Temperature drift correction method based on passive Fourier transform infrared spectrometer

Abstract

The present invention relates to the field of spectral analysis technology, and in particular to a temperature drift correction method based on a passive Fourier infrared spectrometer, the method comprising: obtaining original interference pattern data and temperature sensor data of the passive Fourier infrared spectrometer; screening features to construct a training sample set, and training a convolutional neural network to obtain a spectral temperature drift correction model for a specific ambient temperature. Taking a specific model as a benchmark, adjacent temperature combinations are input into a benchmark network to calculate loss values, and whether they are applicable is judged to obtain a spectral temperature drift correction model for a wider temperature range. Temperature and spectral data are input in real time, and the corrected light intensity-wavelength is output to achieve temperature drift correction. The present invention reduces the amount of calculation and sample data for model training through a predictive network fusion strategy and a dynamic time warping algorithm, improves the model's adaptability and generalization ability to data under different temperature conditions, and achieves accurate temperature drift correction.

Description

Technical Field

[0001] The present invention relates to the field of spectrum analysis technology, and in particular to a temperature drift correction method based on a passive Fourier transform infrared spectrometer. Background Art

[0002] A passive Fourier transform infrared spectrometer is a spectral analysis instrument based on the Fourier transform principle. It measures and analyzes infrared radiation naturally emitted by a target object. This instrument is used in a variety of fields, including environmental monitoring, industrial safety, and scientific research. However, in actual use, especially when operating in open environments, passive Fourier transform infrared spectrometers are susceptible to temperature fluctuations, resulting in temperature drift. Temperature drift refers to the phenomenon in which the instrument's measurement results deviate due to temperature changes, which can seriously affect the accuracy and reliability of spectral analysis.

[0003] When operating in an open environment, passive Fourier transform infrared spectrometers are susceptible to temperature fluctuations, resulting in temperature drift. Temperature variations can cause optical path alignment deviations, reduce the precision of the interferometer's moving mirrors, and alter detector sensitivity, leading to nonlinear shifts in the spectral curve, affecting the accuracy of the analysis results. Furthermore, heat generated during instrument operation can alter the local temperature field, further exacerbating optical system instabilities. These combined factors can cause nonlinear intensity shifts in the spectral curve at varying temperatures, severely impacting the accuracy and reliability of the analysis results.

[0004] Currently, when passive Fourier transform infrared spectrometers operate under different environmental conditions, the ambient temperature and instrument operating temperature constantly fluctuate. These temperature fluctuations affect the instrument's optical components and detectors, leading to varying degrees of spectral drift. However, existing correction methods have limitations, requiring the establishment of separate predictive neural networks for each temperature combination. This not only results in excessive computational effort but also increases sample requirements. Furthermore, predictive neural networks with a fixed temperature range lack flexibility, resulting in large errors in temperature drift correction, making it difficult to meet the requirements for precise correction. Summary of the Invention

[0005] To address the problem that temperature changes can cause spectral curve deviation in a passive Fourier transform infrared spectrometer when operating under different environmental conditions, the existing correction method requires separate modeling for each temperature combination, resulting in large computational complexity, large sample requirements, poor flexibility, and difficulty in achieving accurate correction. The present invention provides solutions in the following aspects.

[0006] The temperature drift correction method based on a passive Fourier infrared spectrometer comprises the following steps: obtaining original interference pattern data and temperature sensor data of the passive Fourier infrared spectrometer, wherein the original interference pattern data comprises wavelength and light intensity, and the temperature sensor data comprises ambient temperature and instrument operating temperature; screening and feature extraction are performed on the obtained original interference pattern data and temperature sensor data, and a training sample set is constructed, and a preset convolutional neural network is trained using the training sample set to obtain a spectral temperature drift correction model for a specific ambient temperature after screening; the spectral temperature drift correction model for a specific ambient temperature is selected as a reference network, and a corresponding spectral temperature drift correction model is obtained. Adjacent temperature combinations are used as temperature combinations to be measured, and each temperature combination to be measured corresponds to a light intensity-wavelength feature set. The light intensity-wavelength feature set at a moment is put into the reference network, and the loss value of the light intensity-wavelength feature set corresponding to the temperature combination to be measured in the reference network is calculated. According to the loss value, it is judged whether the reference network can be used, and a spectral temperature drift correction model for a range of ambient temperatures and instrument operating temperatures is obtained; the real-time ambient temperature and instrument operating temperature belong to the spectral temperature drift correction model corresponding to the range, and the wavelength, light intensity, ambient temperature and instrument operating temperature are input, and the light intensity-wavelength at the next moment is output as the result of temperature drift correction.

[0007] By obtaining the raw interference pattern data and temperature sensor data of the passive Fourier infrared spectrometer, screening and feature extraction are performed, a training sample set is constructed, and these samples are used to train a preset convolutional neural network to obtain a spectral temperature drift correction model under a specific ambient temperature. Furthermore, the correction model for a specific ambient temperature is selected as the baseline network, and the adjacent temperature combinations are evaluated. The loss value is calculated to determine whether the baseline network can be used, thereby obtaining a correction model applicable to a wider temperature range. Finally, the real-time ambient temperature and instrument operating temperature are input into the corresponding correction model, and the light intensity-wavelength mapping at the next moment is output to achieve accurate temperature drift correction. This effectively reduces the amount of calculation and sample requirements, improves the flexibility and generalization ability of the model, and significantly enhances the accuracy and adaptability of temperature drift correction.

[0008] Preferably, the constructing of the training sample set includes:

[0009] The median of all light intensities at each wavelength is taken as the characteristic light intensity-wavelength. According to the similarity measurement method, the discrete distance between each light intensity-wavelength mapping and the characteristic light intensity-wavelength is calculated, and the discrete distance is binary classified. The light intensity-wavelength mapping with the smaller discrete distance after binary classification is retained, and the filtered light intensity-wavelength feature set and the corresponding temperature sensor data are constructed as the training sample set.

[0010] By effectively removing noise and abnormal data, the quality of training samples is improved, making the trained convolutional neural network model have higher accuracy and generalization ability, thereby improving the performance and reliability of spectral temperature drift correction.

[0011] Preferably, the discrete distance includes:

[0012] The dynamic time warping algorithm is used to calculate the similarity between each intensity-wavelength mapping in the intensity-wavelength mapping set and the characteristic intensity-wavelength mapping. The similarity is attenuated using a negative exponential function to obtain the discrete distance between each intensity-wavelength mapping and the characteristic intensity-wavelength.

[0013] The Dynamic Time Warping (DTW) algorithm calculates the similarity between each intensity-wavelength mapping and the characteristic intensity-wavelength mapping, effectively handling nonlinear variations and wavelength shifts in spectral data. A negative exponential function is used to attenuate the similarity and further convert it into discrete distances, ensuring that similar mappings have smaller discrete distances and dissimilar mappings have larger discrete distances. This not only improves the accuracy of feature extraction but also enhances the model's adaptability and robustness to data under different temperature conditions, thereby enhancing the precision and reliability of spectral temperature drift correction.

[0014] Preferably, the step of training a preset convolutional neural network includes:

[0015] The filtered light intensity-wavelength feature set and the corresponding temperature sensor data are used as the training set. The combination of the light intensity-wavelength mapping and temperature sensor data at the previous moment in the training set is used as the input feature of the convolutional neural network, and the light intensity-wavelength mapping at the next moment is used as the output result.

[0016] The training sample set is input into the preset convolutional neural network, and the preset loss function is used to quantitatively evaluate the difference between the predicted output of the preset convolutional neural network and the actual target output. The network parameters are iteratively updated according to the optimization algorithm. When the loss value is less than the preset threshold or reaches the preset number of training times, the training is completed and the spectral temperature drift correction model is obtained.

[0017] Preferably, the network structure of the preset convolutional neural network is as follows: the input layer receives the light intensity-wavelength mapping and temperature sensor data, extracts features through multiple convolutional layers and pooling layers, flattens the feature map and inputs it into the fully connected layer for global feature combination, and predicts the light intensity-wavelength mapping at the next moment based on the output of the output layer.

[0018] Through the convolutional neural network structure, it is not only possible to capture complex patterns in spectral data, but also to adapt to changes in different ambient temperatures and instrument operating temperatures, thereby improving the generalization ability of the model and the accuracy of temperature drift correction.

[0019] Preferably, the loss value is calculated by:

[0020] Taking any moment as the target moment, calculate the absolute value of the difference between the intensity-wavelength mapping at the target moment and the intensity-wavelength mapping at the moment after the target moment, and obtain the degree of change of the intensity-wavelength mapping between consecutive moments. Divide the value of 1 minus the degree of change by the number of wavelength data in the intensity-wavelength mapping as the normalized similarity weight;

[0021] Each target moment corresponds to a light intensity-wavelength mapping, where a mapping contains light intensities at multiple wavelengths. The total light intensity difference is calculated as the average of the sum of the absolute differences between the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the target moment and the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the moment after the target moment.

[0022] The product of the normalized similarity weight and the total light intensity difference is taken as the weighted light intensity difference, and the sum of the weighted light intensity differences at all moments in the training sample is calculated as the neural network loss value.

[0023] Preferably, the adjacent temperature combination includes:

[0024] The specific ambient temperature and instrument operating temperature are regarded as a temperature combination, and the spectral temperature drift correction model of any specific ambient temperature is used as the reference network. All other temperature combinations adjacent to the temperature combination corresponding to the reference network are identified. When the sum of the absolute values ​​of the differences between two groups of temperature combinations is less than a preset threshold, they are considered to be adjacent temperature combinations.

[0025] Preferably, judging whether the reference network can be used according to the loss value includes:

[0026] In response to the loss value being greater than the preset loss threshold, the temperature combination to be measured can use the reference network. Otherwise, if it is less than or equal to the preset loss threshold, the temperature combination to be measured cannot use the reference network and needs to obtain another reference network.

[0027] The present invention has the following effects:

[0028] 1. This invention uses a prediction network fusion strategy to merge sample data from adjacent temperatures and retrain the neural network, resulting in a prediction neural network applicable to a wider temperature range. This effectively reduces the computational effort and sample data required to train a separate model for each temperature combination, improving model training efficiency.

[0029] 2. The present invention uses a dynamic time warping algorithm for feature extraction and screening. The model can handle nonlinear changes and wavelength offsets, improves the adaptability to data under different temperature conditions, and enhances the generalization ability, so that it can perform accurate temperature drift correction under various environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a method flow chart of steps S1 to S4 in a temperature drift correction method based on a passive Fourier transform infrared spectrometer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0032] Reference Figure 1 The temperature drift correction method based on the passive Fourier transform infrared spectrometer includes steps S1 to S4, which are specifically as follows:

[0033] It should be noted that passive FTIR spectrometers collect data in an open environment, making them susceptible to operating in non-standard temperature environments. In such cases, temperature changes can cause the measured light intensity to deviate from the actual value, a phenomenon known as temperature drift.

[0034] The optical components (such as the beam splitter, moving mirror, and fixed mirror) and electronic components within the instrument are extremely sensitive to temperature fluctuations. Specifically, temperature changes can cause thermal expansion or contraction of materials, leading to optical path alignment deviations or reduced interferometer mirror movement accuracy. Insufficient liquid nitrogen cooling or ambient temperature fluctuations can alter the vacuum level or response characteristics, causing signal distortion. Heat accumulation generated during instrument operation can alter the local temperature field, further exacerbating optical system instability. These combined factors can cause nonlinear intensity shifts in the spectral curve at different temperatures, affecting the accuracy of analytical results.

[0035] To correct for this temperature drift, a predictive neural network with a strong correlation between ambient temperature and instrument operating temperature is required. This network can predict and correct light intensity measurements based on current temperature conditions, thereby improving the accuracy of analysis results. The specific implementation process is as follows:

[0036] S1: Obtain raw interferogram data and temperature sensor data of a passive Fourier transform infrared spectrometer, wherein the raw interferogram data includes wavelength and light intensity, and the temperature sensor data includes ambient temperature and instrument operating temperature.

[0037] Wavelength: During the operation of a passive Fourier transform infrared spectrometer, when infrared light emitted by the object being measured passes through the instrument's optical system (such as an interferometer), an interference pattern is formed. Wavelength is a fundamental physical property of light, reflecting its frequency and energy. In infrared spectroscopy, different substances exhibit specific absorption or emission characteristics at different wavelengths. For example, in the mid-infrared band (typically the wavelength range of 2.5-25 microns), many organic compounds exhibit characteristic absorption peaks. The location (i.e., wavelength) and intensity of these absorption peaks can be used to identify the substance's type and concentration.

[0038] Light intensity: Light intensity is a physical quantity that describes the intensity of light. In the original interference pattern, light intensity corresponds to the strength of light at different wavelengths. It reflects the ability of the object being measured to emit or reflect infrared light at corresponding wavelengths. For example, a high-temperature object will have relatively high light intensity in certain infrared bands because high-temperature objects emit more intense thermal radiation. In spectral analysis, the magnitude and variation of light intensity are crucial for quantitative analysis of parameters such as the content of a substance.

[0039] Ambient temperature: This is the temperature of the spectrometer's environment. Changes in ambient temperature may affect the spectrometer's optical and electronic components. For example, temperature fluctuations can cause optical components to expand and contract, affecting the stability of the optical path. For electronic components, temperature changes can affect their performance, such as altering detector sensitivity. Recording ambient temperature helps to compensate for temperature in spectral measurements, improving measurement accuracy and reliability.

[0040] Instrument Operating Temperature: This is the actual operating temperature of key components within the spectrometer, such as the detector and interferometer. The stability of the instrument's operating temperature is crucial for spectral measurements. For example, detector characteristics like dark current and noise vary with temperature. If the instrument's operating temperature is unstable, errors in measured data such as light intensity can occur. By recording the instrument's operating temperature, these temperature-related effects can be corrected, ensuring stable and accurate spectrometer operation under varying temperature conditions.

[0041] S2: The acquired raw interferogram data and temperature sensor data are screened and feature extracted, and a training sample set is constructed. The training sample set is used to train a preset convolutional neural network to obtain a spectral temperature drift correction model for the specific ambient temperature after screening.

[0042] The median of all light intensities at each wavelength is taken as the characteristic light intensity-wavelength. According to the similarity measurement method, the discrete distance between each light intensity-wavelength mapping and the characteristic light intensity-wavelength is calculated, and the discrete distance is binary classified. The light intensity-wavelength mapping with the smaller discrete distance after binary classification is retained, and the filtered light intensity-wavelength feature set and the corresponding temperature sensor data are constructed as the training sample set.

[0043] Discrete distances include:

[0044] The dynamic time warping algorithm is used to calculate the similarity between each intensity-wavelength mapping in the intensity-wavelength mapping set and the characteristic intensity-wavelength mapping. The similarity is attenuated using a negative exponential function to obtain the discrete distance between each intensity-wavelength mapping and the characteristic intensity-wavelength.

[0045] Specifically, the discrete distance satisfies the following relationship:

[0046] ;

[0047] Where, Represents the discrete distance from the intensity-wavelength mapping to the characteristic intensity-wavelength, represents the characteristic light intensity-wavelength mapping, Indicates the first Intensity-wavelength mapping, represents the time warping function, Represented by natural numbers An exponential function with base .

[0048] Alternatively, the Pearson correlation coefficient can be used to calculate the similarity between each intensity-wavelength mapping in the set of intensity-wavelength mappings and the characteristic intensity-wavelength mapping. However, it should be noted that DTW is computationally complex, especially for large datasets, and can take a long time to compute, but it offers higher accuracy. In contrast, the Pearson correlation coefficient is a simple and efficient statistical method with fast computation speed, suitable for large datasets. However, it only measures linear correlation and has limited ability to detect nonlinear relationships. If there are nonlinear shifts between the intensity-wavelength mappings, the Pearson correlation coefficient may not accurately reflect the similarity between them.

[0049] Pearson correlation or dynamic time warping can be used independently for screening. Alternatively, the Pearson correlation can be used for preliminary screening to quickly exclude data with low correlation to the characteristic intensity-wavelength mapping. After preliminary screening, dynamic time warping is used to perform precise similarity calculations and corrections to account for nonlinear variations and wavelength shifts.

[0050] The steps for training a pre-set convolutional neural network include:

[0051] The filtered light intensity-wavelength feature set and the corresponding temperature sensor data are used as the training set. The combination of the light intensity-wavelength mapping and temperature sensor data at the previous moment in the training set is used as the input feature of the convolutional neural network, and the light intensity-wavelength mapping at the next moment is used as the output result.

[0052] The training sample set is input into the preset convolutional neural network, and the preset loss function is used to quantitatively evaluate the difference between the predicted output of the preset convolutional neural network and the actual target output. The network parameters are iteratively updated according to the optimization algorithm. When the loss value is less than the preset threshold or reaches the preset number of training times, the training is completed and the spectral temperature drift correction model is obtained.

[0053] The network structure of the preset convolutional neural network is as follows: the input layer receives the light intensity-wavelength mapping and temperature sensor data, extracts features through multiple convolutional layers and pooling layers, flattens the feature map and inputs it into the fully connected layer for global feature combination, and predicts the light intensity-wavelength mapping at the next moment based on the output of the output layer.

[0054] The loss value is calculated by:

[0055] Taking any moment as the target moment, calculate the absolute value of the difference between the intensity-wavelength mapping at the target moment and the intensity-wavelength mapping at the moment after the target moment, and obtain the degree of change of the intensity-wavelength mapping between consecutive moments. Divide the value of 1 minus the degree of change by the number of wavelength data in the intensity-wavelength mapping as the normalized similarity weight;

[0056] Each target moment corresponds to a light intensity-wavelength mapping, where a mapping contains light intensities at multiple wavelengths. The total light intensity difference is calculated as the average of the sum of the absolute differences between the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the target moment and the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the moment after the target moment.

[0057] The product of the normalized similarity weight and the total light intensity difference is taken as the weighted light intensity difference, and the sum of the weighted light intensity differences at all moments in the training sample is calculated as the neural network loss value.

[0058] Specifically, the loss value satisfies the following relationship:

[0059] ;

[0060] Where, represents the neural network loss value, represents the total number of moments in the training sample, Indicates the Time intensity-wavelength mapping, Indicates the Time intensity-wavelength mapping, Indicates the number of wavelength data in the intensity-wavelength mapping. Indicates the input The intensity-wavelength mapping at time The light intensity corresponding to the wavelength is Indicates the predicted The moment intensity-wavelength mapping The corresponding light intensity.

[0061] In other words, in spectral analysis, the intensity-wavelength mapping is the set of light intensities at different wavelengths at a specific moment. This mapping can be expressed as a function or a data set, where each wavelength corresponds to a specific light intensity value.

[0062] It is further explained that, given that different ambient temperatures and instrument operating temperatures each correspond to a prediction neural network, this not only results in a huge amount of computation and sample requirements, but also makes the prediction neural network within a fixed temperature range lack flexibility, which in turn leads to an increase in the temperature drift correction error. Therefore, it is necessary to implement a prediction network fusion strategy for samples of adjacent temperatures. By merging the sample data of these adjacent temperatures and retraining the neural network on the merged data set, a prediction neural network suitable for a wider temperature range can be obtained. This fusion process effectively reduces the consumption of computing resources and the demand for sample data, while enhancing the generalization ability of the model and significantly improving the accuracy and adaptability of temperature drift correction. The specific steps are as follows:

[0063] S3: Select the spectral temperature drift correction model of a specific ambient temperature as the reference network, obtain adjacent temperature combinations as the temperature combinations to be measured, each temperature combination to be measured corresponds to a light intensity-wavelength feature set, put the light intensity-wavelength feature set at a moment into the reference network, calculate the loss value of the light intensity-wavelength feature set corresponding to the temperature combination to be measured in the reference network, determine whether the reference network can be used based on the loss value, and obtain the spectral temperature drift correction model for a range of ambient temperatures and instrument operating temperatures.

[0064] Adjacent temperature combinations, including:

[0065] The specific ambient temperature and instrument operating temperature are regarded as a temperature combination, and the spectral temperature drift correction model of any specific ambient temperature is used as the reference network. All other temperature combinations adjacent to the temperature combination corresponding to the reference network are identified. When the sum of the absolute values ​​of the differences between two groups of temperature combinations is less than a preset threshold, they are considered to be adjacent temperature combinations.

[0066] In this embodiment, the preset threshold is 5, which can be adjusted according to specific circumstances.

[0067] For example, if the baseline network is trained for an ambient temperature of 25°C and an instrument operating temperature of 30°C, then the combination of an ambient temperature of 26°C and an instrument operating temperature of 31°C is considered close because the sum of the absolute values ​​of their differences is 2, which is less than 5. For each such close temperature combination, we check whether it can be accurately predicted using the baseline network. If so, we save the time and resources of training a separate model for each temperature combination.

[0068] In response to the loss value being greater than the preset loss threshold, the temperature combination to be measured can use the reference network. Otherwise, if the loss value is less than or equal to the preset loss threshold, the temperature combination to be measured cannot use the reference network and needs to obtain another reference network.

[0069] Exemplarily, the preset loss threshold is 0.01.

[0070] S4: The real-time ambient temperature and instrument operating temperature belong to the spectral temperature drift correction model corresponding to the range, input wavelength, light intensity, ambient temperature and instrument operating temperature, and output the light intensity-wavelength at the next moment as the result of temperature drift correction.

[0071] It should be noted that those skilled in the art may make various modifications and improvements without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be based on the appended claims.

Claims

1. A temperature drift correction method based on a passive Fourier transform infrared spectrometer, characterized in that: include: Obtaining raw interference pattern data and temperature sensor data of a passive Fourier transform infrared spectrometer, wherein the raw interference pattern data includes wavelength and light intensity, and the temperature sensor data includes ambient temperature and instrument operating temperature; The acquired raw interferogram data and temperature sensor data are screened and feature extracted, and a training sample set is constructed. The training sample set is used to train a preset convolutional neural network to obtain a spectral temperature drift correction model for a specific ambient temperature after screening. A specific ambient temperature and instrument operating temperature are considered as a temperature combination. The spectral temperature drift correction model of any specific ambient temperature is used as a reference network. All other temperature combinations adjacent to the temperature combination corresponding to the reference network are identified. When the sum of the absolute values ​​of the differences between two groups of temperature combinations is less than a preset threshold, they are considered to be adjacent temperature combinations. Each temperature combination to be measured corresponds to a light intensity-wavelength feature set. The light intensity-wavelength feature set at a moment is put into the reference network. The loss value of the light intensity-wavelength feature set corresponding to the temperature combination to be measured in the reference network is calculated. Based on the loss value, it is determined whether the reference network can be used, and the spectral temperature drift correction model for a range of ambient temperatures and instrument operating temperatures is obtained. The real-time ambient temperature and instrument operating temperature belong to the spectral temperature drift correction model corresponding to the range. The wavelength, light intensity, ambient temperature and instrument operating temperature are input, and the light intensity-wavelength at the next moment is output as the result of temperature drift correction.

2. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 1, wherein The constructing of the training sample set includes: The median of all light intensities at each wavelength is taken as the characteristic light intensity-wavelength. According to the similarity measurement method, the discrete distance between each light intensity-wavelength mapping and the characteristic light intensity-wavelength is calculated, and the discrete distance is binary classified. The light intensity-wavelength mapping with the smaller discrete distance after binary classification is retained, and the filtered light intensity-wavelength feature set and the corresponding temperature sensor data are constructed as the training sample set.

3. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 2, wherein: The discrete distances include: The dynamic time warping algorithm is used to calculate the similarity between each intensity-wavelength mapping in the intensity-wavelength mapping set and the characteristic intensity-wavelength mapping. The similarity is attenuated using a negative exponential function to obtain the discrete distance between each intensity-wavelength mapping and the characteristic intensity-wavelength.

4. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 1, wherein: The step of training the preset convolutional neural network includes: The filtered light intensity-wavelength feature set and the corresponding temperature sensor data are used as the training set. The combination of the light intensity-wavelength mapping and temperature sensor data at the previous moment in the training set is used as the input feature of the convolutional neural network, and the light intensity-wavelength mapping at the next moment is used as the output result. The training sample set is input into the preset convolutional neural network, and the preset loss function is used to quantitatively evaluate the difference between the predicted output of the preset convolutional neural network and the actual target output. The network parameters are iteratively updated according to the optimization algorithm. When the loss value is less than the preset threshold or reaches the preset number of training times, the training is completed and the spectral temperature drift correction model is obtained.

5. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 1, wherein: The network structure of the preset convolutional neural network is as follows: the input layer receives the light intensity-wavelength mapping and temperature sensor data, extracts features through multiple convolutional layers and pooling layers, flattens the feature map and inputs it into the fully connected layer for global feature combination, and predicts the light intensity-wavelength mapping at the next moment based on the output of the output layer.

6. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 4, wherein: The loss value is calculated as follows: Taking any moment as the target moment, calculate the absolute value of the difference between the intensity-wavelength mapping at the target moment and the intensity-wavelength mapping at the moment after the target moment, and obtain the degree of change of the intensity-wavelength mapping between consecutive moments. Divide the value of 1 minus the degree of change by the number of wavelength data in the intensity-wavelength mapping as the normalized similarity weight; Each target moment corresponds to a light intensity-wavelength mapping, where a mapping contains light intensities at multiple wavelengths. The total light intensity difference is calculated as the average of the sum of the absolute differences between the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the target moment and the light intensities corresponding to all wavelength data in the light intensity-wavelength mapping at the moment after the target moment. The product of the normalized similarity weight and the total light intensity difference is taken as the weighted light intensity difference, and the sum of the weighted light intensity differences at all moments in the training sample is calculated as the neural network loss value.

7. The temperature drift correction method based on a passive Fourier transform infrared spectrometer according to claim 1, wherein: The determining whether the benchmark network can be used according to the loss value includes: In response to the loss value being greater than the preset loss threshold, the temperature combination to be measured can use the reference network. Otherwise, if it is less than or equal to the preset loss threshold, the temperature combination to be measured cannot use the reference network and needs to obtain another reference network.

Images