CN111414504A - Building site flood detection system based on block chain and CIM - Google Patents
Building site flood detection system based on block chain and CIM Download PDFInfo
- Publication number
- CN111414504A CN111414504A CN202010383908.0A CN202010383908A CN111414504A CN 111414504 A CN111414504 A CN 111414504A CN 202010383908 A CN202010383908 A CN 202010383908A CN 111414504 A CN111414504 A CN 111414504A
- Authority
- CN
- China
- Prior art keywords
- encoder
- construction site
- area
- unit
- flood
- 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.)
- Withdrawn
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 71
- 238000010276 construction Methods 0.000 claims abstract description 104
- 238000011176 pooling Methods 0.000 claims abstract description 27
- 238000013528 artificial neural network Methods 0.000 claims description 42
- 238000012544 monitoring process Methods 0.000 claims description 33
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 28
- 230000007246 mechanism Effects 0.000 claims description 20
- 238000005516 engineering process Methods 0.000 claims description 18
- 238000012800 visualization Methods 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 7
- 125000004122 cyclic group Chemical group 0.000 claims description 7
- 238000010586 diagram Methods 0.000 claims description 6
- 238000009877 rendering Methods 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000002159 abnormal effect Effects 0.000 claims description 3
- 238000009792 diffusion process Methods 0.000 claims description 3
- 238000000034 method Methods 0.000 abstract description 24
- 238000012545 processing Methods 0.000 abstract description 6
- 230000008569 process Effects 0.000 abstract description 5
- 230000005540 biological transmission Effects 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 7
- 238000013461 design Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 7
- 230000000875 corresponding effect Effects 0.000 description 5
- 238000007726 management method Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000001965 increasing effect Effects 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 230000008447 perception Effects 0.000 description 3
- 230000002829 reductive effect Effects 0.000 description 3
- 230000011218 segmentation Effects 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000003491 array Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000036961 partial effect Effects 0.000 description 2
- 238000012805 post-processing Methods 0.000 description 2
- 230000002265 prevention Effects 0.000 description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 238000006424 Flood reaction Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000004566 building material Substances 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/48—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/483—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/27—Replication, distribution or synchronisation of data between databases or within a distributed database system; Distributed database system architectures therefor
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/08—Construction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Library & Information Science (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Economics (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Computing Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a construction site flood detection system based on a block chain and a CIM. The system comprises a shared feature encoder, a flood area suggestion first encoder, a weight acquisition unit, a weight addition unit, a fraction encoder, a candidate frame encoder, a pooling unit and a first full-connection network, and further comprises a computing cluster. By using the method and the device, not only are the results fed back to multiple elements in the construction site flood detection, but also the practicability of the detection result, the precision of the detection result and the safety performance in the data processing process are improved.
Description
Technical Field
The invention relates to the technical field of block chains, artificial intelligence, intelligent construction sites and CIM, in particular to a construction site flood detection system based on the block chains and the CIM.
Background
With the rapid development of the economy in China, real estate and building related industries are playing more and more important roles. Therefore, the number of construction sites is increasing at present, and with it, the management problems inside the sites are becoming more and more prominent. Particularly, a construction site is often subjected to a flood accident due to a complex field environment. Particularly, at night, normal construction personnel can not find accumulated water in a construction area in time when having a rest off duty. Although there are people on duty at night, the light is dark, the visible range is small, and the water overflowing area of the construction site is not easy to be found.
There are several reasons for site flooding: burst of a water pipe in a construction area, blockage of a water outlet in a construction site, incomplete closing of a water valve caused by that a constructor does not check the water pipe according to regulations after work, continuous leakage of water to form accumulated water and the like. The construction site is affected by water overflow in the construction site, the original building foundation can be corroded, the foundation is unstable, and potential safety hazards can be caused to circuits used in site construction.
The traditional detection algorithm generally uses the difference between the flooding area and the background pixel to obtain the position of the flooding area and the area of the flooding area through fixed threshold segmentation. However, the construction environment of the construction site is complex, the illumination is slightly dark, the fixed threshold value can become unsuitable along with the change of the environment, and the stability and the accuracy of the system using the method are to be improved. In addition, the hardware system used for calculation is easy to leak information and has low safety performance. Therefore, the existing field of construction site flood detection has the problems of low detection precision, poor robustness, single result feedback and low safety performance.
Disclosure of Invention
The invention aims to provide a construction site overflowing water monitoring system based on a block chain and a CIM (common information model), aiming at the defects in the prior art, not only can result be fed back to multiple units, but also the detection precision, robustness and safety performance in the data processing process are improved.
A worksite flood detection system based on block chains and CIMs, the system comprising:
the shared characteristic encoder is used for encoding the monitoring area image and extracting a shared characteristic graph;
the flood area suggests a first encoder for further encoding the shared characteristic diagram and extracting characteristics;
the weight obtaining unit is used for obtaining channel weights of output features of the first encoder suggested by the flood area based on the attention mechanism;
the weight adding unit is used for adding channel weight to a channel of the output characteristic of the first encoder suggested by the water diffusion area to obtain an attention enhancement characteristic diagram;
the score encoder is used for encoding the attention enhancement feature map and extracting candidate frame category score information;
the candidate frame encoder is used for encoding the attention enhancement feature map and extracting coordinate information of a candidate frame in the flooding area;
the pooling unit is used for performing pooling operation on the feature map obtained after the candidate frame of the flooding area is mapped to the shared feature map, and converting the feature map into a fixed size;
the first full-connection network is used for performing regression on the candidate frames of the flooding area on the feature map output by the candidate frame pooling unit to obtain an accurate flooding area surrounding frame;
the system further comprises a computing cluster, a construction site overflowing water detection block chain private chain is configured in the computing cluster according to the image data sent by the image acquisition unit, network reasoning is executed, a construction site overflowing water grade estimation result is obtained, and the construction site overflowing water grade estimation result is sent to the urban construction site area information model.
The weight obtaining unit specifically includes:
the global pooling module is used for performing addition averaging on the eigenvalues in each channel of the output tensor of the first encoder suggested by the flood area;
the bottleneck module is used for adopting the correlation among the channels output by the full-connection layer modeling global pooling module and outputting a group of channel weights;
and the activation module is used for normalizing the channel weight output by the bottleneck module.
For the image data sent by the image acquisition unit, configuring a worksite detection block chain private chain in the computing cluster comprises:
the shared feature encoder, the overflowing region suggestion first encoder, the weight obtaining unit, the weight adding unit, the fraction encoder, the candidate frame encoder, the pooling unit and the first full-connection network are main component modules of the construction site overflowing detection deep neural network;
calculating parameters required by loading the construction site flood detection deep neural network on all nodes in the cluster;
for image data sent by an image acquisition unit, selecting a plurality of available nodes from a computing cluster, taking parameters required by a construction site flooding detection deep neural network, namely a shared feature encoder, a flooding area suggestion first encoder, a weight acquisition unit, a weight addition unit, a score encoder, a candidate frame encoder, a pooling unit and a first full-connection network which are distributed in different available nodes, as block data of corresponding nodes, and connecting the node blocks according to a construction site flooding detection deep neural network reasoning sequence to generate a construction site flooding detection block chain private chain.
And the blocks in the private chain of the block chain encrypt the neural network inference intermediate result data to be transmitted to the next block, and decrypt the neural network inference intermediate result data received from the previous block.
And carrying out encryption and decryption operations on the data transmitted between the blocks by adopting a tensor confusion encryption mechanism.
Tensor obfuscated encryption mechanisms include:
randomly generating C groups of random numbers, wherein each group comprises two random numerical values, and C is the channel number of the output tensor of the node where the block is located;
and performing cyclic shift operation on the width direction and the height direction of the tensor according to the two random numerical values.
The method comprises the steps of respectively and properly subdividing a shared feature encoder, a flood area suggestion first encoder, a weight obtaining unit, a weight adding unit, a fraction encoder, a candidate frame encoder, a pooling unit and a first full-connection network, taking parameters of each subdivided module distributed in different nodes as block data of corresponding nodes, connecting the node blocks according to a construction site flood detection depth neural network reasoning sequence, and generating a construction site flood detection block chain private chain.
The system further comprises a visualization unit, the urban construction site region information model comprises construction site scene modeling information, monitoring region information and construction site overflowing level estimation results in the urban area, and the visualization unit is used for visualizing the urban construction site region information model.
The visualization unit includes:
the initialization module is used for acquiring the modeling information of each construction site scene from the urban construction site region information model, rendering the urban construction site region information model by combining the WebGIS technology, and displaying the rendered urban construction site region information model on a foreground Web page to obtain an initial display result of the urban construction site region information model;
the data acquisition module is used for acquiring monitoring area information and a construction site overtopping detection result from the urban construction site area information model;
and the visualization module is used for matching the monitoring area information and the construction site overtopping detection result to the initial display result from the geographical position according to the monitoring area information, and carrying out warning marking on the monitoring area according to the abnormal early warning information.
Compared with the prior art, the invention has the following beneficial effects:
1. the method adopts the deep neural network to detect the construction site flood, and compared with the traditional detection technology based on threshold segmentation, the method uses a large number of samples, has better generalization performance, and improves the stability and accuracy of the system.
2. Compared with the traditional network structure, the deep neural network introduces an attention mechanism, focuses attention on an interested area, extracts features which can represent the features of the flooding candidate frame, and has high detection efficiency and high detection precision.
3. The method is based on the block chain technology, reasonably divides the construction site flood detection deep neural network, dynamically generates the block chain private chain aiming at each network reasoning request, and compared with the traditional single machine execution, not only improves the parallel performance of the system, but also has better fault tolerance performance.
4. The block chain private chain is generated in real time according to available nodes in the server cluster, is not easy to be tampered and attacked, and has higher confidentiality.
5. The invention encrypts the data between the private chains of the network inference block chain, prevents the leakage of the transmission data between the private chain blocks of the block chain, ensures the confidentiality of the transmission data, has convenient tensor confusion encryption operation and small calculation amount, and can not increase the system burden while improving the confidentiality.
6. The city construction site region information model is designed based on the CIM technology, construction site flood detection results are stored, the city construction site region information model is visualized, compared with the traditional result feedback, the feedback result is more diversified, the construction site region three-dimensional display, the warning mark and the monitoring region image are included, and the monitoring personnel can more clearly and definitely know the wall crack condition in the construction site.
Drawings
FIG. 1 is a diagram of a neural network architecture of the system of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The invention provides a construction site flood detection system based on a block chain and a CIM. The quality of the area suggestion candidate frame is improved by adopting an area suggestion method based on an attention mechanism, and the flooding information of the construction area is monitored in real time based on a two-stage deep learning target detection method. The attention mechanism aligns the internal experience with the external perception to increase the fineness of the observation of the partial regions to focus on the true candidate regions. The method establishes a CIM information model, displays the detected construction site flood information on a Web end through a Web GIS and sends out early warning. In order to prevent the deep neural network from being tampered, the invention adopts the idea of a block chain to encrypt each module of the deep neural network and realizes the distributed calculation of each module, thereby enhancing the data security and stability of the deep neural network. FIG. 1 is a diagram of a neural network architecture of the system of the present invention. The following description will be made by way of specific examples.
The first embodiment is as follows:
building site flood detection system based on block chain and CIM, this system carries out building site flood detection based on city building site information model. The city construction site information model comprises construction site scene modeling information, monitoring area information and construction site water overflowing detection data of a city area.
Specifically, a Building Information model (Building Information Modeling) is a new datamation tool applied to design, construction and management of Building engineering, and integrates all Building data in a Building project cycle. The visualization of the building information model can be realized by using visualization software in combination with technologies such as WebGIS and the like, and convenience is provided for building design and building management. The CIM (City Information Modeling, City Information model) is a further upgrade of the BIM, extends the Modeling range from a single building model or a plurality of building models to the three-dimensional Modeling of the whole City, and can perform three-dimensional Modeling, displaying and managing on the scenes of buildings, traffic, roads and the like of the whole City.
The method and the system are combined with CIM to display the detection result of the construction site flood in real time and provide early warning information for the monitoring personnel. Therefore, the invention designs an urban construction site area information model. The city construction site information model is based on CIM technology and comprises construction site scene modeling information, monitoring area information and construction site water overflowing detection results in a city or a certain area of the city. The building site scene modeling information comprises building site building information, building site building material placement information, worker work area information and other various building site scene information, the type information simultaneously contains corresponding geographic position information, and the three-dimensional scene of the building site can be restored and displayed through visual software by combining a Web GIS technology. The monitoring area information comprises monitoring area geographical position information, monitoring visual sensor geographical position information, images shot by the monitoring visual sensors and a coordinate transformation matrix. The monitoring area information is used for restoring a monitoring area image in a visualized city building site area information model. The coordinate change matrix is used for converting the monitoring image into an image with better visualization effect, such as a top view and the like. And integrating the detection result of the construction site flood into the matched monitoring area for the supervision personnel to check. In the invention, the construction site overtopping detection result is transmitted to the urban construction site information model in real time.
The following is a detailed description of how the site flood detection results are obtained. The invention is realized by adopting a deep neural target detection network based on two stages: through processing a series of data collected by the sensor and adopting an attention mechanism-based region suggestion method, a high-quality flooding region candidate frame is generated, interest information is focused, and interference information is filtered, so that a plurality of irrelevant objects are reduced for a subsequent classification task of the detector, the recall rate is increased, and the accuracy of target detection is improved.
In the traditional two-stage detection network, a rough rectangular frame is predicted through a region suggestion network structure, and then the rectangular frame is input into a regression network to predict the accurate position and type of a target. The area proposal network of the first stage plays a role in coarse screening of accurate targets, so that the classifier of the second stage is easier to converge compared with the detector of the first stage, thereby having higher accuracy. However, if the area recommendation network predicts a rectangular box that does not contain the true target, then the second stage coordinate regression is also not easily converged. Due to insufficient light at night, the flood area is easy to be confused with background information, the original area suggestion network can generate some wrong suggestion areas, and a plurality of irrelevant objects are added for the subsequent classification task of the detector in the second stage, so that the detection accuracy is reduced.
In addition, the night illumination condition of the construction site is relatively poor, and the characteristics of the flood area are similar to other surrounding background characteristics and are not obvious in difference. If the traditional two-stage detector is adopted, the original area suggestion network generates a plurality of wrong suggestion areas, so that a plurality of irrelevant objects are added for the subsequent classification tasks of the second-stage detector, the network computing burden is increased, and the detection efficiency and accuracy are reduced.
Therefore, the present invention improves the quality of the region suggestion candidate box using a region suggestion method based on an attention mechanism that aligns internal experience with external perception to increase the observation fineness of partial regions to focus on real candidate regions. Therefore, a high-quality candidate frame is provided for the detector in the second stage, the network is easier to converge, and the recall rate and the accuracy rate are improved.
The deep neural network of the present invention is shown in fig. 1. And the shared characteristic encoder is used for carrying out characteristic encoding on the acquired monitoring area image and extracting a shared characteristic image. Specifically, a picture is input into a network, and image features are extracted through a shared encoder, wherein the shared features are used as input of both a region suggestion network and a target detection network.
In a convolutional neural network, after different convolution kernel operations, each channel generates a new signal, for example, each channel of a picture feature is convolved with 32 kernels, and a matrix (H, W,32) of 32 new channels is generated, wherein H, W represents the height and width of the picture feature respectively.
In the above method, the characteristics of each channel represent the components of the picture at different convolution kernels. If according to the traditional method, each component is not processed, namely the contribution of each component to the flooding area proposal is the same, more area proposal candidate boxes irrelevant to the flooding area are obtained.
According to the method, the contribution degree of each channel of the feature map to the flooding candidate area is predicted through an attention mechanism, so that the characterization capability of the channel related to the flooding candidate area features is improved, and the characterization capability of backgrounds, such as the ground, trees and the like, which are unrelated to the candidate area is inhibited. After sharing the encoder, extracting a candidate frame by adopting a region suggestion method based on an attention mechanism, then mapping the extracted candidate frame to the shared encoder, obtaining the characteristics of the candidate region, and regressing the category and the position of the object through the characteristics. According to the idea, the invention designs a flood area proposal first encoder, a channel weight acquisition unit and a weight addition unit.
And the flooding area suggests a first encoder for carrying out feature encoding on the shared feature map and further extracting features. The channel weight acquiring unit is used for acquiring the channel weight of the output characteristic of the first encoder suggested by the flood area based on the attention mechanism, and specifically comprises the following steps: the global pooling module is used for performing addition averaging on the eigenvalues in each channel of the output tensor of the first encoder suggested by the flood area; the bottleneck module is used for adopting the correlation among the channels output by the full-connection layer modeling global pooling module and outputting a group of channel weights; and the activation module is used for normalizing the channel weight output by the bottleneck module. And the weight adding unit is used for adding channel weight to the channel of the output characteristic of the first encoder suggested by the water diffusion area to obtain the attention enhancement characteristic map.
Specifically, firstly, the first encoder is suggested to extract the features through the flood area, and an attention mechanism is applied to the extracted feature map, namely, a weight w is added to each channel of the suggested features of the flood areacRepresenting the correlation of the channel with the flood candidate area, the weight wcThe larger the correlation degree of the channel information is, the higher the correlation degree of the channel information is, the more important the channel information is. On the contrary, if wcThe smaller the size, the less highly correlated the representation with the flooding candidate area, and the information that needs to be suppressed. w is acIs a hyper-parameter and can be updated by iteration of the network, where n represents the channel index.
Firstly, a global pooling module is utilized to add all the characteristic values in each channel and then average the characteristic values, namely, the current characteristic graph is subjected to global average pooling. And then two full connection layers form a bottleneck module structure to model the correlation between channels, determine the significance of a certain channel, and output the weight with the same number as that of the input characteristic channel. Further, the activation module normalizes the weight obtained in the previous step by using an activation function to obtain a normalized weight between 0 and 1:
wherein, the function is activation sigmoid, UcFor an input profile, wcThe weights that need to be learned for the current feature map. Finally, the weight adding unit weights the normalized weight to the characteristic of each channel through a Scale operation: xc＝Sc*UcTherefore, attention enhancement on the key channel domain can be achieved, the region suggestions of the types irrelevant to the water surface characteristics are filtered, and the region suggestions of the targets of the real flooding region are reserved. Thereby gaining attentionAnd (4) strong feature maps.
And continuously extracting category score information of the candidate frame by adopting a score encoder, wherein the category refers to the flooding area and the background, extracting coordinate information by adopting the candidate frame encoder, and performing post-processing operation to obtain a final area suggestion rectangular frame. The post-processing operation refers to outputting the rectangular frame as a final region proposal rectangular frame when the predicted type is the flooding region and the intersection ratio of the rectangular frame and the ground channel surrounding frame calculated by the coordinate information extracted by the candidate frame encoder box encoder is larger than a certain threshold value. Preferably, the threshold value is set to 0.7 in this embodiment.
In the training process, the loss of the area recommendation network comprises two parts, namely a candidate box type prediction score generated by a fraction encoder, calculating the loss of the candidate box type through a cross entropy loss function, and optimizing the candidate box type, and a candidate box coordinate prediction (4 values) generated by a candidate box encoder, calculating the position loss of the candidate box type prediction score and the ground channel through an L1 loss function.
Therefore, the first stage, namely the extraction of the rectangular frame of the proposal of the flood area is completed, and the extracted high-quality proposal frames are all the proposal frames by adopting the area proposal method based on the attention mechanism, and the high quality refers to the intersection ratio with the real target position and is greatly improved compared with the condition before the attention mechanism is not added, thereby laying an accurate foundation for the accurate regression of the target position in two stages. It should be noted that the area suggested rectangular box is only roughly located to the flood area, and the coordinates thereof cannot really reflect the position of the flood area.
After the regional suggestion network based on the attention mechanism generates the candidate box of the flood region, mapping the candidate box to a shared feature map to obtain the features of the candidate box region. Note that the candidate regions are different in size, and the full-link layer needs to accept input of a fixed length. Therefore, the frame candidate area needs to be converted into a fixed size through the region-of-interest pooling roi posi posing operation. Preferably, the sizes of the characteristic maps after the roi posing in this example all become fixed 7 × 7. Therefore, the pooling unit is adopted to perform pooling operation on the feature map obtained after the candidate frames are mapped to the shared feature map, and the feature map is converted into a feature map with a fixed size. It should be noted that the principle of the roi posing algorithm is well known and is not specifically described herein, nor is it intended to be protected by the present invention.
The extracted candidate frame feature graph is flattened by adopting a first full-connection network, on one hand, the extracted candidate frame feature graph is used for predicting the category information (the flood area or the background) of the candidate frame, a cross entropy loss function is adopted during training, on the other hand, the extracted candidate frame feature graph is used for predicting the real coordinate information of the flood area, an L1 loss function is adopted during training, a gradient descending mode is adopted, a model is optimized, and the accurate prediction of the coordinates of the flood area is completed.
In the neural network, an encoder is used for carrying out feature extraction on input multi-channel two-dimensional data, and a decoder is used for carrying out decoding reconstruction on a feature map. The encoder and the decoder have various implementations, and CNN Block, Res Block and the like can be adopted. In order to take account of the size of a large target and a small target, the invention proposes to adopt an hourglass network to extract features, and an implementer can also select a proper module design in a neural network, such as Residual Block, Bottleneck Block, CNN Block and the like according to the size of an image and the occupation of a video memory. The encoder, the decoder and the full-connection network of the invention adopt which network design, an implementer can select according to the specific implementation requirements, and the modularization idea is the protection content of the invention.
In order to improve the confidentiality of the system and prevent data leakage, the invention designs the private chain of the block chain by combining the block chain technology.
The present invention is described in detail herein in connection with the block chain technique and the DNN technique. The block chain adopts block division data, a chain data structure is used, the data are used as blocks for verification and storage, the whole data structure is summarized, centralized hardware and management mechanisms do not exist, and decentralization is achieved. The block chain technology of the first generation is mainly applied to a distributed account book, the block chain technology of the second generation mainly realizes an intelligent contract, and the block chain idea and other field technologies can be combined by the block chain technology of the third generation, so that more and more presentation forms exist, and more emphasis is placed on system function service. The block chain private chain completely inherits the characteristics of the public chain, is not bound by a game mechanism, focuses more on data transmission and encryption and decryption processing of practical application, and can be better combined with technologies in other fields. For deep neural network computation in artificial intelligence, it is not necessary to store intermediate result data, and chained logic is retained to match the principle of neural network forward propagation.
The invention considers the problem that the contents of the image data in the plaintext are leaked in the uploading process and the processing process when the data are directly uploaded to be processed, so that the invention uses the form of a block chain private chain, takes different modules of a deep neural network as blocks, carries out dispersed reasoning and carries out encryption processing on the data transmitted among the blocks, thereby realizing excellent performances of parallel reasoning, fault tolerance and data leakage prevention.
Based on the idea, firstly, a shared feature encoder, a first encoder suggested by a flood area, a weight obtaining unit, a weight adding unit, a fraction encoder, a candidate frame encoder, a pooling unit and a first fully-connected network are used as different modules of the network. Thus, according to the inference sequence of the neural network shown in fig. 1, the deep neural network inference chain for the construction site flood detection can be obtained.
The system also comprises a server cluster, wherein all nodes in the server cluster are loaded with the weight and parameters required by the construction site overflowing water detection deep neural network; according to each construction site overtopping detection deep neural network inference request, selecting a plurality of available nodes from a server cluster, taking a shared feature encoder, an overtopping area suggestion first encoder, a weight obtaining unit, a weight adding unit, a score encoder, a candidate frame encoder, a pooling unit and weights and parameters required by a first full-connection network, which are respectively distributed in different available nodes, as block data, generating a construction site overtopping detection block chain private chain according to a construction site overtopping detection deep neural network inference sequence, and executing construction site overtopping detection deep neural network inference.
When selecting available nodes and performing node sorting, preferably, the available nodes in the server cluster are randomly sorted, and the number of computing nodes is selected from the nodes, wherein the number of the computing nodes is the same as the number of the blocks. Specifically, two random numbers are generated, the number of the random numbers is the number N of the available nodes in the server cluster, and the value of each bit in the random numbers is 0-N. Two random number arrays are added together bitwise to obtain a string of numbers, for example: the generated random number arrays are [1,3,12,5.. 8] and [2,10,1,0.. 4], and bit-wise addition is performed to obtain [3,13,13,5.. 12 ]. And according to a module needing calculation, taking 8 bits in the array after the summation as the node of the forward calculation. The value-taking method may be to traverse the array from left to right, and if the value is equal to the previous value, skip the value, thereby obtaining the selected node index. For example, there are 10 available nodes, 8 nodes have been selected according to the random number sequence, the first node is taken, and parameters such as the weight required by the shared feature encoder in the nodes are used as block data; and randomly taking another node, taking parameters such as the weight required by the first encoder in the flood area in the node as block data, linking the block data with the previous block, and by analogy, generating a corresponding private chain of the construction site flood detection block according to a neural network reasoning sequence. Therefore, a plurality of construction site flood detection block private chains generated aiming at different requests can exist in the server cluster at the same time, and the block private chains are dynamically generated, are not easy to crack by attack, and have better confidentiality. And after the block chain private chain is obtained according to the inference sequence, performing network inference calculation on the inference request according to the inference sequence.
Some modules of the neural network are difficult to be put into one node at one time, namely, the computation of some modules is large, and the computation is difficult to be completed in a short time. Therefore, the blocks, i.e., the above-mentioned encoder, decoder, etc., units can be further divided in advance. For those skilled in the art, inside the encoder and decoder is a neural network containing a multi-layer neuron structure, which can be cut by layers. The granularity of the further segmentation can be adjusted by the implementer according to the specific implementation. In this embodiment, preferably, the shared feature encoder is divided into 3 blocks, the first encoder in the flooding area is divided into 3 blocks, the weight obtaining unit is divided into 3 blocks (the global pooling module, the bottleneck module, and the activation module), the fractional encoder is divided into 2 blocks, the candidate frame encoder is divided into 2 blocks, and the first fully-connected network is divided into 4 blocks. Therefore, according to the inference sequence of the neural network, a more subdivided construction site overtopping detection deep neural network inference chain can be obtained. Correspondingly, selecting a plurality of available nodes from the server cluster according to each building site overflowing detection deep neural network reasoning request, taking weights and parameters required by sub-modules which are respectively distributed in different available nodes after subdivision as block data, and generating a building site overflowing detection block chain private chain according to a building site overflowing detection deep neural network reasoning sequence.
Meanwhile, the image acquisition unit can be used as a block and added into the block chain private chain, that is, a camera terminal is added into the block chain private chain, and camera parameters are used as block data. The purpose of doing so is to guarantee that the output of image acquisition unit is secret through subsequent encryption strategy, is difficult for intercepting, falsifying. The node where the monitoring center receiving the construction site overflowing detection result is located can also be added into the block chain private chain, and the purpose of doing so is to ensure the safety of data transmission between the monitoring center and the server cluster.
Further, in order to ensure the integrity and security of data received by each block and prevent data from being attacked and tampered during transmission, data transmitted between blocks needs to be encrypted. That is, data transmission from private link node to node of the block chain requires the use of encryption, preferably using a tensor obfuscated encryption mechanism.
Specifically, the encryption is completed by carrying out tensor confusion on the feature maps output by the current block or the computing node, wherein the specific confusion mode is that for each channel of the current feature map, the size of the channel is H × W, a group of random number combinations are selected to carry out cyclic shift on the current feature map, each group has two random numerical values, C groups are shared, the operation is respectively carried out on the width and the height of the feature map, and C represents the number of the channels of the current feature map.
For example, assuming that a shift operation is currently performed on the feature map with index i, the current random number combination is (m, n), where m < W and n < H. And when m is an odd number, circularly shifting to the left by the shift step length of m, otherwise, circularly shifting to the right. Assuming that the data in the horizontal direction of the original image is 011010001, when the value of m is 2 and even, the step length of right cyclic shift is 2 according to the rule, and after the shift, the step length is 010110100; when m is 3, odd, left shifted 110100010. The image vertical direction is the same as that: and when n is an odd number, performing downward cyclic shift, otherwise performing upward cyclic shift, wherein the shift step size is n. The intermediate result is one-dimensional, and cyclic shift in the left-right direction may be performed using the parameter m.
And applying the encryption algorithm to the data to be transmitted among all the blocks to finish the encryption operation of the data. After the data is input into the next block, the data is decrypted according to the reverse reasoning of the encryption rule, so that the encrypted transmission of the data between the computing nodes is completed.
It should be noted that the encryption parameters required for transmitting data between the modules of the deep neural network for site flood detection, namely between blocks, should be generated by the trusted nodes. The trusted node may be fixed, for example, a master node may be configured to broadcast encryption parameters and allocation tables periodically to prevent the encryption parameters from being broken. However, the fixed trusted node is vulnerable to attack, and therefore, preferably, the encryption parameter required by the next inference request is generated by the node where the last block of the server cluster executing the private chain of the blockchain is located, and is broadcast to all nodes in the server cluster, the terminal cluster and the monitoring center.
In order to visually present the CIM information state of a construction site area and enable a construction site manager to visually acquire the camera perception information of the current area and the information of whether the construction site floods through a network forecast, the CIM information model is displayed on the Web through a WebGIS technology by combining with the Web GIS visualization technology.
And the visualization unit is used for acquiring data from the urban construction site region information model, rendering the urban construction site region information model by combining the Web GIS, and displaying construction site water overflow monitoring conditions on a foreground Web page. The visualization unit includes: the system comprises an initialization module, a data acquisition module and a construction site water overflowing visualization module.
And the initialization module is used for acquiring the building site scene modeling information from the urban building site region information model, rendering the urban building site information model by combining the Web GIS technology, and displaying on a foreground Web page to obtain an initial display result of the urban building site region information model.
And the data acquisition module is used for acquiring monitoring area information and a construction site overflowing monitoring result from the urban construction site area information model.
And the construction site overflowing water visualization module is used for matching the monitoring area information and the construction site overflowing water detection result to the initial display result from the geographical position according to the monitoring area information, and carrying out warning marking on the monitoring area according to the abnormal early warning information.
Meanwhile, when the fact that the water overflowing area of the construction site is large is detected, an alarm can be sent out, and therefore the supervisor can take corresponding prevention and emergency measures according to the geographic position coordinates of the supervisor. Therefore, after the first fully-connected network obtains an accurate overflow area surrounding frame, the area calculation is carried out, and the area of the overflow area can be estimated.
The above embodiments are merely preferred embodiments of the present invention, which should not be construed as limiting the present invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (9)
1. A building site flood detection system based on block chains and CIMs (common information model), which is characterized by comprising:
the shared characteristic encoder is used for encoding the monitoring area image and extracting a shared characteristic graph;
the flood area suggests a first encoder for further encoding the shared characteristic diagram and extracting characteristics;
the weight obtaining unit is used for obtaining channel weights of output features of the first encoder suggested by the flood area based on the attention mechanism;
the weight adding unit is used for adding channel weight to a channel of the output characteristic of the first encoder suggested by the water diffusion area to obtain an attention enhancement characteristic diagram;
the score encoder is used for encoding the attention enhancement feature map and extracting candidate frame category score information;
the candidate frame encoder is used for encoding the attention enhancement feature map and extracting coordinate information of a candidate frame in the flooding area;
the pooling unit is used for performing pooling operation on the feature map obtained after the candidate frame of the flooding area is mapped to the shared feature map, and converting the feature map into a fixed size;
the first full-connection network is used for performing regression on the candidate frames of the flooding area on the feature map output by the candidate frame pooling unit to obtain an accurate flooding area surrounding frame;
the system further comprises a computing cluster, a construction site overflowing water detection block chain private chain is configured in the computing cluster according to the image data sent by the image acquisition unit, network reasoning is executed, a construction site overflowing water grade estimation result is obtained, and the construction site overflowing water grade estimation result is sent to the urban construction site area information model.
2. The system of claim 1, wherein the weight obtaining unit specifically comprises:
the global pooling module is used for performing addition averaging on the eigenvalues in each channel of the output tensor of the first encoder suggested by the flood area;
the bottleneck module is used for adopting the correlation among the channels output by the full-connection layer modeling global pooling module and outputting a group of channel weights;
and the activation module is used for normalizing the channel weight output by the bottleneck module.
3. The system of claim 1, wherein configuring a worksite detection block chain private chain in a computing cluster for image data sent by an image acquisition unit comprises:
the shared feature encoder, the overflowing region suggestion first encoder, the weight obtaining unit, the weight adding unit, the fraction encoder, the candidate frame encoder, the pooling unit and the first full-connection network are main component modules of the construction site overflowing detection deep neural network;
calculating parameters required by loading the construction site flood detection deep neural network on all nodes in the cluster;
for image data sent by an image acquisition unit, selecting a plurality of available nodes from a computing cluster, taking parameters required by a construction site flooding detection deep neural network, namely a shared feature encoder, a flooding area suggestion first encoder, a weight acquisition unit, a weight addition unit, a score encoder, a candidate frame encoder, a pooling unit and a first full-connection network which are distributed in different available nodes, as block data of corresponding nodes, and connecting the node blocks according to a construction site flooding detection deep neural network reasoning sequence to generate a construction site flooding detection block chain private chain.
4. The system of any one of claims 1-3, wherein a tile in the private chain of tiles encrypts neural network inference intermediate result data that it is to transmit to a next tile, and decrypts neural network inference intermediate result data that it receives from a previous tile.
5. The system of claim 4, wherein tensor obfuscated encryption is used to encrypt and decrypt data transmitted between blocks.
6. The system of claim 5, wherein the tensor obfuscated encryption mechanism comprises:
randomly generating C groups of random numbers, wherein each group comprises two random numerical values;
and performing cyclic shift operation on the width direction and the height direction of the tensor according to the two random numerical values.
7. The system according to any one of claims 1 to 6, wherein the shared feature encoder, the flood area recommendation first encoder, the weight acquisition unit, the weight addition unit, the score encoder, the candidate frame encoder, the pooling unit and the first fully-connected network are respectively and properly subdivided, parameters of each subdivided module respectively distributed in different nodes are used as block data of corresponding nodes, and the node blocks are connected according to a building site flood detection deep neural network inference sequence to generate a building site flood detection block chain private chain.
8. The system of claim 1, wherein the urban worksite area information model is constructed based on CIM technology, and comprises various worksite scene modeling information, monitoring area information, worksite flood level estimation results in the urban area; the system further comprises a visualization unit, and the city construction site region information model is visualized by the visualization unit.
9. The system of claim 8, wherein the visualization unit comprises:
the initialization module is used for acquiring the modeling information of each construction site scene from the urban construction site region information model, rendering the construction site region building information model by combining the Web GIS technology, and displaying the construction site region building information model on a foreground Web page to obtain an initial display result of the construction site region building information model;
the data acquisition module is used for acquiring monitoring area information and a construction site overtopping detection result from the urban construction site area information model;
and the visualization module is used for matching the monitoring area information and the construction site overtopping detection result to the initial display result from the geographical position according to the monitoring area information, and carrying out warning marking on the monitoring area according to the abnormal early warning information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010383908.0A CN111414504A (en) | 2020-05-08 | 2020-05-08 | Building site flood detection system based on block chain and CIM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010383908.0A CN111414504A (en) | 2020-05-08 | 2020-05-08 | Building site flood detection system based on block chain and CIM |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111414504A true CN111414504A (en) | 2020-07-14 |
Family
ID=71493881
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010383908.0A Withdrawn CN111414504A (en) | 2020-05-08 | 2020-05-08 | Building site flood detection system based on block chain and CIM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111414504A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113421299A (en) * | 2021-06-25 | 2021-09-21 | 南京云创大数据科技股份有限公司 | Water depth measuring system and method based on water level gauge and camera |
| CN113657534A (en) * | 2021-08-24 | 2021-11-16 | 北京经纬恒润科技股份有限公司 | Classification method and device based on attention mechanism |
-
2020
- 2020-05-08 CN CN202010383908.0A patent/CN111414504A/en not_active Withdrawn
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113421299A (en) * | 2021-06-25 | 2021-09-21 | 南京云创大数据科技股份有限公司 | Water depth measuring system and method based on water level gauge and camera |
| CN113421299B (en) * | 2021-06-25 | 2024-03-29 | 南京云创大数据科技股份有限公司 | Water depth measuring system and method based on water level gauge and camera |
| CN113657534A (en) * | 2021-08-24 | 2021-11-16 | 北京经纬恒润科技股份有限公司 | Classification method and device based on attention mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tingzon et al. | Mapping poverty in the Philippines using machine learning, satellite imagery, and crowd-sourced geospatial information | |
| Liu et al. | Simulation of urban expansion patterns by integrating auto-logistic regression, Markov chain and cellular automata models | |
| Voigt et al. | Feature-based watermarking of 2D vector data | |
| CN109255743A (en) | Conserve total management system | |
| CN111563433A (en) | Wisdom building site is monitored system of overflowing water based on block chain | |
| CN111428694A (en) | Abnormal smoke monitoring system of wisdom building site based on block chain | |
| CN111414504A (en) | Building site flood detection system based on block chain and CIM | |
| Schmidtlein et al. | Sensitivity of tsunami evacuation modeling to direction and land cover assumptions | |
| CN111611948A (en) | Urban garbage can overflow detection method based on CIM and block chain | |
| Kucsicsa et al. | Urban growth in the Bucharest metropolitan area: Spatial and temporal assessment using logistic regression | |
| CN111414499A (en) | Operation personnel safety belt wearing detection system based on block chain and BIM | |
| CN111611949A (en) | Intelligent urban water area garbage density detection method based on CIM and block chain | |
| Aina et al. | Using multisource data and the VIS model in assessing the urban expansion of Riyadh city, Saudi Arabia | |
| Chen et al. | Classification of soft-story buildings using deep learning with density features extracted from 3D point clouds | |
| CN111414501A (en) | Wood flammability risk prediction system based on block chain and CIM | |
| Su et al. | Visualized multiple image selection encryption based on log chaos system and multilayer cellular automata saliency detection | |
| CN111414503A (en) | Wood accumulation detection system based on block chain and CIM | |
| CN111524131A (en) | Wall crack detection system based on block chain and BIM | |
| CN114913422A (en) | Karst rocky mountain area soil and water loss field monitoring device | |
| CN111523434A (en) | Building site personnel abnormal gathering detection system based on block chain and BIM | |
| Chern et al. | Impact of loss functions on semantic segmentation in far‐field monitoring | |
| WO2021084944A1 (en) | Information processing system, information processing method, imaging device, and information processing device | |
| CN111414500A (en) | Steel wire rope breakage early warning system based on block chain and BIM | |
| CN111476208A (en) | Intelligent building site constructor gathering monitoring system utilizing block chains | |
| CN111553269A (en) | Wisdom building site vehicle blocks up monitoring system based on block chain |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20200714 |