NiP binary alloy smelting method and system
By using heat flux density calculation and reinforcement learning algorithms to generate differentiated heating strategies during the NiP binary alloy smelting process, the problems of energy waste and alloy quality inhomogeneity under traditional uniform heating methods are solved, achieving more efficient energy utilization and temperature control.
Patent Information
- Application Number
- CN202511424893.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2025-12-16
AI Technical Summary
In the existing NiP binary alloy smelting process, the heat dissipation differences in different areas of the smelting furnace are ignored, resulting in energy waste and uneven alloy quality.
By acquiring multi-point temperature data, a differentiated power allocation strategy is generated using heat flux density calculation algorithms and reinforcement learning algorithms to dynamically adjust the heating power of each region to match the heat dissipation characteristics.
It improves the energy utilization efficiency and temperature uniformity of the NiP binary alloy smelting process, thereby enhancing the consistency of alloy quality.
Smart Images
Figure CN121140407A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of metal alloy smelting technology, and more specifically, to a NiP binary alloy smelting method and system. Background Technology
[0002] NiP binary alloys are widely used in electronic components, machinery manufacturing, and chemical equipment due to their excellent corrosion resistance, wear resistance, and magnetic properties. In the industrial production of NiP binary alloys, the smelting process is the key step that determines the alloy's performance and quality.
[0003] Currently, the most common technology for controlling the temperature of melting furnaces is a heating method with uniform power distribution. This technology attempts to achieve a uniform temperature distribution within the furnace by providing the same power input to all heating zones. This method has a relatively simple control strategy and is easy to implement.
[0004] However, existing uniform heating technologies have significant drawbacks: they ignore the different heat dissipation characteristics of different areas within the melting furnace. Because the heat dissipation conditions vary at different locations such as the furnace wall, bottom, and top, uniform power distribution can lead to overheating of colder areas that dissipate heat quickly to maintain the target temperature, while hotter areas that dissipate heat more slowly may overheat. This mismatch between heat dissipation differences and the heating method directly reduces the overall energy efficiency of the melting process and affects the uniformity of the alloy quality. Summary of the Invention
[0005] This invention provides a NiP binary alloy smelting method, which solves the technical problem of energy waste caused by ignoring regional heat dissipation differences in related technologies.
[0006] This invention discloses a NiP binary alloy melting method, comprising: acquiring multi-point temperature data within a furnace; collecting ambient temperature, furnace wall temperature, and furnace interior temperature data using temperature sensors; organizing the temperature measurements from different locations into a multi-layer temperature dataset according to spatial coordinates and time series; analyzing the temperature dataset based on a heat flux density calculation algorithm to calculate the heat dissipation rate values for each region within the furnace, generating a heat dissipation distribution map describing the heat dissipation characteristics of different regions; processing the heat dissipation distribution data using a reinforcement learning algorithm, using the heat dissipation distribution map and corresponding operating parameters as training data to learn the optimal power compensation relationship for each region under different operating conditions, generating a strategy function; inputting real-time operating parameters into the strategy function to obtain the current furnace temperature, ambient temperature, and other real-time operating parameters, calculating and outputting differentiated power allocation values for each region.
[0007] This invention discloses a system for performing the above-described NiP binary alloy melting method, comprising a temperature sensor module for collecting ambient temperature, furnace wall temperature, and furnace interior temperature data; a data processing module for organizing multi-layer temperature datasets and executing heat flux density calculation algorithms; a reinforcement learning module for processing heat dissipation distribution data and generating a strategy function; and a power control module for outputting differentiated power allocation schemes according to the strategy function and controlling the heating power of each region.
[0008] Furthermore, the heat flux density calculation algorithm includes calculating the temperature gradient field, calculating the temperature gradient vector of each region based on the temperature difference and distance between adjacent temperature measurement points; calculating the heat flux density based on Fourier's law, calculating the heat flux density of each region based on the negative gradient relationship of Fourier's law; statistically analyzing the heat dissipation rate of each region using a region weighted average method, and using a dimensionless region weight coefficient to perform a weighted summation of the heat flux density to generate heat dissipation rate distribution data.
[0009] Furthermore, the reinforcement learning algorithm employs a deep Q-network model, including an input layer that receives a state vector containing heat dissipation distribution data, ambient temperature, target temperature, and current power allocation vector; a hidden processing layer that performs feature extraction and policy calculation on the state information through a multi-layer neural network, using the ReLU activation function; and an output layer that outputs the Q-value vectors of each region, representing the value assessment of different power allocation actions.
[0010] Furthermore, the components in the state vector are preprocessed for standardization: heat dissipation distribution data are normalized to a preset range based on the maximum value; temperature data are standardized to a set temperature range; and power data are standardized to rated power, in order to eliminate the influence of differences in the dimensions of different types of data.
[0011] Furthermore, the training method of the deep Q-network includes employing an experience replay mechanism, using offline batch training as the training mode, and using the Adam optimizer as the optimization strategy; the loss function is defined as mean squared error, calculated based on immediate reward, discount factor, and target network output; the immediate reward function is calculated based on energy efficiency and temperature uniformity, wherein energy efficiency is determined by the ratio of ideal power to actual power and the degree of temperature deviation, and temperature uniformity is determined by normalizing the temperature variance of each region.
[0012] Furthermore, the step of outputting differentiated power allocation values for each region includes selecting the optimal action index from the Q-value vector using a greedy strategy, selecting the action with the largest Q-value; decoding the action index into a specific power value; and converting the action value into a specific power allocation value within the power adjustment range of each region through a linear mapping.
[0013] Furthermore, it also includes optimization steps: performing time-series filtering on the heat dissipation distribution map, using the Kalman filter algorithm to eliminate random noise in temperature measurements, and generating a smooth heat dissipation characteristic curve. This algorithm recursively estimates the true heat dissipation distribution state through two steps: state prediction and measurement update.
[0014] Furthermore, it also includes establishing a multi-condition training database, collecting heat dissipation data and optimal power allocation schemes under different alloy composition ratios and environmental conditions, expanding the training sample range of reinforcement learning algorithms, and improving the adaptability of policy functions to complex operating conditions.
[0015] Furthermore, the multi-layer temperature dataset is organized according to different spatial locations and time series within the furnace, and includes three data layers: ambient temperature layer, furnace wall temperature layer, and internal temperature layer.
[0016] This invention solves the technical problem of energy waste caused by neglecting regional heat dissipation differences in traditional uniform heating methods by establishing a differentiated power allocation mechanism based on differences in heat dissipation characteristics, achieving significant technical results. Specifically, the heat flux density calculation algorithm can accurately quantify the differences in heat dissipation rates in each region, providing a scientific basis for differentiated heating; the reinforcement learning algorithm learns the mapping relationship between heat dissipation characteristics and optimal power allocation, generating an adaptive strategy function, realizing the technical transformation from fixed uniform allocation to dynamic differentiated allocation, effectively improving the energy utilization efficiency and temperature uniformity control accuracy of the NiP binary alloy smelting process. Attached Figure Description
[0017] Figure 1 This is a flowchart of the NiP binary alloy melting method of the present invention, which includes the complete process of acquiring multi-point temperature data in the furnace to generate a multi-layer temperature dataset, analyzing the temperature dataset based on the heat flux density calculation algorithm to output a heat dissipation distribution map, using reinforcement learning algorithm to process the heat dissipation distribution data to generate the optimal compensation strategy function, and inputting real-time operating parameters into the strategy function to output a differentiated power allocation scheme. Figure 2 This is a flowchart of the heat flux density calculation algorithm of the present invention, which shows in detail the specific steps of calculating the temperature gradient field, calculating the heat flux density based on Fourier's law, and statistically analyzing the heat dissipation rate distribution in each region. Figure 3 This is an optimization process flowchart of the present invention, which includes optimization steps such as performing time-series filtering on the heat dissipation distribution diagram and establishing a multi-condition training database. Detailed Implementation
[0018] In the NiP binary alloy smelting process, traditional uniform heating methods neglect the different heat dissipation characteristics of different areas within the smelting furnace. Because the heat dissipation conditions vary at different locations such as the furnace walls, bottom, and top, a uniform power distribution heating method can lead to overheating of colder areas that dissipate heat quickly to maintain the target temperature, while hotter areas that dissipate heat more slowly may overheat. This mismatch between heat dissipation differences and the heating method directly reduces the overall energy efficiency of the smelting process.
[0019] The method of this embodiment includes the following steps: Step 100: Acquire multi-point temperature data inside the furnace and generate a multi-layer temperature dataset. Ambient temperature, furnace wall temperature, and furnace interior temperature data are collected using temperature sensors. The temperature measurements from different locations are organized into a multi-layer temperature dataset according to spatial coordinates and time series.
[0020] Step 200: Analyze the temperature dataset based on the heat flux density calculation algorithm and output heat dissipation distribution maps for each region. Input the multi-layer temperature dataset into the heat flux density calculation algorithm to calculate the heat dissipation rate values for each region inside the furnace, and generate heat dissipation distribution maps describing the heat dissipation characteristics of different regions.
[0021] Step 300: Process the heat dissipation distribution data using a reinforcement learning algorithm to generate the optimal compensation policy function. Using the heat dissipation distribution map and corresponding operating parameters as training data, the optimal power compensation relationship for each region under different operating conditions is learned through a reinforcement learning algorithm, and the policy function is output.
[0022] Step 400: Input real-time operating parameters into the strategy function and output differentiated power allocation schemes for each region. Obtain current real-time operating parameters such as furnace temperature and ambient temperature, input these parameters into the strategy function generated in step 300, calculate and output differentiated power allocation values for each region. The strategy function selects the optimal action index from the Q-value vector using an ε-greedy strategy: Then the action index is decoded into a specific power value: ,in , These are the minimum and maximum power adjustment values for each region, respectively. The action value corresponding to the i-th region. This represents the maximum value of the action space.
[0023] It should be noted that the above heat flux density calculation algorithm includes the following sub-steps: Step 201: Calculate the temperature gradient field. Based on the temperature difference and distance between adjacent temperature measurement points, calculate the temperature gradient vector for each region. ,in This represents the temperature field function.
[0024] Step 202: Calculate the heat flux density based on Fourier's law, according to the formula... Calculate the heat flux density of each region. ,in is the thermal conductivity coefficient of the alloy.
[0025] Step 203: Statistically analyze the heat dissipation rate distribution in each region. Analyze the calculated heat flux density values according to spatial regions to generate heat dissipation rate distribution data.
[0026] The aforementioned heat flux density calculation algorithm is an improved calculation method based on spatial difference. The input of this algorithm is a multi-layer temperature dataset. The output is a heat dissipation distribution diagram. The algorithm first calculates the temperature gradient using the spatial difference method: ,in , The unit direction vector is used; then, the heat dissipation rate of each region is statistically analyzed using a region-weighted average method: ,in For regional indexes, These are dimensionless regional weighting coefficients. The heat flux density (unit: W / m²) is normalized to ensure that the weighting coefficients satisfy the following conditions. To ensure dimensional consistency, This represents the total number of regions.
[0027] It should be noted that the reinforcement learning algorithm described above uses a deep Q-network model, which includes the following layer structure: Input layer: Receives the state vector ,in This is the heat dissipation distribution data. For ambient temperature, For the target temperature, This is the current power allocation vector. To eliminate the influence of differences in the dimensions of different types of data, each component in the state vector is standardized and preprocessed: heat dissipation distribution data is normalized to the [0,1] interval according to the maximum value, temperature data is standardized according to the set temperature range, and power data is standardized according to the rated power.
[0028] First hidden layer: Fully connected layer, activation function is ReLU, output ,in This is the weight matrix. This is the bias vector.
[0029] Second hidden layer: Fully connected layer, activation function is ReLU, output .
[0030] Output layer: Fully connected layer, outputting the Q-value vectors of each region. ,in This indicates a power distribution action.
[0031] The training method for this deep Q-network employs an empirical replay mechanism, with offline batch training mode, and the Adam optimizer is used for optimization. The loss function is defined as mean squared error: ,in For network parameters, For instant rewards, As a discount factor, Output to the target network. The instantaneous reward function is based on energy efficiency. and temperature uniformity calculate: ,in , These are the weighting coefficients.
[0032] Energy efficiency The specific calculation method is as follows: First, calculate the deviation between the actual average temperature and the target temperature; then, calculate the ratio of total power consumption to ideal power consumption; finally, measure the efficiency using the reciprocal of this ratio. The specific form is... ,in The power required for ideal and uniform heating This represents the actual total power consumption. This represents the average actual temperature of each region.
[0033] Temperature uniformity The specific calculation method is as follows: First, calculate the temperature variance of all temperature measurement areas, and then perform normalization to eliminate the influence of temperature magnitude. The specific form is... ,in For regional indexes, For the first The actual temperature of each area The total number of regions is represented by this index, which reflects the dispersion of temperature distribution in each region. The smaller the value, the more uniform the temperature distribution.
[0034] Target network function The specific implementation is as follows: It adopts the same network structure as the main network, i.e., the input layer receives the state vector, features are extracted through two hidden layers, and the output layer generates the Q-value, but its parameters... Regularly update the main network parameters using a soft update method. The copied formula is updated as follows: ,in This is the update rate constant, which is usually set to a small value to maintain the stability of the target network, thereby providing a stable training objective.
[0035] In this embodiment of the application, in order to improve the accuracy of heat dissipation analysis, the following optimization steps are included in addition to step 200: Step 210: Perform time-series filtering on the heat dissipation distribution map. Use the Kalman filter algorithm to eliminate random noise in the temperature measurements, generating a smooth heat dissipation characteristic curve and improving the stability of the heat dissipation rate calculation. The input to the Kalman filter algorithm is the noisy heat dissipation distribution time-series data. The output is a filtered and smoothed heat dissipation distribution sequence. ,in For time indexing, The algorithm recursively estimates the true heat dissipation distribution state through two steps: state prediction and measurement update. The total number of time steps is denoted as .
[0036] In this embodiment of the application, in order to enhance the adaptability of the policy function, the following optimization steps are included in addition to step 300: Step 310: Establish a multi-condition training database, collect heat dissipation data and optimal power allocation schemes under different alloy composition ratios and environmental conditions, expand the training sample range of the reinforcement learning algorithm, and improve the adaptability of the policy function to complex conditions.
[0037] Technical effects of this embodiment: This implementation overcomes the technical deficiency of traditional uniform heating methods that ignore regional heat dissipation differences by establishing a differentiated power allocation mechanism based on differences in heat dissipation characteristics. Specifically, the heat flux density calculation algorithm can accurately quantify the differences in heat dissipation rates in each region, providing a scientific basis for differentiated heating; the reinforcement learning algorithm learns the mapping relationship between heat dissipation characteristics and optimal power allocation to generate an adaptive strategy function, realizing the technical transformation from fixed uniform allocation to dynamic differentiated allocation. Therefore, this method solves the energy waste problem caused by heat dissipation differences and improves the energy utilization efficiency of the NiP binary alloy smelting process.
[0038] Application scenario: A NiP binary alloy smelting production line of an electronic component manufacturing company. The smelting furnace is divided into 4 power control zones with a target temperature of 1200°C and an ambient temperature of 25°C.
[0039] Raw data collection: Table 1. Temperature sensor data (°C) collected in each area
[0040] Implementation process example: Step 100 execution result: The above temperature data is organized according to spatial coordinates to generate a multi-layer temperature dataset.
[0041]
[0042] Step 200 execution result: Based on the heat flux density calculation algorithm, the heat dissipation rate of each region is calculated. Taking region 1 as an example, the temperature gradient... K / m, heat flux density W / m².
[0043] Step 300 execution result: The deep Q-network processes the heat dissipation distribution data and generates a state vector. The Q-value vector is calculated and output through a neural network. .
[0044] Step 400 execution result: Select the optimal action through a greedy strategy and calculate the differentiated power allocation for each region.
[0045] Differential power allocation results: Table 2 Comparison of power allocation schemes in different regions;
[0046] Implementation results data: Table 3. Temperature distribution in each region after implementation;
[0047] Through the above implementation process, temperature uniformity is achieved. Compared to traditional uniform heating methods It increased by 75%.
[0048] Terminology definition: Multi-layer temperature dataset: refers to a temperature measurement data structure organized according to different spatial locations and time series within the furnace, including three data layers: ambient temperature layer, furnace wall temperature layer, and internal temperature layer.
[0049] Heat dissipation distribution map: This refers to a data map that describes the numerical distribution of heat dissipation rate in various areas of the furnace, and is used to characterize the differences in heat dissipation characteristics at different locations.
[0050] Policy function: refers to a mathematical function generated through reinforcement learning algorithms, which can calculate and output the optimal power allocation scheme for each region based on the input operating parameters.
[0051] Differentiated power allocation: This refers to a heating control method that allocates different power values to each area based on the different heat dissipation characteristics of each area, which is different from the traditional uniform power allocation method.
Claims
1. A method for smelting a NiP binary alloy, characterized in that, Includes the following steps: Acquire temperature data at multiple points inside the furnace. Collect ambient temperature, furnace wall temperature and furnace interior temperature data through temperature sensors. Organize the temperature measurement values at different locations into a multi-layer temperature dataset according to spatial coordinates and time series. The temperature dataset is analyzed based on the heat flux density calculation algorithm to calculate the heat dissipation rate of each region in the furnace and generate a heat dissipation distribution map describing the heat dissipation characteristics of different regions. The heat dissipation distribution data is processed using a reinforcement learning algorithm. The heat dissipation distribution map and the corresponding operating parameters are used as training data to learn the optimal power compensation relationship of each region under different operating conditions and generate a strategy function. Input the real-time operating parameters into the strategy function to obtain the current furnace temperature and ambient temperature real-time operating parameters, calculate and output the differentiated power allocation values for each region.
2. The method according to claim 1, characterized in that, The heat flux density calculation algorithm includes: Calculate the temperature gradient field by calculating the temperature gradient vector of each region based on the temperature difference and distance between adjacent temperature measurement points. Heat flux density is calculated based on Fourier's law, and the heat flux density of each region is calculated based on the negative gradient relationship of Fourier's law. The heat dissipation rate of each region is statistically analyzed by using a region-weighted average method. The heat flux density is then weighted and summed using a dimensionless region weight coefficient to generate heat dissipation rate distribution data.
3. The method according to claim 1, characterized in that, The reinforcement learning algorithm employs a deep Q-network model, including: The input layer receives a state vector containing heat dissipation distribution data, ambient temperature, target temperature, and current power allocation vector. The hidden processing layer performs feature extraction and policy calculation on the state information through a multi-layer neural network, using the ReLU activation function; The output layer outputs the Q-value vectors for each region, representing the value assessment of different power allocation actions.
4. The method according to claim 3, characterized in that, Perform standardization preprocessing on each component of the state vector: The heat dissipation distribution data is normalized to a preset range based on the maximum value. Temperature data is standardized according to the set temperature range; Power data is standardized to rated power to eliminate the impact of differences in the units of different data types.
5. The method according to claim 3, characterized in that, The training method for the deep Q-network includes: An experience replay mechanism is adopted, the training mode is offline batch training, and the optimization strategy uses the Adam optimizer; The loss function is defined as mean squared error and is calculated based on immediate reward, discount factor, and target network output; The instant reward function is calculated based on energy efficiency and temperature uniformity. Energy efficiency is determined by the ratio of ideal power to actual power and the degree of temperature deviation, while temperature uniformity is determined by normalizing the temperature variance of each region.
6. The method according to claim 1, characterized in that, The steps for outputting the differentiated power allocation values for each region include: The optimal action index is selected from the Q-value vector using a greedy strategy, choosing the action with the maximum Q-value. The action index is decoded into a specific power value, and the action value is converted into a specific power allocation value within the power adjustment range of each region through linear mapping.
7. The method according to claim 1, characterized in that, It also includes optimization steps: The heat dissipation distribution map is processed by time-series filtering, and the Kalman filter algorithm is used to eliminate random noise in temperature measurement and generate a smooth heat dissipation characteristic curve. The algorithm recursively estimates the true heat dissipation distribution state through two steps: state prediction and measurement update.
8. The method according to claim 1, characterized in that, Also includes: Establish a multi-condition training database to collect heat dissipation data and optimal power allocation schemes under different alloy composition ratios and environmental conditions, expand the training sample range of reinforcement learning algorithms, and improve the adaptability of policy functions to complex operating conditions.
9. The method according to claim 1, characterized in that, The multi-layer temperature dataset is organized according to different spatial locations and time series within the furnace, and includes three data layers: ambient temperature layer, furnace wall temperature layer, and internal temperature layer.
10. A system for performing the NiP binary alloy smelting method according to any one of claims 1 to 9, characterized in that, include: Temperature sensor module, used to collect ambient temperature, furnace wall temperature and furnace internal temperature data; The data processing module is used to organize multi-layer temperature datasets and execute heat flux density calculation algorithms; The reinforcement learning module is used to process heat dissipation distribution data and generate policy functions. The power control module is used to output differentiated power allocation schemes based on the strategy function and control the heating power of each area.