A mine shovel multi-working-condition stress prediction method based on a composite residual connection neural network

Abstract

A kind of mining electric shovel multi-working condition stress prediction method based on composite residual connection neural network belongs to the technical field of engineering machinery structure health monitoring. First, key data is collected and processed to obtain high-fidelity working condition data of the front-end working mechanism of the electric shovel. Second, a digital twin of the front-end working mechanism of the mining electric shovel is preliminarily constructed. Third, the definitions and contents of the working condition vector and the stress vector are clarified, including the working condition and the to-be-solved stress data. Finally, a mapping relationship model of the working condition vector and the stress vector based on the composite residual connection neural network is constructed on the basis of the digital twin, and the trained mapping relationship model is obtained after training. By constructing a high-fidelity composite residual connection neural network, the stress of the key points of the front-end working mechanism of the mining electric shovel is solved in real time under the given working condition, and the operation and maintenance mode of the mining electric shovel is changed from regular maintenance and passive maintenance after failure to advanced prediction, accurate prediction and taking targeted measures in advance.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring technology for engineering machinery, and relates to a multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network. Background Technology

[0002] The mining industry produces raw materials essential to all sectors of modern society's production and daily life, making it a fundamental and strategic industry for the nation, directly impacting national resource security and the stability of industrial and supply chains. Electric shovels are core equipment in open-pit mining operations, undertaking the crucial tasks of stripping overburden, extracting ore, and loading. Their operational efficiency directly affects mine output and economic benefits. Currently, developing health monitoring technology for electric shovels to ensure their long-term safe, reliable, and efficient operation, extending their maintenance cycle, and reducing repair and replacement costs has become an important goal of modern mine management. Stress impact at critical nodes of electric shovels is one of the main causes of wear and even damage; achieving reliable and real-time stress prediction is a key aspect of electric shovel health monitoring.

[0003] Currently, the operation and maintenance of mining shovels in the engineering field mainly relies on planned preventive maintenance with fixed cycles and passive repair after failure. This method has obvious drawbacks. On the one hand, components that have not yet reached the end of their service life may be replaced prematurely, increasing the cost of spare parts. On the other hand, abnormal conditions such as structural fatigue and damage accumulation may not be detected in time, leading to sudden or even catastrophic failures. The current proposed stress prediction approach mainly uses finite element simulation for simulation, and then uses it as a training dataset to introduce data-driven calculations for proxying. A typical technology is the health monitoring method for mining electric shovel working device based on multidisciplinary digital twin proposed by Zhang Tianci et al. of Yanshan University (Chinese Invention Patent 202510880273.8). Existing technologies face the following challenges: (1) Traditional methods are limited by mechanical structures, and the stress and strain of a large number of key points cannot be directly measured by sensors, resulting in high costs for physical machine experiments; (2) The neural network used by existing methods has insufficient fitting ability and insufficient fitting fidelity; (3) Existing methods are prone to gradient vanishing or gradient explosion during gradient backpropagation.

[0004] In summary, there is an urgent need for a neural network that can predict stress at key points in real time based on operating conditions, has high fidelity, and can stably backpropagate gradients. Summary of the Invention

[0005] To address the problems existing in current stress prediction technologies for mining electric shovels, this invention provides a multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network. This method constructs a high-fidelity composite residual connection neural network to achieve real-time stress calculation at key points of the front-end working mechanism of the mining electric shovel under given working conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network, the method comprising the following steps: Step 1: Collect and process key data to obtain high-fidelity operating condition data of the electric shovel's front-end working mechanism; specifically: Step 1.1: Collect key data using angle sensors, torque sensors, speed sensors, and a binocular camera. The angle sensors collect the boom tilt angle θ1 and stick tilt angle θ2; the torque sensors collect the output torque T1 of the hoisting mechanism and the output torque T2 of the pushing mechanism; the speed sensor collects the hoisting mechanism speed r; and the binocular camera collects the stick extension length d.

[0007] Step 1.2: Process the key data collected in Step 1.1 to obtain high-fidelity operating condition data. Use least squares fitting and Lagrange interpolation to reasonably correct the operating condition information and obtain high-fidelity operating condition data.

[0008] Step 2: Utilize high-fidelity operating data to initially construct a digital twin of the front-end working mechanism of the mining electric shovel; specifically: Step 2.1: Construct a virtual 3D model of the front-end working mechanism of the mining electric shovel. A 3D model of the front-end working mechanism of the electric shovel is constructed using its geometric parameters. These geometric parameters include the absolute spatial positions and pitch angles of the hinge points of the bucket, stick, and boom.

[0009] Step 2.2: Based on the virtual 3D model, realize the virtual-real fusion and initialization of the digital twin.

[0010] The angle sensor, torque sensor, speed sensor, and binocular camera from step 1.1 are connected to the virtual 3D model from step 2.1 via IP protocol to ensure that the information collected by each sensor and the binocular camera can be updated to the virtual 3D model in real time. Before the electric shovel starts operating, the information collected by each sensor and the binocular camera is loaded into the virtual 3D model via IP protocol as initial state information, completing the initialization of the virtual 3D model and initially constructing a digital twin of the electric shovel's front-end working mechanism.

[0011] Step 3: Define and specify the working condition vector and stress vector. Specifically: Define the working condition vector and stress vector. The digital twin constructed in the second step can only passively adjust based on information collected by various sensors and binocular cameras, and cannot predict the stress at key points of the electric shovel's front-end working mechanism. This third step explicitly defines the basis for solving the stress in mathematical form, including the working condition and the stress data to be calculated. Specifically: The working condition vector consists of six parameters: boom tilt angle θ1, stick tilt angle θ2, stick extension length d, lifting mechanism output torque T1, pushing mechanism output torque T2, and lifting mechanism speed r. These parameters are combined sequentially to form a 6-dimensional working condition vector (θ1, θ2, d, T1, T2, r).

[0012] The stress vector includes stresses at six locations: the pushing mechanism, the lifting mechanism, the sheave, the boom-base hinge point A, the stick-bucket junction point B, and the traction rope. These stresses are combined sequentially to form a 6-dimensional stress vector (σ). c , σ h , σ s , σ A , σ B ,σ r These six positions are the six key points of the working mechanism at the front end of the mining electric shovel. Among them, σ c σ represents the stress at the pushing mechanism. h This indicates the stress at the lifting mechanism, σ. s σ represents the stress at the sheave. A σ represents the stress at the hinge point A between the boom and the base. B σ represents the stress at point B, the junction of the boom and the bucket. r This indicates the stress in the traction rope.

[0013] Step 4: Based on the digital twin obtained in Step 2, to achieve efficient stress calculation, a mapping model between the working condition vector and the stress vector based on a composite residual connection neural network is constructed. This model is then trained to possess reliable proxy capabilities, resulting in a well-trained mapping model. Specifically: Step 4.1: Initially construct a composite residual connection neural network mapping relationship model.

[0014] Step 4.1.1: Determine the forward propagation formula.

[0015] To solve for the stress vector from a given load condition vector, a neural network capable of reliably representing the mapping relationship between the two is needed. A composite residual connection neural network combines fitting ability with gradient propagation stability, making it suitable for large-scale basic training and able to accurately reflect the relationship between load conditions and stress. Taking any layer in this neural network, denoted as layer k, and the next layer as layer (k+1), the mathematical expression for the single-layer forward propagation of the composite residual connection neural network is as follows: x k+1 =P M res ( )x k + F k ( x k Wk (1) Where, x k x k+1 These are the input and output vectors of the k-th layer, respectively; Let x be the learnable residual mapping matrix at layer k, used to process the input x of the current layer. k Weighted mixing of different dimensions enhances the flexibility of model learning, thereby improving the model's fitting ability; operator P M res It is the sinkhorn iteration function, responsible for... Project onto a specific manifold and constrain it to a birandom matrix. It is the front-dimensional adjustment matrix of the k-th layer, which is related to the input vector x. k Multiplication changes the dimension of the input data to satisfy the core processing function F. k The requirement is thus used as F k The independent variable. It is the post-dimensional adjustment matrix of the k-th layer, which is related to the output F of the core processing function. k ( x k W k Multiply by P to make the dimension of the output result equal to that of P. M res ( )x k The terms must be consistent for easy addition later; F k This is the core processing function of the k-th layer, enabling the neural network to proxy nonlinear mapping relationships. Its specific form is designed by the developers, and it contains a learnable weight matrix W. k .

[0016] Because of formula (1) Without constraints, gradient vanishing or exploding may occur during error backpropagation, preventing the parameters within the neural network from optimizing towards smaller errors. Therefore, this invention introduces a function P... M res The residual mapping matrix is ​​projected onto a specific manifold. The projected P... M res ( ) is a double random matrix, in which all elements are non-negative and the sum of each row and each column is 1, thus ensuring stable gradient propagation. The single-layer forward propagation formula for each layer is recursively applied to obtain the overall end-to-end forward propagation formula, as shown in formula (2): (2) Where x0 is the input working condition vector; The output stress vector; L is the number of layers in the neural network; i is the index of the current layer; j is the index when calculating the multiplication. It is the residual mapping matrix of the Li-th layer; Represents the residual mapping matrix of the Lj-th layer; It is the adjustment matrix of the i-th layer in the later dimension; This represents the adjustment matrix of the i-th front dimension; This represents the input of the i-th layer; This represents the weight matrix of the i-th layer; This is the core processing function for the i-th layer.

[0017] More specifically, in this invention, its core processing function F k The radial basis function design is adopted, as shown in formula (3): (3) in, F k This is the core processing function on the k-th layer; It is the weight matrix of the k-th layer; These are kernel functions, and their specific forms are designed by developers according to different needs. Commonly used ones include Gaussian kernel function, multi-quadratic surface kernel function, etc. … They are n of the kth layer. c The vector of each center point has dimensions and Consistent, n c This refers to the number of center points, which is set by the developer. (m is the index, m=1,2,…,n) c )yes and The Euclidean distance between them.

[0018] Step 4.1.2, refine the forward propagation formula. , , First, the input x of this layer (the kth layer) is... k Perform root mean square normalization to obtain , recorded as , where RMSNorm is the root mean square normalization function. The , , The dynamic definitions of the three matrices are shown in equations (4) to (6), respectively: (4) (5) (6) in, , , For the scaling scalar of the k-th layer, It is a pre-scaled scalar. It is a post-scaled scalar. It is a residual scaling scalar; , , Let be the linear projection weight matrix of the k-th layer. It is the prelinear projection weight matrix. It is the post-linear projection weight matrix. It is the residual linear projection weight matrix; , , Let k be the bias vector of the k-th layer. It is the front bias vector. It is the back bias vector. It is the residual bias vector; all nine quantities mentioned above are learnable, meaning they need to be determined during training based on the training dataset.

[0019] Step 4.2: Obtain the training dataset and train the initially constructed neural network. Each working condition vector and the actual stress vector under that working condition together constitute a training dataset. The collection of massive training data is called the training dataset. Training, which is the process of making the neural network have reliable proxy capabilities, is essentially the process of determining the values ​​of its various learnable parameters based on the training dataset. Specifically: Step 4.2.1: Obtain the work condition vector from the training dataset using Latin hypercube sampling. Latin hypercube sampling involves dividing the value range of each dimension into N layers within a solution space of dimension p, where the range of values ​​for any dimension is finite. This ensures that each layer is sampled only once in any dimension, and then the sampled values ​​from each dimension are randomly combined. For the work condition vector... x Latin hypercube sampling is performed on the vector 0 = (θ1, θ2, d, T1, T2, r), with the number of dimensions p = 6 for the work condition vector. In Latin hypercube sampling, the number of data points in the collected training dataset is equal to the number of layers N for each dimension. Based on this, a series of work condition vectors as described in 4.2.2 are obtained, denoted as... , , ..., .

[0020] Step 4.2.2: The stress vectors corresponding to the working condition vectors obtained in Step 4.2.1 are obtained using computer finite element simulation and combined into a training dataset. Specifically, firstly, the working conditions are automatically input through a parameter interface. Then, a variable mesh density optimization strategy is used, with local mesh refinement at the six key points mentioned in Step 3 and sparse meshing at the remaining areas. A parallel computing framework is used to simultaneously solve for the stress under different working conditions until the stress vectors for all working conditions are derived, denoted as... , ,…, N is the number of data points in the training dataset described in 4.2.1. The stress vector and the working condition vector are combined to form the training dataset. 75% of the set is evenly selected as the training set, and the remainder is used as the test set.

[0021] Step 4.2.3: Using the training dataset described in 4.2.2, train the neural network with the goal of reducing the loss function. Take any layer in the neural network, denoted as the k-th layer, and the learnable parameters on this layer are those described in step 4.1.2. , , , , , , , , and the weight matrix W k The neural network has L layers, which are denoted as layer 0 to layer (L-1). The set of learnable parameters across all layers is then denoted as... , ∈ , where L is the number of layers in the neural network; k = 0, 1, 2…L-1; Refers to α k θ k b k The superscript of represents different parameters when its value is pre, post, or res.

[0022] First, a loss function describes the error between the predicted and actual values ​​of a neural network. The mean squared error (MSE) function, as a function sensitive to error, is a commonly used loss function in neural network training. This invention defines the mean squared loss function. This is the average of the squared Euclidean distances between the actual stress vector value and the predicted value. The loss function is: (7) in, Indicates parameters; This is the index used when calculating the summation; N is the number of data points in the training dataset described in step 4.2.1. It is the stress vector predicted by the neural network; It is the stress vector obtained from finite element simulation; It is the square of the Euclidean distance between the two; Secondly, using the automatic differentiation system in the deep learning framework, the loss MSEloss is calculated with respect to all parameters. The gradient, denoted as The automatic differentiation system calculates the gradient of each parameter backwards along the forward propagation path using the chain rule.

[0023] Subsequently, using an optimizer, the automatic differentiation system utilizes the gradient descent algorithm to express the formula... Update parameters. Among them, The parameter is the updated value; the equals sign indicates that the value on the right is assigned to the value on the left. These are the parameters before the update; The learning rate is set by the developer. This process iterates continuously until the loss condition is met. Then, a new working condition vector is input into the neural network. Within the allowable error range, the stress vector output by the neural network can be considered to be the stress vector of the front working mechanism of the mining electric shovel under real conditions. In other words, the network has reliable proxy capabilities. It is a very small number, that is .

[0024] In summary, this invention provides a multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network. This method can predict the stress in real time at the pushing mechanism, lifting mechanism, sheave, boom-base hinge point, stick-bucket junction point, and traction rope based on the working conditions collected by sensors.

[0025] The beneficial effects of this invention are: (1) This invention achieves real-time prediction, stable gradient propagation during training, and high-fidelity prediction. Compared with computer finite element simulation, neural networks have lower computational overhead and faster computation speed, enabling real-time prediction. Compared with traditional neural networks, this method uses the sinkhorn iterative algorithm to transform H... res The constraint is a double random matrix P M res ( This invention achieves stable gradient propagation, avoiding the obstruction of learnable parameter optimization caused by gradient vanishing or exploding. Compared to residual connection neural networks, this invention uses the current layer input x... k Expand to P M res ( )x k , and introduce learnable H resMatrix pair x k Weighted mixing of different dimensions can significantly improve the model's ability to fit the functional relationship between stress vector and working condition vector, ensuring high fidelity.

[0026] (2) This invention is used to predict the stress at key points of the front working mechanism of a mining electric shovel, providing an important basis for the health monitoring of the mining electric shovel and promoting the optimization of the operation and maintenance mode of the mining electric shovel. Through a neural network based on composite residual connections, the stress at six key points can be calculated in real time: the pushing mechanism, the lifting mechanism, the sheave, the hinge point A between the boom and the base, the junction point B between the stick and the bucket, and the traction rope. These six key points are areas in the front working mechanism of the mining electric shovel that are sensitive to stress and have a high risk of failure. The stress vector measured by this invention can provide an important basis for the maintenance and failure risk assessment of these areas, enabling the operation and maintenance mode of the mining electric shovel to change from regular inspection and passive maintenance after failure to proactive prediction, accurate judgment, and early targeted measures. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of a mining electric shovel in an embodiment of the present invention.

[0028] Figure 2 This is a flowchart of the present invention.

[0029] Figure 3 This is a schematic diagram of a composite residual connection neural network.

[0030] In the diagram: 1. Sheave; 2. Boom; 3. Traction rope; 4. Pushing mechanism; 5. Bucket; 6. Joint between bucket and stick (B); 7. Stick; 8. Hinge between boom and base (A); 9. Lifting mechanism. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0032] A multi-condition stress prediction method for mining electric shovels based on composite residual connection neural networks, such as... Figure 2 As shown, the specific steps include:

[0033] Step 1: Collect and process key data to obtain high-fidelity operating condition data of the electric shovel's front-end working mechanism; specifically: Step 1.1: Key data is collected using angle sensors, torque sensors, speed sensors, and a binocular camera. The front-end working mechanism of the mining electric shovel is as follows... Figure 1As shown, 1 is the sheave; 2 is the boom; 3 is the traction rope; 4 is the pushing mechanism; 5 is the bucket; 6 is the junction of the bucket and the stick (B); 7 is the stick; 8 is the hinge point between the boom and the base (A); and 9 is the lifting mechanism. An angle sensor collects the boom tilt angle θ1 and the stick tilt angle θ2; a torque sensor collects the output torque T1 of the lifting mechanism and the output torque T2 of the pushing mechanism; a speed sensor collects the speed r of the lifting mechanism; and a binocular camera collects the stick extension length d.

[0034] Step 1.2: Process the key data collected in Step 1.1 to obtain high-fidelity operating condition data. Use least squares fitting and Lagrange interpolation to reasonably correct the operating condition information and obtain high-fidelity operating condition data.

[0035] Step 2: Utilize high-fidelity operating data to initially construct a digital twin of the electric shovel's front-end working mechanism; specifically: Step 2.1: Construct a virtual 3D model of the front-end working mechanism of the mining electric shovel. A 3D model of the front-end working mechanism of the electric shovel is constructed using its geometric parameters. These geometric parameters include the absolute spatial positions and pitch angles of the hinge points of the bucket, stick, and boom.

[0036] Step 2.2: Based on the virtual 3D model, realize the virtual-real fusion and initialization of the digital twin.

[0037] The angle sensor, torque sensor, speed sensor, and binocular camera from step 1.1 are connected to the virtual 3D model from step 2.1 via IP protocol to ensure that the information collected by each sensor and the binocular camera can be updated to the virtual 3D model in real time. Before the electric shovel starts operating, the information collected by each sensor and the binocular camera is loaded into the virtual 3D model via IP protocol as initial state information, completing the initialization of the virtual 3D model and initially constructing a digital twin of the electric shovel's front-end working mechanism.

[0038] Step 3: Define and specify the working condition vector and stress vector. Specifically: Define the working condition vector and stress vector. The digital twin constructed in the second step can only passively adjust based on information collected by various sensors and binocular cameras, and cannot predict the stress at key points of the electric shovel's front-end working mechanism. This third step explicitly defines the basis for solving the stress in mathematical form, including the working condition and the stress data to be calculated. Specifically: The working condition vector consists of six parameters: boom tilt angle θ1, stick tilt angle θ2, stick extension length d, lifting mechanism output torque T1, pushing mechanism output torque T2, and lifting mechanism speed r. These parameters are combined sequentially to form a 6-dimensional working condition vector (θ1, θ2, d, T1, T2, r).

[0039] The stress vector includes stresses at six locations: the pushing mechanism, the lifting mechanism, the sheave, the boom-base hinge point A, the stick-bucket junction point B, and the traction rope. These stresses are combined sequentially to form a 6-dimensional stress vector (σ). c , σ h , σ s , σ A , σ B ,σ r These six positions are the six key points of the working mechanism at the front end of the mining electric shovel. Among them, σ c σ represents the stress at the pushing mechanism. h This indicates the stress at the lifting mechanism, σ. s σ represents the stress at the sheave. A σ represents the stress at the hinge point A between the boom and the base. B σ represents the stress at point B, the junction of the boom and the bucket. r This indicates the stress in the traction rope.

[0040] Step 4: To achieve efficient stress calculation, a mapping model between the working condition vector and the stress vector based on a composite residual connection neural network is constructed and trained to possess reliable proxy capabilities, resulting in a well-trained mapping model. Specifically: Step 4.1: Initially construct a composite residual connection neural network mapping relationship model.

[0041] Step 4.1.1: Determine the forward propagation formula.

[0042] To solve for the stress vector from a given load condition vector, a neural network capable of reliably representing the mapping relationship between the two is needed. A composite residual connection neural network combines fitting ability with gradient propagation stability, making it suitable for large-scale basic training and capable of accurately reflecting the relationship between load conditions and stress. In this neural network, let's take any layer, denoted as layer k, and the next layer as layer (k+1). The computational flow of the composite residual connection neural network at layer k is as follows... Figure 3 As shown, the mathematical expression for its single-layer forward propagation is as follows: x k+1 =P M res ( )x k + F k ( x k W k (1) Where, x k x k+1 These are the input and output vectors of the k-th layer, respectively; Let x be the learnable residual mapping matrix at layer k, used to process the input x of the current layer.k Weighted mixing of different dimensions enhances the flexibility of model learning, thereby improving the model's fitting ability; operator P M res It is the sinkhorn iteration function, responsible for... Project onto a specific manifold and constrain it to a birandom matrix. It is the front-dimensional adjustment matrix of the k-th layer, which is related to the input vector x. k Multiplication changes the dimension of the input data to satisfy the core processing function F. k The requirement is thus used as F k The independent variable. It is the post-dimensional adjustment matrix of the k-th layer, which is related to the output F of the core processing function. k ( x k W k Multiply by P to make the dimension of the output result equal to that of P. M res ( )x k The terms must be consistent for easy addition later; F k This is the core processing function of the k-th layer, enabling the neural network to proxy nonlinear mapping relationships. Its specific form is designed by the developers, and it contains a learnable weight matrix W. k .

[0043] Because of formula (1) Without constraints, gradient vanishing or exploding may occur during error backpropagation, preventing the parameters within the neural network from optimizing towards smaller errors. Therefore, this invention introduces a function P... M res The residual mapping matrix is ​​projected onto a specific manifold. The projected P... M res ( ) is a double random matrix, in which all elements are non-negative and the sum of each row and each column is 1, thus ensuring stable gradient propagation. The single-layer forward propagation formula for each layer is recursively applied to obtain the overall end-to-end forward propagation formula, as shown in formula (2): (2) Where x0 is the input working condition vector; The output stress vector; L is the number of layers in the neural network; i is the index of the current layer; j is the index when calculating the multiplication. It is the residual mapping matrix of the Li-th layer; Represents the residual mapping matrix of the Lj-th layer; It is the adjustment matrix of the i-th layer in the later dimension; This represents the adjustment matrix of the i-th front dimension; This represents the input of the i-th layer; This represents the weight matrix of the i-th layer; This is the core processing function for the i-th layer.

[0044] More specifically, in this invention, its core processing function F k The radial basis function design is adopted, as shown in formula (3): (3) in, F k This is the core processing function on the k-th layer; It is the weight matrix of the k-th layer; It is a kernel function, which is taken in this embodiment. , Pointer function The independent variable; … They are n of the kth layer. c The center point vectors are defined by the developer. In this embodiment, the definition of the center point vectors is achieved by selecting the training dataset, as detailed in step 4.2.1. c The number of center points is also set by the developers. (m is the index, m=1,2,…,n) c )yes and The Euclidean distance between them.

[0045] Step 4.1.2, on the arbitrarily chosen k-th layer, refine the forward propagation formula. , , The dynamic definition of these three matrices is based on the root mean square normalized input, denoted as... , where RMSNorm is the root mean square normalization function. The , , The dynamic definitions of the three matrices are shown in equations (4) to (6), respectively: (4) (5) (6) in, , , For the scaling scalar of the k-th layer, It is a pre-scaled scalar. It is a post-scaled scalar. It is a residual scaling scalar; , , Let be the linear projection weight matrix of the k-th layer. It is the prelinear projection weight matrix. It is the post-linear projection weight matrix. It is the residual linear projection weight matrix; , , Let k be the bias vector of the k-th layer. It is the front bias vector. It is the back bias vector. It is the residual bias vector; all nine quantities mentioned above are learnable, meaning they need to be determined during training based on the training dataset.

[0046] Step 4.2: Obtain the training dataset and train the initially constructed neural network. Each working condition vector and the actual stress vector under that working condition together constitute a training dataset. The collection of massive training data is called the training dataset. Training, which is the process of making the neural network have reliable proxy capabilities, is essentially the process of determining the values ​​of its various learnable parameters based on the training dataset. Specifically: Step 4.2.1: Obtain the work condition vector from the training dataset using Latin hypercube sampling. Latin hypercube sampling involves dividing the value range of each dimension into N layers within a solution space of dimension p, where the range of values ​​for any dimension is finite. This ensures that each layer is sampled only once in any dimension, and then the sampled values ​​from each dimension are randomly combined. For the work condition vector... x Latin hypercube sampling is performed on the vector 0 = (θ1, θ2, d, T1, T2, r), with the number of dimensions p = 6 for the work condition vector. In Latin hypercube sampling, the number of data points in the collected training dataset is equal to the number of layers N for each dimension. Based on this, a series of work condition vectors as described in 4.2.2 are obtained, denoted as... , , ..., The center point vector mentioned in step 4.1.1 is also defined here. = x m m is the index of the center point vector, m=1,2…N.

[0047] Step 4.2.2: The stress vectors corresponding to the working condition vectors obtained in Step 4.2.1 are obtained using computer finite element simulation and combined into a training dataset. Specifically, firstly, the working conditions are automatically input through a parameter interface. Then, a variable mesh density optimization strategy is used, with local mesh refinement at the six key points mentioned in Step 3 and sparse meshing at the remaining areas. A parallel computing framework is used to simultaneously solve for the stress under different working conditions until the stress vectors for all working conditions are derived, denoted as... , ,…, The stress vector and the working condition vector are combined to form the training dataset. 75% of the set is evenly selected as the training set, and the remainder is used as the test set.

[0048] Step 4.2.3: Using the training dataset described in 4.2.2, train the neural network with the goal of reducing the loss function. Take any layer in the neural network, denoted as the k-th layer, and the learnable parameters on this layer are those described in step 4.1.2. , , , , , , , , and the weight matrix W k The neural network has L layers, which are denoted as layer 0 to layer (L-1). The set of learnable parameters across all layers is then denoted as... ,in ∈ , where L is the number of layers in the neural network; k = 0, 1, 2…L-1; Refers to α k θ k b k The superscript of , when it is pre, post or res, indicates different parameters.

[0049] First, a loss function describes the error between the predicted and actual values ​​of a neural network. The mean squared error (MSE) function, as a function sensitive to error, is a commonly used loss function in neural network training. This invention defines the mean squared loss function. This is the average of the squared Euclidean distances between the actual stress vector value and the predicted value. The loss function is: (7) in, Indicates parameters; It is the index used when performing consecutive addition. N is the number of data points in the training dataset described in step 4.2.1. It is the stress vector predicted by the neural network; It is the stress vector obtained from finite element simulation; It is the square of the Euclidean distance between the two; Secondly, in Python, library functions are called to use the automatic differentiation system in the deep learning framework to calculate the loss MSEloss over all parameters. The gradient, denoted as The automatic differentiation system calculates the gradient of each parameter backwards along the forward propagation path using the chain rule.

[0050] Subsequently, the Python optimizer is invoked, and the automatic differentiation system utilizes the gradient descent algorithm, expressed as... Update parameters. Among them, The parameter is the updated value; the equals sign indicates that the value on the right is assigned to the value on the left. These are the parameters before the update; The learning rate is set by the developer. This process iterates continuously until the loss condition is met. Then, a new working condition vector is input into the neural network. Within the allowable error range, the stress vector output by the neural network can be considered to be the stress vector of the front working mechanism of the mining electric shovel under real conditions. In other words, the network has reliable proxy capabilities. It is a very small number, take .

[0051] The above-described embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A multi-condition stress prediction method for mining electric shovels based on composite residual connection neural networks, characterized in that, The multi-condition stress prediction method for mining electric shovels includes the following steps: Step 1: Collect and process key data to obtain high-fidelity working condition data of the electric shovel's front-end working mechanism; Step 2: Using the high-fidelity operating data obtained in Step 1, a preliminary digital twin of the front-end working mechanism of the mining electric shovel is constructed; specifically: Step 2.1: Construct a virtual 3D model of the front working mechanism of the mining electric shovel; Step 2.2: Based on the virtual 3D model, realize the virtual-real fusion and initialization of the digital twin to obtain the digital twin; Step 3: Define and specify the contents of the working condition vector and the stress vector; define the basis for solving the stress in mathematical form, including the working condition and the stress data to be determined. Step 4: Based on the digital twin obtained in Step 2, construct a mapping model between the working condition vector and the stress vector using a composite residual connection neural network, and train it to have reliable proxy capabilities, thus obtaining a well-trained mapping model; specifically: Step 4.1: Initially construct a composite residual connection neural network mapping relationship model; Step 4.2: Obtain the training dataset and train the initially constructed neural network; each working condition vector and the actual stress vector under that working condition are combined to form a training data, and the set of massive training data is called the training dataset.

2. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 1, characterized in that, The first step is specifically: Step 1.1: Collect key data using angle sensors, torque sensors, speed sensors, and binocular cameras; collect the boom tilt angle θ1 and stick tilt angle θ2 of the mining electric shovel using angle sensors; collect the output torque T1 of the hoisting mechanism and the output torque T2 of the pushing mechanism using torque sensors; collect the speed r of the hoisting mechanism using speed sensors; and collect the stick extension length d using binocular cameras. Step 1.2: Process the key data collected in Step 1.1 to obtain high-fidelity operating condition data; use the least squares fitting method and Lagrange interpolation method to reasonably correct the operating condition information to obtain high-fidelity operating condition data.

3. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 2, characterized in that, In the second step mentioned above: In step 2.1: a three-dimensional model of the electric shovel's front-end working mechanism is constructed using the geometric parameter information of the electric shovel; the geometric parameter information includes the absolute spatial position and pitch angle of the hinge point of the bucket, stick, and boom; In step 2.2: the angle sensor, torque sensor, speed sensor, binocular camera in step 1.1 and the virtual 3D model in step 2.1 are connected via IP protocol to ensure that the information collected by each sensor and binocular camera can be updated to the virtual 3D model in real time; Before the electric shovel starts operating, the information collected by each sensor and binocular camera is loaded into the virtual 3D model via IP protocol as initial state information, thus completing the initialization of the virtual 3D model and initially constructing a digital twin of the electric shovel's front-end working mechanism.

4. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 3, characterized in that, The third step is specifically as follows: The working condition vector consists of six parameters: boom tilt angle θ1, stick tilt angle θ2, stick extension length d, lifting mechanism output torque T1, pushing mechanism output torque T2, and lifting mechanism speed r. These parameters are combined in sequence to form a 6-dimensional working condition vector (θ1, θ2, d, T1, T2, r). The stress vector includes stresses at six locations: the pushing mechanism, the lifting mechanism, the sheave, the boom-base hinge point A, the stick-bucket junction point B, and the traction rope. These stresses are combined sequentially to form a 6-dimensional stress vector (σ). c , σ h , σ s , σ A , σ B ,σ r These six positions are the six key points of the front working mechanism of the mining electric shovel; among them, σ c σ represents the stress at the pushing mechanism. h This indicates the stress at the lifting mechanism, σ. s σ represents the stress at the sheave. A σ represents the stress at the hinge point A between the boom and the base. B σ represents the stress at point B where the boom and bucket meet. r This indicates the stress in the traction rope.

5. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 4, characterized in that, Step 4.1 specifically involves: Step 4.1.1: Determine the forward propagation formula; In a composite residual connected neural network, take any layer, denoted as layer k, and the next layer as layer (k+1). The mathematical expression for the single-layer forward propagation of the composite residual connected neural network is as follows: x k+1 =P M res ( )x k + F k ( x k ,W k )(1) Where, x k x k+1 These are the input and output vectors of the k-th layer, respectively; Let P be the learnable residual mapping matrix at the k-th layer; operator P M res It is a sinkhorn iteration function used to... Project onto a specific manifold and constrain it to a birandom matrix; It is the front-dimensional adjustment matrix of the k-th layer, and the input vector x k Multiplication changes the dimension of the input data, resulting in F. k The independent variable; It is the post-dimensional adjustment matrix of the k-th layer, which is related to the output F of the core processing function. k ( x k W k Multiply by P to make the dimension of the output result equal to that of P. M res ( )x k Items consistent; F k This is the core processing function of the k-th layer, which internally contains a learnable weight matrix W. k ; Introducing function P M res The residual mapping matrix is ​​projected onto a specific manifold; the projected P M res ( ) is a double random matrix, all of which are non-negative and the sum of each row and each column is 1; the single-layer forward propagation formula of each layer is recursively applied to finally obtain the overall end-to-end forward propagation formula, as shown in formula (2): （2） Where x0 is the input working condition vector; The output stress vector; L is the number of layers in the neural network; i is the index of the current layer; j is the index when calculating the multiplication. It is the residual mapping matrix of the Li-th layer; Represents the residual mapping matrix of the Lj-th layer; It is the adjustment matrix of the i-th layer in the later dimension; This represents the adjustment matrix of the i-th front dimension; This represents the input of the i-th layer; This represents the weight matrix of the i-th layer; This is the core processing function of the i-th layer; Step 4.1.2, refine the forward propagation formula. , , First, input x of the kth layer k Perform root mean square normalization to obtain , recorded as , where RMSNorm is the root mean square normalization function.

6. The method for multi-condition stress prediction of mining electric shovels based on composite residual connection neural networks according to claim 5, characterized in that, In step 4.1.1, the core processing function F k The radial basis function design is adopted, as shown in formula (3): ，（3） in, F k This is the core processing function at the k-th layer; It is the weight matrix of the k-th layer; It is a kernel function; … They are n of the kth layer. c The vector of each center point has dimensions and Consistent, n c It is the number of center points; yes and The Euclidean distance between the two points, where m is the index, m=1,2,…,n c .

7. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 5, characterized in that, In step 4.1.2, the... , , The dynamic definitions of the three matrices are shown in equations (4) to (6), respectively: （4） （5） （6） in, , , For the scaling scalar of the k-th layer, It is a pre-scaled scalar. It is a post-scaled scalar. It is a residual scaling scalar; , , Let be the linear projection weight matrix of the k-th layer. It is the prelinear projection weight matrix. It is the post-linear projection weight matrix. It is the residual linear projection weight matrix; , , Let k be the bias vector of the k-th layer. It is the front bias vector. It is the back bias vector. It is the residual bias vector; all nine quantities mentioned above are learnable and are determined during training based on the training dataset.

8. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 5, characterized in that, Step 4.2 specifically involves: Step 4.2.1: Obtain the working condition vector from the training dataset using Latin hypercube sampling; For the working condition vector x Latin hypercube sampling is performed on 0=(θ1,θ2,d,T1,T2,r), with the number of dimensions of the work condition vector p=6. In Latin hypercube sampling, the number of data points in the collected training dataset is equal to the number of layers N in each dimension. Based on this, a series of work condition vectors as described in 4.2.2 are obtained, denoted as... , , ..., ; Step 4.2.2: Use computer finite element simulation to obtain the stress vectors corresponding to the working condition vectors obtained in step 4.2.1, and combine them into a training dataset; The working conditions are automatically input through the parameter interface. Then, a variable mesh density optimization strategy is used to locally refine the mesh at the six key points in the third step, while the rest of the area uses a sparse mesh. A parallel computing framework is used to simultaneously solve for stresses under different working conditions until the stress vectors for all working conditions are derived, denoted as . , ,…, N represents the number of data points in the training dataset described in 4.2.1; the stress vector and the working condition vector are combined to form the training dataset. They were divided into training and testing sets. Step 4.2.3: Train the neural network based on the training set with the goal of reducing the loss function; First, take any layer in the neural network, denoted as layer k. The learnable parameters on this layer are those described in step 4.1.

2. , , , , , , , , and the weight matrix W k The neural network has L layers, which are denoted as layers 0 to L-1. The set of learnable parameters across all layers is denoted as... , ∈ , where L is the number of layers in the neural network; k = 0, 1, 2…L-1; Refers to α k θ k b k The superscript value of `pre`, `post`, or `res` indicates different parameters. Secondly, define the mean squared loss function. It is the average of the squares of the Euclidean distances between the actual stress vector value and the pre-test value; Finally, using the automatic differentiation system in the deep learning framework, the mean squared loss function MSEloss is calculated for all parameters. The gradient, denoted as The automatic differentiation system will calculate the gradient of each parameter in reverse order along the forward propagation computation path using the chain rule. Using an optimizer, the automatic differentiation system utilizes the gradient descent algorithm, in the form of... Update parameters; among which, The parameter is the updated value; the equals sign indicates that the value on the right is assigned to the value on the left. These are the parameters before the update; The learning rate is set by the developer. This process iterates continuously until the loss condition is met. Then, a new working condition vector is input into the neural network. Within the allowable error range, the stress vector output by the neural network is considered to be the stress vector of the front working mechanism of the mining electric shovel under real conditions.

9. The multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 8, characterized in that, In step 4.2.3, the mean squared loss function is: （7） in, Indicates parameters; This is the index used when calculating the summation; N is the number of data points in the training dataset described in step 4.2.

1. It is the stress vector predicted by the neural network; It is the stress vector obtained from finite element simulation; It is the square of the Euclidean distance between the two.

10. A multi-condition stress prediction method for mining electric shovels based on a composite residual connection neural network according to claim 8, characterized in that, In step 4.2.3 It is a very small number, that is .

Images