Low-temperature measured value traceability method based on LSTM and multimode calibration system

Abstract

The invention discloses an LSTM-based low-temperature measured value traceability method and a multimode calibration system, and the method achieves the quick switching of modules through the cooperative control of an LSTM model and a segmented PID, the continuous adjustment of a wide temperature region, the high-precision temperature control, the design of double cavities, and a quick disassembly structure, and is compatible with the calibration of a long-rod / short low-temperature thermometer (-196 to-80) DEG C. Through cooperative control of an LSTM prediction model and segmented PID, the heating rate reaches 5 DEG C / min, the overshoot is reduced to 0.1 DEG C from 0.5 DEG C of traditional PID, the stabilization time is shortened to 15 minutes (long rod module) from 25 minutes, the calibration efficiency is improved by 40%, and the liquid nitrogen consumption rate is reduced. A temperature and pressure multi-modal data fusion algorithm is integrated, calibration data correction processing is carried out, and the precision is improved compared with traditional single temperature parameter calibration.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of low-temperature calibration, in particular to a low-temperature temperature metrological value traceability method based on LSTM and a multi-mode calibration system. BACKGROUND

[0002] At present, there is a lack of suitable low-temperature temperature field standard device for the calibration of ultra-low-temperature thermometers in the field of biological medicine, such as -86℃ refrigerator monitoring and biological sample storage thermometers. The existing domestic systems are mainly used for liquid helium temperature zone -269℃, and cannot cover the continuous calibration requirements of the wide temperature range from -196℃ to -80℃. Foreign equipment such as the British Eutronics 459 low-temperature thermostat can be adjusted in temperature, but the price is as high as 300,000 yuan per unit, and the temperature range only covers -180℃ to -80℃, which cannot meet the calibration requirements of the low-temperature range from -196℃ to -180℃.

[0003] In the prior art, domestic low-temperature thermometer calibration systems are mostly used for liquid helium temperature zone, and have the problems of complex device structure, high operation cost, and non-adjustable temperature. For example, the (-180~-80)℃ ultra-low-temperature adjustable thermostat developed by Chengdu Measurement and Detection Research Institute is only suitable for long-stem low-temperature thermometers, and cannot meet the calibration requirements of short low-temperature thermometers. The ultra-low-temperature thermometer calibration device developed by Shanghai Measurement Institute is only suitable for short ultra-low-temperature thermometers, and has poor technical indicators, with a constant temperature block uniformity of only better than ±20mK and a temperature control stability of ≤50mK / h. Foreign low-temperature thermostat equipment such as British Eutronics and European Cambi is expensive, and the design is mostly suitable for long-stem low-temperature thermometer calibration, with poor applicability to short low-temperature thermometers. In addition, some of the existing technologies require a refrigeration machine and large equipment, which are complex to install and use, have high operation and maintenance costs, have a narrow temperature range, and cannot meet the calibration requirements of the wide temperature range from -196℃ to -80℃. At the same time, there is a lack of related calibration specifications as technical guidance, making it difficult to realize accurate value traceability of ultra-low-temperature thermometers.

[0004] Limitations of long-stem thermometers: The (-180~-80)℃ thermostat described in the patent A kind of continuously adjustable low-temperature thermostat (Patent No. ZL202223263681X) is only suitable for long-stem thermometers with an insertion depth of >400mm, and cannot be compatible with short probes such as biological sample library microsensors with a length of ≤300mm.

[0005] Limitations of short thermometers: The calibration system based on GM refrigeration machine only supports short thermometers, and the constant temperature block has a small size (insertion depth ≤200mm), which cannot accommodate long-stem thermometers, and has poor technical indicators (uniformity ≤±20mK, temperature control stability ≤50mK / h).

[0006] The traditional system relies on a refrigerator + vacuum device such as a GM refrigerator system developed by Shanghai Jiaotong University, which needs complex operations such as vacuumizing and cold screen installation, and the single calibration takes more than 2 hours and the cost is more than 3 times the hardware cost.

[0007] The existing value traceability method has the following problems: there is no unified calibration specification for the temperature range of (-196 to -80) DEG C, the existing technology mainly refers to the liquid helium temperature range (-269 DEG C) or the normal temperature calibration method, and the latent heat of phase change of liquid nitrogen, the change of low-temperature heat conduction characteristics and other factors are not considered, so that the calibration result is unreliable.

[0008] The existing calibration does not consider the influence of mechanical deformation and contact thermal resistance on temperature measurement when the thermometer is inserted. For example, different insertion depths of long rod thermometers can cause an axial temperature field deviation of up to 50mK, and the contact thermal resistance between the short probe and the heat block can cause an error of ±20mK. SUMMARY

[0009] To solve the above technical problems, the application designs a low-temperature thermometer value traceability method based on LSTM and a multi-mode calibration system.

[0010] The application adopts the following technical scheme:

[0011] The low-temperature thermometer value traceability method based on LSTM adopts a multi-mode calibration system, which includes the following steps:

[0012] (1) Supply stable cold source: inject liquid nitrogen into the Dewar flask, and heat the heat block to -196 DEG C through heat conduction;

[0013] (2) Cooperatively control the temperature rise through the LSTM model and the segmented PID: start the heating wire according to the multi-segment PID program to rise to the target temperature (such as -80 DEG C), and the radiation shield is heated synchronously to maintain thermal balance;

[0014] (3) Calibrate the temperature meter to be calibrated: when the temperature field is stable (fluctuation≤10mK / 10min), insert the temperature meter to be calibrated and the standard platinum resistance for calibration;

[0015] The calibration process in step (3) includes:

[0016] (3-1) Synchronous data acquisition of the temperature meter to be calibrated and the standard platinum resistance;

[0017] (3-2) Data preprocessing;

[0018] (3-3) Based on the LSTM model, integrate the temperature and pressure multi-modal data fusion algorithm to perform calibration data correction, indication error calculation, correction and evaluation.

[0019] Further, before the operation of step (1), the multi-mode calibration system performs multi-mode switching operation, which includes the following steps:

[0020] Close the heating system, and wait for the temperature of the heat block to rise above -100℃.

[0021] Remove the flange plate, replace the long rod / short module, and reseal before injecting liquid nitrogen for pre-cooling.

[0022] Start the controller and adjust the temperature field according to the corresponding module's PID parameters.

[0023] Further, in step (1), the liquid nitrogen level is maintained at ≥1.5 times the height of the heat block, ensuring stable supply of -196℃ cold source. Regarding the control of the liquid nitrogen level:

[0024] When a single calibration temperature point is set, a fixed threshold is used to control the liquid level: when the liquid level sensor detects that the liquid level is <1.5 times the height of the heat block, the controller will give an electrical signal to open the electromagnetic valve to start the vacuum pump for liquid replenishment. Liquid nitrogen is injected at a rate of 2L / min until the liquid level ≥2 times the height of the heat block, and the liquid replenishment is automatically turned off.

[0025] When multiple temperature calibration points are set and the system automatically adjusts the temperature, the liquid replenishment operation is combined with the LSTM model to predict the temperature change of the heat block. When the predicted temperature >-150℃ and the liquid level <threshold, the electromagnetic valve is opened in advance to replenish nitrogen. The amount of liquid replenishment is dynamically adjusted according to the temperature prediction value of the LSTM model to avoid excessive liquid replenishment causing temperature field fluctuations. In the extremely low temperature zone, maintain a high start threshold of 1.5 times to ensure sufficient liquid nitrogen cold reserve. In the medium and low temperature zone, reduce the start threshold to 1.2 times to reduce unnecessary liquid nitrogen consumption. At the same time, according to the predicted warming rate of the LSTM model, adjust the liquid replenishment rate from 2L / min to 1.5L / min to avoid temperature field overshoot.

[0026] Further, during the liquid replenishment process, the system uses a dual control strategy of liquid level and temperature to ensure the stability of the cold source. When the heat block temperature fluctuates >20mK during the liquid replenishment process, the controller automatically pauses the liquid replenishment, and continues the liquid replenishment after the temperature stabilizes to prevent the temperature field from dropping sharply due to liquid nitrogen injection.

[0027] Further, in step (2), the LSTM model is combined with a segmented PID program. The segmented PID program automatically switches between 3 sets of PID parameters based on the differences in thermodynamic characteristics between -196℃ and -80℃ temperature zones.

[0028] When the constant temperature section temperature is -196~-150℃, K p =1.8, K i =0.08, K d =1.2, warming rate 5℃ / min, overshoot ≤0.5℃.

[0029] PID program is segmented, when the temperature of constant temperature section is -150~ -100℃, the parameters are set: K p = 1.2, K i = 0.05, K d = 0.8;

[0030] When the temperature of constant temperature section is -100~ -80℃, the parameters are set: K p = 0.8, K i = 0.03, K d = 0.5, and the temperature fluctuation is maintained ≤10mK / 10min.

[0031] Further, the sampling frequency of the temperature gauge to be calibrated in the step (3-1) is 1Hz, and the stable section is collected for 10min, and a total of 600 data points are collected, the digital temperature gauge directly collects the temperature value, and the non-digital recording output signal voltage or resistance is converted into the temperature value.

[0032] Further, in the step (3-2):

[0033] First, the temperature field stability is screened, the data of which the temperature field fluctuation exceeds 10mK / 10min period is removed, and the continuous 10min stable data is retained;

[0034] Then, the noise is removed, a 5th order Butterworth low-pass filter is applied, and the cut-off frequency is 0.1Hz, and the high-frequency noise caused by the boiling of liquid nitrogen is eliminated;

[0035] Then, the data alignment is performed, and the standard device and the temperature gauge to be calibrated are aligned based on the time stamp.

[0036] Further, the temperature data is removed by the 5th order Butterworth low-pass filter, the pressure data is standardized to the interval [-1, 1], and the formula is:

[0037]

[0038] Wherein, x is the original measurement value of the pressure, and the pressure unit is MPa; x min is the minimum value of the feature in the training data, and the pressure unit is MPa; x max is the maximum value of the feature in the training data, and the pressure unit is MPa.

[0039] Further, in the step (3-3), after the temperature and pressure data are preprocessed by the 24-bit ADC with a frequency of 10Hz, every 10 seconds is aggregated into a group of feature vectors, and a time window containing 100 groups of data is formed;

[0040] The collected 100 groups of historical feature vectors are input into the LSTM model to predict the temperature in the next 30 seconds, and the deviation between the predicted value and the current value is calculated.

[0041] If the prediction deviation is greater than 5mK, the PID parameters are adjusted 10 seconds in advance. When the temperature is predicted to rise, the Kp value is reduced to suppress overshoot. When the temperature is predicted to drop, the Ki value is increased to speed up the response.

[0042] Further, the LSTM model uses an attention mechanism to calculate the attention weight of each time step hidden state, dynamically focusing on key features, temperature mutation points or pressure abnormal fluctuation points, so that the model pays more attention to data segments that have a greater impact on temperature field prediction, and improves the prediction accuracy of the LSTM model for ultra-low temperature temperature field changes

[0043] The attention mechanism formula is:

[0044]

[0045] Where h is the LSTM hidden state, h i is the hidden state of the i-th time step (i = 1, 2, 3 … n), and assuming that the LSTM has 128 neural units per layer, then h and h i have a dimension of 128*1 (column vector); W, w are weight matrices, W is used to map the hidden state h to the attention score space; w is the attention weight vector, which is used to calculate the final attention score; b is the bias.

[0046] Further, in the calculation of the indicated value error, the standardizer temperature value is converted, and according to the resistance value R s of the standard platinum resistance, the actual temperature T s is calculated by the ITS-90 temperature scale formula,

[0047] ΔT i = T x -T s -ΔT grad

[0048] ΔT i is the indicated value error of the calibrated thermometer at this temperature point; ΔT grad is the axial temperature gradient correction value;

[0049] When the calibrated thermometer is inserted into the short rod module, the axial temperature gradient correction value ΔT grad = -0.002℃;

[0050] When the calibrated thermometer is inserted into the long rod module, the axial temperature gradient correction value ΔT grad = +0.003℃.

[0051] A multi-mode calibration system for implementing the above-mentioned low-temperature temperature measurement value traceability method based on LSTM, comprising a multi-mode constant temperature source module, a liquid nitrogen cold source system, an intelligent temperature control heating system and a controller;

[0052] The multi-mode constant temperature source module adopts a double-cavity integrated design and is provided with a heat block of two types of calibration cavities, the heat block comprising a long rod heat block and a short type heat block, the long rod heat block being provided with a long rod calibration cavity and being adapted to a long rod thermometer with an insertion depth of ≥400 mm, the short type heat block being provided with a short type calibration cavity and being adapted to a short type thermometer with an insertion depth of ≤300 mm, the long rod heat block and the short type heat block being quickly replaced through a flange plate.

[0053] The liquid nitrogen cold source system adopts a high-vacuum Dewar bottle configured with a liquid nitrogen automatic supply device, which contains a float type liquid level sensor and an electromagnetic valve installed in a liquid nitrogen pipeline, the float type liquid level sensor collects liquid level information, the liquid level height is converted into a dry reed tube on-off signal, which is transmitted to the controller end through a cable, and the controller end controls the electromagnetic valve and the liquid supply channel through a control logic to maintain the liquid nitrogen liquid level ≥1.5 times the height of the heat block, ensuring stable supply of the -196℃ cold source.

[0054] The intelligent temperature control heating system comprises a surrounding heating wire, an anti-radiation screen heating layer and a plurality of sensors, the sensors comprising a traceability sensor composed of a two-grade standard platinum resistance, a temperature control sensor composed of 2 groups of thin film thermocouples arranged at the center and the edge of the heat block, and a pressure sensor embedded in the heat block.

[0055] The controller supports parallel computation of the LSTM model, the controller connects the multi-mode constant temperature source module through 16-channel 24-bit ADC for collecting temperature signals, the controller connects the intelligent temperature control heating system through 8-way PWM output to control the heating power, and Thermostat body The controller is connected to the liquid nitrogen cold source system through a control port and a CAN bus.

[0056] Preferably, the heat block adopts a pure oxygen-free copper heat block with a purity of ≥99.95%, a gold-plated inner wall with a thickness of 5 μm and an emissivity of <0.02; the long rod heat block has a diameter of 150 mm and a height of 450 mm, the long rod calibration cavity has a diameter of 20 mm and a depth of 430 mm, and 3 groups of heating wires are built-in;

[0057] The short type heat block has a diameter of 150 mm and a height of 300 mm, the short type calibration cavity has a diameter of 15 mm and a depth of 280 mm, and a partition temperature control copper block is adopted, and 3 groups of heating wires are built-in.

[0058] Preferably, the heating wires are 6 groups of heating wires spirally wound on the outer wall of the heat block, a nickel-chromium alloy with a diameter of 0.5 mm, a total power of 200 W, and a 0.1 W level power adjustment realized through a solid-state relay SSR-25DA.

[0059] Preferably, the anti-radiation screen heating layer is a gold-plated copper screen wrapped by a uniform heating block outer layer, with a thickness of 0.1 mm, and a surface printed resistance value of 100Ω heating film as the main heating layer to reduce heat radiation loss.

[0060] Preferably, the traceable sensor is inserted into the calibration cavity in parallel with the temperature meter to be calibrated, with a sampling rate of 1 Hz.

[0061] The pressure sensor synchronously collects pressure data for correcting temperature measurement deviation, and the pressure sensor can be preferably a fiber Bragg grating pressure sensor with a range of 0-10 MPa and a resolution of 0.1% FS.

[0062] The controller adopts a four-core Cortex-A73 with a main frequency of 2.0 GHz and an external Xilinx Alveo U50 FPGA acceleration card.

[0063] Compared with the prior art, the present application has the following beneficial effects:

[0064] Wide temperature range continuous adjustable and high precision temperature control, double cavity design, fast module switching through quick release structure, compatible with long rod / short type low temperature thermometer(-196~ -80) ℃ range calibration.

[0065] Through the cooperative control of the LSTM prediction model and the segmented PID, the temperature rising rate reaches 5℃ / min, the overshoot is reduced from 0.5℃ of the traditional PID to 0.1℃, the stable time is shortened from 25 minutes to 15 minutes (long rod module), the calibration efficiency is improved by 40%, and the liquid nitrogen consumption rate is reduced.

[0066] Integrating temperature, pressure multi-modal data fusion algorithm, the calibration data correction processing is carried out, and the calibration precision is improved compared with the traditional single temperature parameter. DETAILED DESCRIPTION

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0068] Figure 1 It is a schematic diagram of a multi-modal calibration system.

[0069] Figure 2 It is a schematic diagram of a long type constant temperature source module.

[0070] Figure 3 It is a flow chart of a low temperature thermometer value traceability method based on LSTM.

[0071] In the figure: 1, liquid level sensor, 2, quick connector, 3, short type heat block, 4, high vacuum dewar flask, 5, heating wire, 6, radiation shield heating layer, 7, vacuum pump, 8, electromagnetic valve, 9, short type calibration cavity, 3', long rod heat block, 5', long rod heating wire, 6', long rod radiation shield heating layer, 9', long rod calibration cavity. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings of the embodiments of the present application. Figures 1-3 The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings of the embodiments of the present application.

[0073] Embodiment 1

[0074] As shown in the figure, the low-temperature temperature measurement value traceability method based on LSTM adopts a multi-mode calibration system, including the following steps: Figure 3

[0075] (1) Supply a stable cold source: inject liquid nitrogen into the dewar flask, and the heat block is cooled to-196℃ through heat conduction;

[0076] (2) Cooperatively control the temperature rise through the LSTM model and the segmented PID: the controller starts the heating wire, and raises the temperature to the target temperature (such as-80℃) according to the multi-segment PID program, and the radiation shield is synchronously heated to maintain thermal equilibrium;

[0077] (3) Calibrate the temperature meter to be calibrated: when the temperature field is stable (fluctuation≤10mK / 10min), insert the temperature meter to be calibrated and the standard platinum resistance for calibration;

[0078] The calibration process in step (3) includes:

[0079] (3-1) Synchronous data acquisition of the temperature meter to be calibrated and the standard platinum resistance;

[0080] (3-2) Data preprocessing;

[0081] (3-3) Based on the LSTM model, integrate the temperature and pressure multi-modal data fusion algorithm, and perform calibration data correction processing, indication error calculation, correction and evaluation.

[0082] Before the operation of step (1), the multi-mode calibration system performs multi-mode switching operation, including the following steps:

[0083] Turn off the heating system, and wait for the heat block temperature to rise to above-100℃;

[0084] Dismount the flange plate, replace the long rod / short type module, and reseal and inject liquid nitrogen for precooling;

[0085] ​The starting controller adjusts the temperature field according to the PID parameters of the corresponding module.

[0086] In step (1), the liquid level of liquid nitrogen is maintained at 1.5 times the height of the heating block to ensure stable supply of the-196℃ cold source. The control of the liquid level is as follows:

[0087] When a single calibration temperature point is set, a fixed threshold is used to control the liquid level. When the liquid level detected by the liquid level sensor is less than 1.5 times the height of the heating block, the liquid level threshold is 450mm for a short heating block with a height of 300mm. The controller will give an electrical signal to open the electromagnetic valve to start the vacuum pump for liquid supplement. Liquid nitrogen is injected at a rate of 2L / min until the liquid level is greater than or equal to 2 times the height of the heating block, which is 600mm in this example. The liquid supplement is automatically turned off.

[0088] When multiple temperature calibration points are set and the system needs to automatically adjust the temperature, the liquid supplement operation is also combined with the LSTM model to predict the temperature change of the heating block. When the predicted temperature is greater than-150℃ and the liquid level is less than the threshold, the electromagnetic valve is opened in advance to supplement liquid nitrogen. The amount of liquid supplement is dynamically adjusted according to the temperature prediction value (e.g., 5L for a predicted temperature rise rate of 1℃ / min and 3L for a temperature rise rate of 0.5℃ / min), avoiding excessive liquid supplement that causes temperature field fluctuations. In the extremely low temperature zone (e.g.,-196℃ to-150℃), a high starting threshold of 1.5 times is maintained to ensure sufficient liquid nitrogen cold storage. In the medium and low temperature zone (e.g.,-150℃ to-80℃), the starting threshold is reduced to 1.2 times to reduce unnecessary liquid nitrogen consumption. That is, when the temperature of the heating block rises from-196℃ to-100℃, the starting threshold is dynamically reduced from 450mm (300mm x 1.5) to 360mm (300mm x 1.2), and the stop threshold remains at 600mm (300mm x 2). The starting threshold is reduced, the stop threshold remains unchanged, the liquid supplement interval is reduced, and the liquid supplement frequency is reduced. At the same time, the LSTM predicts the temperature rise rate, and the liquid supplement rate is adjusted from 2L / min to 1.5L / min to avoid temperature field overshoot.

[0089] During the liquid supplement process, the system uses a dual control strategy of liquid level and temperature to ensure the stability of the cold source. When the temperature of the heating block fluctuates more than 20mK during the liquid supplement process, the controller automatically pauses the liquid supplement, and continues after the temperature stabilizes to prevent sudden temperature drop caused by liquid nitrogen injection.

[0090] Further, in step (2), the LSTM model is used in combination with a segmented PID program. The segmented PID program automatically switches between three groups of PID parameters based on the differences in thermodynamic characteristics in the-196℃ to-80℃ temperature zone.

[0091] When the temperature of the constant temperature section is-196℃ to-150℃, K p = 1.8, K i = 0.08, and K d= 1.2, heating rate 5℃ / min, overshoot≤0.5℃;

[0092] Segmented PID program, when the temperature of the constant temperature section is-150-100℃, set the parameters: K p = 1.2, K i = 0.05, K d = 0.8.

[0093] When the temperature of the constant temperature section is-100-80℃, set the parameters: K p = 0.8, K i = 0.03, K d = 0.5, maintain temperature fluctuation≤10mK / 10min.

[0094] Table 1: Temperature intelligent adjustment table

[0095] Temperature zone range K p ]]> K i ]]> K d ]]> Dead zone (mK) Ramp rate (°C / min) -196℃～-150℃ 1.8 0.08 / min 1.2 10 5 -150℃～-100℃ 1.2 0.05 / min 0.8 10 5 -100℃～-80℃ 0.8 0.03 / min 0.5 20 3

[0096] In step (3-1), the platinum and the temperature gauge are synchronously collected data, the sampling frequency is 1Hz (default, can be set), the stable section is continuously collected for 10min, a total of 600 data points, the digital temperature gauge directly collects the temperature value, and the non-digital type records the output signal voltage or resistance and converts it into the temperature value.

[0097] In the step (3-2), first, the temperature field stability screening is performed, the data of the temperature field fluctuation exceeding 10mK / 10min period is removed, and the continuous 10min steady-state data is retained;

[0098] Then, the noise is removed, a 5th order Butterworth low-pass filter is applied, the cutoff frequency is 0.1Hz, and the high-frequency noise caused by the boiling of liquid nitrogen is eliminated;

[0099] After that, the data alignment is performed, and the standard device and the calibrated temperature gauge data are synchronously aligned based on the time stamp.

[0100] The temperature data is removed by the 5th order Butterworth low-pass filter, the pressure data is standardized to the interval [-1, 1], and the formula is adopted:

[0101]

[0102] Wherein, x is the original measured value of the pressure, the pressure unit is MPa; x min is the minimum value of the feature in the training data, the pressure unit is MPa; x max is the maximum value of the feature in the training data, the pressure unit is MPa.

[0103] If the original measured value of the pressure is 7.5MPa, the measurement range is 0-10MPa, then x norm = 0.5.

[0104] In step (3-3), the temperature and pressure data collected by the 24-bit ADC at a frequency of 10 Hz are preprocessed, and every 10 seconds are aggregated into a feature vector to form a time window containing 100 groups of data;

[0105] The 100 groups of historical feature vectors collected are input into the LSTM model to predict the temperature in the next 30 seconds, and the deviation between the predicted value and the current value is calculated;

[0106] If the predicted deviation is greater than 5 mK, adjust the PID parameters 10 seconds in advance, reduce the Kp value when predicting warming to suppress overshoot, and increase the Ki value when predicting cooling to speed up the response.

[0107] The LSTM model uses an attention mechanism formula:

[0108]

[0109] where h is the LSTM hidden state, h i is the hidden state at the i-th time step (i = 1, 2, 3,..., n), and it is assumed that the LSTM has 128 neural units per layer, then the dimensions of h and h i are 128*1 (column vector); W and w are weight matrices, W is used to map the hidden state h to the attention score space; w is an attention weight vector used to calculate the final attention score; and b is a bias.

[0110] This formula calculates the attention weight of each time step hidden state, dynamically focusing on key features such as temperature mutation points during liquid nitrogen phase change or pressure abnormal fluctuation points, so that the model pays more attention to data segments that have a greater impact on temperature field prediction, improving the prediction accuracy of LSTM for ultra-low temperature temperature field changes. Through the model, 100 groups of historical features are input every second, and the temperature prediction value in the next 30 seconds is output. If the predicted deviation is greater than 5 mK, adjust the PID parameters 10 seconds in advance, reduce the Kp value (such as from 1.8 to 1.5) when predicting warming to suppress overshoot, and increase the Ki value (such as from 0.08 to 0.1) when predicting cooling to speed up the response.

[0111] In the calculation of the indication error, the standard temperature value is converted according to the resistance value R s of the standard platinum resistance, and the actual temperature T s is calculated by the ITS-90 temperature scale formula,

[0112] ΔT i = T x - T s - ΔT grad

[0113] ΔT i is the indication error of the temperature point of the thermometer being calibrated; and ΔT gradfor the axial temperature gradient correction value;

[0114] The calibrated thermometer is inserted into the short rod module with an insertion depth of 280 mm, and the axial temperature gradient correction value AT grad = -0.002°C;

[0115] The calibrated thermometer is inserted into the long rod module with an insertion depth of 400 mm, and the axial temperature gradient correction value AT grad = +0.003°C.

[0116] The precise calculation of the indication error avoids the indirect error of temperature coefficient fitting, directly provides reliable deviation evaluation for the metrological traceability of the ultra-low temperature thermometer, and meets the high-precision requirements of temperature measuring equipment in the fields of biological medicine, deep space exploration, etc.

[0117] Example 2:

[0118] A multi-mode calibration system, as shown in Figures 1-2 for implementing the low-temperature temperature metrological value traceability method based on the LSTM in Example 1, the LSTM model is suitable for calibration of ultra-low temperature thermometers in the range of (-196 to -80) °C, and the core is composed of four parts: a double-cavity multi-mode constant temperature source module, a liquid nitrogen cold source system, an intelligent temperature control heating system, and a room temperature controller. The overall device adopts a vertical structure, the bottom is a high-vacuum Dewar bottle with a built-in heat block assembly, and the top is connected to the controller and sensor cable interface through a flange.

[0119] The multi-mode constant temperature source module adopts a double-cavity integrated design, and is provided with two types of heat blocks for calibration cavities. The heat block is made of oxygen-free copper with a purity of ≥99.95%, the inner wall is gold-plated, the thickness is 5 μm, and the emissivity is <0.02. The long rod heat block 3 has a diameter of 150 mm and a height of 450 mm, and is provided with a long rod calibration cavity 9 with a diameter of 20 mm and a depth of 430 mm. It is suitable for long rod thermometers with an insertion depth of ≥400 mm, such as long rod low-temperature thermometers for cold chain transportation monitoring. Three groups of heating wires (nickel-chromium alloy, resistance 50Ω) are built-in.

[0120] The short heat block 3 has a diameter of 150 mm and a height of 300 mm, and is provided with a short calibration cavity 9 with a diameter of 15 mm and a depth of 280 mm. It adopts a partitioned temperature control copper block, and is built-in with three groups of heating wires (nickel-chromium alloy, resistance 50Ω). It is suitable for short thermometers with an insertion depth of ≤300 mm, such as micro sensors for biological sample libraries. The long rod / short type module is quickly replaced through a flange (M40x1.5 thread) and a quick connector 2, and the switching time is <5 minutes. The module contact surface adopts a fluororubber low-temperature sealing ring to prevent cold leakage.

[0121] The liquid nitrogen cooling system employs a high-vacuum Dewar flask 4 equipped with an automatic liquid nitrogen replenishment device, including a float-type liquid level sensor 1 and a solenoid valve 8. The float-type liquid level sensor 1 collects liquid level information, converting the liquid level height into a reed switch on / off signal, which is transmitted to the controller via cable. The controller controls the solenoid valve and the liquid replenishment channel through control logic. A vacuum pump is also installed on the liquid replenishment channel to create a vacuum environment to reduce heat conduction and improve the insulation effect of the constant temperature source. The liquid nitrogen level is maintained at ≥ 1.5 times the height of the heat spreader, ensuring a stable supply of -196℃ cold source.

[0122] The solenoid valve is a cryogenic, normally closed type with a DN25 diameter. It connects the high-vacuum Dewar flask and the replenishment pipeline and is sealed with fluororubber. The replenishment pipeline is a 25mm diameter stainless steel bellows wrapped with MLI insulation and has a filter at the inlet.

[0123] Intelligent temperature control heating system: Heating and temperature control consist of a surrounding heating wire 5 and a radiation shield heating layer 6. Six sets of heating wires, 0.5mm diameter nickel-chromium alloy, are spirally wound around the outer wall of the heat spreader, with a total power of 200W. Power adjustment in 0.1W increments is achieved through a solid-state relay SSR-25DA. Figure 1 As shown, the heat spreader (a short heat spreader) is wrapped with a gold-plated copper screen with a thickness of 0.1 mm. A heating film with a resistance value of 100Ω is printed on the surface, serving as the main heating layer to reduce heat radiation loss. If it is a long-rod heat spreader, it is also equipped with a long-rod heating wire 5' and a long-rod radiation shielding heating layer 6'.

[0124] A second-order standard platinum resistance thermometer (PT1000, uncertainty ≤0.015℃) and the thermometer to be calibrated were inserted in parallel into the calibration chamber, with a sampling rate of 1Hz, as traceability sensors. Two sets of thin-film thermocouples were arranged at the center and edge of the heat spreader as temperature control sensors. Simultaneously, an auxiliary sensor, i.e., a pressure sensor, was embedded inside the heat spreader to synchronously collect pressure data and correct for temperature measurement deviations. A fiber Bragg grating pressure sensor with a range of 0-10MPa and a resolution of 0.1%FS was preferred as the pressure sensor.

[0125] The controller uses a quad-core Cortex-A73 with a main frequency of 2.0GHz and is connected to an external Xilinx Alveo U50 FPGA acceleration card. It supports parallel computing of LSTM models. The control module includes a 16-channel 24-bit ADC for acquiring temperature signals, 8-channel PWM output to control heating power, and is connected to the thermostat body through a control port. The CAN bus is connected to the liquid nitrogen replenishment system.

[0126] Example 3:

[0127] Calibrate a 150℃ long rod thermometer with an insertion depth of 400mm and a target temperature of -150℃. Control requirements: temperature field uniformity ≤10mK, stabilization time ≤15min.

[0128] LSTM model deployment

[0129] Input data example: Temperature sequence: [-196℃, -195.8℃,..., -155℃] (100 time steps, interval 10s); Pressure sequence: [0.1MPa, 0.102MPa,..., 0.11MPa] Sampling interval is the same, the control system calculates, model output: Future 30s temperature prediction value: [-150.2℃, -150.1℃, -150.0℃], single step inference time 0.8ms, meet the prediction frequency of 1Hz.

[0130] PID regulation process

[0131] Warming-up stage (-196℃→-150℃), PID_COLD parameters (K p =1.8, K i =0.08, K d =1.2) are used, warming-up rate 5℃ / min, time-consuming 9.2min; overshoot: 0.3℃ (-150℃→-149.7℃), LSTM predicts overshoot 5min in advance, K p is automatically reduced to 1.5, overshoot is reduced to 0.1℃.

[0132] Constant temperature stage, switch to PID_MID parameters, collect temperature field data every 10s, center control sensor temperature -150.002℃±0.005℃, edge temperature -149.998℃±0.007℃, temperature field uniformity: 6mK, meet the requirement of ≤10mK; stability within 10min fluctuation ≤8mK, better than the index requirement (≤10mK / 10min).

[0133] Table 2 Comparison of traditional PID control, LSTM model and segmented PID collaborative control

[0134] Indicator Traditional PID control LSTM + segmented PID Overshoot 0.5℃ 0.1℃ Stabilization time 25 minutes 15 minutes Temperature field uniformity 15 mK 6 mK Calibration total liquid nitrogen consumption 2.5L 1.8L

[0135] As shown in Table 2 above, through the collaborative control of LSTM model and segmented PID, high-precision temperature field prediction and dynamic adjustment are realized in the wide temperature range of -96℃ to -80℃, which significantly improves the efficiency and accuracy of ultra-low temperature thermometer calibration compared with the traditional scheme.

[0136] Wide temperature range continuous adjustable and high-precision temperature control, double-cavity design, through quick-release structure to realize module quick switching, compatible with long rod / short type low temperature thermometer (-196~-80)℃ range calibration.

[0137] Through the LSTM prediction model and the segmented PID collaborative control, the temperature rising rate reaches 5℃ / min, the overshoot is reduced from 0.5℃ of the traditional PID to 0.1℃, the stable time is shortened from 25 minutes to 15 minutes (long rod module), the calibration efficiency is improved by 40%, and the liquid nitrogen consumption rate is reduced.

[0138] The integrated temperature and pressure multi-modal data fusion algorithm is used for calibration data correction processing, and the calibration precision is improved compared with the traditional single temperature parameter calibration.

[0139] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary technical personnel; when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope of the present application.

Claims

1. A method for low temperature metrological value traceability based on LSTM, characterized in that, Adopting a multi-mode calibration system, comprising the following steps: (1) Supplying a stable cold source: injecting liquid nitrogen into a Dewar flask, and heating the block to -196℃ through heat conduction; (2) Controlling the temperature by the LSTM model and the segmented PID: starting the heating wire, and heating to the target temperature (such as -80℃) according to the multi-segmented PID program, and synchronously heating the radiation shield to maintain thermal balance; (3) Calibrating the temperature meter to be calibrated: when the temperature field is stable (fluctuation ≤10mK / 10min), inserting the temperature meter to be calibrated and the standard platinum resistance for calibration; The calibration process in step (3) comprises: (3-1) Synchronously collecting data of the temperature meter to be calibrated and the standard platinum resistance; (3-2) Data preprocessing; (3-3) Based on the LSTM model, integrating the temperature and pressure multi-modal data fusion algorithm, and performing calibration data correction processing, indicating error calculation, correction and evaluation.

2. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, Before the operation of step (1), the multi-mode calibration system performs multi-mode switching operation, comprising the following steps: Turning off the heating system, and waiting for the temperature of the heating block to rise to above -100℃; Dismantling the flange plate, replacing the long rod / short module, and resealing and then injecting liquid nitrogen for precooling; Starting the controller, and adjusting the temperature field according to the PID parameters of the corresponding module.

3. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, In step (1), the liquid nitrogen level is maintained to be ≥1.5 times the height of the heating block, to ensure stable supply of the -196℃ cold source, and the control of the liquid nitrogen level is as follows: When a single calibration temperature point is set, a fixed threshold is used to control the liquid level: when the liquid level sensor detects that the liquid level is <1.5 times the height of the heating block, the controller will give an electrical signal to open the electromagnetic valve to start the vacuum pump for liquid supplementing, the liquid nitrogen is injected at a rate of 2L / min, and the liquid supplementing is automatically stopped until the liquid level is ≥2 times the height of the heating block; When multiple temperature calibration points are set, and the system automatically adjusts the temperature, the liquid supplementing operation is combined with the LSTM model to predict the temperature change of the heating block: when the predicted temperature is >-150℃ and the liquid level is <the threshold, the electromagnetic valve is opened in advance to supplement liquid nitrogen, and the amount of liquid supplementing is dynamically adjusted according to the temperature prediction value of the LSTM model, to avoid excessive liquid supplementing causing temperature field fluctuation; in the extremely low temperature zone, the starting threshold is maintained to be 1.5 times, to ensure sufficient liquid nitrogen cold energy reserve; in the medium and low temperature zone, the starting threshold is reduced to 1.2 times, to reduce unnecessary liquid nitrogen consumption; and according to the predicted heating rate of the LSTM model, the liquid supplementing rate is adjusted from 2L / min to 1.5L / min, to avoid temperature field overshoot.

4. The LSTM-based low-temperature metrological value traceability method according to claim 3, characterized in that, During the liquid supplementing process, the system performs double control strategy of liquid level and temperature to ensure the stability of the cold source: when the temperature fluctuation of the heating block is >20mK during the liquid supplementing process, the controller automatically suspends the liquid supplementing, and continues the liquid supplementing after the temperature is stable, to prevent sudden drop of the temperature field caused by liquid nitrogen injection.

5. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, In step (2), the LSTM model is combined with the segmented PID program, and the segmented PID program is automatically switched among 3 groups of PID parameters according to the differences in thermodynamic characteristics of the temperature zones from -196℃ to -80℃: When the temperature of the constant temperature section is -196 to -150°C, K p = 1.8, K i = 0.08, K d = 1.2, the heating rate is 5°C / min, and the overshoot is ≤0.5°C; Segmented PID program, when the constant temperature section temperature is -150 to -100°C, set parameters: K p = 1.2, K i = 0.05, K d = 0.8; When the constant temperature section temperature is -100 to -80℃, set parameters: K p = 0.8, K i = 0.03, K d = 0.5, maintain temperature fluctuation ≤10mK / 10min.

6. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, The sampling frequency of the temperature meter to be calibrated in step (3-1) is 1 Hz, and the stable segment is collected for 10 minutes, a total of 600 data points. The digital temperature meter directly collects temperature values, and the non-digital type records output signal voltage or resistance and converts it into temperature values.

7. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, In step (3-2): First, the temperature field stability is screened, and data with a temperature field fluctuation exceeding 10 mK / 10 min period is removed, and continuous 10 min steady-state data is retained; Then, noise is removed, and a 5th order Butterworth low-pass filter is applied with a cutoff frequency of 0.1 Hz to eliminate high-frequency noise caused by liquid nitrogen boiling; Then, data alignment is performed, and the standard and the temperature meter data are aligned based on the time stamp.

8. The LSTM-based low-temperature metrological value traceability method according to claim 7, characterized in that, The temperature data is filtered through a 5th order Butterworth low-pass filter to eliminate high-frequency noise, and the pressure data is standardized to the interval [-1, 1], using the formula: where x is the raw measurement of pressure in MPa; x min is the minimum value of this feature in the training data in MPa; x max is the maximum value of this feature in the training data in MPa.

9. The LSTM-based cryogenic temperature meter value traceability method of claim 1, wherein, In step (3-3), after pre-processing the temperature and pressure data collected by the 24-bit ADC at a frequency of 10 Hz, 10 seconds of data are aggregated into a feature vector, forming a time window containing 100 groups of data; The 100 groups of historical feature vectors collected are input into the LSTM model to predict the temperature 30 seconds in the future, and the deviation between the predicted value and the current value is calculated; If the prediction deviation is >5 mK, adjust the PID parameters 10 seconds in advance. When predicting a temperature rise, reduce the Kp value to suppress overshoot, and when predicting a temperature drop, increase the Ki value to accelerate the response.

10. The LSTM-based low-temperature metrological value traceability method according to claim 9, characterized in that, The LSTM model uses an attention mechanism to calculate the attention weight of each time step hidden state, dynamically focusing on key features, temperature mutation points or pressure abnormal fluctuation points, so that the model pays more attention to data segments that have a greater impact on temperature field prediction, improving the prediction accuracy of the LSTM model for ultra-low temperature temperature field changes The attention mechanism formula is: where h is the LSTM hidden state, h i is the hidden state at the i-th time step (i = 1, 2, 3,... n), assuming that there are 128 neural units in each layer of LSTM, then h and h i have the dimension of 128 * 1 (column vector); W, w are weight matrices, W is used to map the hidden state h to the attention score space; w is the attention weight vector, used to calculate the final attention score; b is the bias.

11. The LSTM-based low-temperature metrological value traceability method according to claim 1, characterized in that, The standard temperature value conversion is carried out in the indication error calculation, and the actual temperature T is calculated by the ITS-90 temperature scale formula according to the resistance value R of the standard platinum resistance s , The actual temperature T is calculated by the ITS-90 temperature scale formula according to the resistance value R of the standard platinum resistance s , ΔT i = T x - T s - ΔT grad ΔT i is the indicated error of the temperature point being calibrated; ΔT grad is the axial temperature gradient correction value; The calibrated thermometer is inserted into the short rod module, and the axial temperature gradient correction value ΔT grad = -0.002°C; The temperature of the thermometer to be calibrated is inserted into the long rod module, and the axial temperature gradient correction value ΔT grad = +0.003°C.

12. A multimode calibration system, characterized by The method for realizing the low-temperature temperature metrological value traceability based on LSTM according to any one of claims 1-11 comprises a multi-mode constant temperature source module, a liquid nitrogen cold source system, an intelligent temperature control heating system, and a controller; The multi-mode constant temperature source module adopts a double-cavity integrated design and is provided with two types of calibration cavities, i.e., a long rod heat block and a short type heat block. The long rod heat block is provided with a long rod calibration cavity, and is suitable for a long rod thermometer with an insertion depth of ≥400 mm. The short type heat block is provided with a short type calibration cavity, and is suitable for a short type thermometer with an insertion depth of ≤300 mm. The long rod heat block and the short type heat block can be quickly replaced through a flange plate. The liquid nitrogen cold source system adopts a high-vacuum Dewar bottle configured with a liquid nitrogen automatic supply device, which comprises a float type liquid level sensor and an electromagnetic valve installed in a liquid nitrogen pipeline. The float type liquid level sensor collects liquid level information, and the liquid level height is converted into a dry reed tube on-off signal, which is transmitted to the controller end through a cable. The controller end controls the electromagnetic valve and the liquid supplement channel through a control logic to maintain the liquid nitrogen liquid level ≥1.5 times the height of the heat block, ensuring stable supply of the-196℃ cold source. The intelligent temperature control heating system comprises a surrounding heating wire, a radiation-proof screen heating layer and a plurality of sensors, the sensors comprising a traceable sensor composed of a two-standard platinum resistance, a temperature control sensor composed of two groups of thin film thermocouples arranged at the center and the edge of the heat block, and a pressure sensor embedded in the heat block; The controller supports parallel calculation of the LSTM model, and is connected with the multi-mode constant temperature source module through 16-channel 24-bit ADC for collecting temperature signals, and is connected with the intelligent temperature control heating system through 8-way PWM output for controlling heating power, and is connected with the multi-mode constant temperature source module through a control port, and is connected with the liquid nitrogen cold source system through a CAN bus.

13. The multi-mode calibration system of claim 12, wherein, The heat block adopts a pure oxygen-free copper heat block with a purity of greater than or equal to 99.95%, a gold-plated inner wall, a thickness of 5 microns and an emissivity of less than 0.02; the long rod heat block has a diameter of 150 mm, a height of 450 mm, a long rod calibration cavity with a diameter of 20 mm and a depth of 430 mm, and three groups of heating wires built-in; The short heat block has a diameter of 150 mm, a height of 300 mm, a short calibration cavity with a diameter of 15 mm and a depth of 280 mm, and adopts a zoned temperature control copper block with three groups of heating wires built-in.

14. The multi-mode calibration system of claim 12, wherein, The heating wire is six groups of heating wires spirally wound on the outer wall of the heat block, with a diameter of 0.5 mm nickel-chromium alloy and a total power of 200 W, and the power is adjusted by a solid-state relay SSR-25DA to 0.1 W level.

15. The multi-mode calibration system of claim 12, wherein, The outer layer of the heat block is wrapped with a radiation-proof screen heating layer, which is a gold-plated copper screen with a thickness of 0.1 mm and a surface printed resistance value of 100Ω heating film as the main heating layer to reduce heat radiation loss.

16. The multi-mode calibration system of claim 12, wherein, The traceable sensor and the temperature meter to be calibrated are inserted into the calibration cavity in parallel, and the sampling rate is 1 Hz. The pressure sensor synchronously collects pressure data for correcting temperature measurement deviation, and the pressure sensor is a fiber Bragg grating pressure sensor with a range of 0-10 MPa and a resolution of 0.1% FS.

Images