A high-precision temperature prediction method for a cryogenic sensor based on deep learning

Abstract

The present invention provides a high-precision temperature prediction method for a cryogenic sensor based on deep learning. The method includes: obtaining a data set of the temperature-resistance correspondence of the cryogenic sensor; preprocessing the data set by using Chebyshev polynomials and the normalization method to obtain a processed data set; constructing a deep learning model, adjusting hyperparameters, segmenting the processed data set into a low-temperature region and a high-temperature region, where the low-temperature region and the high-temperature region have an overlapping interval, and training the processed data set; predicting the temperature value according to the training result. By combining the deep learning method with Chebyshev polynomials, the present invention can improve the temperature prediction accuracy of the cryogenic sensor.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence prediction, and specifically, to a high-precision temperature prediction method for a cryogenic sensor based on deep learning. Background Art

[0002] Before a negative temperature coefficient cryogenic sensor is used, it needs to be calibrated, that is, the relationship between resistance and temperature is obtained, and a fitting polynomial needs to be selected during the calibration process. The standard for measuring the fitting effect is the standard deviation. Currently, the best fitting equation is the Chebyshev polynomial. The fitting order of the fitting equation and the number of sampling points affect the fitting accuracy. Therefore, an automatic segmented Chebyshev polynomial fitting method has been developed, which better realizes the purpose of curve fitting. However, there are still some problems. The prediction result accuracy of this method still has room for improvement in actual application; and in the overlapping area of the segments, the predicted value of this method takes the average of the two segments. Since the accuracy of the segment boundary points is relatively low, taking the average cannot obtain the best effect; moreover, different fitting orders need to be traversed during the segmentation process, and the process is relatively cumbersome.

[0003] Xiaohe Tang et al. (An Optimal Piecewise Chebyshev Fitting Method to Calibrate Cryogenic Temperature Sensors[C], Proc. SPIE 11617, International Conference on Optoelectronic and Microelectronic Technology and Application, 116173D, 4 December 2020) studied an optimal piecewise Chebyshev fitting equation (OPCFM) for calibrating cryogenic temperature sensors (especially negative temperature coefficient resistors). Compared with other commonly used fitting polynomials, the Chebyshev polynomial has the best fitting performance in the low-temperature range of 14.5 - 40 K. OPCFM can well find the optimal segmentation, overlapping intervals, and the optimal fitting order to minimize the standard variance of each segment. The Chebyshev polynomial can theoretically approximate any function, but the method mentioned in the article uses a non-deep learning algorithm for fitting, and the achievable accuracy is limited. In the treatment of the overlapping region of two segments, the method in the article directly takes the average of the two segments as the predicted value of the overlapping region. However, since the closer to the boundary point, the greater the deviation of the predicted value, the larger deviation of one segment may affect the final result. The original method needs to determine the fitting order while determining the segmentation boundary. For example, if there are n boundary points to traverse and m fitting orders to traverse, then finally m * n calculations are required. The process is relatively cumbersome.

[0004] After retrieval, the Chinese invention patent with the authorization announcement number CN105588667B discloses a high-precision thermistor thermometer calibration device. A temperature data fitting system is added to the calibration device, which can select different fitting methods to fit the obtained temperature data, compare the fitting accuracy, and obtain the most accurate fitting result, so as to achieve the precise calibration of the thermistor thermometer. And during the actual calibration process, multiple calibration points and segmented fittings can be carried out according to different fitting methods to verify the fitting accuracy. However, this invention only performs temperature fitting through the fitting equation, and the temperature segmentation of the temperature measurement range is not precise enough, and the fitting accuracy still needs to be improved. Summary of the Invention

[0005] Aiming at the defects in the prior art, the purpose of the present invention is to provide a high-precision temperature prediction method for cryogenic sensors based on deep learning, which can improve the temperature prediction accuracy of cryogenic sensors.

[0006] According to one aspect of the present invention, there is provided a high-precision temperature prediction method for cryogenic sensors based on deep learning, including:

[0007] Obtain a data set of the corresponding relationship between the temperature and resistance of the low-temperature sensor;

[0008] Preprocess the data set by using Chebyshev polynomials and normalization method to obtain a processed data set;

[0009] Build a deep learning model. After adjusting the hyperparameters, segment the processed data set into a low-temperature region and a high-temperature region. The low-temperature region and the high-temperature region have an overlapping interval, and train the processed data set;

[0010] Predict the temperature value according to the training result.

[0011] Preferably, the preprocessing of the data set by using Chebyshev polynomials and normalization method includes:

[0012] The input of the deep learning model is resistance and the output is temperature. Add feature terms to the resistance by using Chebyshev polynomials to accelerate the convergence speed. The Chebyshev polynomial is:

[0013]

[0014] where T is the absolute temperature, R is the resistance, x is a variable whose absolute value is not greater than 1, A and B are normalization constants, n is the fitting order, and a i is the fitting coefficient;

[0015] Perform normalization processing on the input resistance, the generated feature terms, and the output temperature respectively to obtain a processed data set.

[0016] Preferably, the fitting order of the Chebyshev polynomial is set to half of the number of calibration points.

[0017] Preferably, the maximum-minimum normalization method is selected for the normalization processing:

[0018]

[0019] where x' is the normalized sample data, x is the sample data before normalization, max(x) is the maximum value of the sample data, and min(x) is the minimum value of the sample data.

[0020] Preferably, the building of the deep learning model, after adjusting the hyperparameters, segmenting the processed data set into a low-temperature region and a high-temperature region, where the low-temperature region and the high-temperature region have an overlapping interval, and training the processed data set includes:

[0021] Build a deep learning model, add an LSTM layer and a fully connected layer, and add the L1 regularization method;

[0022] Divide the processed data set into a training set and a validation set, then send the training set into the deep learning model for training, and use the validation set to adjust the hyperparameters related to the deep learning model;

[0023] Segment the processed data set to form a low-temperature region and a high-temperature region. The low-temperature region and the high-temperature region have an overlapping interval, and the segmentation boundary points are determined through the complete processed data set.

[0024] Preferably, when building the deep learning model, adding an LSTM layer and a fully connected layer, and adding the L1 regularization method, the objective function is:

[0025]

[0026] where L is the objective function, n is the number of samples, T is the true value of the calibration point temperature, t is the predicted value of the calibration point temperature, k is the number of feature terms, and w is the weight; the deep learning model is used to minimize the objective function.

[0027] Preferably, the hyperparameters include the regularization coefficient.

[0028] Preferably, after segmenting the processed data set into a low-temperature region and a high-temperature region, the fitting degrees of the Chebyshev polynomials are respectively set to half of the number of calibration points in the low-temperature region and the high-temperature region.

[0029] Preferably, the determination of the segmentation boundary points through the complete processed data set includes:

[0030] The temperature of the processed data set decreases from left to right. Take the leftmost end of the processed data set as the left boundary point of the low-temperature region, and traverse the calibration points in the middle for the right boundary point. After training the model, calculate the mean square error of the entire low-temperature region, and select the right endpoint with the lowest mean square error as the right boundary point of the low-temperature region;

[0031] Shift 2-3 points to the left from the right boundary point of the low-temperature region as the left boundary point of the high-temperature region, and take the rightmost end of the processed data set as the right boundary point of the high-temperature region.

[0032] Preferably, the prediction of the temperature value according to the training result includes:

[0033] For the low-temperature region and the high-temperature region, multiple models are trained. Each model performs forward inference to obtain predicted values. After removing the maximum and minimum values, take the average as the final result to obtain the low-temperature region temperature value and the high-temperature region temperature value;

[0034] For the overlapping interval, select the midpoint of the overlapping interval as the demarcation point between the low-temperature region and the high-temperature region. For the data within the overlapping interval to the left of the demarcation point, use the temperature value of the low-temperature region as the predicted value, and for the data within the overlapping interval to the right of the demarcation point, use the temperature value of the high-temperature region as the predicted value.

[0035] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0036] 1. The high-precision temperature prediction method for a low-temperature sensor based on deep learning of the present invention can improve the prediction accuracy of the temperature sensor by combining the Chebyshev polynomial through the deep learning method.

[0037] 2. The high-precision temperature prediction method for a low-temperature sensor based on deep learning of the present invention can improve the temperature prediction accuracy of the overlapping interval by taking the demarcation point for the overlapping interval between the low-temperature region and the high-temperature region, and using the predicted values of the segments with smaller deviations on both the left and right sides as the results.

[0038] 3. The high-precision temperature prediction method for a low-temperature sensor based on deep learning of the present invention can simplify the complexity of the method by adding a regularization method and taking half of the number of calibration points as the fitting degree. Description of the Drawings

[0039] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, objectives, and advantages of the present invention will become more apparent:

[0040] Figure 1 It is a schematic flowchart of the high-precision temperature prediction method for a low-temperature sensor based on deep learning in an embodiment of the present invention. Detailed Embodiments

[0041] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several modifications and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0042] An embodiment of the present invention provides a high-precision temperature prediction method for a low-temperature sensor based on deep learning, as Figure 1 shown, the method includes:

[0043] S1. Obtain a data set of the correspondence between the temperature and resistance of the low-temperature sensor.

[0044] First, measure the cryogenic sensor to obtain a data set in which temperature and resistance correspond one by one. In the data set, temperature and resistance are roughly inversely proportional. In some specific embodiments, the calibration point range of a commercial cryogenic sensor is generally 14K to 273.16K. About 48 calibration point temperatures and their corresponding resistances can be measured within this range. Of course, the calibration point range and the number of calibration points can also be other values, which can be determined according to specific circumstances and actual situations. To ensure the accuracy of data fitting, the measurement system should maintain temperature stability during measurement.

[0045] S2. Preprocess the data set using Chebyshev polynomials and the normalization method to obtain a processed data set.

[0046] In some embodiments, preprocessing the data set using Chebyshev polynomials and the normalization method includes:

[0047] S21. After obtaining a data set in which temperature and resistance correspond one by one, it is necessary to construct features of the data set. This process can be achieved by adding some non-linear features of the input data. Polynomial features are usually used to obtain higher dimensions of features and terms of mutual relationships. The input of the deep learning model is resistance and the output is temperature. Chebyshev polynomials are used to add feature terms to the resistance to accelerate the convergence speed. The Chebyshev polynomial is:

[0048]

[0049] where T is the absolute temperature, R is the resistance, x is a variable whose absolute value is not greater than 1, A and B are normalization constants, n is the fitting order, and a i is the fitting coefficient.

[0050] To simplify the complexity of the method, in some preferred embodiments, the fitting order of the Chebyshev polynomial can be set to half of the number of calibration points. Then the input resistance R is expanded to: a1 cos[icos -1 (AlnR + B)]…a n cos[icos -1 (AlnR + B)], a total of n + 1 terms.

[0051] S22. Normalize the input resistance, the generated feature terms, and the output temperature respectively to obtain a processed data set.

[0052] In some preferred embodiments, the maximum-minimum normalization method is selected for normalization processing:

[0053]

[0054] Among them, x' is the normalized sample data, x is the sample data before normalization, max(x) is the maximum value of the sample data, and min(x) is the minimum value of the sample data. Through normalization processing, the convergence speed can be improved, and the accuracy can also be improved.

[0055] S3. Construct a deep learning model. After adjusting the hyperparameters, segment the processed data set into a low-temperature region and a high-temperature region. There is an overlapping interval between the low-temperature region and the high-temperature region, and then train the processed data set. This step further includes:

[0056] S31. Build a deep learning model, add an LSTM layer and a fully connected layer, and add the L1 regularization method.

[0057] In some preferred embodiments, in building a deep learning model, adding an LSTM layer and a fully connected layer, and adding the L1 regularization method, the objective function is:

[0058]

[0059] Where L is the objective function, n is the number of samples, T is the true value of the calibration point temperature, t is the predicted value of the calibration point temperature, k is the number of feature terms, and w is the weight; the objective function is minimized through the deep learning model.

[0060] S32. Divide the processed data set into a training set and a validation set, then send the training set into the deep learning model for training, and use the validation set to adjust the hyperparameters related to the deep learning model;

[0061] Dividing the processed data set into a training set and a validation set is only used for hyperparameter tuning. Subsequent predictions still use the complete processed data set. Generally, a calibration point near the middle of the temperature range is selected as the validation set, that is, a calibration point near the middle of the processed data set is selected. This calibration point includes the temperature point and its corresponding resistance value; then the training set is sent into the deep learning model for training. The training set includes temperature and resistance data, and the validation set is used to adjust the hyperparameters related to the deep learning model. The hyperparameters can include the regularization coefficient to reduce the complexity of the method.

[0062] S33. Segment the processed data set into a low-temperature region and a high-temperature region. There is an overlapping interval between the low-temperature region and the high-temperature region, and determine the segmentation boundary points through the complete processed data set.

[0063] Since using the same set of Chebyshev polynomial coefficients for measuring the entire temperature range does not yield good results, it is necessary to segment the complete processed data set into a low-temperature region and a high-temperature region. It can be understood that low temperature and high temperature are relative to the entire data set and there is no absolute standard. At the same time, the two segments have an overlapping interval to establish a smooth connection, and then the segmentation boundary points are determined through the complete processed data set. In some embodiments, determining the segmentation boundary points through the complete processed data set further includes:

[0064] S331: If the temperature of the processed data set decreases from left to right, the leftmost end of the processed data set is used as the left boundary point of the low-temperature region, and the right boundary point is traversed at the calibration point in the middle. It should be noted that the calibration point in the middle refers to an interval in the middle of the processed data set. After training the model, calculate the mean square error of the entire low-temperature region, and select the right endpoint with the lowest mean square error as the right boundary point of the low-temperature region.

[0065] The mean square error is the expectation of the square of the difference between the true value and the predicted value. The calculation formula for the mean square error is:

[0066]

[0067] In the above formula, y m represents the true value, represents the predicted value, and m represents the number of calibration points.

[0068] Specifically, the segmented data is sent into the deep learning model for training, and then the trained model is used for prediction to obtain a series of predicted temperature values. Substitute these values into the above formula to obtain the mean square error, which will vary with the selection of the right boundary point. Select the right boundary point with the minimum mean square error as the right boundary point of the low-temperature region.

[0069] To simplify the complexity of the method, in some preferred embodiments, after segmenting the processed data set into a low-temperature region and a high-temperature region, the fitting orders of the Chebyshev polynomials are respectively set to half of the number of calibration points in the low-temperature region and the high-temperature region.

[0070] It can be understood that considering the differences in low-temperature sensors in different batches, in some other embodiments, the segmentation method and fitting order of each low-temperature sensor can be adjusted arbitrarily and appropriately.

[0071] S332: After determining the first segment, move 2 - 3 points to the left from the right boundary point of the low-temperature region as the left boundary point of the high-temperature region, and use the rightmost end of the processed data set as the right boundary point of the high-temperature region.

[0072] S4. Predict the temperature value according to the training result.

[0073] To improve the accuracy of temperature prediction, in some embodiments, predicting the temperature value according to the training result includes:

[0074] S41. For the low-temperature region and the high-temperature region, train multiple models. Each model performs forward inference to obtain predicted values. After removing the maximum value and the minimum value, take the average as the final result to obtain the low-temperature region temperature value and the high-temperature region temperature value.

[0075] Specifically, for non-overlapping regions, select the hyperparameters and the model with the best segmentation effect for training. 6 to 8 models can be trained. Use these models to predict the temperature value, that is, first train the models well, and then input the resistance to obtain the predicted temperature value. The resistance can be input multiple times to obtain multiple predicted temperature values. To exclude singular values, remove one maximum value and one minimum value, and take the average of the remaining predicted values as the final result.

[0076] S42. For the overlapping interval, select the midpoint of the overlapping interval as the demarcation point between the low-temperature region and the high-temperature region. Use the low-temperature region temperature value as the predicted value for the data within the overlapping interval to the left of the demarcation point, and use the high-temperature region temperature value as the predicted value for the data within the overlapping interval to the right of the demarcation point.

[0077] For the overlapping region, the original method was to directly take the average of the results of the two segments. However, since the closer to the segment boundary, the lower the accuracy of the predicted value, only take one segment as the final result to improve the accuracy of the predicted value.

[0078] In the above embodiments of the present invention, through the method of deep learning and in combination with the Chebyshev polynomial, the accuracy of temperature sensor prediction can be improved. Moreover, by taking the demarcation point for the overlapping interval between the low-temperature region and the high-temperature region, and taking the predicted values of the segments with smaller deviations on both the left and right sides as the results, the temperature prediction accuracy of the overlapping interval can be further improved. Additionally, by adding a regularization method and taking half of the number of calibration points as the fitting degree, the complexity of the method can be simplified.

[0079] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be combined arbitrarily without conflict.

Claims

1. A high-precision temperature prediction method for a cryogenic sensor based on deep learning, characterized in that, Including: Obtaining a data set of the temperature-resistance correspondence relationship of a low-temperature sensor; Preprocessing the data set by using Chebyshev polynomials and the normalization method, including: The input of the deep learning model is resistance and the output is temperature. Feature terms are added to the resistance by using Chebyshev polynomials to accelerate the convergence speed. The Chebyshev polynomial is: where T is the absolute temperature, R is the resistance, x is a variable with an absolute value not greater than 1, A and B are normalization constants, n is the fitting order, and a i is the fitting coefficient; Performing normalization processing on the input resistance, the generated feature terms, and the output temperature respectively to obtain a processed data set; Constructing a deep learning model. After adjusting the hyperparameters, segmenting the processed data set into a low-temperature region and a high-temperature region. The low-temperature region and the high-temperature region have an overlapping interval, and training the processed data set; including: Building a deep learning model, adding an LSTM layer and a fully connected layer, and adding an L1 regularization method; where the objective function is: Where L is the objective function, n is the number of samples, T is the true value of the calibration point temperature, t is the predicted value of the calibration point temperature, k is the number of feature terms, and w is the weight; the objective function is minimized through the deep learning model; Dividing the processed data set into a training set and a validation set, then sending the training set into the deep learning model for training, and using the validation set to adjust the hyperparameters related to the deep learning model; Segmenting the processed data set into a low-temperature region and a high-temperature region. The low-temperature region and the high-temperature region have an overlapping interval, and determining the segmentation boundary points through the complete processed data set; including: The temperature of the processed data set decreases from left to right. Taking the leftmost end of the processed data set as the left boundary point of the low-temperature region, traversing the right boundary point at the calibration point in the middle. After training the model, calculating the mean square error of the entire low-temperature region, and selecting the right endpoint with the lowest mean square error as the right boundary point of the low-temperature region; Shifting 2-3 points to the left from the right boundary point of the low-temperature region as the left boundary point of the high-temperature region, and taking the rightmost end of the processed data set as the right boundary point of the high-temperature region; Predicting the temperature value according to the training result.

2. The high-precision temperature prediction method of the cryogenic sensor based on deep learning according to claim 1, wherein The fitting degree of the Chebyshev polynomial is set to half of the number of calibration points.

3. The high-precision temperature prediction method for a cryogenic sensor based on deep learning according to claim 1, characterized in that, The normalization processing selects the maximum-minimum normalization method: Where x' is the normalized sample data, x is the sample data before normalization, max(x) is the maximum value of the sample data, and min(x) is the minimum value of the sample data.

4. The high-precision temperature prediction method of the cryogenic sensor based on deep learning according to claim 1, wherein The hyperparameters include the regularization coefficient.

5. The high-precision temperature prediction method for a cryogenic sensor based on deep learning according to claim 1, characterized in that After segmenting the processed data set into a low-temperature region and a high-temperature region, the fitting degrees of the Chebyshev polynomials are respectively set to half of the number of calibration points in the low-temperature region and the high-temperature region.

6. The high-precision temperature prediction method for a cryogenic sensor based on deep learning according to claim 1, wherein The predicting the temperature value according to the training result includes: For the low-temperature region and the high-temperature region, training multiple models. Each model performs forward inference to obtain a predicted value. After removing the maximum value and the minimum value, taking the average value as the final result to obtain the low-temperature region temperature value and the high-temperature region temperature value; For the overlapping interval, select the midpoint of the overlapping interval as the demarcation point between the low-temperature region and the high-temperature region. For the data within the overlapping interval to the left of the demarcation point, use the temperature value of the low-temperature region as the predicted value, and for the data within the overlapping interval to the right of the demarcation point, use the temperature value of the high-temperature region as the predicted value.

Images