Gas detector calibration system

Abstract

The application discloses a gas detector calibration system, and relates to the field of gas detection.The system comprises a measurement module, a feature module, a conversion module and a calibration module.The measurement module measures the concentration of zero point and full range two gases, and the temperature, humidity and air pressure of the environment multiple times, and generates two correction sequences after correction.The feature module extracts two main feature vectors from the environmental coupling information in the two correction sequences, the mean and variance of the two concentrations in the window, and establishes the mapping of the main feature vectors and the concentration sequence using gradient boosting trees, and generates two key feature vectors based on feature split gain screening.The conversion module determines a linear conversion model based on the two key feature vectors using an attention gate network to convert the mean of the concentration in the two correction sequences into two standard concentrations.The calibration module constructs a two-point linear calibration model according to the two standard concentrations and calibrates, thereby achieving high-precision calibration while maintaining a linear relationship.

Description

Technical Field

[0001] The present invention relates to the field of gas detection, and in particular to a gas detector calibration system. Background Art

[0002] In the gas detection industry, especially in underground spaces such as underground parking lots, subway tunnels, and mines where extremely high accuracy is required for harmful gas detection, traditional calibration methods have the following problems:

[0003] 1. Only zero point calibration is performed, and coordination with full scale calibration is neglected, resulting in imbalance or even disorder of the linear relationship of the instrument;

[0004] 2. It is difficult to effectively deal with the cumulative errors of the measuring equipment and the nonlinear effects of ambient temperature, humidity and air pressure on the measurement results, which seriously limits the calibration accuracy of the gas detector. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a gas detector calibration system to achieve coordinated calibration of zero point and full scale by combining measurement error correction of measuring equipment and adaptive standard conversion of measurement data based on ambient temperature, ambient humidity and ambient pressure.

[0006] The technical solutions for achieving the purpose of the present invention are:

[0007] Gas detector calibration system, including measurement module, characteristic module, conversion module and calibration module;

[0008] The measurement modules measure The concentration of sub-zero gas and full-scale gas, as well as the ambient temperature, ambient humidity and ambient pressure, combined with the impact of ambient temperature changes on the measurement of ambient temperature and ambient humidity and altitude Correct the measurement influence of ambient air pressure and generate zero point correction sequence and full-scale correction sequence , is the number of unilateral measurements;

[0009] The characteristic module calculates the zero point correction sequence and full-scale correction sequence Zero-point concentration in windows of different sizes and digits and full-scale concentration The mean and variance of the corresponding environmental coupling information are combined and two main feature vectors are generated through principal component analysis. , using gradient boosting tree to build the main feature vector With the corresponding concentration series The mapping relationship is established and feature screening is performed based on feature splitting gain to generate two key feature vectors ;

[0010] The conversion module converts two key feature vectors Input the corresponding attention gating network to determine the coefficient vector of the corresponding linear transformation model , the zero point correction sequence and full-scale correction sequence The average zero-point concentration and average full-scale concentration Corresponding conversion to standard zero point concentration under standard environment and standard full-scale concentration , standard environment including standard temperature , standard humidity and standard atmospheric pressure ;

[0011] Calibration module based on reference zero concentration , reference full-scale concentration With standard zero point concentration , standard full-scale concentration A two-point linear calibration model is constructed for the measuring range and the gas detector is calibrated.

[0012] Furthermore, the measurement module includes an asynchronous measurement unit, a temperature and humidity correction unit, and an air pressure correction unit;

[0013] Asynchronous measuring cell for zero gas Measure and recycle the zero point to obtain the zero point parameter sequence , for full range gas Full-scale measurement and recovery to obtain full-scale parameter sequence , zero-point parameter sequence and full range parameter sequence It includes the concentration, ambient temperature, ambient humidity and ambient pressure at each measurement. The concentration includes the zero point concentration and the full range concentration. is the total number of measurements;

[0014] The temperature and humidity correction unit takes into account the impact of the accumulated ambient temperature changes on the ambient temperature measurement, based on the pre-fitted temperature correction coefficient With the previous The cumulative temperature The first ambient temperature Corrected to the corresponding corrected ambient temperature , in Corrected ambient temperature With standard temperature The difference between the two is used to introduce the second Time accumulation factor , will ambient humidity Corrected to Corrected ambient humidity , time accumulation factor It is pre-fitted and shows a trend of slow increase, rapid increase and convergence with the increase of temperature accumulation;

[0015] The air pressure correction unit is based on the principle of atmospheric statics, combined with the ambient temperature With altitude The change in ambient air pressure Corrected to the corrected ambient pressure at sea level .

[0016] Furthermore, the feature module includes a feature construction unit and a feature screening unit;

[0017] Feature construction unit sorting and correction sequence The concentration in windows of different sizes and quantiles The mean and variance of , extract the feature vector through principal component analysis With the correction sequence The splicing result of the environmental coupling information is used to generate the main feature vector , corrected sequence Including zero point correction sequence and full-scale correction sequence ,concentration Including zero point concentration and full-scale concentration , feature vector Including zero-point eigenvector and full-range characteristic vector , the main feature vector Including the zero-point main eigenvector and the full range main characteristic vector ;

[0018] Feature screening unit extracts and corrects sequence Concentration in To form a concentration sequence , train the gradient boosting tree to establish the main feature vector With concentration series The mapping is used to evaluate the importance of each feature based on the feature split gain, remove the least important features and record the changes in the training loss of the gradient boosting tree, obtain the reference score of each feature and filter it to generate the key feature vector , concentration series Including zero-point concentration series and full-scale concentration series , key feature vector Including zero-point key feature vector and full range key feature vector .

[0019] Furthermore, the feature construction unit generates a feature vector , including the following steps:

[0020] Set the scale to 、 and 3 windows and placed in the correction sequence The starting side and calculate the concentration in 3 windows respectively The mean and variance of

[0021] Stop moving the fully traversed window and right-shift all incompletely traversed windows, and calculate the concentration in the right-shifted window The mean and variance of until all three windows stop moving;

[0022] Concentrate the same-scale window in different digits The mean and variance of the initial splicing and the secondary splicing according to the scale size are used as feature vectors .

[0023] Furthermore, the sequence The temperature correlation gradient is obtained by dividing the difference between each two adjacent corrected ambient temperatures, corrected ambient humidity, and corrected ambient pressure by the corresponding concentration difference and taking the average value. , humidity-related gradient Pressure-related gradient , where the temperature-dependent gradient Including zero point temperature dependent gradient and full-scale temperature-related gradient , humidity-related gradient Including zero point humidity correlation gradient and full-scale humidity gradient , pressure-related gradient Including zero pressure related gradient Gradient associated with full-scale pressure .

[0024] Furthermore, the feature screening unit generates key feature vectors , including the following steps:

[0025] According to the correction sequence middle The concentration measured Constructing concentration sequences And as the main feature vector The training labels of the gradient boosting tree are the main feature vectors of , predict concentration series and predicted concentration series With concentration series The mean square error of the predicted concentration series Including zero-point predicted concentration series and full-scale predicted concentration series ;

[0026] Based on the node splitting strategy, the gradient boosting tree is trained, and each node is based on the main feature vector The feature with the largest node splitting gain is split. The node splitting gain is the reduction in training loss before and after the node split. The splitting stops when the training loss is less than the minimum value or the tree depth reaches the maximum tree depth.

[0027] Eliminate the main feature vector The feature with the smallest feature split gain is selected and a new round of training is started to calculate the training loss. The difference between the training loss of the new round and the previous round is used as the reference score of the removed feature. The feature split gain is equal to the sum of the split gains of all nodes that select the feature for splitting.

[0028] Construct a key feature vector based on features with a reference score greater than a score threshold .

[0029] Furthermore, the attention gating network includes a zero-point attention gating network and a full-range attention gating network with the same network structure, which includes a gated recurrent layer, an attention layer, and a nonlinear output layer;

[0030] The gated recurrent layer is based on the key feature vector The dimension of determines the total number of loop steps. In each step, the update gate and reset gate are used in synergistic manner to fuse the state of the previous step with the key features of the current step and use the nonlinear mapping of the retained part of the state of the previous step to correct and generate the state of the current step. The state of each step is obtained and a state set is constructed. The state set includes a zero-point state set and a full-scale state set.

[0031] The attention layer uses a self-attention mechanism to re-represent the state of each step in the state set using the weighted sum of the states of other steps. The weight is generated by substituting the correlation between the state of each step and the states of other steps into the activation function, converting the state set into a context vector. The context vector includes a zero-point context vector and a full-range context vector.

[0032] The nonlinear output layer converts the context vector into a coefficient vector through linear modulation and cubic nonlinear activation function. wherein the third non-linear activation function is a Softmax activation function, and the coefficient vector comprises a zero-point coefficient vector and a full-scale coefficient vector .

[0033] Specifically, the linear conversion model converts the average zero-point concentration and the average full-scale concentration in the zero-point correction sequence and the full-scale correction sequence into the standard zero-point concentration and the standard full-scale concentration , i.e., the standard zero-point concentration / standard full-scale concentration is equal to the average zero-point concentration / average full-scale concentration plus the inner product of the standard ambient error vector and the corresponding coefficient vector , the standard ambient error vector comprises a standard temperature difference , a standard humidity difference and a standard pressure difference , and specifically, the difference between the average of the corrected ambient temperature, the corrected ambient humidity and the corrected ambient pressure in the zero-point correction sequence and the full-scale correction sequence and the standard temperature , the standard humidity and the standard pressure .

[0034] Further, the construction of the two-point linear calibration model comprises the following steps:

[0035] subtracting the reference zero-point concentration from the reference full-scale concentration to obtain a reference range, and subtracting the standard zero-point concentration from the standard full-scale concentration to obtain a standard measurement range;

[0036] taking the ratio of the reference range to the standard measurement range as a calibration coefficient , and subtracting the product of the standard full-scale concentration and the calibration coefficient from the reference zero-point concentration to obtain an offset ;

[0037] Therefore, the two-point linear calibration model is specifically that the reference full-scale concentration / reference zero-point concentration is equal to the standard full-scale concentration / Standard zero point concentration and calibration coefficients The product plus the offset .

[0038] Compared with the prior art, the present invention has the following significant advantages:

[0039] 1. Determine the range by using the standard zero-point concentration and the standard full-scale concentration under the standard environment, and build a two-point linear calibration model by comparing it with the range of the reference zero-point concentration and the reference full-scale concentration to perform linear calibration on the gas detector and solve the linear relationship imbalance and disorder problems during calibration;

[0040] 2. Considering the impact of accumulated ambient temperature changes on the measurement of ambient temperature and humidity, the ambient temperature and humidity are corrected. Based on the principle of atmospheric statics, the ambient air pressure measured each time is converted into the corrected ambient air pressure at sea level to reduce the influence of measurement errors of the measuring equipment. The main feature vectors of the zero-point correction sequence and the full-scale correction sequence are extracted and combined with the gradient boosting tree for feature screening to generate the corresponding key feature vectors to improve the conversion efficiency. The corresponding linear conversion model is determined through the attention gating network, and the mean of the zero-point concentration and the full-scale concentration in the zero-point correction sequence and the full-scale correction sequence is converted into the standard zero-point concentration and the standard full-scale concentration under the standard environment, eliminating the calibration effects caused by ambient temperature, humidity and pressure, and improving the calibration accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of the gas detector calibration system;

[0042] Figure 2 generating a feature vector schematic for a feature building unit;

[0043] Figure 3 Generate key feature vector flow chart for feature screening unit;

[0044] Figure 4 This is the attention gating network model diagram. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0046] Example 1

[0047] like Figure 1 As shown, a specific embodiment of the present invention discloses a gas detector calibration system, including a measurement module, a feature module, a conversion module and a calibration module;

[0048] The measurement module measures the The concentration of sub-zero gas and full-scale gas and the corresponding ambient temperature, ambient humidity and ambient pressure are used to obtain the zero-point parameter sequence. and full range parameter sequence , considering the influence of the accumulated changes in ambient temperature on the measurement of ambient temperature and ambient humidity, the zero point parameter sequence and full range parameter sequence The ambient temperature and humidity in the image are corrected, and the ambient pressure measured each time is converted into the corrected ambient pressure at sea level based on the principle of atmospheric statics to generate a zero point correction sequence. and full-scale correction sequence , is the number of measurements;

[0049] Characteristic module corrects sequence at zero point and full-scale correction sequence Move windows of different scales and calculate the zero-point concentration in the window in turn and full-scale concentration The mean and variance of , combined with the zero point correction sequence and full-scale correction sequence The corresponding environmental coupling information is obtained by principal component analysis to generate the corresponding main feature vectors , using gradient boosting tree to build the main feature vector With the corresponding concentration series The mapping relationship is based on the feature splitting gain, and the importance of each feature is recursively evaluated for feature screening, and the corresponding key feature vectors are generated respectively. ;

[0050] The conversion module converts the zero point correction sequence and full-scale correction sequence The corresponding key feature vector Input the corresponding attention gating network respectively, and fit the zero-point coefficient vector of the linear conversion model and the full-range coefficient vector , combined with the linear conversion model to correct the zero point sequence and full-scale correction sequence The average zero-point concentration in and average full-scale concentration Based on standard temperature difference , standard humidity difference , standard air pressure difference Converted to standard zero point concentration under standard environment and standard full-scale concentration , where the standard temperature difference , standard humidity difference Difference from standard atmospheric pressure Refers to the average difference between the corrected ambient temperature, corrected ambient humidity, corrected ambient pressure and the standard environment during measurement. The standard environment includes the standard temperature , standard humidity and standard atmospheric pressure , standard atmospheric pressure The height at sea level ;

[0051] Calibration module based on reference zero concentration , reference full-scale concentration With standard zero concentration , standard full-scale concentration The calibration factor is determined by the measuring range And calculate the offset , construct a two-point linear calibration model to calibrate the gas detector so that the measurement values ​​of the gas detector for zero point gas and full range gas are equal to the reference zero point concentration respectively and reference full-scale concentration .

[0052] Furthermore, the measurement module includes an asynchronous measurement unit, a temperature and humidity correction unit, and an air pressure correction unit;

[0053] The asynchronous measuring cell is fed with zero gas and measured at a fixed frequency. Secondary zero point measurement, obtain zero point parameter sequence And recycle the zero point gas, introduce the full range gas and repeat it at a fixed frequency Full-scale measurement to obtain full-scale parameter sequence And recycle the full-process gas, among which, For the The zero-point parameter vector of the measurement, including the Zero-point concentration of sub-zero point measurement , ambient temperature , ambient humidity and ambient air pressure , since the zero point measurement is performed first, The second measurement is Sub-zero point measurement, For the The full-scale parameter vector of the measurement, including the Full-scale concentration of sub-full-scale measurement , ambient temperature , ambient humidity and ambient air pressure , because it has been Second zero point measurement, The second measurement is essentially the Full-scale measurement, let Indicates the total number of measurements, These measurements are all achieved through the gas detector and its internal capacitive temperature and humidity sensor and piezoresistive air pressure sensor;

[0054] The temperature and humidity correction unit takes into account the impact of the accumulated ambient temperature changes on the ambient temperature measurement, based on the temperature correction coefficient With the previous The cumulative temperature For the first ambient temperature Perform temperature correction to offset the measurement error gradually accumulated during continuous measurement of the temperature sensor, and obtain the Corrected ambient temperature And replace the corresponding ambient temperature , the temperature correction formula is as follows:

[0055] ,

[0056] Among them, the temperature correction coefficient Obtained through preliminary experimental fitting, Corrected ambient temperature With standard temperature Based on the influence of the difference in humidity on the change of Time accumulation factor In order to quantify the synergistic effect of the accumulated changes in ambient temperature during the measurement process, the ambient humidity Replaced by humidity correction Corrected ambient humidity , the humidity correction formula is as follows:

[0057] ,

[0058] in, and Control time accumulation factor The control coefficient of the rate of change is obtained through preliminary experimental fitting and is based on the time accumulation factor It can be seen that in the early stage of the capacitive temperature and humidity sensor, the influence of temperature change accumulation on humidity measurement is small, and the humidity measurement error grows slowly. As the temperature change accumulates further, the humidity measurement error changes nonlinearly, and the humidity measurement error grows rapidly. When the temperature change accumulates to a certain extent, the capacitive temperature and humidity sensor reaches the limit of physical performance change, and the humidity measurement error tends to be flat.

[0059] The air pressure correction unit is based on the principle of atmospheric statics. ambient air pressure Replaced by the corrected ambient air pressure at sea level via the barometric pressure correction To eliminate altitude The impact on air pressure measurement, the air pressure correction formula is as follows:

[0060] ,

[0061] in, is the sea level temperature, specifically 15°C, is the vertical lapse rate of atmospheric temperature, specifically 0.0065℃ / m, altitude It can be obtained by adding a GPS positioning component to the piezoresistive air pressure sensor. is the gas constant, as follows:

[0062] ,

[0063] in, 、 and are the acceleration due to gravity, the molar mass of air and the universal gas constant, respectively.

[0064] Furthermore, the feature module includes a feature construction unit and a feature screening unit;

[0065] Feature building unit for correction sequence , calculate the concentration in windows of different scales and different digits The mean and variance of the eigenvector are constructed by sorting , calculate the corrected sequence The environmental coupling information and the characteristic vector Splicing and removing redundant information through principal component analysis to generate the main feature vector , corrected sequence Including zero point correction sequence and full-scale correction sequence ,concentration Including zero point concentration and full-scale concentration , feature vector Including zero-point eigenvector and full-range characteristic vector , the main feature vector Including the zero-point main eigenvector and the full range main characteristic vector ;

[0066] The feature screening unit extracts the correction sequence according to the measurement order Concentration in To form a concentration sequence , select the gradient boosting tree to establish the main feature vector With concentration series For the gradient boosting tree, the importance of each feature is evaluated based on the feature split gain. The least important features are recursively removed and the training loss changes of the gradient boosting tree are recorded. The reference score of each feature is obtained and the key feature vector is generated by comprehensive screening. , concentration series Including zero-point concentration series and full-scale concentration series , key feature vector Including zero-point key feature vector and full range key feature vector .

[0067] like Figure 2 As shown, further, the feature construction unit generates a feature vector , including the following steps:

[0068] Set the scale to 、 and 3 windows and place them synchronously in the correction sequence The far left side;

[0069] Perform statistical calculations and calculate the zero-point concentration in each of the three windows and full-scale concentration The mean and variance of

[0070] Determine whether there is a fully traversed correction sequence middle The concentration measured Window;

[0071] If it exists, perform a partial move, stop moving the fully traversed window and move the incompletely traversed window one position to the right; if it does not exist, perform a global move, moving all three windows one position to the right;

[0072] After performing a partial or global move, statistical calculations are performed on the right-shifted window until all three windows stop moving;

[0073] According to the order of window movement, the concentration in each window is The mean and variance of the three windows are arranged into corresponding window feature vectors, and the window feature vectors of the three windows are sequentially concatenated into feature vectors. .

[0074] Furthermore, the feature construction unit calculates the correction sequence The environmental coupling information includes the following steps:

[0075] Calculate the correction sequence sequentially The temperature correlation gradient is obtained by taking the average of the ratio of the difference between the corrected ambient temperatures of each two adjacent values ​​to the corresponding concentration difference. , temperature-dependent gradient Including zero point temperature dependent gradient and full-scale temperature-related gradient ;

[0076] Calculate the correction sequence sequentially The humidity correlation gradient is obtained by taking the average of the ratio of the difference between the corrected ambient humidity of each two adjacent values ​​and the difference between the corresponding concentrations. , humidity-related gradient Including zero point humidity correlation gradient and full-scale humidity gradient ;

[0077] Calculate the correction sequence sequentially The pressure correlation gradient is obtained by taking the average of the ratio of the difference between the corrected ambient pressures of each two adjacent values ​​and the corresponding concentration difference. , pressure-related gradient Including zero pressure related gradient Gradient associated with full-scale pressure .

[0078] like Figure 3 As shown, further, the feature screening unit generates a key feature vector , including the following steps:

[0079] Extract correction sequence based on measurement order middle The concentration obtained by the measurement To construct a concentration sequence , the concentration sequence As the main feature vector The training labels of the gradient boosting tree are the main feature vectors and predicted concentration series , the training loss is the predicted concentration sequence With concentration series The mean square error of the predicted concentration series Including zero-point predicted concentration series and full-scale predicted concentration series ;

[0080] Based on the node splitting strategy, the gradient boosting tree is trained. For each node, the main feature vector is traversed. all features in the main feature vector, taking the training loss reduction amount before and after splitting the node as the node split gain, selecting the feature pair with the maximum node split gain to split the node, until the training loss is less than the minimum value or the tree depth of the gradient boosting tree reaches the maximum tree depth;

[0081] taking each feature in the main feature vector as the target feature in turn, searching for the node split with the target feature and summing the corresponding node split gain as the feature split gain of the target feature;

[0082] eliminating the feature with the minimum feature split gain in the main feature vector and starting a new round of training, calculating the training loss of the new round, and taking the difference between the training loss of the new round and the training loss of the last round as the reference score of the eliminated feature;

[0083] until the reference scores of all features in the main feature vector are obtained, retaining the features with reference scores greater than the score threshold to form the key feature vector , the key feature vector simplifies the main feature vector that removes redundant information, and retains the key features that can reflect the concentration sequence .

[0084] As shown in Figure 4 , further, the attention gate network includes a zero-point attention gate network and a full-range attention gate network, which respectively generate a zero-point coefficient vector and a full-range coefficient vector based on the zero-point key feature vector and the full-range key feature vector , have the same network structure, and the training loss adopted during training is the error between the standard zero-point concentration and the standard full-range concentration linearly converted from the model output and the concentration of the same zero-point gas and full-range gas actually measured under the standard environment, and the network structure includes a gate recurrent layer, an attention layer, and a nonlinear output layer;

[0085] The gate recurrent layer performs time series modeling on the key feature vector through the combination of the update gate and the reset gate, determines the total number of cycles according to the dimension of the key feature vector , in each step, the update gate fuses the state of the last step and the key feature of the current step to generate the preliminary state of the current step, the reset gate controls the retention ratio of the state of the last step, and the preliminary state of the current step is corrected by combining the retention ratio to obtain the state of the current step through nonlinear mapping, through the cooperative action of the update gate and the reset gate, the state is updated step by step to strengthen the key feature vector The context association of different key features in the process is used to combine the states of each step to generate a state set, which includes a zero-point state set and a full-scale state set.

[0086] The attention layer uses the self-attention mechanism to re-represent the state of each step in the state set using the weighted sum of the states of other steps. The weight is generated by substituting the correlation between the state of each step and the state of other steps into the activation function, converting the state set into a context vector that carries the importance of the state, enhancing the quality of feature representation, and enabling the attention gate network to more accurately capture the standard temperature difference. , standard humidity difference and standard atmospheric pressure difference Impact on concentration, the context vector includes the zero-point context vector and the full-scale context vector;

[0087] The nonlinear output layer reduces the dimension of the context vector to 3 through linear modulation and further transforms the generated coefficient vector through a 3-fold nonlinear activation function. , where the first two nonlinear activation functions are arbitrarily selected, and the third nonlinear activation function selects the Softmax activation function to normalize the elements in the vector, and the coefficient vector Including zero coefficient vector and the full-range coefficient vector .

[0088] Specifically, the linear transformation model is used to transform the zero-point correction sequence and full-scale correction sequence The average zero-point concentration in and average full-scale concentration Corresponding to the standard zero point concentration converted to the standard environment and standard full-scale concentration , specifically the standard concentration Equal to the average concentration Add standard temperature difference , standard humidity difference and standard atmospheric pressure difference The standard environmental error vector With the corresponding coefficient vector The inner product of , where the standard environmental error vector With coefficient vector The inner product is equal to the standard temperature difference , standard humidity difference and standard atmospheric pressure difference The sum of the products of the corresponding coefficients, the standard temperature difference , standard humidity difference and standard atmospheric pressure difference Zero point correction sequence and full-scale correction sequence The average of the corrected ambient temperature, corrected ambient humidity and corrected ambient pressure is the same as the standard temperature. , standard humidity and standard atmospheric pressure The difference between the standard concentration Including standard zero point concentration and standard full-scale concentration , average concentration Including average zero point concentration and average full-scale concentration .

[0089] Furthermore, the construction of the two-point linear calibration model includes the following steps:

[0090] The reference full-scale concentration Subtract reference zero concentration To obtain the reference range, the standard full-scale concentration Subtract the standard zero concentration To obtain the standard measurement range;

[0091] The ratio of the reference range to the standard measurement range is used as the calibration factor ;

[0092] The reference zero concentration Subtract the standard full-scale concentration and calibration coefficients The product of gets the offset ;

[0093] A two-point linear calibration model is established to describe the linear relationship between the reference value and the standard measurement value, that is, the reference concentration is equal to the standard measurement concentration and the calibration coefficient. The product plus the offset , where the reference concentration includes the reference full-scale concentration and reference zero concentration , standard measurement concentration includes standard full-scale concentration and standard zero concentration .

[0094] The present invention discloses a gas detector calibration system, which includes a measurement module, a feature module, a conversion module and a calibration module; the measurement module repeatedly measures the concentration of zero-point gas and full-scale gas as well as the ambient temperature, ambient humidity and ambient air pressure, considers the influence of the accumulated ambient temperature change on the measurement of the ambient temperature and ambient humidity, corrects the ambient temperature and ambient humidity, converts the ambient air pressure measured each time into the corrected ambient air pressure at sea level based on the principle of atmospheric statics, and generates a zero-point correction sequence and a full-scale correction sequence; the feature module extracts the corresponding environmental coupling information in the zero-point correction sequence and the full-scale correction sequence and the mean and variance of the zero-point concentration and the full-scale concentration in windows of different scales and digits through principal component analysis. The main feature vector is mapped to the corresponding concentration sequence using a gradient boosting tree and the corresponding key feature vector is generated based on feature splitting gain screening; the conversion module inputs each key feature vector into the corresponding attention gating network to determine the corresponding linear conversion model, and converts the mean of the zero-point concentration and the full-scale concentration in the zero-point correction sequence and the full-scale correction sequence into the standard zero-point concentration and the standard full-scale concentration under the standard environment through the linear conversion model; the calibration module constructs a two-point linear calibration model based on the ratio of the range of the standard zero-point concentration and the standard full-scale concentration to the range of the reference zero-point concentration and the reference full-scale concentration, and performs linear calibration on the gas detector, achieving high-precision calibration that maintains a linear relationship.

[0095] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. Gas detector calibration system, characterized in that, Including measurement module, feature module, conversion module and calibration module; The measurement module repeatedly measures the concentration of zero-point gas and full-scale gas as well as the ambient temperature, ambient humidity and ambient pressure, and generates a zero-point correction sequence and a full-scale correction sequence based on the measurement error caused by the accumulation of ambient temperature changes and the influence of altitude on the measurement of ambient pressure; The feature module calculates the mean and variance of the zero-point concentration and the full-scale concentration in windows of different scales and digits in the zero-point correction sequence and the full-scale correction sequence, combines them with the corresponding environmental coupling information, and generates two main feature vectors through principal component analysis. A gradient boosting tree is used to establish a mapping between the main feature vectors and the corresponding concentration sequence, and feature screening is performed based on feature splitting gain to generate two key feature vectors. The conversion module inputs the two key feature vectors into the corresponding attention gating network to determine the corresponding linear conversion model, and converts the average zero-point concentration and the average full-scale concentration of the zero-point correction sequence and the full-scale correction sequence into the standard zero-point concentration and the standard full-scale concentration under standard temperature, standard humidity and standard pressure through the linear conversion model; The calibration module constructs a two-point linear calibration model based on the reference zero concentration, the reference full-scale concentration and the standard zero concentration, and the standard full-scale concentration range and calibrates the gas detector; Among them, the feature module extracts the concentration in the correction sequence to form a concentration sequence, trains the gradient boosting tree to establish a mapping between the main feature vector and the concentration sequence, and evaluates the importance of each feature based on the feature split gain. The least important feature is removed and the change in the training loss of the gradient boosting tree is recorded. The reference score of each feature is obtained and filtered to generate a key feature vector. The concentration sequence includes a zero-point concentration sequence and a full-scale concentration sequence, and the key feature vector includes a zero-point key feature vector and a full-scale key feature vector. Generate a key feature vector, including: construct a concentration sequence based on the concentration in the correction sequence and use it as a training label for the main feature vector, the input, output and training loss of the gradient boosting tree are the main feature vector, the predicted concentration sequence and the mean square error between the predicted concentration sequence and the concentration sequence, respectively, and the predicted concentration sequence includes a zero-point predicted concentration sequence and a full-scale predicted concentration sequence; train the gradient boosting tree based on a node splitting strategy, each node is split according to the feature with the largest node splitting gain in the main feature vector, the node splitting gain is the amount of training loss reduction before and after the node split, and the splitting is stopped until the training loss is less than a minimum value or the tree depth reaches a maximum tree depth; remove the feature with the smallest feature splitting gain in the main feature vector and start a new round of training to calculate the training loss, and use the difference between the training loss of the new round and the previous round as the reference score of the removed feature, wherein the feature splitting gain is equal to the sum of the node splitting gains of all nodes that select the feature for splitting; construct a key feature vector based on features with a reference score greater than a score threshold.

2. The gas detector calibration system according to claim 1, wherein: The measurement module includes a temperature and humidity correction unit and an air pressure correction unit; The temperature and humidity correction unit corrects each ambient temperature to a corrected ambient temperature based on a pre-fitted temperature correction coefficient and a temperature accumulation amount before each measurement, introduces a corresponding time accumulation factor based on the difference between each corrected ambient temperature and the standard temperature, and corrects each ambient humidity to a corrected ambient humidity; The air pressure correction unit is based on the principle of atmospheric statics and combines the change of the ambient temperature with the altitude to correct the ambient pressure at each time.

3. The gas detector calibration system according to claim 1, wherein: The attention gating network is specifically divided into zero-point attention gating network and full-range attention gating network, including gated recurrent layer, attention layer and nonlinear output layer; The gated recurrent layer determines the total number of loop steps and collaboratively utilizes the update gate and the reset gate in each step, fuses the state of the previous step with the key features of the current step, and uses the nonlinear mapping of the retained portion of the state of the previous step to correct and generate the state of the current step, obtains the state of each step and constructs a state set, which includes a zero-point state set and a full-scale state set; The attention layer uses a self-attention mechanism to re-represent the state of each step in the state set using the weighted sum of the states of other steps. The weight is generated by substituting the correlation between the state of each step and the states of other steps into the activation function, and converts the state set into a context vector. The context vector includes a zero-point context vector and a full-range context vector. The nonlinear output layer converts the context vector into a coefficient vector through linear modulation and multiple nonlinear activation functions, wherein the last nonlinear activation function is a Softmax activation function, and the coefficient vector includes a zero-point coefficient vector and a full-range coefficient vector.

4. The gas detector calibration system according to claim 1, wherein: The linear conversion model is specifically that the standard zero-point concentration / standard full-scale concentration is equal to the average zero-point concentration / average full-scale concentration in the zero-point correction sequence / full-scale correction sequence plus the inner product of the standard environmental error vector and the zero-point coefficient vector / full-scale coefficient vector. The standard environmental error vector includes the standard temperature difference, the standard humidity difference and the standard air pressure difference, and is specifically the difference between the average of the corrected ambient temperature, corrected ambient humidity and corrected ambient pressure in the zero-point correction sequence and the full-scale correction sequence and the standard temperature, standard humidity and standard air pressure.

5. The gas detector calibration system according to claim 1, wherein: The feature module organizes the mean and variance of the concentration in windows of different scales and digits in the correction sequence to construct a feature vector, and extracts the splicing results of the feature vector and the environmental coupling information of the correction sequence through principal component analysis to generate a main feature vector. The correction sequence includes a zero-point correction sequence and a full-scale correction sequence, the concentration includes a zero-point concentration and a full-scale concentration, the feature vector includes a zero-point feature vector and a full-scale feature vector, and the main feature vector includes a zero-point main feature vector and a full-scale main feature vector.

6. The gas detector calibration system according to claim 5, wherein: Constructing a feature vector includes the following steps: Based on the factor of the number of measurements, multiple windows of different sizes are set and placed at the starting side of the correction sequence, and the mean and variance of the concentration in each window are calculated; Stop moving the completely traversed window and right-shift all the incompletely traversed windows by one position, and calculate the mean and variance of the concentration in the right-shifted window until each window stops moving; The mean and variance of the concentration in different digits of the same scale window are initially spliced ​​and then secondary spliced ​​into feature vectors according to the scale size.

7. The gas detector calibration system according to claim 5, wherein: The difference between every two adjacent corrected ambient temperatures, corrected ambient humidity, and corrected ambient pressures in the correction sequence is divided by the corresponding concentration difference and the average is taken to obtain the temperature-related gradient, humidity-related gradient, and pressure-related gradient, wherein the temperature-related gradient includes a zero-point temperature-related gradient and a full-scale temperature-related gradient, the humidity-related gradient includes a zero-point humidity-related gradient and a full-scale humidity-related gradient, and the pressure-related gradient includes a zero-point pressure-related gradient and a full-scale pressure-related gradient.

8. The gas detector calibration system according to claim 5, wherein: Constructing a two-point linear calibration model involves the following steps: Subtract the reference zero concentration from the reference full-scale concentration to obtain the reference range, and subtract the standard zero concentration from the standard full-scale concentration to obtain the standard measurement range; The ratio of the reference range to the standard measurement range is used as the calibration coefficient, and the offset is obtained by subtracting the product of the standard full-scale concentration and the calibration coefficient from the reference zero concentration; The two-point linear calibration model is specifically that the reference full-scale concentration / reference zero-point concentration is equal to the product of the standard full-scale concentration / standard zero-point concentration and the calibration coefficient plus the offset.

Images