Hydrogen concentration measuring method based on multi-mode sensor

Abstract

The invention discloses a hydrogen concentration measurement method based on a multi-mode sensor, and relates to the technical field of gas detection and industrial process safety monitoring. Hydrogen sensors based on different physical principles are introduced in the same measurement link, and environmental variables such as temperature, humidity, pressure and the like and sensor output are collected together; a multi-dimensional feature is formed after normalization and time alignment, and then fusion is performed by using a weight adaptive to the change of an environment, so that instantaneous errors generated by a single sensor due to temperature drift, wet drift or pressure fluctuation are shared or suppressed, and the stability and repeatability of a fusion result are improved; furthermore, a cross sensitivity compensation module arranged before fusion identifies response differences between hydrogen and interference components such as carbon monoxide and methane by using a machine learning model obtained by training under a multi-gas mixing working condition, and the probability of false signals and false alarms caused by coexisting gas is reduced on the premise that a special detection channel is not added.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas detection and industrial process safety monitoring, and in particular to a hydrogen concentration measurement method based on a multi-modal sensor. BACKGROUND

[0002] In a chemical plant environment, a pipeline system involves multiple gases coexisting and dynamic working conditions, and hydrogen leakage monitoring is a key link to ensure safe operation. In such a scenario, the gas composition is complex and the concentration changes frequently, and the environmental parameters such as temperature and pressure fluctuate significantly, which requires high real-time and accuracy for hydrogen concentration measurement. The existing monitoring method needs to achieve reliable detection in the background of interference factors, but in actual application, there are cross-sensitivity and dynamic interference, which can easily cause measurement results to deviate.

[0003] In recent years, hydrogen concentration measurement technologies based on new sensing materials have been gradually applied, such as organic semiconductor sensors and optical fiber surface plasmon resonance sensors. Such solutions show high sensitivity and response speed in single gas or stable environments, but in the pipeline monitoring scenario of a chemical plant, when hydrogen coexists with carbon monoxide, methane and other gases, the sensor is easily affected by cross-sensitivity, producing false signals. At the same time, sudden changes in pipeline pressure or rapid changes in environmental temperature can cause the sensor output to drift, causing unstable readings. Some methods use hardware filtering or static calibration to alleviate the problem, but it is difficult to adapt to real-time changes in dynamic interference.

[0004] To address the above-mentioned disadvantages, some existing technologies introduce multi-sensor fusion or environmental parameter compensation mechanisms. For example, combining optical and electrical sensor data, or using a temperature and humidity module to correct signals. Such methods improve anti-interference ability to some extent, but still have limitations in complex working conditions: multi-sensor data synchronization processing has high delay, and the compensation model relies on preset parameters, which cannot fully adapt to sudden fluctuations and long-term drift in pipeline leakage monitoring. In addition, the compensation effect of cross-sensitivity is limited by sensor characteristics and algorithm generalization ability, and is prone to false negatives or false positives in a changing environment. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides a hydrogen concentration measurement method based on a multi-modal sensor to solve the problem that existing hydrogen monitoring is prone to cross-sensitivity, large drift, and insufficient real-time accuracy in multi-gas, temperature and pressure rapid change working conditions.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides a hydrogen concentration measurement method based on a multi-modal sensor, which includes:

[0009] Step S1, synchronously collecting hydrogen response signals and environmental parameters of a measured environment by using a multi-modal sensor array, the multi-modal sensor array at least including a hydrogen sensor based on a first physical principle and a hydrogen sensor based on a second physical principle, and the environmental parameters at least including one of temperature and humidity;

[0010] Step S2, pre-processing each signal output by the multi-modal sensor array, the pre-processing including signal normalization and time alignment, to obtain multi-dimensional feature data for fusion;

[0011] Step S3, determining weights of feature data of each sensor based on reliability indexes corresponding to the sensors, and performing multi-source data fusion on the multi-dimensional feature data to obtain a fused hydrogen concentration value;

[0012] Step S4, comparing the fused hydrogen concentration value with a preset threshold and outputting a measurement result.

[0013] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor, the multi-modal sensor array includes an organic semiconductor hydrogen sensor and a surface plasmon resonance optical sensor, which respectively constitute the hydrogen sensor based on the first physical principle and the hydrogen sensor based on the second physical principle.

[0014] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor, the multi-source data fusion in step S3 includes:

[0015] The weights of the feature data of each sensor are dynamically adjusted according to the environmental parameters, so as to reduce the influence of temperature and / or humidity changes on the fused hydrogen concentration value.

[0016] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor, the dynamic adjustment of the weights of the feature data of each sensor is based on confidence degrees of the sensors, and the confidence degrees are determined according to a deviation degree of historical output data of the sensors from current environmental parameters.

[0017] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor, step S3 further includes cross-sensitivity compensation on the multi-dimensional feature data, so as to distinguish hydrogen signals from signals of at least one interfering gas.

[0018] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor according to the present application, wherein: the cross-sensitivity compensation is realized through a machine learning model, the machine learning model is trained offline based on multi-modal sensor data under multiple gas mixing conditions, and outputs the compensated hydrogen concentration component during online measurement.

[0019] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor according to the present application, wherein: further comprising:

[0020] At a preset time interval or when sensor drift is detected, a reference gas with a known hydrogen volume fraction is introduced for automatic calibration, and the fusion parameters used in the multi-source data fusion are updated according to the deviation between the actual output of the sensor to the reference gas and the reference value.

[0021] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor according to the present application, wherein: the reference gas is zero gas or standard gas containing target hydrogen concentration, and the automatic calibration is carried out without affecting online measurement.

[0022] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor according to the present application, wherein: the multi-modal sensor array further includes a sensor for collecting pressure, and the preprocessing step further includes density or refractive index compensation based on pressure for each signal.

[0023] As a preferred scheme of the hydrogen concentration measurement method based on a multi-modal sensor according to the present application, wherein: the output measurement result includes:

[0024] When the fusion hydrogen concentration value exceeds the preset threshold, an alarm signal is generated and sent to an external safety control system.

[0025] The application has the beneficial effects that: the application introduces hydrogen sensors based on different physical principles at the same measurement link at the same time, and collects environmental quantities such as temperature, humidity, and pressure together with sensor output, forms a multi-dimensional feature after normalization and time alignment, and then uses adaptive weights that change with the environment to fuse, so that the instantaneous error of a single sensor caused by temperature drift, humidity drift, or pressure fluctuation is shared or suppressed, and the stability and repeatability of the fusion result are improved. Further, the cross-sensitivity compensation module set before fusion uses the machine learning model trained under the mixed gas condition to identify the response difference of hydrogen and interfering components such as carbon monoxide and methane, reduces the false signal and false alarm probability caused by coexisting gases without increasing the dedicated detection channel. Because the system supports automatic calibration triggered by time or drift, the real concentration corresponding to the reference gas can be written back to the fusion parameters, environmental correction factors, or model coefficients, thereby compensating for long-term drift caused by sensor aging, pollution, and zero drift, so that the device still maintains measurement accuracy matching the safety alarm threshold under long-term online monitoring conditions. The final threshold comparison and alarm output make the method can be directly embedded in the safety control link of the chemical device, and still has the ability to discover hydrogen leakage in time under dynamic conditions such as device start-stop, material switching, or short-term emptying, which improves the adaptability and robustness of the system, reduces false positives and false negatives caused by environmental disturbances, and is suitable for deployment in actual industrial scenes with multiple gases and dynamic temperature and pressure. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope of the present application.

[0027] Figure 1 The flowchart of the hydrogen concentration measurement method based on multi-modal sensors in the embodiments. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will further illustrate the present application in detail by combining with the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, and are not used to limit the present application.

[0029] All terms used in the present application (including technical and scientific terms) have the meanings generally understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0030] For example, the terms "first", "second", etc. used in the present application are only used to distinguish similar objects, to distinguish the first object from another object, and are not used to describe a specific order or sequence, nor can it be understood as indicating or implying relative importance.

[0031] The present application proposes a hydrogen concentration measurement method based on a multi-modal sensor, which combines Figure 1 As shown in the figure, the method comprises:

[0032] Step S1, synchronously collecting hydrogen response signals and environmental parameters of the measured environment by using a multi-modal sensor array, the multi-modal sensor array at least comprising a hydrogen sensor based on a first physical principle and a hydrogen sensor based on a second physical principle, and the environmental parameters at least comprising one of temperature, humidity;

[0033] Step S2, pre-processing each signal output by the multi-modal sensor array, the pre-processing comprising signal normalization and time alignment, to obtain multi-dimensional feature data for fusion;

[0034] Step S3, performing multi-source data fusion on the multi-dimensional feature data based on the reliability indicators corresponding to each sensor, to obtain a fused hydrogen concentration value;

[0035] Step S4, comparing the fused hydrogen concentration value with a preset threshold value and outputting a measurement result;

[0036] In the present embodiment, the hydrogen response signals of the measured environment are preferably obtained by the aforementioned two types of hydrogen sensors based on different physical principles at the same sampling time, and the environmental parameters are preferably given in real time by temperature detection elements, humidity detection elements or pressure detection elements installed adjacent to the sensor array, or can be provided by existing process measurement points in the upper monitoring system after writing the current time identifier, so as to ensure that the sensor outputs and environmental quantities at the same time can be one-to-one corresponding. The sampling period can be set to one second by default to meet the needs of discovering short-term pressure fluctuations and small-scale leaks in chemical pipelines; when the monitored object changes slowly, it can be relaxed to within ten seconds, and when there is a fast start-stop or emptying operation, it can be tightened to half a second. The adjustment basis is to ensure that the change of the environmental quantity within any sampling period does not exceed the allowed error band given by the sensor under that range. The preset threshold value can be set once according to the safety instrument level, the control requirements of the production device for hydrogen volume fraction or the enterprise internal standard, or can be fine-tuned according to the actual background gas level during the device debugging period. When adjusting, do not change the subsequent fusion and compensation processing flow. Alternatively, when a certain channel fails to return a valid value within the current period, the system temporarily sets the participation weight of the channel to zero, still uses the valid data of the remaining channels to complete the preprocessing and fusion, and automatically restores the participation weight of the channel after the channel returns to normal output in the subsequent period, so as to ensure that the entire measurement process can still be executed in the case of single-point missing measurement.

[0037] In one embodiment, the multi-modal sensor array comprises an organic semiconductor hydrogen sensor and a surface plasmon resonance optical sensor, which respectively constitute a hydrogen sensor based on a first physical principle and a hydrogen sensor based on a second physical principle;

[0038] In this specification, the hydrogen sensor based on the first physical principle refers to a sensor that uses the change in electrical quantity (such as conductance, resistance, voltage, current) generated after the gas interacts with the sensitive material to represent the hydrogen concentration; the hydrogen sensor based on the second physical principle refers to a sensor that uses the change in optical quantity (such as reflected light intensity, resonance wavelength, phase, or refractive index change) caused by the presence of gas to represent the hydrogen concentration. The measurement objects of the above two types of sensors are the same, but the sensing mechanisms and output dimensions are different, so they can be used as multi-modal data sources to participate in subsequent fusion.

[0039] Specifically, the organic semiconductor hydrogen sensor can be selected from devices that have a specific chemical adsorption or conductance change response to hydrogen, and its output is usually an electrical signal or resistance change with the change of hydrogen concentration, which is suitable for direct collection under normal pressure to medium pressure field conditions; the surface plasmon resonance optical sensor can be selected from structures that have a light path with the measured gas or are led out to the detection cavity through an optical fiber, and its output is usually the change in reflected light intensity, resonance wavelength, or phase shift, which is suitable for use in places where the ambient temperature fluctuates and high selectivity is required. In order to reduce the measurement error caused by space difference, it is preferred to arrange the two types of sensors in the same sampling gas path or within a range of not more than one tenth of the pipe diameter, and set a preheating or stabilization time of not less than three minutes after starting, so that the zero point and initial reliability of the sensor are consistent with the subsequent calibration value. When one type of sensor is temporarily unavailable, the basic reliability of the other type of sensor can be increased to the upper limit set in advance in this embodiment, but the output format required for multi-source data fusion is not changed.

[0040] In one embodiment, the multi-source data fusion in step S3 comprises:

[0041] Adjusting the weight of each sensor feature data according to the environmental parameters to reduce the influence of temperature and / or humidity change on the fused hydrogen concentration value;

[0042] In this embodiment, the environmental parameters can be stored in the cache area in the order of sampling time before entering the weight adjustment, so that the temperature, humidity and pressure values of the recent several cycles can be directly read when time alignment or historical deviation calculation is required. By default, the cache depth can be set to the last sixty sampling points, which is equivalent to a one-minute environmental quantity window when the sampling period is one second. When the external environment fluctuates quickly, the cache depth can be shortened to twenty sampling points to improve the sensitivity of weight update. The above environmental parameters maintain the same time identifier as the corresponding sensor channel when writing to the cache, and do not need to introduce a new time synchronization module. As long as the data entering the fusion module is completed within one sampling period, the real-time value can be restored after the next successful sampling when the environmental quantity is missing for a short time.

[0043] The weight adjustment mode of each sensor feature data is as follows:

[0044] Step D1, obtaining environmental quantity and channel quantity at the current sampling time to obtain a set ; wherein, is the current environmental temperature, unit is , is the current environmental relative humidity, unit is %, is the current environmental pressure, unit is or the same pressure unit as the system, is the current output of the channel in the multi-modal sensor array, is the basic reliability index of the channel obtained under the standard reference environment, dimensionless quantity, is the channel index, the value range is , is the total number of channels contained in the array, which is consistent with the above-mentioned array structure;

[0045] In this embodiment, the current environmental temperature can be obtained by a digital temperature acquisition element in the same shell as the sensor array, or can be obtained by a temperature measurement point provided by an external control system after time alignment; the relative humidity can be obtained by a capacitive or impedance type humidity acquisition element, if the monitored environment is dry or treated gas, the relative humidity can be fixed as the last valid measurement value; the environmental pressure can be obtained from the pressure transmitter installed on the same pipeline, and when the output unit is inconsistent with the unit set by the system, it needs to be converted by the known proportion coefficient. The basic reliability index can be measured once according to the standard reference environment at the manufacturing stage or initial installation stage and written into the storage space. The index is only updated when automatic calibration or obvious drift is detected in subsequent operation, so as to avoid parameter inconsistency caused by frequent writing.

[0046] Step D2, for each channel, calculate the un-normalized weight according to the current environment and its own state, the calculation formula is:

[0047] ,

[0048] wherein, is the un-normalized weight of the i-th channel, dimensionless quantity, is the un-normalized weight of the i-th channel, dimensionless quantity, is the basic reliability of the i-th channel, is the basic reliability of the i-th channel, is the temperature correction factor of the i-th channel, dimensionless quantity, is the temperature correction factor of the i-th channel, dimensionless quantity, is the humidity correction factor of the i-th channel, dimensionless quantity, is the humidity correction factor of the i-th channel, dimensionless quantity, is the pressure correction factor of the i-th channel, dimensionless quantity, is the pressure correction factor of the i-th channel, dimensionless quantity, is the correction factor based on the deviation of the current output of the channel from the expected output, dimensionless quantity;

[0049] For example, when the sampling period is short and the environmental quantity changes little, the temperature correction factor, the humidity correction factor and the pressure correction factor can be directly taken as the calculation results of the previous period to reduce repeated operation, and only when it is detected that the environmental quantity changes more than a pre-set small threshold, the calculation is re-performed; when there is no valid reading of a certain environmental quantity in the current period, the correction factor can be temporarily taken as a neutral value of one to ensure that the un-normalized weight can still be calculated. The basic reliability of each channel determines the contribution of different physical principle sensors under normal working conditions, and it is suggested that the basic reliability of the channel with high sensitivity to hydrogen and low signal noise be set to a normalized value of one to two under the reference environment, and the basic reliability of the channel which is sensitive to the environmental quantity or has large long-term drift be set to a normalized value of 0.5 to 1, so that the subsequent environmental correction factor can amplify or suppress it.

[0050] In the formula, the temperature correction factor is defined as:

[0051] ,

[0052] wherein, is the temperature sensitivity coefficient of the i-th channel, unit is , the larger the value, the more sensitive the channel is to temperature deviation, is the reference temperature of the i-th channel, unit is , is the natural exponential function; is the natural exponential function;

[0053] ​Further, the temperature sensitivity coefficient in the above temperature correction factor can be obtained according to the actual temperature drift characteristics of different types of sensors in the automatic calibration stage. For the organic semiconductor channel which is sensitive to temperature change, a larger coefficient can be taken to form a steeper decay curve. For the optical channel which has good temperature stability, a smaller coefficient can be taken to allow a certain range of temperature fluctuations. The reference temperature can be the standard room temperature agreed upon in device design, or the reference temperature given in the sensor factory calibration report. If the device is operated in high or low temperature conditions for a long time, the stable temperature of the current working condition can also be written as a new reference temperature when it is first put into operation, so that the correction factor remains maximum near this temperature. To prevent the correction factor from suddenly dropping due to occasional temperature measurement anomalies, when the calculated temperature deviation exceeds the range that can physically occur in the device, the temperature deviation can be limited to the maximum absolute value allowed before participating in the operation.

[0054] If it is necessary to write the rationality of the real-time output into the weight, it is defined as:

[0055] ,

[0056] Wherein, is the output deviation decay coefficient of the i th channel, and the unit is consistent with the inverse of is the expected output of the i th channel obtained according to the historical calibration or training model under the current temperature , humidity , pressure ;

[0057] Optionally, the decay coefficient in the correction factor obtained based on the deviation of the current output and the expected output can be set according to the noise level of the channel. For channels with high noise, a smaller decay coefficient can be used so that slight output fluctuations do not immediately cause a significant decrease in the weight. For channels with stable output but which may indicate a fault once deviated, a larger decay coefficient can be used to quickly reduce its participation when atypical readings occur. The expected output can be obtained through the historical calibration data, offline training model or statistical average value formed during long-term operation. When the current combination of environmental quantities has not appeared in the history, the expected output corresponding to the combination of environmental quantities closest to it can be used instead. If the current output deviates significantly from all known expected outputs and the deviation lasts for more than 3 to 5 sampling periods, the channel can be marked as a maintenance state, and the correction factor of the channel is forced to zero during the marking period.

[0058] Step D3, to make the sum of the weights of all channels equal to 1, normalize all the unnormalized weights obtained in step D2, and the calculation formula is written as: ​​​

[0059] ,

[0060] wherein, is the normalized weight of the i-th channel, dimensionless, which can be directly sent into the multi-source data fusion module, is the summation result of all the un-normalized weights of the channels, is the summation index, whose value range is , is the total number of channels; the normalized is the direct output form of the weight or confidence of each channel;

[0061] Similarly, in actual operation, the un-normalized weights of all channels may approach zero at the same time, for example, all external environmental quantities fail or most channels fail at the same time. At this time, in order to avoid the denominator being zero, a small constant of fixed magnitude can be added to the summation result to ensure that the normalization can be completed, and at the same time, the fusion result of this period is marked as a low confidence output for the upper system to reference. For the normalized weight, an upper limit value and a lower limit value per channel can also be set, so that any single channel cannot obtain all the weights in normal working condition, and will not completely lose the weight due to a slight fluctuation, thereby maintaining the smoothness of the fusion result.

[0062] In the specific implementation process, the three landing ways of the update strategy are as follows:

[0063] 1. Interval lookup table method: for each channel, a corresponding table of to is preset, and each correction factor is directly read according to the current measured temperature, humidity and pressure landing point, and then steps D2 and D3 are calculated;

[0064] 2. Function relationship method: the exponential or segmented expression given in step D2 is adopted, and , or the segmented coefficients are written in the automatic calibration stage;

[0065] 3. Online regression method: according to the difference between the actual fusion output and the reference output during the long-term operation of the system, the coefficients in , or the table are written back;

[0066] ​In this embodiment, online regression can be triggered at fixed time intervals or after detecting systematic deviation for several consecutive cycles to prevent excessive parameter adjustment caused by accidental environmental disturbances; the fixed time interval can be set to 1-24 hours according to the continuous operation characteristics of the device, and can be temporarily shortened to ten minutes when the device is in a stage of frequent start-stop or frequent switching of gas sources. When online regression, it is preferred to do weighted smoothing on the newly calculated coefficients and the original coefficients, so that the update of the parameters is in a gradual manner and will not cause a sudden change in weight distribution after a single update. If there is insufficient storage space or the current cycle data volume does not meet the requirements of the regression algorithm during online regression, this update can be skipped, the original coefficients remain unchanged, and the next time the conditions are met, the execution continues.

[0067] In one embodiment, the dynamic adjustment of the weight of each sensor feature data is based on the confidence of each sensor, which is determined according to the deviation of the historical output data of the sensor from the current environmental parameter;

[0068] In one embodiment, step S3 further comprises cross-sensitivity compensation of the multi-dimensional feature data to distinguish the hydrogen signal from the signal of at least one interfering gas;

[0069] In one embodiment, the cross-sensitivity compensation is realized by a machine learning model, which is trained offline based on multi-modal sensor data under multiple gas mixing conditions, and outputs the compensated hydrogen concentration component during online measurement;

[0070] The machine learning model construction step comprises:

[0071] Step E1, construct a training set, introduce hydrogen and at least one interfering gas in the experimental device in sequence or in groups to obtain a sample sequence ; wherein, is the total number of training samples; each sample records:

[0072] 1. The current ambient temperature , unit ;

[0073] 2. The current ambient relative humidity , unit %

[0074] 3. The current ambient pressure , unit ;

[0075] 4. The output of each channel in the multi-modal sensor array at this time , wherein , is the total number of array channels;

[0076] 5. Real hydrogen gas volume fraction at the time of input , in % (volume fraction);

[0077] 6. If there are interfering gases at the same time, record the real volume fraction of each interfering gas , where , is the number of interfering gas types;

[0078] The output here is a labeled sample pair ;

[0079] where the feature vector ;

[0080] In this embodiment, in order to ensure that the training set can cover the possible multi-gas combinations in the field, the experimental device should at least contain pure hydrogen gas, pure interfering gas and several mixed gas scenes with different volume fractions when constructing samples. The temperature, humidity and pressure under each scene are preferably taken at different levels within the expected operating range, for example, several stable points are selected at low temperature, medium temperature and high temperature, so that the trained model can make reasonable estimates within the full range. The total number of samples can be determined according to the model complexity and the desired generalization ability. When using a deep learning model, the number of samples should be greater than several thousand, and when using a lightweight model such as gradient boosting tree, several hundred samples can be used. If individual channels temporarily output abnormal values during the collection process, the abnormal values of the channel can be treated as missing in the sample labeling stage, while the valid outputs of the remaining channels are retained, so that the model can learn the ability to estimate the hydrogen concentration even in the case of partial channel missing.

[0081] Step E2, in the cross-sensitivity compensation scene, the sensor and the environmental quantity can be written into the model input, and the target concentration can be written into the model output; Two output forms can be selected:

[0082] Single output regression form: the model only outputs the hydrogen concentration, denoted as:

[0083] ,

[0084] where is the hydrogen concentration estimate given by the model for the th sample, in %, is a machine learning model with parameter vector , which can be a deep neural network, a gradient boosting tree or other supervised regression model, is the parameter vector to be solved in the training stage, which is a dimensionless quantity;

[0085] Multi-output regression form: the model simultaneously outputs the hydrogen and interfering gas components, denoted as:

[0086] ,

[0087] in, For the model to the first The first sample given Estimates of the volume fraction of interfering gases, in %. For multi-output machine learning models, For interfering gas indexes, the value range is: ;

[0088] For example, when using a multi-output approach, a single-output approach can be used first to complete the regression training for hydrogen concentration and obtain a relatively stable network structure or tree model parameters. Then, output nodes for each interfering gas component can be added. This reduces training time and improves the convergence of the training process. During model training, the samples can be divided into training and validation sets in an 8:2 or 7:3 ratio. The validation set does not participate in parameter updates and is only used to evaluate the error level under unknown combinations. Training can be terminated early when the validation error no longer decreases within several training epochs to avoid overfitting. If only the hydrogen concentration is needed in practical applications and the output of each interfering gas component is not required, only the output node corresponding to the hydrogen concentration can be retained during the deployment phase, and the other nodes can be excluded from forward computation, thus saving runtime resources.

[0089] Step E3: If a single-output regression is used, the following formula can be used as the objective function for all samples during the training phase:

[0090] ,

[0091] in, Let be the scalar of the loss function for this training. The integral number of hydrogen gas estimated by the model. The integral of the actual hydrogen gas in the sample. The regularization coefficient is a dimensionless quantity used to suppress excessively large model parameters. The square of the L2 norm of the parameter vector. This is the sample index, and its value range is... If a multi-output format is used, then add the following to the summation term: item;

[0092] Optionally, the regularization coefficient can be automatically estimated based on the training set size and model capacity. A larger regularization coefficient can be used when the training set has a small number of samples or a large number of model layers to limit model complexity. Conversely, a smaller regularization coefficient can be used when the training set has a large number of samples and a low input dimensionality to improve fitting accuracy. The training process can be limited by two constraints: a maximum number of epochs and an early stopping condition on the validation set. Training is immediately terminated and the current optimal parameters are saved when either constraint is reached, ensuring that training time remains within a controllable range. If the loss function value is found to increase abnormally over several consecutive epochs during training, the process can revert to the previously saved stable parameters, reduce the learning rate, or increase the batch size before retraining.

[0093] In the specific implementation process, the model is deployed at the following locations:

[0094] 1. If the computing power at the edge meets the computational requirements of the model, the trained... Downloaded to the terminal side and received directly by the terminal side. and output or The advantage of edge deployment is short response time;

[0095] 2. If edge resources are limited, the trained model can be deployed on the host computer. The current feature vector is uploaded from the edge, and the host computer completes the forward calculation and returns the compensated hydrogen component to the fusion module. Both methods only affect the location of the model and do not change the mathematical relationship between steps E2 and E3.

[0096] In this embodiment, to avoid the impact of communication jitter on online measurements, a set of model parameters from the previous version or a set of rule-based simplified compensation logic can be retained on the terminal side. When the host computer fails to return the compensated hydrogen component within a specified time window, the terminal side immediately calls the backup logic to complete the fusion calculation for this cycle, and automatically switches back to the host computer's calculation results after communication is restored. The size of the communication time window can be set between tens of milliseconds and hundreds of milliseconds, depending on network conditions and host computer load. The principle for setting it is to ensure that results can still be provided within one sampling cycle even in the worst case. If no results are obtained from the host computer within multiple consecutive sampling cycles, an alarm can be triggered and the current measurement can be marked as a degraded operation state.

[0097] In addition, the output of step E3 The hydrogen concentration component, after cross-sensitivity compensation, can be incorporated into the multi-source data fusion process, along with the weights already given in the previous round. This forms a path that first compensates and then weights the results. If a multi-output approach is chosen, then from... Winning The components are input into fusion; further, if the hydrogen concentration component compensated by cross-sensitivity and the weighted results of each original channel are obtained at the same time in one fusion cycle, the compensated hydrogen concentration component can be used as the output of the cycle according to the preset confidence order, and the result obtained by weighting the original channel is saved as a backup value for subsequent self-checking, calibration or online regression. When the difference between the compensated hydrogen concentration component and the backup value continuously exceeds the allowed range, the compensation path can be temporarily disabled and the fusion mode based on weight only can be used to ensure the continuity and safety of the output.

[0098] In one embodiment, the method further comprises:

[0099] At a preset time interval or when sensor drift is detected, a reference gas with a known hydrogen volume fraction is introduced for automatic calibration, and the fusion parameters used in multi-source data fusion are updated according to the deviation between the actual output of the sensor to the reference gas and the reference value;

[0100] In this embodiment, the introduction of the reference gas can be performed by bypassing small flow injection or short time switching. The injection time can be set to ten to thirty seconds to ensure that all types of sensors can complete the response and output stable values. The calibration period can be set to once a day or once a week according to the long-term drift level of the sensor, and a calibration can also be triggered immediately when it is detected that the fusion output presents a deviation in the same direction in multiple consecutive cycles. When updating the fusion parameters, it is preferred to adjust the basic reliability and environmental correction coefficient of each channel respectively in a component manner, and it is not recommended to replace all parameters at once to avoid sudden changes in measurement characteristics due to abnormal single calibration. If a certain channel does not respond, responds too slowly or responds in the wrong direction during the reference gas injection stage, only the parameters of the remaining channels can be updated, and the abnormal channel is marked as a state that needs to be checked and maintained.

[0101] In one embodiment, the reference gas is zero gas or a standard gas containing a target hydrogen concentration, and the automatic calibration is performed without affecting online measurement;

[0102] In one embodiment, the multi-modal sensor array further includes a sensor for collecting pressure, and the preprocessing step further includes density or refractive index compensation based on pressure for each signal;

[0103] In one embodiment, the output measurement result comprises:

[0104] When the fused hydrogen concentration value exceeds a preset threshold, an alarm signal is generated and sent to an external safety control system.

[0105] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the present application; even though the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some or all of the technical features thereof can be replaced equivalently; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0106] In addition, those skilled in the art will appreciate that a combination of features of different embodiments means that such combination is within the scope of the present application and forms a different embodiment. For example, all of the above embodiments can be used in any combination. The information disclosed in this section is merely intended to deepen the understanding of the general background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes the prior art known to those skilled in the art.

Claims

1. A hydrogen concentration measurement method based on a multi-modal sensor, characterized by, The method comprises the following steps: Step S1, synchronously collecting hydrogen response signals and environmental parameters of a measured environment by using a multi-modal sensor array, wherein the multi-modal sensor array comprises at least a hydrogen sensor based on a first physical principle and a hydrogen sensor based on a second physical principle, and the environmental parameters comprise at least one of temperature and humidity; Step S2, pre-processing each signal output by the multi-modal sensor array, wherein the pre-processing comprises signal normalization and time alignment, and multi-dimensional feature data for fusion is obtained; Step S3, determining the weight of each sensor feature data based on the reliability index corresponding to each sensor, and performing multi-source data fusion on the multi-dimensional feature data to obtain a fused hydrogen concentration value; Step S4, comparing the fused hydrogen concentration value with a preset threshold value and outputting a measurement result.

2. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, The multi-modal sensor array comprises an organic semiconductor hydrogen sensor and a surface plasmon resonance optical sensor, which respectively constitute the hydrogen sensor based on the first physical principle and the hydrogen sensor based on the second physical principle.

3. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, The multi-source data fusion in step S3 comprises: The weight of each sensor feature data is dynamically adjusted according to the environmental parameters to reduce the influence of temperature and / or humidity changes on the fused hydrogen concentration value.

4. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 3, wherein, The dynamic adjustment of the weight of each sensor feature data is based on the confidence of each sensor, and the confidence is determined according to the deviation of historical output data of the sensor from the current environmental parameters.

5. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, Step S3 further comprises cross-sensitivity compensation on the multi-dimensional feature data to distinguish the hydrogen signal from the signal of at least one interfering gas.

6. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 5, wherein, The cross-sensitivity compensation is realized by a machine learning model, which is trained offline based on multi-modal sensor data under multiple gas mixing conditions, and outputs a compensated hydrogen concentration component during online measurement.

7. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, Further comprising: At a preset time interval or when sensor drift is detected, a reference gas with a known hydrogen volume fraction is introduced for automatic calibration, and fusion parameters used in the multi-source data fusion are updated according to the deviation between the actual output of the sensor to the reference gas and the reference value.

8. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 7, wherein, The reference gas is zero gas or standard gas containing a target hydrogen concentration, and the automatic calibration is performed without affecting online measurement.

9. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, The multi-modal sensor array further comprises a sensor for collecting pressure, and the pre-processing step further comprises density or refractive index compensation based on pressure for each signal.

10. The hydrogen concentration measurement method based on a multi-modal sensor according to claim 1, wherein, The output measurement result comprises: When the fused hydrogen concentration value exceeds the preset threshold value, an alarm signal is generated and sent to an external safety control system.

Images