An edge-computing-based group-fog monitoring method and monitoring system

Abstract

A kind of group fog monitoring method based on edge computing, the method comprises: (one), using data mining technology to train group fog early warning model;(two), train group fog monitoring model;(three), group fog early warning model and group fog monitoring model are evaluated, update and joint, form real-time group fog computing model;(four), the environmental data of monitored area are collected, the environmental data of monitored area are substituted into real-time group fog computing model, and current group fog monitoring result is calculated by real-time group fog computing model;A kind of group fog monitoring system based on edge computing, the system includes sensing end, edge device, data transmission module, server end;Wherein, the signal output end of sensing end is connected with the signal input end of edge device, the signal output end of edge device is connected with the signal input end of data transmission module, the signal output end of data transmission module is connected with the signal input end of server end.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of based on edge computing's group fog monitoring method and monitoring system, belong to image processing and edge computing field. BACKGROUND

[0002] Group fog is influenced by local microclimate environment, in the local range of dozens of meters to hundreds of meters in heavy fog, the fog that appears is more "thick", visibility is lower;Group fog appears in a small range, only a few kilometers or even hundreds of meters, the visibility is good outside group fog, the visibility is extremely low inside group fog, the coverage of single group fog is often only dozens of meters or even tens of meters, belongs to strong thick fog, with suddenness, local, small scale, concentration, difficult to predict and forecast characteristics;The appearance of group fog brings a lot of inconvenience to people's travel, such as: affected by wind force, group fog can move, when group fog moves to highway, it will lead to sudden change of visibility, vision becomes blurred, it is difficult to identify isolation pier and guardrail, and driver is prone to judgment and operation error, resulting in major traffic accidents, therefore, group fog is also called "mobile killer" of highway.

[0003] In view of the harmfulness of group fog, group fog monitoring work is very important, early group fog monitoring is often reminded by the operation and maintenance experience of relevant personnel (such as, when group fog on highway is monitored, the operation and maintenance experience of highway management and maintenance unit is often relied on) in group fog season, but there is lack of professional, efficient and accurate group fog monitoring system or method;With the development of science and technology, the method of target detection and classification based on deep learning becomes mainstream, by training group fog monitoring model on computer end, deploying trained group fog monitoring model on cloud server, cloud server calculates the received image according to the deployed group fog monitoring model, and feeds back the calculation result to user, which can meet the group fog monitoring requirements under certain conditions, but there are also defects, for example, the response speed of collected image and video data is slow due to large magnitude, which increases data transmission cost, and the monitoring efficiency of single model is not high and the monitoring precision is low;At the same time, the existing group fog monitoring technology can only monitor the presence or absence of group fog, and does not warn group fog and subdivide the specific level when group fog appears, which is not conducive to the development of more refined law enforcement means by traffic department. SUMMARY

[0004] The present application proposes a kind of based on edge computing's group fog monitoring method and monitoring system, its purpose aims at solving the problem that whether group fog occurs cannot be warned when the prior art monitors group fog.

[0005] The technical solution of the present application: a kind of based on edge computing's group fog monitoring method, the method includes:

[0006] (1), group fog warning model is trained using data mining technology;

[0007] (II) Training fog monitoring model;

[0008] (III) Evaluate, update and combine the fog early warning model and the fog monitoring model to form a real-time fog calculation model;

[0009] (iv) Collect environmental data of the monitored area, substitute the environmental data of the monitored area into the real-time fog calculation model, and calculate the current fog monitoring result through the real-time fog calculation model.

[0010] Furthermore, the process of training the fog early warning model using data mining techniques specifically includes:

[0011] 1) Collect meteorological data values ​​of the occurrence of fog in the monitored area in history, including wind speed, humidity, temperature, air pressure, and duration of fog;

[0012] 2) Establish a fog warning index model, as shown in formula (1):

[0013] R f = v*V + h * H + f *F + p *P + t * T (1);

[0014] Among them, R f The fog warning index is defined as follows: V represents wind speed, H represents humidity, F represents temperature, P represents air pressure, and T represents the duration of fog. V, h, f, p, and t are the weighting coefficients for each variable. Specifically, v is the weighting coefficient for wind speed, h is the weighting coefficient for humidity, f is the weighting coefficient for temperature, p is the weighting coefficient for air pressure, and t is the weighting coefficient for the duration of fog.

[0015] 3) Classify the fog warning levels, specifically including: calculating the fog warning index R using the fog warning index model. f The standard deviation classification method was used, with the average value of the fog warning index as the basis. Based on the standard deviation s of the fog warning index, fog warnings are divided into several fog warning levels.

[0016] Furthermore, when The fog warning level is Level 1. The fog warning level is Level 2. The fog warning level is level 3. The fog warning level is level 4; the higher the value of the fog warning level, the greater the harm caused by the fog.

[0017] Furthermore, the weight coefficients corresponding to each variable are determined using the analytic hierarchy process (AHP), specifically including:

[0018] (1) Construct the judgment matrix;

[0019] (2) Perform a consistency check on the judgment matrix;

[0020] (3) Calculate the weight coefficients for each variable.

[0021] Furthermore, the construction of the judgment matrix specifically includes: selecting wind speed, humidity, temperature, air pressure, and fog duration from meteorological data values ​​as five variables, assigning variable numbers to each of the five variables; and comparing each of the five variables pairwise to obtain the relative importance 'a' between the different variables. ij i = 1, 2, 3, 4, 5; j = 1, 2, 3, 4, 5; The judgment matrix A is obtained, and its mathematical expression is:

[0022]

[0023] Among them, a ij This indicates the relative importance between variables i and j; wind speed is set as variable 1, humidity as variable 2, temperature as variable 3, air pressure as variable 4, and fog duration as variable 5.

[0024] Furthermore, the consistency check of the judgment matrix specifically includes:

[0025] 2-1) Calculate the eigenvalues ​​λ of the judgment matrix A. max， Calculate the consistency index CI and find the average random consistency index RI.

[0026]

[0027] 2-2) Calculate the consistency ratio (CR):

[0028]

[0029] When CR < 0.1, the judgment matrix A is considered to have satisfactory consistency; when CR ≥ 0.1, the judgment matrix A is considered to lack satisfactory consistency, and in this case, the relative importance 'a' between different variables is readjusted. ij The specific value of a, and then the adjusted a ij Substitute the specific value into the judgment matrix A, and repeat steps 2-1) and 2-2) until CR < 0.1 and the judgment matrix A has satisfactory consistency.

[0030] Further, the weighting coefficients for each variable are calculated, specifically including:

[0031] 3-1) Normalize the judgment matrix A by column, that is, divide each element of the judgment matrix A by the sum of its column to generate a normalized matrix A.5×5 ;

[0032] 3-2) Normalize matrix A 5×5 Adding the columns together generates matrix A. 5×1 ;

[0033] 3-3) The matrix A obtained after addition 5×1 Dividing each element by n, where n = 5, yields the weight vector ω = [ω1, ω2, ω3, ω4, ω5]. T Wherein, ω1 is the weighting coefficient corresponding to wind speed, ω2 is the weighting coefficient corresponding to humidity, ω3 is the weighting coefficient corresponding to temperature, ω4 is the weighting coefficient corresponding to air pressure, and ω5 is the weighting coefficient corresponding to the duration of fog.

[0034] Furthermore, the training fog monitoring model specifically includes:

[0035] 1) Collect historical images of fog patches in the monitored areas to form a fog patch dataset;

[0036] 2) The collected fog images in the fog dataset are filtered according to certain filtering rules, retaining fog images with clear image quality and clear targets. The fog images in the filtered fog dataset are then subjected to random rotation, random color transformation, random cropping, random scaling, and random flipping to augment the data and form a new fog dataset.

[0037] 3) Use a visibility meter to mark the visibility value and location coordinates of each fog image in the new fog dataset, and save the visibility value and fog location coordinates in YOLO format;

[0038] 4) The new fog dataset is randomly divided into training set, validation set and test set according to a certain ratio, and the number of fog images in each training set, validation set and test set is counted respectively.

[0039] 5) In the PyTorch framework, select the YOLOv5 model, adjust the fog detection parameters, and train the fog detection model using the training set through the YOLOv5 model;

[0040] 6) Selection of fog monitoring model: After each iteration, the fog monitoring model trained is validated and tested on the validation and test sets in the new fog dataset, and the model with the highest accuracy is selected as the fog monitoring model.

[0041] Furthermore, the evaluation, updating, and combination of the fog early warning model and the fog monitoring model to form a real-time fog calculation model specifically includes:

[0042] 1) Calculate the updated weights:

[0043] L represents the fog warning model, and M represents the fog monitoring model. The fog warning model and the fog monitoring model are normalized separately and then divided. The ratio α represents the updated weight.

[0044] 2) Generate a real-time fog calculation model:

[0045] The fog early warning model and the fog monitoring model are evaluated, updated, and combined according to formula (3) to form a real-time fog calculation model C:

[0046] C = Le -α +M(1-e -α ), α∈[0,1] (3).

[0047] A fog monitoring system based on edge computing is disclosed. The system is suitable for running fog monitoring methods based on edge computing. The system includes a sensing end, an edge device, a data transmission module, and a server. The signal output end of the sensing end is connected to the signal input end of the edge device, the signal output end of the edge device is connected to the signal input end of the data transmission module, and the signal output end of the data transmission module is connected to the signal input end of the server.

[0048] The beneficial effects of this invention are:

[0049] 1) This invention utilizes a fog early warning model, a fog monitoring model, and a fog calculation model for data processing and analysis. By using the fog early warning model and the fog monitoring model for joint verification and updating, it is possible to provide early warning of whether fog will occur, thereby improving the efficiency of fog monitoring and solving the problems of low accuracy and poor performance of existing fog monitoring models.

[0050] 2) Through further design, the fog patches are classified into different levels during the fog patch warning stage, which solves the problem of the single classification of fog patch levels in the existing fog patch monitoring model. Attached Figure Description

[0051] Appendix Figure 1 This is a flowchart illustrating a fog monitoring method based on edge computing.

[0052] Appendix Figure 2 This is a schematic diagram of a fog monitoring system based on edge computing.

[0053] Appendix Figure 3 This is a schematic diagram of the training process of the fog monitoring model in this invention. Detailed Implementation

[0054] A fog detection method based on edge computing, comprising the following steps:

[0055] (I) Training a fog early warning model using data mining techniques;

[0056] (II) Training fog monitoring model;

[0057] (III) Evaluate, update and combine the fog early warning model and the fog monitoring model to form a real-time fog calculation model;

[0058] (iv) Collect environmental data of the monitored area, substitute the environmental data of the monitored area into the real-time fog calculation model, and calculate the current fog monitoring result through the real-time fog calculation model.

[0059] The fog monitoring model is preferably trained using a deep learning neural network; more preferably, the YOLOv5 model is used to train the fog monitoring model.

[0060] The environmental data of the monitored area is collected through the sensing end; during operation, the sensing end collects the environmental data of the monitored area; the monitored area is a geographical area that needs to be monitored and determined whether fog occurs, such as sections of highways that are prone to fog; the environmental data includes images, videos, temperature, humidity, wind speed, air pressure and other data of the monitored area.

[0061] The sensing end includes monitoring instruments such as cameras, temperature sensors, and humidity sensors; the sensing end also includes a small weather instrument that can measure wind speed, wind direction, air pressure, rainfall, etc. The sensing end acquires images, videos, wind speed, humidity, temperature, air pressure, and fog duration data around the edge computing-based fog monitoring system in real time.

[0062] The process of training a fog early warning model using data mining techniques specifically includes the following steps:

[0063] 1) Collect meteorological data values ​​of the occurrence of fog patches in the monitored area in history. The meteorological data values ​​include wind speed, humidity, temperature, air pressure, and duration of fog patches. Preferably, collect meteorological data values ​​of the monitored area in the past 3 years. The past 3 years refer to the 3 years before the monitoring year. For example, if monitoring the occurrence of fog patches in the monitored area in 2023, it is preferable to collect meteorological data values ​​of the monitored area in 2020, 2021, and 2022. It is even more preferable to collect meteorological data values ​​of the corresponding months of the monitored area in the past 3 years. For example, if monitoring the occurrence of fog patches in the monitored area from October to December 2023, it is even more preferable to collect meteorological data values ​​of the monitored area from October to December 2020, 2021, and 2022.

[0064] 2) Establish a fog warning index model, as shown in formula (1):

[0065] R f= v*V + h * H + f *F + p *P + t * T (1);

[0066] Among them, R f The fog warning index is defined as follows: V represents wind speed, H represents humidity, F represents temperature, P represents air pressure, and T represents the duration of fog. V, h, f, p, and t are the weighting coefficients for each variable. Specifically, v is the weighting coefficient for wind speed, h is the weighting coefficient for humidity, f is the weighting coefficient for temperature, p is the weighting coefficient for air pressure, and t is the weighting coefficient for the duration of fog.

[0067] 3) Classify the fog warning levels, specifically including: calculating the fog warning index R using the fog warning index model. f The standard deviation classification method was used, with the average value of the fog warning index as the basis. Based on the standard deviation s of the fog warning index, fog warnings are divided into four levels: when R... f The fog warning level is Level 1. The fog warning level is Level 2. The fog warning level is level 3. The fog warning level is level 4.

[0068] The higher the value of the fog warning level, the greater the hazard of the fog. Specifically, a fog warning level of 1 indicates a slightly hazardous fog; a level 2 indicates a relatively hazardous fog; a level 3 indicates a very hazardous fog; and a level 4 indicates a severely hazardous fog. For example, when monitoring fog on a highway, a level 1 fog warning indicates a slightly hazardous fog, a level 2 fog indicates a relatively significant fog, a level 3 fog indicates a very significant fog, and a level 4 fog indicates a severely hazardous fog.

[0069] The weight coefficients corresponding to each variable are determined using the analytic hierarchy process (AHP), specifically including the following steps:

[0070] (1) Construct a judgment matrix, specifically including: selecting wind speed, humidity, temperature, air pressure, and fog duration from meteorological data values ​​as 5 variables, assigning variable numbers to each of the 5 variables; comparing the 5 variables pairwise to obtain the relative importance a between different variables. ij i = 1, 2, 3…5; j = 1, 2, 3…5; The judgment matrix A is obtained, and its mathematical expression is:

[0071]

[0072] Among them, aij This indicates the relative importance between variable i and variable j; preferably, wind speed is set as variable 1, humidity as variable 2, temperature as variable 3, air pressure as variable 4, and fog duration as variable 5; for example: a 12 This indicates the relative importance of wind speed and humidity compared to each other; a 21 This indicates the relative importance of humidity and wind speed compared to each other.

[0073] (2) Perform a consistency check on the judgment matrix A, which includes the following steps:

[0074] 2-1) Calculate the eigenvalues ​​λ of the judgment matrix A. max The consistency index CI is obtained, and the average random consistency index RI is found through the average random consistency index reference table.

[0075]

[0076] 2-2) Calculate the consistency ratio (CR):

[0077]

[0078] When CR < 0.1, the judgment matrix A is considered to have satisfactory consistency; when CR ≥ 0.1, the judgment matrix A is considered to lack satisfactory consistency, and in this case, the relative importance 'a' between different variables is readjusted. ij The specific value of a, and then the adjusted a ij Substitute the specific value into the judgment matrix A, and repeat steps 2-1) and 2-2) until CR < 0.1 and the judgment matrix A has satisfactory consistency.

[0079] (3) Calculate the weight coefficients for each variable:

[0080] 3-1) Normalize the judgment matrix A by column, that is, divide each element of the judgment matrix A by the sum of its column to generate a normalized matrix A. 5×5 ;

[0081] 3-2) Normalize matrix A 5×5 Adding the columns together, i.e., summing the rows, generates matrix A. 5×1 ;

[0082] 3-3) The matrix A obtained after addition 5×1 Dividing each element by n, where n = 5, yields the weight vector ω = [ω1, ω2, ω3, ω4, ω5]. TWherein, ω1 is the weighting coefficient corresponding to wind speed, ω2 is the weighting coefficient corresponding to humidity, ω3 is the weighting coefficient corresponding to temperature, ω4 is the weighting coefficient corresponding to air pressure, and ω5 is the weighting coefficient corresponding to the duration of fog.

[0083] The training fog monitoring model specifically includes the following steps:

[0084] 1) Collection of fog patch dataset; The collection of the fog patch dataset includes collecting historical images of fog patches in the monitored area to form a fog patch dataset; preferably, the dataset is formed by acquiring images of fog patches from the past 3 years through cameras installed in the monitored area, or by expanding the dataset with publicly available images of fog patches online; the past 3 years refers to the 3 years prior to the monitoring year. For example, if monitoring the occurrence of fog patches in the monitored area in 2023, it is preferable to collect images of fog patches from the monitored area in 2020, 2021, and 2022 to form a fog patch dataset; further preferably, images of fog patches from the monitored area in the corresponding months of the past 3 years are collected to form a fog patch dataset. For example, if monitoring the occurrence of fog patches in the monitored area from October to December 2023, it is preferable to collect images of fog patches from the monitored area from October to December 2020, 2021, and 2022 to form a fog patch dataset; when collecting images of fog patches on highways, continuous shooting can be selected, or shooting can be performed by setting a certain sampling interval;

[0085] 2) Creation of the fog dataset; The creation of the fog dataset includes filtering the fog images in the collected fog dataset according to certain filtering rules, retaining fog image data with clear image quality and clear target, and performing random rotation, random color transformation, random cropping, random scaling and random flipping on the fog images in the filtered fog dataset to augment the data and form a new fog dataset.

[0086] 3) Fog dataset annotation;

[0087] The annotation of the fog dataset includes using a visibility meter to annotate the visibility value and fog location coordinates of each fog image in the new fog dataset, and saving the visibility value and fog location coordinates in YOLO format;

[0088] 4) Partitioning of the fog dataset;

[0089] The partitioning of the fog dataset specifically includes: randomly dividing the new fog dataset into a training set, a validation set, and a test set at a certain ratio, and counting the number of fog images in each of the training set, validation set, and test set respectively; preferably, the new fog dataset is randomly divided into a training set, a validation set, and a test set at a ratio of 7:2:1.

[0090] 5) Training of the fog monitoring model;

[0091] The training of the fog monitoring model specifically includes: selecting the YOLOv5 model in the PyTorch framework, adjusting fog monitoring parameters such as the number of iterations (step), batch size, confidence level (conf), and learning rate (lr), and training the fog monitoring model using the training set through the YOLOv5 model. Training the fog monitoring model can improve the real-time performance and accuracy of fog monitoring.

[0092] The iteration number step is preset to an initial step = 300. When the number of iterations step < 300 during the training of the fog monitoring model, the rate of change of the loss function loss has approached zero. The number of iterations step can be appropriately reduced from 300 at the beginning of the next fog monitoring model training to save training time. When the number of iterations step = 300 during the training of the fog monitoring model, the rate of change of the loss function loss still does not approach zero. The number of iterations step can be appropriately increased from 300 at the beginning of the next fog monitoring model training.

[0093] The batch size is preset to 32. At the beginning of each iteration, the training set from the new fog dataset is fed into the YOLOv5 model for training according to the preset batch size. The larger the batch size, the faster the training speed. The confidence level is preset to 0.6. In practical applications, when the required false negative rate for fog detection is low, the confidence level can be preset to >0.6; when the required false negative rate for fog detection is high, the confidence level can be preset to ≤0.6. The learning rate is preset to 0.1 to observe the loss function loss of the YOLOv5 model in the initial stage. If the loss function loss shows gradient explosion in the early stage of training, it means that the initial learning rate is too large. The initial learning rate can be reduced by 10 times and tried again. If the loss function loss decreases slowly in the early stage of training, it means that the initial learning rate is too small. The initial learning rate can be increased by 5 times or 10 times and tried again.

[0094] 6) Selection of fog monitoring model:

[0095] After each iteration, the fog monitoring model trained is validated and tested on the validation and test sets of the new fog dataset. The model with the highest accuracy and precision is selected as the fog monitoring model.

[0096] The evaluation, updating, and combination of the fog early warning model and the fog monitoring model to form a real-time fog calculation model specifically includes the following steps:

[0097] 1) Calculate the updated weights:

[0098] L represents the fog warning model, and M represents the fog monitoring model. The fog warning model L and the fog monitoring model M are normalized and then divided. The ratio α represents the updated weight.

[0099] 2) Generate a real-time fog calculation model:

[0100] The fog warning model and the fog monitoring model are evaluated, updated, and combined according to the following formula (3) to form a real-time fog calculation model C:

[0101] C = Le -α +M(1-e -α ), α∈[0,1] (3).

[0102] The process of calculating the current fog monitoring results using a real-time fog calculation model specifically includes: real-time collection of data such as wind speed, humidity, temperature, air pressure, fog duration, and video images of the monitored road section, and inputting the data into the real-time fog calculation model, which can then calculate the current fog monitoring results, i.e., whether fog has occurred.

[0103] A fog monitoring system based on edge computing includes a sensing end, an edge device, a data transmission module, and a server. The signal output end of the sensing end is connected to the signal input end of the edge device, the signal output end of the edge device is connected to the signal input end of the data transmission module, and the signal output end of the data transmission module is connected to the signal input end of the server.

[0104] A fog monitoring system based on edge computing, the system also includes a data platform, and the signal output terminal of the server is connected to the signal input terminal of the data platform.

[0105] The edge computing-based fog monitoring system can be used to run edge computing-based fog monitoring methods.

[0106] The sensing terminal collects highway video and environmental data, and connects to edge devices to deploy fog warning and monitoring models to edge device nodes, enabling rapid fog warning and monitoring and reducing the occurrence of major traffic accidents.

[0107] The edge device is used to receive data of various structural types collected by the sensing end. The edge device uses data mining technology to train a fog early warning model and uses the YOLOv5 model to train a fog monitoring model at the edge, which can provide real-time early warning and monitoring of fog occurrence. This invention deploys the fog early warning model, fog monitoring model and real-time fog calculation model on the edge device, performs data processing and analysis on the edge device, and only returns the monitoring results to the server, which can reduce network bandwidth and transmission pressure and improve monitoring efficiency.

[0108] The edge device returns the fog monitoring results to the server in real time through the data transmission module.

[0109] The edge device includes a hardware development board that can process incoming images and videos in real time, and features low power consumption, small size, high computing power, and easy compatibility.

[0110] The data platform is used to receive fog monitoring results sent by the server and to perform visualization analysis and display.

[0111] Example 1

[0112] The following is a detailed description of an application example of the fog early warning model trained using data mining techniques in this invention, which includes the following steps:

[0113] 1) Collect meteorological data values ​​of the occurrence of fog in the monitored area in history, including wind speed, humidity, temperature, air pressure, and duration of fog;

[0114] 2) Establish a fog warning index model, as shown in formula (1):

[0115] R f =v*V + h *H + f *F + p *P + t *T (1);

[0116] Among them, R f The fog warning index is defined as follows: V is wind speed, H is humidity, F is temperature, P is air pressure, and T is the duration of fog. The weighting coefficients for wind speed, humidity, temperature, air pressure, and duration of fog are: v = weighting coefficient, h = weighting coefficient, f = weighting coefficient, p = weighting coefficient, and t = weighting coefficient.

[0117] 3) Establish a hierarchical model: "Determine the fog warning index" is set as the target layer; "wind speed, humidity, temperature, air pressure, and fog duration" are set as the criteria layer; and "fog warning index" is set as the scheme layer.

[0118] 4) Constructing the judgment matrix: Wind speed, humidity, temperature, air pressure, and fog duration are used as the five elements in the criterion layer. The relative importance of each of the five elements in the criterion layer is calculated by comparing them pairwise, as shown in Table 1.

[0119]

[0120] The relative importance value in Table 1 represents the scale of the element corresponding to the vertical direction and the element corresponding to the horizontal direction. The calculation of relative importance should follow the calculation rules in Table 2:

[0121] Scale Meaning 1 Also important 3 Slightly important 5 Quite important 7 Very important 9 Extremely important 2、4、6、8 The midpoint of the two adjacent judgments Reciprocal If the scale is 3, then B is 1 / 3 as important as A

[0122] We obtain the judgment matrix A:

[0123]

[0124] 5) Consistency check of the judgment matrix:

[0125] Calculate the eigenvalues ​​λ of the judgment matrix A max To obtain the consistency index And find the average random consistency index RI using the following reference table of average random consistency indices:

[0126] n 1 2 3 4 5 6 7 8 9 10 RI 0 0 0.52 0.89 1.12 1.26 1.36 1.41 1.46 1.49

[0127] When n=5, RI=1.12. Calculate the consistency ratio. When CR < 0.1, the judgment matrix A is considered to have satisfactory consistency; after calculation, CR = 0.06, CR < 0.1, the judgment matrix A has satisfactory consistency.

[0128] 6) Normalize the judgment matrix A by column, that is, divide each element of the judgment matrix A by the sum of its column to generate a normalized matrix A. 5×5 ;

[0129] 7) Normalize matrix A 5×5 Adding the columns together, i.e., summing the rows, generates matrix A. 5×1 ;

[0130] 8) The matrix A obtained after addition 5×1 Dividing each element by 5 yields the weight vector ω = [ω1, ω2, ω3, ω4, ω5]. T Where ω1 is the weighting coefficient corresponding to wind speed, ω2 is the weighting coefficient corresponding to humidity, ω3 is the weighting coefficient corresponding to temperature, ω4 is the weighting coefficient corresponding to air pressure, and ω5 is the weighting coefficient corresponding to the duration of fog.

Claims

1. A fog monitoring method based on edge computing, characterized in that: include: (i) Training a fog early warning model using data mining techniques; (II) Training fog monitoring model; (iii) Evaluate, update and combine the fog patch early warning model and the fog patch monitoring model to form a real-time fog patch calculation model; (iv) Collect environmental data of the monitored area, substitute the environmental data of the monitored area into the real-time fog calculation model, and calculate the current fog monitoring result through the real-time fog calculation model; The method of training a fog early warning model using data mining techniques specifically includes: 1) Collect meteorological data values ​​of the occurrence of fog in the monitored area in history, including wind speed, humidity, temperature, air pressure, and duration of fog; 2) Establish a fog warning index model, as shown in formula (1): （1）； in, R f The fog warning index is [missing information]. V For wind speed, H For humidity, F For temperature, P For air pressure, T Duration of the fog patch; v、 h, f, p, t These are the weight coefficients for each variable; specifically: v The weighting coefficient corresponding to wind speed. h The weighting coefficients corresponding to humidity. f These are the weighting coefficients corresponding to temperature. p The weighting coefficient corresponding to air pressure. t The weighting coefficient corresponding to the duration of the fog patch; 3) Classify the fog warning levels, specifically including: calculating the fog warning index using the fog warning index model. R f The standard deviation classification method was used, with the average value of the fog warning index as the basis. Standard deviation of the fog warning index s Based on this, fog warnings are divided into several fog warning levels.

2. The fog monitoring method based on edge computing according to claim 1, characterized in that: when The fog warning level is Level 1. The fog warning level is Level 2. The fog warning level is level 3. The fog warning level is level 4; the higher the value of the fog warning level, the greater the harm caused by the fog.

3. The fog monitoring method based on edge computing according to claim 1, characterized in that: The weight coefficients corresponding to each variable are determined using the analytic hierarchy process (AHP), specifically including: (1) Construct the judgment matrix; (2) Perform a consistency check on the judgment matrix; (3) Calculate the weight coefficients for each variable.

4. The edge computing-based fog monitoring method according to claim 3, characterized in that: The construction of the judgment matrix specifically includes: selecting wind speed, humidity, temperature, air pressure, and fog duration from meteorological data values ​​as five variables, assigning variable numbers to each of the five variables; and comparing each of the five variables pairwise to obtain the relative importance between different variables. ; ; ; Obtain the judgment matrix A Its mathematical expression is: ; in, express Variable number and The relative importance of the variables; wind speed is set as variable 1, humidity as variable 2, temperature as variable 3, air pressure as variable 4, and fog duration as variable 5.

5. The edge computing-based fog monitoring method according to claim 4, characterized in that: The consistency check of the judgment matrix specifically includes: 2-1) Calculate the judgment matrix A eigenvalues λ max To obtain the consistency index CI And find the average random consistency index RI , ， ； 2-2) Calculate the consistency ratio CR : ； when CR When <0.1, the judgment matrix is ​​considered to be A There is satisfactory consistency; when CR When ≥0.1, the judgment matrix is ​​considered to be... A If satisfactory consistency is not achieved, the relative importance of different variables should be readjusted. The specific value, and then the adjusted Substitute the specific value into the judgment matrix A Repeat steps 2-1 and 2-2 until... CR <0.1 judgment matrix A Continue until satisfactory consistency is achieved.

6. The edge computing-based fog monitoring method according to claim 5, characterized in that: The calculation of the weight coefficients corresponding to each variable specifically includes: 3-1) The judgment matrix A According to column normalization, i.e., the judgment matrix A Divide each element in the matrix by the sum of its columns to generate a normalized matrix. A 5×5 ; 3-2) Normalize the matrix A 5×5 Adding the columns together generates a matrix. A 5×1 ; 3-3) The matrix obtained after addition A 5×1 Divide each element by , The weight vector can then be obtained. ;in, This represents the weighting coefficient corresponding to the wind speed. The value of the weighting coefficient corresponding to humidity. The values ​​of the weighting coefficients corresponding to temperature. This represents the weighting coefficient corresponding to the air pressure. This represents the weighting coefficient corresponding to the duration of the fog patch.

7. The edge computing-based fog monitoring method according to claim 1, characterized in that: The training fog monitoring model specifically includes: 1) Collect historical images of fog patches in the monitored areas to form a fog patch dataset; 2) The fog images in the collected fog dataset are filtered according to certain filtering rules, retaining fog images with clear image quality and clear targets. The fog images in the filtered fog dataset are then subjected to random rotation, random color transformation, random cropping, random scaling, and random flipping to augment the data and form a new fog dataset. 3) Use a visibility meter to mark the visibility value and location coordinates of each fog image in the new fog dataset, and save the visibility value and fog location coordinates in YOLO format; 4) The new fog dataset is randomly divided into training set, validation set and test set according to a certain ratio, and the number of fog images in each training set, validation set and test set is counted respectively; 5) In the PyTorch framework, select the YOLOv5 model, adjust the fog detection parameters, and train the fog detection model using the training set through the YOLOv5 model; 6) Selection of fog monitoring model: After each iteration, the fog monitoring model trained is validated and tested on the validation and test sets in the new fog dataset, and the model with the highest accuracy is selected as the fog monitoring model.

8. The fog monitoring method based on edge computing according to claim 1, characterized in that: The evaluation, updating, and combination of the fog early warning model and the fog monitoring model to form a real-time fog calculation model specifically includes: 1) Calculate the updated weights: L Fog warning model for delegations M The fog monitoring model is represented by a fog early warning model and a fog monitoring model. The two models are then normalized and divided. The ratio of their values ​​is... α This represents updating the weights; 2) Generate a real-time fog calculation model: The fog early warning model and the fog monitoring model are evaluated, updated, and combined according to formula (3) to form a real-time fog calculation model. C : C = Le -α +M (1- e -α ) ，α ∈[0,1] （3）。

Images