A tunnel jet fan state monitoring and prediction method

Abstract

This invention provides a method for monitoring and predicting the status of tunnel jet fans. The method includes periodically acquiring images of the tunnel and data from jet fan sensors; preprocessing the images and inputting the preprocessed images into a graph neural network model for training; setting the activation conditions of duct ventilation devices and roof ventilation devices based on the output results to monitor the status of the tunnel jet fans; and introducing location information by inputting data with location information into an encoder-decoder model to train the model and predict the status of the tunnel jet fans. The purpose of predicting the jet fan status is to determine the jet fan status using the prediction results when the image acquisition device malfunctions, and to activate additional ventilation devices in the tunnel to reduce jet fan losses when the jet fan is about to operate at high power, thus completing auxiliary monitoring of the jet fan.

Description

Technical Field

[0001] This invention belongs to the field of tunnel monitoring and prediction technology, and relates to a method for monitoring and predicting the status of tunnel jet fans. Background Technology

[0002] Jet ventilation is a device that injects air to enhance airflow in tunnels. Enhanced airflow primarily aims to improve ventilation within the tunnel, increase visibility, reduce harmful gas levels, ensure safety during tunnel use, and enhance comfort. However, jet ventilation has limitations. In special circumstances, such as excessive traffic or accidents within the tunnel, jet ventilation systems may operate under heavy load, shortening the lifespan of the jet fans.

[0003] There are currently some methods for monitoring tunnel jet fans, but these methods have the following shortcomings: (1) Most existing jet fan monitoring systems require manual monitoring, which may lead to deviations in judging the high-power operation time of the jet fan due to different monitoring personnel, resulting in abnormal situations not being detected in time, which in turn reduces the service life of the jet fan; (2) Currently, jet fan monitoring is not equipped with corresponding prediction software and hardware systems. When the monitoring instrument fails, the data predicting the jet fan status can be used to judge the status of the jet fan and monitor the jet fan status in real time; (3) Most existing monitoring methods have low accuracy, poor environmental adaptability, and low degree of automation, resulting in low reliability of monitoring results. Moreover, when the monitoring instrument fails, there is no method to ensure that the jet fan status is monitored during the instrument failure period. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method for monitoring and predicting the condition of tunnel jet fans, which can overcome the problems existing in the background art.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for monitoring and predicting the condition of a tunnel jet fan includes the following steps:

[0007] Step S1: Activate the jet fan inside the tunnel and periodically acquire images inside the tunnel and sensor data from the jet fan.

[0008] Step S2: Preprocess the acquired images and input them into a graph neural network-based model for training;

[0009] Step S3: Automatically set the opening conditions for the duct ventilation device and the top ventilation device based on the training results;

[0010] Step S4: Preprocess the sensor data of the jet fan and input it into the time series model for training.

[0011] Furthermore, step S1 specifically includes:

[0012] Step S101: Activate the jet fan and image acquisition device inside the tunnel to collect training data;

[0013] Step S102: Set the image acquisition frequency for the image acquisition device and save the images collected by the image acquisition device;

[0014] Step S103: Save the sensor data in the jet fan.

[0015] Furthermore, step S2 specifically includes:

[0016] Step S201: Preprocess the image in the image acquisition device by first converting the image into a raster image, and then converting the raster image into a graph structure. The graph structure consists of three types of vectors: node vectors, edge vectors, and global vectors.

[0017] Step S202: Input the preprocessed graph structure into the graph neural network model for training.

[0018] Furthermore, step S3 specifically includes:

[0019] Step S301: Collect the training results in the graph neural network model;

[0020] Step S302: Automatically set the opening conditions for the duct ventilation device and the top ventilation device based on the training results.

[0021] Furthermore, step S4 specifically includes:

[0022] Step S401: Preprocess the sensor data in the jet fan;

[0023] Step S402: Add location information to the preprocessed data and then send it into the time series model for training.

[0024] Furthermore, the vector representation method in step S201 is as follows:

[0025] First, there's the node vector representation. In the bitmap, five different sizes of points are formed based on the vehicle type. These points are represented by numbers 0-4 in the node vector, with larger numbers indicating larger vehicles. The length of the node vector is determined by the number of vehicles in the image. Next, there's the edge vector representation. During the conversion of the bitmap into a graph, a coordinate system is established, and the distances between points are calculated. The elements in the edge vector represent the distances between nodes, and the number of elements in the edge vector is determined by the number of distances between points. Finally, there's the global vector representation, which is obtained by multiplying the node vectors and edge vectors by a set of learnable parameters and then summing them together.

[0026] The formula for calculating the elements in the edge vector is:

[0027]

[0028]

[0029] Where D represents the distance between nodes a and b; (x a ,y a ), (x b ,y b ) represents the position coordinates of nodes a and b in the coordinate system, r1 and r2 are the radii of the two nodes; γ is a learnable parameter, initially set to 1; κ is a scaling parameter, the i-th scaling parameter is κ. i It means that D i This represents the distance from the current node to a certain node. D i [j] represents the sum of distances from the current node to all other nodes, and n represents the number of nodes; after processing by the exp() function, the parameter κ can be controlled between 0 and 1; the distance between the center coordinates of two nodes minus r1 and r2 can be obtained by subtracting r1 and r2 from the distance between the center coordinates of two nodes, but due to the observation angle of the image acquisition device, there is a gap between this distance and the distance between vehicles in reality, so the κ scaling parameter is introduced to reduce the gap; at the same time, a learnable parameter γ is introduced to adjust the scaling parameter during training;

[0030] The formula for expressing the overall vector is:

[0031] a=ω1·v-ω2·e

[0032] Where v represents the node vector, e represents the edge vector, a represents the global vector, and ω1 and ω2 are matrices composed of the learnable parameters corresponding to the node vector and edge vector, respectively; if the lengths of the node vector and edge vector are different, they are padded with 0 to make them of equal length; the larger the element value obtained by adding the negative edge vector to the node vector, the easier it is for the jet fan to operate under high load. Further, the preprocessing in step S401 is as follows:

[0033] If the data collected from the jet fan varies significantly, it is necessary to control the range of data variation to avoid a decrease in the efficiency of training the model using the stochastic gradient descent algorithm. The formula for controlling the range of data variation is:

[0034]

[0035] Where, x i These are the values ​​that show the largest changes in the data collected from the jet fan before the conversion. It is the converted value, x min To x max These are the minimum and maximum values ​​of the data collected in the jet fan before the conversion. and These are the minimum and maximum values ​​of the converted data.

[0036] Furthermore, the formula involving the location information in step S402 is as follows:

[0037]

[0038]

[0039] Among them, Position i,j This represents the position information of row i and column j in a table in the database. The parameter τ represents a random number; τ is odd when i is odd and even when i is even. The parameter σ is a fixed value, assigned to 2. 16 The purpose is to make the value of the sin() function approach 0. The position information of the i-th row and j-th column of the data is encoded into a number approaching 0 through the sin() function and the parameter σ. The parameter τ is used to change the trend of the sin() function and avoid the repetition of position encoding.

[0040] Furthermore, the formulas involved in the timing model in step S402 are as follows:

[0041] Control0 = ReLU(X) t ·W1+H t-1 ·W2+b1)

[0042] Control1 = ReLU(X) t ·W3+H t-1 ·W4+b2)

[0043] …

[0044] Control n =ReLU(X) t ·W 2n-1 +H t-1 ·W 2n +b n )

[0045] Remember t =(Control0) ij ·(Remember t-1 ) ij +(Control0) ij (Control1) ij ·(Remember t-1 ) ij +…

[0046] +(Control0) ij (Control1) ij ·…·(Control n-1 ) ij ·(Remember t-1 ) ij

[0047] Y t =H t =(Remember t ) ij ·(Control n ) ij

[0048] Among them, Control0 to Control n This represents n record unit modules in a time series model, each containing its own learnable parameter tensors W and b, which can respectively adapt to X. t and H t-1 To process and jointly decide on X t With H t-1 Which parts are transmitted to the next time step to calculate Y? t and H t ReLU() is the activation function; Remember t Represents X t With H t-1 The data retained in the memory, when t=0, is remembered. t Set all elements in (Control0) to 1; ij ·(Remember t-1 ) ij The meaning is: multiplying the corresponding elements of two tensors, that is, multiplying the elements corresponding to the i-th row and j-th column.

[0049] The beneficial effects of this invention are:

[0050] (1) This invention introduces a combination of graph neural network and time series model. The graph neural network monitors the status of the jet fan through the image acquisition device; the time series model predicts the status of the jet fan through the sensor data on the jet fan. When the image acquisition device fails, the time series model can be used to predict the status of the jet fan, and the additional ventilation device can be turned on in time to avoid the jet fan from working at high power, ensure the safe use of the tunnel, and increase the service life of the jet fan.

[0051] (2) When training the graph neural network model, this invention preprocesses the input in the graph neural network by introducing a bitmap, converting the image collected in the image acquisition device into a bitmap, reducing the storage space required for the input data, increasing the amount of data that can be stored, and thus improving the running efficiency of the algorithm.

[0052] (3) When training the time series model, the present invention adds a control module to the time series model. Each control module has its own learnable parameters, which increases the generalization of the model and filters the input data. The participation of multiple control modules in the calculation can increase the reliability of the training results.

[0053] (4) When training the time series model, the present invention adds position encoding to the input of the time series model. The position encoding operation can encode the position information of the data into a number close to 0, that is, ensure that the original data has position encoding information. Moreover, since the position encoding is close to 0, the original data can be directly added to the position encoding without affecting the training result. Furthermore, the application of position encoding can preserve the original position information when the data is processed in parallel in multiple control modules, enabling the data to be processed in parallel, thereby improving the training efficiency.

[0054] (5) Traditional jet fan status monitoring and prediction methods are mainly implemented by manual operation. Due to limitations, manual identification may have errors and inaccurate judgments. This invention uses a graph neural network model to intelligently monitor jet fan status changes. It is more sensitive to the judgment of excessive jet fan power and can promptly activate additional ventilation devices to achieve intelligent monitoring of jet fan status.

[0055] (6) Traditional jet fan status monitoring technology does not integrate jet fan status prediction function, which means that when the image acquisition device in the jet fan monitoring equipment fails, the jet fan status cannot be monitored in real time, and thus the jet fan status cannot be predicted. This invention uses a sequence model to predict the jet fan status. When the image acquisition device fails, an additional ventilation device is activated in time to avoid the jet fan from operating at high power.

[0056] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained from the following description and the foregoing claims. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the tunnel jet fan status monitoring and prediction process in this embodiment;

[0058] Figure 2 This is a schematic diagram illustrating the generation and preprocessing of the dot matrix image in this embodiment;

[0059] Figure 3 This is a schematic diagram of the graph neural network model structure in this embodiment;

[0060] Figure 4 This is a schematic diagram of the tunnel internal pipe ventilation device and the top ventilation device in this embodiment;

[0061] Figure 5 This is a schematic diagram of the timing model structure in this embodiment. Detailed Implementation

[0062] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0063] A method for monitoring and predicting the condition of tunnel jet fans, such as Figure 1 As shown, it includes the following steps:

[0064] Step S1: Activate the jet fan inside the tunnel and periodically acquire images inside the tunnel and sensor data from the jet fan.

[0065] Step S2: Preprocess the acquired images and input them into a graph neural network-based model for training;

[0066] Step S3: Automatically set the opening conditions for the duct ventilation device and the top ventilation device based on the training results;

[0067] Step S4: Preprocess the sensor data of the jet fan and input it into the time series model for training.

[0068] In this embodiment, step S1 includes the following sub-steps:

[0069] Step S101: Activate the jet fan and image acquisition device inside the tunnel to collect training data;

[0070] Step S102: Set the image acquisition frequency for the image acquisition device and save the images collected by the image acquisition device;

[0071] Step S103: Save the sensor data in the jet fan.

[0072] In this embodiment, step S2 includes the following sub-steps:

[0073] Step S201: Preprocess the image in the image acquisition device by first converting the image into a raster image, and then converting the raster image into a graph structure. The graph structure consists of three types of vectors: node vectors, edge vectors, and global vectors.

[0074] Step S202: Input the preprocessed graph structure into the graph neural network model for training. The vector representation process in step S201 is as follows: Figure 2 As shown, the specific method is as follows:

[0075] First, there's the node vector. In the dot matrix, five types of points of varying sizes are formed based on vehicle type: The first type is motorcycles and electric vehicles; the second type is family cars, such as sedans, SUVs, and vans; the third type is utility vehicles, which should be avoided, such as fire trucks and ambulances; the fourth type is large transport vehicles, which travel at relatively slow speeds and are prone to causing congestion; the fifth type is construction vehicles, such as cranes, which may appear due to damage within the tunnel requiring repairs, also causing congestion and potentially leading to high-power operation of the jet fan. These points are represented by numbers 0-4 in the node vector. The larger the character, the larger the vehicle size. The length of the node vector is determined by the number of vehicles in the image. For example, the node vector [0,0,4,4,0] means that there are five vehicles in the current image, three of which are of the first type and two of which are of the fifth type. Then there are the edge vectors. During the conversion of the raster image into a graph, a coordinate system is established for the raster image, and the distances between the points are calculated. These distances are the elements in the edge vectors, and the number of elements in the edge vectors is determined by the number of distances between the points. Finally, there is the global vector, which is obtained by multiplying the node vectors and edge vectors by a set of learnable parameters and then adding them together.

[0076] The formula for calculating the elements in the edge vector is:

[0077]

[0078]

[0079] Where D represents the distance between nodes a and b; (x a ,y a ), (xb ,y b ) represents the position coordinates of nodes a and b in the coordinate system, r1 and r2 are the radii of the two nodes; γ is a learnable parameter, initially set to 1; κ is a scaling parameter, the i-th scaling parameter is κ. i It means that D i This represents the distance from the current node to a certain other node. D i [j] represents the sum of distances from the current node to all other nodes, and n represents the number of nodes; after processing by the exp() function, the parameter κ can be controlled between 0 and 1; the distance between the center coordinates of two nodes minus r1 and r2 can be obtained by subtracting r1 and r2 from the distance between the center coordinates of two nodes, but due to the observation angle of the image acquisition device, there is a gap between this distance and the distance between vehicles in reality, so the κ scaling parameter is introduced to reduce the gap; at the same time, a learnable parameter γ is introduced to adjust the scaling parameter during training;

[0080] The formula for expressing the overall vector is:

[0081] a=ω1·v–ω2·e

[0082] v=[2,4,2,2,1],e=-[1,2,1,1,1,2,1,1]

[0083] v=[2,4,2,2,1,0,0,0],e=-[1,2,1,1,1,2,1,1]

[0084] ω1=[1,2,2,4,7,8,1,5],ω2=[1,1,2,1,4,5,3,4]

[0085] a = [2]

[0086] Where v = [2,4,2,2,1] represents the node vector, e = -[1,2,1,1,1,2,1,1] represents the edge vector, and ω1 and ω2 are vectors composed of the learnable parameters corresponding to the node vector and the edge vector, respectively. If the node vector and the edge vector have different dimensions and lengths, they are padded with 0 to make the dimensions and lengths equal, that is, v is padded to [2,4,2,2,1,0,0,0]. The larger the overall value in v (the more large vehicles there are, the greater the possibility of traffic obstruction in the tunnel), the smaller the overall value in e (the shorter the distance between vehicles, the greater the possibility of traffic accidents in the tunnel), the easier it is to cause the jet fan to work at high power. When a is expressed by v and e together, the value in the edge vector e is negative. The larger the element value obtained after adding the negative value to the node vector, the easier it is to cause the jet fan to work at high power. In this embodiment, the overall vector a = [2] is calculated, which is small, indicating that the possibility of the jet fan working at high power in the tunnel is small.

[0087] The neural network model diagram in step S202 is shown below. Figure 3 As shown, its workflow is as follows:

[0088] First, after obtaining the preprocessed graph using the graph neural network model, vectors are created for nodes, edges, and the overall vectors of nodes and edges. In PyTorch, the `dgl` package is used to create the graph neural network model. `dgl.DGLGraph()` is used to create the graph, and `g.add_nodes()` is used to create node vectors, and `g.add_edges()` is used to create edge vectors. The node vectors and edge vectors are multiplied by their weights to obtain the overall vector. These vectors are then input into the hidden layer. The `defforward()` function defines how these vectors are computed in the hidden layer. In this example, a fully connected neural network is used as the hidden layer, employing `nn.Parameter(torch.randn(num_inputs, num_hiddens)` and `nn.Parameter(torch.randn(num_inputs, num_hiddens)`. `randn(num_hiddens, num_outputs)` creates the weights in the hidden layer, and `nn.Parameter(torch.zeros(num_hiddens, require_grad = True))` creates the biases in the hidden layer. The output structure is similar to the input, but the internal features of the output graph are different. Finally, the `Softmax()` function is defined to classify the graph, and `defsoftmax(x, y, z)` processes the nodes, edges, and the overall vector separately. During training, the loss function is set to cross-entropy. In this invention, another criterion for evaluating the model's performance is set: manual sampling is used to check whether the model correctly judges whether the duct ventilation device and the top ventilation device should be turned on. The result of manual sampling is called accuracy.

[0089] The optimization process can utilize the simple, easy-to-use, and computationally inexpensive Adam optimization algorithm to optimize the model. If training shows underfitting, consider labeling the images to assist model training. For example: if congestion in the tunnel is caused by special vehicles (such as ambulances, fire trucks, or transport vehicles), causing the jet fans to operate at high power, the image can be labeled "Huge"; if congestion in the tunnel is caused by a car accident, the jet fans may operate at high power, so the image can be labeled "TA" (Traffic accident); if all vehicles avoid driving on the right, but there are no vehicles on the right, it may be due to construction on the right, causing slow vehicle movement in the tunnel and the jet fans to operate at high power, so the image can be labeled "Eg" (Engineering).

[0090] In this embodiment, step S3 includes the following sub-steps:

[0091] Step S301: Collect the training results in the graph neural network model;

[0092] Step S302: Automatically set the opening conditions for the duct ventilation device and the top ventilation device based on the training results. Whether to open or not is determined by comparing the node vector, edge vector, and overall vector of the acquired image with the classification results of the Softmax() function in step S2. For example: the current graph node vector returned by the image acquisition device is (0,4,2,0,1,1,2,3), the edge vector is -(1,2,1,1,1,2,1,1), ω1 is (1,2,2,4,7,8,1,5), and the overall vector of ω2 is (1,1,2,1,4,5,3,4). The overall vector calculation result is (17). According to the previous training results of the graph neural network, the node vector shows that there is currently a fifth type of engineering vehicle (represented by 4) in the tunnel. The edge vector shows that the distance between the vehicles is relatively close, and the overall vector value is relatively high. Therefore, it is judged that the jet fan may work at high power, and the duct ventilation device and the top ventilation device need to be opened. The schematic diagram of the duct ventilation device and the top ventilation device is shown below. Figure 4 As shown.

[0093] In this embodiment, step S4 includes the following sub-steps:

[0094] Step S401: Preprocess the sensor data in the jet fan;

[0095] Step S402: Add location information to the preprocessed data and then send it into the time series model for training.

[0096] Furthermore, the preprocessing in step S401 is as follows:

[0097] In this embodiment, date-related information can be supplemented with fields [year, month, day, day_of_year, day_of_week]. These fields allow for quick identification of important information. For example, day_of_week indicates whether today is a weekday or a weekend, and day_of_year indicates whether today is a holiday. This information can be used to observe changes in the status of the jet fan within the tunnel on weekends, holidays, and weekdays. If the data collected from the jet fan varies significantly, for example, the jet fan's power varies from 0-60W, but its output airflow varies from 0-600,000 cm³ / h. 3 If the output airflow is / h, then it is necessary to control the range of data variation so that the range of output airflow variation is similar to that of power, to avoid a decrease in the efficiency of training the model using the stochastic gradient descent algorithm. The formula for controlling the range of data variation is:

[0098]

[0099] Where, xi These are the values ​​that show the largest changes in the data collected from the jet fan before the conversion. It is the converted value, x min To x max These are the minimum and maximum values ​​of the data collected in the jet fan before the conversion. and These are the minimum and maximum values ​​of the converted data. For the above airflow changes, the output airflow x i Is in x min =0, x max Given a number between x and 600000, x i =540000cm 3 / h, settings Control the airflow variation between 0-100, after conversion

[0100] Furthermore, the formula involving the location information in step S402 is as follows:

[0101]

[0102]

[0103] Among them, Position i,j This represents the position information of row i and column j in a table in the database. The parameter τ represents a random number; τ is odd when i is odd and even when i is even. The parameter σ is a fixed value, assigned to 2. 16 The goal is to make the value of the sin() function approach 0. The position information of the i-th row and j-th column of the data is encoded into a number approaching 0 through the sin() function and the parameter σ. The parameter τ is used to change the trend of the sin() function and avoid duplicate position encoding.

[0104] Furthermore, the formulas involved in the timing model in step S402 are as follows:

[0105] Control0 = ReLU(X) t ·W1+H t-1 ·W2+b1)

[0106] Control1 = ReLU(X) t ·W3+H t-1 ·W4+b2)

[0107] …

[0108] Control n =ReLU(X) t ·W 2n-1 +H t-1 ·W 2n +bn )

[0109] Remember t =(Control0) ij ·(Remember t-1 ) ij +(Control0) ij (Control1) ij ·(Remember t-1 ) ij +…

[0110] +(Control0) ij (Control1) ij ·…·(Control n-1 ) ij ·(Remember t-1 ) ij

[0111] Y t =H t =(Remember t ) ij ·(Control n ) ij

[0112] Among them, Control0 to Control n This represents n record unit modules in a time series model, each containing its own learnable parameter tensors W and b, which can respectively adapt to X. t and H t-1 To process and jointly decide on X t With H t-1 Which parts are transmitted to the next time step to calculate Y? t and H t ReLU() is the activation function; Remember t Represents X t With H t-1 The data retained in the memory, when t=0, is remembered. t Set all elements in (Control0) to 1; ij ·(Remember t-1 ) ij The meaning is: multiplying the corresponding elements of two tensors, that is, multiplying the elements corresponding to the i-th row and j-th column.

[0113] The timing model diagram in step S402 is as follows: Figure 5 As shown, its workflow is as follows:

[0114] This embodiment uses three control modules as an example to describe the workflow. First, the preprocessed data is added with position information and then input into the encoder. The entire encoder consists of multiple encoder blocks. Each encoder block contains a control module and a feedforward neural network, which transforms the input into a hidden state and sends it to the decoder. The decoder is similar to the encoder, also containing a control module and a feedforward neural network, but it also needs to receive the output from its previous time step. Assume X... t H t-1 The control module is represented in PyTorch as a 2x3 tensor:

[0115] Control0=torch.relu(torch.matmul(X t ,W1)+torch.matmul(H t-1 ,W2)+b1))

[0116] Control1=torch.relu(torch.matmul(X t ,W3)+torch.matmul(H t-1 ,W4)+b2))

[0117] Control2=torch.relu(torch.matmul(X t ,W5)+torch.matmul(H t-1 ,W6)+b3))

[0118] Remember0 = torch.ones(2,3)

[0119] Remember1=torch.matmul(Control0,Remember0)

[0120] Remember2=torch.matmul(Control0,Remember1)+torch.matmul(torch.matmul(Control0,

[0121] Control1), Remember1)

[0122] Y t =H t =torch.matmul(Remember2,Control2)

[0123] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0127] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method of tunnel-ducted-fan condition monitoring and prediction, characterized by, The method comprises the following steps: Step S1, enabling the jet fan and the image collector in the tunnel, regularly acquiring the jet fan sensor data and the images in the tunnel; Step S2, pre-processing the collected images, inputting the images into a graph neural network model for training, and completing the state monitoring of the jet fan; The step S2 specifically comprises: Step S201, pre-processing the images in the image collector, first converting the images into dot matrix images, and then converting the dot matrix images into graph structures, the graph structures being composed of three vectors: a node vector, an edge vector and an overall vector; Step S202, inputting the pre-processed graph structures into the graph neural network model for training; The vector representation method in the step S201 is as follows: Firstly, the node vector representation, in the dot matrix image, five types of points with different sizes are formed according to the types of vehicles, the above nodes are represented by numbers 0-4 in the node vector, the larger the number, the larger the size of the vehicle, and the dimension of the node vector is determined by the number of vehicles in the image; then the edge vector representation, a coordinate system is established for the dot matrix image in the process of converting into a graph, the distance between points is calculated, and the elements in the edge vector are represented by the distance between nodes, the number of elements in the edge vector is determined by the number of distances between points; finally, the overall vector representation, the overall vector is obtained by adding the node vector and the edge vector multiplied by a set of learnable parameters respectively; The calculation formula of the element in the edge vector is: wherein, D represents the distance between nodes a and b ; x a , y a ),( x b , y b ) are the position coordinates of nodes a and b in the coordinate system, r 1 and r 2 are the radii of the two nodes; The expression formula of the overall vector is: is a learnable parameter, and the initial value is 1; is a scale parameter, and the first i scale parameter is denoted by , wherein D i represents the distance from the current node to a certain node, represents the sum of the distances from the current node to all other nodes, n represents the number of nodes; after processing by the Step S3, automatically setting the opening conditions of the pipeline ventilation device and the top ventilation device according to the training result; () function, the parameter can be controlled between 0 and 1; the distance between the two node centers in the coordinate system can be obtained by subtracting r 1 and r 2, but because of the observation angle of the image collector, there is a gap between this distance and the distance between the vehicles in reality, so the scale parameter is introduced to reduce the gap; at the same time, the learnable parameter Step S4, pre-processing the sensor data of the jet fan, inputting the data into a time series model for training, and completing the state prediction of the jet fan. is introduced to adjust the scale parameter during the training process; The step S1 specifically comprises: Step S101, enabling the jet fan and the image collector in the tunnel, collecting training data; 1• Step S102, setting the frequency of collecting images for the image collector, and saving the collected images in the image collector; 2• e wherein, v represents a node vector, e represents an edge vector, a represents an overall vector, Step S103, saving the sensor data in the jet fan. 1 and The step S3 specifically comprises: 2 are matrices composed of learnable parameters corresponding to the node vector and the edge vector, respectively; if the dimensions of the node vector and the edge vector are different, 0 is supplemented to the same dimension; the greater the element value obtained by taking the negative of the edge vector and adding the node vector, the easier the jet fan works under high load. Step S301, collecting the training result in the graph neural network model; Step S302, automatically setting the opening conditions of the pipeline ventilation device and the top ventilation device according to the training result.

2. A method of tunnel-duct fan condition monitoring and prediction according to claim 1, wherein, The step S4 specifically comprises: Step S401, pre-processing the sensor data in the jet fan; Step S402, adding position information to the pre-processed data and sending the data into a time series model for training. The formula related to the position information in the step S402 is:

3. A method of tunnel-duct fan condition monitoring and prediction according to claim 1, wherein, Step S3, automatically setting the opening conditions of the pipeline ventilation device and the top ventilation device according to the training result; Step S4, pre-processing the sensor data of the jet fan, inputting the data into a time series model for training, and completing the state prediction of the jet fan. The step S1 specifically comprises:

4. The method of claim 1, wherein, Step S101, enabling the jet fan and the image collector in the tunnel, collecting training data; Step S102, setting the frequency of collecting images for the image collector, and saving the collected images in the image collector; Step S103, saving the sensor data in the jet fan.

5. A method of tunnel-duct fan condition monitoring and prognostics according to claim 4, wherein, The step S3 specifically comprises: Step S301, collecting the training result in the graph neural network model; = wherein, Step S302, automatically setting the opening conditions of the pipeline ventilation device and the top ventilation device according to the training result. i,j represents the position information of the position on the table in the database as the first i row and the first j column, the parameter The step S4 specifically comprises: represents a random number, when i is odd Step S401, pre-processing the sensor data in the jet fan; takes an odd number, when i is even Step S402, adding position information to the pre-processed data and sending the data into a time series model for training. takes an even number, the parameter The formula related to the position information in the step S402 is: is a fixed value, and is assigned a value of 2 16 , the purpose is to make ​ () function value tends to 0, the position information of the first i row j column of data is encoded as a number that tends to 0 through ​ () function and parameter ​ , the parameter ​ is used to change the trend of ​ () function, to avoid position encoding duplication.

Images