A deep learning-based defatting intelligent temperature control method

Abstract

The application discloses a kind of based on deep learning's defatting intelligent temperature control method, comprising: A1: acquisition furnace multiple area temperature time series data, blank three-dimensional point cloud data, blank arrangement layout image, forming agent component mole fraction and respectively pre-process;A2: calculate furnace temperature hidden state feature and wall thickness geometric feature, adjacent topological feature and component coding feature, fusion coupling feature;A3: calculate temperature stage constraint embedding feature and temperature stage boundary matrix, decoder initial hidden state and stage context vector, temperature rising rate sequence and holding time length sequence;A4: calculate continuous temperature control curve;A5: continuous temperature control curve is discretized as temperature set value sequence and is issued to each temperature zone proportional-integral-derivative temperature controller by cycle, and loop control is executed.The application can solve the problem that traditional fixed constant temperature process is difficult to adapt to wall thickness difference and forming agent multi-component decomposition characteristics, and process adaptation cost is relatively high.

Description

Technical Field

[0001] This invention relates to the field of powder molding technology, and in particular to a deep learning-based intelligent temperature control method for degreasing. Background Technology

[0002] In both metal powder forming and ceramic forming processes, the debinding process lies between green body forming and high-temperature sintering. Its core task is to fully remove the organic binder system from the gaps between powder particles while maintaining the geometric integrity of the green body. However, differences in green body wall thickness lead to uneven heat conduction paths from the surface to the core. Furthermore, the forming agent is typically composed of multiple components such as paraffin, polyethylene, and stearic acid, each with a wide range of thermal decomposition temperature ranges and varying decomposition rate characteristics. This is particularly problematic in mass production across different wall thicknesses and material systems. Metal green bodies have high thermal conductivity and rapid heat transfer, making thin-walled areas prone to rapid heating and instantaneous vaporization and cracking of low-boiling-point binders. In contrast, ceramic green bodies have lower thermal conductivity, and heat conduction in the thick-walled core is delayed. Under the same isothermal regime, the decomposition cycle of high-boiling-point forming agents is significantly prolonged, making internal carbon residue black spots more likely. The two types of green bodies have significantly different requirements for temperature control regarding heating slope and holding window, making a uniform, fixed process unsuitable for both types of products.

[0003] Currently, the industry commonly employs an experience-based segmented isothermal process for the thermal debinding stage. This involves operators determining a fixed stepped temperature control curve through multiple orthogonal firing trials for a specific molding agent formulation. The heating rate and holding time for each temperature step are not adaptively adjustable. However, this fixed process struggles to adapt to the spatial gradient of the green body wall thickness. Thin-walled regions reach the decomposition temperature of low-boiling-point components first, while the thick-walled core remains unheated. Rapid temperature rise can lead to sudden vaporization and volume expansion of low-boiling-point components, causing the green body to crack. Simultaneously, the uniformly set holding time often fails to fully remove high-boiling-point components from the thick-walled regions, resulting in carbonization and precipitation of residues during subsequent sintering, forming black spots on the product surface. When companies change molding agent formulations or switch to green body products with different geometric characteristics, the aforementioned segmented isothermal curve must be repeatedly adjusted based on manual experience, leading to significant time and material costs for process adaptation.

[0004] In recent years, although a few studies have used shallow machine learning models such as BP neural networks and support vector regression to predict temperature control curves based on single-dimensional inputs such as furnace temperature or molding agent type, these methods have limited ability to fuse multi-source heterogeneous operating data, and the models' generalization performance on variable wall thickness products and novel molding agent systems is insufficient. The Transformer architecture has shown strong multimodal feature interaction capabilities in sequence modeling, but it still lacks mature application solutions in the specific technical field of degreasing temperature control. In particular, how to organically integrate multi-dimensional information such as furnace temperature field distribution, billet geometric parameters, furnace packing density, and molding agent type and ratio to accurately capture the nonlinear coupling relationship between wall thickness, heat conduction, and multi-group decomposition time sequence remains an unsolved technical problem. Summary of the Invention

[0005] In view of this, the present invention aims to provide a deep learning-based intelligent temperature control method for degreasing, in order to solve the problems that traditional fixed constant temperature processes are difficult to adapt to differences in wall thickness and the multi-component decomposition characteristics of molding agents, and the process adaptation cost is relatively high.

[0006] A deep learning-based intelligent temperature control method for degreasing includes: A1: Collect time-series temperature data of multiple zones in the furnace, 3D point cloud data of the billet, image of the billet arrangement, and molar fraction of the molding agent components, and preprocess them respectively to obtain preprocessed furnace temperature time-series data, preprocessed point cloud data with wall thickness attributes, preprocessed billet spatial adjacency map, preprocessed component molar fraction data, and component decomposition temperature parameters. A2: Based on the preprocessed furnace temperature time series data, the preprocessed point cloud data with wall thickness attributes, the preprocessed billet spatial adjacency graph, and the preprocessed component mole fraction data, calculate the furnace temperature latent features and wall thickness geometric features, adjacency topology features and component coding features, and fusion coupling features. A3: Based on the fusion coupling characteristics and component decomposition temperature parameters, calculate the temperature level constraint embedding features and temperature level boundary matrix, decoder initial hidden state and stage context vector, heating rate sequence and holding time sequence in sequence. A4: Based on the heating rate sequence, the holding time sequence, and the fusion coupling characteristics, calculate the temperature step curve segment sequence after smoothing projection, the temperature step curve segment sequence after verification and correction, and the continuous temperature control curve in sequence. A5: Discretize the continuous temperature control curve into a temperature setpoint sequence and send it to the proportional-integral-derivative temperature controller of each temperature zone cycle by cycle. Collect the actual temperature, calculate the deviation, generate the control quantity, and execute closed-loop control cyclically until the holding time of all temperature levels is accumulated, thus completing the temperature control.

[0007] Furthermore, step A1 also includes: A11: Temperature time-series data of multiple zones in the furnace is collected by thermocouple sensor arrays deployed in the industrial field. The data type is multi-dimensional time-series numerical data, including temperature sampling values, temperature setpoints and timestamps of each temperature zone. Outliers are filtered out by a three-fold standard deviation constant removal method for the temperature time-series data of the furnace, missing sampling points are filled in by linear interpolation, and Z-score standardization is used to map the temperature values ​​of each temperature zone to a standard distribution with a mean of 0 and a variance of 1, to obtain the preprocessed furnace temperature time-series data. A12: Obtain the three-dimensional point cloud data of the billet through the CAD model analysis system, and match the corresponding wall thickness attribute value for each point cloud vertex; filter out scanning noise by using the statistical outlier removal method on the three-dimensional point cloud data of the billet, and perform point cloud sparsification processing by using the voxel mesh downsampling method to obtain sparsified point cloud data; perform Z-score standardization on the wall thickness attribute value of each point cloud vertex in the sparsified point cloud data, and map the wall thickness value to a standard distribution with a mean of 0 and a variance of 1 to obtain preprocessed point cloud data with wall thickness attribute; A13: Images of the billet arrangement layout inside the furnace are acquired using an industrial area array camera; the billet arrangement layout images are denoised using Gaussian filtering, and then each billet in the billet arrangement layout image is segmented at the pixel level using an instance segmentation algorithm, and the center coordinates and circumscribed bounding box of each billet are extracted; a preset adjacency distance threshold is set, and the Euclidean distance between the center coordinates of any two billets is calculated. When the distance is less than or equal to the adjacency distance threshold, it is determined that the two billets have a spatial adjacency relationship; an undirected graph of billet arrangement is constructed with a single billet as a graph node and the spatial adjacency relationship between billets as graph edges; an adjacency matrix and a node attribute matrix are generated with the center coordinates and circumscribed bounding box of each billet as node attributes, to obtain the preprocessed billet spatial adjacency graph; A14: Collect the molar fraction of the molding agent components, and then use the component type identifier as the query key to integrate the decomposition start temperature, decomposition peak temperature and decomposition termination temperature corresponding to the component type identifier with the component molar fraction to obtain the preprocessed component molar fraction data and component decomposition temperature parameters.

[0008] Furthermore, step A2 also includes: A21: Based on the preprocessed furnace temperature time series data and the preprocessed point cloud data with wall thickness attributes, perform time series dynamic coding and spatial geometric representation extraction. First, calculate the latent features of furnace temperature; then, based on the billet material, select to calculate the geometric features of metal billet wall thickness or non-metal billet wall thickness. A22: Based on the preprocessed billet spatial adjacency graph and the preprocessed component mole fraction data, graph structure encoding and component attribute embedding are performed to obtain adjacency topology features and component encoding features; A23: Based on the latent characteristics of furnace temperature, geometric characteristics of metal billet wall thickness, geometric characteristics of non-metal billet wall thickness, adjacency topology characteristics, and component coding characteristics, branch gating and heterogeneous features are deeply fused to obtain fused coupling features.

[0009] Furthermore, step A2 also includes: The calculation process of the furnace temperature latent features in step A21 is as follows: First, the preprocessed furnace temperature time series data is processed by a bidirectional long short-term memory network and a linear layer projection process for extracting the furnace temperature latent features. Then, the two processing results are added together, and then the layer normalization operation is performed. Finally, the data is processed by a multilayer perceptron for extracting the furnace temperature latent features to obtain the furnace temperature latent features. When the billet material is metal, the calculation process of the geometric features of the metal billet wall thickness is as follows: the pre-processed point cloud data with wall thickness attributes is first subjected to ball query grouping operation, then the farthest point sampling operation is performed, and then processed by a multi-level point cloud feature extraction network to obtain the geometric features of the metal billet wall thickness. When the material of the blank is non-metallic, the calculation process of the geometric features of the wall thickness of the non-metallic blank is as follows: the pre-processed point cloud data with wall thickness attributes is first processed by K-nearest neighbor mapping, then processed by dynamic graph convolutional neural network, then processed by layer normalization, and then processed by multilayer perceptron for geometric feature extraction to obtain the geometric features of the wall thickness of the non-metallic blank. The calculation process of adjacency topology features and component coding features in step A22 is as follows: First, the preprocessed billet space adjacency graph is processed by graph convolution operation and linear projection layer for graph structure extraction, respectively. Then, the two processing results are added and processed by graph attention network operation, then by layer normalization operation, and then by multilayer perceptron for adjacency topology feature extraction to obtain adjacency topology features. The preprocessed component mole fraction data is processed by embedding layer, then by self-attention operation, then added to the output of embedding layer, then by layer normalization operation, and then by multilayer perceptron for component feature extraction to obtain component coding features. The calculation process of the fusion coupling feature in step A23 is as follows: Based on the wall thickness geometric features, adjacency topology features, and component coding features, a concatenation operation is performed according to the feature dimensions. This concatenation is then processed by a multilayer perceptron for adaptive gating generation, followed by a Sigmoid activation function to obtain the metal-nonmetal adaptive gating weights. A cross-attention operation is performed on the furnace temperature latent features and the concatenated features obtained from the wall thickness geometric features, adjacency topology features, and component coding features. This cross-attention operation is then performed by a multilayer perceptron for fusion coupling feature extraction to obtain the first intermediate result. A self-attention operation is performed on the concatenated features obtained from the wall thickness geometric features, adjacency topology features, and component coding features. This self-attention operation is then performed by a multilayer perceptron for fusion coupling feature extraction to obtain the second intermediate result. The metal-nonmetal adaptive gating weights are multiplied by the first intermediate result, and the result after subtracting the metal-nonmetal adaptive gating weights from 1 is multiplied by the second intermediate result. The two Hadamard product results are then added together to obtain the fusion coupling feature.

[0010] It should be further explained that the sintering quality of the billet is determined by four heterogeneous factors: the influence of the wall thickness distribution of the billet itself on heat conduction, the spatial arrangement of the billet in the furnace, and the decomposition characteristics of each component of the forming agent. Metal billets are mainly affected by heat conduction, and the wall thickness determines the heat penetration rate, while non-metal billets are mainly affected by heat radiation and convection, and the surface geometry has a greater impact. When the two types of billets are fired together in the same furnace, the same temperature control curve will inevitably lead to one type being under-fired and the other being over-fired. There is a lack of a unified feature expression method that can automatically sense material differences and integrate multi-source heterogeneous information. For the extraction of latent furnace temperature features and wall thickness geometric features, preprocessed furnace temperature time-series data is used as input. A bidirectional long short-term memory network is used to capture the dependencies between the heating and cooling phases—linking the causes and consequences of temperature changes temporally to avoid treating temperatures at different time points as isolated samples. Simultaneously, linear projection residual connections are introduced to preserve the original magnitude information, enabling subsequent multilayer perceptron compression operations to accurately model the heat accumulation process based on the original amplitude, thus obtaining the latent furnace temperature features. These features not only reflect the trend of temperature changes but also incorporate the hysteresis effect caused by thermal inertia—providing a physical reference for predicting actual temperature changes inside the billet. Meanwhile, for the extraction of wall thickness spatial geometric features, different processing branches are selected based on the billet material. For metal billets, a multi-level point cloud feature extraction network is used. Representative spatial points are first selected by sampling the farthest point, and then the spatial geometric features of the wall thickness distribution are extracted layer by layer by using sphere query grouping. This captures the continuous gradient of the wall thickness from thin to thick, so that the spatial geometric representation matches the physical heat transfer mechanism of metal, which is mainly based on thermal conduction. For non-metal billets, a dynamic graph convolutional network is used. The response pattern of the local surface geometry to external convective heat transfer is captured by K-nearest neighbor mapping. The convective heat transfer coefficients of the surface concave and convective parts and corners are much different. K-nearest neighbor mapping can identify these subtle morphological differences in the local neighborhood, so that the geometric features correspond to the response characteristics of thermal radiation and convective heat transfer. This distinguishes the different thermal response tendencies of the two types of materials at the feature level, providing a targeted intermediate representation for subsequent heterogeneous fusion. For the extraction of adjacency topology features and component coding features, the preprocessed billet spatial adjacency graph is used as input. First, graph convolution is performed on the spatial arrangement of the billets to extract the arrangement pattern in the local neighborhood. Then, a graph attention network is superimposed to assign differentiated attention weights to billets at different distances. This step can automatically identify the heat shading effect caused by the billet spacing being too close and the heat dissipation bottleneck in the densely arranged area, thus obtaining adjacency topology features. At the same time, the preprocessed component mole fraction data is used as input. First, discrete components are mapped to a continuous semantic space through an embedding layer. Then, the proportion dependence between different molding agent components is captured through a self-attention operation. The volatilization temperature and decomposition rate of the components differ greatly. The self-attention mechanism can identify the cooperative or competitive relationship between each component in the firing process in a global range, thus obtaining component coding features. This process transforms the billet spatial arrangement information and component proportion information, which could not be directly incorporated into a unified representation, into structured adjacency topology features and component coding features. This prepares an effective input for the deep fusion of heterogeneous features and enables the model to use the physical factor of billet spacing to predict heat conduction bottlenecks. For deep fusion and unified representation of heterogeneous features, the latent features of furnace temperature, geometric features of wall thickness, adjacency topology features, and component coding features are concatenated along their feature dimensions. This concatenation is first processed by a multilayer perceptron for adaptive gating generation, and then by a sigmoid activation function to obtain a metal-nonmetal adaptive gating weight between 0 and 1. This weight can automatically determine whether the current billet is more metal- or non-metallic in its thermal response based on the wall thickness geometry, spatial arrangement, and component ratio. The gating weight, along with the results of cross-attention processing of the latent features of furnace temperature and the geometric-topology-component concatenation features, undergoes a Hadamard product operation. Simultaneously, 1 is subtracted from the result of the gating weight and the geometric-topology-component encoding feature. The results of the spliced ​​features after self-attention processing are subjected to Hadamard product operation, and the two are then added element-by-element to obtain the fused coupled features. When processing billets with a metallic tendency, cross-attention enables furnace temperature time-series features as the dominant information and geometry-topology-composition features as auxiliary information to conduct cross-modal interaction. When processing billets with a non-metallic tendency, self-attention autonomously aggregates the synergistic relationship between surface morphology, arrangement position and composition decomposition characteristics in the geometry-topology-composition joint space. This unifies the four types of heterogeneous information, namely material difference, wall thickness distribution, arrangement position and composition ratio, into a single fused coupled feature tensor, providing a unified feature that fully expresses the heating state of the billet for subsequent process control. In existing technologies, the degreasing and firing process often uses a single temperature control curve combined with manual experience for adjustment. This approach essentially assumes that all billets have uniform thermal response characteristics within the furnace. However, for metallic billets, heat conduction is dominant and wall thickness determines the heat penetration rate, while for non-metallic billets, heat radiation and convection are dominant and surface geometry has a greater impact. When both types of billets are fired together in the same furnace, the same temperature control curve inevitably leads to one type being underfired and the other overfired. Furthermore, it fails to perceive differences in the spatial arrangement of the billets within the furnace—when adjacent billets are too close together, they block each other's heat, causing localized temperature rise to lag; when sparsely arranged, heat dissipates too quickly, causing the temperature to drop. These spatial factors can only be judged visually by the operator in traditional empirical methods. This invention utilizes a bidirectional long short-term memory network and linear projection residual connections. The dependence of heating and cooling in the furnace temperature time series and its original magnitude are captured to ensure the ability of time series features to represent thermal inertia effects. Then, according to the metal and non-metal materials, multi-level point cloud networks and dynamic graph convolutional networks are used to encode the wall thickness geometry, distinguishing two different thermal response physical mechanisms at the feature level. Subsequently, the spatial topological information such as billet arrangement density and spacing and the ratio information of component self-attention encoding are extracted in parallel through graph convolution and graph attention networks. Finally, an adaptive gating mechanism is used to automatically determine the material tendency according to the geometry-topology-component features of the current billet, and the cross-attention path dominated by furnace temperature and the geometry-topology-component self-aggregation path are softly weighted and fused, so that the same network structure can cope with different scenarios of metal and non-metal.

[0011] Furthermore, step A3 also includes: A31: Based on the fusion coupling characteristics and component decomposition temperature parameters, perform thermal property constraint encoding and temperature order boundary embedding to obtain temperature order constraint embedding characteristics and temperature order boundary matrix; A32: Based on the temperature level constraint embedding features, temperature level boundary matrix and fusion coupling features, perform constraint gating initialization and multi-stage decoding state generation to obtain the decoder's initial hidden state and stage context vector; A33: Based on the initial hidden state of the decoder, the stage context vector, and the temperature order boundary matrix, perform autoregressive temperature order sequence decoding with boundary constraints to obtain the heating rate sequence and the holding time sequence.

[0012] Furthermore, step A3 also includes: The calculation process of the temperature-order constraint embedding feature and the temperature-order boundary matrix in step A31 includes: firstly, embedding the component decomposition temperature parameters, then processing them by a first constraint-encoding multilayer perceptron after layer normalization; simultaneously, performing a linear projection of the fused coupling features for temperature-order constraints; performing cross-attention on the results of the above two processes, then adding the embedding result of the component decomposition temperature parameters, and then processing them sequentially by layer normalization and a second constraint-encoding multilayer perceptron to obtain the temperature-order constraint embedding feature; concatenating the temperature-order constraint embedding feature with the result of performing a linear projection of the fused coupling feature for generating the temperature-order boundary matrix, then processing it by a multilayer perceptron for generating the temperature-order boundary matrix, and finally processing it by the Sigmoid function to obtain the temperature-order boundary matrix; The calculation process of the decoder's initial hidden state and stage context vector in step A32 includes: concatenating the fused coupling feature with the temperature-order constraint embedding feature, processing it through a first hidden state multilayer perceptron after layer normalization, adding the result of residual linear projection on the fused coupling feature, and then processing it through a second hidden state multilayer perceptron in sequence after layer normalization to obtain the decoder's initial hidden state; concatenating the temperature-order constraint embedding feature with the fused coupling feature, processing it first through a first context multilayer perceptron, then processing it through a second context multilayer perceptron after layer normalization, and performing a Hadamard product with the temperature-order boundary matrix to obtain the stage context vector; The decoding process of the heating rate sequence and holding time sequence in step A33 includes: for each time step, concatenating the position encoding, stage context vector, and boundary vector corresponding to the current time step in the temperature order boundary matrix with the result of the Hadamard product of the decoded hidden state of the previous time step, and inputting the concatenated result and the decoded hidden state of the previous time step into the gated recurrent unit to obtain the decoded hidden state of the current time step; after layer normalization, the decoded hidden state of the current time step is processed by the first output multilayer perceptron, and the result of linear projection of the decoded hidden state of the current time step for the residual is added, and then processed by the second output multilayer perceptron in sequence, and then the heating rate and holding time of the current time step are obtained by channel segmentation operation; after decoding, the heating rate and holding time of each time step are concatenated in chronological order to obtain the heating rate sequence and the holding time sequence.

[0013] It is important to further explain that the decomposition temperature of each component of the molding agent is a rigid physical boundary—a specific component decomposes only within a certain temperature range. If the heating rate is too fast, the decomposition gases will not have enough time to diffuse and escape from the interior of the green body, causing micro-cracks or even cracks to form under gas pressure. Insufficient holding time will result in incomplete decomposition of the component, and the residual carbon elements will be transformed into carbide defects that are difficult to remove in the subsequent sintering stage, affecting the density and mechanical properties of the final product. Due to fluctuations in the source of raw materials, the molar fraction of each component of the molding agent varies from batch to batch. The wall thickness and arrangement of the green body will also vary with the production batch, making it difficult for the same set of manually defined temperature step parameters to always fall within the safe window of the component decomposition temperature in different batches. In addition, a single degreasing process often includes three to five temperature steps. The heating rate and holding time of each temperature step need to be aligned with the decomposition temperature range of the specific component. Manually adjusting the parameters one by one among thousands of possible temperature step combinations is both time-consuming and difficult to guarantee the global optimum. This invention uses component decomposition temperature parameters and fusion coupling features as inputs—the former carrying rigid boundary information such as decomposition start temperature, peak temperature, and termination temperature, and the latter carrying the thermal response capability of the billet in the current batch, including the billet's thermal conductivity, gas permeability, and thermal hysteresis effect caused by geometric arrangement. First, a first-constraint-coded multilayer perceptron is used to perform layer normalization and linear embedding of the component decomposition temperature parameters. Then, a cross-attention mechanism is used to perform cross-modal interaction between the decomposition temperature embedding and the linear projection of the fusion coupling features, aligning the decomposition temperature boundary information and thermal response features in the semantic space. The residual connection retains the original decomposition temperature embedding, and after being processed by a second-constraint-coded multilayer perceptron and layer normalization, the output temperature level constraints are determined. Embedded features; simultaneously, the temperature level constraint embedded features are concatenated with another linear projection of the fused coupling features, and after processing by a multilayer perceptron and a sigmoid function, a temperature level boundary matrix is ​​generated. Each element of this matrix corresponds to the heating / holding mode switching probability of a time step, marking which time steps are in the rapid heating safety window and which time steps must switch to the holding stage. This transforms rigid physical boundaries into soft constraints that can be referenced step by step, avoiding the problem of boundary ambiguity under batch fluctuations of a single temperature threshold. The temperature level boundary matrix suppresses physically meaningless decoding directions in the feature space from the first sub-step, so that the subsequently generated temperature level sequence is strictly constrained by the component decomposition temperature range from the beginning. After obtaining the temperature-order constraint embedding features and the temperature-order boundary matrix, they are concatenated with the fused coupling features: The first hidden-state multilayer perceptron performs layer normalization and feature interaction; the residual connection path retains the original fused coupling features through a residual linear projection layer; and then, after passing through the second hidden-state multilayer perceptron and layer normalization, the initial hidden state of the decoder is obtained—this state simultaneously integrates global thermal response information and component decomposition temperature constraint information, providing a physically semantically correct starting point for autoregressive decoding. Simultaneously, the temperature-order constraint embedding features and the fused coupling features are concatenated, processed by the first context multilayer perceptron, and then the processing result is gated using the temperature-order boundary matrix—the boundary matrix... Time steps with a probability close to 1 retain complete contextual information, while time steps with a probability close to 0 are suppressed. After passing through a second contextual multilayer perceptron and layer normalization, a stage context vector is output. This vector retains only information within the allowable range of the current temperature level. Features exceeding the decomposition temperature range of the sound-emitting temperature level are filtered out by a gating mechanism, so that the decoder automatically focuses on the feature region related to the decomposition of the components at each temperature level switch. When the decoder enters the decomposition temperature level of low-boiling-point components, the thermal response features corresponding to high-boiling-point components are effectively suppressed, thereby avoiding the residual information of the previous temperature level from interfering with the temperature control parameter setting of the current stage, so that the decoder focuses on matching the decomposition rate and exhaust demand of the current temperature level. After the initial hidden state and stage context vector of the decoder are ready, an autoregressive temperature sequence decoding is performed using a gated recurrent unit. At each time step, the previous hidden state is gated with the boundary vector of the corresponding time step in the temperature boundary matrix using a Hadamard product. This gated state is then concatenated with the position code and stage context vector as input to the gated recurrent unit, generating the decoded hidden state for the current time step. This hidden state is then processed by a two-layer residual multilayer perceptron and channel segmentation, synchronously outputting the heating rate and holding time for that time step. The boundary matrix applies progressive constraints to the hidden state at each time step, ensuring that the decoding result does not exceed the safe boundary of the component decomposition temperature—when decoding progresses to a time step approaching the peak temperature of a component decomposition temperature. At that time, the boundary matrix sets the probability of switching the heating mode to zero and the probability of the holding mode to one, forcing the decoder to output the holding time instead of the heating rate at that time step. After decoding, the heating rate and holding time of each time step are spliced ​​together in chronological order to obtain a complete heating rate sequence and holding time sequence. This autoregressive mechanism makes the parameter values ​​of each time step in the temperature step sequence form a causal chain with the state of the previous time step, ensuring the physical continuity of the heating rate and holding time in the time dimension. The degree of residual decomposition gas discharge in the previous temperature step will naturally affect the heating permit conditions of the next temperature step, avoiding the problem of the parameters of each stage being isolated and unable to be adjusted in linkage when statically dividing the temperature steps. In existing technologies, process engineers typically manually divide temperature levels based on the decomposition temperature range of each component and set the heating rate and holding time for each level. This method is highly sensitive to fluctuations in the mole fraction of components between batches. Once the actual concentration of a component deviates from the nominal value, the manually set fixed temperature level may fall into the dangerous region of incomplete decomposition or excessively rapid heating. Furthermore, process engineers cannot quickly locate the optimal parameters among thousands of possible temperature level combinations. This invention uses the component decomposition temperature parameter as a rigid constraint and performs cross-modal semantic alignment with the fusion coupling feature through cross-attention to generate a temperature level boundary matrix as a physical constraint signal for each time step. During the autoregressive decoding process, the boundary matrix performs Hadamard product gating on each hidden state, so that the heating rate and holding time are constrained by the current temperature level decomposition temperature range at each time step. The rigid physical boundary is embedded into each step of parameter generation inference, rather than being truncated only at the final output. This allows the decoding result to adaptively adjust the temperature level division position and duration. Concentration fluctuations will drive the boundary matrix to redistribute the temperature level switching probability through changes in the component decomposition temperature parameter, thereby automatically matching the safe process window for the new batch.

[0014] Furthermore, step A4 also includes: A41: Based on the heating rate sequence, holding time sequence and fusion coupling characteristics, the temperature step curve segments are parametrically reconstructed and heat conduction is smoothly projected to obtain the preliminary temperature step curve segment sequence. A42: Based on the temperature level curve segment sequence after smooth projection and the fusion coupling characteristics, perform energy balance consistency verification and curve segment adaptive correction to obtain the verified and corrected temperature level curve segment sequence. A43: Based on the verified and corrected temperature level curve segment sequence, perform cross-temperature level continuous splicing and global temperature curve reconstruction to obtain a continuous temperature control curve.

[0015] Furthermore, step A4 also includes: Based on the heating rate sequence, holding time sequence, and fusion coupling features, parameterized reconstruction and heat conduction smoothing projection of temperature level curve segments are performed to obtain an initial temperature level curve segment sequence. Specifically, the heating rate of the current temperature level in the heating rate sequence, the holding time of the current temperature level in the holding time sequence, and the fusion coupling features are concatenated after processing by a linear projection layer for context. This concatenation is then processed by a first curve segment multilayer perceptron, with the addition of a position encoding function for the current temperature level, followed by layer normalization. Finally, it is processed by a second curve segment multilayer perceptron to obtain the initial curve segment of the current temperature level. The initial curve segment is concatenated with the results obtained from the first-order difference operation of the curve segments of the previous temperature level and the current temperature level. The concatenation is then processed by a gated weighted multilayer perceptron and then by a Sigmoid function to obtain the smoothing gate weight of the current temperature level. The smoothing gate weight of the current temperature level is then multiplied by the Hadamard product of the initial curve segment of the current temperature level. At the same time, the smoothing gate weight of the current temperature level is subtracted from 1 and multiplied by the Hadamard product of the results obtained from the thermal diffusion smoothing operation of the initial curve segments of the previous temperature level and the current temperature level. The two Hadamard product results are then added together to obtain the smoothed projected curve segment of the current temperature level. Based on the smoothed projected temperature level curve segment sequence and the fused coupling features, energy balance consistency verification and curve segment adaptive correction are performed to obtain the verified and corrected temperature level curve segment sequence. Specifically, the smoothed projected current temperature level curve segment is processed by a fourth constraint-coded multilayer perceptron to obtain the first intermediate feature. Simultaneously, the fused coupling features are processed by a linear projection layer used as the query key to obtain the second intermediate feature. The first and second intermediate features are subjected to cross-attention operation, then layer normalization, and finally added to the result obtained by the smoothed projected current temperature level curve segment processed by the fourth constraint-coded multilayer perceptron. Then, the result is processed by a third constraint-coded multilayer perceptron to obtain... The energy deviation characteristics of the current temperature level are obtained; the energy deviation characteristics of the current temperature level are concatenated with the smoothed projected current temperature level curve segment, processed by a gated weighted multilayer perceptron, and then passed through a Sigmoid function to obtain the adaptive correction intensity coefficient of the current temperature level; the energy deviation characteristics of the current temperature level are concatenated with the first-order difference operation result of the curve segment between the smoothed projected current temperature level curve segment and the smoothed projected temperature level curve segment of the previous temperature level, processed by a curve segment incremental multilayer perceptron, and then Hadamard product is performed with the adaptive correction intensity coefficient of the current temperature level. This product is then added to the smoothed projected current temperature level curve segment to obtain the current temperature level curve segment after verification and correction. Based on the verified and corrected temperature level curve segment sequence, continuous splicing across temperature levels and global temperature curve reconstruction are performed to obtain a continuous temperature control curve. Specifically, for each temperature level and each time step, the joint position encoding function and indicator function of the current time step and the current temperature level are spliced ​​together. After processing by a splicing weight multilayer perceptron, the Softmax function is calculated along the temperature level dimension to obtain the splicing weight of the current temperature level curve segment at the current time step. The indicator function is used to indicate whether the current time step belongs to the interval between the start timestamp of the current temperature level and the start timestamp of the next temperature level. For each temperature level from the first temperature level to the last temperature level, the value of the current temperature level curve segment at the current time step minus the start timestamp of the current temperature level is multiplied by the splicing weight of the current temperature level curve segment at the current time step. The product results of all temperature levels are accumulated to obtain the temperature value of the continuous temperature control curve at the current time step.

[0016] It should be further explained that step A41 of this invention takes the heating rate sequence, the holding time sequence, and the fusion coupling feature as inputs—this feature carries the physical state information of the entire sintering stage of the billet; firstly, the heating rate and holding time corresponding to each temperature level are spliced ​​with the linear projection of the fusion feature, and the position code of the temperature level is superimposed to enhance the ability to distinguish the temperature level number. Then, the initial curve segment of the temperature level is parameterized and reconstructed by a double-layer curve segment multilayer perceptron—transforming the discrete parameters into a curve segment shape with a continuous time structure, providing a numerical space for subsequent fine adjustment; then, the first-order difference between the initial curve segments of adjacent temperature levels is calculated—reflecting the drastic change in the curve shape, and the input gate weight multilayer perceptron is processed by the Sigmoid function to obtain the smooth gate weight, whose physical meaning is the difference between the current temperature level and the previous temperature level. The required degree of thermal diffusion smoothing between temperature levels: When adjacent temperature levels change drastically, the gating weights tend to use thermal diffusion smoothing—the curve shape of the previous temperature level is naturally extended to the current temperature level through thermal conduction diffusion, so that the temperature gradient at the temperature level boundary is controlled within the range acceptable to the furnace's thermal inertia performance, avoiding large adjustments to the controller triggered by temperature jumps; when adjacent temperature levels change gradually, the gating weights tend to retain the initial curve segment to maintain the original response speed, so that the smoothing process no longer relies on fixed window filtering, but is adaptively adjusted according to the operating conditions—in temperature levels with drastic transitions between heating and holding, the smoothing projection effectively avoids frequent valve actions caused by setpoint jumps, thereby saving fuel and stabilizing the furnace operating conditions; in temperature levels with gradual changes, the process details of the initial curve segment are preserved to the greatest extent, so as not to obscure the fine-tuning intention due to excessive smoothing. After initial smoothing in step A41, each temperature level curve segment achieves a smooth transition at the boundary. However, whether the heat input of each temperature level matches the heat absorption capacity of the billet in that temperature range has not yet been verified. Step A42 takes the smoothed projected temperature level curve segment sequence and the fusion coupling features as input. First, the local features of the current temperature level curve segment are extracted through a fourth constraint-coded multilayer perceptron. The global context of the fusion coupling features is extracted through a linear projection layer used as a query key. Then, a cross-attention operation is performed to allow the curve segment to interact with the global context—allowing the curve segment to perceive the heat absorption characteristics and ideal heat input pattern of the billet implied in the fusion coupling features, thereby extracting the energy deviation features between the current curve segment and the actual ideal heat input. Subsequently, the energy deviation features are concatenated with the current curve segment and processed by a gated weighted multilayer perceptron and a sigmoid function to generate an adaptive correction intensity coefficient—this coefficient determines the correction increment applied to the current temperature level. The amplitude is then calculated. Finally, the energy deviation feature is spliced ​​with the first-order difference between the current curve segment and the previous curve segment. After processing by the curve segment incremental multilayer perceptron, it is multiplied element-by-element by the adaptive correction intensity coefficient to achieve point-by-point incremental correction. This ensures that the heat input corresponding to each temperature level matches the actual heat absorption capacity of the billet in that temperature range. If a certain curve segment heats up too quickly, causing the accumulated heat to exceed the upper limit that the billet can absorb, the energy deviation feature will show a positive deviation. The correction intensity coefficient will prompt the incremental correction to compensate in the opposite direction, so that the curve forms a gentle heating segment to avoid overheating and cracking of the billet. If the heating is too slow, resulting in insufficient heat to drive the phase change reaction in that temperature range, the correction will cause the curve to heat up with a larger slope to make up for the heat gap. After verification and correction, each temperature level curve segment is no longer an independently generated segment, but a physical curve with thermodynamic self-consistency. This curve can adapt to the fluctuations in the heat absorption characteristics of different batches of billets, thus broadening the applicability of the process curve. Step A43 employs joint position coding, encoding the relative time relationship and indicator function between the time step and the temperature level. This ensures that in the temperature level transition region, the weights are distributed continuously between adjacent temperature levels, rather than undergoing abrupt switching at time boundaries. Finally, the values ​​of each temperature level curve segment at the corresponding relative time are weighted and summed according to the splicing weights to reconstruct a continuously verifiable and smooth continuous temperature control curve, which serves as the target set trajectory for the temperature controller. This allows each temperature level curve segment to maintain its own physical trend within its dominant range, while the splicing weights in the transition region are adaptively adjusted based on the indicator function and position coding, avoiding temperature inflection points caused by hard switching.

[0017] Furthermore, step A5 also includes: A51: The continuous temperature control curve is used as the target set trajectory of the temperature controller. The continuous temperature control curve is discretized into a temperature set value sequence by the logic controller according to a fixed control cycle, and then sent to the proportional-integral-derivative temperature controller corresponding to each temperature zone cycle by cycle. A52: The temperature controller collects the current actual temperature fed back by the corresponding thermocouple sensor in each control cycle, calculates the deviation between the current actual temperature and the current cycle temperature setpoint, inputs the proportional-integral-derivative control algorithm, generates the output control quantity for the current cycle, and converts the output control quantity into a gas valve opening command, which is then sent to the actuator of the corresponding temperature zone via the industrial fieldbus. A53: The above closed-loop control process is executed repeatedly until the holding time of all temperature levels in the current firing batch is accumulated. The controller outputs a zero-power command and closes the main gas valve or cuts off the heating power supply to complete the temperature control of the firing of this batch of billets.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Traditional fixed constant temperature process cannot take into account the difference in wall thickness between metal and ceramic blanks and the decomposition and release characteristics of multiple components of molding agent. The process switching relies on manual trial and error, which is costly and lacks precision. This invention can dynamically capture the nonlinear influence relationship between wall thickness, heat conduction and multi-component decomposition time sequence by synchronous standardization of multi-source heterogeneous information and fusion of multi-dimensional working condition characteristics. This allows for adaptive generation of heating rate and holding time sequence matching blank geometry and composition for each temperature level. On this basis, the temperature control curve is further smoothed by projection and boundary verification correction to avoid blank cracking or residual black spots caused by temperature change. After discretization, the control quantity is sent to the PID controller of each temperature zone in cycle by cycle. The actual temperature deviation is fed back in real time to generate control quantity, so that the heating rhythm of each temperature level always fits the current heat conduction state of the blank. The thick-walled core is fully heat-preserved while the thin-walled area avoids overheating expansion. This can adapt to blank characteristics and formula changes, so that metal blanks and ceramic blanks can stably complete degreasing under the same process path, breaking through the upper limit of the adaptation of traditional single process to material type and geometric characteristics.

[0019] (2) To address the issue of temperature control fragmentation caused by the difference in thermal response between metal and non-metal billets in the same furnace, this invention first uses a bidirectional long short-term memory network to capture the sequential dependence of furnace temperature and retain the original magnitude, so that the thermal inertia effect can be accurately characterized. Then, for metal billets, a multi-level point cloud network is used to extract the continuous gradient of wall thickness, and for non-metal billets, a dynamic graph convolutional network is used to capture the surface morphology convection thermal response, thus distinguishing the two types of thermal response physical mechanisms at the feature level. At the same time, the spatial topological information such as billet arrangement spacing and density is extracted in parallel with the component self-attention encoded ratio information through graph convolution and graph attention networks. Finally, an adaptive gating mechanism is used to automatically determine the material tendency based on the geometric, topological, and component characteristics of the current billet, and unify the four types of heterogeneous information, namely material difference, wall thickness distribution, arrangement position, and component ratio, into a single fused coupled feature tensor, providing a basis for subsequent process control to integrate multi-source physical information in one calculation.

[0020] (3) This invention uses the component decomposition temperature parameter as the constraint source, performs cross-modal semantic alignment through cross-attention and fusion coupling features, generates a temperature level boundary matrix as a time-step physical constraint signal, and transforms the rigid boundary into a soft constraint that can be referenced step by step; in the autoregressive decoding process, the boundary matrix performs Hadamard product gating on each hidden state to ensure that the heating rate and holding time are constrained by the current temperature level decomposition temperature range at each time step; when decoding enters the low boiling point component temperature level, the features corresponding to the high boiling point component are automatically suppressed to avoid interference from the previous residual information, embedding the physical boundary into each step of inference rather than just truncating it at the output end, and when the batch fluctuates, the concentration change drives the boundary matrix to redistribute the temperature level switching probability through the decomposition temperature parameter, and automatically matches the new batch safety process window.

[0021] (4) The present invention parameterizes and reconstructs the heating rate and holding time of each temperature level into continuous curve segments. Then, through gating weights, it adaptively selects heat diffusion smoothing or retains the original curve according to the degree of change of adjacent temperature levels, so that the temperature gradient at the temperature level boundary is controlled within the acceptable range of furnace thermal inertia. Then, through cross attention, the curve segments perceive the heat absorption characteristics of the billet implied in the fusion coupling characteristics, extract the energy deviation characteristics and generate an adaptive correction intensity coefficient, and perform incremental correction for each temperature level point by point. When the heating is too fast, reverse compensation slows down the curve, and when the heating is too slow, the slope is increased to make up for the heat gap, so as to ensure that the heat input of each segment matches the actual heat absorption capacity of the billet. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a deep learning-based intelligent temperature control method for degreasing provided by this invention; Figure 2 The experimental comparison diagram provided for this invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0024] Example 1: A deep learning-based intelligent temperature control method for degreasing, such as... Figure 1 As shown, it includes the following steps: A1: Collect time-series temperature data of multiple zones in the furnace, 3D point cloud data of the billet, billet arrangement image, and molar fraction of forming agent components, and preprocess them respectively to obtain preprocessed furnace temperature time-series data, preprocessed point cloud data with wall thickness attributes, preprocessed billet spatial adjacency map, preprocessed component molar fraction data, and component decomposition temperature parameters, including: A11: Temperature time-series data of multiple zones in the furnace is collected by thermocouple sensor arrays deployed in the industrial field. The data type is multi-dimensional time-series numerical data, including temperature sampling values, temperature setpoints and timestamps of each temperature zone. Outliers are filtered out by a three-fold standard deviation constant removal method for the temperature time-series data of the furnace, missing sampling points are filled in by linear interpolation, and Z-score standardization is used to map the temperature values ​​of each temperature zone to a standard distribution with a mean of 0 and a variance of 1, to obtain the preprocessed furnace temperature time-series data. A12: Obtain the three-dimensional point cloud data of the billet through the CAD model analysis system, and match the corresponding wall thickness attribute value for each point cloud vertex; filter out scanning noise by using the statistical outlier removal method on the three-dimensional point cloud data of the billet, and perform point cloud sparsification processing by using the voxel mesh downsampling method to obtain sparsified point cloud data; perform Z-score standardization on the wall thickness attribute value of each point cloud vertex in the sparsified point cloud data, and map the wall thickness value to a standard distribution with a mean of 0 and a variance of 1 to obtain preprocessed point cloud data with wall thickness attribute; A13: Images of the billet arrangement layout inside the furnace are acquired using an industrial area array camera; the billet arrangement layout images are denoised using Gaussian filtering, and then each billet in the billet arrangement layout image is segmented at the pixel level using an instance segmentation algorithm, and the center coordinates and circumscribed bounding box of each billet are extracted; a preset adjacency distance threshold is set, and the Euclidean distance between the center coordinates of any two billets is calculated. When the distance is less than or equal to the adjacency distance threshold, it is determined that the two billets have a spatial adjacency relationship; an undirected graph of billet arrangement is constructed with a single billet as a graph node and the spatial adjacency relationship between billets as graph edges; an adjacency matrix and a node attribute matrix are generated with the center coordinates and circumscribed bounding box of each billet as node attributes, to obtain the preprocessed billet spatial adjacency graph; A14: Collect the molar fraction of the molding agent components, and then use the component type identifier as the query key to integrate the decomposition start temperature, decomposition peak temperature and decomposition termination temperature corresponding to the component type identifier with the component molar fraction to obtain the preprocessed component molar fraction data and component decomposition temperature parameters.

[0025] A2: Based on the preprocessed furnace temperature time-series data, preprocessed point cloud data with wall thickness attributes, preprocessed billet spatial adjacency graph, and preprocessed component mole fraction data, calculate the latent features of furnace temperature and geometric features of wall thickness, adjacency topology features and component coding features, and fusion coupling features, including: A21: Based on the preprocessed furnace temperature time series data and the preprocessed point cloud data with wall thickness attributes, perform time series dynamic coding and spatial geometric representation extraction. First, calculate the latent features of furnace temperature; then, based on the billet material, select to calculate the geometric features of metal billet wall thickness or non-metal billet wall thickness. The calculation method for the latent characteristics of furnace temperature is as follows: ; When the billet is metallic, the geometric characteristics of the billet wall thickness are calculated as follows: ; When the blank is non-metallic, the geometric characteristics of the non-metallic blank wall thickness are calculated as follows: ; in, This is a latent characteristic of furnace temperature. A multilayer perceptron for extracting latent features of furnace temperature. For layer normalization operation, This is a bidirectional long short-term memory network operation. This is the preprocessed furnace temperature time series data. For linear layer projection operations used to extract hidden features of furnace temperature, This refers to the geometric characteristics of the wall thickness of the metal billet. For multi-level point cloud feature extraction network operations, For the farthest point sampling operation, Grouping operations for ball queries, This refers to preprocessed point cloud data with wall thickness attributes. This refers to the geometric characteristics of the wall thickness of a non-metallic billet. For multilayer perceptrons used for geometric feature extraction, This is for dynamic graph convolutional neural network operations. Graph construction operation for K nearest neighbors; A22: Based on the preprocessed billet spatial adjacency graph and the preprocessed component mole fraction data, graph structure encoding and component attribute embedding are performed to obtain adjacency topology features and component encoding features. The calculation method is as follows: ; ; in, This is an adjacency topology feature. This is a multilayer perceptron used for adjacency topology feature extraction. For graph attention network operations, For graph convolution operations, This is the spatial adjacency diagram of the pre-processed billet. For linear projection layers used for graph structure extraction, For component coding features, For multilayer perceptrons used for component feature extraction, For self-attention operations, For embedding layer, The data are the preprocessed component mole fractions. A23: Based on the latent characteristics of furnace temperature, the geometric characteristics of metal billet wall thickness, the geometric characteristics of non-metal billet wall thickness, the adjacency topology characteristics, and the component coding characteristics, branch gating and heterogeneous features are deeply fused to obtain fused coupled features. The calculation method is as follows: ; ; in, Metal Non-metallic adaptive gating weights, for Activation function For use in adaptive gating generation of multilayer perceptrons, This is a feature dimension concatenation operation. For the geometric characteristics of wall thickness, take the following according to the type of billet material: or ; To integrate coupling features, For the Hadamard operation, For multilayer perceptrons used for fusion and coupled feature extraction, This is a cross-attention operation.

[0026] A3: Based on the fusion coupling characteristics and component decomposition temperature parameters, the temperature level constraint embedding features and temperature level boundary matrix, the decoder initial hidden state and stage context vector, the heating rate sequence and the holding time sequence are calculated sequentially, including: A31: Based on the fusion coupling characteristics and component decomposition temperature parameters, thermophysical property constraint encoding and temperature order boundary embedding are performed to obtain the temperature order constraint embedding features and temperature order boundary matrix. The calculation method is as follows: ; ; in, For temperature-level constrained embedding features, For the second constraint encoding of the multilayer perceptron, Encode the multilayer perceptron for the first constraint. For component decomposition temperature parameters, For linear projection layers used for temperature level constraints, The temperature order boundary matrix, For multilayer perceptron operations used to generate temperature order boundary matrices, For the linear projection layer used to generate the temperature order boundary matrix; A32: Based on the temperature-order constraint embedding features, temperature-order boundary matrix, and fusion coupling features, constraint gating initialization and multi-stage decoding state generation are performed to obtain the decoder's initial hidden state and stage context vector. The calculation method is as follows: ; ; in, This is the initial hidden state of the decoder. This is a multilayer perceptron with a second hidden state. This is a multilayer perceptron with the first hidden state. For residual linear projection layer, For stage context vectors, For a second-context multilayer perceptron, For a first-context multilayer perceptron; A33: Based on the decoder's initial hidden state, stage context vector, and temperature order boundary matrix, perform autoregressive temperature order sequence decoding with boundary constraints to obtain the heating rate sequence and holding time sequence. The calculation method is as follows: ; ; in, This represents the hidden decoding state at time step t, where t is the time step index. [0,T], where T is the total number of time steps. When t=0... ; For gated loop unit operation, This represents the hidden decoding state at time step (t-1). This is the position encoding function for time step t. The corresponding temperature order boundary matrix Boundary vectors at each time step Let be the heating rate at time step t. Let t be the insulation duration at the t-th time step. For channel splitting operation, For the second output multilayer perceptron, For layer normalization operation, For the first output multilayer perceptron, For linear projection layers used for residuals; After decoding, the heating rate and holding time of each time step are spliced ​​together in chronological order to obtain the heating rate sequence and the holding time sequence.

[0027] A4: Based on the heating rate sequence, holding time sequence, and fusion coupling characteristics, the smoothed projected temperature step curve segment sequence, the verified and corrected temperature step curve segment sequence, and the continuous temperature control curve are calculated sequentially, including: A41: Based on the heating rate sequence, holding time sequence, and fusion coupling characteristics, parameterized reconstruction and heat conduction smoothing projection of temperature step curve segments are performed to obtain a preliminary temperature step curve segment sequence. The calculation method is as follows: ; ; ; in, For the first The initial curve segment of each temperature level, For the first The initial curve segment of each temperature level, For the second curve segment, a multilayer perceptron. For the first curve segment, a multilayer perceptron. The first in the heating rate sequence The heating rate for each temperature level, where k is the temperature level index. The first in the heat preservation time sequence The duration of heat preservation at each temperature level, For the linear projection layer used for context, For the position encoding function of temperature order k, For the smoothed projection of the first A temperature grade curve segment, For the first Smoothing gate weights for each temperature level, For smooth thermal diffusion operation, For gated weighted multilayer perceptrons, This is a first-order difference operation on the curve segment; when k=0, it is stipulated that... That is, the first temperature level is referenced by itself for a period of time preceding it; A42: Based on the smoothed projected temperature level curve segment sequence and the fusion coupling characteristics, energy balance consistency verification and curve segment adaptive correction are performed to obtain the verified and corrected temperature level curve segment sequence. The calculation method is as follows: ; ; ; in, For the first Energy deviation characteristics at each temperature level Encode the third constraint for the multilayer perceptron. For layer normalization operation, For cross-attention operations, For the fourth constraint-coded multilayer perceptron, For the linear projection layer used for the query key, For adaptive correction intensity multilayer perceptron, To verify the corrected first A temperature grade curve segment, For the first Adaptive correction intensity coefficients for each temperature level For curve segment incremental multilayer perceptron; A43: Based on the verified and corrected temperature level curve segment sequence, perform cross-temperature level continuous splicing and global temperature curve reconstruction to obtain the continuous temperature control curve. The calculation method is as follows: ; ; in, For the continuous temperature control curve in the first Temperature values ​​at each time step The total number of temperature levels. For the first The temperature grade curve segment at the first The concatenation weights of each time step To verify the corrected first Each temperature level curve segment is in The value at each time step, For the first The start timestamp of each temperature level The softmax function is calculated along the temperature dimension. To construct a weighted multilayer perceptron, Regarding the first The joint position encoding function of time step and temperature order k This is an indicator function.

[0028] Specifically, for operating scenarios involving rapid switching between multiple temperature levels and abrupt changes in the temperature control curve, this invention also provides a method for calculating continuous temperature control curves based on two-dimensional convolutional local temporal modeling, to replace step A43. The calculation method is as follows: ; ; in, This refers to the concatenation weights of the k-th temperature level curve segment at time step t, based on the output of 2D convolution. ( ) represents a weighted two-dimensional convolutional layer.

[0029] A5: Discretize the continuous temperature control curve into a sequence of temperature setpoints and send them to the proportional-integral-derivative temperature controllers of each temperature zone cycle by cycle. Collect the actual temperature, calculate the deviation, generate the control quantity, and execute closed-loop control cyclically until the holding time of all temperature levels is accumulated, thus completing the temperature control, including: A51: The continuous temperature control curve is used as the target set trajectory of the temperature controller. The continuous temperature control curve is discretized into a temperature set value sequence by the logic controller according to a fixed control cycle, and then sent to the proportional-integral-derivative temperature controller corresponding to each temperature zone cycle by cycle. A52: The temperature controller collects the current actual temperature fed back by the corresponding thermocouple sensor in each control cycle, calculates the deviation between the current actual temperature and the current cycle temperature setpoint, inputs the proportional-integral-derivative control algorithm, generates the output control quantity for the current cycle, and converts the output control quantity into a gas valve opening command, which is then sent to the actuator of the corresponding temperature zone via the industrial fieldbus. A53: The above closed-loop control process is executed repeatedly until the holding time of all temperature levels in the current firing batch is accumulated. The controller outputs a zero-power command and closes the main gas valve or cuts off the heating power supply to complete the temperature control of the firing of this batch of billets.

[0030] The neural network described in this invention comprises the following modules, and the parameters of each module are configured as follows: The furnace temperature hidden state feature extraction module is configured with two BiLSTM layers with a hidden layer size of 128 and a linear projection layer with an output size of 256. Combined with layer normalization layer and two fully connected layers , Both layers have inputs and outputs of 256, 256 and 256, 128, respectively, and use the ReLU activation function throughout without normalization or dropout processing. The metal billet wall thickness geometric feature extraction module is set up with an FPS sampling layer with 512 sampling points, a BallQuery grouping layer with a neighborhood radius of 0.2 and 32 sampling points per group, and two PointNet++ SA layers. The two SA layers have 4-channel and 64-channel inputs, respectively, and output 64-channel and 128-channel outputs. The non-metallic billet wall thickness geometric feature extraction module uses a KNN layer with a K value of 20, paired with two DGCNN EdgeConv layers. The input and output channels are the same as the SA layer for metallic billet feature extraction. Each layer is equipped with a LeakyReLU activation function with a negative slope of 0.2. It also features a layer normalization layer and a double-layer 128-dimensional output MLP_G with ReLU activation. The adjacency topology feature extraction module uses a hidden layer and a GraphConv layer with 128 outputs, a Proj_A projection layer with 128 outputs, and a layer normalization layer. It is paired with an 8-attention head, a GAT layer with 16 outputs per head and a total output of 128, and a double-layer ReLU activation. The final output dimension is 64; The component encoding feature extraction module encodes the mole fraction data of the four components through an embedding layer of size 32. It is configured with a SelfAttn layer with 4 attention heads, a key query size of 32, and an output of 128, along with a layer normalization layer, and a double-layer ReLU activation, resulting in a final output of 64. The fusion coupling feature generation module is set to a two-layer structure with an input of 256 and an output of 128. The final layer activates the output gating weights using a Sigmoid function, paired with an 8-attention head, a query weight of 128, a key-value weight of 256, and an output weight of 256 (CrossAttn layer), an 8-attention head 256-dimensional SelfAttn layer, and two 256-dimensional ReLU-activated layers. ; The temperature level constraint embedding and temperature level boundary matrix generation module encodes six temperature point parameters through an Emb layer with an embedding size of 64, combined with a multi-layer fully connected constraint encoding network and projection layer, multiple attention layers, and normalization layers, and finally activated by a Sigmoid function. Output the temperature level boundary matrix, and perform overall dimension matching feature fusion and constraint encoding features; The decoder initial hidden state and stage context vector generation module relies on fused features and constraint features, and is configured with multiple sets of two-layer and single-layer ReLU activated fully connected coding networks, combined with a residual projection layer with an output of 256 and a layer normalization layer, to complete the generation of the decoder hidden state and stage context vector respectively. The autoregressive temperature sequence decoding module employs a sine-cosine position encoding function with a size of 32, configured with a single-layer GRU layer with 256 hidden layers, and combined with a multi-layer ReLU-activated fully connected output network and residual projection and normalization structures. This ultimately decomposes the 64-dimensional output into 32-dimensional features for heating rate and holding time, with a total of 12 decoding time steps. The temperature curve segment parameterization reconstruction module uses a position encoding function with a size of 32, a SmoothProj operation with a thermal diffusion smoothing kernel size of 3, and multiple sets of ReLU-activated fully connected networks and Sigmoid-activated gated weight networks. The total number of temperature levels is set to 6, completing the parameterization reconstruction and smoothing optimization of the temperature control curve. The energy balance consistency verification module is configured with multiple sets of constraint-encoded fully connected layers, a CrossAttn layer with 4 attention heads, and a normalization layer. It generates correction parameters based on a feature projection network and calculates the curve segment increment and adaptive correction intensity through a multi-layer network activated by ReLU and Sigmoid, achieving energy balance verification. The cross-temperature-level continuous splicing module adopts a 32-dimensional joint position encoding function, configures a ReLU-activated two-layer splicing weight network, and relies on the Softmax operation to complete the weight normalization and continuous splicing of the 6 temperature levels. This invention employs an end-to-end training approach for iterative model training. The Adam optimizer is used, with an initial learning rate of 0.0005. A step-wise learning rate decay strategy is adopted, reducing the current learning rate by 0.5 every 20 training rounds. The batch size is set to 16, and the total number of training rounds is 200. The model training uses a composite loss function with multiple weights, including three types of loss branches. The mean square error loss for predicting heating rate and holding time, and the smoothing L1 loss for fitting the continuous temperature control curve and the target curve are all weighted at 1.0. The binary cross-entropy loss for predicting the temperature boundary matrix is ​​weighted at 0.5. Multi-dimensional loss constraints ensure the comprehensive accuracy of the model's temperature control curve prediction and feature matching.

[0031] like Figure 2 As shown, Figure 2 The thermal degreasing process (approximately 1200 minutes) was demonstrated, encompassing four heating stages and four holding plateaus. The method of this invention closely tracked the ideal curve across the entire temperature range, with smaller deviations at temperature transitions compared to the comparative method. The BP neural network method exhibited moderate overall deviation, but with larger fluctuations in the high-temperature range. Notably, in the initial heating stage (t=0→60min), the response lag of the traditional method resulted in a slightly smaller deviation than the method of this invention (counterexamples accounted for approximately 32.5%). This is because the temperature change was monotonous in this stage, and the adaptive gating mechanism of the method of this invention had not yet accumulated sufficient temporal context. Quantitatively, the method of this invention achieved RMSE=15.79°C, MAE=12.51°C, and R... 2 =0.9909, which reduces the RMSE by 59.2% compared to the traditional segmented isothermal process and by 38.2% compared to the BP neural network method; the results show that the multi-source heterogeneous feature fusion temperature control strategy based on deep learning can effectively capture the nonlinear coupling relationship between wall thickness-heat conduction-multi-group decomposition time sequence, and shows better tracking accuracy than the fixed process in both temperature level conversion and heat preservation stages.

[0032] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0033] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0034] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep learning-based intelligent temperature control method for degreasing, characterized in that, Includes the following steps: A1: Collect time-series temperature data of multiple zones in the furnace, 3D point cloud data of the billet, image of the billet arrangement, and molar fraction of the molding agent components, and preprocess them respectively to obtain preprocessed furnace temperature time-series data, preprocessed point cloud data with wall thickness attributes, preprocessed billet spatial adjacency map, preprocessed component molar fraction data, and component decomposition temperature parameters. A2: Based on the preprocessed furnace temperature time series data, the preprocessed point cloud data with wall thickness attributes, the preprocessed billet spatial adjacency graph, and the preprocessed component mole fraction data, calculate the furnace temperature latent features and wall thickness geometric features, adjacency topology features and component coding features, and fusion coupling features. A3: Based on the fusion coupling characteristics and component decomposition temperature parameters, calculate the temperature level constraint embedding features and temperature level boundary matrix, decoder initial hidden state and stage context vector, heating rate sequence and holding time sequence in sequence. A4: Based on the heating rate sequence, the holding time sequence, and the fusion coupling characteristics, calculate the temperature step curve segment sequence after smoothing projection, the temperature step curve segment sequence after verification and correction, and the continuous temperature control curve in sequence. A5: Discretize the continuous temperature control curve into a temperature setpoint sequence and send it to the proportional-integral-derivative temperature controller of each temperature zone cycle by cycle. Collect the actual temperature, calculate the deviation, generate the control quantity, and execute closed-loop control cyclically until the holding time of all temperature levels is accumulated, thus completing the temperature control.

2. The intelligent temperature control method for degreasing based on deep learning according to claim 1, characterized in that, Step A1 includes: A11: Temperature time-series data of multiple zones in the furnace is collected by thermocouple sensor arrays deployed in the industrial field. The data type is multi-dimensional time-series numerical data, including temperature sampling values, temperature setpoints and timestamps of each temperature zone. Outliers are filtered out by a three-fold standard deviation constant removal method for the temperature time-series data of the furnace, missing sampling points are filled in by linear interpolation, and Z-score standardization is used to map the temperature values ​​of each temperature zone to a standard distribution with a mean of 0 and a variance of 1, to obtain the preprocessed furnace temperature time-series data. A12: Obtain the three-dimensional point cloud data of the billet through the CAD model analysis system, and match the corresponding wall thickness attribute value for each point cloud vertex; filter out scanning noise by using the statistical outlier removal method on the three-dimensional point cloud data of the billet, and perform point cloud sparsification processing by using the voxel mesh downsampling method to obtain sparsified point cloud data; perform Z-score standardization on the wall thickness attribute value of each point cloud vertex in the sparsified point cloud data, and map the wall thickness value to a standard distribution with a mean of 0 and a variance of 1 to obtain preprocessed point cloud data with wall thickness attribute; A13: Images of the billet arrangement layout inside the furnace are acquired using an industrial area array camera; the billet arrangement layout images are denoised using Gaussian filtering, and then each billet in the billet arrangement layout image is segmented at the pixel level using an instance segmentation algorithm, and the center coordinates and circumscribed bounding box of each billet are extracted; a preset adjacency distance threshold is set, and the Euclidean distance between the center coordinates of any two billets is calculated. When the distance is less than or equal to the adjacency distance threshold, it is determined that the two billets have a spatial adjacency relationship; an undirected graph of billet arrangement is constructed with a single billet as a graph node and the spatial adjacency relationship between billets as graph edges; an adjacency matrix and a node attribute matrix are generated with the center coordinates and circumscribed bounding box of each billet as node attributes, to obtain the preprocessed billet spatial adjacency graph; A14: Collect the molar fraction of the molding agent components, and then use the component type identifier as the query key to integrate the decomposition start temperature, decomposition peak temperature and decomposition termination temperature corresponding to the component type identifier with the component molar fraction to obtain the preprocessed component molar fraction data and component decomposition temperature parameters.

3. The intelligent temperature control method for degreasing based on deep learning according to claim 1, characterized in that, Step A2 includes: A21: Based on the preprocessed furnace temperature time series data and the preprocessed point cloud data with wall thickness attributes, perform time series dynamic coding and spatial geometric representation extraction. First, calculate the latent features of furnace temperature; then, based on the billet material, select to calculate the geometric features of metal billet wall thickness or non-metal billet wall thickness. A22: Based on the preprocessed billet spatial adjacency graph and the preprocessed component mole fraction data, graph structure encoding and component attribute embedding are performed to obtain adjacency topology features and component encoding features; A23: Based on the latent characteristics of furnace temperature, geometric characteristics of metal billet wall thickness, geometric characteristics of non-metal billet wall thickness, adjacency topology characteristics, and component coding characteristics, branch gating and heterogeneous features are deeply fused to obtain fused coupling features.

4. The intelligent temperature control method for degreasing based on deep learning according to claim 3, characterized in that, Step A2 includes: The calculation process of the furnace temperature latent features in step A21 is as follows: First, the preprocessed furnace temperature time series data is processed by a bidirectional long short-term memory network and a linear layer projection process for extracting the furnace temperature latent features. Then, the two processing results are added together, and then the layer normalization operation is performed. Finally, the data is processed by a multilayer perceptron for extracting the furnace temperature latent features to obtain the furnace temperature latent features. When the billet material is metal, the calculation process of the geometric features of the metal billet wall thickness is as follows: the pre-processed point cloud data with wall thickness attributes is first subjected to ball query grouping operation, then the farthest point sampling operation is performed, and then processed by a multi-level point cloud feature extraction network to obtain the geometric features of the metal billet wall thickness. When the material of the blank is non-metallic, the calculation process of the geometric features of the wall thickness of the non-metallic blank is as follows: the pre-processed point cloud data with wall thickness attributes is first processed by K-nearest neighbor mapping, then processed by dynamic graph convolutional neural network, then processed by layer normalization, and then processed by multilayer perceptron for geometric feature extraction to obtain the geometric features of the wall thickness of the non-metallic blank. The calculation process of adjacency topology features and component coding features in step A22 is as follows: First, the preprocessed billet space adjacency graph is processed by graph convolution operation and linear projection layer for graph structure extraction, respectively. Then, the two processing results are added and processed by graph attention network operation, then by layer normalization operation, and then by multilayer perceptron for adjacency topology feature extraction to obtain adjacency topology features. The preprocessed component mole fraction data is processed by embedding layer, then by self-attention operation, then added to the output of embedding layer, then by layer normalization operation, and then by multilayer perceptron for component feature extraction to obtain component coding features. The calculation process of the fusion coupling feature in step A23 is as follows: Based on the wall thickness geometric features, adjacency topology features, and component coding features, a concatenation operation is performed according to the feature dimensions. This concatenation is then processed by a multilayer perceptron for adaptive gating generation, followed by a Sigmoid activation function to obtain the metal-nonmetal adaptive gating weights. A cross-attention operation is performed on the furnace temperature latent features and the concatenated features obtained from the wall thickness geometric features, adjacency topology features, and component coding features. This cross-attention operation is then performed by a multilayer perceptron for fusion coupling feature extraction to obtain the first intermediate result. A self-attention operation is performed on the concatenated features obtained from the wall thickness geometric features, adjacency topology features, and component coding features. This self-attention operation is then performed by a multilayer perceptron for fusion coupling feature extraction to obtain the second intermediate result. The metal-nonmetal adaptive gating weights are multiplied by the first intermediate result, and the result after subtracting the metal-nonmetal adaptive gating weights from 1 is multiplied by the second intermediate result. The two Hadamard product results are then added together to obtain the fusion coupling feature.

5. The intelligent temperature control method for degreasing based on deep learning according to claim 3, characterized in that, Step A3 includes: A31: Based on the fusion coupling characteristics and component decomposition temperature parameters, perform thermal property constraint encoding and temperature order boundary embedding to obtain temperature order constraint embedding characteristics and temperature order boundary matrix; A32: Based on the temperature level constraint embedding features, temperature level boundary matrix and fusion coupling features, perform constraint gating initialization and multi-stage decoding state generation to obtain the decoder's initial hidden state and stage context vector; A33: Based on the initial hidden state of the decoder, the stage context vector, and the temperature order boundary matrix, perform autoregressive temperature order sequence decoding with boundary constraints to obtain the heating rate sequence and the holding time sequence.

6. The intelligent temperature control method for degreasing based on deep learning according to claim 5, characterized in that, Step A3 includes: The calculation process of the temperature-order constraint embedding feature and the temperature-order boundary matrix in step A31 includes: firstly, embedding the component decomposition temperature parameters, then processing them by a first constraint-encoding multilayer perceptron after layer normalization; simultaneously, performing a linear projection of the fused coupling features for temperature-order constraints; performing cross-attention on the results of the above two processes, then adding the embedding result of the component decomposition temperature parameters, and then processing them sequentially by layer normalization and a second constraint-encoding multilayer perceptron to obtain the temperature-order constraint embedding feature; concatenating the temperature-order constraint embedding feature with the result of performing a linear projection of the fused coupling feature for generating the temperature-order boundary matrix, then processing it by a multilayer perceptron for generating the temperature-order boundary matrix, and finally processing it by the Sigmoid function to obtain the temperature-order boundary matrix; The calculation process of the decoder's initial hidden state and stage context vector in step A32 includes: concatenating the fused coupling feature with the temperature-order constraint embedding feature, processing it through a first hidden state multilayer perceptron after layer normalization, adding the result of residual linear projection on the fused coupling feature, and then processing it through a second hidden state multilayer perceptron in sequence after layer normalization to obtain the decoder's initial hidden state; concatenating the temperature-order constraint embedding feature with the fused coupling feature, processing it first through a first context multilayer perceptron, then processing it through a second context multilayer perceptron after layer normalization, and performing a Hadamard product with the temperature-order boundary matrix to obtain the stage context vector; The decoding process of the heating rate sequence and holding time sequence in step A33 includes: for each time step, concatenating the position encoding, stage context vector, and boundary vector corresponding to the current time step in the temperature order boundary matrix with the result of the Hadamard product of the decoded hidden state of the previous time step, and inputting the concatenated result and the decoded hidden state of the previous time step into the gated recurrent unit to obtain the decoded hidden state of the current time step; after layer normalization, the decoded hidden state of the current time step is processed by the first output multilayer perceptron, and the result of linear projection of the decoded hidden state of the current time step for the residual is added, and then processed by the second output multilayer perceptron in sequence, and then the heating rate and holding time of the current time step are obtained by channel segmentation operation; after decoding, the heating rate and holding time of each time step are concatenated in chronological order to obtain the heating rate sequence and the holding time sequence.

7. The intelligent temperature control method for degreasing based on deep learning according to claim 5, characterized in that, The A4 step includes: A41: Based on the heating rate sequence, holding time sequence and fusion coupling characteristics, the temperature step curve segments are parametrically reconstructed and heat conduction is smoothly projected to obtain the preliminary temperature step curve segment sequence. A42: Based on the temperature level curve segment sequence after smooth projection and the fusion coupling characteristics, perform energy balance consistency verification and curve segment adaptive correction to obtain the verified and corrected temperature level curve segment sequence. A43: Based on the verified and corrected temperature level curve segment sequence, perform cross-temperature level continuous splicing and global temperature curve reconstruction to obtain a continuous temperature control curve.

8. The intelligent temperature control method for degreasing based on deep learning according to claim 7, characterized in that, The A4 step includes: Based on the heating rate sequence, holding time sequence, and fusion coupling features, parameterized reconstruction and heat conduction smoothing projection of temperature level curve segments are performed to obtain an initial temperature level curve segment sequence. Specifically, the heating rate of the current temperature level in the heating rate sequence, the holding time of the current temperature level in the holding time sequence, and the fusion coupling features are concatenated after processing by a linear projection layer for context. This concatenation is then processed by a first curve segment multilayer perceptron, with the addition of a position encoding function for the current temperature level, followed by layer normalization. Finally, it is processed by a second curve segment multilayer perceptron to obtain the initial curve segment of the current temperature level. The initial curve segment is concatenated with the results obtained from the first-order difference operation of the curve segments of the previous temperature level and the current temperature level. The concatenation is then processed by a gated weighted multilayer perceptron and then by a Sigmoid function to obtain the smoothing gate weight of the current temperature level. The smoothing gate weight of the current temperature level is then multiplied by the Hadamard product of the initial curve segment of the current temperature level. At the same time, the smoothing gate weight of the current temperature level is subtracted from 1 and multiplied by the Hadamard product of the results obtained from the thermal diffusion smoothing operation of the initial curve segments of the previous temperature level and the current temperature level. The two Hadamard product results are then added together to obtain the smoothed projected curve segment of the current temperature level. Based on the smoothed projected temperature level curve segment sequence and the fused coupling features, energy balance consistency verification and curve segment adaptive correction are performed to obtain the verified and corrected temperature level curve segment sequence. Specifically, the smoothed projected current temperature level curve segment is processed by a fourth constraint-coded multilayer perceptron to obtain the first intermediate feature. Simultaneously, the fused coupling features are processed by a linear projection layer used as the query key to obtain the second intermediate feature. The first and second intermediate features are subjected to cross-attention operation, then layer normalization, and finally added to the result obtained by the smoothed projected current temperature level curve segment processed by the fourth constraint-coded multilayer perceptron. Then, the result is processed by a third constraint-coded multilayer perceptron to obtain... The energy deviation characteristics of the current temperature level are obtained; the energy deviation characteristics of the current temperature level are concatenated with the smoothed projected current temperature level curve segment, processed by a gated weighted multilayer perceptron, and then passed through a Sigmoid function to obtain the adaptive correction intensity coefficient of the current temperature level; the energy deviation characteristics of the current temperature level are concatenated with the first-order difference operation result of the curve segment between the smoothed projected current temperature level curve segment and the smoothed projected temperature level curve segment of the previous temperature level, processed by a curve segment incremental multilayer perceptron, and then Hadamard product is performed with the adaptive correction intensity coefficient of the current temperature level. This product is then added to the smoothed projected current temperature level curve segment to obtain the current temperature level curve segment after verification and correction. Based on the verified and corrected temperature level curve segment sequence, continuous splicing across temperature levels and global temperature curve reconstruction are performed to obtain a continuous temperature control curve. Specifically, for each temperature level and each time step, the joint position encoding function and indicator function of the current time step and the current temperature level are spliced ​​together. After processing by a splicing weight multilayer perceptron, the Softmax function is calculated along the temperature level dimension to obtain the splicing weight of the current temperature level curve segment at the current time step. The indicator function is used to indicate whether the current time step belongs to the interval between the start timestamp of the current temperature level and the start timestamp of the next temperature level. For each temperature level from the first temperature level to the last temperature level, the value of the current temperature level curve segment at the current time step minus the start timestamp of the current temperature level is multiplied by the splicing weight of the current temperature level curve segment at the current time step. The product results of all temperature levels are accumulated to obtain the temperature value of the continuous temperature control curve at the current time step.

9. The intelligent temperature control method for degreasing based on deep learning according to claim 7, characterized in that, Step A5 includes: A51: The continuous temperature control curve is used as the target set trajectory of the temperature controller. The continuous temperature control curve is discretized into a temperature set value sequence by the logic controller according to a fixed control cycle, and then sent to the proportional-integral-derivative temperature controller corresponding to each temperature zone cycle by cycle. A52: The temperature controller collects the current actual temperature fed back by the corresponding thermocouple sensor in each control cycle, calculates the deviation between the current actual temperature and the current cycle temperature setpoint, inputs the proportional-integral-derivative control algorithm, generates the output control quantity for the current cycle, and converts the output control quantity into a gas valve opening command, which is then sent to the actuator of the corresponding temperature zone via the industrial fieldbus. A53: The above closed-loop control process is executed repeatedly until the holding time of all temperature levels in the current firing batch is accumulated. The controller outputs a zero-power command and closes the main gas valve or cuts off the heating power supply to complete the temperature control of the firing of this batch of billets.

Images