Component state sensing method, system, medium and equipment

Abstract

The invention discloses a step-by-step embedded physical information neural network component state sensing method and system, a medium and equipment, and the method comprises the steps: constructing a component geometric model, and carrying out the first grid unit division and the second grid unit division of the component geometric model; collecting first strain data under a first node coordinate and second strain data under a second node coordinate of a part in a static state under the action of external force, training an extreme learning machine model by using the training set and the test set, constructing three multi-layer perceptron networks, training the multi-layer perceptron networks by using the data of the training set, and constructing a multi-layer perceptron network model; the test set tests the trained multi-layer sensor network, the second node coordinates are adopted as input, target second strain data are calculated and output through the multi-layer sensor network, the target second strain data are compared with the second strain data, the equivalent load of the second node coordinates of the force application face of the component is solved according to the node load expression, and the force application face of the component is obtained. And calculating a component deformation state cloud picture.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical structure component state recognition, and in particular to a step-by-step embedded physical information neural network component state perception method, system, medium and equipment. Background Art

[0002] In the field of mechanical engineering, the state identification of key mechanical component structures under static models is an important part of design, analysis, and optimization. With the advancement of science and technology and the rapid development of the manufacturing industry, the complexity and precision requirements of mechanical structural components are increasing, and the working environment is becoming increasingly harsh. The algorithm-driven static analysis method faces many challenges. In order to accurately identify the operating status of components, it is often necessary to build complex physical models and use large-scale computing resources for data processing and model training. This process is not only time-consuming and labor-intensive, but also due to the complexity of the model itself, it is often difficult to maintain high recognition accuracy while ensuring calculation speed. In addition, with the increasing complexity and intelligence of industrial equipment, higher requirements are placed on the real-time and accuracy of component status identification, making it increasingly difficult for traditional methods to meet actual needs.

[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0004] The present invention provides a method, system, medium, and device for sensing component states using a step-by-step embedded physical information neural network. These methods are designed to accelerate the computation time required to identify component states using machine learning algorithms and significantly improve recognition accuracy. This method can be applied to cantilever beams, thick plates, blocks, and other structures by constructing component geometric models for these structures and sensing component states using a step-by-step embedded physical information neural network. Specifically, this method can be applied to vibration transfer path analysis and testing for aircraft wings, gas turbine casings, and feedwater pump bodies.

[0005] A step-by-step embedded physical information neural network component state perception method, the method is used to construct a component geometric model of a cantilever beam, a thick plate, and a block, and a step-by-step embedded physical information neural network component state perception,

[0006] The method comprises,

[0007] Step 1: constructing a component geometric model and dividing the component geometric model into first grid units, numbering first nodes and recording first node coordinates, and dividing the component geometric model into second grid units, numbering second nodes and recording second node coordinates, wherein the first grid size is larger than the second grid size;

[0008] Step 2: Collect first strain data at the first node coordinate and second strain data at the second node coordinate of the component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate, and a strain gauge is installed at the second node coordinate. The first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set;

[0009] Step 3: Use the training set and the test set to train the extreme learning machine model respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as the displacement-strain transformation matrix.

[0010] Step 4: Build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ;

[0011] Step 5: Use the training set data to train the multilayer perceptron network, take the first node coordinate as input, and the first strain data as input data of the loss function. The loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss. The loss function is calculated based on the displacement-strain transformation matrix.

[0012] Step 6: Testing the trained multilayer perceptron network with a test set, using the second node coordinates as input for the multilayer perceptron network to calculate and output target second strain data, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step 7 is executed; otherwise, the process returns to step 1 to adjust the first grid size, the second grid size, step 3 to adjust the extreme learning machine model parameters, step 4 to adjust the network parameters of the multilayer perceptron network, and / or step 5 to adjust the training parameters of the multilayer perceptron network.

[0013] Step 7: Solve the equivalent load of the second node coordinate of the force-applying surface of the component according to the node load expression. The equivalent load of the second node coordinate is the stress magnitude of the second node on the outer surface of the component in a direction perpendicular to the outer surface.

[0014] Step 8: Calculate the component deformation state cloud map, calculate the second node displacement size, second node strain size, second node stress size, second node deformed coordinates, strain energy size, boundary support reaction size of the component according to the displacement at the second node coordinate and draw a cloud map.

[0015] In the method described, step three includes,

[0016] Step 3.1: Establish a finite element displacement-strain calculation model and define the element type function expression as:

[0017]

[0018] in, , 、 、 is the shape function The natural coordinate system coordinates of the type function are expressed as:

[0019]

[0020] According to the definition, the relationship between node displacement and element displacement can be obtained:

[0021]

[0022] in, is the value of the type function at the kth node, is the type function matrix, is the displacement of any node of the element, is the unit node displacement vector, calculate the unit Jacobian matrix And calculate its inverse matrix :

[0023]

[0024] in, is the unit Jacobian matrix, is the inverse matrix of the Jacobian matrix, and the displacement partial derivative is calculated. The displacement partial derivative calculation formula is:

[0025]

[0026] Among them, B is the partial differential operator, which is used to calculate the partial differential of the displacement of any node in the element through the node displacement and calculate the element deformation. Its expression is:

[0027]

[0028] in, is the positive strain in the x-axis direction, is the positive strain in the y-axis direction, is the positive strain in the z-axis direction, is the shear strain in the xy direction, is the shear strain in the yz direction, is the shear strain in the xz direction,

[0029] According to the above process, the model is simplified as follows:

[0030]

[0031] in, is the strain vector, is the column vector of the element node displacement and strain The transformation matrix has 6 rows and 24 columns, and is only related to the coordinate position of the unit node.

[0032] Step 3.2: Generate the model calculation data set, take the grid unit node coordinates as the input data set, and use the input data set to calculate the unit node displacement column vector and strain The transformation matrix As the output dataset,

[0033] Step 3.3: Build the extreme learning machine model, and construct the input layer, hidden layer and output layer of the extreme learning machine model in sequence. The input layer has 24 neurons and the hidden layer has neurons, the output layer has 144 neurons, and the activation value of the hidden layer neurons is Expressed as:

[0034]

[0035] in, is the activation function, It is The input weight vector of hidden layer neurons is self-set or randomly generated. is the bias term, which is 1,

[0036] For the weights from the hidden layer to the output layer , assuming that the output layer uses linear combination output, the output Expressed as:

[0037]

[0038] In order to minimize the overall prediction error of all samples, the output weights are found by the least squares method or the steepest descent method. , by minimizing the mean square error and solving the normal equation :

[0039]

[0040] in, is the hidden layer activation matrix, is the target output The 144 elements in the matrix are used to form a matrix, and the activation value calculation expression of the hidden layer neurons is calculated using the 24 elements in each group of the input data set. ,

[0041] Step 3.4: Through the process from step 1 to step 3, the trained extreme learning machine model is obtained, which satisfies the mapping relationship:

[0042]

[0043] in, is the element node strain magnitude, is the element node displacement vector, It is an extreme learning machine model that uses the unit node coordinates as input to calculate the displacement-strain transformation matrix G.

[0044] In the described method, the activation function is sigmoid and ReLU functions.

[0045] In the method described, in step 5,

[0046] Step 5.1: Calculate the unit displacement-strain transformation matrix based on the extreme learning machine model parameters trained in step 3 , the calculation expression is

[0047]

[0048] in, The activation values ​​of the hidden layer neurons, is the weight from the hidden layer to the output layer, is the final learning parameter result obtained in step 3,

[0049] Step 5.2: Calculate the unit deformation, the expression is:

[0050] .

[0051] In the method described, the multilayer perceptron network model is divided into four parts: input layer, subnetwork, connection layer, and output layer. The input layer is connected to three independent and identical subnetworks respectively. The output data of the three subnetworks are combined into a three-channel vector through the connection layer and output through the output layer. The input layer input data is a 4-row n-column matrix, where n is equal to the total number of nodes, and each column is sequentially represented as the x-axis coordinate of the node , y-axis coordinate , z-axis coordinate and time coordinates , where i is the grid node number, j is the sampling time point number, and each sub-network consists of a normalization layer, multiple fully connected layers, and an activation layer connected in sequence. The number of neurons in the last fully connected layer is 1, and the output layer outputs a 3-row n-column matrix, where each row represents the node displacement in the x-axis direction. , displacement in the y-axis direction , z-axis displacement .

[0052] In the method, the component includes a mechanical structure component.

[0053] In the method, the external force is the component being squeezed, stretched, or twisted.

[0054] A step-by-step embedded physical information neural network component state perception system, the system is used to build component geometric models of cantilever beams, thick plates, and blocks, and to perceive component states through a step-by-step embedded physical information neural network.

[0055] The system comprises,

[0056] a component geometric model construction unit, which constructs the component geometric model and divides the component geometric model into first grid units, numbers first nodes and records first node coordinates, and divides the component geometric model into second grid units, numbers second nodes and records second node coordinates, wherein the first grid size is larger than the second grid size;

[0057] an acquisition unit, configured to acquire first strain data at a first node coordinate and second strain data at a second node coordinate of a component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate and a strain gauge is installed at the second node coordinate, the first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set;

[0058] An extreme learning machine model unit is used to train an extreme learning machine model using a training set and a test set respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as a displacement-strain transformation matrix.

[0059] Multilayer perceptron network unit, which is used to build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ;

[0060] A training unit is used to train a multilayer perceptron network using training set data, taking the first node coordinate as input and the first strain data as input data of a loss function, wherein the loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss, wherein the loss function is calculated based on the displacement-strain transformation matrix.

[0061] a testing unit, configured to test a trained multilayer perceptron network using a test set, wherein the multilayer perceptron network calculates and outputs target second strain data using the second node coordinates as input, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step seven is executed; otherwise, the process returns to step one to adjust the first grid size and the second grid size, returns to step three to adjust the extreme learning machine model parameters, returns to step four to adjust the network parameters of the multilayer perceptron network, and / or returns to step five to adjust the training parameters of the multilayer perceptron network;

[0062] a calculation unit, configured to solve an equivalent load of a second node coordinate of the force-applying surface of the component according to a node load expression, wherein the equivalent load of the second node coordinate is a stress magnitude of the second node on the outer surface of the component in a direction perpendicular to the outer surface;

[0063] A cloud map unit is used to calculate the deformation state cloud map of the component, calculate the displacement size of the second node of the component, the strain size of the second node, the stress size of the second node, the coordinates of the second node after deformation, the strain energy size, and the boundary support reaction force size according to the displacement at the second node coordinate and draw a cloud map.

[0064] A computer storage medium includes computer instructions, which, when executed on a computer, cause the computer to execute the method described above.

[0065] An electronic device, comprising:

[0066] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0067] When the processor executes the program, the method described is implemented.

[0068] Compared with the existing technology, the present invention has the following advantages: the present invention combines the experimentally measured strain data to identify the current displacement change and load state of the component structure, and while complying with the finite element constitutive relationship, it reduces the impact of sensor noise on the model. It provides a real-time analysis method for accurate status detection of key components, early prevention of component failures, and digital twinning of component structures. It also solves the problem that there are fewer measuring points on complex equipment components and it is difficult to obtain more comprehensive sensor measurement information. The calculation speed of the present invention is significantly improved compared to the statics-based perception method in the existing technology, and the average calculation time is reduced by about 20-35%. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are intended only to illustrate preferred embodiments and are not to be construed as limiting the present invention. It should be understood that the drawings described below are merely examples of the present invention, and that those skilled in the art will be able to derive other drawings from these drawings without inventive effort. Throughout the drawings, identical reference numerals are used to denote identical components.

[0070] In the attached figure:

[0071] Figure 1 This is a structural diagram of a component state recognition acceleration method based on an extreme learning machine;

[0072] Figure 2 This is a schematic diagram of the network structure of the component state recognition acceleration method based on the extreme learning machine;

[0073] Figure 3 This is a schematic diagram of the results of component state recognition using the extreme learning machine-based component state recognition acceleration method.

[0074] The present invention will be further explained below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0075] Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0076] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the invention. The scope of protection of the present invention shall be as defined in the attached claims.

[0077] To facilitate understanding of the embodiments of the present invention, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0078] like Figures 1 to 3 As shown, the method includes,

[0079] Step 1: constructing a component geometric model and dividing the component geometric model into first grid units, numbering first nodes and recording first node coordinates, and dividing the component geometric model into second grid units, numbering second nodes and recording second node coordinates, wherein the first grid size is larger than the second grid size;

[0080] Step 2: Collect first strain data at the first node coordinate and second strain data at the second node coordinate of the component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate, and a strain gauge is installed at the second node coordinate. The first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set;

[0081] Step 3: Use the training set and the test set to train the extreme learning machine model respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as the displacement-strain transformation matrix.

[0082] Step 4: Build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ;

[0083] Step 5: Use the training set data to train the multilayer perceptron network, take the first node coordinate as input, and the first strain data as input data of the loss function. The loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss. The loss function is calculated based on the displacement-strain transformation matrix.

[0084] Step 6: Testing the trained multilayer perceptron network with a test set, using the second node coordinates as input for the multilayer perceptron network to calculate and output target second strain data, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step 7 is executed; otherwise, the process returns to step 1 to adjust the first grid size, the second grid size, step 3 to adjust the extreme learning machine model parameters, step 4 to adjust the network parameters of the multilayer perceptron network, and / or step 5 to adjust the training parameters of the multilayer perceptron network.

[0085] Step 7: Solve the equivalent load of the second node coordinate of the force-applying surface of the component according to the node load expression. The equivalent load of the second node coordinate is the stress magnitude of the second node on the outer surface of the component in a direction perpendicular to the outer surface.

[0086] Step 8: Calculate the component deformation state cloud map, calculate the second node displacement size, second node strain size, second node stress size, second node deformed coordinates, strain energy size, boundary support reaction size of the component according to the displacement at the second node coordinate and draw a cloud map.

[0087] In a preferred embodiment of the method, step three comprises:

[0088] Step 3.1: Establish a finite element displacement-strain calculation model and define the element type function expression as:

[0089]

[0090] in, , 、 、 is the shape function The natural coordinate system coordinates of the type function are expressed as:

[0091]

[0092] According to the definition, the relationship between node displacement and element displacement can be obtained:

[0093]

[0094] in, is the value of the type function at the kth node, is the type function matrix, is the displacement of any node of the element, is the unit node displacement vector, calculate the unit Jacobian matrix And calculate its inverse matrix :

[0095]

[0096] in, is the unit Jacobian matrix, is the inverse matrix of the Jacobian matrix, and the displacement partial derivative is calculated. The displacement partial derivative calculation formula is:

[0097]

[0098] Among them, B is the partial differential operator, which is used to calculate the partial differential of the displacement of any node in the element through the node displacement and calculate the element deformation. Its expression is:

[0099]

[0100] in, is the positive strain in the x-axis direction, is the positive strain in the y-axis direction, is the positive strain in the z-axis direction, is the shear strain in the xy direction, is the shear strain in the yz direction, is the shear strain in the xz direction,

[0101] According to the above process, the model is simplified as follows:

[0102]

[0103] in, is the strain vector, is the column vector of the element node displacement and strain The transformation matrix has 6 rows and 24 columns, and is only related to the coordinate position of the unit node.

[0104] Step 3.2: Generate the model calculation data set, take the grid unit node coordinates as the input data set, and use the input data set to calculate the unit node displacement column vector and strain The transformation matrix As the output dataset,

[0105] Step 3.3: Build the extreme learning machine model, and construct the input layer, hidden layer and output layer of the extreme learning machine model in sequence. The input layer has 24 neurons and the hidden layer has neurons, the output layer has 144 neurons, and the activation value of the hidden layer neurons is Expressed as:

[0106]

[0107] in, is the activation function, It is The input weight vector of hidden layer neurons is self-set or randomly generated. is the bias term, which is 1,

[0108] For the weights from the hidden layer to the output layer , assuming that the output layer uses linear combination output, the output Expressed as:

[0109]

[0110] In order to minimize the overall prediction error of all samples, the output weights are found by the least squares method or the steepest descent method. , by minimizing the mean square error and solving the normal equation :

[0111]

[0112] in, is the hidden layer activation matrix, is the target output The 144 elements in the matrix are used to form a matrix, and the activation value calculation expression of the hidden layer neurons is calculated using the 24 elements in each group of the input data set. ,

[0113] Step 3.4: Through the process from step 1 to step 3, the trained extreme learning machine model is obtained, which satisfies the mapping relationship:

[0114]

[0115] in, is the element node strain magnitude, is the element node displacement vector, It is an extreme learning machine model that uses the unit node coordinates as input to calculate the displacement-strain transformation matrix G.

[0116] In a preferred embodiment of the method, the activation function is a sigmoid or ReLU function.

[0117] In a preferred embodiment of the method, in step 5,

[0118] Step 5.1: Calculate the unit displacement-strain transformation matrix based on the extreme learning machine model parameters trained in step 3 , the calculation expression is

[0119]

[0120] in, The activation values ​​of the hidden layer neurons, is the weight from the hidden layer to the output layer, is the final learning parameter result obtained in step 3,

[0121] Step 5.2: Calculate the unit deformation, the expression is:

[0122] .

[0123] In a preferred embodiment of the method, the multilayer perceptron network model is divided into four parts: input layer, subnetwork, connection layer, and output layer. The input layer is connected to three independent and identical subnetworks respectively. The output data of the three subnetworks are combined into a three-channel vector through the connection layer and output through the output layer. The input layer input data is a 4-row n-column matrix, where n is equal to the total number of nodes, and each column is sequentially represented as the x-axis coordinate of the node. , y-axis coordinate , z-axis coordinate and time coordinates , where i is the grid node number, j is the sampling time point number, and each sub-network consists of a normalization layer, multiple fully connected layers, and an activation layer connected in sequence. The number of neurons in the last fully connected layer is 1, and the output layer outputs a 3-row n-column matrix, where each row represents the node displacement in the x-axis direction. , displacement in the y-axis direction , z-axis displacement .

[0124] In a preferred embodiment of the method, the component comprises a mechanical structural component.

[0125] In a preferred embodiment of the method, the external force is applied to the component by squeezing, stretching or twisting.

[0126] A step-by-step embedded physical information neural network component state perception system includes:

[0127] a component geometric model construction unit, which constructs the component geometric model and divides the component geometric model into first grid units, numbers first nodes and records first node coordinates, and divides the component geometric model into second grid units, numbers second nodes and records second node coordinates, wherein the first grid size is larger than the second grid size;

[0128] an acquisition unit, configured to acquire first strain data at a first node coordinate and second strain data at a second node coordinate of a component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate and a strain gauge is installed at the second node coordinate, the first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set;

[0129] An extreme learning machine model unit is used to train an extreme learning machine model using a training set and a test set respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as a displacement-strain transformation matrix.

[0130] Multilayer perceptron network unit, which is used to build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ;

[0131] A training unit is used to train a multilayer perceptron network using training set data, taking the first node coordinate as input and the first strain data as input data of a loss function, wherein the loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss, wherein the loss function is calculated based on the displacement-strain transformation matrix.

[0132] a testing unit, configured to test a trained multilayer perceptron network using a test set, wherein the multilayer perceptron network calculates and outputs target second strain data using the second node coordinates as input, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step seven is executed; otherwise, the process returns to step one to adjust the first grid size and the second grid size, returns to step three to adjust the extreme learning machine model parameters, returns to step four to adjust the network parameters of the multilayer perceptron network, and / or returns to step five to adjust the training parameters of the multilayer perceptron network;

[0133] a calculation unit, configured to solve an equivalent load of a second node coordinate of the force-applying surface of the component according to a node load expression, wherein the equivalent load of the second node coordinate is a stress magnitude of the second node on the outer surface of the component in a direction perpendicular to the outer surface;

[0134] A cloud map unit is used to calculate the deformation state cloud map of the component, calculate the displacement size of the second node of the component, the strain size of the second node, the stress size of the second node, the coordinates of the second node after deformation, the strain energy size, and the boundary support reaction force size according to the displacement at the second node coordinate and draw a cloud map.

[0135] The specific method implemented in the stepped block solid component instance includes:

[0136] like Figure 1 As shown, in step 1, the geometric model of the stepped block solid component is meshed, and the component morphology structure is converted into a mesh form using a three-dimensional isoparametric unit (such as a C3D8 unit or a C3D20 unit), and the node coordinates and numbers are recorded;

[0137] like Figure 1 As shown, in step 2, the strain data of the component during operation is collected, a load is applied to the edge of the raised portion of the component, and the strain magnitude at the node on the outer surface of the component in a static equilibrium state is collected;

[0138] like Figure 1 As shown, in step three, the extreme learning machine model is trained, as Figure 2 As shown on the right, an extreme learning machine model is established with 24 neuron nodes in the input layer and 288 neuron nodes in the hidden layer. The node coordinates in each unit are formed into a vector in the order of the node numbers as input, and the displacement-strain transformation matrix corresponding to the finite element displacement-strain calculation process is calculated as output. The extreme learning machine model is trained to realize the mapping function relationship from the unit node coordinates to the displacement-strain transformation matrix.

[0139] like Figure 1 As shown, in step 4, a multilayer perceptron network is constructed, such as Figure 2 The structure shown establishes three multilayer perceptron networks, and the inputs are all node x-axis coordinates , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ;

[0140] like Figure 1 As shown in Figure 2, in step 5, the multilayer perceptron network is trained using the training set data. The unit node coordinates and numbers in step 1 are used as the input data of the multilayer perceptron network. The component surface node strain data collected in step 2 is used as the input data of the loss function. The loss function consists of four parts: observation loss, initial value loss, boundary loss, and virtual work loss. Figure 2 As shown, the obtained unit node displacement is used to form a vector according to the unit node order , and the displacement strain function calculated by the extreme learning machine Perform matrix multiplication to obtain the unit node strain, and then add it to the material matrix Perform matrix multiplication to obtain the stress magnitude of the unit node;

[0141] like Figure 1 As shown in , in step 6, the training effect of the multilayer perceptron network is tested, and the test set node coordinates (the fine grid node coordinates in step 1) are used as network input to calculate the displacement at the test node, as shown in Figure 2 The strain magnitude at the nodes of the test set is calculated by displacement and compared with the strain magnitude measured in step 2;

[0142] like Figure 1 As shown, in step seven, the external node equivalent load of the force-applying surface is solved, and the stress at the component node is calculated according to the test node coordinates. The external node equivalent load is the stress at the component outer surface node perpendicular to the outer surface.

[0143] like Figure 1 As shown, in step eight, the deformation state cloud map of the component is calculated. According to the displacement at the test point calculated in step five, the component node displacement, node strain, node stress, node coordinate after deformation, strain energy, boundary support reaction force are calculated and a cloud map is drawn, as shown in FIG. Figure 3 Displacement and principal strain contours of the stepped block component shown under compressive load.

[0144] A computer storage medium includes computer instructions, which, when executed on a computer, cause the computer to execute the method described above.

[0145] An electronic device, comprising:

[0146] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0147] When the processor executes the program, the method described is implemented.

[0148] In practice, the calculation speed of the present invention is significantly improved compared with the statics-based perception method in the prior art, and the average calculation time is reduced by about 20-35%.

[0149] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. A step-by-step embedded physical information neural network component state perception method, characterized in that: The method is used to construct component geometric models of cantilever beams, thick plates, and blocks, as well as component state perception of step-by-step embedded physical information neural networks. The method comprises the following steps: Step 1: constructing a component geometric model and dividing the component geometric model into first grid units, numbering first nodes and recording first node coordinates, and dividing the component geometric model into second grid units, numbering second nodes and recording second node coordinates, wherein the first grid size is larger than the second grid size; Step 2: Collect first strain data at the first node coordinate and second strain data at the second node coordinate of the component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate, and a strain gauge is installed at the second node coordinate. The first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set; Step 3: Use the training set and the test set to train the extreme learning machine model respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as the displacement-strain transformation matrix. Step 4: Build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ; Step 5: Use the training set data to train the multilayer perceptron network, take the first node coordinate as input, and the first strain data as input data of the loss function. The loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss. The loss function is calculated based on the displacement-strain transformation matrix. Step 6: Test the trained multilayer perceptron network with a test set, and use the second node coordinate as input for the multilayer perceptron network to calculate and output target second strain data, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step 7 is executed, otherwise, the method returns to step 1 to adjust the first grid size, the second grid size, the extreme learning machine model parameters, the network parameters of the multilayer perceptron network, and / or the training parameters of the multilayer perceptron network are adjusted. Step 7: Solve the equivalent load of the second node coordinate of the force-applying surface of the component according to the node load expression, and the equivalent load of the second node coordinate is the stress magnitude of the second node on the outer surface of the component perpendicular to the outer surface. Step 8: Calculate the component deformation state cloud map, calculate the second node displacement size, second node strain size, second node stress size, second node deformed coordinates, strain energy size, boundary support reaction size of the component according to the displacement at the second node coordinate and draw a cloud map.

2. The method according to claim 1, characterized in that Preferably, step three includes, Step 3.1: Establish a finite element displacement-strain calculation model and define the element type function expression as: ， in, , 、 、 is the shape function The natural coordinate system coordinates of the type function are expressed as: ， According to the definition, the relationship between node displacement and element displacement can be obtained: ， in, is the value of the type function at the kth node, is the type function matrix, is the displacement of any node of the element, is the unit node displacement vector, calculate the unit Jacobian matrix And calculate its inverse matrix : ， in, is the unit Jacobian matrix, is the inverse matrix of the Jacobian matrix, and the displacement partial derivative is calculated. The displacement partial derivative calculation formula is: ， Among them, B is the partial differential operator, which is used to calculate the partial differential of the displacement of any node in the element through the node displacement and calculate the element deformation. Its expression is: ， in, is the positive strain in the x-axis direction, is the positive strain in the y-axis direction, is the positive strain in the z-axis direction, is the shear strain in the xy direction, is the shear strain in the yz direction, is the shear strain in the xz direction, According to the above process, the model is simplified as follows: ， in, is the strain vector, is the column vector of the element node displacement and strain The transformation matrix has 6 rows and 24 columns, and is only related to the coordinate position of the unit node. Step 3.2: Generate the model calculation data set, take the grid unit node coordinates as the input data set, and use the input data set to calculate the unit node displacement column vector and strain The transformation matrix As the output dataset, Step 3.3: Build the extreme learning machine model, and construct the input layer, hidden layer and output layer of the extreme learning machine model in sequence. The input layer has 24 neurons and the hidden layer has neurons, the output layer has 144 neurons, and the activation value of the hidden layer neurons is Expressed as: ， in, is the activation function, It is The input weight vector of hidden layer neurons is self-set or randomly generated. is the bias term, which is 1, For the weights from the hidden layer to the output layer , assuming that the output layer uses linear combination output, the output Expressed as: ， In order to minimize the overall prediction error of all samples, the output weights are found by the least squares method or the steepest descent method. , by minimizing the mean square error and solving the normal equation : ， in, is the hidden layer activation matrix, is the target output The 144 elements in the matrix are used to form a matrix, and the activation value calculation expression of the hidden layer neurons is calculated using the 24 elements in each group of the input data set. , Step 3.4: Through the process from step 1 to step 3, the trained extreme learning machine model is obtained, which satisfies the mapping relationship: ， in, is the element node strain magnitude, is the element node displacement vector, It is an extreme learning machine model that uses the unit node coordinates as input to calculate the displacement-strain transformation matrix G.

3. The method according to claim 1, characterized in that The activation functions are sigmoid and ReLU functions.

4. The method according to claim 1, wherein In step five, Step 5.1: Calculate the unit displacement-strain transformation matrix based on the extreme learning machine model parameters trained in step 3 , the calculation expression is ， in, The activation values ​​of the hidden layer neurons, is the weight from the hidden layer to the output layer, is the final learning parameter result obtained in step 3, Step 5.2: Calculate the unit deformation, the expression is: 。 5. The method according to claim 1, characterized in that The multilayer perceptron network model is divided into four parts: input layer, subnetwork, connection layer, and output layer. The input layer is connected to three independent and identical subnetworks. The output data of the three subnetworks are combined into a three-channel vector through the connection layer and output through the output layer. The input layer input data is a 4-row n-column matrix, where n is equal to the total number of nodes, and each column is sequentially represented as the x-axis coordinate of the node. , y-axis coordinate , z-axis coordinate and time coordinates , where i is the grid node number, j is the sampling time point number, and each sub-network consists of a normalization layer, multiple fully connected layers, and an activation layer connected in sequence. The number of neurons in the last fully connected layer is 1, and the output layer outputs a 3-row n-column matrix, where each row represents the node displacement in the x-axis direction. , displacement in the y-axis direction , z-axis displacement .

6. The method according to claim 1, characterized in that Components include mechanical structural components.

7. The method according to claim 1, characterized in that The external force acts as compression, stretching and twisting on the components.

8. A step-by-step embedded physical information neural network component state perception system, characterized in that: The system is used to construct component geometric models of cantilever beams, thick plates, and blocks, as well as component state perception of step-by-step embedded physical information neural networks. The system comprises, a component geometric model construction unit, which constructs the component geometric model and divides the component geometric model into first grid units, numbers first nodes and records first node coordinates, and divides the component geometric model into second grid units, numbers second nodes and records second node coordinates, wherein the first grid size is larger than the second grid size; an acquisition unit, configured to acquire first strain data at a first node coordinate and second strain data at a second node coordinate of a component in a stationary state under the action of an external force, wherein a strain gauge is installed at the first node coordinate and a strain gauge is installed at the second node coordinate, the first node coordinate and its first strain data constitute a training set, and the second node coordinate and its second strain data constitute a test set; An extreme learning machine model unit is used to train an extreme learning machine model using a training set and a test set respectively. The extreme learning machine model includes an input layer, a hidden layer, and an output layer. The input layer inputs the first node coordinates and the second node coordinates. The hidden layer calculates and the output layer outputs the calculation results as a displacement-strain transformation matrix. Multilayer perceptron network unit, which is used to build three multilayer perceptron networks, with the input being the x-axis coordinates of the nodes , y-axis coordinate , z-axis coordinate and time nodes , the outputs are x-axis displacement , displacement in the y-axis direction , z-axis displacement , combining the three multilayer perceptron network outputs into a node displacement vector ; A training unit is used to train a multilayer perceptron network using training set data, taking the first node coordinate as input and the first strain data as input data of a loss function, wherein the loss function consists of four parts: observation loss, constitutive loss, boundary loss, and virtual work loss, wherein the loss function is calculated based on the displacement-strain transformation matrix. a testing unit, configured to test a trained multilayer perceptron network using a test set, wherein the multilayer perceptron network calculates and outputs target second strain data using the second node coordinates as input, wherein the target second strain data is calculated based on the displacement-strain transformation matrix, and the target second strain data is compared with the second strain data. If the difference is less than a threshold, step seven is executed; otherwise, the process returns to step one to adjust the first grid size and the second grid size, returns to step three to adjust the extreme learning machine model parameters, returns to step four to adjust the network parameters of the multilayer perceptron network, and / or returns to step five to adjust the training parameters of the multilayer perceptron network; a calculation unit, configured to solve an equivalent load of a second node coordinate of the force-applying surface of the component according to a node load expression, wherein the equivalent load of the second node coordinate is a stress magnitude of the second node on the outer surface of the component in a direction perpendicular to the outer surface; A cloud map unit is used to calculate the deformation state cloud map of the component, calculate the displacement size of the second node of the component, the strain size of the second node, the stress size of the second node, the coordinates of the second node after deformation, the strain energy size, and the boundary support reaction force size according to the displacement at the second node coordinate and draw a cloud map.

9. A computer storage medium, characterized in that The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

Images