CN115795996A - Real network diagram-based city data real-time management method and system of graph neural network - Google Patents
Real network diagram-based city data real-time management method and system of graph neural network Download PDFInfo
- Publication number
- CN115795996A CN115795996A CN202211213714.1A CN202211213714A CN115795996A CN 115795996 A CN115795996 A CN 115795996A CN 202211213714 A CN202211213714 A CN 202211213714A CN 115795996 A CN115795996 A CN 115795996A
- Authority
- CN
- China
- Prior art keywords
- neural network
- data
- sensor
- node
- graph
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 84
- 238000007726 management method Methods 0.000 title claims abstract description 13
- 238000010586 diagram Methods 0.000 title claims abstract description 12
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims abstract description 35
- 230000002159 abnormal effect Effects 0.000 claims abstract description 30
- 238000012544 monitoring process Methods 0.000 claims abstract description 26
- 238000000034 method Methods 0.000 claims abstract description 20
- 230000008447 perception Effects 0.000 claims abstract description 19
- 238000013523 data management Methods 0.000 claims abstract description 5
- 239000013598 vector Substances 0.000 claims description 52
- 238000012549 training Methods 0.000 claims description 34
- 230000006870 function Effects 0.000 claims description 25
- 238000012360 testing method Methods 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 17
- 238000012795 verification Methods 0.000 claims description 16
- 238000011176 pooling Methods 0.000 claims description 14
- 230000017105 transposition Effects 0.000 claims description 14
- 230000007246 mechanism Effects 0.000 claims description 10
- 230000002441 reversible effect Effects 0.000 claims description 9
- 238000009826 distribution Methods 0.000 claims description 8
- 230000005856 abnormality Effects 0.000 claims description 7
- 239000012634 fragment Substances 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000004931 aggregating effect Effects 0.000 claims description 6
- 238000012217 deletion Methods 0.000 claims description 6
- 230000037430 deletion Effects 0.000 claims description 6
- 238000005259 measurement Methods 0.000 claims description 6
- 238000002372 labelling Methods 0.000 claims description 5
- 238000003860 storage Methods 0.000 claims description 5
- 238000009434 installation Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 239000003086 colorant Substances 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 239000007789 gas Substances 0.000 claims 4
- 230000003213 activating effect Effects 0.000 claims 1
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical group C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 description 10
- 238000005457 optimization Methods 0.000 description 6
- 230000002776 aggregation Effects 0.000 description 3
- 238000004220 aggregation Methods 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 239000000809 air pollutant Substances 0.000 description 2
- 231100001243 air pollutant Toxicity 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007664 blowing Methods 0.000 description 1
- 210000000476 body water Anatomy 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001010 compromised effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000006062 fragmentation reaction Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000001932 seasonal effect Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
Images
Landscapes
- Testing Or Calibration Of Command Recording Devices (AREA)
Abstract
The invention relates to a real network diagram-based city data real-time management method and a system thereof of a neural network, wherein the method comprises the following steps: s1, establishing a 3D model of an urban sensor real network; s2, establishing a 3D graph neural network according to the 3D model established in the S1; s3, transmitting multi-protocol data represented by various sensors in the S1 to a server, constructing an attention model based on a 3D (three-dimensional) graph neural network, and establishing a perception model of data abnormity; and S4, the server utilizes the perception model established in the S3 to realize abnormal monitoring of the multi-protocol data generated by various sensors. The efficient and accurate multi-protocol big data management of atmospheric temperature and ground temperature, atmospheric humidity and road air humidity, water level, water quality, harmful gas leakage, automobile exhaust, local temperature and humidity indication in cities is realized.
Description
Technical Field
The invention relates to a real-time access method of urban data, in particular to a method for realizing the access and the visit management of all sensing data in a city by using a graph neural network algorithm based on a real network graph, belonging to the field of artificial intelligence processing and analysis of big data.
Background
The city as a space carrier of life relates to the most concerned parameter indexes such as temperature, humidity, water level of water body, water quality, air pollutant indexes, electric energy, gas energy, water consumption and the like in the essential elements of daily life. These parameter indices are particularly governed by a number of government functions or social units (including institutions, enterprise buildings, cells, buildings). Unified management of data cannot be realized, and when a user needs to know the data, the user needs to disperse energy to a plurality of platforms to access and inquire so as to obtain or download recorded data, thereby realizing statistical research and providing guiding significance for life production.
On the other hand, the sensors on which the data depend, particularly the parameters of air temperature and humidity are not required or can be measured by other water body water levels, water quality, air pollutant indexes, electric energy, gas energy and water consumption, and the measurement modes are quite different due to the fact that the measurement is realized by professional departments or units for a long time. Any natural human individual can detect the temperature and humidity in as little as one room by having a thermometer, a hygrometer or a thermo-hygrometer. However, the prior art only focuses on the air temperature and humidity of the atmospheric environment, and does not specifically detail the air temperature and humidity of local places in various activities of people in each building and any critical occasions. The body feeling of people usually depends on specific temperature and humidity in the environment around the human body, and the temperature and humidity reference significance of the whole city does not very accurately reflect the body feeling of people. When people need to know the temperature and the humidity of a place where the people are located, the temperature and humidity meter is generally checked, so that whether the temperature and/or the humidity need to be changed or not is known or determined, for example, monitoring and changing under the requirement of keeping the temperature and the humidity in a laboratory and a factory building often need respective independent adjusting means, and unified adjustment cannot be realized among different social units.
Since monitoring of the parameter is often not unusual from a single data point, it is assumed that an anomaly has occurred. For example, due to the ground temperature difference caused by direct sunlight and tree shading, it cannot be considered that the place where the direct sunlight is generated has an abnormally high temperature, but the whole performance of the machine learning area in the specified seasonal period of the longitude and latitude of the city can be used for correctly judging whether the ground temperature is abnormal in the concerned period. Therefore, if simple sensor data is combined with the positioning coordinates as a method for determining whether the parameters of each location are abnormal, a great probability may cause erroneous determination. Therefore, how to monitor the abnormal occurrence in real time from the whole and the specific place needs to be embodied by a new algorithm.
Therefore, in order to solve the problem of the smart city, how to integrate the data is considered, and the key data corresponding to the most directly sensed parameters around the human body is actually reflected, and the management of the data becomes a problem to be solved in the internet of things structure.
The sensors for measuring the parameters in the city have the characteristic of fixed positions, and the fixed sensors stably exist in the geographic coordinates of the city, so that a node dot matrix graph is formed in the city, and a real network graph is formed. When the nodes belong to the same concerned region or the same directly-affiliated social unit, the graph neural network is formed when the connection edges between the unit nodes represent the concerned or directly-affiliated logical relationship. Therefore, the prior art only belongs to an abstract relational network and is not a space-time network although considering the relationship formed by nodes and edges in the graph neural network. It is therefore not considered that a real network graph can be essentially viewed as a real version of a graph neural network.
Disclosure of Invention
In order to solve the above problems, the prior art has provided macroscopic monitoring means for monitoring leakage of atmospheric temperature, atmospheric humidity, water level, water quality and harmful gas, and water, electricity and gas monitoring means represented by electric energy, gas energy and water consumption, sensors corresponding to these means have relatively fixed positions and belong to public monitoring data, and microscopic temperature and humidity data do not realize real-time statistics, and two parameters are indexes capable of most reflecting actual human indoor body feeling. Therefore, the following two aspects are considered to carry out real-time access, access and management on all the data, and the intelligent unified management method for the big data is realized. Firstly, a 3D model of a real network of an urban sensor is constructed, secondly, a 3D graph neural network is constructed based on the 3D model, and protocol data of various sensors are accessed uniformly, so that multi-protocol data are obtained in real time, and thirdly, the graph neural network is used for realizing abnormal monitoring of the multi-protocol data. It should be noted that anomalies of the present invention are not merely data that are out of acceptable range or that would be in the potential for an accident or loss, but also data that deviate from the desired range of values and require timely correction to within the normal or desired range. The present invention with a sensor indicating local temperature, humidity means that such sensor is capable of detecting the temperature and/or humidity of the local location.
In order to suggest the above consideration, in one aspect, the present invention provides a city data real-time management method for a graph neural network based on a real network diagram, which is characterized by comprising the following steps:
s1, establishing a 3D model of an urban sensor reality network;
s2, establishing a 3D graph neural network according to the 3D model established in the S1;
s3, multi-protocol data represented by various sensors in the S1 are transmitted to a server, an attention model is built based on a 3D (three-dimensional) graph neural network, and a perception model of data abnormity is built;
and S4, the server realizes abnormal monitoring of the multi-protocol data generated by various sensors by using the perception model established in the S3.
With respect to S1
S1-1, an urban positioning system (CCS) is arranged in various sensors for monitoring or detecting atmospheric temperature, ground temperature, atmospheric humidity, air humidity on a road section, water level, water quality, harmful gas leakage and automobile exhaust, the CCS carries out accurate sensor positioning according to a Beidou satellite system to obtain the positions of various sensors, or the positions of the various sensors in the city or the places near the various sensors (such as 50cm-5 m) are measured by positioning personnel, and the measuring method comprises the following steps: the positioning personnel measures whether the distance between the optional point on the surface of the sensor and the optional point detected by the laser range finder is between 50cm and 5m or not through the laser range finder, and if yes, the positioning personnel acquires the position of the positioning personnel through Beidou satellite positioning and uses the position of the positioning personnel as the measured position of the sensor;
it can be understood that the CCS measuring method needs to install CCS which can communicate with the Beidou satellite to obtain positioning in various sensors, the position is accurate, positioning modification can be timely obtained along with the change of the installation position of the sensors, and people are needed to walk for measurement of positioning personnel, so that the measurement is troublesome, the CCS does not need to be installed, the product cost is reduced, and the labor cost is increased. Both of them are good and bad. Set up ground temperature sensor and cool down in order to guide the watering lorry in summer, reduce the vehicle danger of blowing out.
S1-2, arranging a humidity sensor with local temperature indication in an item optionally specified by a user, and arranging a CCS in or near the humidity sensor with local temperature indication (local) temperature indication (such as 50cm-5 m), wherein the item comprises but is not limited to a mobile terminal (including a smart phone and a tablet computer), a television, a computer display, a wall body in a distance range of 50cm-5m near an illuminating device and/or an air conditioner, a handheld portable temperature and humidity detection device;
s1-3, constructing an urban 3D model;
wherein S1-3 specifically comprises the following steps:
s1-3-1, collecting remote sensing satellite images of a plurality of specified cities, artificially labeling road profiles of roads in the remote sensing satellite images and filling the interior of the roads with colors to form artificial labeling image layers;
s1-3-2, dividing a remote sensing satellite map with an overlapped part with an artificial mark map layer into fragments with uniform size of 225 x 225-1000 x 1000, and dividing a plurality of most fragments formed by the remote sensing satellite maps of a plurality of cities into a training set and a verification set, wherein the proportion is 5-3;
s1-3-3, inputting the training set into a D-LinkNet encoder, performing 8 x 8 convolution, with stride of 2, maximizing the three-order residual after pooling, performing the maximal pooling on the result to continue the four-order residual, performing the maximal pooling on the result to continue the six-order residual, performing the maximal pooling on the result to continue the three-order residual, and entering the middle part after the last three-order residual after the maximal pooling on the result;
s1-3-4 enters the middle part of the D-LinkNet, and then is subjected to cavity convolution stacked in a cascade mode to increase the receptive field of the characteristics of the central part of the network and retain detailed information, wherein the cascade mode comprises the steps of continuously performing convolution with the expansion value of 1,2,4,8, the expansion value of 1,2,4 and the expansion value of 1,2 and the expansion value of 1 by 4, and adding the three-order residual results without the expansion convolution into a decoder;
in the S1-3-5D-LinkNet decoder, performing double sampling on the result after summation by adopting a transposition convolutional layer, performing double sampling on the result after summation with a six-order residual error result by continuously adopting the transposition convolutional layer, then continuously performing double sampling on the result after summation with a four-order residual error result by continuously adopting the transposition convolutional layer, and then continuously performing double sampling on the result after summation with a three-order residual error result and finally performing double sampling by adopting the transposition convolutional layer;
s1-3-6, performing 5 x 5 transposition convolution on a double sampling result obtained by adopting a transposition convolution layer at the last time, performing 4 x 4 convolution with an expansion value of 1, outputting an identification result through an SOFTMAX function, verifying the correct rate and obtaining a loss function value by using a verification set, adjusting D-LinkNet network parameters through back propagation until the correct rate and the loss function value are stable, and finishing training; during specific calculation, fragmenting the remote sensing satellite images to be recognized in the same size, sequentially inputting the fragmentized remote sensing satellite images into D-LinkNet to form road recognition, and sequentially splicing and restoring the road recognition according to the input sequence to form a road recognition result, wherein the road comprises a motor vehicle road, a non-motor vehicle road, a pedestrian road, a river, a natural lake and an artificial lake; the pedestrian roads include roads which can be traveled by people or road vehicles or work tasks (such as wheeled vehicles, fire trucks, ambulances, police vehicles and the like) in building group areas such as streets in cities, pedestrian roads beside non-motor vehicle lanes, cells or factory buildings and the like.
The loss function adopts the difference between the area of the road part of the recognition result and the area of the road part in the artificial marking map layer, so that the accuracy of the predicted boundary contour is obtained through the difference. If the inaccurate road profile can generate the situation that the sensor is mistakenly identified as other types of sensors, for example, the tail gas sensor on the road can be mistakenly mutually serialized with the pavement temperature sensor of the opposite road, so that the access of data with different protocols from the target detection data fails, and further abnormal false alarm is caused with a certain probability.
In the conventional RNN, the identification of the road needs to be carried out by adopting two steps of node connection and road path widening, and because the widening is roughly specified for different road type standards, the method cannot be accurately applied to occasions needing relatively accurate positioning of sensors. When the widening is not proper, the sensor is likely to be identified incorrectly, so that the multiprotocol data is mutually serialized. The work of manual broadening is transferred to manual road marking in the training period, and more accurate road identification model establishment is achieved on the premise that the work total amount is not changed obviously.
S1-3-7, based on a training set formed by remote sensing satellite maps of multiple cities in S1-3-1, extracting a series of feature maps obtained by using a VGG-16 algorithm without an added layer as a CNN main network, wherein the feature maps are 1/2-1/10, preferably 1/8, of the size of an input image;
meanwhile, a characteristic pyramid is constructed by using different layers of a CNN main network through an image pyramid algorithm FPN, and the frames of a plurality of buildings are predicted,
s1-3-8, for each building in a plurality of buildings, obtaining a local feature map F of the building by using a RoIAlign algorithm on the feature maps obtained by the series of different convolutional layers and the corresponding frame of the building;
s1-3-9, performing convolutional layer processing on the local characteristic diagram F of each building to form a polygonal boundary cover M, then performing convolutional layer processing to form P predicted vertexes of the boundary cover M, and optimizing a CNN network model by using the position mean square error between the CNN network model and the real vertexes as a judgment standard until the mean square error is smaller than a preset value, wherein the CNN network model and a trained D-LinkNet model form a real network 3D model together.
With respect to S2
S2-1, sequentially dividing various sensors of atmospheric temperature, ground temperature, atmospheric humidity, road air humidity, water level, water quality, harmful gas leakage and automobile exhaust into eight regions, namely an atmospheric temperature map region (namely a first map region, and sequencing the regions), a ground temperature map region, an atmospheric humidity map region, a road air humidity map region, a water level map region, a water quality map region, a harmful gas leakage map region and an automobile exhaust map region;
setting a sensor with local temperature and humidity as a ninth region;
s2-2, edges (representing monitored and monitored relations) of each sensor node-server node in each graph area are formed inside each graph area, and a sub-graph neural network inside the graph areas is formed; therefore, a double graph network with the spatial distribution of sensors and the relation distribution represented by a sub-graph neural network among nodes is formed, the sub-graph neural network in each graph area is used as a graph neural network layer to monitor each parameter respectively, all the graph neural network layers form a 3D graph neural network, and the data access, access and abnormal monitoring of each type of sensor are realized to realize multi-protocol data management.
With respect to S3
For each graph neural network layer, transmitting multi-protocol data represented by various types of sensors in S1 to a server at regular time intervals to form a data set, dividing the data set into a training set Train = { (u, yu) }, a verification set Val = { (u ', yu') }, and a prediction anomaly probability Test set Test = { (u ", yu") }, wherein the ratio of the three data is 1-1, yu ', yu "respectively represents the labels of nodes u, u', u", and the data of all the three data has data of corresponding sensor sequence numbers and data acquisition times, and an attention model is constructed based on each graph neural network layer, and the steps are as follows:
s3-1, aggregating the server nodes in a specified period, specifically comprising:
s3-1-1 for each sensor and server within the neural network layer of the graph, a node pair is formed (i) Nk ,j k ) Wherein N represents the sequential number of the sensors in the neural network layer, k belongs to X represents the number of the servers serving one neural network layer, X is the number set of the servers, the number of elements is the total number 9 of all the servers, and the importance is defined as
This means that the server node j k For target sensor node i Nk Measure of importance of. Node pair (i) based on the edge to which it belongs in S2-2 as a path Nk ,j k ) The importance formula of (c) can be expressed as follows:
wherein
Respectively representing each sensor node i in the neural network layer of the graph Nk And server node j k The multi-protocol data in the data set is used as one of elements in the vector to represent, and a vector with a certain dimension (4-8 dimensions) is formed, A node (. Is a deep neural network that performs a nodal attention mechanism; it is understood that other elements in the vector may be represented by the time of data acquisition (i.e., the time point obtained by dividing the predetermined time period by the predetermined time interval) and the sensor node sequence number, so that the time when the anomaly is sensed and the location from which the sensing data is transmitted are traced back in the subsequent sensing result, and the higher the importance is, the higher the doubtful nature of the anomaly is.
In fact, all server nodes j given the same one of said edges k Are sampled by the same connection pattern, so that for a given said edge (i.e. a sensor within a given said graph neural network layer), all edge-based node pairs share a node (·)。
S3-1-2 predicted by the training set Train through the step S3-1-1
Verifying accuracy using a verification set Val, back-propagating using a cross-entropy loss function, training A node (. O) until the accuracy stabilizes and the cross entropy loss function reaches a minimum;
s3-1-3 importance pairs Using the softmax function
Are normalized into
The following:
s3-1-4, aggregating the feature vector of the node i based on the edge b with the corresponding coefficient through the projection feature of the neighbor as follows:
wherein
Representing a node j k A feature vector learned on the edge, σ being the activation function LeakyReLU;
s3-1-5 node attention mechanism A node (. K is 1-5) times and concatenating each learned vector representation into the graph neural network layer for all edge sets { b } 1k ,b 2k ,b 3k ，…,b Nk Treated as a given type of path b (i.e. sensor-server), vector features for all sensor nodes
Here, we realize the acquisition of attention feature vectors in a layer of the neural network of the graph only through a node level aggregation in the layer according to the structural characteristics of a specially constructed 3D neural network, so as to obtain input vectors for realizing abnormal perception prediction in the layer, and the specific perception mechanism is as follows:
s3-2 for eachA graph neural network layer, which uses a training set Train to obtain a plurality of corresponding characteristic vectors p according to S3-1-1-S3-1-5 bk Inputting k belongs to X into a multilayer perceptron (MLP), training the MLP by taking whether the corresponding abnormality is abnormal or not as classification, verifying by using a verification set Val to obtain the accuracy acc, and calculating the cross entropy as a loss function
MLP network parameters are optimized by back propagation until acc stabilizes and loss is minimized, where y L Is a set of node indexes with labels, Y and
labels and vector features p, respectively, of nodes bk W is the classifier's parameters in MLP, and g is superscript to indicate
The middle summation item corresponds to the sensor node.
Preferably, the specified time interval is 1s-1min, and the specified period is 1 week-1 year.
Preferably, according to different specified periods T ∈ [4,12 ]]A plurality of attention systems { A } corresponding to different predetermined periods are formed node1 (·)，A node2 (·)，…，A nodeT (-) and a system of perceptual models { MPL } 1 ,MPL 2 ,…,MPL T }。
It is worth emphasizing that the server node is attentive to the mechanism A node Training separately for (t) and MPL, can avoid calibration competition between network parameters. This is because, if A is to be used node If the (. Cndot.) and MPL are optimized as a whole, then in separate optimization, a suitable A is optimized node A when the parameter may not be an overall optimization node (. Cndot.) corresponding optimization results, as well as for MPL. Thus, optimization of the parameters of one is actually compromised to some extent by the other, resulting in less than optimal ideal for both models as a whole. That is, if a back propagation of the loss function at the MLP end is used at A node Adjusting network parameters simultaneously in (-) and MPL does not solve the problem of co-optimization of parameters of both models. For this we use a split training scheme.
With respect to S4
S4-1, continuously collecting data and updating a data set;
s4-2, collecting Test data of the Test set in a preset time period (1 day to 1 year), and obtaining a feature vector to be predicted according to S3-1-1-S3-1-5
Since data is transmitted at predetermined time intervals, the obtained feature vector corresponds to the time point.
S4-3 will
Inputting the training MPL to obtain the prediction result, if the MPL is abnormal, the abnormal time is found in real time in a preset time period, and according to the corresponding feature vector
And reversely checking the characteristic vectors in each input test set, finding the marked places in the abnormal characteristic vectors, and inquiring the places corresponding to the suspected abnormal data so as to further eliminate the places corresponding to the abnormal data determined on the basis of the sensor abnormality.
Preferably, the reverse check is realized by temporarily storing the input test set node vectors in the server, the input vectors in the temporary storage are automatically deleted when the data is normal, and the deletion operation is not performed once the data is abnormal, so as to supply the reverse check until the reverse check and the deletion after the vectors are stored.
Preferably, the attention system { A } is derived for different prescribed time periods node1 (·)，A node2 (·)，…，A nodeT (-) and a perception model system { MPL } 1 ,MPL 2 ,…,MPL T S4-2-S4-3 is also performed in each of the specified time periods.
The invention also provides a system for realizing the city data real-time management method of the neural network based on the graph of the real network diagram, which is characterized by comprising nine types of sensors for monitoring or detecting the atmospheric temperature and the ground temperature, the atmospheric humidity and the air humidity on a road section, the water level, the water quality, the harmful gas leakage, the automobile exhaust and indicating the local (local) temperature and humidity, wherein each type of sensor comprises at least one sensor, nine servers, and each server serves each type of sensor and is used for collecting data of different protocols of each type of sensor to form a data set; wherein,
the server is also used for constructing and establishing a 3D model of the urban sensor real network, establishing a 3D graph neural network based on the established 3D model, establishing an attention model based on the 3D graph neural network, establishing a perception model of data abnormity, and realizing abnormity monitoring of multi-protocol data sent by various sensors by utilizing the established perception model.
A third aspect of the present invention provides a non-transitory storage medium, in which a computer readable program executable by the server to implement the real-time city data management method for the graph neural network based on the real network diagram is stored.
Advantageous effects
1. The method of D-LinkNet coding and decoding and auxiliary artificial marking is adopted to accurately identify the range of the road, and a 3D space model basis is provided for the distribution of accurate sensors;
2. the concepts of the graph network and the graph neural network are combined, the reality graph network is related, and the graph neural network is embodied, so that the access and the visit of multi-protocol data and the monitoring of the time-space distribution of data abnormity are realized;
3. based on various sensors for monitoring parameters under different scenes, a graph neural network layer is formed, an attention model and a perception model of data abnormity of various protocols are trained separately based on the accumulation of multi-protocol data sets in the graph neural network layer, so that the optimization of model parameters is more perfect,
4. and monitoring is respectively carried out according to the types of the multi-protocol data, only one-level node-level aggregation is carried out in the aggregation aspect, and the model is more simplified and also conforms to the structure of a 3D (three-dimensional) graph neural network.
5. The beneficial effects 3 and 4 of the invention are that the high-efficiency and accurate management of the multi-protocol big data is realized on the whole.
Drawings
FIG. 1 is a flow chart of the construction of a real network 3D model for urban road and building identification in embodiment 1 of the present invention,
FIG. 2 is a flow chart of the construction of a multi-protocol data forming data set for attention model and perception model MLP of data abnormality in city according to embodiment 2 of the present invention,
FIG. 3 is a partial schematic diagram of a 3D model of a real network based on the attention model in FIG. 2 and the construction of a perception model MLP of data anomaly, and a schematic diagram of a 3D neural network structure of corresponding three regions A, B and I,
fig. 4 is a flowchart illustrating an anomaly monitoring process of the server according to the multi-protocol data generated by the nine types of sensors by using the sensing model MLP established in S3 in embodiment 4 of the present invention.
Detailed Description
Example 1
This embodiment explains step S1.
S1-1, an urban positioning system (CCS) is arranged in various sensors for monitoring or detecting atmospheric temperature, ground temperature, atmospheric humidity, air humidity on road sections, water level, water quality, harmful gas leakage and automobile exhaust, the CCS carries out accurate sensor positioning according to a Beidou satellite system so as to obtain the positions of various sensors, or the positions of various sensor installation places or places within a 1m range in a city are obtained according to measurement of positioning personnel, and the measuring method comprises the following steps: the positioning personnel measures whether the distance between the optional point on the surface of the sensor and the optional point detected by the laser range finder is within 1m or not through the laser range finder, and if so, the positioning personnel acquires the position of the positioning personnel through Beidou satellite positioning and uses the position of the positioning personnel as the position of the measured sensor;
s1-2, arranging a humidity sensor with local temperature indication in a wall body within 1m range near an air conditioner, and arranging a CCS in the humidity sensor with local temperature indication or 1m near the humidity sensor;
s1-3, constructing a city 3D model.
As shown in fig. 1, S1-3 specifically includes the following steps:
s1-3-1, collecting remote sensing satellite images of a plurality of specified cities, artificially labeling road profiles of roads in the remote sensing satellite images and filling the interior of the roads with colors to form artificial labeling image layers;
s1-3-2, dividing the remote sensing satellite images with the overlapped parts with the artificial marking image layers into fragments with uniform size of 500 x 500, and dividing a plurality of fragments formed by the remote sensing satellite images of a plurality of cities into a training set and a verification set, wherein the proportion is 3;
s1-3-3, inputting the training set into a D-LinkNet encoder, performing 8 x 8 convolution, with stride of 2, maximizing the three-order residual after pooling, performing the maximal pooling on the result to continue the four-order residual, performing the maximal pooling on the result to continue the six-order residual, performing the maximal pooling on the result to continue the three-order residual, and entering the middle part after the last three-order residual after the maximal pooling on the result;
s1-3-4 enters the middle part of the D-LinkNet and is subjected to cavity convolution stacked in a cascade mode to increase the receptive field of the characteristics of the central part of the network and retain detailed information, wherein the cascade mode comprises the steps of continuously performing convolution with the expansion value of 1,2,4,8, continuously performing convolution with the expansion value of 1,2,4 multiplied by 4 with the expansion value of 1 and continuously performing convolution with the expansion value of 1, and adding the three-order residual results without performing the expansion convolution into a decoder;
in the S1-3-5D-LinkNet decoder, performing double sampling on the result after summation by adopting a transposition convolutional layer, performing double sampling on the result after summation with a six-order residual error result by continuously adopting the transposition convolutional layer, then continuously performing double sampling on the result after summation with a four-order residual error result by continuously adopting the transposition convolutional layer, and then continuously performing double sampling on the result after summation with a three-order residual error result and finally performing double sampling by adopting the transposition convolutional layer;
s1-3-6, performing 5 × 5 transposition convolution on the result of performing double sampling by adopting the transposition convolution layer for the last time, performing 4 × 4 convolution with an expansion value of 1, outputting an identification result through an SOFTMAX function, verifying the accuracy and obtaining a loss function value by using a verification set, and adjusting D-LinkNet network parameters by back propagation until the accuracy and the loss function value are stable, and finishing training; during specific calculation, fragmentations with the same size are carried out on the remote sensing satellite images to be identified, the fragmented numbers 1 and 2 and the second lines x and x +1 in figure 1 are sequentially input into the D-LinkNet to form road identification, and then the fragments are sequentially spliced according to the input sequence to restore to form road identification results.
Continuing to refer to FIG. 1, S1-3-7 uses a VGG-16 algorithm without an added layer as a CNN backbone network to extract a series of feature maps obtained by different convolutional layers based on a training set formed by remote sensing satellite maps of a plurality of cities in S1-3-1, wherein the feature maps are 1/8 of the size of an input image;
meanwhile, a characteristic pyramid is constructed by using different layers of a CNN main network through an image pyramid algorithm FPN, and the frames of a plurality of buildings are predicted,
s1-3-8, for each building in a plurality of buildings, obtaining a local feature map F of the building by using a RoIAlign algorithm on the feature maps obtained by the series of different convolutional layers and the corresponding frame of the building;
s1-3-9, for the local feature map F of each building, convolutional layer processing is adopted to form a polygonal boundary cover M, P predicted vertexes of the boundary cover M are formed by utilizing the convolutional layer processing, a CNN network model is optimized by taking the position mean square deviation between the CNN network model and a real vertex as a judgment standard until the mean square deviation is smaller than a preset value, and the CNN network model and a trained D-LinkNet model form a real network 3D model.
Example 2
This embodiment will explain step S2
S2-1, the sensors of the atmospheric temperature A, the ground temperature B, the atmospheric humidity C, the air humidity D on a road section, the water level E, the water quality F, the harmful gas leakage G and the automobile exhaust H are respectively and sequentially divided into eight regions, and a ninth region (shown in figure 2) is set as the sensor I with the local temperature and humidity indication, so that nine types of multi-protocol data are formed.
S2-2 is specifically as shown in fig. 3, a local schematic view of a real network 3D model formed in embodiment 1, where six sensors of ground temperature are set on a road and numbered with numbers 1-6, building 1 shows three temperature and humidity sensors at three layers bcd, and building 2 is set with an atmospheric temperature sensor, so that three regions, i.e., a region a, a region B, and a region I, are formed.
Edges of sensor nodes-server nodes in each graph area, namely a server A, a server B and a server I, are formed inside each graph area, and a sub-graph neural network inside the graph area is formed respectively with a node a, nodes 1-6, nodes B, c and d; therefore, a double graph network with the spatial distribution of sensors and the relation distribution represented by a sub-graph neural network among nodes is formed, the sub-graph neural network in each graph area is used as a graph neural network layer to monitor each parameter respectively, all the graph neural network layers form a 3D graph neural network, and the data access, access and abnormal monitoring of each type of sensor are realized to realize multi-protocol data management.
Example 3
This embodiment will explain step S3
As shown in fig. 2, for each of four seasons in a year, for each of the neural network layers of example 2, multi-protocol data represented by each type of sensor in S1 is transmitted to a server at regular time intervals of 30S to form a data set, the data set is divided into a training set Train = { (u, yu) }, a verification set Val = { (u ', yu') }, and a prediction anomaly probability Test set Test = { (u ", yu") }, the ratio of the training set Train to the verification set Val = { (u, yu) } is 1, yu ', yu "represents the labels of nodes u, u', u", respectively, and the data of the three sets have corresponding sensor sequence numbers and data acquisition times, and an attention model is constructed based on each of the neural network layers, and the following steps are performed:
s3-1, aggregating the server nodes in a specified period, specifically comprising:
s3-1-1 for each sensor and server within the neural network layer of the graph, a node pair is formed (i) Nk ,j k ) Wherein N represents the sequential number of the sensors in the neural network layer, k belongs to X represents the number of the servers serving one neural network layer, X is the number set of the servers, the element number is the total number 9 of all the servers, and the importance is defined as
This indicates that server node j k For target sensor node i Nk Is measured. Node pair (i) based on the edge to which it belongs in S2-2 as a path Nk ,j k ) The significance formula of (a) can be expressed as follows:
wherein
Respectively representing each sensor node i in the neural network layer of the graph Nk And server node j k Using the multi-protocol data in the dataset as one of the elements in the vector to form a 4-dimensional (not all shown) vector, a node (. Cndot.) is a deep neural network that performs a node attention mechanism, all node pairs that share A based on edges including those exemplarily given in embodiment 2 node (·)。
The other elements in the vector are represented by the time of data acquisition, that is, the time point obtained by dividing the predetermined one season in units of the predetermined time interval 30s, and the sensor node sequence number.
S3-1-2 predicted by the training set Train through the step S3-1-1
Verifying accuracy using a verification set Val, back-propagating using a cross-entropy loss function, training A node (. O) until the accuracy stabilizes and the cross entropy loss function reaches a minimum;
s3-1-3 uses the softmax function to significance
Is normalized to
The following:
s3-1-4, aggregating the feature vector of the node i based on the edge b with the corresponding coefficient through the projection feature of the neighbor as follows:
wherein
Representing a node j k Feature vectors learned on the edge;
s3-1-5 node attention mechanism A node (. The) repeat 3 times, and splice each learned vector representation into the graph neural network layer, for all edge sets { b } 1k ,b 2k ,b 3k ，…,b Nk Treated as a given type of path b (i.e. sensor-server), for fig. 3 the ABI trigram areas are a, 1,2,3,4,5,6, b, c, d, respectively, and the vector characteristics for all sensor nodes
The specific perception mechanism is as follows:
as shown in FIG. 2, S3-2 uses the training set Train to obtain a plurality of corresponding feature vectors p according to S3-1-1-S3-1-5 for each graph neural network layer bk Inputting k belongs to X into a multilayer perceptron (MLP), training the MLP by taking whether the corresponding anomaly exists as classification, verifying by using a verification set Val to obtain the accuracy acc, and calculating the cross entropy as a loss function
MLP network parameters are optimized by back propagation until acc stabilization and loss are minimized, where y L Is a set of node indexes with labels, Y and
labels and vector features p of nodes, respectively bk W is the classifier's parameters in MLP, g is superscript to indicate
The middle summation item corresponds to the sensor node.
The same execution of the above steps according to different predetermined periods T =4 forms 4 { a corresponding to different seasons (spring, summer, autumn and winter) { a } node1 (·)，A node2 (·)，A node3 (·)，A node4 (. To) } and { MPL 1 ,MPL 2 ,MPL 3 ,MPL 4 }。
Example 4
This embodiment will explain step S4
As shown in fig. 4, S4-1 continuously collects data and updates the data set;
s4-2 within 1 day, collecting Test data of the Test set, and obtaining the feature vector to be predicted according to the S3-1-1-S3-1-5 of the embodiment 3
S4-3 will
Inputting the training MPL to obtain a prediction result, obtaining four continuous normal results of t1, t2, t3 and t4, and determining the result as abnormal when monitoring the t5 th result.
Then obtaining the time for finding the abnormality at the t5 moment in the day in real time and according to the corresponding characteristic vector
Reverse checking out the characteristic vector in the test set of each input to find out abnormal characteristicsAnd inquiring the position corresponding to the suspected abnormal data by the position marked in the vector so as to further exclude the position corresponding to the abnormal data determined on the basis of the sensor abnormality.
The reverse check is realized by temporarily storing the input test set node vectors in the server, the input vectors in the temporary storage are automatically deleted when the data are normal, and the deletion operation is not carried out once the data are abnormal, so that the reverse check is carried out until the reverse check and the deletion after the vectors are stored.
Obtained for different prescribed periods of time { A node1 (·)，A node2 (·)，…，A nodeT (·) } and { MPL } 1 ,MPL 2 ,…,MPL T S4-2-S4-3 is also performed during each of the specified time periods.
Example 5
This embodiment will explain the implementation of the system, and a system capable of implementing the city data real-time management method based on the neural network of the graph of the real network diagram of the above embodiments 1 to 4 is characterized in that the system comprises nine types of sensors for monitoring or detecting the atmospheric temperature and the ground temperature, the atmospheric humidity and the air humidity on the road, the water level, the water quality, the harmful gas leakage, the automobile exhaust gas, and the local (local) temperature and humidity, and at least one sensor for each type, nine servers, and each server serves each type of sensor, and is used for collecting the data of different protocols of each type of sensor to form a data set; wherein,
the server is also used for constructing and establishing a 3D model of the urban sensor real network, establishing a 3D graph neural network based on the established 3D model, establishing an attention model based on the 3D graph neural network, establishing a perception model of data abnormity, and realizing abnormity monitoring of multi-protocol data sent by various sensors by utilizing the established perception model.
A partial system configuration is schematically shown in fig. 3.
Claims (10)
1. A city data real-time management method of a graph neural network based on a real network graph is characterized by comprising the following steps:
s1, establishing a 3D model of an urban sensor reality network;
s2, establishing a 3D graph neural network according to the 3D model established in the S1;
s3, transmitting multi-protocol data represented by various sensors in the S1 to a server, constructing an attention model based on a 3D (three-dimensional) graph neural network, and establishing a perception model of data abnormity;
and S4, the server utilizes the perception model established in the S3 to realize abnormal monitoring of the multi-protocol data generated by various sensors.
2. The method according to claim 1, characterized in that S1 comprises in particular the steps of: s1-1, an urban positioning system (CCS) is arranged in various sensors for monitoring or detecting atmospheric temperature, ground temperature, atmospheric humidity, air humidity on road sections, water level, water quality, harmful gas leakage and automobile exhaust, the CCS carries out accurate sensor positioning according to a Beidou satellite system so as to obtain the positions of various sensors, or the positions of the installation places of various sensors or the places of 50cm-5m of the installation places of various sensors in a city are obtained according to the measurement of positioning personnel, and the measuring method comprises the following steps: the positioning personnel measures whether the distance between the optional point on the surface of the sensor and the optional point detected by the laser range finder is 50cm-5m or not through the laser range finder, and if so, the positioning personnel acquires the position of the positioning personnel through Beidou satellite positioning and uses the position of the positioning personnel as the measured position of the sensor;
s1-2, arranging a humidity sensor with local temperature indication in an item optionally specified by a user, and arranging a CCS in or near the humidity sensor with local temperature indication (local), such as 50cm-5m, wherein the item comprises a mobile terminal, a television, a computer display, a wall body in a distance range of 50cm-5m near an illuminating device and/or an air conditioner, a handheld portable temperature and humidity detection device;
s1-3, constructing a city 3D model.
3. The method according to claim 2, wherein S1-3 specifically comprises the steps of:
s1-3-1, collecting remote sensing satellite images of a plurality of specified cities, artificially labeling road contours and filling colors in the remote sensing satellite images to form artificially labeled image layers;
s1-3-2, dividing a remote sensing satellite map with an overlapped part with an artificial mark map layer into fragments with uniform size of 225 x 225-1000 x 1000, and dividing a plurality of most fragments formed by the remote sensing satellite maps of a plurality of cities into a training set and a verification set, wherein the proportion is 5-3;
s1-3-3, inputting the training set into a D-LinkNet encoder, carrying out 8 multiplied by 8 convolution, obtaining stride of 2, carrying out three-order residual error after maximum pooling, carrying out maximum pooling on the result, continuing to obtain four-order residual error, carrying out maximum pooling on the result, continuing to obtain six-order residual error, carrying out maximum pooling on the result, carrying out last three-order residual error, and entering the middle part;
s1-3-4 enters the middle part of the D-LinkNet, and then cavity convolution is stacked in a cascade mode to increase the receptive field of the characteristics of the central part of the network and retain detailed information, wherein the cascade mode comprises that expansion values of 1,2,4 and 8, expansion values of 1,2 and 4 multiplied by 4 of the expansion value of 1 are continuously carried out respectively, and the three-order residual results without the expansion convolution are added together and enter a decoder;
in the S1-3-5D-LinkNet decoder, performing double sampling on the summed result by adopting a transposed convolution layer, performing double sampling on the summed result by continuously adopting the transposed convolution layer with the sum of the sixth-order residual error result, performing double sampling on the summed result by continuously adopting the transposed convolution layer with the sum of the fourth-order residual error result, and performing double sampling on the summed result by continuously adopting the transposed convolution layer with the sum of the third-order residual error result and the final-time transposed convolution layer;
s1-3-6, performing 5 x 5 transposition convolution on a double sampling result obtained by adopting a transposition convolution layer at the last time, performing 4 x 4 convolution with an expansion value of 1, outputting an identification result through an SOFTMAX function, verifying the correct rate and obtaining a loss function value by using a verification set, adjusting D-LinkNet network parameters through back propagation until the correct rate and the loss function value are stable, and finishing training; during specific calculation, fragmenting the remote sensing satellite images to be recognized in the same size, sequentially inputting the fragmentized remote sensing satellite images into D-LinkNet to form road recognition, and sequentially splicing and restoring the road recognition according to the input sequence to form a road recognition result, wherein the road comprises a motor vehicle road, a non-motor vehicle road, a pedestrian road, a river, a natural lake and an artificial lake; wherein the pedestrian road comprises roads which can be walked by people or road vehicles or operation task tools in building group areas such as streets in cities, footpaths beside non-motor vehicle lanes, districts or factory buildings and the like;
the loss function adopts the difference value of the area of the road part of the identification result and the area of the road part in the manual marking layer;
s1-3-7, based on a training set formed by remote sensing satellite maps of multiple cities in S1-3-1, extracting a series of feature maps obtained by different convolutional layers by using a VGG-16 algorithm without an added layer as a CNN (compressed natural gas) backbone network, wherein the feature maps are 1/2-1/10 of the size of an input image;
meanwhile, a characteristic pyramid is constructed by using different layers of a CNN main network through an image pyramid algorithm FPN, and the frames of a plurality of buildings are predicted,
s1-3-8, for each building in a plurality of buildings, obtaining a local feature map F of the building by using a RoIAlign algorithm on the feature maps obtained by the series of different convolutional layers and the corresponding frame of the building;
s1-3-9, for the local feature map F of each building, convolutional layer processing is adopted to form a polygonal boundary cover M, P predicted vertexes of the boundary cover M are formed by utilizing the convolutional layer processing, a CNN network model is optimized by taking the position mean square deviation between the CNN network model and a real vertex as a judgment standard until the mean square deviation is smaller than a preset value, and the CNN network model and a trained D-LinkNet model form a real network 3D model.
4. The method according to any one of claims 1 to 3, wherein S2-1 divides each sensor of the atmospheric temperature and the ground temperature, the atmospheric humidity and the air humidity on the road section, the water level, the water quality, the harmful gas leakage and the automobile exhaust into eight regions in sequence, namely an atmospheric temperature region (i.e. a first region, and then sorting according to the sequence), a ground temperature region, an atmospheric humidity region, an air humidity region on the road section, a water level region, a water quality region, a harmful gas leakage region and an automobile exhaust region;
setting a sensor with local temperature and humidity as a ninth region;
s2-2, edges of each sensor node-server node in each graph area are formed inside each graph area, and a sub-graph neural network inside the graph area is formed; thereby forming a double graph network with spatial distribution of sensors and relationship distribution among nodes represented by a subgraph neural network, and taking the subgraph neural network in each graph region as a graph neural network layer to monitor each parameter respectively, wherein all the graph neural network layers form a 3D graph neural network.
5. The method according to claim 4, wherein S3 specifically comprises:
for each graph neural network layer, transmitting multi-protocol data represented by various types of sensors in S1 to a server at regular time intervals to form a data set, dividing the data set into a training set Train = { (u, yu) }, a verification set Val = { (u ', yu') }, and a prediction anomaly probability Test set Test = { (u ", yu") }, wherein the ratio of the three data is 1-3, yu ', yu "respectively represents the labels of nodes u, u', u", and the data has corresponding sensor sequence numbers and data acquisition times, and an attention model is constructed based on each graph neural network layer, and the following steps are performed:
s3-1, aggregating the server nodes in a specified period, specifically comprising:
s3-1-1 for each sensor and server within the neural network layer of the graph, a node pair is formed (i) Nk ,j k ) Wherein N represents the sequential number of the sensors in the neural network layer, k belongs to X represents the number of the servers serving one neural network layer, X is the number set of the servers, the number of elements is the total number 9 of all the servers, and the importance is defined as
This indicates that server node j k For target sensor node i Nk Is measured. Based on S2Node pair (i) with the edge in-2 as the path Nk ,j k ) The importance formula of (c) can be expressed as follows:
wherein
Respectively representing each sensor node i in the neural network layer of the graph Nk And server node j k The multi-protocol data in the data set is taken as one of elements in the vector to be represented to form a 4-8-dimensional vector, A node (. Is a deep neural network that performs a nodal attention mechanism; wherein,
other elements in the vector may be represented by the time of data acquisition, and the sensor node order number, all edge-based node pairs sharing A node (·)；
S3-1-2 predicted by the training set Train through the step S3-1-1
Verifying accuracy using a verification set Val, back-propagating using a cross-entropy loss function, training A node (. Until the accuracy stabilizes and the cross entropy loss function reaches a minimum;
s3-1-3 uses the softmax function to significance
Are normalized into
The following:
wherein ∑ M Summing the serial numbers M of all the sensors in the neural network layer of the diagram; activating a function AF comprises ReLU, leakyReLU, sigmoid and tanh;
s3-1-4, aggregating the feature vector of the node i based on the edge b with the corresponding coefficient through the projection feature of the neighbor as follows:
wherein
Representing a node j k The feature vector learned on the edge, σ, is the activation function LeakyReLU;
s3-1-5 node attention mechanism A node (. K) repeat 1-5 times, and concatenate each learned vector representation into the graph neural network layer, for all edge sets { b } 1k ,b 2k ,b 3k ，…,b Nk As a given type of path b, vector features for all sensor nodes
S3-2 for each graph neural network layer, a plurality of corresponding feature vectors p are obtained according to S3-1-1-S3-1-5 by using a training set Train bk Inputting k belongs to X into a multilayer perceptron (MLP), training the MLP by taking whether the corresponding abnormality is abnormal or not as classification, verifying by using a verification set Val to obtain the accuracy acc, and calculating the cross entropy as a loss function
MLP network parameters are optimized by back propagation until acc stabilizes and loss is minimized, where y L Is a set of node indexes with labels, Y and
labels and vector features p, respectively, of nodes bk W is classifier in MLPThe parameter, g, is superscript to indicate
The middle summation item corresponds to the sensor node.
6. The method of claim 5, wherein said specified time interval is 1s-1min, said specified time period is 1 week-1 year, and according to different specified time periods T e [4,12 ]]A plurality of attention systems { A) corresponding to different predetermined periods are formed node1 (·)，A node2 (·)，…，A nodeT (-) and a system of perceptual models { MPL } 1 ,MPL 2 ,…,MPL T }。
7. The method according to claim 5 or 6, wherein S4 specifically comprises:
s4-1, continuously collecting data and updating a data set;
s4-2, collecting Test data of the Test set in a preset time period (1 day-1 year), and obtaining a feature vector to be predicted according to S3-1-1-S3-1-5
Since data is transmitted at predetermined time intervals, the obtained feature vector corresponds to the time point.
S4-3 will
Inputting the training MPL to obtain the prediction result, if finding abnormal, finding abnormal time in preset time period, according to the corresponding characteristic vector
And reversely checking the characteristic vectors in each input test set, finding the marked places in the abnormal characteristic vectors, and inquiring the places corresponding to the suspected abnormal data so as to further eliminate the places corresponding to the abnormal data determined on the basis of the sensor abnormality.
8. The method of claim 7, wherein the reverse check is implemented by temporarily storing the input test set node vectors in the server, the input vectors in the temporary storage are automatically deleted when the data is normal, and the deletion operation is not performed once the data is abnormal, so as to be used for the reverse check until the reverse check and the deletion after the vector is stored;
attention mechanism { A } derived for different prescribed periods node1 (·)，A node2 (·)，…，A nodeT (-) and a system of perceptual models { MPL } 1 ,MPL 2 ,…,MPL T S4-2-S4-3 is also performed during each of the specified time periods.
9. A system capable of implementing the method according to any one of claims 1 to 8, comprising nine servers, and each server serving each type of sensors, for monitoring or detecting atmospheric temperature and ground temperature, atmospheric humidity and road air humidity, water level, water quality, harmful gas leakage, automobile exhaust, sensors indicating local temperature and humidity, and at least one sensor per type, and for collecting data of different protocols for each type of sensor, forming a data set; wherein,
the server is also used for building and establishing a 3D model of the urban sensor reality network, building a 3D graph neural network based on the built 3D model, building an attention model based on the 3D graph neural network, building a perception model of data abnormity, and realizing abnormity monitoring of multi-protocol data sent by various sensors by utilizing the built perception model.
10. A non-transitory storage medium having stored therein a computer readable program executable by the server to implement the real-time city data management method for the real-world network graph-based neural network of any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211213714.1A CN115795996A (en) | 2022-09-30 | 2022-09-30 | Real network diagram-based city data real-time management method and system of graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211213714.1A CN115795996A (en) | 2022-09-30 | 2022-09-30 | Real network diagram-based city data real-time management method and system of graph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115795996A true CN115795996A (en) | 2023-03-14 |
Family
ID=85432504
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211213714.1A Pending CN115795996A (en) | 2022-09-30 | 2022-09-30 | Real network diagram-based city data real-time management method and system of graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115795996A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116994427A (en) * | 2023-07-04 | 2023-11-03 | 重庆邮电大学 | Road condition prediction method based on big data |
| CN117074627A (en) * | 2023-10-16 | 2023-11-17 | 三科智能(山东)集团有限公司 | Medical laboratory air quality monitoring system based on artificial intelligence |
| CN117332120A (en) * | 2023-08-29 | 2024-01-02 | 泰瑞数创科技(北京)股份有限公司 | Geographic entity relation construction and expression method based on space calculation |
| CN118314464A (en) * | 2024-06-11 | 2024-07-09 | 华侨大学 | Historical urban area boundary judgment method, system and equipment based on graph neural network |
| CN118503332A (en) * | 2024-05-31 | 2024-08-16 | 北京普巴大数据有限公司 | Knowledge management method, terminal and server based on block chain |
| CN118551801A (en) * | 2024-05-21 | 2024-08-27 | 北京普巴大数据有限公司 | Industrial Internet of things equipment collaborative management and control system based on data and knowledge driving |
-
2022
- 2022-09-30 CN CN202211213714.1A patent/CN115795996A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116994427A (en) * | 2023-07-04 | 2023-11-03 | 重庆邮电大学 | Road condition prediction method based on big data |
| CN117332120A (en) * | 2023-08-29 | 2024-01-02 | 泰瑞数创科技(北京)股份有限公司 | Geographic entity relation construction and expression method based on space calculation |
| CN117332120B (en) * | 2023-08-29 | 2024-04-30 | 泰瑞数创科技(北京)股份有限公司 | Geographic entity relation construction and expression method based on space calculation |
| CN117074627A (en) * | 2023-10-16 | 2023-11-17 | 三科智能(山东)集团有限公司 | Medical laboratory air quality monitoring system based on artificial intelligence |
| CN117074627B (en) * | 2023-10-16 | 2024-01-09 | 三科智能(山东)集团有限公司 | Medical laboratory air quality monitoring system based on artificial intelligence |
| CN118551801A (en) * | 2024-05-21 | 2024-08-27 | 北京普巴大数据有限公司 | Industrial Internet of things equipment collaborative management and control system based on data and knowledge driving |
| CN118503332A (en) * | 2024-05-31 | 2024-08-16 | 北京普巴大数据有限公司 | Knowledge management method, terminal and server based on block chain |
| CN118314464A (en) * | 2024-06-11 | 2024-07-09 | 华侨大学 | Historical urban area boundary judgment method, system and equipment based on graph neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115795996A (en) | Real network diagram-based city data real-time management method and system of graph neural network | |
| Wang et al. | Spatio-temporal urban growth dynamics of Lagos Metropolitan Region of Nigeria based on Hybrid methods for LULC modeling and prediction | |
| CA2561939C (en) | Method and system for forecasting events and results based on geospatial modeling | |
| WO2018214060A1 (en) | Small-scale air quality index prediction method and system for city | |
| Yang et al. | ImgSensingNet: UAV vision guided aerial-ground air quality sensing system | |
| Lin et al. | Mining public datasets for modeling intra-city PM2. 5 concentrations at a fine spatial resolution | |
| US11776104B2 (en) | Roof condition assessment using machine learning | |
| US20080104005A1 (en) | Method and system for spatial behavior modification based on geospatial modeling | |
| US20080082472A1 (en) | Event, threat and result change detection system and method | |
| CN110738354B (en) | Method and device for predicting particulate matter concentration, storage medium and electronic equipment | |
| US11232582B2 (en) | Visual localization using a three-dimensional model and image segmentation | |
| CN114912707B (en) | Air quality prediction system and prediction method based on multi-mode fusion | |
| CN112288156A (en) | Air quality prediction method based on meta-learning and graph attention space-time neural network | |
| CN115775085B (en) | Digital twinning-based smart city management method and system | |
| US20230004903A1 (en) | Methods of greening management in smart cities, system, and storage mediums thereof | |
| CN109828578A (en) | A kind of instrument crusing robot optimal route planing method based on YOLOv3 | |
| CN110346518A (en) | A kind of traffic emission pollution visualization method for early warning and its system | |
| CN113092684A (en) | Air quality inference method based on space-time matrix decomposition | |
| CN118070619B (en) | Urban waterlogging model sensitive parameter identification optimization method by utilizing machine learning | |
| CN115629160A (en) | Air pollutant concentration prediction method and system based on space-time diagram | |
| Turukmane et al. | Multispectral image analysis for monitoring by IoT based wireless communication using secure locations protocol and classification by deep learning techniques | |
| CN115880466A (en) | Urban engineering surveying and mapping method and system based on unmanned aerial vehicle remote sensing | |
| CN113744541B (en) | Road network discharge loss space-time distribution reconstruction method and system for confrontation graph convolution network | |
| Wang et al. | Hybrid model for prediction of carbon monoxide and fine particulate matter concentrations near a road intersection | |
| Chen et al. | A Spatiotemporal Interpolation Graph Convolutional Network for Estimating PM₂. ₅ Concentrations Based on Urban Functional Zones |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |