Additive manufacturing device water temperature regulation method based on recurrent neural network

Abstract

The application discloses a water temperature regulation and control method for an additive manufacturing device based on a recurrent neural network, and comprises the following steps: collecting multi-source time series data and performing pretreatment; based on a time series sample set, performing multi-period training on an improved IndRNN model to generate a water temperature prediction model; collecting multi-source time series data in real time, performing prediction processing through the water temperature prediction model to obtain a water temperature prediction result; calculating a control deviation, adjusting a control parameter, and generating a water temperature regulation and control instruction; performing parameter adjustment, calculating a prediction error, and combining online thermal coupling time series characteristics to form an incremental sample; based on the incremental sample, updating the water temperature prediction model and performing forward reasoning to generate a new water temperature regulation and control instruction. The application utilizes a multi-period recurrent neural network to model thermal coupling characteristics, realizes accurate water temperature prediction and dynamic regulation and control, and has the advantages of fast response, high stability and strong self-adaptation.

Description

Technical Field

[0001] This invention relates to the field of intelligent temperature control technology in additive manufacturing, and in particular to a water temperature control method for additive manufacturing equipment based on recurrent neural networks. Background Technology

[0002] During prolonged operation, current additive manufacturing equipment experiences continuous heat accumulation due to factors such as nozzle power, printing path, and material deposition rhythm, leading to constant variations in the heat dissipation load of the water cooling system. Traditional water temperature control methods often rely on fixed threshold control or temperature prediction models based on a single time scale, which typically fail to simultaneously capture the coupling relationship between rapid thermal fluctuations, medium-term heat accumulation, and long-term temperature rise trends, resulting in low prediction accuracy. In actual operation, the water cooling inlet temperature, water cooling outlet temperature, and water cooling flow rate exhibit multi-scale, nonlinear characteristics with varying loads, making it difficult for recurrent neural network models with a single update rhythm to accurately reflect changes across different time scales.

[0003] Existing methods generally lack recursive structure partitioning for multiple time update cycles, making it impossible to model short-term, medium-term, and long-term temperature changes separately. They also lack mechanisms to guide the model towards a reasonable thermal equilibrium through physical constraints. In water temperature control, traditional methods typically generate control commands based on single-point predictions, failing to incorporate future trends across multiple time steps to establish a more comprehensive control bias. Furthermore, most models lack online incremental learning mechanisms for prediction errors, preventing them from self-correcting based on the latest operating conditions, leading to a rapid decline in temperature control performance with environmental changes.

[0004] Therefore, how to provide a water temperature control method for additive manufacturing equipment based on recurrent neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a water temperature control method for additive manufacturing equipment based on recurrent neural networks. This invention utilizes multi-period recurrent neural networks to model thermal coupling characteristics, thereby achieving accurate water temperature prediction and dynamic control, and possesses the advantages of fast response, high stability, and strong adaptability.

[0006] The water temperature control method for additive manufacturing equipment based on recurrent neural networks according to embodiments of the present invention includes the following steps:

[0007] Temperature sensors, flow sensors, and power acquisition modules are deployed in additive manufacturing equipment and water cooling systems to collect multi-source time-series data, perform preprocessing, and generate time-series sample sets.

[0008] Based on the time series sample set, the improved IndRNN model is trained for multiple cycles to generate a water temperature prediction model.

[0009] Multi-source time-series data are collected in real time during the additive manufacturing process, and the water temperature prediction results are obtained through a water temperature prediction model.

[0010] The control deviation is calculated based on the water temperature prediction results and the preset water temperature target curve. The control parameters are adjusted based on the control deviation, and water temperature control instructions are generated.

[0011] The water temperature control command is parsed and the water cooling system execution components are driven to adjust relevant parameters. After the adjustment is performed, the prediction error is calculated, and incremental samples are generated by combining online thermal coupling time series characteristics and written into the sample library.

[0012] Based on incremental samples, the internal parameters of each updated sub-module of the water temperature prediction model are incrementally trained, and forward inference is performed based on the updated water temperature prediction model to generate new water temperature control commands.

[0013] Optionally, the multi-source time-series data includes printing parameters, nozzle power, water-cooled inlet temperature, water-cooled outlet temperature, and water-cooled flow rate. The printing parameters include scanning speed, scanning path, layer thickness, material feed speed, and printing layer index. The preprocessing includes noise reduction, time alignment, missing value completion, and numerical normalization.

[0014] Optionally, the generation of the water temperature prediction model specifically includes:

[0015] Based on the time-series sample set, construct the thermal load feature sequence and the water cooling state feature sequence, and splice them together to generate the thermally coupled time-series feature sequence;

[0016] An improved IndRNN model is constructed, which divides the hidden layer into a fast update sub-module, a medium update sub-module, and a slow update sub-module. The fast update sub-module is set to update at every time step, the medium update sub-module is set to update within a preset time step, and the slow update sub-module is set to update after a preset number of time steps.

[0017] The improved IndRNN model includes dividing the hidden layer into multiple sub-modules based on different time update periods, and adding physical constraint terms that constrain the model's prediction behavior to the loss function during model training.

[0018] Perform temporal unrolling on the thermally coupled temporal feature sequence, and drive the three sub-modules to update their hidden states at different time steps;

[0019] For each time step, perform forward calculations, integrate the intermediate states generated by the fast update submodule, medium update submodule, and slow update submodule at the current time step according to a preset method, transform the integrated intermediate states according to a preset mapping relationship, and generate a predicted value representing the future trend of water temperature change.

[0020] The sum of the squares of the numerical differences between the actual value and the corresponding predicted value of the water-cooled outlet temperature is taken as the main loss, and the sum of the squares of the numerical differences between the increase in heat caused by the nozzle power and the amount of cooling carried away by the water-cooled flow rate is taken as the additional loss. The main loss and the additional loss are weighted and summed to construct a loss function.

[0021] Backpropagation is performed at each time step to calculate the gradient generated by the loss function, and the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule are updated respectively.

[0022] The above operations are repeated in chronological order during the training period, forming a new set of parameters in each iteration, until the training error converges and the final set of parameters is solidified, thus generating a water temperature prediction model.

[0023] Optionally, obtaining the water temperature prediction result specifically includes:

[0024] During the operation of the additive manufacturing task, multi-source time-series data are collected in real time according to the preset sampling period, preprocessed, and combined and spliced ​​in chronological order to generate an online thermally coupled time-series feature sequence.

[0025] Feature alignment and dimension matching are performed on the online thermal coupling time series feature sequence, and then input into the water temperature prediction model. Forward inference is performed according to the preset number of future time steps to obtain the water cooling outlet temperature prediction sequence for each future time step.

[0026] Based on the predicted water-cooled outlet temperature sequence, the numerical difference and average rate of change between the predicted water-cooled outlet temperatures of adjacent future time steps are calculated in chronological order, and the numerical difference and average rate of change are combined to generate the water temperature change trend result.

[0027] The predicted water-cooled outlet temperature sequence is correlated with the water temperature change trend results according to the time index of the future time step to obtain the water temperature prediction result.

[0028] Optionally, the generation of the water temperature control command specifically includes:

[0029] Extract the predicted water-cooled outlet temperature values ​​corresponding to each future time step from the water temperature prediction results, and extract the target water temperature values ​​under the same time index from the preset water temperature target curve. Align the water-cooled outlet temperature prediction value sequence with the target water temperature value sequence one by one according to the time index.

[0030] At each future time step, the control deviation value for the corresponding time step is obtained by subtracting the predicted water-cooled outlet temperature value from the corresponding target water temperature value, and the control deviation values ​​for each future time step are arranged in chronological order to form a control deviation sequence.

[0031] Based on the control deviation sequence and the preset time weight, the control deviation values ​​of each future time step are weighted and summed to generate a comprehensive control deviation index that represents the overall water temperature deviation.

[0032] Based on the comprehensive control deviation index and the preset cooling flow rate adjustment coefficient, cooling power adjustment coefficient and water pump speed adjustment coefficient, the basic cooling flow rate, basic cooling power and basic water pump speed are numerically adjusted respectively to generate the cooling flow rate set value, cooling power set value and water pump speed set value, and the three are combined to generate the water temperature control command.

[0033] Optionally, the generation of the incremental samples specifically includes:

[0034] The system analyzes water temperature control commands and sends them to the flow control device, refrigeration device, and circulating water pump.

[0035] The control system's actuators adjust the cooling flow rate, cooling power, and circulation pressure according to the water temperature control command, and maintain the adjusted operating conditions for a preset duration.

[0036] After the adjustment is performed, the actual water-cooled outlet temperature is collected according to the preset sampling period, and the corresponding predicted water-cooled outlet temperature value is extracted from the water temperature prediction result according to the time index.

[0037] At each time index, the actual water-cooled outlet temperature is subtracted from the corresponding predicted water-cooled outlet temperature to generate a prediction error value. The prediction error value is then combined with the online thermal coupling time series feature of the corresponding time step in the online thermal coupling time series feature sequence according to the time index to generate incremental samples.

[0038] The incremental samples are written into the sample library, and the corresponding water temperature control command information and actual water cooling outlet temperature information are recorded.

[0039] Optionally, the incremental training of the water temperature prediction model specifically includes:

[0040] Incremental samples are read from the sample library in chronological order, and the online thermally coupled temporal features in the incremental samples are used as model inputs, while the prediction error values ​​are used as supervision signals.

[0041] Within each training batch, the online thermally coupled temporal features are input into the water temperature prediction model based on the time index of the incremental samples, and the loss value of each batch is calculated based on the loss function.

[0042] At each time step, the gradient is calculated on the loss function and applied to the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule.

[0043] At time steps that meet the update cycle conditions, the internal parameters of the medium-speed update submodule and the slow-speed update submodule are numerically adjusted. At time steps that do not meet the update cycle conditions, their internal parameters remain unchanged. The internal parameters of the fast update submodule are numerically adjusted at each time step.

[0044] At the end of the training batch, the updated internal parameters of the fast update submodule, medium update submodule, and slow update submodule are written into the parameter storage unit.

[0045] The beneficial effects of this invention are:

[0046] This invention collects multi-source time-series data from additive manufacturing equipment and water-cooling systems, and constructs a thermally coupled time-series feature sequence including printing parameters, nozzle power, water-cooling inlet temperature, water-cooling outlet temperature, and water-cooling flow rate. This enables the model to comprehensively acquire core information reflecting real-time heat load changes. Based on this, the improved recurrent neural network model proposed in this invention divides the hidden layer into fast update sub-modules, medium-speed update sub-modules, and slow update sub-modules. Different sub-modules update the hidden state at different time scales. This structure can simultaneously capture multi-scale characteristics such as short-term thermal disturbances, medium-term heat accumulation, and long-term temperature rise trends during additive manufacturing, solving the problem that existing single-time-scale prediction models cannot adequately account for different thermal change rhythms, and significantly improving the accuracy and stability of water temperature prediction for multiple future time steps.

[0047] This invention further constructs a loss function that includes both primary and secondary losses, and introduces physical constraints based on nozzle power and water cooling flow rate. This ensures that the model, during training and incremental learning, not only relies on data-driven error backpropagation but is also constrained by thermal equilibrium relationships, thus preventing the model from exhibiting abnormal convergence directions that deviate from the true thermal evolution. Simultaneously, based on the inference results of the water temperature prediction model, this invention constructs a control deviation sequence and a comprehensive control deviation index. Based on this index, cooling flow rate, cooling power, and pump speed are simultaneously adjusted. This means that the generation of control commands no longer depends on the predicted value at a single moment but is based on the overall regulation of temperature trends over multiple future time steps. This improves the coordination and foresight of water temperature control, avoiding the regulation lag and over-response phenomena that are common in traditional methods.

[0048] Furthermore, this invention constructs an online incremental learning mechanism. It generates a prediction error by comparing the actual water-cooled outlet temperature after water temperature control is executed with the predicted value. This prediction error is then combined with corresponding online thermal coupling time-series features to generate incremental samples, enabling continuous updating and adaptive optimization of the model during operation. This mechanism allows the water temperature prediction model to absorb the latest operating conditions and temperature response differences in real time, continuously improving its adaptability to environmental changes, load fluctuations, and equipment thermal inertia, effectively overcoming the problem of declining prediction accuracy over time in traditional models. In summary, this invention has significant technical advantages in multi-timescale thermal coupling modeling, physical constraint-driven training, forward-looking water temperature control command generation, and online incremental learning, achieving higher accuracy, higher stability, and faster response speed in water temperature control. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of the water temperature control method for additive manufacturing equipment based on recurrent neural networks proposed in this invention;

[0051] Figure 2 This is a flowchart illustrating the multi-cycle training and physical constraint coupling of the improved IndRNN model for the water temperature control method of additive manufacturing equipment based on recurrent neural networks proposed in this invention.

[0052] Figure 3 This is a flowchart of the incremental training process for the water temperature prediction model of the water temperature control method for additive manufacturing equipment based on recurrent neural networks proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1-3 A water temperature control method for additive manufacturing equipment based on recurrent neural networks includes the following steps:

[0055] Temperature sensors, flow sensors, and power acquisition modules are deployed in additive manufacturing equipment and water cooling systems to collect multi-source time-series data, perform preprocessing, and generate time-series sample sets.

[0056] Based on the time series sample set, the improved IndRNN model is trained for multiple cycles to generate a water temperature prediction model.

[0057] Multi-source time-series data are collected in real time during the additive manufacturing process, and the water temperature prediction results are obtained through a water temperature prediction model.

[0058] The control deviation is calculated based on the water temperature prediction results and the preset water temperature target curve. The control parameters are adjusted based on the control deviation, and water temperature control instructions are generated.

[0059] The water temperature control command is parsed and the water cooling system execution components are driven to adjust relevant parameters. After the adjustment is performed, the prediction error is calculated, and incremental samples are generated by combining online thermal coupling time series characteristics and written into the sample library.

[0060] Based on incremental samples, the internal parameters of each updated sub-module of the water temperature prediction model are incrementally trained, and forward inference is performed based on the updated water temperature prediction model to generate new water temperature control commands.

[0061] In this embodiment, the multi-source time-series data includes printing parameters, nozzle power, water-cooled inlet temperature, water-cooled outlet temperature, and water-cooled flow rate. The printing parameters include scanning speed, scanning path, layer thickness, material feed speed, and printing layer index. The preprocessing includes noise reduction, time alignment, missing value completion, and numerical normalization.

[0062] In this embodiment, the generation of the water temperature prediction model specifically includes:

[0063] Based on the time-series sample set, a thermal load feature sequence and a water-cooling state feature sequence are constructed and spliced ​​together to generate a thermally coupled time-series feature sequence. The construction of the thermal load feature sequence and the water-cooling state feature sequence includes combining printing parameters and nozzle power in time order to form a thermal load feature sequence, and combining water-cooling inlet temperature, water-cooling outlet temperature and water-cooling flow rate in time order to form a water-cooling state feature sequence.

[0064] An improved IndRNN model is constructed, which divides the hidden layer into a fast update sub-module, a medium update sub-module, and a slow update sub-module. The fast update sub-module is set to update at every time step, the medium update sub-module is set to update within a preset time step, and the slow update sub-module is set to update after a preset number of time steps.

[0065] The improved IndRNN model includes dividing the hidden layer into multiple sub-modules based on different time update periods, and adding physical constraint terms that constrain the model's prediction behavior to the loss function during model training.

[0066] The specific division includes: when constructing the improved IndRNN model, the hidden layer node set of the model is grouped according to the update frequency. Hidden layer nodes used to represent short-term changes are assigned to the fast update submodule and set to be updated at every time step. Hidden layer nodes used to represent medium-term changes are assigned to the medium-speed update submodule and set to be updated within a preset number of time steps. Hidden layer nodes used to represent long-term change trends are assigned to the slow update submodule and set to be updated after crossing a preset number of time steps. During the model initialization process, the recursive connection structure of the three submodules and the corresponding update scheduling rules are established respectively.

[0067] The recursive connection structure includes: establishing a recursive association between the hidden state of the previous time step and the input features of the current time step in three sub-modules respectively; passing the hidden state of the previous cycle to the hidden state update process of the next cycle through the internal recursive unit; and setting up an independent hidden state storage unit in each sub-module to save the state information under the corresponding time scale.

[0068] Perform temporal unrolling on the thermally coupled temporal feature sequence, and drive the three sub-modules to update their hidden states at different time steps;

[0069] The update of the hidden state specifically includes: inputting the hot-coupled temporal features into the fast update submodule, the medium-speed update submodule, and the slow update submodule according to the current time step index; reading the hidden state formed in the previous update cycle from the hidden state storage unit of each of the three submodules as the initial state; and performing state transformation between the current input features and the corresponding initial state in the recursive unit within each submodule to generate a new hidden state. Specifically, the fast update submodule performs hidden state update at each time step; the medium-speed update submodule performs hidden state update at time steps that meet the medium-speed update cycle conditions, and keeps the hidden state of the previous cycle unchanged at time steps that do not meet the medium-speed update cycle conditions; the slow update submodule performs hidden state update at time steps that meet the slow update cycle conditions, and keeps the hidden state of the previous cycle unchanged at time steps that do not meet the slow update cycle conditions; and writing the hidden state generated in this time step into the hidden state storage unit of the corresponding submodule for recursive calculation in the next time step.

[0070] For each time step, perform forward calculations, integrate the intermediate states generated by the fast update submodule, medium update submodule, and slow update submodule at the current time step according to a preset method, transform the integrated intermediate states according to a preset mapping relationship, and generate a predicted value representing the future trend of water temperature change.

[0071] The sum of the squares of the numerical differences between the actual value and the corresponding predicted value of the water-cooled outlet temperature is taken as the main loss, and the sum of the squares of the numerical differences between the increase in heat caused by the nozzle power and the amount of cooling carried away by the water-cooled flow rate is taken as the additional loss. The main loss and the additional loss are weighted and summed to construct a loss function. The loss function simultaneously constrains the convergence direction of the parameters of the fast update submodule, the medium update submodule and the slow update submodule in different time update cycles.

[0072] Backpropagation is performed at each time step to calculate the gradient generated by the loss function, and the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule are updated respectively.

[0073] The update of the internal parameters specifically includes: performing parameter matching and numerical adjustment on the internal parameters of the fast update submodule based on the gradient of the loss function at the current time step; performing parameter matching and numerical adjustment on the internal parameters of the medium-speed update submodule based on the gradient of the loss function at the time step that meets the medium-speed update cycle condition, and keeping the internal parameters of the medium-speed update submodule unchanged at the time step that does not meet the medium-speed update cycle condition; performing parameter matching and numerical adjustment on the internal parameters of the slow update submodule based on the gradient of the loss function at the time step that meets the slow update cycle condition, and keeping the internal parameters of the slow update submodule unchanged at the time step that does not meet the slow update cycle condition, and writing the update results into the corresponding parameter storage unit for training calculation in the next time step after completing the update processing of the internal parameters of the three submodules;

[0074] The above operations are repeated in chronological order during the training period, forming a new set of parameters in each iteration, until the training error converges and the final set of parameters is solidified, thus generating a water temperature prediction model.

[0075] In this embodiment, obtaining the water temperature prediction result specifically includes:

[0076] During the operation of the additive manufacturing task, multi-source time-series data are collected in real time according to the preset sampling period, preprocessed, and combined and spliced ​​in chronological order to generate an online thermally coupled time-series feature sequence.

[0077] Feature alignment and dimension matching are performed on the online thermal coupling time series feature sequence, and then input into the water temperature prediction model. Forward inference is performed according to the preset number of future time steps to obtain the water cooling outlet temperature prediction sequence for each future time step.

[0078] Based on the predicted water-cooled outlet temperature sequence, the numerical difference and average rate of change between the predicted water-cooled outlet temperatures of adjacent future time steps are calculated in chronological order, and the numerical difference and average rate of change are combined to generate the water temperature change trend result.

[0079] The predicted water-cooled outlet temperature sequence is correlated with the water temperature change trend results according to the time index of the future time step to obtain the water temperature prediction result.

[0080] In this embodiment, the generation of the water temperature control command specifically includes:

[0081] Extract the predicted water-cooled outlet temperature values ​​corresponding to each future time step from the water temperature prediction results, and extract the target water temperature values ​​under the same time index from the preset water temperature target curve. Align the water-cooled outlet temperature prediction value sequence with the target water temperature value sequence one by one according to the time index.

[0082] At each future time step, the control deviation value for the corresponding time step is obtained by subtracting the predicted water-cooled outlet temperature value from the corresponding target water temperature value, and the control deviation values ​​for each future time step are arranged in chronological order to form a control deviation sequence.

[0083] Based on the control deviation sequence and the preset time weight, the control deviation values ​​of each future time step are weighted and summed to generate a comprehensive control deviation index that represents the overall water temperature deviation.

[0084] Based on the comprehensive control deviation index and the preset cooling flow rate adjustment coefficient, cooling power adjustment coefficient and water pump speed adjustment coefficient, the basic cooling flow rate, basic cooling power and basic water pump speed are numerically adjusted to generate cooling flow rate setpoint, cooling power setpoint and water pump speed setpoint. The three are then combined to generate a water temperature control command. The cooling flow rate setpoint, cooling power setpoint and water pump speed setpoint are generated based on the same comprehensive control deviation index and maintain an interrelated adjustment relationship during the numerical adjustment process.

[0085] In this embodiment, the generation of incremental samples specifically includes:

[0086] The system analyzes water temperature control commands and sends them to the flow control device, refrigeration device, and circulating water pump.

[0087] The control system's actuators adjust the cooling flow rate, cooling power, and circulation pressure according to the water temperature control command, and maintain the adjusted operating conditions for a preset duration.

[0088] After the adjustment is performed, the actual water-cooled outlet temperature is collected according to the preset sampling period, and the corresponding predicted water-cooled outlet temperature value is extracted from the water temperature prediction result according to the time index.

[0089] At each time index, the actual water-cooled outlet temperature is subtracted from the corresponding predicted water-cooled outlet temperature to generate a prediction error value. The prediction error value is then combined with the online thermal coupling time series feature of the corresponding time step in the online thermal coupling time series feature sequence according to the time index to generate incremental samples.

[0090] The incremental samples are written into the sample library, and the corresponding water temperature control command information and actual water cooling outlet temperature information are recorded.

[0091] In this embodiment, the incremental training of the water temperature prediction model specifically includes:

[0092] Incremental samples are read from the sample library in chronological order, and the online thermally coupled temporal features in the incremental samples are used as model inputs, while the prediction error values ​​are used as supervision signals.

[0093] Within each training batch, the online thermally coupled temporal features are input into the water temperature prediction model based on the time index of the incremental samples, and the loss value of each batch is calculated based on the loss function.

[0094] At each time step, the gradient is calculated on the loss function and applied to the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule.

[0095] At time steps that meet the update cycle conditions, the internal parameters of the medium-speed update submodule and the slow-speed update submodule are numerically adjusted. At time steps that do not meet the update cycle conditions, their internal parameters remain unchanged. The internal parameters of the fast update submodule are numerically adjusted at each time step. During incremental training, the fast update submodule, the medium-speed update submodule, and the slow update submodule absorb the gradient generated by the same incremental sample in different time periods according to their respective update cycles, thereby maintaining the consistency of the model at different time scales.

[0096] At the end of the training batch, the updated internal parameters of the fast update submodule, medium update submodule, and slow update submodule are written into the parameter storage unit.

[0097] Example 1:

[0098] To verify the feasibility of this invention in practice, it was applied to a metal laser selective melting equipment in an additive manufacturing center located in East China. This center performs continuous metal printing tasks daily. During high-load operation, the equipment frequently experiences issues such as lag in the water cooling system response, significant fluctuations in the water cooling outlet temperature, and difficulty in matching cooling capacity to changes in heat load. This leads to unstable interlayer temperatures, affecting forming accuracy and equipment stability. Before trying this invention, the manufacturing center relied mainly on fixed operating condition settings and manual adjustments based on experience. This often resulted in lag when dealing with rapid changes in nozzle power, scanning path switching, and material feed speed, especially during unattended nighttime operations, making real-time control difficult and frequently causing localized overheating in the forming area.

[0099] In practice, temperature sensors, flow sensors, and power acquisition modules are deployed in the additive manufacturing equipment and water cooling system to continuously collect multi-source time-series data, including printing parameters, nozzle power, water cooling inlet temperature, water cooling outlet temperature, and water cooling flow rate. This data is then denoised, aligned, and normalized in chronological order to form a time-series sample set, which is used to train the improved recurrent neural network model. The model employs a multi-timescale update structure: a fast update submodule captures instantaneous fluctuations in nozzle power, a medium-speed update submodule learns mid-term heat accumulation, and a slow update submodule models the overall temperature rise trend within the printing cycle. During training, a physical constraint on the consistency of heat changes is introduced, enabling the model to maintain a heat conservation trend across different time scales, guiding the parameters to converge in a physically reasonable direction.

[0100] When the equipment continuously performs high-power printing tasks in the second quarter, the water temperature prediction model constructed by this invention can receive online-collected thermal coupling time-series characteristics in real time and predict the water-cooled outlet temperature trend for multiple future time steps. In East China, where local temperatures are high and the equipment operates continuously beyond daytime shifts, traditional water temperature control methods may experience insufficient cooling due to the continuous increase in printhead power. In this scenario, this invention can predict in advance the trend of the water-cooled outlet temperature exceeding the target value for future time periods, compare the prediction result with the preset water temperature target curve, and automatically calculate the control deviation. Based on this deviation, the cooling flow rate, cooling power, and circulation pressure are dynamically adjusted, allowing the water-cooling system to enter an active control state before the actual heat load increases.

[0101] Throughout the printing process, water temperature control commands drive the water cooling system's execution components to continuously adjust relevant parameters. After each adjustment, the actual water cooling outlet temperature is collected in real time and compared with the predicted value. This difference is combined with the online thermal coupling time-series features of the corresponding time step to generate incremental samples, which are written back to the sample library for continuous updates to the water temperature prediction model. This allows the model to adapt to dynamic changes caused by factors such as daily ambient temperature, equipment thermal decay status, and material property variations.

[0102] During continuous operation from spring to early summer, the invention demonstrated its online adaptive capabilities in real-world scenarios. Through continuous incremental training, the model gradually converged to the thermal load structures of different printing tasks, maintaining high stability in the face of heat fluctuations caused by rapid power changes, complex scanning strategies, and high material feed rates. Field engineers reported that the equipment's water temperature control response was faster, the water cooling system's adjustment behavior was smoother, and it could maintain stable temperatures even during unattended nighttime operation, effectively reducing the risk of printing defects caused by overheating and improving the surface quality of the printed product.

[0103] Table 1. Performance Comparison of the Invention and Traditional Additive Manufacturing Equipment Water Temperature Control Method

[0104] Average fluctuation range of water-cooled outlet temperature (°C) 3.6 1.2 Maximum instantaneous temperature deviation (°C) 5.1 1.9 Cooling parameter response delay (seconds) 14 4 Control command adjustment frequency (times / hour) 8 23 Temperature stability rate (%) during periods of high heat load variation 72 95 Average temperature deviation (°C) during nighttime periods without monitoring 4.4 1.5 Mean prediction error (°C) after incremental training 1.7 0.6

[0105] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0106] In controlling the conventional fluctuations of water-cooled outlet temperature, traditional methods result in an average fluctuation of 3.6℃, ​​while this invention achieves only 1.2℃. The main reason for this is that this invention uses an improved recurrent neural network model to predict temperature trends in advance and combines this with a multi-timescale update mechanism to adjust system parameters in a timely manner, resulting in more precise temperature control and a smaller fluctuation range.

[0107] In scenarios involving instantaneous thermal shock, the maximum instantaneous temperature deviation of traditional methods is 5.1℃, while that of this invention is only 1.9℃. This indicates that when nozzle power changes rapidly or scanning strategies switch abruptly, this invention can identify future temperature trends in advance and generate more reasonable control commands by predicting deviations, significantly reducing the magnitude of temperature abrupt changes. The improved prediction performance stems from the addition of a physical constraint on the difference between heat increase and cooling in the loss function, making the model convergence direction more consistent with thermophysical laws and fundamentally improving prediction quality.

[0108] Regarding cooling response speed, the traditional method has a 14-second response delay for cooling parameters, while this invention only takes 4 seconds. This is because the invention generates control deviations in advance based on water temperature prediction results, allowing the system to adjust cooling flow, cooling power, and circulation pressure in advance when it identifies potential future temperature rise trends. Furthermore, the multi-timescale submodule structure can more quickly capture short-term change signals, making the control actions more timely.

[0109] Regarding the sensitivity of control actions, traditional methods adjust control commands at a frequency of only 8 times per hour, while this invention achieves 23 times per hour, indicating that the water cooling system of this invention exhibits more proactive and adaptive behavior. The higher command frequency is achieved because the framework of this invention can generate online thermal coupling timing features in real time and continuously optimize the model through incremental training, ensuring that the prediction results closely follow changes in equipment status.

[0110] During the high heat load stage, the temperature stability rate of traditional methods is only 72%, while that of this invention reaches 95%. This is because the three types of update sub-modules handle temperature changes on short, medium, and long time scales respectively, enabling the model to maintain stable control when facing continuous printing, stepped power output, or changes in material feed rate.

[0111] In nighttime unattended monitoring scenarios, the average temperature deviation of traditional methods is 4.4℃, while that of this invention is only 1.5℃. This invention maintains high stability in unattended conditions thanks to the predictive power and automatic parameter adjustment mechanism of recurrent neural networks, enabling the system to remain adaptive even without human intervention.

[0112] Regarding prediction quality, the average prediction error after incremental training for traditional methods is 1.7℃, while that of this invention is only 0.6℃, significantly lower than that of traditional methods. The actual reason lies in the synergistic effect of the physical constraint loss term and the multi-period update structure in this invention, which continuously reduces the prediction bias during training.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A water temperature control method for additive manufacturing equipment based on recurrent neural networks, characterized in that, Includes the following steps: Temperature sensors, flow sensors, and power acquisition modules are deployed in additive manufacturing equipment and water cooling systems to collect multi-source time-series data, perform preprocessing, and generate time-series sample sets. Based on the time series sample set, the improved IndRNN model is trained for multiple cycles to generate a water temperature prediction model. The generation of the water temperature prediction model specifically includes: Based on the time-series sample set, construct the thermal load feature sequence and the water cooling state feature sequence, and splice them together to generate the thermally coupled time-series feature sequence; An improved IndRNN model is constructed, which divides the hidden layer into a fast update sub-module, a medium update sub-module, and a slow update sub-module. The fast update sub-module is set to update at every time step, the medium update sub-module is set to update within a preset time step, and the slow update sub-module is set to update after a preset number of time steps. The improved IndRNN model includes dividing the hidden layer into multiple sub-modules based on different time update periods, and adding physical constraint terms that constrain the model's prediction behavior to the loss function during model training. Perform temporal unrolling on the thermally coupled temporal feature sequence, and drive the three sub-modules to update their hidden states at different time steps; For each time step, perform forward calculations, integrate the intermediate states generated by the fast update submodule, medium update submodule, and slow update submodule at the current time step according to a preset method, transform the integrated intermediate states according to a preset mapping relationship, and generate a predicted value representing the future trend of water temperature change. The sum of the squares of the numerical differences between the actual value and the corresponding predicted value of the water-cooled outlet temperature is taken as the main loss, and the sum of the squares of the numerical differences between the increase in heat caused by the nozzle power and the amount of cooling carried away by the water-cooled flow rate is taken as the additional loss. The main loss and the additional loss are weighted and summed to construct a loss function. Backpropagation is performed at each time step to calculate the gradient generated by the loss function, and the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule are updated respectively. The above operations are repeated in chronological order during the training cycle, forming a new set of parameters in each iteration, until the training error converges and the final set of parameters is solidified, thus generating a water temperature prediction model. Multi-source time-series data are collected in real time during the additive manufacturing process, and the water temperature prediction results are obtained through a water temperature prediction model. The water temperature prediction results are obtained specifically through: During the operation of the additive manufacturing task, multi-source time-series data are collected in real time according to the preset sampling period, preprocessed, and combined and spliced ​​in chronological order to generate an online thermally coupled time-series feature sequence. Feature alignment and dimension matching are performed on the online thermal coupling time series feature sequence, and then input into the water temperature prediction model. Forward inference is performed according to the preset number of future time steps to obtain the water cooling outlet temperature prediction sequence for each future time step. Based on the predicted water-cooled outlet temperature sequence, the numerical difference and average rate of change between the predicted water-cooled outlet temperatures of adjacent future time steps are calculated in chronological order, and the numerical difference and average rate of change are combined to generate the water temperature change trend result. The predicted water-cooled outlet temperature sequence is correlated with the water temperature change trend results according to the time index of the future time step to obtain the water temperature prediction result; The control deviation is calculated based on the water temperature prediction results and the preset water temperature target curve. The control parameters are adjusted based on the control deviation, and water temperature control instructions are generated. The water temperature control command is parsed and the water cooling system execution components are driven to adjust relevant parameters. After the adjustment is performed, the prediction error is calculated, and incremental samples are generated by combining online thermal coupling time series characteristics and written into the sample library. Based on incremental samples, the internal parameters of each updated sub-module of the water temperature prediction model are incrementally trained, and forward inference is performed based on the updated water temperature prediction model to generate new water temperature control commands. The incremental training of the water temperature prediction model specifically includes: Incremental samples are read from the sample library in chronological order, and the online thermally coupled temporal features in the incremental samples are used as model inputs, while the prediction error values ​​are used as supervision signals. Within each training batch, the online thermally coupled temporal features are input into the water temperature prediction model based on the time index of the incremental samples, and the loss value of each batch is calculated based on the loss function. At each time step, the gradient is calculated on the loss function and applied to the internal parameters of the fast update submodule, the medium update submodule, and the slow update submodule. At time steps that meet the update cycle conditions, the internal parameters of the medium-speed update submodule and the slow-speed update submodule are numerically adjusted. At time steps that do not meet the update cycle conditions, their internal parameters remain unchanged. The internal parameters of the fast update submodule are numerically adjusted at each time step. At the end of the training batch, the updated internal parameters of the fast update submodule, medium update submodule, and slow update submodule are written into the parameter storage unit.

2. The water temperature control method for additive manufacturing equipment based on recurrent neural networks according to claim 1, characterized in that, The multi-source time-series data includes printing parameters, nozzle power, water-cooled inlet temperature, water-cooled outlet temperature, and water-cooled flow rate. The printing parameters include scanning speed, scanning path, layer thickness, material feed speed, and printing layer index. The preprocessing includes noise reduction, time alignment, missing value completion, and numerical normalization.

3. The water temperature control method for additive manufacturing equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the water temperature control command specifically includes: Extract the predicted water-cooled outlet temperature values ​​corresponding to each future time step from the water temperature prediction results, and extract the target water temperature values ​​under the same time index from the preset water temperature target curve. Align the water-cooled outlet temperature prediction value sequence with the target water temperature value sequence one by one according to the time index. At each future time step, the control deviation value for the corresponding time step is obtained by subtracting the predicted water-cooled outlet temperature value from the corresponding target water temperature value, and the control deviation values ​​for each future time step are arranged in chronological order to form a control deviation sequence. Based on the control deviation sequence and the preset time weight, the control deviation values ​​of each future time step are weighted and summed to generate a comprehensive control deviation index that represents the overall water temperature deviation. Based on the comprehensive control deviation index and the preset cooling flow rate adjustment coefficient, cooling power adjustment coefficient and water pump speed adjustment coefficient, the basic cooling flow rate, basic cooling power and basic water pump speed are numerically adjusted respectively to generate the cooling flow rate set value, cooling power set value and water pump speed set value, and the three are combined to generate the water temperature control command.

4. The water temperature control method for additive manufacturing equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the incremental samples specifically includes: The system analyzes water temperature control commands and sends them to the flow control device, refrigeration device, and circulating water pump. The control system's actuators adjust the cooling flow rate, cooling power, and circulation pressure according to the water temperature control command, and maintain the adjusted operating conditions for a preset duration. After the adjustment is performed, the actual water-cooled outlet temperature is collected according to the preset sampling period, and the corresponding predicted water-cooled outlet temperature value is extracted from the water temperature prediction result according to the time index. At each time index, the actual water-cooled outlet temperature is subtracted from the corresponding predicted water-cooled outlet temperature to generate a prediction error value. The prediction error value is then combined with the online thermal coupling time series feature of the corresponding time step in the online thermal coupling time series feature sequence according to the time index to generate incremental samples. The incremental samples are written into the sample library, and the corresponding water temperature control command information and actual water cooling outlet temperature information are recorded.

Images