CN108733893A - The public building burst pollution of coupling depth learning method is traced to the source - Google Patents

Abstract

The present invention provides a method for tracing the source of sudden pollution in public buildings coupled with a deep learning method, which includes: using the forward simulation of the computational fluid dynamics model and the concentration value of the hypothetical sensor position obtained from monitoring to define the traceability model family and obtain the matching pair set; expand the traceability method in the traceability model family into a deep neural network; train and test the deep neural network; finally solve the problem of sudden pollution traceability; the present invention aims at the uncertainty of sudden pollution of public buildings and the timeliness of traceability. Introducing the deep learning method into the traceability theory of sudden pollution in public buildings greatly improves the timeliness of the traceability problem and also lays the foundation for engineering application of the building emergency technology system; in addition, this method couples the accuracy of the model-based solution with the The timeliness of data deep learning allows data optimization based on matching in the traceability model family, which can not only tolerate the uncertainty in modeling and traceability, but also output accurate traceability results.

Description

Traceability of sudden pollution in public buildings coupled with deep learning method

technical field

The invention belongs to the technical field of traceability of sudden pollution of buildings, and in particular relates to a traceability of sudden pollution of public buildings coupled with a deep learning method.

Background technique

In recent years, terrorists have begun to use highly lethal, unconventional biological and chemical weapons to attack buildings and civilian facilities, and have challenged the health and safety of human living environments through the deliberate release of chemical or biological warfare agents.

Due to the high density of people in most public buildings, it may become a potential target of terrorists. Based on the requirements of building environment safety and health control, when an air pollution incident occurs, it is necessary to reversely search for source information such as the pollution emission location and pollution intensity in the building environment, so as to remove the pollution source as soon as the incident occurs and implement Emergency ventilation and other means to achieve rapid and efficient pollution hazard control. Compared with solving the propagation process of pollutants in the building environment under the condition of known pollution source information, pollution source traceability is a reverse process. It does not stop at exceeding the standard or safety alarm when the sensor monitors the pollution concentration information, but needs to be further traced. , Timely and accurate identification of pollution source information. However, in the actual pollution traceability process, the data usually come from the monitoring results of pre-arranged sensors. Due to the limited number of sensors and the data may contain human and instrument errors, it may be difficult to solve the problem when using such data to trace the source of pollution. Existence, non-unique, and discontinuously dependent data issues.

The solution of the traditional pollution traceability problem is based on the model. First, it is necessary to use domain knowledge to model the air pollution transmission process inside the building in detail, and then develop an accurate and efficient simulation method based on the sensor monitoring values under the determined model. Finally, Combining various prior conditions to select the optimal variational form, the solution of the traceability problem is obtained. Such completely model-based inverse problem solving methods are full of difficulties and uncertainties. First of all, it is difficult to achieve accurate modeling. Even under various approximations and assumptions, approximate modeling is possible, but it is usually quite difficult; second, the development of high-precision forward modeling (usually represented by For the numerical solution of a partial differential equation), it is inevitable to introduce new approximations; finally, due to many difficulties such as the existence and uncertainty of observation errors, and the difficulty in describing the prior conditions of the solution, the direct model-based traceability method takes a long time and is difficult to apply in actual engineering. Another type of traceability methods, such as genetic algorithm and particle swarm optimization algorithm, have fast solution speed and do not depend on the model (mainly depends on data quality), but their effectiveness essentially depends on the hypothesis space, and the determination of the hypothesis space is often random. It is easy to cause a large deviation between the traceability results and the actual pollution source parameters. How to make full use of the accuracy of the traditional model-based traceability method and the timeliness of the data-based traceability method at the same time, there is currently no public technology that can be used for reference.

Deep learning methods are widely used in the fields of image recognition and speech recognition. Common deep learning methods such as Convolution Neural Networks (CNN) for image processing and Recurrent Neural Networks (Recurrent Neural Networks) for sequence data processing , RNN). The convolutional neural network continuously stacks the convolutional layer and the pooling layer to form several layers of fully connected layers. The convolutional layer can capture regional connection characteristics, and the application of the weight sharing principle greatly reduces the number of parameters to be trained by the model. Pooling Layers combine similar features by merging adjacent nodes, reducing the amount of training data. The cyclic neural network introduces directional loops on the basis of the traditional feedforward neural network. The nodes between the hidden layers are connected, and the input of the hidden layer includes not only the output of the input layer, but also the output of the hidden layer at the previous moment. . Compared with data-based traceability methods such as legacy algorithms, the multi-hidden-layer artificial neural network provided by deep learning methods has better feature learning capabilities and higher timeliness, which can be expanded during the construction of traceability methods. It is a W-layer deep neural network, and uses the feature learning and identification of the neural network to realize the identification of pollution source information. At present, there is no public technology that can be used for reference to apply deep learning methods to the field of building accident pollution traceability.

Contents of the invention

The present invention aims at the deficiencies in the prior art, and aims to provide a traceability of sudden pollution of public buildings coupled with a deep learning method.

To achieve the above object, the solution of the present invention is:

A source of sudden pollution in public buildings coupled with a deep learning method, which includes the following steps:

(1) Construct a traceability model family and obtain a matching pair set;

(2) Expand the traceability method in the traceability model family into a deep neural network;

(3), training and testing deep neural networks;

(4) Trace the source of sudden pollution.

Preferably, in step (1), the specific process of constructing a traceability model family and obtaining a matching pair set is as follows:

(1.1), according to different building types and structures, use the technology of computational fluid dynamics (Computational FluidDynamics, CFD) model to establish the corresponding physical model and carry out the forward simulation of sudden pollution propagation, obtain the observation operator, and monitor the The concentration value at the hypothetical sensor position of is the observed value, and the observed operator and the observed value jointly define an optional traceability model family;

(1.2), the traceability method in the traceability model family takes the concentration value (i.e. the actual observed value C) of the actual sensor position as the input value, and uses the final pollution source parameter S solved by the iterative method as the output value;

(1.3), A set of matching pairs G=(C ₁ , S ₁ ), (C ₂ , S ₂ ), ..., (C _N , S _N );

(1.4) Randomly split the matching pair set G into two parts: the training set G _Train and the test set G _Test .

Preferably, the matching pairs in the matching pair set G all obey the unknown distribution and are independent and identically distributed.

Preferably, in step (2), the specific process of developing the traceability method into a deep neural network (Deep Neural Network, DNN) is:

The deep neural network (DNN) determines the hypothesis space of the learning process through the traceability model family in step (1), and determines the topology structure of the deep neural network (DNN) through the traceability method in the traceability model family.

Preferably, the deep neural network (DNN) is composed of W layer operation units in series, and each operation unit includes four network layers: traceability layer, convolution layer, nonlinear transformation layer and multiplier update layer.

Preferably, in step (3), the concrete process of training and testing deep neural network (DNN) is:

After using the empirical risk minimization method to train the deep neural network based on the training set, it is then tested based on the test set.

Among them, the training set training DNN is to determine the optimal connection weights of each layer of DNN, and its main function is to learn and train the distribution implied by each matching pair.

The empirical risk minimization method is solved by the reverse gradient method.

Preferably, in step (4), the specific process of solving sudden pollution traceability is as follows:

For the actual observed value C of the input concentration of the deep neural network after the test, the output DNN (C) is used as the final solution of the traceability problem.

Owing to adopting said scheme, the beneficial effect of the present invention is:

First, the present invention provides a pollution traceability method based on model solving and deep learning based on data. This method allows data to be optimized based on matching in the traceability model family, and can tolerate uncertainty in modeling and traceability It can also output accurate traceability results.

Second, the traceability method of the present invention can be expanded into a W-layer deep neural network DNN. The topology of the DNN is uniquely determined by the forward simulation and the model-based traceability method. The introduction of DNN greatly improves the solution efficiency of the traceability method, making it possible for The rapid traceability and engineering of building pollution incidents become possible.

In short, the present invention aims at the uncertainty of sudden pollution in public buildings and the timeliness of traceability. For the first time, the deep learning method is introduced into the theory of traceability of sudden pollution in public buildings. The application basis of public building sudden pollution emergency technical system engineering supported by monitoring network and traceability theory and method.

Description of drawings

Fig. 1 is a schematic flow chart of tracing the source of sudden pollution in public buildings coupled with the deep learning method of the present invention.

Fig. 2 is a schematic diagram of the topological structure of the deep neural network DNN in the traceability of sudden pollution of public buildings coupled with the deep learning method of the present invention.

Fig. 3 is a schematic structural diagram of a three-dimensional building model in an embodiment of the present invention.

FIG. 4 is a schematic two-dimensional plan view of a three-dimensional architectural model in an embodiment of the present invention.

Reference signs: SA1-air-conditioning air supply system, RA1-first air-conditioning return air system, RA2-second air-conditioning return air system, a-standard height of air-conditioning ventilation system to ground and b-standard height of target public building.

Detailed ways

The invention provides a method for tracing the source of sudden pollution in public buildings coupled with a deep learning method.

<Tracing the Source of Sudden Pollution in Public Buildings by Coupling Deep Learning Method>

As shown in Figure 1, the traceability of public building sudden pollution of the coupling deep learning method of the present invention includes the following steps:

(1), Construct the traceability model family and obtain the matching pair set stage:

(1.1) According to different building types and structures, use Computational Fluid Dynamics (CFD) model technology to establish corresponding physical models and carry out forward simulation of sudden pollution propagation, and obtain observation operators, while monitoring The concentration value at the hypothetical sensor position is the observation value, and the observation operator and the observation value jointly define an optional traceability model family;

(1.2), the traceability method in the traceability model family takes the concentration value (i.e. the actual observed value C) of the actual sensor position as the input value, and uses the final pollution source parameter S solved by the iterative method as the output value;

(1.3), the matching pair set G of pollution source parameters and observation values in a group of actual public buildings has been obtained through the observation operator;

(1.4) Randomly split the matching pair set G into two parts: the training set G _Train and the test set G _Test .

That is, the matching pair set G is:

G＝G _Train UG _Test

Among them, the matching pairs in the matching pair set G all obey the unknown distribution and are independent and identically distributed.

The above-mentioned matching pair set G is split into a training set G _Train and a test set G _Test to learn the network parameters of the deep neural network.

(2) Expand the traceability method into a deep neural network stage:

The deep neural network (DNN) determines the hypothesis space of the learning process through the traceability model family in step (1), and determines the topology structure of the deep neural network (DNN) through the traceability method in the traceability model family.

Let ^S∈SN be the possible pollution source information, and C∈C ^N '(N'<N) be the discontinuous sampling information of the sensor. According to the air pollution propagation theory, the air pollution traceability model can be constructed as the following optimization problem:

Among them, arg min{} represents the collection of all pollution source information S when the function in {} obtains the minimum value, D is the observation operator, A _i is the linear transformation, j( ) is the non-convex regularization function, and λ _i is Weight coefficient, I is a possible variable grouping parameter, all parameters (A _i , j(·), λ _i , I) in the optimization problem are uncertain and optional. The observation operator D formed by CFD technology, the observed value C formed by monitoring the concentration value of the hypothetical sensor position, and the above model parameters define a traceability model family. In this traceability model family, when I≤N, The corresponding regularizer is The regularizer describes the traceability process under the discontinuous prior.

The above optimization problem is mapped to a W-layer deep neural network. The topological structure of the deep neural network (DNN) is shown in Figure 2. It is composed of W-layer operation units in series, and each operation unit includes 4 network layers: traceability layer, convolutional layer, nonlinear transformation layer, and multiplier update layer.

Among them, the traceability layer S ⁽ⁿ⁾ implements the traceability operation of the traceability model family. After the operation of the traceability layer, the pollution source information S can be input and output through the given observation value C; the convolution layer H ⁽ⁿ⁾ performs the convolution operation; The nonlinear transformation layer L ⁽ⁿ⁾ realizes the nonlinear projection function operation; the multiplier update layer R ⁽ⁿ⁾ realizes the multiplier update operation in the deep neural network.

(3), the training and testing phase of the deep neural network:

First of all, it is necessary to initialize each parameter in the deep neural network, take the initial linear transformation A _i sine transformation, and initialize the regularizer as The corresponding nonlinear projection function is a soft threshold function, and the rest of the network parameters are initialized randomly. Then take a loss function as the optimization goal, and its corresponding empirical risk is:

Among them, n is the matching logarithm contained in the training set G _Train , θ=(θ ₁ ,θ ₂ ,...θ _W ) is the weight of each layer of DNN, y is the loss function, usually taken as the least square function Among them, ||·‖ represents a functional.

Therefore, training a DNN means learning to solve the following optimization problem:

The above learning problem is solved using the reverse gradient descent method.

In fact, the calculation formula of the reverse gradient descent method is:

Among them, M is the empirical risk to be sought, θ is a possible pollution source parameter, f is an intermediate variable, w is the total number of pollution source parameters (such as location, intensity, initial release time, etc.), and j is the jth pollution source parameter.

The trained DNN is tested for its effectiveness by the test set G _Test .

(4) The solution stage of sudden pollution traceability:

In the traceability stage of sudden pollution, for the tested deep neural network input concentration actual observation value C, the output DNN(C) is used as the final solution of the traceability problem, which contains all the information of the pollution source (namely, location, intensity and Release the initial moment).

The present invention will be further described below in conjunction with embodiment.

Example:

The traceability of sudden pollution in public buildings coupled with the deep learning method of this embodiment includes the following steps:

(1) Construct the three-dimensional CFD geometric model of the target public building, as shown in Figure 3, the area shown by the cuboid is the large space of the target public building, and the pipeline above is the air conditioning and ventilation system equipped with the public building, and the air conditioning and ventilation system is one of the standing The air-conditioning ventilation system may not only be a potential transmission channel of air pollution, but also serve as a discharge channel of air pollution afterwards. In fact, the air-conditioning ventilation system includes the air-conditioning air supply system SA1, the first air-conditioning return air system RA1 and the second air-conditioning return air system RA2, wherein the total air volume of the air-conditioning air supply system SA1 is 15800m ³ /h, and the first air-conditioning return air system The total air volume of the system RA1 and the second air-conditioning return system RA2 is 6300m ³ /h, the standard height a of the air-conditioning ventilation system to the ground is 5.0m, and the standard height b of the target public building is 4.5m. There are five air pollution monitoring sensors (indicated by black triangles) in the three-dimensional public building space, and the specific positions are: D1 (6.0, 22.5, 0.2), D2 (20.0, 7.5, 0.5), D3 (20.0, 22.5, 1.5), D4 (12.5, 15.0, 5.2) and D5 (6.0, 7.5, 2.5).

(2) As shown in Figure 4, at the 36 potential pollution source locations (indicated by black solid circles) and actual pollution source locations (indicated by black hollow circles) in the forward simulation stage of CFD technology, it is assumed that air pollutants are instantaneous State release, the simulation time of CFD technology is 300s, and each air pollution monitoring sensor returns 5 monitoring data in one monitoring cycle, then the obtained pollution source parameters and monitoring concentration form a total of 5×5×36 matching pairs, the matching pairs The set contains 900 matching pairs of data, 600 of which are used as training data, and the remaining 300 matching pairs of data are used as test data.

(3), get the degree of depth W=20 of depth neural network DNN, the number of neurons of each layer is 4 (being 4 network layers): traceability layer, convolution layer, non-linear transformation layer and multiplier update layer, each The layer topology is shown in Figure 2.

(4) The concentration data of the monitoring sensor is used as the input value, and the pollution source parameters in the normalized form are used as the output value, which is the final solution of the traceability problem.

The obtained results were compared with those of the traditional model-based backward probability method traceability and the data-based genetic algorithm traceability for the same traceability problem. The mean square error (MSE) was used as the comparison criterion. The comparison results are shown in Table 1.

Table 1 Comparison of traceability results of different traceability methods for the same pollution source

It can be seen from Table 1 that the coupling deep learning method traceability of the present invention is significantly better than the genetic algorithm and the backward probability method in terms of traceability accuracy and solution speed. Model selection method, which greatly improves the traceability solution speed while ensuring the traceability accuracy, which is of great value in solving the problem of timeliness of traceability in practice, and also verifies the feasibility of coupling deep learning method to trace the source of sudden pollution in public buildings and efficiency.

The above description of the embodiments is for those of ordinary skill in the art to understand and use the present invention. It is obvious that those skilled in the art can easily make various modifications to these embodiments, and apply the general principles described here to other embodiments without creative efforts. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the principles of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims (7)

1. a kind of public building burst pollution of coupling deep learning method is traced to the source, it is characterised in that：It includes the following steps：

(1), it builds and traces to the source model race and obtain a pairing set；

(2), the source tracing method in the model race that traces to the source is launched into deep neural network；

(3), train and test the deep neural network；

(4), burst pollution is traced to the source solution.

2. the public building burst pollution of coupling deep learning method according to claim 1 is traced to the source, it is characterised in that： In step (1), the detailed process for building trace to the source model race and acquisition pairing set is：

(1.1), the computational fluid dynamics model of public building is built, and with the technology pair of the computational fluid dynamics model Public building burst pollution carries out positive simulation and obtains Observation Operators, and the concentration value for monitoring the hypothesis sensing station of gained is For observation, the Observation Operators constitute the model race that traces to the source with the observation；

(1.2), the source tracing method in the model race that traces to the source is using the monitoring concentration of actual sensing station as input value, fortune Use the final pollution sources parameter of solution by iterative method as output valve；

(1.3), the matching pair of pollution sources parameter and the observation in practical public building has been obtained by the Observation Operators Collection；

(1.4), described pairing set is split as training set and test set.

3. the public building burst pollution of coupling deep learning method according to claim 1 or 2 is traced to the source, feature exists In：The matching is to all matchings of concentration to obeying unknown distribution and being independent same distribution.

4. the public building burst pollution of coupling deep learning method according to claim 1 is traced to the source, it is characterised in that： In step (2), the detailed process that source tracing method is launched into deep neural network is：

The deep neural network determines the hypothesis space of learning process by the model race that traces to the source described in step (1), passes through institute State the topological structure that the source tracing method in tracing to the source model race determines the deep neural network.

5. the public building burst pollution of coupling deep learning method according to claim 4 is traced to the source, it is characterised in that： The deep neural network is composed in series by multilayer operation unit, and each operating unit includes four network layers：Trace to the source layer, Convolutional layer, nonlinear transformation layer and multiplier update step.

6. the public building burst pollution of coupling deep learning method according to claim 1 or 2 is traced to the source, feature exists In：In step (3), the detailed process of the training and test depth neural network is：

After training the deep neural network based on the training set field experience risk minimization method, then it is based on the test Collection is tested；

The empirical risk minimization method is solved by reversed gradient method.

7. the public building burst pollution of coupling deep learning method according to claim 1 is traced to the source, it is characterised in that： In step (4), the trace to the source detailed process of solution of burst pollution is：

Actual concentrations observation is inputted for the deep neural network after test and exports the last solution for problem of tracing to the source.