Early warning method and system for load deformation of steel structure roof

Abstract

The invention relates to the technical field of building monitoring, in particular to an early warning method and system for load deformation of a steel structure roof, and the method comprises the following steps: determining a spatial point location for monitoring the load deformation degree of the roof on the steel structure roof; monitoring feature data for quantifying the deformation degree of the roof load on the spatial point location; performing convolutional learning on the relevance between the feature data and the roof load deformation degree by using a teacher-student model, and constructing a deformation lightweight model; and according to the deformation lightweight model, deformation point locations are screened out from the space point locations on the steel structure roof, and deformation early warning is carried out. According to the method, a high-performance and high-generalization teacher model for predicting the load deformation degree is trained by using a large-scale network, and the method is applied to guidance and training of a student model for predicting the load deformation degree in a lightweight manner, so that the student model has prediction performance equivalent to that of the teacher model, model acceleration is realized, and the requirement of roof deformation monitoring for timeliness is met.

Description

Technical Field

[0001] The present invention relates to the technical field of building monitoring, and in particular to an early warning method and system for load deformation of a steel structure roof. Background Art

[0002] Load deformation of steel structure roofs usually occurs in stress concentration areas where stress mutations are prone to occur. Therefore, monitoring stress concentration areas can determine whether the roof has load deformation and provide timely warnings when the roof is deformed to avoid accidents.

[0003] In the prior art, deep learning intelligent algorithms are usually used to monitor and classify the load deformation conditions of roofs. In order to ensure the model performance of load deformation monitoring, existing intelligent algorithms often use large-scale neural network structures. However, the slow computing speed of large-scale networks will lead to the failure to meet the timeliness of monitoring, that is, the phenomenon of untimely warning will occur. Summary of the invention

[0004] The purpose of the present invention is to provide an early warning method for load deformation of steel structure roofs, so as to solve the technical problem that in the prior art, intelligent algorithms often use large-scale neural network structures to ensure the model performance of load deformation monitoring, but the slow computing speed of large-scale networks leads to the failure to meet the monitoring timeliness requirements.

[0005] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

[0006] An early warning method for load deformation of a steel structure roof comprises the following steps:

[0007] On the steel structure roof, determine the spatial points for monitoring the load deformation of the roof;

[0008] At the spatial point, characteristic data for quantifying the degree of deformation of the roof load is monitored by monitoring sensors;

[0009] The teacher-student model is used to perform convolutional learning on the correlation between the characteristic data and the degree of deformation of the roof load, and a deformation quantification lightweight model for quantifying the degree of deformation of the roof load is constructed;

[0010] According to the roof load deformation degree quantified by the deformation quantification lightweight model, the deformation points are screened out from various spatial points on the steel structure roof, and deformation warning is carried out.

[0011] As a preferred solution of the present invention, the method for determining the spatial point position includes:

[0012] Finite element modeling and analysis are used to divide the steel structure roof into dense and sparse grids according to stress concentration;

[0013] Set the grid density threshold, mark the grids on the steel structure roof whose size is greater than or equal to the grid density threshold as sparse grids, and mark the grids on the steel structure roof whose size is less than the grid density threshold as fine grids, where the fine grids are stress concentration areas on the steel structure roof, and the sparse grids are stress non-concentration areas on the steel structure roof;

[0014] The center of the fine grid on the steel structure roof is marked as the spatial point.

[0015] As a preferred solution of the present invention, the characteristic data monitoring method includes:

[0016] Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points;

[0017] Use X-ray instrument to measure stress data at spatial points;

[0018] The image data of the steel structure member, the image data of the abnormal load and the stress data are used as the characteristic data.

[0019] As a preferred solution of the present invention, the method for constructing the deformation quantized lightweight model includes:

[0020] Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data;

[0021] At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data;

[0022] The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data;

[0023] The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data;

[0024] The teacher network and the student network are jointly trained by using sample input data and sample output data, wherein the teacher network provides training guidance to the student network;

[0025] The trained student network is used as the deformed lightweight model;

[0026] The teacher network is: D_t=BP_t(MSC(S));

[0027] The student network is: D_s=BP_s(S);

[0028] The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s);

[0029] Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

[0030] As a preferred solution of the present invention, the weight coefficient is associated with the receptive field of the multi-scale convolutional network;

[0031] Wherein, the weight coefficient is: k = 1 + [1-Var ({R f |f∈[1,n]})]log(Var({R f |f∈[1,n]}));

[0032] Where k is the weight coefficient of CE(D_t,D_s), R f is the receptive field of the fth scale of the multi-scale convolutional network, n is the total number of receptive field scales in the multi-scale convolutional network, {R f |f∈[1,n]} is a data sequence consisting of n-scale receptive fields in a multi-scale convolutional network, Var is a variance operator, Var({R f |f∈[1,n]}) is {R f |The variance calculation value of f∈[1,n]};

[0033] R f =2(r f -1)*(m-1)+m;

[0034] In the formula, r f is the number of holes at the fth scale of the multi-scale convolutional network, and m is the convolution kernel size.

[0035] As a preferred embodiment of the present invention, the method for screening deformation points includes:

[0036] The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point;

[0037] A deformation threshold is set, and the predicted value of the load deformation degree at each spatial point is compared with the deformation threshold, wherein:

[0038] When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point;

[0039] When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

[0040] As a preferred solution of the present invention, the present invention provides an early warning system for load deformation of a steel structure roof. According to the early warning method for load deformation of a steel structure roof, the system includes:

[0041] Monitoring sensors, including visual sensors and X-ray devices, are used to monitor characteristic data used to quantify the degree of deformation of the roof load at the spatial point;

[0042] The deep learning unit uses a teacher-student model to perform convolutional learning on the correlation between the characteristic data and the degree of deformation of the roof load, and constructs a deformation quantification lightweight model for quantifying the degree of deformation of the roof load;

[0043] The deformation warning unit selects deformation points from various spatial points on the steel structure roof according to the roof load deformation degree quantified by the deformation quantification lightweight model, and performs deformation warning.

[0044] As a preferred solution of the present invention, the method for constructing a deformation quantization lightweight model by a deep learning unit includes:

[0045] Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data;

[0046] At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data;

[0047] The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data;

[0048] The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data;

[0049] The teacher network and the student network are jointly trained by using sample input data and sample output data, wherein the teacher network provides training guidance to the student network;

[0050] The trained student network is used as the deformed lightweight model;

[0051] The teacher network is: D_t=BP_t(MSC(S));

[0052] The student network is: D_s=BP_s(S);

[0053] The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s);

[0054] Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

[0055] As a preferred solution of the present invention, the method for monitoring characteristic data by the monitoring sensor includes:

[0056] Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points;

[0057] Use X-ray instrument to measure stress data at spatial points;

[0058] The image data of the steel structure member, the image data of the abnormal load and the stress data are used as the characteristic data.

[0059] As a preferred solution of the present invention, the method for screening deformation points by the deformation warning unit includes:

[0060] The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point;

[0061] A deformation threshold is set, and the predicted value of the load deformation degree at each spatial point is compared with the deformation threshold, wherein:

[0062] When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point;

[0063] When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] The present invention utilizes a teacher-student model and a large-scale network to train a high-performance and highly generalized teacher model for predicting the degree of load deformation, which is applied to guide the training of a lightweight student model for predicting the degree of load deformation, so that the student model has the same prediction performance as the teacher model, and the parameter processing process and number are greatly reduced, thereby achieving model compression and acceleration, and meeting the timeliness requirements of roof deformation monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0067] Figure 1 A flow chart of a method for early warning of load deformation of a steel structure roof provided by an embodiment of the present invention;

[0068] Figure 2 A block diagram of an early warning system for load deformation of a steel structure roof provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] like Figure 1 As shown, the present invention provides an early warning method for load deformation of a steel structure roof, comprising the following steps:

[0071] On the steel structure roof, determine the spatial points for monitoring the load deformation of the roof;

[0072] At the spatial point, the characteristic data used to quantify the degree of roof load deformation is monitored by monitoring sensors;

[0073] The teacher-student model is used to conduct convolutional learning on the correlation between feature data and roof load deformation, and a deformation quantification lightweight model for quantifying roof load deformation is constructed.

[0074] According to the roof load deformation degree quantified by the deformation quantification lightweight model, the deformation points are screened out from various spatial points on the steel structure roof, and deformation warning is carried out.

[0075] In order to establish a high-performance and highly generalized lightweight roof deformation monitoring model, the present invention utilizes a teacher-student model, wherein a high-performance and highly generalized load deformation degree prediction network trained based on a large-scale network is used as a teacher network to guide a lightweight load deformation degree prediction network (i.e., a student network) trained based on a small-scale network, so that the student network has a prediction performance (high performance and high generalization) comparable to that of the teacher network, and the parameter processing process and quantity are greatly reduced, thereby achieving model compression and acceleration, and meeting the timeliness requirements of roof deformation monitoring.

[0076] Specifically, the large-scale network of the teacher network in the present invention is composed of a multi-scale convolutional network and a BP neural network. The multi-scale convolutional network is used to form a multi-scale receptive field, and more scale information is extracted from the feature data used to quantify the degree of deformation of the roof load. By extracting features at different scale levels, a wider range of contextual information can be captured. Deeper features usually have a better grasp of global semantic information, while shallower features can provide more local detail information. Through multi-scale convolution, the algorithm can use global and local information at different levels at the same time, so as to better understand the feature content and improve the prediction accuracy. Moreover, in the deep convolutional neural network, as the number of network layers increases, the size of the feature map gradually decreases, which may lead to information loss, especially for smaller data details. By fusing features of multiple scales, the impact of information loss can be reduced and the network capacity can be improved. Therefore, the multi-scale convolutional network can provide richer data information and improve the performance of the model in quantifying the degree of deformation of the roof load.

[0077] After mining the feature data for richer information through a multi-scale convolutional network, the correlation between the feature data and the roof load deformation degree is mapped and learned through a BP neural network, which enables the teacher network to quantify the roof load deformation degree based on the feature data and complete the load deformation monitoring.

[0078] The present invention only sets the network structure of the student network with a BP neural network. Compared with the teacher network, there is no need to perform multi-scale convolution. The network structure is lightweight, and a large amount of data processing process and a large amount of parameter data generated during the convolution process are avoided, thereby achieving lightweight prediction of roof load deformation degree.

[0079] Although the student network has achieved lightweight, it still has the predictive performance of the teacher network due to the guidance of the teacher network during training. From the loss function, we can see that during training, CE(D_r,D_s) in the loss function ensures the accuracy between the student network prediction value and the true value, and CE(D_t,D_s) ensures the accuracy between the student network prediction value and the teacher network prediction value, reflecting the guidance of the teacher model. Therefore, the student network output can ensure the predictive output performance of the teacher network, that is, high precision and high generalization performance, and is also lightweight.

[0080] The present invention adds CE(D_r, D_s) in the loss function to ensure the accuracy between the predicted value of the student network and the true value, so that the output of the student network can ensure the predicted output performance of the teacher network and will not deviate from the authenticity prediction, that is, it limits the guidance direction of the teacher network to the student model and avoids the phenomenon of chaotic guidance. That is, although the student model has the predicted output performance of the teacher network, it deviates from the real result and belongs to invalid training. Therefore, adding CE(D_r, D_s) ensures that the student network has practical training significance.

[0081] The present invention adds a weight coefficient to CE(D_t, D_s) and associates the weight coefficient with the receptive field of the multi-scale convolutional network, so that the richer the receptive field of the multi-scale convolutional network, the higher the model performance of the teacher network and the higher the learning difficulty. Therefore, the higher the CE(D_t, D_s) weight coefficient is set, the higher the learning priority of the student network is, and more learning resources are occupied to ensure that the student network learns the teacher network to the greatest extent, or to increase the degree of guidance of the teacher network to the student network, so that the student network can have the performance of the teacher network to the greatest extent.

[0082] Therefore, the present invention realizes the adaptive learning of the student network to the teacher network, or the adaptive guidance of the teacher network to the student network according to its own structure, which is more robust and further ensures the prediction performance of the student network.

[0083] The present invention determines the stress concentration area on the steel structure roof where stress mutation is prone to occur through finite element analysis. Since the load deformation of the steel structure roof usually occurs in the stress concentration area where stress mutation is prone to occur, the stress concentration area can be used as a monitoring position (i.e., the spatial point position in the present invention) to monitor the degree of load deformation of the roof.

[0084] The methods for determining the spatial point include:

[0085] Finite element modeling and analysis are used to divide the steel structure roof into dense and sparse grids according to stress concentration;

[0086] Set the grid density threshold, mark the grids on the steel structure roof whose size is greater than or equal to the grid density threshold as sparse grids, and mark the grids on the steel structure roof whose size is less than the grid density threshold as fine grids, where the fine grids are stress concentration areas on the steel structure roof, and the sparse grids are stress non-concentration areas on the steel structure roof;

[0087] Mark the center of the fine grid on the steel structure roof as the spatial point.

[0088] The monitoring methods of characteristic data include:

[0089] Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points;

[0090] Use X-ray instrument to measure stress data at spatial points;

[0091] The image data of the steel structure member, the image data of the abnormal load and the stress data are used as feature data.

[0092] In order to establish a high-performance, highly generalized lightweight roof deformation monitoring model, the present invention uses a teacher-student model, wherein a high-performance and highly generalized load deformation degree prediction network trained based on a large-scale network is used as a teacher network to guide a lightweight load deformation degree prediction network (i.e., a student network) trained based on a small-scale network, so that the student network has the same prediction performance (high performance, high generalization) as the teacher network, and the parameter processing process and number are greatly reduced, thereby achieving model compression and acceleration, and meeting the timeliness requirements of roof deformation monitoring, as follows:

[0093] The method for constructing a deformation quantized lightweight model includes:

[0094] Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data;

[0095] At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data;

[0096] The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data;

[0097] The large-scale network of the teacher network in the present invention is composed of a multi-scale convolutional network and a BP neural network. The multi-scale convolutional network is used to form a multi-scale receptive field, and more scale information is extracted from the feature data used to quantify the degree of deformation of the roof load. By extracting features at different scale levels, a wider range of contextual information can be captured. Deeper features usually have a better grasp of global semantic information, while shallower features can provide more local detail information. Through multi-scale convolution, the algorithm can use global and local information at different levels at the same time, so as to better understand the feature content and improve the prediction accuracy. Moreover, in the deep convolutional neural network, as the number of network layers increases, the size of the feature map gradually decreases, which may lead to information loss, especially for smaller data details. By fusing features of multiple scales, the impact of information loss can be reduced and the network capacity can be improved. Therefore, the multi-scale convolutional network can provide richer data information and improve the performance of the model in quantifying the degree of deformation of the roof load.

[0098] After mining the feature data for richer information through a multi-scale convolutional network, the correlation between the feature data and the roof load deformation degree is mapped and learned through a BP neural network, which enables the teacher network to quantify the roof load deformation degree based on the feature data and complete the load deformation monitoring.

[0099] The present invention only sets the network structure of the student network with a BP neural network. Compared with the teacher network, there is no need to perform multi-scale convolution. The network structure is lightweight, and a large amount of data processing process and a large amount of parameter data generated during the convolution process are avoided, thereby achieving lightweight prediction of roof load deformation degree.

[0100] The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data;

[0101] The teacher network and the student network are jointly trained through sample input data and sample output data, wherein the teacher network provides training guidance for the student network;

[0102] The trained student network is used as a deformable lightweight model;

[0103] The teacher network is: D_t = BP_t (MSC (S));

[0104] The student network is: D_s = BP_s(S);

[0105] The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s);

[0106] Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

[0107] Although the student network has achieved lightweight, it still has the predictive performance of the teacher network due to the guidance of the teacher network during training. From the loss function, we can see that during training, CE(D_r,D_s) in the loss function ensures the accuracy between the student network prediction value and the true value, and CE(D_t,D_s) ensures the accuracy between the student network prediction value and the teacher network prediction value, reflecting the guidance of the teacher model. Therefore, the student network output can ensure the predictive output performance of the teacher network, that is, high precision and high generalization performance, and is also lightweight.

[0108] The present invention adds CE(D_r, D_s) in the loss function to ensure the accuracy between the predicted value of the student network and the true value, so that the output of the student network can ensure the predicted output performance of the teacher network and will not deviate from the authenticity prediction, that is, it limits the guidance direction of the teacher network to the student model and avoids the phenomenon of chaotic guidance. That is, although the student model has the predicted output performance of the teacher network, it deviates from the real result and belongs to invalid training. Therefore, adding CE(D_r, D_s) ensures that the student network has practical training significance.

[0109] The weight coefficients are associated with the receptive field of the multi-scale convolutional network;

[0110] Among them, the weight coefficient is: k = 1 + [1-Var ({R f |f∈[1,n]})]log(Var({R f |f∈[1,n]}));

[0111] Where k is the weight coefficient of CE(D_t,D_s), R f is the receptive field of the fth scale of the multi-scale convolutional network, n is the total number of receptive field scales in the multi-scale convolutional network, {R f |f∈[1,n]} is a data sequence consisting of n-scale receptive fields in a multi-scale convolutional network, Var is a variance operator, Var({R f |f∈[1,n]}) is {R f |The variance calculation value of f∈[1,n]};

[0112] R f =2(r f -1)*(m-1)+m;

[0113] In the formula, r f is the number of holes at the fth scale of the multi-scale convolutional network, and m is the convolution kernel size.

[0114] For example, when r f When set to 1, 2, and 3, the receptive fields of the convolution kernels in the multi-scale convolutional network are 3×3×3, 7×7×7, and 11×11×11, respectively. Multi-scale convolutional networks with different hole numbers are combined into a multi-scale convolutional network to extract information from feature data.

[0115] Specifically, the feature extraction operation in the multi-scale convolutional network is implemented by 3D convolution, where the calculation formula for calculating the value at position (x, y, z) on the jth feature map in the i-th layer of the 3D convolutional network is:

[0116]

[0117] in, is the value at position (x, y, z) on the jth feature map of the i-th layer, H i , W i and D i is the height, width and depth of the convolution kernel of the i-th layer, f is the activation function of the i-th layer, is the weight parameter at the convolution kernel position (h,w,d) connected to the ma-th feature map of the previous layer, is the value at position (x+h,y+w,z+d) on the ma-th feature map of the previous layer, b ij is the bias of the jth feature map in the i-th layer.

[0118] The present invention adds a weight coefficient to CE(D_t, D_s) and associates the weight coefficient with the receptive field of the multi-scale convolutional network, so that the richer the receptive field of the multi-scale convolutional network, the higher the model performance of the teacher network and the higher the learning difficulty. Therefore, the higher the CE(D_t, D_s) weight coefficient is set, the higher the learning priority of the student network is, and more learning resources are occupied to ensure that the student network learns the teacher network to the greatest extent, or to increase the degree of guidance of the teacher network to the student network, so that the student network can have the performance of the teacher network to the greatest extent.

[0119] Therefore, the present invention realizes the adaptive learning of the student network to the teacher network, or the adaptive guidance of the teacher network to the student network according to its own structure, which is more robust and further ensures the prediction performance of the student network.

[0120] The methods for screening deformation points include:

[0121] The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point;

[0122] Set the deformation threshold and compare the predicted value of the load deformation degree at each spatial point with the deformation threshold, where:

[0123] When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point;

[0124] When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

[0125] like Figure 2 As shown, the present invention provides an early warning system for load deformation of a steel structure roof, according to an early warning method for load deformation of a steel structure roof, the system comprises:

[0126] Monitoring sensors, including visual sensors and X-ray devices, are used to monitor characteristic data used to quantify the degree of roof load deformation at spatial points;

[0127] The deep learning unit uses the teacher-student model to conduct convolutional learning on the correlation between feature data and roof load deformation, and constructs a deformation quantification lightweight model for quantifying the roof load deformation degree;

[0128] The deformation warning unit selects deformation points from various spatial points on the steel structure roof according to the roof load deformation degree quantified by the deformation quantification lightweight model, and performs deformation warning.

[0129] The method for constructing a deformation quantization lightweight model by the deep learning unit includes:

[0130] Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data;

[0131] At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data;

[0132] The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data;

[0133] The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data;

[0134] The teacher network and the student network are jointly trained through sample input data and sample output data, wherein the teacher network provides training guidance for the student network;

[0135] The trained student network is used as a deformable lightweight model;

[0136] The teacher network is: D_t = BP_t (MSC (S));

[0137] The student network is: D_s = BP_s(S);

[0138] The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s);

[0139] Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

[0140] The monitoring method of the characteristic data by the monitoring sensor includes:

[0141] Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points;

[0142] Use X-ray instrument to measure stress data at spatial points;

[0143] The image data of the steel structure member, the image data of the abnormal load and the stress data are used as feature data.

[0144] The deformation warning unit screens deformation points by:

[0145] The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point;

[0146] Set the deformation threshold and compare the predicted value of the load deformation degree at each spatial point with the deformation threshold, where:

[0147] When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point;

[0148] When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

[0149] The present invention utilizes a teacher-student model and a large-scale network to train a high-performance and highly generalized teacher model for predicting the degree of load deformation, which is applied to guide the training of a lightweight student model for predicting the degree of load deformation, so that the student model has the same prediction performance as the teacher model, and the parameter processing process and number are greatly reduced, thereby achieving model compression and acceleration, and meeting the timeliness requirements of roof deformation monitoring.

[0150] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.

Claims

1. A method for early warning of load deformation of a steel structure roof, characterized in that: The following steps are involved: On the steel structure roof, determine the spatial points for monitoring the load deformation of the roof; At the spatial point, characteristic data for quantifying the degree of deformation of the roof load is monitored by monitoring sensors; The teacher-student model is used to perform convolutional learning on the correlation between the characteristic data and the degree of deformation of the roof load, and a deformation quantification lightweight model for quantifying the degree of deformation of the roof load is constructed; According to the roof load deformation degree quantified by the deformation quantification lightweight model, the deformation points are screened out from various spatial points on the steel structure roof, and deformation warning is carried out.

2. The early warning method for load deformation of a steel structure roof according to claim 1, characterized in that: The method for determining the spatial point position includes: Finite element modeling and analysis are used to divide the steel structure roof into dense and sparse grids according to stress concentration; Set the grid density threshold, mark the grids on the steel structure roof whose size is greater than or equal to the grid density threshold as sparse grids, and mark the grids on the steel structure roof whose size is less than the grid density threshold as fine grids, where the fine grids are stress concentration areas on the steel structure roof, and the sparse grids are stress non-concentration areas on the steel structure roof; The center of the fine grid on the steel structure roof is marked as the spatial point.

3. The early warning method for load deformation of a steel structure roof according to claim 1, characterized in that: The characteristic data monitoring method comprises: Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points; Use X-ray instrument to measure stress data at spatial points; The image data of the steel structure member, the image data of the abnormal load and the stress data are used as the characteristic data.

4. The early warning method for load deformation of a steel structure roof according to claim 1, characterized in that: The method for constructing the deformation quantization lightweight model includes: Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data; At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data; The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data; The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data; The teacher network and the student network are jointly trained by using sample input data and sample output data, wherein the teacher network provides training guidance for the student network; The trained student network is used as the deformed lightweight model; The teacher network is: D_t=BP_t(MSC(S)); The student network is: D_s=BP_s(S); The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s); Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

5. The early warning method for load deformation of a steel structure roof according to claim 1, characterized in that: The weight coefficients are associated with the receptive field of the multi-scale convolutional network; Wherein, the weight coefficient is: k = 1 + [1-Var ({R f |f∈[1,n]})]log(1 / Var({R f |f∈[1,n]})); Where k is the weight coefficient of CE(D_t,D_s), R f is the receptive field of the fth scale of the multi-scale convolutional network, n is the total number of receptive field scales in the multi-scale convolutional network, {R f |f∈[1,n]} is a data sequence consisting of n-scale receptive fields in a multi-scale convolutional network, Var is a variance operator, Var({R f |f∈[1,n]}) is {R f |The variance calculation value of f∈[1,n]}; R f =2(r f -1)*(m-1)+m; In the formula, r f is the number of holes at the fth scale of the multi-scale convolutional network, and m is the convolution kernel size.

6. The early warning method for load deformation of a steel structure roof according to claim 1, characterized in that: The method for screening the deformation points includes: The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point; A deformation threshold is set, and the predicted value of the load deformation degree at each spatial point is compared with the deformation threshold, wherein: When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point; When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

7. An early warning system for load deformation of steel structure roof, characterized in that: According to an early warning method for load deformation of a steel structure roof according to any one of claims 1 to 6, the system comprises: Monitoring sensors, including visual sensors and X-ray devices, are used to monitor characteristic data used to quantify the degree of deformation of the roof load at the spatial point; The deep learning unit uses a teacher-student model to perform convolutional learning on the correlation between the characteristic data and the degree of deformation of the roof load, and constructs a deformation quantification lightweight model for quantifying the degree of deformation of the roof load; The deformation warning unit selects deformation points from various spatial points on the steel structure roof according to the roof load deformation degree quantified by the deformation quantification lightweight model, and performs deformation warning.

8. The early warning system for load deformation of a steel structure roof according to claim 1, characterized in that: The method for constructing a deformation quantization lightweight model by the deep learning unit includes: Select multiple moments as sample moments, and record image data of steel structure components at spatial points, image data of abnormal loads, and stress data as input sample data; At each sample moment, the load deformation degree at the spatial point is manually calculated as the output sample data; The multi-scale convolutional network and BP neural network are combined to form a teacher network that quantifies the degree of load deformation according to the characteristic data; The BP neural network is constructed into a student network that quantifies the degree of load deformation according to the characteristic data; The teacher network and the student network are jointly trained by using sample input data and sample output data, wherein the teacher network provides training guidance for the student network; The trained student network is used as the deformed lightweight model; The teacher network is: D_t=BP_t(MSC(S)); The student network is: D_s=BP_s(S); The loss function of the student network is: loss = CE (D_r, D_s) + kCE (D_t, D_s); Where D_t is the load deformation degree output by the teacher network, D_s is the load deformation degree output by the student network, D_r is the true value of the load deformation degree, S is the feature data, loss is the loss function value, MSC is the multi-scale convolutional network, BP_t and BP_s are both BP neural networks, CE is the cross entropy, CE(D_r,D_s) is the cross entropy between D_r and D_s, CE(D_t,D_s) is the cross entropy between D_t and D_s, and k is the weight coefficient of CE(D_t,D_s).

9. The early warning system for load deformation of a steel structure roof according to claim 1, characterized in that: The monitoring method of the characteristic data by the monitoring sensor includes: Using visual sensors to collect image data of steel structure components at spatial points, and to collect image data of abnormal loads at spatial points; Use X-ray instrument to measure stress data at spatial points; The image data of the steel structure member, the image data of the abnormal load and the stress data are used as the characteristic data.

10. The early warning system for load deformation of a steel structure roof according to claim 1, characterized in that: The deformation warning unit screens deformation points by: The load deformation degree of each spatial point is predicted using the deformation quantization lightweight model to obtain the predicted value of the load deformation degree at each spatial point; A deformation threshold is set, and the predicted value of the load deformation degree at each spatial point is compared with the deformation threshold, wherein: When the predicted value of the load deformation degree is greater than or equal to the deformation threshold, the corresponding spatial point is marked as a deformation point; When the predicted value of the load deformation degree is less than the deformation threshold, the corresponding spatial point is marked as an undeformed point.

Images