Quick cooling system and method for calcined sand cooling kiln based on deep reinforcement learning

Abstract

The application discloses a calcined sand cooling kiln rapid cooling system and method based on deep reinforcement learning, and belongs to the technical field of cooling control of granular materials. The system comprises a temperature sensing device, a first heat exchange medium injection unit, a second heat exchange medium purging unit, a rotating speed sensing device and a controller. The method obtains the internal transient temperature distribution and the maximum internal temperature gradient of the material as the thermal stress constraint through a physical information neural network; a high-level reinforcement learning model dynamically divides the flash evaporation zone, the sweeping zone and the slow cooling heat transfer zone; a low-level reinforcement learning model generates the first and second continuous wave signals, synchronizes the injection pulse with the radial tumbling frequency of the material through a phase-locked loop, and adjusts the wave form to form an overlapping zone; and after safety constraint verification, the output is output to the actuator. Through the application, the problems of violent cooling burst, dehumidification difficulty and actuator impact caused by the inability of traditional discrete control to perceive internal thermal stress are solved.

Description

Technical Field

[0001] This invention relates to a rapid cooling system and method for calcined sand cooling kilns based on deep reinforcement learning, belonging to the field of granular material cooling control technology. Background Technology

[0002] This invention relates to the field of industrial material heat treatment and cooling control technology, and more specifically, to a rapid cooling system and method for a calcined sand cooling kiln based on deep reinforcement learning.

[0003] In modern industrial production, granular materials such as calcined sand must undergo rapid and uniform cooling after high-temperature calcination. This process directly determines key quality indicators of the final product, such as mechanical strength and moisture content. Currently, the mainstream cooling technologies in the industry mainly rely on air cooling or water cooling processes. Traditional air cooling primarily relies on the sensible heat of the air for heat exchange, and its cooling efficiency has a natural physical limit. To meet the temperature reduction requirements of large-scale industrial production, extremely long cooling equipment is often required, which not only results in a large footprint but also significantly increases infrastructure and operating costs.

[0004] To overcome the low efficiency of air cooling, researchers often introduce direct water cooling technology, attempting to significantly increase the cooling rate by utilizing the enormous latent heat absorbed during the vaporization of water during phase change. However, this research approach, aimed at improving efficiency, has led to more intractable and deeper problems in practical applications. When liquid water directly contacts the surface of high-temperature particles or penetrates their micropores, the instantaneous and intense vaporization causes a sharp drop in the particle surface temperature, while the particle core remains at an extremely high temperature due to lag in heat conduction. This extreme temperature difference between the surface and the interior generates a huge thermal stress gradient within the material, leading to a "chilling cracking" phenomenon in material mechanics. This results in numerous microcracks within the material, severely compromising the overall mechanical strength of the product.

[0005] Meanwhile, efficient vaporization inevitably generates a large amount of high-humidity steam cloud, which adheres to the material. If it cannot be removed in time, it can easily cause the cooled particles to become damp and condense, failing to meet the strict moisture content requirements of the final product. To remove this high-humidity vapor, existing technologies typically employ a mechanical control strategy that increases the intensity of cold air blowing. This involves a programmable logic controller that hard switches and discretely starts and stops between water spraying and strong air blowing. However, this abrupt change in airflow and discrete square wave control logic not only causes severe secondary cold shocks to the surface of the particles that have just undergone initial cooling, further exacerbating the generation of thermal stress cracks, but also subjectes the physical actuators within the system to the fluid impact effect caused by sudden start-stop operations for extended periods, drastically increasing the mechanical wear and maintenance costs of the equipment.

[0006] A deeper examination reveals that the fundamental challenge of simultaneously achieving cooling efficiency, material strength, and moisture control lies in the inherent limitations of existing cooling systems in terms of state perception and execution. In terms of perception, existing physical sensor networks can only monitor the temperature field distribution on the material surface or in the external environment, completely failing to observe the actual temperature gradient changes within dynamically moving particles. This lack of crucial internal state parameters prevents the control system from accurately predicting the risk of impending thermal stress damage in real time. In terms of execution, constrained by traditional discrete control frameworks, the system cannot continuously and smoothly coordinate the physical quantities of different cooling media within minute time scales, inevitably leading to severe fluctuations in the heat exchange environment.

[0007] In summary, existing cooling technologies are caught in a dilemma where efficiency and quality are mutually constrained. Therefore, under the objective condition that the internal thermal stress of materials cannot be directly measured, how to overcome the thermal shock and environmental fluctuations caused by traditional mechanical discrete control, and seek a control mechanism that can achieve a dynamic balance between extreme phase change cooling and smooth release of internal thermal stress of materials, thereby completely resolving the contradiction between efficient cooling, absolute moisture protection, and ensuring high product strength, has become a key technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a rapid cooling system and method for calcined sand cooling kilns based on deep reinforcement learning, which solves the problems of rapid cooling and cracking, dehumidification difficulties and actuator impact caused by the inability of traditional discrete control to sense internal thermal stress.

[0009] The technical problem to be solved by this invention is achieved by the following technical solution: On the one hand, this invention provides a rapid cooling method for calcined sand cooling kilns based on deep reinforcement learning, comprising: Obtain real-time status information and cooling status of the target object; Based on the real-time status information and the target cooling status, a first continuous waveform signal and a second continuous waveform signal are generated. The first continuous waveform signal is used to control the injection intensity of the first heat exchange medium, and the second continuous waveform signal is used to control the purging intensity of the second heat exchange medium. The first heat exchange medium is used to undergo a phase change to absorb heat when it comes into contact with the target object. The first continuous waveform signal and the second continuous waveform signal are respectively output to the corresponding actuators to control the first heat exchange medium and the second heat exchange medium to overlap with non-discontinuous continuous waveforms with a fixed timing relationship on the time axis, so that the temperature of the target object approaches the target cooling state.

[0010] The present invention is further configured such that: the target object is located inside a rotary cooling device, moves axially and tumbles radially; The pulse frequency of the first continuous waveform signal is associated with the radial tumbling frequency of the target object within the rotary cooling device, so that the jet pulse of the first heat exchange medium is aligned with the moment when the target object is exposed to the surface.

[0011] The present invention is further configured such that: the real-time status information includes the surface temperature distribution of the target object; after acquiring the real-time status information of the target object, it further includes: Based on the tilt angle, rotation speed, and feed flow rate of the rotary cooling equipment, calculate the axial moving speed and residence time distribution of the target object; The surface temperature distribution, the axial moving speed, and the current values ​​of the first and second continuous waveform signals are input into the physical information neural network model. The physical information neural network model is trained by using the partial differential equation describing the unsteady heat conduction process inside the target object as a constraint term of the loss function, and outputs the transient temperature distribution of the target object from the surface to the core. The maximum internal temperature gradient is calculated based on the transient temperature distribution, and the maximum internal temperature gradient is used as a thermal stress constraint condition for the generation of subsequent control signals.

[0012] The present invention is further configured such that: the generation of the first continuous waveform signal and the second continuous waveform signal specifically includes: The high-level reinforcement learning model dynamically divides the spatial proportion of virtual temperature zones based on the feed temperature, target discharge temperature, feed flow rate, current rotation speed and filling rate of the rotary cooling equipment, and in combination with the axial moving speed and residence time distribution. The virtual temperature zones include flash zone, sweep zone and slow cooling heat transfer zone. The space ratio of the flash zone is positively correlated with the feed temperature and the target cooling rate, and is used to allocate the resource ratio of phase change cooling; the space ratio of the slow cooling heat transfer zone is positively correlated with the residence time and the thermal stress release demand represented by the maximum internal temperature gradient, and is used to allocate the resource ratio of internal heat conduction. Based on the spatial proportion of the virtual temperature zone, the multiple independently controlled first heat exchange medium injection units and second heat exchange medium purging units arranged along the axial direction are dynamically divided into groups corresponding to the flash zone, sweeping zone and slow cooling heat transfer zone, respectively. Within the dynamically divided group, the underlying reinforcement learning model generates the first continuous waveform signal and the second continuous waveform signal based on the maximum internal temperature gradient, exhaust humidity, and radial tumbling frequency of the target object at the current moment. The radial tumble frequency is calculated in real time based on the rotational speed, filling rate of the rotary cooling equipment and the dynamic rest angle of the target object. The underlying reinforcement learning model synchronizes the pulse frequency of the first continuous waveform signal with the radial tumble frequency through a phase-locked loop algorithm, so that the peak of the first continuous waveform signal is aligned with the moment when the target object tumbles from the passive core area to the active surface layer.

[0013] The present invention is further configured such that: the underlying reinforcement learning model also dynamically adjusts the waveform shape of the first continuous waveform signal and the second continuous waveform signal according to the maximum internal temperature gradient and the exhaust humidity, generating an asymmetric sine wave or quasi-sine wave, so that the rising phase and falling phase of the two overlap in time to form a waveform overlap area. Specifically, when the maximum internal temperature gradient is greater than a preset safety threshold, the duration of the waveform overlap region is increased; when the exhaust humidity is greater than a preset dew point threshold, the duration of the waveform overlap region is decreased.

[0014] The present invention is further configured such that, through waveform shape control of the first continuous waveform signal and the second continuous waveform signal, the following stages are sequentially formed within one cooling cycle: The rising edge of the first continuous waveform signal corresponds to the flash evaporation endothermic period, during which the injection rate of the first heat exchange medium increases. The waveform overlap region corresponds to the transition stage from the flash evaporation endothermic period to the mild sweeping period. During this period, the injection amount of the first heat exchange medium decreases and the purging amount of the second heat exchange medium increases. The peak of the second continuous waveform signal corresponds to the mild sweeping period, during which the second heat exchange medium is subjected to strong purging, and the second heat exchange medium is the waste heat warm air recovered from the tail of the rotary cooling device. The falling edge of the second continuous waveform signal corresponds to the internal heat conduction period, during which the purging intensity of the second heat exchange medium is reduced.

[0015] The present invention is further configured such that: the underlying reinforcement learning model is trained using a multi-objective reward function, the reward function comprising: The positive reward is positively correlated with the instantaneous temperature drop of the target object; The first penalty term is positively correlated with the square of the maximum internal temperature gradient exceeding the dynamic temperature gradient threshold. The second penalty is positively correlated with the degree to which the exhaust humidity approaches the saturation dew point; The dynamic temperature gradient threshold is calculated in real time based on the real-time average temperature of the target object and the nonlinear coupling relationship between the material's yield strength, elastic modulus, and thermal expansion coefficient as the temperature changes, so that the dynamic temperature gradient threshold has a higher value in the high-temperature range and a lower value in the low-temperature range.

[0016] The present invention is further configured such that, before outputting the first continuous waveform signal and the second continuous waveform signal to the corresponding actuator, a safety constraint step is included: Calculate the maximum theoretical water carrying capacity under the current operating conditions based on real-time wind speed, exhaust temperature, and dew point temperature. If the instantaneous jet volume corresponding to the amplitude of the first continuous waveform signal exceeds the maximum theoretical water-carrying capacity, the amplitude of the first continuous waveform signal is corrected to the upper limit value corresponding to the maximum theoretical water-carrying capacity, and a negative reward is fed back to the control model that generated the waveform signal.

[0017] The present invention is further configured such that the generation models of the first continuous waveform signal and the second continuous waveform signal are obtained through the following steps: Historical operational data was collected, and the conservative Q-learning algorithm was used for offline pre-training to obtain the initial model. The initial model is optimized through reinforcement learning in a virtual twin environment constructed using physical information neural networks and computational fluid dynamics to update the model parameters; The updated model was deployed to the edge computing gateway and fine-tuned online using the flexible actor-critic algorithm, under the constraints of physical security rules.

[0018] On the other hand, the present invention provides a cooling system for use in rotary cooling equipment, characterized in that it comprises: Temperature sensing device, used to acquire the real-time surface temperature distribution and spatial temperature field of the target object; Multiple axially independently controlled first heat exchange medium injection units, each injection unit including at least one nozzle and a corresponding servo proportional valve, are used to independently adjust the injection intensity of the area according to the first continuous waveform signal; Multiple axially independently controlled second heat exchange medium purging units, each purging unit including at least one damper or fan, for independently adjusting the purging intensity of the area according to the second continuous waveform signal; A rotation speed and tilt angle sensing device is used to acquire the real-time rotation speed and tilt angle parameters of the rotary cooling equipment; The controller is communicatively connected to the temperature sensing device, the speed and tilt sensing device, the plurality of first heat exchange medium injection units and the plurality of second heat exchange medium purging units, and is configured to perform the cooling method as described in any one of claims 1 to 9.

[0019] The beneficial effects of this invention are: This invention employs the non-discontinuous overlap of a first continuous waveform signal and a second continuous waveform signal on the time axis. Through dynamic adjustment of the waveform shape and adaptive variation of the waveform overlap region duration, the two heat exchange media form a coupling relationship of mutual restraint during the rising and falling phases, avoiding oscillations in the heat exchange environment and fluid impacts on the actuator caused by sudden changes in the physical quantities of the media, thereby maintaining the stability of the heat and mass exchange boundary on a microscopic time scale. By introducing a physical information neural network into the control system, the partial differential equations of the unsteady heat conduction process inside the material are embedded as constraints in the model training. This enables the system to inverse the transient temperature distribution from the surface to the core based on the measurable surface temperature distribution, and then calculate the maximum internal temperature gradient as a thermal stress constraint. On this basis, the high-level reinforcement learning model dynamically divides the virtual temperature zone according to the feeding conditions and rationally allocates the resource ratio between phase change cooling and internal heat conduction. The low-level reinforcement learning model combines real-time states such as radial tumbling frequency, maximum internal temperature gradient, and exhaust humidity to generate a continuous waveform signal with frequency and phase lock, so that the injection pulse is precisely aligned with the moment when the material tumbles to the surface. The above-mentioned hierarchical architecture achieves the decoupling of macroscopic resource scheduling and microscopic waveform control, ensuring cooling efficiency while keeping the internal thermal stress of the material within a controllable range. The underlying reinforcement learning model is trained using a multi-objective reward function. A positive reward term is associated with the cooling rate, a first penalty term is associated with the degree to which the maximum internal temperature gradient exceeds a dynamic threshold, and a second penalty term is associated with the degree to which exhaust humidity approaches the saturation dew point. The dynamic temperature gradient threshold is calculated in real-time based on the nonlinear relationship between the material's yield strength, elastic modulus, and coefficient of thermal expansion with temperature, allowing for a larger temperature gradient in the high-temperature range and strictly limiting it in the low-temperature range. Furthermore, the system sets a safety constraint step before outputting the control signal, calculating the maximum theoretical water-carrying capacity based on real-time wind speed, exhaust temperature, and dew point, limiting the instantaneous injection volume, and providing negative rewards. This mechanism enables the cooling process to automatically adjust the control strategy under different temperature ranges and operating conditions, maintaining the boundary conditions of internal material stress and surface dryness while meeting the cooling rate requirements, thus achieving a dynamic balance between cooling efficiency, product strength, and moisture resistance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the overall method of the present invention.

[0021] Figure 2 This is a schematic diagram of the material state inside the rotary cooling device of the present invention.

[0022] Figure 3 This is a timing diagram showing the overlap of the first and second medium waveforms during the cooling cycle of this invention. Detailed Implementation

[0023] The following detailed description of the specific implementation methods further elaborates on the technical details and implementation principles of the rapid cooling system and method for calcined sand cooling kilns based on deep reinforcement learning of the present invention.

[0024] In modern industrial production, granular materials such as calcined sand must undergo rapid and uniform cooling after high-temperature calcination. This process directly determines key quality indicators of the final product, such as mechanical strength and moisture content. To overcome the low efficiency of traditional air cooling and the problems of thermal shock cracking and material condensation caused by direct water cooling, such as... Figure 1 As shown, this invention provides a rapid cooling method for calcined sand cooling kilns based on deep reinforcement learning. This method achieves a dynamic balance between extreme phase change cooling and smooth release of internal thermal stress in the material by coordinating the physical quantities of different cooling media. This flexible cooling method includes the following main steps.

[0025] S1 obtains the real-time status information of the target object and the target cooling status.

[0026] In actual industrial production environments, the target object is usually a granular material that has just undergone a high-temperature calcination process, such as calcined quartz sand or bauxite at temperatures as high as 800 to 1000 degrees Celsius. The real-time status information covers the dynamic physical parameters of the target object inside the cooling equipment, including but not limited to the surface temperature distribution, spatial position, movement speed of the target object, and the temperature and humidity parameters of the environment. The target cooling state is set according to the discharge temperature range and the upper limit standard of material moisture content according to the issued production process requirements. It is usually required to safely reduce the material temperature to below 100 degrees Celsius and keep the surface absolutely dry.

[0027] S2 generates a first continuous waveform signal and a second continuous waveform signal based on real-time status information and target cooling status.

[0028] The first continuous waveform signal is configured to control the injection intensity of the first heat exchange medium, which is usually atomized liquid water. Its core function is to utilize the huge latent heat absorbed by the violent vaporization phase change when the liquid water comes into contact with the surface of the high-temperature target object to achieve rapid stripping of the surface temperature of the target object.

[0029] The second continuous waveform signal is configured to control the purging intensity of the second heat exchange medium, which is usually a dry gas with a certain pressure. Its core function is to quickly remove the local high humidity vapor cloud generated by the phase change of the first heat exchange medium, and at the same time remove some sensible heat through convection heat exchange. The core purpose of generating these two waveform signals is to abandon the mechanical discrete start-stop operation in traditional control, and instead achieve smooth and continuous injection of cold energy within a small time scale, thereby avoiding violent fluctuations in the heat exchange environment.

[0030] S3 outputs the first continuous waveform signal and the second continuous waveform signal to the corresponding actuators to control the first heat exchange medium and the second heat exchange medium to overlap with non-discontinuous continuous waveforms with a fixed timing relationship on the time axis so that the temperature of the target object approaches the target cooling state.

[0031] The actuators include precision adjustment devices such as servo proportional valves and variable frequency fans that can respond to analog signals. The non-discontinuous continuous waveform overlap with a fixed timing relationship means that within any tiny time slice, the injection of the first heat exchange medium and the purging of the second heat exchange medium are not isolated, but exhibit a dynamic coupling relationship of mutual restraint and cooperation. This overlap can maintain an efficient and stable heat and mass exchange boundary layer on the surface of the target object, ensuring both the high efficiency of phase change heat absorption and completely eliminating the local accumulation of high-humidity water vapor.

[0032] To further adapt to the unique motion characteristics of granular materials within a rotary cooling device, the target object is located within the rotary cooling device, moving axially and tumbling radially. The rotary cooling device is typically a large cylinder with a certain angle of inclination. During rotation, the material is lifted to a certain height under the combined action of friction and gravity and then slides down, forming an active surface layer and a relatively static passive core area. The heat transfer efficiency inside the material and the external heat exchange efficiency are highly dependent on this radial tumbling behavior. Therefore, the pulse frequency of the first continuous waveform signal is correlated with the radial tumbling frequency of the target object within the rotary cooling device, so that the jet pulse of the first heat exchange medium is aligned with the moment when the target object is exposed to the surface.

[0033] The radial tumbling frequency depends not only on the mechanical rotation speed of the rotating cylinder, but also on the current filling rate of the material, the particle size distribution, and the dynamic angle of repose of the material that changes dynamically with temperature. By precisely aligning the injection pulse with the moment when the material tumbles to the surface, it is possible to ensure that the phase change cooling medium acts directly on the surface of the fresh material that has just tumbled out from the inside and carries extremely high heat, maximizing the heat exchange temperature difference, while avoiding the cold medium being sprayed on the already cooled material surface, which would lead to localized overcooling and resource waste.

[0034] In implementing the aforementioned flexible cooling process, the key to preventing material bursting due to rapid cooling lies in the accurate control of internal thermal stress. Since existing physical sensing networks can only monitor the surface temperature of the material and cannot observe the actual temperature gradient changes within the particles, this invention introduces physical information neural network technology. Real-time status information includes the surface temperature distribution of the target object. After acquiring the real-time status information of the target object, the following sub-steps are further included to assess internal thermal stress.

[0035] S31 calculates the axial moving speed and residence time distribution of the target object based on the tilt angle, rotation speed and feed flow rate of the rotary cooling equipment.

[0036] The tilt angle and rotation speed determine the basic movement trend of the material under the action of the axial component of gravity, while the change in bed thickness caused by the feed flow rate will change the geometric boundary of the material sliding. Based on the kinematic model, the average moving speed of a specific batch of material in each axial section of the cooling kiln and the probability density distribution of residence time in the entire cooling equipment can be accurately calculated.

[0037] S32 inputs the surface temperature distribution, axial moving speed, and the current values ​​of the first and second continuous waveform signals into the physical information neural network model. The physical information neural network model is trained by embedding a partial differential equation describing the unsteady heat conduction process inside the target object as a constraint term of the loss function, and outputs the transient temperature distribution of the target object from the surface to the core.

[0038] Traditional data-driven deep learning models are prone to violating fundamental laws of thermodynamics when predicting temperature fields. However, the physical information neural network model used in this invention deeply integrates physical mechanisms into its structure. This physical information neural network model can employ a fully connected feedforward neural network structure, for example, containing 8 hidden layers with 256 neurons per layer. The activation function is the Tanh function, which has smooth derivative characteristics. The partial differential equations describing the unsteady heat conduction process inside the target object are typically constructed based on Fourier's law of heat conduction, expressed as follows: ,in Represents material density, Represents specific heat capacity. This represents the thermal conductivity.

[0039] The input layer of the neural network receives data from multiple sensor sources, and the output layer outputs the predicted temperature values ​​on the three-dimensional spatial grid nodes. During model training, the total loss function includes the mean square error of data matching, the mean square error of partial differential equation residuals, and the mean square error of boundary condition constraints.

[0040] Its total loss function can be further defined as The weight coefficients can be set to 5:3:2 in the initial state, and adaptive weight adjustment is used at the end of the training convergence period to eliminate the influence of stiffness gradient. The physical information neural network model adds the automatic differentiation result of the above partial differential equation as a penalty term to the loss function, which forces the network to strictly follow the physical laws of energy conservation and heat conduction when fitting the observed data. Thus, it can reverse-engineer the transient temperature distribution profile from the surface to the core of the particle with high fidelity when only surface temperature observations and external cold and heat source boundary conditions are input.

[0041] S33 calculates the maximum internal temperature gradient based on the transient temperature distribution and uses the maximum internal temperature gradient as a thermal stress constraint condition for the generation of subsequent control signals.

[0042] According to solid mechanics theory, the internal thermal stress of a material is positively correlated with the temperature gradient, the coefficient of linear expansion of the material, and the elastic modulus. The maximum internal temperature gradient is the maximum value of the absolute value of the spatial derivative in the transient temperature distribution field. Using it as a hard constraint condition for subsequent control can fundamentally prevent the phenomenon of rapid cooling and cracking in the material mechanics and ensure the overall mechanical strength of the product.

[0043] After obtaining the key internal thermal stress constraint indicators, this invention employs a hierarchical deep reinforcement learning model to generate high-dimensional continuous control signals to address the complex multivariable coupled nonlinear control challenges of the cooling system. The generation of the first and second continuous waveform signals specifically includes the following sub-steps: The S41 high-level reinforcement learning model dynamically divides the spatial proportion of virtual temperature zones based on the feed temperature, target discharge temperature, feed flow rate, current rotation speed and fill rate of the rotary cooling equipment, combined with axial movement speed and residence time distribution. Figure 2 As shown, the virtual temperature zone includes a flash zone, a sweep zone, and a slow cooling heat transfer zone.

[0044] The high-level reinforcement learning model is mainly responsible for the scheduling and allocation of macro-level resources. Its action space is a continuous spatial length ratio. These three virtual temperature zones are not static partitions on physical equipment, but rather control strategy execution areas dynamically mapped by the system at the logical level according to the current thermodynamic state of the material.

[0045] The spatial proportion of the flash zone is positively correlated with the feed temperature and the target cooling rate, and is used to allocate the proportion of resources for phase change cooling. When the feed temperature is extremely high or a large cooling is urgently needed, the high-level reinforcement learning model will expand the proportion of the flash zone and mobilize more axial spray units to engage in liquid water phase change heat absorption. The spatial proportion of the slow cooling heat transfer zone is positively correlated with the residence time and the thermal stress release demand represented by the maximum internal temperature gradient, and is used to allocate the proportion of resources for internal heat conduction. When the internal temperature gradient of the material approaches the destruction limit, the system must provide enough space and time for the high temperature of the particle core to be transferred smoothly to the surface through natural heat conduction. At this time, the high-level reinforcement learning model will significantly increase the length of the slow cooling heat transfer zone, and in this area, the forced cooling of the external cold source will be greatly weakened or even suspended.

[0046] S42 dynamically divides multiple independently controlled first heat exchange medium injection units and second heat exchange medium purging units arranged along the axial direction into groups corresponding to the flash zone, sweep zone, and slow cooling heat transfer zone, respectively, according to the spatial proportion of the virtual temperature zone.

[0047] Through this dynamic grouping mechanism, hundreds of actuators along the axial direction of the cooling equipment are logically integrated into several control domains that work together, completely breaking the traditional fixed cooling section division mode of equipment and realizing flexible area reconstruction based on the real-time status of materials.

[0048] The S43 underlying reinforcement learning model generates a first continuous waveform signal and a second continuous waveform signal within the dynamically divided groups based on the maximum internal temperature gradient, exhaust humidity, and radial tumbling frequency of the target object at the current moment.

[0049] The underlying reinforcement learning model acts as a precise operator directly controlling the actuators. Its task is to generate microsecond-level high-frequency continuous control waveforms for each specific group within the policy framework defined by the higher-level model. The radial tumble frequency is calculated in real time based on the rotational speed, fill rate, and dynamic repose angle of the target object of the rotary cooling equipment. The underlying reinforcement learning model synchronizes the pulse frequency of the first continuous waveform signal with the radial tumble frequency using a phase-locked loop (PLL) algorithm, aligning the peak of the first continuous waveform signal with the moment when the target object tumbles from the passive core region to the active surface layer. The application of the PLL algorithm here is crucial for achieving precise cooling.

[0050] In practical implementation, a Fast Fourier Transform (FFT) can be performed on the radar point cloud data or motor current inside the rotary kiln to extract the main frequency characterizing the material bed tumbling cycle as a reference signal. By extracting the characteristic frequency of the material bed tumbling pattern as a reference signal, the voltage-controlled oscillator logic module within the phase-locked loop continuously adjusts the output frequency and phase of the first continuous waveform signal until they maintain a strictly fixed phase difference. To ensure targeting efficiency, the system can require the phase error of the output waveform pulse to be locked within a certain range. Within this range. This means that no matter how the kiln rotation speed fluctuates, the maximum spray volume of atomized water always precisely hits the material that is crossing the material peak and exposing a large amount of fresh, high-temperature surface, greatly improving the targeting efficiency of phase change cooling.

[0051] The underlying reinforcement learning model not only controls the frequency and phase, but also focuses on adjusting the microscopic shape of the waveform. The underlying reinforcement learning model also dynamically adjusts the waveform shape of the first and second continuous waveform signals according to the maximum internal temperature gradient and exhaust humidity to generate asymmetric sine waves or quasi-sine waves, so that the rising and falling phases of the two overlap in time to form a waveform overlap region.

[0052] In a preferred embodiment, the mathematical expression of this asymmetric sine wave or quasi-sine wave can be abstracted as the superposition of the fundamental wave and the second harmonic: The amplitude and phase parameters are generated in real time by the control model. This asymmetric sinusoidal waveform design can provide a smoother rate of change of physical quantities than discrete square waves, significantly reducing the mechanical impact on the fluid pipeline of the actuator. The duration of the waveform overlap region is a highly dynamic adaptive parameter.

[0053] When the maximum internal temperature gradient exceeds a preset safety threshold, the duration of the waveform overlap zone is increased. Within this overlap zone, the water spray intensity gradually decreases while the purging intensity gradually increases. Extending this zone slows the surface cooling rate, providing a buffer time for internal heat to conduct outwards, thus effectively mitigating the thermal stress peak generated by rapid cooling. Conversely, when the exhaust humidity exceeds a preset dew point threshold, the duration of the waveform overlap zone is decreased. This means the system will more quickly switch to a powerful purging state, using a short, intense airflow to break through the vapor coating layer, accelerating the extraction of high-humidity gas and preventing water vapor condensation.

[0054] The waveform shapes of the first and second continuous waveform signals are controlled to sequentially form the following four subdivided physical stages within one cooling cycle: In the first stage, the rising edge of the first continuous waveform signal corresponds to the flash evaporation endothermic period, during which the injection rate of the first heat exchange medium increases. At this time, the material has just rolled to the surface, and the high temperature triggers the instantaneous vaporization of liquid water, absorbing a large amount of latent heat, causing the surface temperature of the material to drop sharply.

[0055] In the second stage, the overlapping waveform zone corresponds to the transition from the flash evaporation endothermic period to the mild sweeping period. During this period, the injection rate of the first heat exchange medium decreases while the purging rate of the second heat exchange medium increases. As vaporization proceeds, a steam cloud forms on the surface, hindering further heat exchange. At this point, reducing water spray can prevent moisture from remaining in the material pores, while increasing air blowing begins to strip away the steam cloud.

[0056] In the third stage, the peak of the second continuous waveform signal corresponds to the mild sweeping period. During this period, the second heat exchange medium is strongly swept. The second heat exchange medium is the waste heat warm air recovered from the tail of the rotary cooling equipment. Using waste heat warm air for sweeping has two benefits. On the one hand, its low relative humidity has a strong water carrying and dehumidification capacity. On the other hand, its own warm properties avoid the secondary cold impact on the material surface caused by room temperature cold air, making the cooling process smoother.

[0057] In the fourth stage, the falling edge of the second continuous waveform signal corresponds to the internal heat conduction period. During this period, the purging intensity of the second heat exchange medium weakens. At this time, the forced heat exchange effect of the external environment on the material is reduced to a minimum. The material continues to tumble and mix inside the bed. The high-temperature heat in the core area is continuously transferred to the lower-temperature surface through the internal heat conduction mechanism of the solid, so that the overall temperature field of the material tends to be homogenized and the thermal stress is fully released.

[0058] like Figure 3 As shown, within a complete cooling cycle, the first continuous waveform signal (curve A, corresponding to the first heat exchange medium injection intensity) and the second continuous waveform signal (curve B, corresponding to the second heat exchange medium purging intensity) overlap in an asymmetric sinusoidal shape, forming four stages in sequence: The first stage is the flash evaporation endothermic period, corresponding to the rising edge of the first continuous waveform signal. At this time, the injection volume of the first heat exchange medium increases, and the high temperature on the material surface triggers the instantaneous vaporization of liquid water, absorbing a large amount of latent heat; the second stage is the transition zone, corresponding to the overlap zone of the two waveforms, where the first continuous waveform signal decreases while the second continuous waveform signal decreases. As the waveform signal rises, the injection rate of the first heat exchange medium decreases while the purging rate of the second heat exchange medium increases, achieving a smooth transition from phase change cooling to purging dehumidification. The third stage is the gentle sweeping period, corresponding to the peak of the second continuous waveform signal. During this period, the second heat exchange medium performs a strong purging, rapidly carrying away the high-humidity vapor cloud generated by the phase change. The fourth stage is the internal heat conduction period, corresponding to the falling edge of the second continuous waveform signal. The purging intensity of the second heat exchange medium weakens, the external forced heat exchange effect is minimized, and the internal heat of the material is transferred to the surface through the heat conduction mechanism, making the overall temperature field tend to be homogenized. The shaded area in the figure represents the waveform overlap region, the duration of which is a dynamically adjustable parameter: when the maximum internal temperature gradient is greater than a preset safety threshold, the overlap region duration is increased to slow down the surface cooling rate; when the exhaust humidity is greater than a preset dew point threshold, the overlap region duration is decreased to accelerate dehumidification.

[0059] To guide the reinforcement learning model to autonomously learn the aforementioned complex collaborative control strategies, the underlying reinforcement learning model is trained using a multi-objective reward function. This reward function is designed to fully balance cooling efficiency, material quality, and moisture resistance requirements, and comprises the following three core components: Positive reward items are positively correlated with the instantaneous temperature drop of the target object.

[0060] The faster the model causes the material temperature to drop per unit time, the greater the numerical reward it receives. This drives the model to continuously explore combinations of heat exchange actions with higher intensity to ensure the overall output and efficiency of the cooling kiln.

[0061] The first penalty term is positively correlated with the square of the maximum internal temperature gradient exceeding the dynamic temperature gradient threshold.

[0062] The use of a squared penalty mechanism means that once the temperature gradient exceeds the limit, the penalty will be amplified exponentially, thereby imposing a severe constraint on the behavior that damages the material strength. The dynamic temperature gradient threshold is calculated in real time based on the real-time average temperature of the target object and the nonlinear coupling relationship between the material's yield strength, elastic modulus and thermal expansion coefficient that change with temperature. This results in a higher value for the dynamic temperature gradient threshold in the high-temperature range and a lower value in the low-temperature range.

[0063] The specific solution formula is constructed based on the thermodynamic fracture criterion: ,in The yield strength is dynamically reduced with temperature. For elastic modulus, The coefficient of thermal expansion is... It is Poisson's ratio.

[0064] The system can pre-set fitting curve matrices of the physical properties of different materials (such as quartz sand and bauxite) as a function of temperature for real-time reference. At high temperatures, the materials typically exhibit certain plastic characteristics, capable of withstanding large thermal deformations without brittle fracture. Therefore, a relaxed threshold allows for more aggressive cooling. However, in the low-temperature range, the materials become completely embrittled, and even a small temperature gradient can trigger cracks. Therefore, the threshold must be significantly tightened, forcing the model to switch to an extremely mild cooling strategy.

[0065] The second penalty is positively correlated with the degree to which exhaust humidity approaches the saturation dew point.

[0066] This penalty-forced model constantly monitors the humidity balance within the system to avoid excessive pursuit of phase change heat absorption, which could lead to a large accumulation of water vapor and ultimately fail to meet the product output moisture content target.

[0067] Before outputting the first and second continuous waveform signals to the corresponding actuators, S5 also includes a safety constraint step to ensure the absolute physical safety of the system under any extreme operating conditions: S51 calculates the maximum theoretical water-carrying capacity under the current operating conditions based on real-time wind speed, exhaust temperature, and dew point temperature. The maximum theoretical water-carrying capacity reflects the limit of water vapor mass that the current exhaust system can remove without condensation. This calculation process derives the saturated vapor pressure corresponding to the current exhaust temperature based on the Antoine equation in thermodynamics, and solves for the mass flow limit by combining real-time wind speed and pipe cross-sectional area.

[0068] The form of the Antoine equation it uses is: ,in It is the saturated vapor pressure. The exhaust temperature, It is a constant, combined with wind speed and pipe cross-sectional area The formula for calculating the maximum mass flow rate limit can be expressed as follows: ,in This is the on-site dust humidity correction factor (e.g., take 0.9).

[0069] S52 If the instantaneous jet volume corresponding to the amplitude of the first continuous waveform signal exceeds the maximum theoretical water carrying capacity, the amplitude of the first continuous waveform signal is corrected to the upper limit value corresponding to the maximum theoretical water carrying capacity, and a negative reward is fed back to the control model that generated the waveform signal.

[0070] This safety truncation mechanism, based on hard physical rules, serves as the last line of defense outside the reinforcement learning model. It ensures that no matter what exploratory action the model outputs, it will not inject more excess into the system than it can expel. This safety mechanism is a hard limit constraint and remains effective during the online fine-tuning phase. When the limit is triggered, the system injects a large constant negative penalty (such as -10) into the gradient of the reinforcement learning model network. This strong feedback negative reward can prompt the reinforcement learning model to automatically avoid such invalid or dangerous action regions that will be truncated in future trial and error, thus accelerating policy convergence.

[0071] Direct trial and error with deep reinforcement learning models in real-world industrial environments is extremely costly and risky. Therefore, the generation models for the first and second continuous waveform signals are trained using a hybrid architecture with the following three stages: Phase 1 involves collecting historical operational data and performing offline pre-training using the conservative Q-learning algorithm to obtain the initial model. The historical data includes records of the system's past operations under manual or traditional logic control, as well as sensor states. The conservative Q-learning algorithm, by adding a regularization penalty term to the objective function, suppresses the model from overestimating unfamiliar actions not present in the dataset, effectively avoiding the curse of out-of-distribution state overestimation common in offline reinforcement learning. This ensures that the pre-trained initial model possesses robust basic control capabilities. In this pre-training phase, the Adam network optimizer is selected, and the initial learning rate can be set to [value missing]. Discount factor for And it is combined with an experience playback mechanism.

[0072] Phase Two involves optimizing the initial model through reinforcement learning in a virtual twin environment constructed using physical information neural networks and computational fluid dynamics (CFD) to update model parameters. This phase automatically initiates when the pre-trained model's output divergence falls below a set threshold. The virtual twin environment is a high-fidelity digital simulation platform. The physical information neural network simulates the internal state evolution of materials, while the CFD software accurately solves the complex two-phase flow and aerodynamic field within the cooling kiln. In this environment free from physical damage risk, the model undergoes millions of virtual interactive scenarios, fully exploring various extreme injection combinations and learning to handle nonlinear thermodynamic boundary conditions, significantly broadening the effective coverage of the strategy.

[0073] Phase Three involves deploying the updated model to the edge computing gateway and fine-tuning it online using the Flexible Actor-Critics Algorithm (FAC), a sophisticated algorithm based on the maximum entropy principle. This algorithm maximizes cumulative rewards while maintaining the randomness of action strategies. This allows the model deployed on the industrial edge computing gateway to have continuous environmental adaptability and self-evolution capabilities under the hard protection of the aforementioned security constraints, automatically compensating for model biases caused by wear and tear of real physical equipment or seasonal climate changes.

[0074] To support the operation of the aforementioned complex control logic and algorithms, this invention provides a cooling system, which includes a hardware perception layer, an execution layer, and a core control decision layer.

[0075] The system is equipped with a temperature sensing device to acquire the real-time surface temperature distribution and spatial temperature field of the target object. The device includes an array of high-resolution infrared thermal imagers and radar temperature probes resistant to high-temperature dust. It can not only draw a two-dimensional temperature topology map of the material surface inside the kiln, but also penetrate the steam layer to measure the ambient temperature distribution of the kiln space, providing a massive high-dimensional data source for calculating the transient temperature field of the material.

[0076] At the execution level, the system is equipped with multiple independently controlled first heat exchange medium injection units along the axial direction. Each injection unit includes at least one nozzle and a corresponding servo proportional valve, which is used to independently adjust the injection intensity of the area according to the first continuous waveform signal. The servo proportional valve has extremely high response bandwidth and can perfectly analyze and execute the continuous asymmetric sine wave command output by the underlying model, driving the specially designed atomizing nozzle to convert high-pressure water into micron-level ultrafine water mist, so as to maximize the specific surface area and promote instantaneous phase change.

[0077] The coordinated execution also includes multiple axially independently controlled second heat exchange medium purging units. Each purging unit includes at least one damper or fan, used to independently adjust the purging intensity of its area according to the second continuous waveform signal. These purging units achieve stepless smooth adjustment of air volume through frequency conversion drive technology. Their air inlets are connected to the waste heat recovery system at the tail of the rotary kiln through insulated pipes, continuously drawing dry hot air for the stripping and sweeping of steam clouds.

[0078] To capture the kinematic state of the material, the system integrates a speed and tilt angle sensing device to obtain the real-time speed and tilt angle parameters of the rotary cooling equipment. This device transmits the mechanical operation posture data of the equipment to the control center at a millisecond frequency through a high-precision rotary encoder and tilt angle sensor, which serves as a key reference for calculating the material tumbling frequency and phase-locked loop synchronization.

[0079] The core of the system is the controller, which is communicatively connected to temperature sensing devices, rotational speed and tilt angle sensing devices, multiple first heat exchange medium injection units, and multiple second heat exchange medium purging units, and is configured to execute the aforementioned flexible cooling method steps. The controller is built using a heterogeneous computing architecture, integrating a high-performance graphics processor cluster for processing massive amounts of visual and sensor data, and a tensor processing unit for performing online inference and fine-tuning computations in reinforcement learning.

[0080] Through the industrial Ethernet protocol, the controller interacts with all peripheral devices with extremely low latency, transforming abstract deep reinforcement learning strategies into smooth waveform commands that drive the efficient operation of heavy equipment weighing thousands of tons in real time. This completely solves the dilemma of thermal shock and environmental fluctuation caused by traditional mechanical discrete control, achieving a perfect dynamic balance between efficient cooling, absolute moisture protection, and ensuring high product strength.

[0081] In practical applications: When faced with extremely high feed temperatures exceeding 900°C, the system can instantly and adaptively expand the flash zone ratio and tighten the dynamic temperature gradient threshold based on internal algorithms. This achieves second-level surface heat stripping while effectively preventing thermal stress microcracks. In extremely humid climates, the underlying algorithm automatically reduces the waveform overlap time and increases the frequency of warm air purging to ensure that the final product output moisture content meets the anti-condensation index. Furthermore, in the face of sudden changes in feed flow or the degradation of fan mechanical performance due to long-term operation, the online fine-tuning mechanism based on the Flexible Actor-Critic (SAC) algorithm continuously updates the network weights, thereby automatically compensating for equipment deviations and demonstrating strong adaptive adjustment capabilities and life-cycle stability.

[0082] The above-described implementation details constitute the technical closed loop of this invention. Without departing from the spirit and scope of this invention, those skilled in the art can make various adaptive changes in the specific network topology selection, thermodynamic parameter correction, and actuator selection, all of which fall within the protection scope of this invention.

Claims

1. A rapid cooling method for a calcined sand cooling kiln based on deep reinforcement learning, characterized in that, include: Obtain real-time status information and cooling status of the target object; Based on the real-time status information and the target cooling status, a first continuous waveform signal and a second continuous waveform signal are generated. The first continuous waveform signal is used to control the injection intensity of the first heat exchange medium, and the second continuous waveform signal is used to control the purging intensity of the second heat exchange medium. The first heat exchange medium is used to undergo a phase change to absorb heat when it comes into contact with the target object. The first continuous waveform signal and the second continuous waveform signal are respectively output to the corresponding actuators to control the first heat exchange medium and the second heat exchange medium to overlap with non-discontinuous continuous waveforms with a fixed timing relationship on the time axis, so that the temperature of the target object approaches the target cooling state.

2. The cooling method according to claim 1, characterized in that, The target object is located inside a rotary cooling device, moving axially and tumbling radially; The pulse frequency of the first continuous waveform signal is associated with the radial tumbling frequency of the target object within the rotary cooling device, so that the jet pulse of the first heat exchange medium is aligned with the moment when the target object is exposed to the surface.

3. The cooling method according to claim 2, characterized in that, The real-time status information includes the surface temperature distribution of the target object; after acquiring the real-time status information of the target object, it further includes: Based on the tilt angle, rotation speed, and feed flow rate of the rotary cooling equipment, calculate the axial moving speed and residence time distribution of the target object; The surface temperature distribution, the axial moving speed, and the current values ​​of the first and second continuous waveform signals are input into the physical information neural network model. The physical information neural network model is trained by using the partial differential equation describing the unsteady heat conduction process inside the target object as a constraint term of the loss function, and outputs the transient temperature distribution of the target object from the surface to the core. The maximum internal temperature gradient is calculated based on the transient temperature distribution, and the maximum internal temperature gradient is used as a thermal stress constraint condition for the generation of subsequent control signals.

4. The cooling method according to claim 3, characterized in that, The generation of the first continuous waveform signal and the second continuous waveform signal specifically includes: The high-level reinforcement learning model dynamically divides the spatial proportion of virtual temperature zones based on the feed temperature, target discharge temperature, feed flow rate, current rotation speed and filling rate of the rotary cooling equipment, and in combination with the axial moving speed and residence time distribution. The virtual temperature zones include flash zone, sweep zone and slow cooling heat transfer zone. The space ratio of the flash zone is positively correlated with the feed temperature and the target cooling rate, and is used to allocate the resource ratio of phase change cooling; the space ratio of the slow cooling heat transfer zone is positively correlated with the residence time and the thermal stress release demand represented by the maximum internal temperature gradient, and is used to allocate the resource ratio of internal heat conduction. Based on the spatial proportion of the virtual temperature zone, the multiple independently controlled first heat exchange medium injection units and second heat exchange medium purging units arranged along the axial direction are dynamically divided into groups corresponding to the flash zone, sweeping zone and slow cooling heat transfer zone, respectively. Within the dynamically divided group, the underlying reinforcement learning model generates the first continuous waveform signal and the second continuous waveform signal based on the maximum internal temperature gradient, exhaust humidity, and radial tumbling frequency of the target object at the current moment. The radial tumble frequency is calculated in real time based on the rotational speed, filling rate of the rotary cooling equipment and the dynamic rest angle of the target object. The underlying reinforcement learning model synchronizes the pulse frequency of the first continuous waveform signal with the radial tumble frequency through a phase-locked loop algorithm, so that the peak of the first continuous waveform signal is aligned with the moment when the target object tumbles from the passive core area to the active surface layer.

5. The cooling method according to claim 4, characterized in that, The underlying reinforcement learning model also dynamically adjusts the waveform shape of the first continuous waveform signal and the second continuous waveform signal according to the maximum internal temperature gradient and the exhaust humidity, generating an asymmetric sine wave or quasi-sine wave, so that the rising phase and falling phase of the two overlap in time to form a waveform overlap area. Specifically, when the maximum internal temperature gradient is greater than a preset safety threshold, the duration of the waveform overlap region is increased; when the exhaust humidity is greater than a preset dew point threshold, the duration of the waveform overlap region is decreased.

6. The cooling method according to claim 5, characterized in that, By controlling the waveform shape of the first and second continuous waveform signals, the following stages are sequentially formed within one cooling cycle: The rising edge of the first continuous waveform signal corresponds to the flash evaporation endothermic period, during which the injection rate of the first heat exchange medium increases. The waveform overlap region corresponds to the transition stage from the flash evaporation endothermic period to the mild sweeping period. During this period, the injection amount of the first heat exchange medium decreases and the purging amount of the second heat exchange medium increases. The peak of the second continuous waveform signal corresponds to the mild sweeping period, during which the second heat exchange medium is subjected to strong purging, and the second heat exchange medium is the waste heat warm air recovered from the tail of the rotary cooling device. The falling edge of the second continuous waveform signal corresponds to the internal heat conduction period, during which the purging intensity of the second heat exchange medium is reduced.

7. The cooling method according to claim 3, characterized in that, The underlying reinforcement learning model is trained using a multi-objective reward function, which includes: The positive reward is positively correlated with the instantaneous temperature drop of the target object; The first penalty term is positively correlated with the square of the maximum internal temperature gradient exceeding the dynamic temperature gradient threshold. The second penalty is positively correlated with the degree to which the exhaust humidity approaches the saturation dew point; The dynamic temperature gradient threshold is calculated in real time based on the real-time average temperature of the target object and the nonlinear coupling relationship between the material's yield strength, elastic modulus, and thermal expansion coefficient as the temperature changes, so that the dynamic temperature gradient threshold has a higher value in the high-temperature range and a lower value in the low-temperature range.

8. The cooling method according to claim 1, characterized in that, Before outputting the first continuous waveform signal and the second continuous waveform signal to the corresponding actuator, a safety constraint step is also included: Calculate the maximum theoretical water carrying capacity under the current operating conditions based on real-time wind speed, exhaust temperature, and dew point temperature. If the instantaneous jet volume corresponding to the amplitude of the first continuous waveform signal exceeds the maximum theoretical water-carrying capacity, the amplitude of the first continuous waveform signal is corrected to the upper limit value corresponding to the maximum theoretical water-carrying capacity, and a negative reward is fed back to the control model that generated the waveform signal.

9. The cooling method according to claim 1, characterized in that, The generation models for the first and second continuous waveform signals are trained through the following steps: Historical operational data was collected, and the conservative Q-learning algorithm was used for offline pre-training to obtain the initial model. The initial model is optimized through reinforcement learning in a virtual twin environment constructed using physical information neural networks and computational fluid dynamics to update the model parameters; The updated model was deployed to the edge computing gateway and fine-tuned online using the flexible actor-critic algorithm, under the constraints of physical security rules.

10. A cooling system applied to a rotary cooling device, characterized in that, include: Temperature sensing device, used to acquire the real-time surface temperature distribution and spatial temperature field of the target object; Multiple axially independently controlled first heat exchange medium injection units, each injection unit including at least one nozzle and a corresponding servo proportional valve, are used to independently adjust the injection intensity of the area according to the first continuous waveform signal; Multiple axially independently controlled second heat exchange medium purging units, each purging unit including at least one damper or fan, for independently adjusting the purging intensity of the area according to the second continuous waveform signal; A rotation speed and tilt angle sensing device is used to acquire the real-time rotation speed and tilt angle parameters of the rotary cooling equipment; The controller is communicatively connected to the temperature sensing device, the speed and tilt sensing device, the plurality of first heat exchange medium injection units and the plurality of second heat exchange medium purging units, and is configured to perform the cooling method as described in any one of claims 1 to 9.

Images