AI-based vacuum heat pipe exhaust endpoint intelligent judgment system and method

Abstract

The application relates to the technical field of model deep learning, and particularly discloses an AI-based intelligent judgment system and method for a vacuum heat pipe exhaust end point, which comprises a heating exhaust adjusting module, a vacuum pumping speed adjusting module, an end point preliminary judgment module and a double physical verification module. The heating exhaust adjusting module dynamically adjusts the heating power of a heating device, so that the heating curve is matched with the exhaust process in real time. The vacuum pumping speed adjusting module adjusts the vacuum pump pumping speed in real time and synchronously generates an adaptive index threshold value. The end point preliminary judgment module triggers a preliminary end point signal and performs a sealing review. The double physical verification module starts mass conservation verification and spectral characteristic verification according to the sealing review result, and only when the two verifications are passed, the final end point judgment result is output and a sealing operation is performed. The AI judgment result is reviewed from two independent dimensions of material conservation and component purity, so that the hidden quality problems caused by residual gas or impurities are fundamentally eliminated.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to an AI-based intelligent judgment system and method for the exhaust endpoint of a vacuum heat pipe. Background Technology

[0002] Vacuum heat pipes are highly efficient heat-conducting components. They utilize the phase change of matter in a vacuum environment, such as a liquid turning into a gas, to transfer heat. The heat transfer speed is extremely fast, and the efficiency is far higher than that of metals such as copper and silver. The intelligent determination of the exhaust endpoint of a vacuum heat pipe involves sensors monitoring key parameters such as temperature and vacuum level in real time. A programmable logic controller (PLC) automatically determines whether the exhaust is complete and automatically performs the sealing operation.

[0003] In existing vacuum heat pipe manufacturing processes, the evacuation process is typically performed before heat pipe sealing. This involves connecting the heat pipe body to a vacuum pumping device to continuously evacuate the internal space. During evacuation, the pipe body or related tooling is often heated to gradually release and remove air, adsorbed gases from the material surface and interior, and dissolved gases from the working fluid. The vacuum state inside the pipe is monitored in real time during evacuation, and the evacuation status is controlled according to the vacuum level setpoints corresponding to different process types. The evacuation process is considered complete when the vacuum level reaches the setpoint or when the vacuum level variation remains within a predetermined range for a continuous period. Some processes also include a pressure holding monitoring phase after evacuation stops, where changes in vacuum level during the pressure holding period confirm the evacuation endpoint. After determining the evacuation endpoint, the heat pipe is sealed to form a closed vacuum heat pipe structure.

[0004] Current methods for determining the exhaust endpoint of vacuum heat pipes typically rely on vacuum level monitoring. However, vacuum level readings only reflect the total amount of gas molecules within the cavity and cannot distinguish whether they originate from free gas in space or adsorbed gas on the pipe wall surface. Therefore, even if the vacuum gauge reading reaches a preset threshold, the inner wall of the heat pipe, especially in areas that have undergone machining or have an oxide layer, may still adsorb a layer of undesorbed gas molecules. These residual gases are gradually released during subsequent heating of the heat pipe, transforming into non-condensable gases, leading to a decline in product performance during use. To promote the desorption of adsorbed gases, existing processes often employ heating methods, but these are mostly based on fixed heating parameters, such as heating time and temperature. This static control logic is essentially an open-loop control, preventing the equipment from sensing the real-time desorption status of adsorbed gases inside the heat pipe.

[0005] Because the desorption process cannot be accurately detected, operators or equipment often tend to extend the heating time to ensure complete venting. This, in turn, exacerbates the temperature imbalance inside the heat pipe, leading to a further widening of the temperature difference between the heating and condensing ends. Areas far from the heat source experience insufficient desorption due to insufficient temperature, while the heating end overheats due to prolonged heating. This practice of extending heating to compensate for insufficient desorption also directly results in excessive loss of working fluid vapor. While the working fluid carries non-condensable gases out, it is also consumed in large quantities. However, current venting endpoint judgment and working fluid control are independent, making it impossible to accurately control the remaining amount of working fluid while judging the endpoint. This causes the working fluid charge of the finished heat pipe to deviate from the design value, thereby affecting its heat transfer power and operational stability. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent judgment system and method for vacuum heat pipe exhaust endpoint, which can effectively solve the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides an AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe, comprising: a heating exhaust adjustment module, used to acquire multi-source sensor signals during the exhaust process of the vacuum heat pipe and input them to a stage recognition AI model, outputting the current exhaust stage of the vacuum heat pipe, and an intelligent judgment unit adjusting the heating power of the heating device; a vacuum pumping speed adjustment module, used to extract characteristic parameters corresponding to the vacuum heat pipe from the multi-source sensor signals based on the exhaust stage to analyze the exhaust sufficiency index, used to adjust the pumping speed of the vacuum pump and simultaneously generate an adaptive index threshold; a preliminary endpoint judgment module, used to determine the preliminary endpoint condition satisfaction based on the adaptive index threshold, and when the preliminary endpoint signal is triggered, the intelligent judgment unit performs a sealing verification; and a dual physical verification module, used to perform dual physical verification based on the sealing verification result, and if the dual physical verification passes, the final endpoint judgment result is obtained, and the intelligent judgment unit performs a sealing operation on the vacuum heat pipe to complete the intelligent judgment of the vacuum heat pipe exhaust endpoint.

[0008] The second aspect of this invention provides an AI-based intelligent method for determining the exhaust endpoint of a vacuum heat pipe, comprising: acquiring multi-source sensor signals during the exhaust process of the vacuum heat pipe and inputting them into a stage recognition AI model, outputting the current exhaust stage of the vacuum heat pipe, and adjusting the heating power of the heating device by an intelligent determination unit; extracting characteristic parameters corresponding to the vacuum heat pipe from the multi-source sensor signals based on the exhaust stage to analyze the exhaust sufficiency index, which is used to adjust the pumping speed of the vacuum pump and generate an adaptive index threshold; performing a preliminary endpoint condition satisfaction determination based on the adaptive index threshold, and performing a sealing verification when the preliminary endpoint signal is triggered; performing a dual physical verification based on the sealing verification result, and obtaining the final endpoint determination result if the dual physical verification passes; and performing a sealing operation on the vacuum heat pipe by the intelligent determination unit to complete the intelligent determination of the vacuum heat pipe exhaust endpoint.

[0009] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0010] (1) This invention provides an AI-based intelligent judgment system and method for the exhaust endpoint of a vacuum heat pipe. The heating exhaust adjustment module acquires multi-source sensor signals and inputs them into a stage recognition AI model, outputting the current exhaust stage. The intelligent judgment unit dynamically adjusts the heating power of the heating device accordingly, so that the heating curve matches the exhaust process in real time, avoiding temperature field imbalance or insufficient desorption caused by fixed power. The vacuum pumping speed adjustment module extracts feature parameters based on the exhaust stage and inputs them into a sufficiency evaluation model, outputting an exhaust sufficiency index for real-time adjustment of the vacuum pumping speed and synchronous generation of an adaptive index threshold. The dynamic optimization of the pumping speed enhances the sensitivity of water vapor concentration detection, and at the same time provides a basis for subsequent endpoint judgment. A personalized baseline is established; the preliminary endpoint determination module monitors the preliminary endpoint conditions in real time based on adaptive index thresholds. When the conditions are met, a preliminary endpoint signal is triggered and the intelligent judgment unit performs a sealing verification. The exhaust endpoint is identified in advance by combining statistical regularity and trend prediction, avoiding false triggering caused by single-point fluctuations or sensor noise. The dual physical verification module initiates mass conservation verification and spectral feature verification based on the sealing verification results. Only when both verifications pass, the final endpoint determination result is output and the sealing operation is performed. The AI ​​judgment result is verified from two independent dimensions: material conservation and component purity, fundamentally eliminating hidden quality problems caused by residual gas or impurities.

[0011] (2) This invention introduces a deep learning model into the exhaust endpoint determination, constructing an end-to-end intelligent sensing system from multi-source sensor signals to control decisions. The AI ​​model can automatically learn the exhaust patterns of different heat pipe specifications and different incoming material conditions from a large amount of historical data, forming an adaptive control strategy for the individual differences of each heat pipe. This fundamentally solves the problem of false compliance caused by the inability to sense the desorption process of adsorbed gas and the inability to identify gas components in traditional methods. It breaks through the limitations of traditional methods that rely on fixed thresholds or human experience, and realizes adaptive control for the individual differences of each heat pipe, significantly improving the consistency of exhaust quality and process stability.

[0012] (3) This invention constructs an independent verification defense line in addition to AI intelligent judgment, fundamentally solving the risk of misjudgment that may exist in a single AI model. Mass conservation verification starts from the macroscopic total quantity level, ensuring the material conservation of the exhaust process by comparing the initial total gas mass with the cumulative exhaust gas mass. Spectral feature verification starts from the microscopic component level, using near-infrared spectroscopy to detect impurities in the exhaust gas in real time, which can identify purity traps where the vacuum degree meets the standard but impurities exceed the standard. The two complement and corroborate each other, effectively solving the risk of misjudgment that may exist in a single AI model, and providing dual protection for the long-term reliable operation of the heat pipe.

[0013] (4) This invention systematically overcomes the fundamental defects of existing technologies in terms of detection dimensions, control logic, and quality assurance modes through a complete technical chain of multi-source sensing, stage identification, adaptive thresholding, and physical verification. Compared with traditional methods that rely on vacuum monitoring, this invention no longer uses vacuum as the sole criterion, but integrates multi-dimensional information such as pressure, temperature, flow rate, water vapor concentration, and near-infrared spectroscopy, fundamentally solving the problem of false compliance where the vacuum level meets the standard but the gas composition is impure. Compared with static control methods that rely on fixed parameters, this invention dynamically divides the exhaust process through a stage identification model and adaptively adjusts the heating power and vacuum pump speed according to the real-time identified stages, so that the control strategy matches the physical process in real time, avoiding insufficient desorption or excessive loss of working fluid due to differences in the state of incoming materials. Compared with traditional processes that rely on human experience, this invention transforms tacit knowledge into a quantifiable and replicable AI model, and achieves continuous optimization of process parameters through cloud data closed loop, so that exhaust quality shifts from experience-driven to data-driven, completely solving the industry pain points of reliance on skills and lack of replicability. Attached Figure Description

[0014] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0015] Figure 1This is a schematic diagram of the system module connections of the present invention.

[0016] Figure 2 This is a schematic diagram of the method steps of the present invention.

[0017] Figure 3 This is a diagram of the AI ​​model architecture for stage identification.

[0018] Figure 4 A diagram of the AI ​​evaluation model architecture for full sufficiency. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] Example 1:

[0021] In a specific embodiment of the vacuum heat pipe production workshop, a heat pipe that has just undergone vacuum pretreatment is clamped onto a specialized device. Its exhaust port is connected to the sensor assembly and vacuum pump inside the device via a sealed pipe. A controllable heating device is arranged on the outside of the evaporation section of the heat pipe. After the device is started, the servo tightening shaft first loosens the nut at the end of the heat pipe to the angle position corresponding to the pre-loosening torque value. The remaining high-pressure gas and water vapor inside the heat pipe begin to be discharged outward through the exhaust pipe. At this time, the intelligent judgment unit controls the heating device to automatically adjust the power according to the current exhaust stage: in the initial stage of exhaust, the heating power is maintained at a low level to avoid the heat pipe temperature field imbalance caused by excessively rapid heating; when the sensor detects that the free gas has been basically exhausted and the stage of adsorbed gas desorption is dominant, the heating power is automatically increased to accelerate the release of adsorbed gas on the pipe wall. At the same time, the temperature gradient of each measuring point along the axis of the heat pipe is monitored in real time. If the temperature difference is too large, the power is adjusted back to prevent local overheating.

[0022] As the exhaust gas enters its later, stable decay phase, the heating power automatically decreases to a heat preservation level to avoid overheating and consuming the working fluid. Simultaneously, the vacuum pump speed is dynamically adjusted based on real-time calculated exhaust adequacy indicators: a high pumping speed is maintained initially to quickly remove free gas; the speed is appropriately reduced in the middle phase to extend the gas residence time in the detection pipeline and improve the sensitivity of water vapor concentration measurement; and the pumping speed is further reduced in the later phase to minimize airflow disturbance and create stable conditions for endpoint determination. When the judgment system analyzes the water vapor concentration decay trend and combines multi-sensor signals to determine that the exhaust gas is nearing its endpoint, a preliminary endpoint signal is triggered. The vacuum pump inlet valve is then closed for a seal check. After confirming no leakage, dual physical verification is initiated: the cumulative exhaust volume is calculated using mass conservation to ensure it matches the initial gas content, and near-infrared spectroscopy analysis is used to determine if the impurities in the exhaust gas meet standards. Once both verifications pass, the judgment system ultimately determines that the exhaust endpoint has been reached. The servo tightening shaft locks the nut at the angle corresponding to the pre-tightening torque value, completing the exhaust sealing operation for the heat pipe. All process data is automatically uploaded to the cloud for subsequent process optimization.

[0023] Reference Figure 1 As shown, the first aspect of this invention provides an AI-based intelligent endpoint determination system for vacuum heat pipe exhaust, comprising: a heating exhaust adjustment module, a vacuum pumping speed adjustment module, an endpoint preliminary determination module, a dual physical verification module, and a cloud database. The cloud database stores preset values ​​for various parameters and process parameters of the vacuum heat pipe exhaust operation.

[0024] The heating and exhaust adjustment module is connected to the vacuum pumping speed adjustment module, which is connected to the endpoint preliminary determination module. The endpoint preliminary determination module is connected to the dual physical verification module. The heating and exhaust adjustment module, the vacuum pumping speed adjustment module, the endpoint preliminary determination module, and the dual physical verification module are all connected to the cloud database.

[0025] The heating and exhaust adjustment module is used to acquire multi-source sensor signals during the exhaust process of the vacuum heat pipe and input them to the stage recognition AI model. It outputs the current exhaust stage of the vacuum heat pipe and the intelligent judgment unit adjusts the heating power of the heating device.

[0026] In the judgment system of this invention, the intelligent judgment unit is specifically manifested as an edge controller deployed on the production site, serving as the core control hub for the entire exhaust endpoint intelligent judgment and sealing operation. This edge controller receives and processes signals from multiple source sensors in real time via a high-speed data interface. It integrates the model required for endpoint judgment, enabling online AI inference calculations and real-time output of the current exhaust stage label and exhaust adequacy index.

[0027] Based on these real-time analysis results, the edge controller autonomously generates and issues control commands: on the one hand, it dynamically adjusts the heating power of the heating device according to the current exhaust stage and triggers a callback mechanism when a temperature gradient exceeding the limit is detected; on the other hand, it adjusts the pumping speed of the vacuum pump in real time according to the exhaust adequacy index to match the gas sampling sensitivity requirements of different exhaust stages. When the AI ​​model determines that the exhaust is nearing its end, the edge controller automatically triggers a sealing verification and dual physical verification process, and decides whether to send a sealing command to the servo tightening shaft based on the verification results. At the same time, the edge controller uploads the entire process data and quality judgment results of each operation to the cloud database in real time for subsequent model iteration and process parameter optimization, and receives the optimized parameter set sent from the cloud to update the local control strategy.

[0028] The multi-source sensor signals include at least pressure signals, temperature signals, exhaust flow signals, and water vapor concentration signals.

[0029] The aforementioned stage recognition AI model is a deep learning classification model based on a Temporal Convolutional Network (TCN). For example... Figure 3 As shown, Figure 3 The diagram illustrates the AI ​​model architecture for stage identification. During model training, a large amount of historical data covering different operating conditions of the vacuum heat pipe exhaust process is first collected as the true labels for the training samples. The model uses a sliding window approach to extract multivariate time-series segments as input. Each input segment contains a continuous sequence of various sensor signals, such as pressure, temperature, exhaust flow rate, and water vapor concentration, within a fixed time window. The temporal convolutional network first performs preliminary feature extraction on the input time series through causal convolutional layers, then stacks multiple residual blocks. Each residual block contains two dilated convolutional layers, sequentially connected to a weight normalization layer, a linear rectified activation layer, and a random deactivation layer. Gradient propagation is achieved smoothly through residual connections. The dilation coefficient of each residual block increases exponentially, thereby exponentially expanding the receptive field and capturing signal change patterns at different time scales.

[0030] During training, the model propagates the input time-series segments layer by layer, ultimately outputting the probability distribution of the current exhaust stage through a softmax classification layer. Cross-entropy is used as the loss function to measure the difference between the model's output probability and the true label, and the weight parameters within the network are continuously optimized through backpropagation. After iterative training with thousands of samples, the model learns to automatically extract feature patterns related to the exhaust stage from multivariate time-series signals. For example, a sharp drop in pressure corresponds to rapid initial exhaust, an increase in water vapor concentration corresponds to desorption dominance in the middle stage, and a slowing decline in concentration corresponds to stable decay in the later stage. In practical online applications, real-time sensor signals are input into the trained model using the same sliding window method. After forward propagation, the model directly outputs the probability distribution of the exhaust stage at the current moment, and the category corresponding to the highest probability is taken as the stage identification result.

[0031] The signals from multiple sensors are preprocessed, including using Kalman filtering to denoise the pressure and temperature signals, and using adaptive wavelet threshold denoising to denoise the exhaust flow and water vapor concentration signals.

[0032] The exhaust phase includes an initial rapid exhaust phase, a mid-term desorption-dominated phase, and a late-term stable decay phase.

[0033] The heating power of the heating device is adjusted, specifically by the intelligent judgment unit adjusting the heating power according to the exhaust stage.

[0034] When the exhaust phase is identified as the initial rapid exhaust phase, the intelligent judgment unit controls the heating device to adjust the current heating power to the first preset power at the first heating adjustment rate.

[0035] When the exhaust phase is identified as the mid-term desorption-dominated phase, the intelligent judgment unit controls the heating device to increase the current heating power to a second preset power at a second heating adjustment rate. During this increase, the axial temperature gradient of the vacuum heat pipe is continuously monitored. If the axial temperature gradient exceeds a preset gradient threshold, the increase is paused, and the heating device is controlled to reduce the heating power back to a target reduction power at a third heating adjustment rate. The target reduction power is between the first and second preset power. In the vacuum heat pipe exhaust process, the axial temperature gradient refers to the rate of temperature change per unit length along the heat pipe axis, typically from the evaporation section to the condensation section. Specifically, it represents the degree of temperature difference between measuring points at different locations on the outer wall of the heat pipe. In actual monitoring, temperature sensors are placed in the evaporation section, adiabatic section, and condensation section of the heat pipe's outer wall to collect temperature values ​​at each point in real time, and their standard deviation or maximum temperature difference is calculated as a quantitative representation of the axial temperature gradient.

[0036] When the exhaust phase is identified as the later stable decay phase, the intelligent judgment unit controls the heating device to reduce the current heating power to the third preset power at the fourth heating adjustment rate.

[0037] In a specific embodiment, the first, second, third, and fourth heating adjustment rates are set according to the following relationship: second heating adjustment rate ≥ first heating adjustment rate > third heating adjustment rate ≥ fourth heating adjustment rate. This ordering is based on the different physical requirements of heating response speed at each exhaust stage: the second heating adjustment rate is set to be the fastest, used in the mid-term desorption-dominant stage, aiming to quickly increase power to establish the thermal field required for enhanced desorption and accelerate the release of adsorbed gas; the first heating adjustment rate is the next fastest, used in the initial rapid exhaust stage, to establish the initial heating state at a moderate speed, ensuring both response speed and avoiding thermal shock; the third heating adjustment rate is set to be relatively slow, used for the callback operation when the temperature gradient exceeds the limit in the mid-term desorption-dominant stage, correcting the temperature field unevenness by gradually reducing power, and preventing thermal stress or desorption interruption caused by a sudden drop in power; the fourth heating adjustment rate is set to be the slowest, used in the power reduction process of the later stable decay stage, to smoothly transition to the heat preservation state at an extremely slow speed, minimizing temperature disturbance to the heat pipe about to be sealed, protecting the working fluid, and ensuring the stability of the endpoint judgment. By setting the graded rate as described above, an optimal balance between response speed and control accuracy is achieved in the adjustment of heating power.

[0038] Similarly, the three preset power values ​​are set according to the relationship: second preset power > first preset power > third preset power. The first, second, and third preset power values ​​correspond to the heating targets of different exhaust stages, and their values ​​are set based on the inherent physical laws governing the thermal field requirements of each stage: the first preset power is used in the initial rapid exhaust stage, set at a relatively low power value to establish a basic heating state to promote the exhaust of free gas, while avoiding excessive axial temperature gradients in the heat pipe due to excessively high initial power; the second preset power is used in the mid-term desorption-dominant stage, set at the highest power value throughout the entire exhaust process, to rapidly increase the heat pipe temperature to enhance the desorption process of adsorbed gas on the pipe wall, ensuring sufficient release of the adsorbed gas layer; the third preset power is used in the later stable decay stage, set at the lowest power value, to maintain the heat pipe in a low-temperature insulation state, preventing excessive loss of working fluid due to overheating and providing a stable thermal environment for endpoint determination. This achieves optimized control of power progression and exit from the initial gentle start-up, the mid-term enhanced desorption, to the later stable conclusion.

[0039] Adjusting the heating power of the heating device also includes:

[0040] The process data and quality assessment results of each exhaust operation are uploaded to the cloud database.

[0041] The process data includes: heat pipe model, actual exhaust time, actual power values ​​at each stage, peak temperature gradient, mass conservation verification deviation value recorded during the dual verification process, and peak absorbance value for spectral characteristic verification. The mass conservation verification deviation value refers to the relative deviation calculated based on the initial total gas mass of the heat pipe and the cumulative exhaust gas mass.

[0042] The quality assessment results include: the final endpoint assessment result, the pass / fail status label after manual review, and the type of non-compliance reason when the non-compliance is determined to be non-compliant after manual review.

[0043] When the cumulative data volume corresponding to the same heat pipe model reaches the preset batch threshold, the cloud database performs statistical analysis on the process data and quality judgment results. The statistical analysis includes:

[0044] Extract samples that meet the quality assessment criteria, and calculate the mean and standard deviation of their actual power values ​​at each stage, as well as the mean and standard deviation of their actual adjustment rates at each stage. Compare the mean power of the qualified samples with the current preset power. If the deviation exceeds the first preset deviation threshold, the mean power of the qualified samples is used as the optimized first, second, or third preset power.

[0045] The average adjustment rate of qualified samples is compared with the current preset adjustment rate. If the deviation between the two exceeds the second preset deviation threshold, the average adjustment rate of qualified samples is used as the optimized first heating adjustment rate, second heating adjustment rate, third heating adjustment rate or fourth heating adjustment rate.

[0046] The optimized preset power and heating adjustment rate are sent to the intelligent judgment unit as an optimized parameter set to replace the original preset parameters in order to control the exhaust process of subsequent heat pipes of the same model.

[0047] The vacuum pumping speed adjustment module is used to extract the characteristic parameters corresponding to the vacuum heat pipe from the multi-source sensor signals during the exhaust stage to analyze the exhaust adequacy index, and to adjust the pumping speed of the vacuum pump while generating an adaptive index threshold.

[0048] The characteristic parameters corresponding to the vacuum heat pipe include the maximum slope of pressure drop within a preset first time window when it is in the initial rapid exhaust phase, and the peak exhaust flow rate within a preset first time window.

[0049] When the desorption is dominant in the middle stage, the peak water vapor concentration is extracted within the preset second time window, the standard deviation of the temperature gradient at the heat pipe axial temperature measurement point is calculated within the preset second time window, and the second derivative of the pressure curve is extracted as the second derivative of pressure decay within the preset second time window.

[0050] When in the later stable decay stage, the slope of the water vapor concentration curve is extracted within the preset third time window as the water vapor concentration decay slope. The standard deviation of the water vapor concentration signal is calculated within the preset third time window as the concentration fluctuation standard deviation. The cumulative exhaust volume is obtained by integrating the exhaust flow signal from the start of exhaust to the current time.

[0051] The exhaust adequacy index is specifically calculated by inputting characteristic parameters into the exhaust adequacy AI evaluation model and outputting the exhaust adequacy index.

[0052] The aforementioned sufficiency AI evaluation model is a deep learning regression model based on a gated recurrent unit (GRU). For example... Figure 4 As shown, Figure 4 To fully evaluate the AI ​​evaluation model architecture, during the model training phase, a large amount of historical qualified exhaust operation process data needs to be collected as supervision labels. Typically, the target value corresponding to the endpoint is set to 1, the exhaust start time is set to 0, and intermediate times are linearly interpolated over time. The model input is an extracted sequence of multidimensional feature vectors, such as the pressure drop slope, peak water vapor concentration, and temperature gradient standard deviation. These feature vectors are arranged in chronological order to form the input sequence.

[0053] These feature vectors are arranged in chronological order to form the input sequence, and the feature vector at each time step is denoted as x. t The gated loop unit resets the gate r. t Control the exhaust characteristic memory of the previous moment h t−1 The impact on the current candidate state is determined by updating gate z. t By controlling the fusion ratio of the previous time-series memory and the current candidate state, the model effectively captures long-term dependencies in the feature vector sequence and learns the mapping from feature evolution patterns to the exhaust process. During training, the model propagates the input sequence forward step by step, ultimately outputting a scalar value between 0 and 1, called the exhaust sufficiency index. Mean squared error is used as the loss function to measure the difference between the model's output value and the supervision label, and the weight parameters within the network are optimized through backpropagation. After iterative training with thousands of lattice samples, the model learns to identify the patterns of the exhaust process from the temporal changes of multi-dimensional feature vectors. For example, when the water vapor concentration continuously decreases and the fluctuation slows down, and the pressure decay rate approaches zero, the sufficiency index output by the model approaches 1, indicating that the exhaust process is nearing its end. In practical online applications, the real-time extracted feature vector sequence is input into the trained GRU model. The model directly calculates the exhaust sufficiency index at the current time through forward propagation, providing a quantitative basis for subsequent adaptive threshold generation and endpoint determination.

[0054] The adaptive metric threshold is generated as follows:

[0055] The exhaust adequacy index is input into the Bayesian online change point detection algorithm in real time. The algorithm recursively calculates the probability distribution of the data segment to which the exhaust adequacy index belongs at the current moment, and dynamically identifies whether the exhaust adequacy index has entered the stable region based on the probability distribution.

[0056] The Bayesian online change point detection algorithm is a probabilistic recursive online sequence analysis model used to dynamically monitor and identify stationary regions of real-time input exhaust adequacy indices. This algorithm treats the exhaust adequacy index sequence as consisting of several stationary data segments following the same probability distribution, with the transition points between segments being the change points. In the initialization phase, the prior probability of belonging to the first data segment at the current moment is set to 1, and it is assumed that the observations within each data segment follow a normal distribution, with the prior distributions of their mean and variance parameterized using conjugate priors. During real-time operation, for each new exhaust adequacy index value collected, the algorithm recursively calculates the posterior probability of belonging to each candidate data segment at the current moment: first, based on the posterior probabilities of each data segment at the previous moment, combined with the growth probability and the change point probability, the prior probability of each data segment at the current moment is predicted; then, the prior probabilities are updated according to the likelihood of the new observation in a given data segment, resulting in the posterior probability distribution of each data segment at the current moment. The data segment with the highest posterior probability is selected as the most likely data segment for the current moment. Stationarity is determined based on the statistical characteristics of continuous sampling points within this data segment. When the duration of this data segment exceeds a preset length threshold, and the maximum fluctuation amplitude of the exhaust adequacy index within the data segment is less than a preset fluctuation threshold, the index is considered to have entered a stable region. Once in a stable region, an adaptive threshold is generated using the mean of the index within the data segment as the baseline value and the standard deviation as the fluctuation amount, calculated by subtracting a preset multiple of the fluctuation amount from the baseline value. This threshold is dynamically adjusted as the stable region is updated until a new change point is detected or the exhaust process ends.

[0057] When the exhaust fullness index is detected to have entered a stable region and is currently in the later stable decay stage, the average value of the data after entering the stable region is used as the benchmark value, and the standard deviation of the data after entering the stable region is used as the fluctuation amount. The adaptive index threshold is set to the benchmark value minus the fluctuation amount of the preset multiple.

[0058] The criteria for determining a stable region are: within a predetermined number of consecutive sampling points, the fluctuation amplitude of the exhaust adequacy index is less than a predetermined fluctuation threshold. The fluctuation amplitude of the exhaust adequacy index refers to the difference between the maximum and minimum values ​​of the exhaust adequacy index over a continuous period, used to quantify the severity of the index's fluctuations during that time period.

[0059] Adjust the pumping speed of the vacuum pump by generating a pumping speed adjustment command based on the real-time exhaust adequacy index.

[0060] When the exhaust adequacy index is less than the first preset threshold, a first pumping speed adjustment command is generated. The intelligent judgment unit controls the vacuum pump to increase the current pumping speed to the rated pumping speed at the first pumping speed adjustment rate based on the first pumping speed adjustment command.

[0061] When the exhaust adequacy index is greater than or equal to a first preset threshold and less than a second preset threshold, a second pumping speed adjustment command is generated. Based on the second pumping speed adjustment command, the intelligent judgment unit controls the vacuum pump to reduce the current pumping speed to the first preset pumping speed at a second pumping speed adjustment rate. In this technical solution, the first preset threshold and the second preset threshold are related in that the first preset threshold is less than the second preset threshold.

[0062] When the exhaust fullness index is greater than or equal to the second preset threshold, a third pumping speed adjustment command is generated. The intelligent judgment unit controls the vacuum pump to reduce the current pumping speed to the second preset pumping speed based on the third pumping speed adjustment command.

[0063] When the increase in the exhaust adequacy index within a preset time period is less than a preset threshold, the increase is calculated by taking a preset time period (e.g., 30 seconds) backward from the current time as the end point. The difference between the exhaust adequacy index value at the end of the window and the value at the beginning of the window is calculated, generating a fourth pumping speed adjustment command. Based on this command, the intelligent judgment unit controls the heating device to add a preset power increment to the current heating power at a first heating adjustment rate, and controls the vacuum pump to reduce the current pumping speed to a third preset pumping speed at the fourth pumping speed adjustment rate. After the temporary increase lasts for the preset time period, the heating device is controlled to restore the original heating power at a second heating adjustment rate. The specific value of the original heating power corresponds to the specific value of the current heating power.

[0064] In a specific embodiment, the first, second, third, and fourth pumping speed adjustment rates are set according to the following relationship: first pumping speed adjustment rate ≥ second pumping speed adjustment rate > third pumping speed adjustment rate ≥ fourth pumping speed adjustment rate. This ordering relationship is based on the different physical requirements of pumping speed response speed at each exhaust stage: the first pumping speed adjustment rate is set to the fastest, used to increase the pumping speed to the rated pumping speed in the initial stage of exhaust, aiming to quickly establish a high pumping speed environment to rapidly remove a large amount of free gas in the heat pipe; the second pumping speed adjustment rate is next, used to reduce the pumping speed to the first preset pumping speed in the middle desorption-dominant stage, completing the pumping speed switching at a moderate speed, ensuring response speed while avoiding disturbances to the desorption process caused by sudden changes in airflow; the third pumping speed adjustment rate is set to a slower rate, used to reduce the pumping speed to the second preset pumping speed in the later stable decay stage, reducing airflow fluctuations through gradual deceleration, creating stable conditions for endpoint judgment; the fourth pumping speed adjustment rate is set to the slowest, used for temporary adjustment when the exhaust fullness index rises slowly, reducing the pumping speed at an extremely slow speed in conjunction with the heating power adjustment, minimizing the dual disturbances to the internal thermal field and airflow of the heat pipe. By setting the graded rate as described above, an optimal balance between rapid response and stable control of the vacuum pump pump speed is achieved.

[0065] Similarly, the four preset pumping speeds are set according to the relationship: rated pumping speed > first preset pumping speed > second preset pumping speed > third preset pumping speed. These values ​​are based on the different requirements for gas sampling sensitivity and process stability at each stage: the rated pumping speed is used in the initial stage of exhaust and is set to the highest pumping speed that the vacuum pump can operate at, aiming to quickly remove free gas from the heat pipe with maximum exhaust capacity; the first preset pumping speed is used in the middle stage dominated by desorption and is set to an intermediate value lower than the rated pumping speed. By appropriately reducing the pumping speed, the residence time of gas in the detection pipeline is extended, improving the measurement sensitivity of the water vapor concentration sensor, while maintaining sufficient... Sufficient venting capacity is provided to maintain the desorption process; a second preset pumping speed is used in the later stable decay stage, set to a lower value than the first preset pumping speed, to further slow down the airflow velocity, reduce the interference of turbulence fluctuations on the water vapor concentration signal, and provide a stable detection environment for endpoint judgment; a third preset pumping speed is used for temporary adjustment when the venting adequacy index rises slowly, set to the lowest value below the second preset pumping speed, and is used in conjunction with the slowest pumping speed while temporarily increasing the heating power, to maximize the contact time between the gas and the heat pipe, promote the desorption process, and at the same time avoid excessive extraction of the working fluid vapor generated by heating at high pumping speed. This achieves a progressive pumping speed control strategy from rapid venting in the initial stage, sensitive detection in the middle stage, stable judgment in the later stage, to coordinated intervention in abnormal conditions.

[0066] The preliminary endpoint determination module is used to determine whether the preliminary endpoint conditions are met based on the adaptive index threshold. When the preliminary endpoint signal is triggered, the intelligent judgment unit performs a sealing verification.

[0067] The preliminary endpoint conditions include: the current exhaust phase is in the late stable decay phase, the exhaust adequacy index is continuously lower than the adaptive index threshold within the first preset time period, and the exhaust adequacy index will not rebound to exceed the adaptive index threshold within the second preset time period as predicted by the time-series prediction AI model.

[0068] The aforementioned time-series prediction AI model is a deep learning prediction model based on Long Short-Term Memory (LSTM) networks. During the model training phase, a large amount of historical exhaust gas sufficiency index data from start to finish during exhaust gas operations is collected. Training samples are constructed using a sliding window approach: each sample uses a sequence of exhaust gas sufficiency indices across multiple consecutive time points as input features, and the maximum value of the exhaust gas sufficiency index or whether it exceeds a specific threshold within a future period is used as the prediction target label. The LSTM network controls the flow of information between time steps through input, forget, and output gates, effectively capturing long-term dependencies and non-linear variation patterns in the exhaust gas sufficiency index sequence, learning the mapping from historical trends to future short-term trends. During training, the model propagates the input sequence forward step by step, ultimately outputting the predicted values ​​for each time point within a second preset time period or the probability of the index exceeding the threshold within a future time period. Mean squared error or cross-entropy loss is used to measure the difference between the model's prediction and the true label, and the network weight parameters are optimized using a backpropagation algorithm.

[0069] After iterative training with a large number of samples, the model learned to identify typical characteristics of an impending rebound from the historical change patterns of the exhaust adequacy index. For example, when the rate of decline of the index gradually slows down and the fluctuation range narrows, the model predicts a lower probability of a future rebound; when there are signs of continuous intensified fluctuations or local rebounds during the decline of the index, the model predicts a higher probability of a future rebound. In actual online applications, the exhaust adequacy index sequence of a preset number of consecutive time intervals before the current time is input into the trained LSTM model. After forward propagation, the model outputs the probability that the index will exceed an adaptive threshold within a second preset time interval. When this probability is lower than a preset confidence threshold, it is determined that there will be no rebound in the future, which serves as one of the criteria for the initial endpoint condition.

[0070] It should be explained that when the preliminary endpoint conditions are not met, the system determines that the current exhaust process has not yet reached the stage where sealing can be triggered. At this time, the subsequent sealing verification and locking operations will not be performed. Instead, the current monitoring and control state will continue: the edge controller will continue to collect multi-source sensor signals in real time, continuously run the stage identification model and the sufficiency evaluation model, dynamically update the exhaust sufficiency index, and continuously adjust the heating power of the heating device and the pumping speed of the vacuum pump according to the latest identified exhaust stage and sufficiency index, until all preliminary endpoint conditions are met in the subsequent monitoring cycle, including being in the later stable decay stage, the exhaust sufficiency index being lower than the adaptive threshold for a continuous first preset time, and the time series prediction model determining that there will be no rebound within the next second preset time.

[0071] The sealing verification includes: closing the vacuum pump inlet valve and monitoring the rate of pressure rise in the vacuum heat pipe within a predetermined preset time. If the rate of rise is less than the preset rate threshold, the sealing verification is deemed to have passed; otherwise, an alarm is triggered.

[0072] When the initial endpoint conditions are met but the sealing verification fails, the system determines that there is a physical leakage risk in the heat pipe or exhaust pipe. At this time, the automatic sealing process is immediately interrupted and an alarm signal is issued: the edge controller controls the vacuum pump inlet valve to remain closed, suspends all subsequent actions, and displays the alarm information of sealing verification failure on the human-machine interface, prompting the operator to check the sealing status of the connection between the heat pipe and the exhaust pipe, whether the sealing ring is aged or damaged, and whether there are physical defects such as micro-cracks in the heat pipe itself. After the leakage fault is manually eliminated and the equipment is reset, the system restarts the exhaust process from the workpiece initialization stage, or returns to the state before the sealing verification according to the preset strategy to re-monitor.

[0073] The dual physical verification module is used to perform dual physical verification based on the sealing verification result. If the dual physical verification passes, the final endpoint determination result is obtained. The intelligent judgment unit then performs a sealing operation on the vacuum heat pipe to complete the intelligent determination of the vacuum heat pipe exhaust endpoint.

[0074] Dual physical verification includes mass conservation verification and spectral feature verification.

[0075] The mass conservation verification specifically includes: retrieving the volume V corresponding to the current heat pipe model from the database and the initial pressure P0 measured before exhaust begins; and calculating the initial total gas mass. Where R represents the gas constant and T represents the thermodynamic temperature of the gas inside the heat pipe, in Kelvin (K). The cumulative mass of discharged gas is obtained by integrating the flow signal. Where ρ(t) is the gas density, calculated from pressure and temperature, and t represents the cumulative time from the start of exhaust to the current moment, in seconds (s), continuously recorded by the edge controller's system clock and used as the independent variable in the integration calculation. F(t) represents the instantaneous value of the exhaust flow rate signal at time t, in L / s or m³ / s. This signal is acquired in real time by a mass flow meter installed on the exhaust pipe and used to integrate over time to calculate the cumulative exhaust volume. The deviation is then calculated based on this. If δ ≤ the preset deviation threshold, the verification passes; otherwise, it fails.

[0076] The spectral feature verification specifically includes: acquiring the near-infrared spectrum of the exhaust gas, extracting the absorbance peak at the preset characteristic wavelength of the impurity gas, and determining that the spectral feature verification is passed if all absorbance peaks are less than the preset absorbance threshold.

[0077] Once the sealing verification passes, dual physical verification is performed. If both verifications pass, the intelligent judgment unit obtains the final endpoint determination result and performs a sealing operation on the vacuum heat pipe. Specifically, it sends a locking command to the servo tightening shaft, tightens the nut to a preset torque, and completes the sealing. If either verification fails, the process parameters are adjusted according to the verification deviation type, and the steps from exhaust stage identification to dual physical verification are re-executed.

[0078] Adjusting process parameters based on the type of verification deviation specifically includes:

[0079] If the mass conservation verification fails, the first adjustment coefficient is calculated by comparing the difference between the current initial total gas mass and the cumulative discharged gas mass and the preset deviation threshold. The difference between the current initial total gas mass and the preset deviation threshold is then compared with the preset deviation threshold. The reference heating power of the heating device is increased, the target threshold of the exhaust adequacy index is increased, and the maximum allowable exhaust time is extended according to the first adjustment coefficient. The increase in the reference heating power, the increase in the target threshold, and the extension of the maximum allowable exhaust time are all positively correlated with the first adjustment coefficient.

[0080] The formula for calculating the first adjustment coefficient k1 is k1 = min((δ - δ)). th ) / δ th k 1max ), where δ is the measured deviation value for mass conservation verification, δth is the preset deviation threshold, and k 1max This is the preset maximum adjustment coefficient. The adjustment amount for the reference heating power is ΔP = P. base *k1*α, where P base K is the reference heating power, and α is the preset power adjustment step size factor. Similarly, the increase in the target threshold and the extension of the maximum allowable exhaust time are both positively correlated with k1.

[0081] If the spectral characteristic verification fails, a second adjustment coefficient is calculated based on the proportion of the current impurity absorbance peak exceeding the preset absorbance threshold. The reference heating power of the heating device is then increased, the rate threshold in the sealing detection standard is tightened, and the maximum allowable exhaust time is extended according to this second adjustment coefficient. The increase in reference heating power, the tightening of the rate threshold, and the extension of the maximum allowable exhaust time are all positively correlated with the second adjustment coefficient. The impurity absorbance peak is the maximum absorbance value detected at the preset impurity gas characteristic wavelength after real-time scanning of the exhaust gas using a near-infrared spectrometer during the spectral characteristic verification process.

[0082] The formula for calculating the second adjustment coefficient k2 is k2 = min((A max- A th ) / A th k 2max ), where A max To measure the peak absorbance of impurities, A th To preset the absorbance threshold, k 2max This is the preset maximum adjustment coefficient. The tightening range of the rate threshold is Δv = v th *k2*β, where v th β is the preset rate threshold, and β is the preset tightening step size factor.

[0083] The reference heating power, target threshold, maximum allowable exhaust time and rate threshold are all set with preset upper limits, and the adjusted values ​​do not exceed their respective preset upper limits.

[0084] The adjusted process parameters are used to control the subsequent exhaust process until the dual physical verification is passed or the preset maximum number of attempts is reached.

[0085] The entire process data of each task is uploaded to the cloud for incremental learning and iterative updates of the stage identification AI model, sufficiency AI evaluation model, and time series prediction AI model.

[0086] Reference Figure 2 As shown, the second aspect of the present invention provides an AI-based intelligent determination method for the exhaust endpoint of a vacuum heat pipe, comprising: acquiring multi-source sensor signals during the exhaust process of the vacuum heat pipe and inputting them into a stage recognition AI model, outputting the current exhaust stage of the vacuum heat pipe, and adjusting the heating power of the heating device by an intelligent determination unit.

[0087] Based on the exhaust phase, the characteristic parameters corresponding to the vacuum heat pipe are extracted from the multi-source sensor signals to analyze the exhaust adequacy index, which is used to adjust the pumping speed of the vacuum pump and generate an adaptive index threshold.

[0088] The initial endpoint condition is determined based on the adaptive index threshold. When the initial endpoint signal is triggered, the intelligent judgment unit performs a sealing verification.

[0089] Based on the sealing verification results, a dual physical verification is performed. If the dual physical verification passes, the final endpoint determination result is obtained. The intelligent judgment unit then performs a sealing operation on the vacuum heat pipe, completing the intelligent determination of the vacuum heat pipe exhaust endpoint.

[0090] Example 2:

[0091] With other conditions remaining unchanged in Example 1, the above exhaust adequacy index can also be obtained through eigenvector mapping, specifically:

[0092] Multiple feature parameters are combined in a fixed order to construct a multidimensional feature vector, where each dimension of the feature vector corresponds to a feature parameter with a clear physical meaning, such as the pressure drop slope, the peak water vapor concentration, and the standard deviation of the temperature gradient. Then, the feature vector is directly converted into an exhaust adequacy index through a mapping function constructed during the offline training phase.

[0093] The process of establishing the mapping function is as follows: collect a large number of feature vector samples of historical qualified exhaust operations, and use the manually determined end time as the benchmark. Map the feature vector corresponding to the end time to the target value 1, the start time to 0, and the intermediate times to linear interpolation or expert experience. Use regression algorithms, such as support vector regression, multiple linear regression, or shallow neural networks, to train the sample data and learn the mapping relationship from the feature vector space to the exhaust fullness index. In actual online application, the real-time constructed feature vector is input into the trained mapping function, and the exhaust fullness index at the current time is directly output through function calculation.

[0094] The essential difference between this method and the aforementioned deep learning model based on gated recurrent units is that the former establishes a static regression relationship between feature vectors and adequacy indices through a shallow mapping function, while the latter captures the temporal evolution pattern in the feature vector sequence through a recurrent neural network. Both can achieve effective conversion from multi-dimensional features to quantitative indicators of the exhaust process, and can be flexibly selected according to different computing resources, real-time requirements and model complexity based on actual application scenarios.

[0095] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An AI-based intelligent system for determining the exhaust endpoint of a vacuum heat pipe, characterized in that: include: The heating exhaust adjustment module is used to acquire multi-source sensor signals during the exhaust process of the vacuum heat pipe and input them to the stage recognition AI model. It outputs the current exhaust stage of the vacuum heat pipe and the intelligent judgment unit adjusts the heating power of the heating device. The vacuum pumping speed adjustment module is used to extract the characteristic parameters corresponding to the vacuum heat pipe from the multi-source sensor signals during the exhaust stage to analyze the exhaust adequacy index, and to adjust the pumping speed of the vacuum pump while generating an adaptive index threshold. The preliminary endpoint determination module is used to determine whether the preliminary endpoint conditions are met based on the adaptive index threshold. When the preliminary endpoint signal is triggered, the intelligent judgment unit performs a sealing verification. The dual physical verification module is used to perform dual physical verification based on the sealing verification result. If the dual physical verification passes, the final endpoint determination result is obtained. The intelligent judgment unit then performs a sealing operation on the vacuum heat pipe to complete the intelligent determination of the vacuum heat pipe exhaust endpoint.

2. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 1, characterized in that, Includes the following steps: The multi-source sensor signals include at least pressure signals, temperature signals, exhaust flow signals, and water vapor concentration signals; The multi-source sensor signals are preprocessed, including using Kalman filtering to denoise the pressure and temperature signals, and using adaptive wavelet threshold denoising to denoise the exhaust flow and water vapor concentration signals. The exhaust phase includes an initial rapid exhaust phase, a mid-term desorption-dominated phase, and a late-term stable decay phase.

3. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 2, characterized in that: The heating power of the heating device is adjusted by the intelligent judgment unit according to the exhaust stage. When the exhaust stage is identified as the initial rapid exhaust stage, the intelligent judgment unit controls the heating device to adjust the current heating power to the first preset power at the first heating adjustment rate. When the exhaust stage is identified as the mid-term desorption-dominant stage, the intelligent judgment unit controls the heating device to increase the current heating power to the second preset power at the second heating adjustment rate, and continuously monitors the axial temperature gradient of the vacuum heat pipe during the increase. If the axial temperature gradient exceeds the preset gradient threshold, the increase is paused, and the heating device is controlled to reduce the heating power to the target reduction power at the third heating adjustment rate. The target reduction power is between the first preset power and the second preset power. When the exhaust phase is identified as the later stable decay phase, the intelligent judgment unit controls the heating device to reduce the current heating power to the third preset power at the fourth heating adjustment rate.

4. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 1, characterized in that, Includes the following steps: The characteristic parameters corresponding to the vacuum heat pipe include the maximum slope of pressure drop and the peak value of exhaust flow rate extracted within a preset time window when it is in the initial rapid exhaust phase. When the desorption is dominant in the middle stage, the peak water vapor concentration, standard deviation of temperature gradient, and second derivative of pressure decay are extracted within a preset time window. When in the later stable decay stage, extract the water vapor concentration decay slope, concentration fluctuation standard deviation and cumulative exhaust volume within the preset time window.

5. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 1, characterized in that, Includes the following steps: The exhaust adequacy index is specifically obtained by inputting the characteristic parameters into the exhaust adequacy AI evaluation model and outputting the exhaust adequacy index. The adaptive index threshold is generated as follows: The exhaust adequacy index is input into the Bayesian online change point detection algorithm in real time. The algorithm recursively calculates the probability distribution of the data segment to which the exhaust adequacy index belongs at the current moment, and dynamically identifies whether the exhaust adequacy index has entered the stable region based on the probability distribution. When the exhaust fullness index is detected to have entered a stable region and is currently in the later stable decay stage, the average value of the data after entering the stable region is used as the benchmark value, and the standard deviation of the data after entering the stable region is used as the fluctuation amount. The adaptive index threshold is set to the benchmark value minus the fluctuation amount of the preset multiple. The condition for determining the stable region is that within a consecutive preset number of sampling points, the fluctuation amplitude of the exhaust adequacy index is less than a preset fluctuation threshold.

6. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 5, characterized in that: The pumping speed of the vacuum pump is adjusted by generating a pumping speed adjustment command based on the real-time exhaust adequacy index. When the exhaust adequacy index is less than the first preset threshold, a first pumping speed adjustment command is generated to increase the current pumping speed to the rated pumping speed. When the exhaust fullness index is greater than or equal to the first preset threshold and less than the second preset threshold, a second pumping speed adjustment command is generated to reduce the current pumping speed to the first preset pumping speed. When the exhaust fullness index is greater than or equal to the second preset threshold, a third pumping speed adjustment command is generated to reduce the current pumping speed to the second preset pumping speed. When the increase in exhaust fullness index within a preset time period is less than a preset threshold, a fourth pumping speed adjustment command is generated. Based on the fourth pumping speed adjustment command, the intelligent judgment unit controls the heating device to temporarily increase the current heating power by a preset power increment at a first heating adjustment rate, and controls the vacuum pump to reduce the current pumping speed to a third preset pumping speed. After the duration of the temporary increase reaches a preset time period, the heating device is controlled to restore the original heating power at a second heating adjustment rate.

7. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 1, characterized in that, Includes the following steps: The preliminary endpoint conditions include: the current exhaust stage is in the late stable decay stage, and the exhaust sufficiency index is continuously lower than the adaptive index threshold within the first preset time period, and the exhaust sufficiency index will not rebound to exceed the adaptive index threshold within the second preset time period as predicted by the time-series prediction AI model. The sealing verification includes: closing the vacuum pump inlet valve and monitoring the rate of pressure rise in the vacuum heat pipe within a predetermined preset time. If the rate of rise is less than a preset rate threshold, the sealing verification is deemed to have passed; otherwise, an alarm is triggered.

8. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 7, characterized in that, Includes the following steps: The dual physical verification includes mass conservation verification and spectral feature verification; Once the sealing verification is passed, dual physical verification is performed. If both verifications pass, the final endpoint determination result is obtained, and the intelligent judgment unit performs a sealing operation on the vacuum heat pipe. If either verification fails, the process parameters are adjusted according to the verification deviation type, and the steps from exhaust stage identification to dual physical verification are re-executed.

9. The AI-based intelligent judgment system for the exhaust endpoint of a vacuum heat pipe according to claim 3, characterized in that: The heating power adjustment of the heating device also includes: Upload the process data and quality assessment results of each exhaust operation to the cloud database; When the cumulative data volume corresponding to the same vacuum heat pipe model reaches the preset batch threshold, the cloud database performs statistical analysis on the process data and generates an optimized parameter set corresponding to the vacuum heat pipe model. The optimized parameter set includes the optimized first preset power, second preset power and third preset power, as well as the optimized first heating adjustment rate, second heating adjustment rate, third heating adjustment rate and fourth heating adjustment rate. The optimized parameter set is sent to the intelligent judgment unit to replace the original preset parameters in order to control the exhaust process of subsequent heat pipes of the same model.

10. An AI-based intelligent method for determining the exhaust endpoint of a vacuum heat pipe, characterized in that: include: The multi-source sensor signals acquired during the vacuum heat pipe exhaust process are input to the stage recognition AI model, which outputs the current exhaust stage of the vacuum heat pipe and the intelligent judgment unit adjusts the heating power of the heating device. Based on the exhaust stage, the characteristic parameters corresponding to the vacuum heat pipe are extracted from the multi-source sensor signals to analyze the exhaust adequacy index, which is used to adjust the pumping speed of the vacuum pump and generate an adaptive index threshold. The initial endpoint condition is determined based on the adaptive index threshold. When the initial endpoint signal is triggered, the intelligent judgment unit performs a sealing verification. Based on the sealing verification results, a dual physical verification is performed. If the dual physical verification passes, the final endpoint determination result is obtained. The intelligent judgment unit then performs a sealing operation on the vacuum heat pipe, completing the intelligent determination of the vacuum heat pipe exhaust endpoint.

Images