Real-time reconstruction method of temperature field of heat source system based on limited measuring points

Abstract

The application discloses a kind of heat source system temperature field real-time reconstruction methods based on limited measuring point, comprising: obtaining multiple training data, training data includes monitoring data and prediction data, monitoring data includes the position and temperature of multiple monitoring points in a component layout, and prediction data includes the position and temperature of at least one prediction point in the same component layout;Vector mapping-based machine learning model is constructed;The mapping relationship between the temperature of monitoring point and the temperature of prediction point is fitted using training data to train vector mapping-based machine learning model;The temperature of prediction point is predicted using trained machine learning model.The method of the application can obtain the temperature of other positions under the component layout using the temperature measured by limited monitoring points, realize the rapid prediction and real-time prediction of the temperature of prediction point under different component layouts, reduce the number of temperature sensors needed to be installed in the component layout area, and reduce the space requirement of sensor installation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device thermal control, and particularly relates to a real-time reconstruction method of a temperature field of a heat source system based on limited measuring points. BACKGROUND

[0002] In existing heat source systems such as microelectronic devices, due to the large number of internal heat generating components and the dense distribution of the heat generating components, the temperature of the heat source system is high and the temperature change is large. At the same time, the heat source system will also adjust the component power in real time according to the changes of the environment and the tasks performed during the normal operation stage, which will also cause the temperature distribution of the entire heat source system to change frequently. Since electronic components are very sensitive to temperature changes, once the temperature of some areas in the heat source system exceeds the working temperature range of the components, the heat source system will be difficult to continue to work, and even may be damaged. Therefore, it is necessary to monitor the temperature of the internal area of the heat source system in real time.

[0003] At present, the temperature monitoring of the internal area of the heat source system mainly adopts a contact type temperature monitoring method, that is, temperature sensors are installed at some positions in the heat source system, and the temperature sensors are used for real-time measurement of the temperature. However, due to the limitation of the internal space of the heat source system, the number of sensors installed is usually limited, and if too many sensors are installed, it will seriously affect the system design and cause unnecessary system redundancy. Therefore, the existing contact type temperature monitoring method can only realize heat monitoring of a small number of positions equipped with sensors, and cannot realize real-time temperature monitoring of other positions of the heat source system. When there are many positions in the heat source system that need to be monitored, it is difficult to meet the actual application requirements. SUMMARY

[0004] In order to solve the above-mentioned technical problems in the prior art, the present application provides a real-time reconstruction method of a temperature field of a heat source system based on limited measuring points.

[0005] The technical scheme of the present application is as follows:

[0006] A real-time reconstruction method of a temperature field of a heat source system based on limited measuring points is provided, and the method comprises:

[0007] Obtaining a plurality of training data, wherein the training data comprises monitoring data and prediction data, the monitoring data comprises the positions and temperatures of a plurality of monitoring points in a component layout, and the prediction data comprises the positions and temperatures of at least one prediction point in the same component layout;

[0008] Constructing a machine learning model based on vector mapping;

[0009] training the vector mapping based machine learning model with the training data to fit a mapping relationship between the temperature of the monitoring points and the temperature of the prediction points;

[0010] performing temperature prediction of the prediction points by using the trained machine learning model.

[0011] In some possible implementation manners, the obtaining the plurality of training data comprises:

[0012] determining positions of the monitoring points and the prediction points in a component layout region of the heat source system, and determining positions of the components in the component layout region;

[0013] for each component, keeping the component position unchanged, randomly sampling a power value from a preset power range corresponding to each component to determine the power of each component, obtaining a component layout, simulating and calculating a temperature field corresponding to the component layout by using a finite element method, extracting the temperature of the monitoring points and the temperature of the prediction points from the temperature field, determining a monitoring data comprising the positions and the temperatures of the monitoring points and a prediction data comprising the positions and the temperatures of the prediction points, and obtaining a training data comprising the monitoring data and the prediction data, and repeating the random sampling and the extracting until a preset number of training data is obtained.

[0014] In some possible implementation manners, the vector mapping based machine learning model is a multilayer perceptron.

[0015] In some possible implementation manners, the training the vector mapping based machine learning model with the training data comprises:

[0016] training the multilayer perceptron by taking the temperature of the monitoring points in the monitoring data as an input of the multilayer perceptron and taking the temperature of the prediction points in the prediction data as an output of the multilayer perceptron.

[0017] In some possible implementation manners, the vector mapping based machine learning model is a conditional neural process model, and the conditional neural process model comprises a first fully connected neural network, a second fully connected neural network, and a feature processing module.

[0018] an output of the first fully connected neural network is connected to the feature processing module, an output of the feature processing module is connected to the second fully connected neural network, the first fully connected neural network is configured to learn features of the monitoring data, the feature processing module is configured to calculate a feature mean value of the output of the first fully connected neural network, and the second fully connected neural network is configured to predict the temperature of the prediction points according to the feature mean value output by the feature processing module.

[0019] In some possible implementation manners, the training of the vector mapping based machine learning model based on the training data comprises:

[0020] The position and temperature of the monitoring point in the monitoring data are taken as the input of the first full connection neural network, the output of the feature processing module and the position of the prediction point in the prediction data are taken as the input of the second full connection neural network, and the temperature of the prediction point in the prediction data is taken as the output of the second full connection neural network, so as to train the conditional neural process model.

[0021] In some possible implementation manners, the vector mapping based machine learning model is a self-attention mechanism model, and the self-attention mechanism model comprises a first full connection neural network, a second full connection neural network, a third full connection neural network, an encoder and a decoder.

[0022] The output of the first full connection neural network is connected with the input of the encoder, the output of the second full connection neural network and the output of the encoder are respectively connected with the input of the decoder, and the output of the decoder is connected with the input of the third full connection neural network, the first full connection neural network is used to learn the features of the monitoring data, the encoder is used to update the features output by the first full connection neural network according to the self-attention mechanism, the second full connection neural network is used to learn the features of the position of the monitoring point in the monitoring data and the features of the position of the prediction point in the prediction data, the decoder is used to generate the features of the prediction point according to the features output by the encoder and the features output by the second full connection neural network according to the self-attention mechanism, and the third full connection neural network is used to predict the temperature of the prediction point according to the features output by the decoder.

[0023] In some possible implementation manners, the training of the vector mapping based machine learning model based on the training data comprises:

[0024] The position and temperature of the monitoring point in the monitoring data are taken as the input of the first full connection neural network, the position of the monitoring point in the monitoring data and the position of the prediction point in the prediction data are taken as the input of the second full connection neural network, and the temperature of the prediction point in the prediction data is taken as the output of the third full connection neural network, so as to train the self-attention mechanism model.

[0025] In some possible implementation manners, the vector mapping based machine learning model is a graph convolutional neural network model.

[0026] In some possible implementation manners, the training of the vector mapping based machine learning model based on the training data comprises:

[0027] establish a graph model of the monitoring points and the predicted points according to positions of the monitoring points in the monitoring data and positions of the predicted points in the predicted data;

[0028] Based on the graph model, the temperature of the monitoring points in the monitoring data is taken as an input of the graph convolutional neural network model, the temperature of the predicted points in the predicted data is taken as an output of the graph convolutional neural network model, and the graph convolutional neural network model is trained.

[0029] The main advantages of the technical solution of the present application are as follows:

[0030] The real-time reconstruction method of the temperature field of the heat source system based on limited monitoring points of the present application trains the machine learning model based on vector mapping by using the monitoring data and the predicted data under different component layouts, and can realize fast prediction and real-time prediction of the temperature of the predicted points under different component layouts by using the trained machine learning model, so as to obtain the temperature of other positions under the component layout by using the temperature measured by limited monitoring points, reduce the number of temperature sensors needed to be installed in the component layout area, and reduce the space requirement for sensor installation. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0032] Figure 1 The flowchart of the temperature field reconstruction method of the heat source system based on limited monitoring points of an embodiment of the present application;

[0033] Figure 2 The network structure schematic diagram of the multi-layer perceptron of an embodiment of the present application;

[0034] Figure 3 The network structure schematic diagram of the conditional neural process model of an embodiment of the present application;

[0035] Figure 4 The structure schematic diagram of the encoder of an embodiment of the present application;

[0036] Figure 5 The structure schematic diagram of the encoder block of an embodiment of the present application;

[0037] Figure 6 The structure schematic diagram of the decoder of an embodiment of the present application;

[0038] Figure 7 The structure schematic diagram of the decoder block of an embodiment of the present application;

[0039] Figure 8 A schematic diagram of a monitoring point and a prediction point of a graph model of an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0041] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the drawings.

[0042] Referring to Figure 1 An embodiment of the present application provides a temperature field reconstruction method of a heat source system based on limited monitoring points, which comprises the following steps:

[0043] Step S1, obtaining a plurality of training data, wherein the training data comprises monitoring data and prediction data, the monitoring data comprises positions and temperatures of a plurality of monitoring points in a component layout, and the prediction data comprises positions and temperatures of at least one prediction point in the same component layout;

[0044] Step S2, constructing a machine learning model based on vector mapping;

[0045] Step S3, training the machine learning model based on vector mapping by using the training data to fit the mapping relationship between the temperatures of the monitoring points and the temperatures of the prediction points;

[0046] Step S4, predicting the temperature of the prediction point by using the trained machine learning model.

[0047] The temperature field real-time reconstruction method of the heat source system based on limited monitoring points provided by an embodiment of the present application can train the constructed machine learning model based on vector mapping by using the monitoring data and the prediction data under different component layouts, and can realize the rapid prediction and real-time prediction of the temperature of the prediction point under different component layouts by using the trained machine learning model, so as to obtain the temperature of other positions under the component layout by using the temperature measured by the limited monitoring points, reduce the number of temperature sensors needed to be installed in the component layout area, and reduce the space requirement for sensor installation.

[0048] Considering the heat conduction application scenario of a two-dimensional plane, a plurality of heat source components are distributed in a certain layout area, and the steady-state temperature field distribution in the layout area satisfies the following Poisson equation:

[0049]

[0050] wherein k represents the heat conduction coefficient of the layout area, (x, y) represents the position coordinates of any point of the layout area, φ(x, y) represents the heat source intensity distribution function, and u represents the temperature field of the layout area.

[0051] The heat source intensity distribution function φ(x, y) is determined by the position of the heat source, and is specifically represented as:

[0052]

[0053] wherein φ i represents the intensity of the i-th heat source, Γ i represents the layout area covered by the heat source.

[0054] In the case of determining the heat source intensity, the temperature field distribution is determined by the following boundary conditions:

[0055]

[0056] wherein u0 represents the temperature value at the isothermal boundary, n' represents the normal direction at the boundary, h represents the surface heat transfer coefficient between the object and the surrounding fluid at the boundary, u=u0 represents the first type of boundary condition (Dirichlet boundary condition), u-u0 represents the second type of boundary condition (Neumann boundary condition), and u-u0 represents the third type of boundary condition (Robin boundary condition).

[0057] The steps and principles of the heat source system temperature field reconstruction method provided by the embodiment of the present application based on a limited number of measuring points are specifically described below based on the layout of the heat source assembly in a two-dimensional plane:

[0058] Step S1, obtaining a plurality of training data.

[0059] In the embodiment of the present application, the plurality of training data can be obtained in the following manner:

[0060] The positions of the monitoring points and the prediction points in the component layout area of the heat source system are determined, and the positions of the components in the component layout area are determined;

[0061] For each component, a power value is randomly sampled from the preset power range corresponding to each component to determine the power of each component, a component layout is obtained, the temperature field corresponding to the component layout is simulated and calculated by using the finite element method, the temperature of the monitoring point and the temperature of the prediction point are extracted from the temperature field, and a monitoring data including the position and temperature of the monitoring point {(x i ,y i ,u i), i = 1, 2, …, n} and a prediction data {(x j ,y j ,u j ), j = n + 1, n + 2, …, n + m} including the position and temperature of the prediction point, to obtain training data including the monitoring data and the prediction data, and the random sampling and extraction process is repeated for multiple times until a preset number of training data is obtained. Wherein, (x i ,y i ) represents the position coordinates of the i-th monitoring point, u i represents the temperature of the i-th monitoring point, n represents the number of monitoring points, (x j ,y j ) represents the position coordinates of the j-n-th prediction point, u j represents the temperature of the j-n-th prediction point, and m represents the number of prediction points.

[0062] Wherein, the position coordinates of the monitoring points and the prediction points can be determined by pre-setting a two-dimensional plane coordinate system in the component layout area.

[0063] Through the above-mentioned manner, training data including monitoring data and prediction data under multiple different component layouts can be obtained. Wherein, two component layouts being the same means that the positions of the components in one component layout in the component layout area are the same as the positions of the components in another component layout in the component layout area, and the powers of the two components located at the same position in the component layout area are also the same; two component layouts being different means that the positions of the components in one component layout in the component layout area are the same as the positions of the components in another component layout in the component layout area, and the powers of at least one pair of components located at the same position in the component layout area are different.

[0064] Wherein, the specific number and position of the monitoring points can be determined according to the number and position of the temperature sensors installed in the component layout area of the heat source system, and the specific number and position of the prediction points can be determined according to the positions requiring real-time temperature monitoring in the component layout area of the heat source system.

[0065] In an embodiment of the present application, the specific number of training data can be set according to the training efficiency and accuracy requirement. Generally speaking, the more the training data is, the higher the prediction accuracy of the finally obtained model is, but the longer the training period required for model training is.

[0066] In an embodiment of the present application, the finite element method can be, for example, finite element method, finite difference method, finite volume method, etc.

[0067] Step S2, constructing a machine learning model based on vector mapping.

[0068] In an embodiment of the present application, the vector mapping-based machine learning model comprises any one of a multi-layer perceptron, a conditional neural process model, a self-attention mechanism model and a graph convolutional neural network model.

[0069] Referring to Figure 2 In an embodiment of the present application, the multi-layer perceptron comprises a plurality of hidden layers for learning a mapping relationship between the temperature of the monitoring point and the temperature of the prediction point.

[0070] Referring to Figure 3 In an embodiment of the present application, the conditional neural process model comprises a first fully connected neural network, a second fully connected neural network and a feature processing module, an output of the first fully connected neural network is connected to the feature processing module, an output of the feature processing module is connected to the second fully connected neural network, the first fully connected neural network is configured to learn features of the monitoring data, the feature processing module is configured to calculate a feature mean of the output of the first fully connected neural network, and the second fully connected neural network is configured to predict the temperature of the prediction point according to the feature mean output by the feature processing module.

[0071] In an embodiment of the present application, the self-attention mechanism model comprises a first fully connected neural network, a second fully connected neural network, a third fully connected neural network, an encoder and a decoder, an output of the first fully connected neural network is connected to an input of the encoder, an output of the second fully connected neural network and an output of the encoder are respectively connected to an input of the decoder, an output of the decoder is connected to an input of the third fully connected neural network, the first fully connected neural network is configured to learn features of the monitoring data, the encoder is configured to update the features output by the first fully connected neural network according to a self-attention mechanism, the second fully connected neural network is configured to learn features of position data of the monitoring point in the monitoring data and features of position data of the prediction point in the prediction data, the decoder is configured to generate features of the prediction point according to the features output by the encoder and the features output by the second fully connected neural network according to the self-attention mechanism, and the third fully connected neural network is configured to predict the temperature of the prediction point according to the features output by the decoder.

[0072] Referring to Figure 4 and Figure 5 In an embodiment of the present application, the encoder comprises a plurality of encoder blocks connected in sequence from an input end to an output end, the encoder block comprises a first multi-head self-attention mechanism, a first normalization layer, a feedforward neural network and a second normalization layer connected in sequence from an input end to an output end, the first multi-head self-attention mechanism comprises a plurality of parallel self-attention mechanism modules, an input of the first multi-head self-attention mechanism is also connected to an input of the feedforward neural network in a skip connection manner, and an output of the first normalization layer is also connected to an output of the second normalization layer in a skip connection manner.

[0073] Referring to Figure 6 and Figure 7In an embodiment of the present application, the decoder comprises a plurality of decoder blocks connected in sequence from an input end to an output end, and each decoder block comprises a second multi-head self-attention mechanism, a third normalization layer, a feedforward neural network and a fourth normalization layer connected in sequence from an input end to an output end, the second multi-head self-attention mechanism comprises a plurality of parallel self-attention mechanism modules, and the input of the second multi-head self-attention mechanism is also connected to the input of the feedforward neural network in a skip connection manner, and the output of the third normalization layer is also connected to the output of the fourth normalization layer in a skip connection manner.

[0074] In an embodiment of the present application, the graph convolutional neural network model comprises a plurality of graph convolutional layers.

[0075] In step S3, the machine learning model based on vector mapping is trained by using the training data to fit the mapping relationship between the temperature of the monitoring point and the temperature of the prediction point.

[0076] From the limited temperature monitoring points to the temperature distribution can be regarded as a function fitting problem based on partial monitoring values, a two-dimensional temperature distribution function can be expressed as u=f(x,y). Based on the existing monitoring point data, a supervised learning method can be used to train and fit an approximate function f:(x,y)→u. In fact, due to the similarity between different component layouts, part of the information between the approximate functions corresponding to different component layouts can be shared. By using the approximate function information under different component layouts and the monitoring point information under the current component layout, the approximate function under the current component layout can be inferred to obtain the relationship between the monitoring point and the prediction point under the current component layout, and the prediction point temperature under the current component layout is predicted.

[0077] In an embodiment of the present application, when the machine learning model based on vector mapping is a multi-layer perceptron, training the machine learning model based on vector mapping by using the training data comprises:

[0078] The temperature of the monitoring point in the monitoring data is used as the input of the multi-layer perceptron, and the temperature of the prediction point in the prediction data is used as the output of the multi-layer perceptron, and the multi-layer perceptron is trained.

[0079] When the machine learning model based on vector mapping is a multi-layer perceptron, the loss function can be set as:

[0080]

[0081] Wherein, represents the temperature value of the j-nth prediction point in the prediction data of the bth training data, represents the temperature prediction value of the j-nth prediction point output by the multi-layer perceptron in the prediction data of the bth training data, B represents the number of training data, and m represents the number of prediction points.

[0082] Further, in an embodiment of the present application, when the vector mapping based machine learning model is a conditional neural process model, training the vector mapping based machine learning model using the training data comprises:

[0083] training the conditional neural process model by taking the positions and temperatures of the monitoring points in the monitoring data as inputs of a first fully connected neural network, taking the output of the feature processing module and the positions of the prediction points in the prediction data as inputs of a second fully connected neural network, and taking the temperatures of the prediction points in the prediction data as outputs of the second fully connected neural network.

[0084] Suppose P represents a function distribution f: (x, y) → u, the monitoring data is denoted as O = {(x i ,y i ,u i )|i = 1, 2,..., n}, and the position data of the prediction data is denoted as T = {(x j ,y j )|j = n + 1, n + 2,..., n + m}, then according to the monitoring data O, the prediction f(x j ,y j ), (x j ,y j ) ∈ T can be regarded as maximizing the conditional distribution probability P(f(T) | O, T).

[0085] In an embodiment of the present application, the conditional neural process model is used to maximize the conditional distribution probability P(f(T) | O, T). When a neural network is used to learn the features r of the monitoring data, it is assumed that the features r contain all the information of the temperature distribution function, and the change of the data order on the monitoring data O has no effect on the features r, i.e. for different order arrangements O' of the same monitoring data, there exists Q θ (O) = Q θ (O'), and Q θ represents the neural network.

[0086] Therefore, in an embodiment of the present application, a fully connected neural network, i.e. a first fully connected neural network, is used to learn the features of the monitoring data, which can be specifically expressed as:

[0087]

[0088] wherein r i represents the features of the i-th monitoring point data in the monitoring data, X represents a horizontal coordinate space, Y represents a vertical coordinate space, and U represents a temperature space, represents a d-dimensional feature space, and the first fully connected neural network is used to implement.

[0089] a feature r corresponding to all monitoring point data output by the first fully connected neural network i The feature processing module averages all features to obtain a feature r of the monitoring data, and the specific expression is as follows:

[0090]

[0091] For (x j , y j ) element of T, another fully connected neural network, namely a second fully connected neural network, is used to make a prediction in combination with the feature r output by the feature processing module, and the specific expression is as follows:

[0092]

[0093] Wherein, u′ j represents a temperature prediction value corresponding to the j-nth prediction point, represents a one-dimensional temperature space.

[0094] In an embodiment of the present application, when the conditional neural process model is trained, the loss function can be set as:

[0095]

[0096]

[0097] Wherein, represents a temperature value of the j-nth prediction point of the prediction data of the bth training data, represents a temperature prediction value output by the conditional neural process model at the j-nth prediction point in the prediction data of the bth training data, represents the position coordinates of the j-nth prediction point in the prediction data of the bth training data, B represents the number of training data, and m represents the number of prediction points, represents the position coordinates of the i-th monitoring point in the monitoring data of the bth training data, represents the temperature of the i-th monitoring point in the monitoring data of the bth training data, and n represents the number of monitoring points.

[0098] In an embodiment of the present application, one training data corresponds to one specific component layout, and when the conditional neural process model is trained, the positions and temperatures of the n monitoring points in the monitoring data of the training data are taken as the input of the first fully connected neural network f θ , and the feature r of the monitoring data is obtained after the output feature of the first fully connected neural network f θ is processed by the feature processing model, the feature r contains the information of the temperature distribution function under the current component layout, and can be combined with the second fully connected neural network gθ Composition temperature distribution function g θ (·,·,r), the positions of the m prediction points in the prediction data of the training data are taken as the input of the second fully connected neural network g θ , and the temperatures of the m prediction points in the prediction data of the training data are taken as the output labels of the second fully connected neural network g θ , the loss function value is calculated, and the model parameters are updated by back propagation.

[0099] Further, in an embodiment of the present application, the conditional neural process model can also be integrated with a Gaussian process. Since any finite random variable in the Gaussian process assumption set obeys a joint normal distribution, in an embodiment of the present application, it is assumed that the temperature value output by the conditional neural process model at each prediction point obeys a Gaussian distribution, which can be specifically expressed as:

[0100] p(u j |x j ,y j ,O)=N(u j ;μ j ,σ j )

[0101] μ j ,σ j =g θ (x j ,y j ,O)

[0102] Where p(u j |x j ,y j ,O) represents the conditional probability distribution of the temperature value u j at (x j ,y j ) based on the observation data O; N(u j ; μ j ,σ j ) represents that u j obeys a normal distribution with an expectation μ j and a variance σ j .

[0103] In an embodiment of the present application, when integrating the conditional neural process model with the Gaussian process, during training of the model, based on the given training data, the model parameters can be updated by maximizing the log-likelihood, which can be specifically expressed as:

[0104]

[0105] Where θ * represents the optimal model parameters obtained by training, θ represents the model parameters, Θ represents the parameter set, and E represents the average expectation.

[0106] In one embodiment of the present invention, when the conditional neural process model is integrated with a Gaussian process, the temperature value u of the predicted point can also be obtained simultaneously using the conditional neural process model. j and the measurement of prediction uncertainty σ j .

[0107] Furthermore, in one embodiment of the present invention, when the machine learning model based on vector mapping is a self-attention mechanism model, training the machine learning model based on vector mapping using training data includes:

[0108] The location and temperature of the monitoring points in the monitoring data are used as inputs to the first fully connected neural network. The location of the monitoring points in the monitoring data and the location of the predicted points in the prediction data are used as inputs to the second fully connected neural network. The temperature of the predicted points in the prediction data is used as the output of the third fully connected neural network to train the self-attention mechanism model.

[0109] Specifically, based on the self-attention mechanism model set above, the positions and temperatures of n monitoring points in the monitoring data {(x i ,y i ,u i The input of each monitoring point (i = 1, 2, ..., n) is fed into a first fully connected neural network. Based on the information from each monitoring point, the first fully connected neural network outputs a feature r of length d for each monitoring point. i The encoder receives n features of length d from the output of the first fully connected neural network. In each encoder block, the first multi-head self-attention mechanism receives features {r} from n monitoring points. i The input |i=1,2,...,n} is processed through multiple self-attention mechanism modules to generate new features for each monitoring point. Within each self-attention mechanism module, each feature r... i Each feature {q} is transformed into three features of length l through three pre-connected neural networks. i ,k i ,v i}, eigenvectors {q i |i=1,2,...,n}、{k i |i=1,2,...,n} and {v i |i=1,2,...,n} are denoted as the query matrix Q, the key matrix K, and the value matrix V, respectively, for subsequent processing.

[0110] In one embodiment of the present invention, the self-attention mechanism is calculated using the following formula:

[0111]

[0112] Where Attention(·) represents the self-attention mechanism module, softmax(·) represents the softmax function, and d k Represents vector k in the key matrix i The length.

[0113] Since the first multi-head self-attention mechanism includes multiple parallel self-attention mechanism modules, after passing through multiple self-attention mechanism modules, it outputs h matrices {H}. i |i=1,2,...,h},H i Let {H} represent the matrix output by the i-th self-attention mechanism module, and h represent the number of parallel self-attention mechanism modules. Then, we can group the h matrices {H}... i After merging |i=1,2,...,h}, the output H of the first multi-head self-attention mechanism can be obtained by transforming the matrix size through matrix operations. The specific calculation process is as follows:

[0114] H = Concat(H1,...,H) h W o

[0115] Where Concat(·) represents the Concat function, W o This indicates a linear transformation used to adjust the length of the concatenated vector.

[0116] Based on the above calculation process, the output H of the first multi-head self-attention mechanism contains n feature vectors {h} of length d. i |i=1,2,...,n}, each feature vector is sequentially passed through the first normalization layer, the feedforward neural network, and the second normalization layer to obtain a new feature r′ for each monitoring point. i The new feature r′ i The final feature r″ can be obtained by stacking multiple encoder blocks. i ,r″ i This refers to the feature corresponding to the i-th monitoring point output by the encoder.

[0117] Furthermore, based on the self-attention mechanism model established above, the positions of n monitoring points in the monitoring data {(x i ,y i The positions of m predicted points in the predicted data {(x) | i = 1, 2, ..., n} and i = 1, 2, ..., n}. j ,y j The input of each monitoring point (j = n+1, n+2, ..., n+m) is processed by a second fully connected neural network. Based on the location information of each monitoring point, the second fully connected neural network outputs a feature r of length d for each monitoring point. 1,i Based on the location information of each prediction point, a feature r of length d is output for each prediction point.2,j The decoder receives the feature r 1,i and the feature r 2,j output by the encoder i The feature r 1,i , the feature r 2,j and the feature r i are respectively taken as a query matrix Q, a key matrix K and a value matrix V of a multi-head self-attention mechanism, and are processed based on the self-attention mechanism to generate a corresponding feature r 2,j for each prediction point. The decoder processes data based on the query matrix Q, the key matrix K and the value matrix V in the same way as the encoder, and the specific manner can refer to the processing procedure of the encoder.

[0118] Further, based on the self-attention mechanism model set above, the feature r 2,j output by the decoder for each prediction point is taken as an input of a third fully connected neural network. 2,j The third fully connected neural network outputs a temperature u j corresponding to the prediction point according to the input feature r 2,j .

[0119] When the vector mapping-based machine learning model is a self-attention mechanism model, the loss function can be set as:

[0120]

[0121] wherein, represents a temperature value of the j-nth prediction point in the prediction data of the bth training data, represents a temperature prediction value of the j-nth prediction point output by the self-attention mechanism model in the prediction data of the bth training data, B represents the number of training data, and m represents the number of prediction points.

[0122] Further, in an embodiment of the present application, when the vector mapping-based machine learning model is a graph convolutional neural network model, training the vector mapping-based machine learning model using the training data comprises:

[0123] establishing a graph model of the monitoring points and the prediction points according to the positions of the monitoring points in the monitoring data and the positions of the prediction points in the prediction data;

[0124] based on the graph model, taking the temperatures of the monitoring points in the monitoring data as inputs of the graph convolutional neural network model, taking the temperatures of the prediction points in the prediction data as outputs of the graph convolutional neural network model, and training the graph convolutional neural network model.

[0125] A graph is a data structure similar to an image, but unlike an image, a graph is irregular and can store and process non-regularized and non-regionalized data, and in an embodiment of the present application, a corresponding graph model is established according to position information of monitoring points in monitoring data and position information of prediction points in prediction data.

[0126] The key to defining a graph is the establishment of nodes and edges in the graph, and in an embodiment of the present application, a graph model of monitoring points and prediction points is established according to positions of monitoring points in monitoring data and positions of prediction points in prediction data, which can include:

[0127] Each monitoring point is taken as a node, each prediction point is taken as a node, and nodes of the graph model are determined;

[0128] According to positions of all monitoring points and positions of all prediction points, distances between any two nodes in all nodes are calculated respectively, and if the distance is less than a threshold value, an undirected edge is established between the corresponding two nodes, all undirected edges of the graph model are determined, and the graph model of the monitoring points and the prediction points is established.

[0129] The threshold value can be set according to actual conditions.

[0130] Referring to Figure 8 , Figure 8 A graph model provided by an embodiment of the present application is shown in the accompanying drawings, in which a square frame represents a boundary of a component layout area of a heat source system, an internal rectangle represents a component, points in the area of the component are nodes corresponding to monitoring points, and points outside the area of the component are nodes corresponding to prediction points.

[0131] In an embodiment of the present application, the graph convolutional neural network model includes multiple graph convolutional layers, and the graph convolutional operation of the graph convolutional neural network is a layer similar to the convolutional operation in image processing, which can extract neighborhood information of nodes and pass it to the next layer through learnable parameters. The graph convolutional operation of the graph convolutional layer includes an aggregation operation and a transmission operation, and can be specifically represented as:

[0132]

[0133] wherein H (l+1) represents a feature of the l+1th layer of the graph convolutional neural network, sigma (·) represents a nonlinear activation function, represents an adjacency matrix with self-loop, A represents an adjacency matrix, represents a degree matrix with self-loop, D represents a degree matrix, I N represents a unit matrix, H (l) represents a feature of the lth layer of the graph convolutional neural network, H (0) represents an input of the graph convolutional neural network, H (K)represents the output of the graph convolutional neural network, K represents the number of layers of the graph convolutional neural network, and W represents the weight of the graph convolutional neural network. (l) represents the parameter of the lth layer of the graph convolutional neural network.

[0134] wherein the adjacency matrix is used to represent the mutual connection relationship between nodes, if there is an undirected edge between node i and node j, then the corresponding adjacency matrix element A ij = 1, otherwise A ij = 0; the degree matrix is used to represent the number of edges existing for each node, which is a diagonal matrix, N represents the number of columns of the adjacency matrix.

[0135] Further, considering that the input of the entire graph convolutional neural network model is only monitoring data, which cannot cover the features of the entire graph. Therefore, in an embodiment of the present application, a one-dimensional convolutional layer is added in front of the entire network model, so as to convert the node features corresponding to the input monitoring points into features on the entire graph.

[0136] Further, in an embodiment of the present application, in order to improve the accuracy and avoid the problem that the graph convolutional neural network model falls into over-smoothing, the strategy of batch normalization (BatchNorm) and the form of residual connection (Residual connection) are added in the graph convolutional neural network.

[0137] Specifically, after adding the strategy of batch normalization (BatchNorm) and the form of residual connection (Residual connection), the graph convolution operation can be represented as:

[0138]

[0139] wherein BatchNorm(·) represents batch normalization.

[0140] In an embodiment of the present application, the input of the graph convolutional neural network model is the temperature of the monitoring point in the monitoring data, and the output is the temperature prediction value of the prediction point in the prediction data. When training the graph convolutional neural network model, the mean absolute error (Mean Absolute Error, MAE) of the temperature of the prediction point and the temperature prediction value can be used as a loss function, and the model parameters are updated by the gradient descent method.

[0141] Step S4, temperature prediction of the prediction point is performed by using the trained machine learning model.

[0142] Specifically, after the training of the machine learning model is completed, the positions and temperatures of the monitoring points in different component layout situations can be directly input into the machine learning model, and the machine learning model outputs the predicted temperature values of the prediction points in the current component layout situation according to the input positions and temperatures of the monitoring points, so as to realize real-time prediction and monitoring of the temperatures of the prediction points.

[0143] It should be noted that, in this document, relational terms such as "first" and "second", and the like, are used solely to distinguish one entity or action from another entity or action, without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Furthermore, "front", "rear", "left", "right", "upper", "lower", and the like, as used in this document, refer to the positions shown in the drawings and are for reference only.

[0144] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for real-time reconstruction of temperature field of a heat source system based on limited measurement points, characterized in that, The method comprises: obtaining a plurality of training data, wherein the training data comprises monitoring data and prediction data, the monitoring data comprises positions and temperatures of a plurality of monitoring points in a component layout, and the prediction data comprises positions and temperatures of at least one prediction point in the same component layout; constructing a vector mapping-based machine learning model; training the vector mapping-based machine learning model by using the training data to fit a mapping relationship between the temperatures of the monitoring points and the temperatures of the prediction points; performing temperature prediction of the prediction points by using the trained machine learning model; the vector mapping-based machine learning model is a conditional neural process model, which comprises a first fully connected neural network, a second fully connected neural network and a feature processing module; an output of the first fully connected neural network is connected to the feature processing module, an output of the feature processing module is connected to the second fully connected neural network, the first fully connected neural network is used to learn features of the monitoring data, the feature processing module is used to calculate a feature mean value of the output of the first fully connected neural network, and the second fully connected neural network is used to predict the temperature of the prediction point according to the feature mean value output by the feature processing module; or the vector mapping-based machine learning model is a self-attention mechanism model, which comprises a first fully connected neural network, a second fully connected neural network, a third fully connected neural network, an encoder and a decoder; an output of the first fully connected neural network is connected to an input of the encoder, an output of the second fully connected neural network and an output of the encoder are respectively connected to inputs of the decoder, an output of the decoder is connected to an input of the third fully connected neural network, the first fully connected neural network is used to learn features of the monitoring data, the encoder is used to update the features output by the first fully connected neural network according to a self-attention mechanism, the second fully connected neural network is used to learn features of positions of the monitoring points in the monitoring data and features of positions of the prediction points in the prediction data, the decoder is used to generate features of the prediction points according to the features output by the encoder and the features output by the second fully connected neural network according to a self-attention mechanism, and the third fully connected neural network is used to predict the temperature of the prediction point according to the features output by the decoder.

2. The method of claim 1, wherein, The method comprises: determining positions of the monitoring points and the prediction points in a component layout area of a heat source system, and determining positions of various components in the component layout area. For each component, a power value is randomly sampled from a preset power range corresponding to each component to determine the power of each component, a component layout is obtained, a temperature field corresponding to the component layout is simulated and calculated by using a finite element method, the temperature of a monitoring point and the temperature of a prediction point are extracted from the temperature field, a monitoring data including the position and temperature of the monitoring point and a prediction data including the position and temperature of the prediction point are determined, a training data including the monitoring data and the prediction data is obtained, and the random sampling and extraction process is repeated for multiple times until a preset number of training data is obtained.

3. The method of claim 1, wherein, When the vector mapping-based machine learning model is a conditional neural process model, the training of the vector mapping-based machine learning model by using the training data comprises: taking the position and temperature of the monitoring point in the monitoring data as the input of the first full connection neural network, taking the output of the feature processing module and the position of the prediction point in the prediction data as the input of the second full connection neural network, and taking the temperature of the prediction point in the prediction data as the output of the second full connection neural network, to train the conditional neural process model; When the vector mapping-based machine learning model is a self-attention mechanism model, the training of the vector mapping-based machine learning model by using the training data comprises: taking the position and temperature of the monitoring point in the monitoring data as the input of the first full connection neural network, taking the position of the monitoring point in the monitoring data and the position of the prediction point in the prediction data as the input of the second full connection neural network, and taking the temperature of the prediction point in the prediction data as the output of the third full connection neural network, to train the self-attention mechanism model.

Images