneously collects image data through a multispectral camera fixed inside the sintering furnace and aimed at the diamond grains to be sintered, forming multimodal raw data 101.

[0145] The data preprocessing module 2 performs unified formatting, noise reduction and time alignment processing on the multimodal raw data 101 output by the sensor module 1, and outputs standardized data 102. Some of the data is also transmitted to the user interaction module 5 as a comparison basis.

[0146] The liquid neural network module 3 uses the principle of continuous-time dynamic modeling to process the standardized data 102. The liquid neural network module 3 includes a data input unit, a continuous state modeling unit, and a closed-form solution unit. The continuous state modeling unit simulates the change of neuron state over time, while the closed-form solution unit uses Taylor expansion and matrix exponential approximation methods to quickly solve the state change equation and output the hidden state data 103.

[0147] Attention fusion module 4 This module has at least four parallel attention heads, which process temperature, humidity, spectral features extracted from images (reflecting material composition, uniformity, defects and dryness), and mechanical force data respectively. Each head generates query, key and value vectors using independent linear transformations. After calculating their respective features through the attention mechanism, they are fused through concatenation and linear mapping layers to form unified multimodal information 104.

[0148] User interaction module 5 allows users to input preset material ratios and pressing parameters through the operating terminal. Simultaneously, user interaction module 5 can read reference process data from the database. User interaction module 5 integrates and outputs user-preset parameters 105. At the same time, it automatically compares some initial detection data from data preprocessing module 2 with the preset parameters and triggers an alarm when the preset threshold is reached.

[0149] The thermal-structure coupling module 6 receives real-time temperature data and image data from the sensor module 1, extracts the appearance change characteristics of diamond grains (such as size shrinkage, surface roughness, etc.) through image analysis, and maps the external data into internal stress, strain and temperature distribution through a preset thermal-structure coupling model and computer simulation, and outputs the internal structural state analysis results 106.

[0150] The real-time feedback control module 7 integrates the multimodal information 104 of the attention fusion module 4, the user preset parameters 105 of the user interaction module 5, and the internal structural state analysis results 106 of the thermal structure coupling module 6. It then uses a closed-loop control strategy to calculate the optimal firing parameters and converts them into control signals 107, which are fed back to the sintering furnace actuator in real time to achieve automatic control and anomaly compensation.

[0151] Online learning module 8 is the core enhancement component of this embodiment. Online learning module 8 can receive real-time feedback data and status data from various parts of the system, mainly including:

[0152] Feedback information from the real-time feedback control module 7 (e.g., the deviation between the target firing parameters and the actual execution parameters, and the dynamic changes of the control signal 107).

[0153] Intermediate state and output data from the liquid neural network module 3 and the attention fusion module 4 (e.g., comparison information between hidden state data 103 and actual measured values).

[0154] Key operating condition data from sensor module 1 or data preprocessing module 2, and (if detectable) final product quality evaluation data.

[0155] The online learning module 8 integrates online learning algorithms such as online gradient descent, Kalman filtering, or reinforcement learning. It uses a preset objective function (such as minimizing control error or improving product quality) to calculate the adjustment of weights and dynamic parameters in the liquid neural network module 3 and the attention fusion module 4. After processing by the online learning module 8, the updated parameters are fed back to the liquid neural network module 3 and the attention fusion module 4 in real time, thereby dynamically adjusting the response characteristics of these core models.

[0156] Work process

[0157] The specific steps of the working process in this embodiment are as follows:

[0158] Sensor module 1 continuously collects on-site temperature, humidity, pressure, mechanical force and image data to form multimodal raw data 101, which is then processed into standardized data 102 by data preprocessing module 2.

[0159] Standardized data 102 is processed in real time by liquid neural network module 3 to generate hidden state data 103; hidden state data 103 enters attention fusion module 4, and each attention head extracts temperature, humidity, image spectrum and mechanical force features and then fuses them into multimodal information 104.

[0160] Meanwhile, the user interaction module 5 receives input from the operator and outputs user-preset parameters 105; the thermal structure coupling module 6 outputs internal structure state analysis results 106 based on real-time temperature and image data; the real-time feedback control module 7 integrates multimodal information 104, user-preset parameters 105 and structural state analysis results 106, calculates the optimal firing parameters through a closed-loop control strategy, and generates control signals 107 to feed back to the actuator, thereby realizing real-time adjustment of the sintering process.

[0161] While the main control loop is running, the online learning module 8 periodically or in real time collects control error information, hidden state data 103, attention fusion intermediate output and key environmental data from the real-time feedback control module 7.

[0162] Based on the collected data and preset learning objectives, the online learning module 8 calculates the adjustment amounts of the weights and dynamic parameters in the liquid neural network module 3 and the attention fusion module 4 through the online learning algorithm.

[0163] The calculated update is fed back to the liquid neural network module 3 and the attention fusion module 4 in real time or periodically, enabling these core modules to have better dynamic response capabilities when processing input data.

[0164] This online learning process runs in parallel with the main control loop, enabling continuous adaptation and optimization of the model.

[0165] The online learning module 8 enables the liquid neural network module 3 and the attention fusion module 4 to dynamically adjust the core parameters based on real-time feedback, thereby adapting to changes in operating conditions caused by raw material batches, environmental changes, or equipment aging, and maintaining the system's optimal control performance under various conditions.

[0166] The system continuously learns from real-time feedback data, enabling online updates of model weights and dynamic parameters. As the running time increases, the system's control accuracy, response speed, and robustness gradually improve, achieving the effect of "getting better with use."

[0167] Online learning mechanisms reduce the need for large amounts of offline training data and periodic manual intervention, enabling the system to automatically adapt to environmental changes and reducing the cost of long-term model maintenance and updates.

[0168] Thanks to optimized liquid neural network modeling and more dynamic multimodal information fusion, the real-time feedback control module 7 can generate more precise control signals 107, ensuring that the sintering process parameters are maintained in the optimal state, thereby improving product consistency and quality.

[0169] The online learning module 8 can also provide the system with long-term data accumulation, which helps to identify and diagnose abnormal operating conditions, provide timely feedback on potential problems, and further ensure production safety and process stability.

[0170] In summary, based on Examples 1 and 2, Example 3 achieves continuous self-learning and real-time model optimization through the online learning module 8, overcoming the shortcomings of traditional pre-trained models in adapting to environmental changes. This results in the entire intelligent control system having higher adaptability, continuous optimization capability, and control precision, which has significant advantages in improving the stability of the sintering process and product quality.

[0171] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be modified within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An intelligent control system for a diamond particle solidification and sintering furnace, characterized in that, include: The sensor module (1) collects data through sensors and a multispectral camera to form multimodal raw data (101). The data preprocessing module (2) performs unified formatting, noise reduction and time alignment on the multimodal raw data (101) and outputs standardized data (102). The liquid neural network module (3) processes standardized data (102) using the principle of continuous-time dynamic modeling. It generates hidden state data by simulating the change of neuron state over time, and uses Taylor expansion and matrix exponential approximation method to quickly solve the state change equation and output the hidden state data (103). The attention fusion module (4) processes the latent state data (103). This module has at least four parallel attention heads, of which: the first head performs a linear transformation on the temperature data to generate query, key, and value vectors, and extracts temperature features using the attention mechanism; the second head performs a linear transformation on the humidity data to generate query, key, and value vectors, and extracts humidity features using the attention mechanism; the third head processes the spectral features extracted from the image that reflect the material composition, uniformity, defects, and humidity; the fourth head processes the mechanical force data; and the outputs of each head are fused by concatenation and linear mapping to form unified multimodal information (104). The user interaction module (5) receives the preset material ratio and pressing parameters input by the operator through the operation terminal, reads the reference process data from the database, and outputs the user preset parameters (105). The thermal structure coupling module (6) receives real-time temperature data and image data collected on site. After extracting the appearance changes of diamond grains through image analysis, it uses a preset physical model and computer simulation to map the external temperature and appearance changes into internal stress, strain and temperature distribution, and outputs the internal structure state analysis results (106). The real-time feedback control module (7) receives multimodal information (104), user preset parameters (105), and structural state analysis results (106), calculates the optimal firing parameters using a closed-loop control strategy, converts the calculation results into control signals (107), and feeds them back to the sintering furnace actuator in real time to achieve automatic adjustment and stable control.

2. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: The sensors in the sensor module (1) specifically collect temperature, humidity, pressure and mechanical force data, and the multispectral camera is fixedly installed inside the sintering furnace and aimed at the diamond particles to be sintered.

3. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: The data preprocessing module (2) also provides some of the preprocessed data to the user interaction module (5) for comparison with the user input information.

4. The intelligent control system for a diamond particle solidification and sintering furnace according to claim 1, characterized in that: The liquid neural network module (3) includes a data input unit, a continuous state modeling unit, and a closed-form solution unit. The continuous state modeling unit is responsible for simulating the evolution of the neuron state over time, and the closed-form solution unit is responsible for quickly approximating the solution of the state change equation.

5. The intelligent control system for a diamond particle solidification and sintering furnace according to claim 1, characterized in that: Each parallel attention head in the attention fusion module (4) generates its query, key, and value vectors using an independent linear transformation unit. The fusion method of each head output is to concatenate the output vectors of each head and then generate the unified multimodal information (104) through a linear mapping layer.

6. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: The user interaction module (5) is also configured to automatically compare the actual material composition, uniformity, defects and moisture information detected and analyzed by the multispectral camera with the preset material ratio input by the operator, and issue an alarm signal through the operation terminal when the deviation between the two exceeds a preset threshold.

7. The intelligent control system for a diamond particle curing and sintering furnace according to claim 1, characterized in that: The thermal structure coupling module (6) includes a data conversion unit, a physical modeling unit, and a simulation analysis unit. The data conversion unit is responsible for converting real-time temperature data and extracted appearance change features into an input format suitable for the physical model. The physical modeling unit performs mapping based on a preset coupling model. The simulation analysis unit is responsible for calculating the internal state.

8. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: The preset physical model and computer simulation used in the thermal structure coupling module (6) can dynamically and accurately reflect the distribution of internal stress, strain and temperature of diamond grains at different stages of the sintering process, providing decision support for the real-time feedback control module (7).

9. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: The real-time feedback control module (7) includes a decision calculation unit and a control signal generation unit. The decision calculation unit is responsible for executing the closed-loop control strategy to calculate the optimal firing parameters. The control signal generation unit is responsible for converting the calculated parameters into control signals (107) compatible with the sintering furnace actuator and implementing adjustment logic including anomaly compensation.

10. The intelligent control system for a diamond particle solidification sintering furnace according to claim 1, characterized in that: It also includes an online learning module (8), which is configured to receive real-time feedback data and control status data generated during system operation, and adaptively update the network weights and dynamic parameters in the liquid neural network module (3) and the attention fusion module (4) to continuously optimize the accuracy and response speed of the control model in order to adapt to different sintering conditions.

Images