A sensor-based carbon measurement method and device

Abstract

The present invention discloses a sensor-based carbon measurement method and device, comprising the following steps: A sensor group is provided, the sensor group includes multiple sensors, and multi-dimensional sensing information in the surrounding environment is collected in real time; Machine learning algorithms are used to comprehensively analyze the sensing information and match it with a pre-trained activity recognition model library to determine various activities occurring in the environment; The monitored activity type, intensity, and duration are substituted into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to each activity. The present invention can measure carbon emissions in real time and continuously, and has strong versatility.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor identification, and particularly relates to a carbon measurement method and device based on sensors. Background Art

[0002] The measurement of carbon emissions is a hot topic in the scientific and technological field. However, different from the widely used electricity meters and water meters, the "carbon meter" that can measure carbon emissions has not been invented yet. Currently, there are mainly two types of carbon emission quantification methods: the accounting-based method and the continuous monitoring-based method. The accounting-based method refers to the accumulation of the product of various activity data and their emission factors, or the quantification of greenhouse gas emissions by calculating the carbon mass balance during the process. The continuous monitoring-based method calculates the greenhouse gas emissions by directly measuring the flue gas flow rate and the carbon dioxide concentration in the flue gas.

[0003] However, both methods have obvious drawbacks. The accounting-based method cannot monitor in real time and cannot comprehensively collect activity data. Because of the lack of direct data vouchers, the measurement results may cause disputes, so it is only suitable for trial calculations, rough calculations, and estimations. The continuous monitoring-based method is not only complex in installation and deployment, with high monitoring costs, but also has inconsistent standards and poor data consistency. In cases where the carbon emission point is inconsistent with the actual usage point, such as central air conditioners, it is impossible to reasonably allocate emissions. Summary of the Invention

[0004] The purpose of the present invention is to provide a carbon measurement method and device based on sensors, which can measure carbon emissions in real time and continuously, and have strong versatility.

[0005] To solve the above technical problems, the present invention provides a carbon measurement method based on sensors, including the following steps:

[0006] S1: A sensor group is set, the sensor group includes multiple sensors, and multi-dimensional sensing information in the surrounding environment is collected in real time;

[0007] S2: The sensing information is comprehensively analyzed using a machine learning algorithm and matched with a pre-trained activity recognition model library to judge various activities occurring in the environment;

[0008] S3: The monitored activity types, intensities, and durations are substituted into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to various activities.

[0009] As a further improvement of the present invention, the step S1 further specifically includes the following steps:

[0010] S11: The information collected by each sensor forms an original data stream;

[0011] S12: The original data stream performs frame segmentation on the data stream according to its own code rate, so that the sensor data with different code rates are aligned in time to obtain each sensor data frame. Among them, for sensors with jitter in the signal, a sliding time window is added during data frame segmentation, and the size of the time window is configured according to the jitter situation;

[0012] S13: Perform feature extraction on each sensor data frame, retain the key information of the original data, and at the same time compress and encrypt the data to obtain a set of feature data vectors;

[0013] S14: Perform virtual grouping on the set of feature data vectors to obtain a virtual sensor group feature set.

[0014] As a further improvement of the present invention, the sensor-acquired data performs feature extraction calculations according to the sampling frequency: for sensor data with a sampling frequency lower than 100 Hz, calculate the time-domain eigenvalue of the maximum value, minimum value, median, average value, standard variance, centroid, first-order difference, and second-order difference respectively; for sensor data with a sampling frequency higher than 100 Hz, calculate the time-domain eigenvalue of the maximum value, minimum value, median, average value, standard variance, centroid, first-order difference, and second-order difference respectively, and use the fast Fourier transform to calculate the frequency-domain eigenvalue. The number of frequency-domain eigenvalues depends on the sampling frequency. The higher the sampling frequency, the more the number of frequency-domain eigenvalues.

[0015] As a further improvement of the present invention, the step S14 specifically includes the following steps: according to the characteristics of real activities, bind multiple relevant sensors into a virtual sensor. One virtual sensor corresponds to at least two physical sensors, and the feature combinations of all virtual sensors together are the virtual sensor group feature set.

[0016] As a further improvement of the present invention, the activity recognition model library includes a set of customized activity models. Each activity model corresponds to a carbon emission activity that needs to be recognized and measured. The activity model is generated by supervised machine learning. The training of the activity model includes the following steps:

[0017] Input the virtual sensor group feature set and its corresponding activity type label and environmental background label. The activity type label is that a certain carbon emission activity that needs to be measured has occurred, and the environmental background label is the environmental state where no carbon emission activity that needs to be measured has occurred;

[0018] A one-to-one mapping relationship is established between the environmental background label and the virtual sensor group feature set, and the environmental background parameters of the virtual sensor group are calculated and updated. The background parameters of each virtual sensor are calculated separately according to its feature vector group. The calculation method includes the arithmetic mean and variance of multiple sets of similar training data and the weighted average and variance of some eigenvalues;

[0019] A many-to-one mapping relationship is established between the virtual sensor group feature set and the activity type label, and the data is merged into a training set of virtual sensor group features and activity labels;

[0020] The training set of virtual sensor group features and activity labels is used to update the virtual sensor activation conditions and train the activity recognition model;

[0021] The training of the activity recognition model is carried out in batch training and cross-validation when the training samples accumulate to a certain number or in incremental training and validation on the basis of the already trained activity model;

[0022] The trained activity recognition model is added to the activity recognition model library for real-time activity recognition in the production environment;

[0023] Among them, before the virtual sensor feature data is substituted into the activity recognition model library, it is first filtered according to the virtual sensor activation conditions, and the feature data of the virtual sensors activated by the corresponding activities is retained. The method for updating the virtual sensor activation conditions includes the steps of: calculating the activity parameters of the virtual sensor feature values corresponding to multiple similar activity labels, and calculating the distance between the activity parameters and the corresponding virtual sensor background parameters.

[0024] As a further improvement of the present invention, the recognition process includes the following steps:

[0025] S21: Pre-load the virtual sensor group activation parameters and the activity recognition model library;

[0026] S22: After collecting data and calculating to obtain the virtual sensor group feature set, it is necessary to filter according to the virtual sensor activation conditions, retain the feature data of the virtual sensors activated by the corresponding activities, and obtain a list of activated virtual sensors and corresponding activities;

[0027] S23: Substitute the feature data of the virtual sensors in the list into the corresponding activity recognition models for operation one by one. After the model recognition is completed, output the activity recognition results and the credibility list;

[0028] Among them, a credibility threshold is set for each type of activity. If the credibility of the recognition result exceeds this threshold, the activity recognition results and the credibility list are output. When an activity that needs to measure carbon emissions is recognized and its credibility exceeds the set threshold, the data generated in S1 and S2 of the activity is substituted into the carbon measurement intelligent algorithm model for carbon measurement operation to calculate the carbon emissions corresponding to the activity.

[0029] As a further improvement of the present invention, step S3 specifically includes the following steps:

[0030] S31: Input the activity recognition results and the credibility list, extract relevant data according to the recognized activity types. The data includes the sensing data related to the activity and the context information of this activity.

[0031] S32: After filtering out mutually exclusive activities, obtain the parameter list of the activities to be measured.

[0032] S33: Substitute the parameter list of the activities to be measured into the carbon measurement intelligent algorithm model item by item in terms of activity units, calculate and update the carbon emission activity list and the unit measurement equivalent.

[0033] The carbon emission activity list and the unit measurement equivalent include all the activity information currently in the measurement state and their unit measurement equivalents. The results of the operation of the carbon measurement intelligent algorithm model update this list, specifically including adding activities to the list, deleting activities from the list, and updating the activity information and measurement equivalents in the list. Among them, each update of this list will trigger the accumulation of the data of each item in the carbon emission activity list, so as to obtain the latest carbon emission measurement values of each activity, and the sum of its members gives the final measurement result of carbon measurement.

[0034] Among them, the automatic clock and user query will trigger the accumulation of the data of each item in the carbon emission activity list regularly and irregularly to ensure the timeliness of the measurement result.

[0035] As a further improvement of the present invention, the carbon measurement intelligent algorithm model includes an environmental background model, a direct emission measurement model, an indirect emission measurement model, and a green emission reduction and recycling model. The environmental background model is the impact of the surrounding environmental changes on the carbon emissions of a certain activity. The direct emission measurement model is the carbon emissions directly generated by a certain activity. The indirect emission measurement model is the carbon emissions indirectly caused by the occurrence of a certain activity. The green emission reduction and recycling model is the carbon emission offset brought by the environmental protection factors of a building or system to a certain activity.

[0036] A sensor-based carbon measurement device includes a circuit board and a sensor group and a chip integrated on the circuit board.

[0037] The sensor group includes multiple sensors for real-time collecting multi-dimensional sensing information in the surrounding environment.

[0038] The embedded software runs on the chip. The embedded software includes an activity recognition model library and a carbon measurement intelligent algorithm model.

[0039] Among them, the machine learning algorithm is used to comprehensively analyze the sensing information and match it with the pre-trained activity recognition model library, so as to judge the activities occurring in the environment. And the monitored activity types, intensities and durations are substituted into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to each activity.

[0040] As a further improvement of the present invention, a power module, a communication module and a display module are also integrated on the circuit board; the power module is used to supply power to the entire device, the communication module is used for data transmission, and the display module is used to display monitoring information and metering data.

[0041] Advantages of the present invention: The present invention uses a variety of general sensors to conduct comprehensive environmental information analysis. In theory, it can detect all activities related to carbon emissions in the surrounding environment. For new carbon emission activities, it is also very easy to add them to the activity model library and carbon metering operation model, with strong expandability. It realizes a "carbon meter" that can measure real-time and continuously like an electricity meter and a water meter, and is universal, inexpensive, and can be used by every household. It has a wide detection range, strong universality, and obvious cost advantages. Moreover, the device is easy to deploy, plug-and-play, suitable for wide promotion, and helps the country's "dual carbon" strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a schematic flow chart of the method of the present invention;

[0043] Figure 2 is a schematic flow chart of sensor data acquisition and processing of the present invention;

[0044] Figure 3 is a schematic flow chart of activity model training of the present invention;

[0045] Figure 4 is a schematic flow chart of activity recognition of the present invention;

[0046] Figure 5 is a schematic flow chart of carbon metering operation of the present invention;

[0047] Figure 6 is a schematic structural diagram of the device of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The present invention will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments given are not intended to limit the present invention.

[0049] Refer to Figure 1 , the embodiment of the present invention provides a sensor-based carbon metering method, including the following steps:

[0050] S1: A sensor group is provided, the sensor group includes a plurality of sensors, and multi-dimensional sensing information in the surrounding environment is collected in real time;

[0051] S2: The sensing information is comprehensively analyzed by using a machine learning algorithm and matched with a pre-trained activity recognition model library, so as to judge various activities occurring in the environment;

[0052] S3: Substitute the monitored activity types, intensities, and durations into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to each activity.

[0053] Specifically, all activities that generate carbon emissions indoors are accompanied by multi-dimensional information output at the same time, such as temperature, humidity changes, sound, light, color, electromagnetic waves, infrared, vibration, magnetic field, air composition, etc. For example, when turning on hot water and cold water in the bathroom, in addition to the change in the sound of running water, there are also different thermal radiations; when the refrigerator starts and stops, there are sound, rotation, vibration, electromagnetic waves, and changes in current and voltage; when turning on and off the light, there are changes in brightness, color temperature, and electromagnetic field. The sensor group collects all these environmental information in real time (first-order data collection), and then uses machine learning algorithms to comprehensively analyze this information and match it with the pre-trained activity model library to determine the activities occurring in the environment (second-order activity recognition); then, by integrating the first-order and second-order data, substitute the monitored activity types, intensities, and durations into the carbon measurement operation model (third-order intelligent operation) to calculate the carbon emissions corresponding to these activities.

[0054] As Figure 2 shown, the first-order data collection and processing process: First, the multi-dimensional environmental information emitted by the surrounding environment is collected in real time by the physical sensor group to form the "original data stream of each sensor". Starting from the characteristics of environmental information, different physical sensors require different data sampling frequencies. For example, sounds, electromagnetic waves, etc. require a very high sampling frequency to maintain more accurate information, while temperature and humidity, dust, etc. only require a relatively low sampling frequency. Based on this, all sensor data collections are divided into three categories: high sampling frequency, medium sampling frequency, and low sampling frequency.

[0055] Immediately afterwards, the collected "original data stream of each sensor" is subjected to data frame splitting processing according to their respective code speeds. The purpose is to align the sensor data with different code speeds in time. For example, the data of each sensor takes 100 milliseconds of sampled data as a data frame. In other words, each sensor will generate 10 data frames per second. For sensors with relatively large signal jitter, a sliding time window can be selectively added during data frame splitting, and the size of the time window can be configured according to the jitter situation, ranging from a few hundred milliseconds to several thousand milliseconds.

[0056] After the "data frames of each sensor" are aligned in time, the next step is to perform feature extraction on these data frames. Sensor data feature extraction is to extract the features of the sensor data frames, which not only retains the key information of the original data but also plays a role in data compression and confidentiality. For example, the data frame of a microphone may have thousands of bytes, and after feature extraction, it only has dozens of bytes. After the original sensor data is subjected to feature extraction processing, it is converted into a "group of feature data vectors" represented by F. F includes the same number of vectors as the number of physical sensors, and one sensor corresponds to one vector.

[0057] It should be noted that the data characterization calculation methods for data with different sampling frequencies are different: for sensor data with a sampling frequency lower than 100 Hz, such as magnetic field, color, illuminance, Doppler radar, temperature and humidity, infrared, air pressure, etc., 8 time-domain characteristic values (maximum value, minimum value, median, average value, standard variance, centroid, first-order difference, and second-order difference) are calculated respectively; for sensors with a sampling frequency higher than 100 Hz, such as accelerometers, gyroscopes, microphones, electromagnetic induction, etc., in addition to calculating 8 time-domain characteristic values, it is also necessary to calculate frequency-domain characteristic values using the fast Fourier transform (FFT). The number of frequency-domain characteristic values depends on the sampling frequency. Generally, the higher the sampling frequency, the more frequency-domain characteristic values should be (i.e., using a filter bank with more frequency bands). In particular, MFCC (Mel-scale Frequency Cepstral Coefficient) is used for microphone data to obtain better characteristics. See Table 1 for specific classification information.

[0058]

[0059] Therefore, the mathematical description of the "characterized data vector group" F is:

[0060] F = [f1 f2 … f i … f n (1)

[0061] f i = [f i，1 f i,2 … f i，x (2)

[0062] where n is the number of physical sensors; f i is the characteristic value of one frame of data of the i-th physical sensor, including x members; according to the sampling frequency of the sensor, typical values of x can be 8, 16, or 24.

[0063] Since F contains up to hundreds of characteristic values, if directly used for training and recognition of activity models, not only is the efficiency low and resource consumption high, which conventional IoT chips cannot bear, but also subsequent calibration or addition of the model will be very troublesome. Therefore, it is necessary to further perform virtual grouping of sensors on F, that is, according to the characteristics of real activities, bind multiple related sensors into a virtual sensor. A virtual sensor corresponds to at least two and at most n physical sensors. In other words, it is to divide F into multiple subsets. There can be intersections between subsets, and no two subsets are exactly the same. Its mathematical description is:

[0064] F = F1 ∪ F2 ∪ F3 … ∪ F m , and F1 ≠ F2 ≠ F3 … ≠ Fm (3)

[0065] Among them, m is the number of virtual sensors, which can cover all activities to be monitored. Note: One virtual sensor may correspond to multiple activities to be monitored. Therefore, m is usually less than the total number of activities to be monitored. For example, for the activity of "turning on the light", the relevant physical sensors may include electromagnetic induction (f e ), illuminance (f l ), color (f c ). Therefore, the definitions of these three physical sensors can be combined into one virtual sensor (F l ), that is:

[0066] F1 = [f e f1f c (4)

[0067] The combined characteristics of all m virtual sensors form the "virtual sensor group feature set" S, that is:

[0068] S = {F1, F2, F3…, F m} (5)

[0069] The activity models (classifiers) in the second-order activity recognition library are generated through supervised machine learning. The machine learning tools are not limited. Specifically, they can be: R Studio, Python NumPy, TensorFlow, Spark ML, etc. The activity model training process is as Figure 3 shown.

[0070] In addition to the aforementioned "virtual sensor group feature set" generated by the first-order sensor group, the input data for activity model training also requires manually provided corresponding labels. The labels are divided into two categories: "activity type labels" (L a ) and "environmental background labels" (L b ). "Activity type labels" refer to the occurrence of a certain carbon emission activity to be measured, such as: turning on the light, turning on the air conditioner. "Environmental background labels" refer to the environmental states where no carbon emission activities to be measured occur. Both types of labels are a set of labels, which can include multiple subclasses: "activity type labels" include all activities to be measured. Assuming there are p activities to be measured; while "environmental background labels" can include environmental backgrounds in different seasons and different time periods. Assuming there are q environmental backgrounds to be distinguished. Therefore, the mathematical description of the labels is:

[0071] L a = {l a1 , l a2 , l a3 …, l bp} (6)

[0072] L b = {l b1 , l b2 , l b3 …, l bq} (7)

[0073] It should be noted that for a "virtual sensor group feature set", it either corresponds to an "activity type label" or an "environmental background label". If it corresponds to an "environmental background label", there can only be a specific one; while if it corresponds to an "activity type label", it may include multiple activity types, that is, multiple activity labels. Further, for the convenience of algorithm processing, the label needs to establish a mapping with the members of the "virtual sensor group feature set".

[0074] When the input label is an "environmental background label", assuming this label is l bz , l bz is a member of L b , then the "virtual sensor group feature set" and the "environmental background label" have a one-to-one mapping relationship, which can be described as:

[0075]

[0076] where m is the number of virtual sensors, that is, the features of each virtual sensor correspond to the same environmental background label l bz .

[0077] After completing the mapping between the "environmental background label" and the virtual sensor, the "virtual sensor group environmental background parameter" B can be calculated and updated. The background parameter of each virtual sensor can be calculated separately according to its feature vector group, and the calculation methods include but are not limited to the arithmetic mean and variance of multiple sets of similar training data, the weighted average and variance of some eigenvalues, etc. Its mathematical expression is described as follows:

[0078]

[0079] where and b σi are the mean and deviation of the environmental background parameter of the i-th virtual sensor respectively. m is the number of virtual sensors.

[0080] When the input label is an "activity background label", assuming the labels of this group of activities are L az , L az is a subset of L a , then the "virtual sensor group feature set" and the "activity background label" have a one-to-many mapping relationship, which can be described as:

[0081]

[0082] After the "activity background label" and the virtual sensor are mapped, the data is merged into the "virtual sensor group feature - activity label training set" D t , which is the "X - Y" training sample in machine learning. On this basis, on the one hand, it is needed to update the virtual sensor activation conditions. On the other hand, it is necessary to train the activity recognition model.

[0083] The purpose of the virtual sensor activation condition is to improve the speed of real - time activity recognition in the production environment. Substituting the virtual sensor feature data into the activity recognition model library for recognition consumes more resources. In the daily use process, most of the time, there are no activities that need to be measured. If all feature data are substituted into the activity recognition model library for operation, it is very uneconomical. Therefore, before substituting into the activity recognition model library, first filter according to the virtual sensor activation conditions and only retain the feature data of the virtual sensors whose corresponding activities are activated, which can greatly reduce the operation scale of the activity recognition model. The method for updating the virtual sensor activation conditions is first to calculate the activity parameters of the virtual sensor feature values corresponding to the same type of activity labels multiple times, including the mean and variance. The algorithm is similar to calculating the "virtual sensor group environment background parameters". Then, calculate the distance between the activity parameters and the corresponding virtual sensor background parameters. The distance algorithm here can be the absolute value, Euclidean distance, geometric distance, etc. Assume that the corresponding virtual sensor F of activity x x , then its activation condition can be expressed as:

[0084]

[0085] Among them, and a σx are the mean distance and deviation of the features of the virtual sensor F x from the corresponding background parameters when activity x occurs. m is the number of virtual sensors.

[0086] The training of the activity recognition model can be carried out in batch training and cross - validation when the training samples accumulate to a certain number, or incremental training and verification can be carried out on the basis of the already trained activity model. The machine learning tool is only used in the model training stage and supports conventional machine learning algorithms (such as: SVM support vector machine, random forest, neural network), and can be an offline tool or used online through an automated script. The trained activity model is added to the "activity recognition model library" M and can be used for real - time activity recognition in the production environment.

[0087] The second - order activity recognition process is as Figure 4Shown as follows: First, pre-load the "virtual sensor group activation parameters" and the "activity recognition model library". After collecting data and calculating to obtain the "virtual sensor group feature set", it is necessary to filter according to the virtual sensor activation conditions and only retain the feature data of the virtual sensors for which the corresponding activities are activated, resulting in the "activated virtual sensors and corresponding activity list". Then, substitute the feature data of the virtual sensors in the list into the corresponding activity recognition models (classifiers) one by one for operation. After the model recognition is completed, the "activity recognition result and confidence list" R is output. Here, a confidence threshold can be set for each type of activity, and only when the confidence of the recognition result exceeds this threshold will it be output.

[0088] If an activity that needs to measure carbon emissions is recognized and its confidence exceeds the set threshold, then substitute the first- and second-order data of the activity into the carbon measurement operation model for third-order intelligent operation to calculate the corresponding carbon emissions of these activities.

[0089] The carbon emission calculation process is as Figure 5As shown, the third-order carbon measurement operation first needs to extract relevant data according to the identified activity type. The data includes both the first-order sensing data related to the activity and the context information of this activity. For example, the context information to be extracted for the activity of turning off the light includes the time of the last activity of turning on the light. The algorithm here also needs to filter out mutually exclusive activities at the same moment. Mutually exclusive activities refer to two activities that logically cannot occur simultaneously, such as turning on the light and turning off the light. For this reason, a list of mutually exclusive activities needs to be set in advance. If mutually exclusive activities are found, the activity with a higher credibility is retained. After this operation is completed, the "parameter list of activities to be measured after filtering out mutually exclusive activities" P is obtained. Then, this list is substituted into the carbon measurement intelligent algorithm model item by item in activity units, and the "carbon emission activity list and unit measurement equivalent" L is calculated and updated. The carbon measurement intelligent algorithm includes four algorithm units: "environmental background model" Ie, "direct emission measurement model" Id, "indirect emission measurement model" Ii, and "green emission reduction and recycling model" Ir. It should be noted that although the parameter list of each type of activity includes three types of data: activity type, first-order sensor data, and context information, the specific quantity and type of parameters can be different for different types of activities. Therefore, for each type of activity, these algorithm models are targeted. The "environmental background model" refers to the impact of changes in the surrounding environment on the carbon emissions of a certain activity, which may increase or decrease the carbon emissions of the activity; the "direct emission measurement model" refers to the carbon emissions directly generated by a certain activity; the "indirect emission measurement model" refers to the carbon emissions indirectly caused by the occurrence of a certain activity; the "green emission reduction and recycling model" refers to the carbon emission offset brought by the environmental protection factors of a building or system to a certain activity. These four algorithm units are automatically loaded after the carbon measurement sensor device is started. The form of the algorithm is not limited, and it can specifically be a compensation constant, a compensation coefficient, a linear equation, a fitting function, a piecewise function, etc. The algorithm data (formulas and parameters) can be input or imported by the user through the "carbon measurement intelligent algorithm model maintenance module".

[0090] The "carbon emission activity list and unit measurement equivalent" L is automatically generated and maintained by the system, including all the activity information currently in the measurement state and its unit measurement equivalent. The operation result of the carbon measurement intelligent algorithm model updates this list. Specifically, it includes adding activities to the list, deleting activities from the list, and updating the activity information and measurement equivalent in the list. Each time this list is updated, it will trigger the "accumulation of data for each item in the carbon emission activity list", thereby obtaining the latest "carbon emission measurement value of each activity" C. The sum of the members of C is the final measurement result of this carbon measurement sensor device. In addition, the automatic clock and user queries will regularly trigger and irregularly trigger the "accumulation of data for each item in the carbon emission activity list" to ensure the timeliness of the measurement result. The mathematical expression of the above process is as follows:

[0091] P = {p1, p2,... p i …, py}, where y ∈ [1, p] (12)

[0092] L = {l1, l2, … l i …, l w}, where w ∈ [0, p] (13)

[0093] Among them, l i = I e (p i ) + I d (p i ) + I i (p i ) + I r (p i ) (14)

[0094] C = {c1, c2, … c i …, c p}, where c i = l i × t (15)

[0095] Therefore, the total carbon emissions

[0096] In the above formula, p is the total number of activities to be measured. y is the number of activities to be measured after filtering out mutually exclusive activities in this identification, and its value is between 1 and p. w is the number of the current activity in the carbon emission activity list, and its value is between 0 and p. l i is the unit measurement equivalent of carbon emission activity i, and t is the duration of carbon emission activity i.

[0097] As Figure 6 shown, the present invention also provides a sensor-based carbon measurement device, including a circuit board and a sensor group and a chip integrated on the circuit board;

[0098] The sensor group includes multiple sensors for real-time collection of multi-dimensional sensing information in the surrounding environment;

[0099] An embedded software runs on the chip, and the embedded software includes an activity recognition model library and a carbon measurement intelligent algorithm model;

[0100] Among them, machine learning algorithms are used to comprehensively analyze the sensing information and match it with a pre-trained activity recognition model library to determine various activities occurring in the environment; and the monitored activity type, intensity, and duration are substituted into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to each activity.

[0101] The sensor device consists of hardware and embedded software. The hardware includes: an Internet of Things chip (specifically, it can be STM32F407), a communication module (specifically, it can be a wireless communication module, such as Wi-Fi, Bluetooth, NB-IoT, 4G, etc.; or it can be a wired interface, such as TTL, USB, 433, etc.), a power module (specifically, it can be DC power supply, AC power supply or battery power supply), a display module (specifically, it can be an LED display screen, an LCD display screen, a touch liquid crystal display screen, etc.), and a sensor group. Among them, the sensor group includes an accelerometer, a gyroscope, a microphone, electromagnetic induction, a magnetic field, temperature and humidity, carbon dioxide, dust, illuminance, color, Doppler radar, infrared induction, air pressure, etc. The embedded software runs on the Internet of Things chip, and the core includes two parts: an activity recognition model library and a carbon measurement intelligent algorithm. Among them, the activity recognition model library includes a group of customizable activity models, and the number and categories of the activity models are not limited. Each activity model corresponds to a carbon emission activity that needs to be recognized and measured, such as: air conditioner startup, shutdown, refrigerator startup, hot water output from the shower, microwave oven operation, and so on. The carbon measurement intelligent algorithm includes four algorithm units: an environmental background model, a direct emission measurement model, an indirect emission measurement model, and a green emission reduction and recycling model. Among them, the activity model training process is mainly completed before the carbon measurement sensor device leaves the factory. If the chip performance meets the requirements, it can also be directly completed in the chip. After the carbon measurement sensor device leaves the factory, if further calibration for specific activities or adding new activity models is required, a small amount of on-site training can be carried out.

[0102] Embodiment 1

[0103] Deploy a sensor-based carbon measurement device as described above in the office. Its installation location is unrestricted, and it only needs to be powered on. Its activity recognition library includes various carbon emission activity models in the office environment, such as: turning on the light, turning off the light, corresponding to virtual sensors (electromagnetic induction, illuminance, color); starting and stopping the air conditioner, corresponding to virtual sensors (microphone, electromagnetic induction, temperature, humidity, infrared, gyroscope); running and stopping the coffee machine, corresponding to virtual sensors (microphone, electromagnetic induction, infrared, gyroscope, accelerometer, magnetic field); working and stopping the printer, corresponding to virtual sensors (microphone, electromagnetic induction, gyroscope, accelerometer, magnetic field); completing one page of copying on the copier, corresponding to virtual sensors (microphone, electromagnetic induction, gyroscope, accelerometer, magnetic field); running at high speed, medium speed, low speed and stopping the fan, corresponding to virtual sensors (microphone, electromagnetic induction, radar, gyroscope, accelerometer, magnetic field); turning on and off the computer, corresponding to virtual sensors (microphone, electromagnetic induction, infrared, illuminance, color); starting and stopping heating on the water dispenser, corresponding to virtual sensors (microphone, electromagnetic induction, infrared, magnetic field), etc. If a new carbon emission activity needs to be recognized, only the sensor data of this activity needs to be collected specifically, trained and tested, and after passing the test, the new activity model can be added to the activity recognition library. When the carbon measurement sensor device is working, the sensor group collects various environmental data in the surrounding environment in real time, and substitutes the data into the activity recognition library for recognition. If a certain activity is detected, the relevant information of this activity is further substituted into the carbon measurement operation model to calculate the carbon emission corresponding to this activity. For example, when the air conditioner starts and stops, there are signal changes such as sound, electromagnetic waves, temperature and humidity, and vibration. After its characteristics are detected by the activity recognition library, the carbon measurement operation model comprehensively considers information such as the surrounding environment changes (environmental background model), air conditioner parameters and running duration (direct emission measurement model), indirect carbon emissions to maintain the normal operation of the air conditioning system (indirect emission measurement model), and emission reduction factors of the air conditioner and the building itself (green emission reduction and recycling model) to calculate the carbon emission measurement value of the air conditioner operation activity. The carbon emission measurement value is displayed on the display module of the sensor device in real time, and at the same time, it is transmitted to the client, server or cloud platform through the communication module.

[0104] Embodiment 2

[0105] Install one of the above-mentioned sensor-based carbon measurement devices in a vehicle, which is powered through the vehicle's cigarette lighter or power port converted to a USB port, and its installation location is not restricted. Its activity recognition library includes various carbon emission activity models during vehicle use, such as: engine start, driving, acceleration, deceleration, braking, waiting, engine shutdown, corresponding to virtual sensors (microphone, electromagnetic induction, temperature, humidity, infrared, gyroscope, accelerometer); opening and closing the car door, corresponding to virtual sensors (microphone, electromagnetic induction, temperature, humidity, infrared, radar); turning on and off the in-vehicle lights, corresponding to virtual sensors (electromagnetic induction, illuminance, color); starting and stopping the in-vehicle air conditioner, corresponding to virtual sensors (microphone, electromagnetic induction, temperature, humidity, infrared, gyroscope); turning on and off the in-vehicle audio, etc., corresponding to virtual sensors (microphone, electromagnetic induction, gyroscope). If a new carbon emission activity needs to be recognized, only the sensor data of this activity needs to be collected specifically, trained and tested, and after passing the test, the new activity model can be added to the activity recognition library. When the carbon measurement sensor device is working, the sensor group collects various in-vehicle environmental data in real time, and substitutes the data into the activity recognition library for recognition. If a certain activity is detected, the relevant information of this activity is further substituted into the carbon measurement operation model to calculate the carbon emissions corresponding to this activity. For example, when the engine starts and stops, there are signal changes such as sound, electromagnetic wave, and vibration. After its characteristics are detected by the activity recognition library, the carbon measurement operation model calculates the carbon emission measurement value of vehicle use by integrating information such as the surrounding environmental changes (environmental background model), engine parameters, driving speed, mileage (direct emission measurement model), indirect carbon emissions for maintaining normal vehicle operation (indirect emission measurement model), and the vehicle's own emission reduction factors (green emission reduction and recycling model). The carbon emission measurement value is displayed in real time on the display module of the sensor device, and at the same time, it is transmitted to the client, server or cloud platform through the communication module.

[0106] The above office scenarios and vehicles are only examples. This method is applicable to various other scenarios and means of transportation, such as: shopping malls, classrooms, kitchens, bathrooms, wards, bedrooms, workrooms, tool rooms, shops, buses, airplanes, trains, etc. Moreover, by networking the carbon measurement sensors, it is possible to monitor the carbon emissions of a building, a community, or even a city in real time.

[0107] The above-described embodiments are only preferred embodiments cited to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A sensor-based carbon measurement device, characterized in that: It includes a circuit board and a sensor group and a chip integrated on the circuit board; The sensor group includes multiple sensors for real-time collection of multi-dimensional sensing information in the surrounding environment; Embedded software runs on the chip, and the embedded software includes an activity recognition model library and a carbon measurement intelligent algorithm model; Among them, machine learning algorithms are used to comprehensively analyze the sensing information and match it with the pre-trained activity recognition model library to judge various activities occurring in the environment; and the monitored activity types, intensities, and durations are substituted into the carbon measurement intelligent algorithm model to calculate the carbon emissions corresponding to each activity; The activity recognition model library includes a set of customized activity models, each activity model corresponding to a carbon emission activity that needs to be recognized and measured. The activity model is generated by supervised machine learning. The training of the activity model includes the following steps: Input the virtual sensor group feature set and its corresponding activity type label and environmental background label. The activity type label is a carbon emission activity that needs to be measured, and the environmental background label is the environmental state where no carbon emission activity that needs to be measured occurs; A one-to-one mapping relationship is established between the environmental background label and the virtual sensor group feature set, and the environmental background parameters of the virtual sensor group are calculated and updated. The background parameters of each virtual sensor are calculated separately according to its feature vector group. The calculation method includes the arithmetic mean and variance of multiple sets of similar training data and the weighted average and variance of some eigenvalues; A one-to-many mapping relationship is established between the virtual sensor group feature set and the activity type label, and the data is merged into the virtual sensor group feature and activity label training set; The virtual sensor group feature and activity label training set are used to update the virtual sensor activation conditions and train the activity recognition model; The training of the activity recognition model is carried out in batches and cross-validated when the training samples accumulate to a certain number or incrementally trained and validated on the basis of the already trained activity model; The trained activity recognition model is added to the activity recognition model library for real-time activity recognition in the production environment; Among them, before the virtual sensor feature data is substituted into the activity recognition model library, it is first filtered according to the virtual sensor activation conditions, and the feature data of the virtual sensors corresponding to the activated corresponding activities is retained. The method for updating the virtual sensor activation conditions includes the steps of calculating the virtual sensor feature value activity parameters corresponding to multiple sets of similar activity labels and calculating the distance between the activity parameters and the corresponding virtual sensor background parameters.

2. The carbon measurement device based on a sensor according to claim 1, wherein: When the multi-dimensional sensing information in the surrounding environment is collected in real time, it includes: S11: The information collected by each sensor forms an original data stream; S12: The original data stream is frame-processed according to its own code rate so that the sensor data with different code rates are aligned in time to obtain each sensor data frame. Among them, for sensors with signal jitter, a sliding time window is added during data framing, and the size of the time window is configured according to the jitter situation; S13: Feature processing is performed on each sensor data frame to retain the key information of the original data, and at the same time, the data is compressed and encrypted to obtain a feature data vector group; S14: Perform virtual grouping on the characterized data vector group to obtain a virtual sensor group feature set.

3. The carbon metering device based on a sensor according to claim 2, wherein: When performing virtual grouping on the characterized data vector group, it includes: according to the characteristics of real activities, binding multiple relevant sensors into a virtual sensor. One virtual sensor corresponds to at least two physical sensors. The combined features of all virtual sensors are the virtual sensor group feature set.

4. The carbon measurement device based on a sensor according to claim 1, characterized in that: The data collected by the sensors are characterized and calculated according to the sampling frequency: for sensor data with a sampling frequency lower than 100 Hz, calculate the time-domain eigenvalue of the maximum value, minimum value, median, average value, standard variance, centroid, first-order difference, and second-order difference respectively; For sensor data with a sampling frequency higher than 100 Hz, calculate the time-domain eigenvalue of the maximum value, minimum value, median, average value, standard variance, centroid, first-order difference, and second-order difference respectively, and use the fast Fourier transform to calculate the frequency-domain eigenvalue. The number of frequency-domain eigenvalues depends on the sampling frequency. The higher the sampling frequency, the more the number of frequency-domain eigenvalues.

5. The carbon measurement device based on a sensor according to claim 1, characterized in that: The recognition process includes the following steps: S21: Preload the virtual sensor group activation parameters and the activity recognition model library; S22: After collecting data and calculating to obtain the virtual sensor group feature set, it is necessary to filter according to the virtual sensor activation conditions, retain the feature data of the virtual sensors whose corresponding activities are activated, and obtain the list of activated virtual sensors and corresponding activities; S23: Substitute the feature data of the virtual sensors in the list into the corresponding activity recognition models one by one for operation. After the model recognition is completed, output the activity recognition result and the credibility list; Among them, a credibility threshold is set for each type of activity. If the credibility of the recognition result exceeds this threshold, output the activity recognition result and the credibility list. When an activity that needs to measure carbon emissions is recognized and its credibility exceeds the set threshold, substitute the data generated above for the activity into the carbon measurement intelligent algorithm model for carbon measurement operation to calculate the carbon emissions corresponding to the activity.

6. The carbon metering device based on a sensor according to claim 1, characterized in that: The specific steps for calculating the carbon emissions corresponding to each activity include the following: S31: Input the activity recognition result and the credibility list, extract relevant data according to the recognized activity type. The data includes the sensing data related to the activity and the context information of this activity; S32: After filtering out mutually exclusive activities, obtain the parameter list of the activities to be measured; S33: Substitute the parameter list of the activities to be measured into the carbon measurement intelligent algorithm model item by item in terms of activity units, calculate and update the carbon emission activity list and the unit measurement equivalent; S34: The carbon emission activity list and the unit measurement equivalent include all the activity information currently in the measurement state and their unit measurement equivalents. The results of the operation of the carbon measurement intelligent algorithm model update this list, specifically including adding activities to the list, deleting activities from the list, and updating the activity information and measurement equivalents in the list; among them, each update of this list will trigger the accumulation of the data of each item in the carbon emission activity list, so as to obtain the latest carbon emission measurement values of each activity, and the sum of its members is the final measurement result of carbon measurement. Among them, the automatic clock and user queries will trigger regularly and irregularly, and accumulate the data of each item in the carbon emission activity list to ensure the timeliness of the measurement results.

7. The carbon metering device based on a sensor according to claim 1, characterized in that: The carbon measurement intelligent algorithm model includes an environmental background model, a direct emission measurement model, an indirect emission measurement model, and a green emission reduction and recycling model; the environmental background model is the impact of surrounding environmental changes on the carbon emissions of a certain activity; the direct emission measurement model is the carbon emissions directly generated by a certain activity; the indirect emission measurement model is the carbon emissions indirectly caused by a certain activity; the green emission reduction and recycling model is the carbon emission offset brought by the environmental protection factors of a building or system to a certain activity.

8. The carbon measurement device based on a sensor according to claim 1, characterized in that: The circuit board is also integrated with a power module, a communication module, and a display module; the power module is used to supply power to the entire device, the communication module is used for data transmission, and the display module is used to display monitoring information and measurement data.

Images