Engineering inspection robot positioning method based on multi-sensor data fusion

Abstract

The present application relates to the field of data processing, in particular to a method for positioning an engineering inspection robot based on multi-sensor data fusion, which comprises: collecting real sensor data through at least two different types of sensors, generating generated sensor data through a meta-learning-based generative adversarial network algorithm, training a feature extraction model using a neural network algorithm optimized based on sea wave energy capture, training a feature dimension reduction model using a dynamic adjustment matrix-based autoencoder algorithm, training a classifier using a limit learning machine algorithm based on an adaptive error correction term, collecting new real sensor data and inputting it into the trained classifier to obtain the position classification of the engineering inspection robot in the position environment. The existing positioning method has relatively inaccurate positioning accuracy due to the reasons of small sample data quantity and weak data processing capability, while the present application expands the sample data quantity and optimizes the data processing, thus having relatively accurate positioning accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information monitoring of digital data processing, and particularly to an engineering patrol robot positioning method based on multi-sensor data fusion. BACKGROUND

[0002] With the continuous improvement of industrial automation and intelligence level, engineering patrol robots play an increasingly important role in monitoring and maintaining industrial environments. These engineering patrol robots need to be accurately positioned in various complex environments to know the specific area they are in, such as office area, production workshop area, etc., in order to effectively perform tasks such as inspection, monitoring and fault diagnosis, etc.

[0003] The existing positioning method usually includes four steps of data collection by sensors, feature extraction of collected data, dimensionality reduction of feature data, and classification of reduced data.

[0004] The collected data is the basis for subsequent training data, but the collection, labeling and preprocessing are time-consuming and labor-intensive. Traditional engineering patrol robot positioning methods rely on a single type of sensor for data collection, such as vision or radar signals, using real sensor data collected by the sensor as sample data. A single type of sensor may face problems such as signal interference in varying environmental conditions, resulting in low data accuracy and small amount of collected data. The lack of sample data limits the training effect and generalization ability of the model, resulting in poor generalization ability of the model, which affects the accuracy of the model, making it difficult for training sample data to fully cover the data diversity of application scenarios.

[0005] However, increasing the number of sensor types and numbers increases the difficulty of data processing, which ultimately affects the positioning results. For example, existing engineering patrol robot positioning methods use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient explosion or getting stuck in local optimal solution may occur. Therefore, the existing positioning method often fails to fully dynamically adjust the extraction strategy of key features, which may result in insufficient learning of the model on specific data features, affecting the stability of the training and the performance of the model, and further affecting the final positioning accuracy. In terms of feature dimensionality reduction, the existing positioning method fails to efficiently encode and reconstruct complex data, and the loss of information during dimensionality reduction may result in a decrease in the performance of the final model. In the process of classifying reduced data, the existing positioning method may have difficulty capturing minor data changes due to the lack of weight initialization and updating strategy when processing complex or low-dimensional data, which limits the classification accuracy and reaction sensitivity, thereby affecting the positioning accuracy of the model.

[0006] Therefore, the existing positioning method has the problem of inaccurate positioning accuracy. SUMMARY

[0007] The technical problem solved by the present application is to overcome the above-mentioned shortcomings of the prior art, and to provide an engineering inspection robot positioning method based on multi-sensor data fusion with high positioning accuracy.

[0008] To solve the above technical problems, the engineering inspection robot positioning method based on multi-sensor data fusion provided by the present application comprises the following steps:

[0009] S1, collecting real sensor data through at least two different types of sensors, classifying and labeling the real sensor data to obtain labeled real sensor data;

[0010] S2, generating generated sensor data based on the meta-learning-based generative adversarial network algorithm according to the labeled real sensor data, wherein the labeled real sensor data and the generated sensor data constitute sample data;

[0011] S3, inputting the sample data into a feature extraction model, training the feature extraction model using a neural network algorithm based on sea wave energy capture optimization to obtain a trained feature extraction model and feature-extracted sample data;

[0012] S4, inputting the feature-extracted sample data into a feature dimension reduction model, training the feature dimension reduction model using a self-encoder algorithm based on a dynamically adjusted matrix to obtain a trained feature dimension reduction model and dimension-reduced sample data;

[0013] S5, inputting the dimension-reduced sample data into a classifier, training the classifier using a limit learning machine algorithm based on an adaptive error correction term to obtain a trained classifier;

[0014] S6, collecting new real sensor data through at least two different types of sensors, inputting the new real sensor data into the trained classifier to obtain the position classification of the engineering inspection robot in the position environment, and realizing the position positioning of the engineering inspection robot.

[0015] As a further improvement of the present application: the labeled categories in S1 include machine room area, office area, production workshop area and warehouse area.

[0016] As a further improvement of the present application: the training steps of the meta-learning-based generative adversarial network algorithm in S2 are as follows:

[0017] S201, initializing generator parameters and discriminator parameters

[0018] S202, the generator receives a random noise vector z c , generates corresponding new generated sensor data;

[0019] S203, according to the feedback provided by the discriminator, the meta-learning optimizer is used to adjust the parameter update strategy of the generator and the discriminator;

[0020] S204, according to the gradient descent strategy, the generator parameters adjusted based on the meta-learning are updated again and the discriminator parameters ;

[0021] S205, repeat the above steps until the preset stop iteration condition is met, that is, the feature model training is completed.

[0022] As a further improvement of the application: the training step of training the feature extraction model in S3 is as follows:

[0023] S301, initialize the weight parameters of the neural network;

[0024] S302, set an adaptive mask layer in the first 3 layers of the network, which dynamically adjusts its transmittance according to the characteristics of the input data;

[0025] S303, simulate the dynamic behavior of sea waves to optimize the weights of the neural network, and the calculation method of the update amount of the neural network weight is:

[0026]

[0027] In the formula, ΔW p is the update amount of the neural network weight; η p is the learning rate of the neural network; is the energy gradient corresponding to the current weight of the neural network; α p is the momentum factor of the neural network; is the update amount of the weight of the last iteration of the neural network;

[0028] S304, the input data passes through the neural network and adopts the forward propagation mode to generate output after 6 layers of processing;

[0029] S305, repeat S302-S304 until the preset stop iteration condition is met, that is, the feature extraction model training is completed.

[0030] As a further improvement of the application: the training step of training the feature extraction model in S3 is as follows:

[0031] S401, initialize the parameters of the autoencoder algorithm based on dynamic adjustment matrix, including the weights and biases of the autoencoder;

[0032] S402. Data is passed layer by layer through the encoder via forward propagation. The encoder output is as follows:

[0033] h r =Re(W r ·X r +b r ),

[0034] In the formula, h r For the encoded hidden layer features, Re() is the ReLU activation function, and W... r X represents the weight parameters of the autoencoder. r For input data, b r These are the bias parameters for the autoencoder;

[0035] S403. Calculate the composite loss function based on data fidelity and coding sparsity control to evaluate model performance. The loss function L of the autoencoder... r The calculation method is as follows:

[0036]

[0037] In the formula, ∥∥ is the L2 norm, X r Xe is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder, and λ is the input data. r Let W be the regularization coefficient of the autoencoder. r ∥ F Weight W r The Frobenius norm, ρ r The coefficient n is used to control the strength of sparsity regularization. g This represents the number of autoencoder layers. As an indicator function, when h rj Greater than θ rj The value is 1 if it is true, and 0 otherwise.

[0038] S404. Repeat S402 to S403 until the preset stopping iteration condition is met, which means that the feature dimensionality reduction model training is complete.

[0039] As a further improvement of the present invention, the training steps for classifier training in S5 are as follows:

[0040] S501, Regarding the weight parameters W of the Extreme Learning Machine u Perform initialization;

[0041] S502. Normalize the reduced low-dimensional feature vectors.

[0042] S503. Input the normalized features into the Extreme Learning Machine classifier;

[0043] S504, loss calculation and back propagation are performed using a loss function based on an adaptive error correction term, to ensure the generalization ability of the model, and the loss function L u The calculation method is as follows:

[0044]

[0045] In the formula, m u is the number of sample data to be classified, L u is the loss function of the extreme learning machine, T ui is the true label of the i-th sample data, Te ui is the predicted label of the i-th sample data, λ u is the regularization parameter of the extreme learning machine, and || is the L2 norm, W u is the weight parameter of the extreme learning machine, β u is the nonlinear regularization coefficient of the extreme learning machine, sin 2 (W u ) is the sine square term for the weight parameter of the extreme learning machine, and Rz(t) is the adaptive error correction term of the t-th iteration.

[0046] S505, repeat the above steps until the preset stopping iteration condition is met, that is, the model training is completed.

[0047] The beneficial effects of this invention are as follows: The positioning method provided by this invention has relatively high positioning accuracy. This invention employs multiple sensors for data acquisition, obtaining multi-sensor data, which enables the acquisition of more accurate and comprehensive location data. In addition to the higher accuracy of the acquired real sensor data, this invention also uses a meta-learning-based generative adversarial network for data augmentation, optimizing the parameter update strategies of the generator and discriminator to generate various types of generated sensor data to enrich the sample data. This addresses the problems of high data acquisition costs and insufficient sample data volume in practical applications. Therefore, the sample database of this invention includes both real sensor data and generated sensor data, resulting in a high sample data volume with relatively little change in acquisition costs, improving the model's generalization ability and thus enhancing the model's accuracy. Meanwhile, this invention employs a neural network algorithm optimized based on wave energy capture for feature extraction, combined with an adaptive masking layer to dynamically adjust the masking layer in the network, focusing on features that significantly impact the positioning of the engineering inspection robot. This improves the effectiveness of feature extraction and the model's generalization ability, enhancing the stability and accuracy of feature extraction. Furthermore, it uses an autoencoder based on a dynamically adjusted matrix for feature dimensionality reduction, strengthening the focus on important features and reducing the learning of noisy or irrelevant features. This allows the model to better extract key information, reducing interference from irrelevant factors and improving the efficiency and accuracy of the data dimensionality reduction process. Finally, it uses an extreme learning machine algorithm based on adaptive error correction terms for classification, which optimizes the weights and biases of the extreme learning machine model in real time during training, making the model's fit to the training data more accurate while avoiding overfitting and improving classification accuracy. Through optimization of data processing in feature extraction, feature data dimensionality reduction, and classification, the invention ultimately achieves a relatively accurate positioning method. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] like Figure 1 As shown, the engineering inspection robot localization method based on multi-sensor data fusion provided by the present invention includes the following steps:

[0052] S1. Multi-sensor data acquisition and annotation:

[0053] Real sensor data is collected by lidar, acoustic wave sensor, inertial measurement unit (IMU), temperature and humidity sensor, and GPS sensor. The real sensor data is manually classified and labeled to obtain labeled real sensor data.

[0054] The real sensor data contains data collected by multiple sensors, and is a kind of multi-sensor data.

[0055] The labeled categories include: machine room area, office area, production workshop area, warehouse area;

[0056] The data collected in this embodiment is stored in a structured JSON format, but in actual application, the number of data attributes is usually more than 10, and the number of data attributes may reach dozens or even hundreds. Users can select other data formats and types according to their needs.

[0057] The data collected by the laser radar includes: distance data of the laser radar, unit: meter; angle data of the laser radar, unit: degree;

[0058] The data collected by the sound wave sensor includes: echo time of the sound wave sensor, unit: second; intensity of the sound wave sensor;

[0059] The data collected by the inertial measurement unit includes: acceleration data of the IMU, unit: meter / s2; angular velocity data of the IMU, unit: degree / s;

[0060] The data collected by the temperature and humidity sensor includes: temperature data, unit: Celsius; humidity data, unit: %;

[0061] The data collected by the GPS sensor includes: GPS position data, represented as latitude and longitude.

[0062] Because multiple sensors are used for data collection, the above data are all assigned time stamps when they are constructed.

[0063] S2, multi-sensor data expansion:

[0064] According to the labeled real sensor data, a generative sensor data is generated by a meta-learning-based generative adversarial network algorithm, and the labeled real sensor data and the generative sensor data constitute sample data;

[0065] The meta-learning-based generative adversarial network includes a generator and a discriminator, the generator is used to generate new generative sensor data, and the discriminator is used to evaluate the similarity between the generative sensor data and the real sensor data. The meta-learning-based generative adversarial network adopts a meta-learning strategy, dynamically adjusts the learning rate of the generator and the discriminator through a learning optimizer, adjusts the parameter update strategy based on real-time feedback, so that the meta-learning-based generative adversarial network algorithm can adapt to the diversity and complexity of data more quickly, and the quality and diversity of the generative sensor data are improved;

[0066] S3, training feature extraction model:

[0067] The feature extraction model is established based on the neural network algorithm optimized for sea wave energy capture, the sample data is input into the feature extraction model for feature extraction of multi-sensor data, the feature extraction model is trained based on the neural network algorithm optimized for sea wave energy capture, and the trained feature extraction model and the sample data after feature extraction are obtained.

[0068] The neural network algorithm optimized for sea wave energy capture dynamically optimizes the weights of the neural network by simulating the dynamic behavior of sea waves, utilizes the momentum characteristics of sea waves including the changes of velocity and acceleration, and combines the first-order and second-order damping behaviors to optimize the update strategy of the neural network weight parameters, that is, the weight update amount of the neural network depends on both the current energy gradient and the historical updated momentum, and specifically, the first-order damping term (the influence of velocity) and the second-order damping term (the influence of acceleration) are adopted to balance the weight update strength, thereby improving the stability and precision of neural network training.

[0069] S4, training a feature dimension reduction model:

[0070] The feature dimension reduction model is established based on the autoencoder algorithm with a dynamically adjusted matrix, the sample data after feature extraction is input into the feature dimension reduction model for feature dimension reduction of multi-sensor data, the feature dimension reduction model is trained based on the autoencoder algorithm with a dynamically adjusted matrix, and the trained feature dimension reduction model and the sample data after dimension reduction are obtained.

[0071] The autoencoder algorithm with a dynamically adjusted matrix updates the weights and biases by monitoring the activation of the hidden layer features through the dynamically adjusted matrix, thereby strengthening the attention to important features and weakening the learning of noise or irrelevant features, so that the model can better extract key information and reduce the interference of irrelevant factors on the results.

[0072] S5, training a classifier:

[0073] The classifier model is established based on the extreme learning machine algorithm with an adaptive error correction term, the sample data after dimension reduction is input into the classifier for classification of multi-sensor data, the classifier is trained based on the extreme learning machine algorithm with an adaptive error correction term, and the trained classifier is obtained.

[0074] The extreme learning machine algorithm with an adaptive error correction term dynamically adjusts the error between the model output and the true label through the adaptive error correction term, which can optimize the weights and biases of the extreme learning machine model in real time during the training process, so that the model fitting of the training data is more accurate, and the overfitting problem is avoided.

[0075] S6, engineering inspection robot positioning:

[0076] The new real sensor data is collected by at least two different types of sensors, and is input into the trained classifier to obtain the position classification of the engineering inspection robot in the position environment, so that the user can know the specific area where the engineering inspection robot is located, thereby realizing the position positioning of the engineering inspection robot. If the classification result is more accurate, the new real sensor data can be input into the trained feature extraction and feature dimension reduction model for feature processing, and the processed features are input into the classifier to obtain the classification result. The classification categories include: machine room area, office area, production workshop area, and warehouse area.

[0077] Example Two

[0078] The overall steps are the same as those in Example One, and the difference lies in that the training steps of the generative adversarial network algorithm based on meta-learning in S2 are as follows:

[0079] S201, initializing generator parameters and discriminator parameters The initialization method is as follows:

[0080]

[0081] In the formula, is the variance of the initialization parameter, represents a normal distribution with a mean of 0 and a standard deviation of .

[0082] A batch of data is randomly selected from the labeled real sensor data set as the training basis, and the selected batch data is denoted as X c,batch , wherein x c,i is the i-th training sample data in the selected batch training data, D c represents the discriminator, N c is the number of batch data.

[0083] S202, the generator receives a random noise vector z c , generates corresponding new generated sensor data, denoted as:

[0084]

[0085] In the formula, G c () is the generator function, is the generator parameter, including the generator weight and the generator bias

[0086] The calculation method of the weight of the generator is as follows:

[0087]

[0088] wherein, Sig() is a Sigmoid activation function, K is the number of basis functions, and the basis functions refer to mixed Gaussian kernel functions with different parameters; z c is an input noise vector, α ck is the weight of the kth kernel function, μ ck is the mean of the kth kernel function, σ ck is the standard deviation of the kth kernel function.

[0089] The discriminator evaluates the difference between the generated sensor data and the real sensor data batch, and outputs a value between 0 and 1, indicating the probability that the input data belongs to the real sensor data, and the evaluation method is represented as:

[0090]

[0091] wherein, D c () is a discriminator function; x represents the independent variable of the function, which represents real sensor data or generated sensor data, is a discriminator parameter, and Sig() is a Sigmoid activation function; α c is the learning rate of the discriminator; and are the weight and bias of the discriminator, respectively;

[0092] The weight of the discriminator is calculated in an adaptive weight adjustment manner, and the calculation method is:

[0093]

[0094] wherein, β c is the first discriminator adjustment parameter, β c is set to 0.1; γ c is the second discriminator adjustment parameter, γ c is set to 2, and tanh() is a hyperbolic tangent activation function; V c is a pre-posed weight matrix, x represents the input of the function, and u c is a pre-posed bias term.

[0095] S203, according to the feedback provided by the discriminator, the meta-learning optimizer is used to adjust the parameter update strategy of the generator and the discriminator, and the adjustment method is represented as:

[0096]

[0097] wherein, is the parameter adjusted based on meta-learning, including the generator parameter adjusted based on meta-learning θ θ c is the current parameter, including generator parameter θ λ c is the meta learning rate, is the loss function of the generative adversarial network about θ c , D c is the discriminator, G c is the generator;

[0098] loss function is calculated as follows:

[0099]

[0100] In the formula, N c is the number of batch data, y ci is the i-th real sensor data label, D c is the discriminator, x ci is the i-th sample data input to the discriminator, G c is the generator, z ci is the i-th noise vector input to the generator; i is the sample number index processed by the current batch of generative adversarial network, λ c is the regularization parameter, λ c is set in an adaptive adjustment manner, and its initial value is set to 3; ∥θ c ∥ 2 is the square of the L2 regularization term of the parameter θ c .

[0101] The adaptive adjustment manner of the regularization parameter λ c is represented as:

[0102]

[0103] In the formula, is the adaptive adjustment regularization parameter, grad prev is the gradient of the last step, ∥∥ is the L2 norm, λ adjust is the latent variable adjustment coefficient, is the latent variable loss function, loss function about the gradient of the parameter θ c ; ∈r is a small amount to prevent the denominator from being zero, and ∈r is set to 0.001.

[0104] The calculation manner of the latent variable adjustment coefficient λ adjust is:

[0105]

[0106] where Var(h c ) denotes the variance of the latent variable, σ threshold is a preset threshold, and σ threshold is set to 0.5.

[0107] A dynamic loss adjustment mechanism based on latent variable regression can be adopted to dynamically adjust the loss according to the internal structure of the data, so as to more finely control the model training process and improve the quality of the generated sensor data and the training efficiency of the model. The latent variable h c is defined to capture the features or attributes not explicitly expressed in the training data. The latent variable is generated by an independent network layer H c , and the network layer H c takes the actual data x real and the generated sensor data G c (z c ) as input. The calculation of the latent variable h c is as follows:

[0108] h c = H c (x real , G c (z c ) ; φ c ),

[0109] where H c () is the latent variable network layer function, G c () is the generator function, z c is the noise vector input to the generator, and φ c denotes the parameters of the latent variable generation layer.

[0110] The regression target ye c is calculated according to the latent variable h c , which represents the internal quality of the data sample data, and the calculation of the latent variable loss function L is as follows:

[0111]

[0112] where N c is the number of batch data, y ci is the i-th real sensor data label, and ye ci is the i-th label predicted based on the latent variable.

[0113] If the latent variable network layer function adopts a Sigmoid function as the activation function, the calculation of the above-mentioned label predicted based on the latent variable ye ci is as follows:

[0114]

[0115] In the formula, Sig() is the Sigmoid activation function. and These represent the weights and biases of the regression layer in the latent variable network, h. ci Intermediate features for generating data for the i-th generator.

[0116] S204. Adjust the generator parameters based on meta-learning according to the gradient descent strategy. With discriminator parameters To perform another update, the update method is as follows:

[0117]

[0118] In the formula, ← represents the parameter update symbol; γ c The learning rate adjusted for meta-learning, γ c Set the value to 0.01;

[0119] The generator parameters for the next iteration. For loss function Regarding the generator parameters adjusted based on meta-learning gradient, The generator loss function; The discriminator parameters for the next iteration. For loss function Regarding the discriminant parameters adjusted based on meta-learning gradient, The loss function of the discriminator;

[0120] S205. Repeat the above steps until the preset stopping iteration condition is met, which means that the feature model training is complete. The preset maximum number of iterations is set to 1000.

[0121]

Example 3

[0122] The overall steps are the same as in Example 1, except that the training steps in S3 for training the feature extraction model using a neural network algorithm optimized based on wave energy capture are as follows:

[0123] S301. Initialize the weight parameters of the neural network. The initialization method is as follows:

[0124]

[0125] In the formula, Let be the weights of the l-th layer of the neural network, where l is the layer index of the neural network; This represents the number of nodes in the (l-1)th layer of the neural network. represents a normal distribution with mean 0 and standard deviation 1.

[0126] S302, an adaptive mask layer is set at the front 3 layers of the network, the mask layer in the network is dynamically adjusted through the adaptive mask enhancement strategy, more attention is paid to the features that have greater influence on the different positioning of the engineering patrol robot in the training process, thereby improving the accuracy of feature extraction and the generalization ability of the model. The adaptive mask layer dynamically adjusts its transmittance according to the characteristics of the input data, and the calculation method is:

[0127] M p =Sig(β p ·X p +b p ),

[0128] In the formula, M p is the output of the mask layer, Sig() is the Sigmoid activation function, β p is the learnable parameter of the mask layer, and the adaptive learning strategy is used to optimize the β p parameter update; X p is the input feature, and b p is the bias term.

[0129] The gradient-based method is used to dynamically adjust β p , and the calculation method of the update amount Δβ p of the parameter β p is:

[0130]

[0131] In the formula, is the learning rate of β p , and is set to 0.01; is the cross-entropy loss L p based on one-hot encoding. The partial derivative of β p .

[0132] The partial derivative of the cross-entropy loss L p based on one-hot encoding with respect to β p is calculated through the chain rule and is expressed as:

[0133]

[0134] In the formula, L p is the cross-entropy loss based on one-hot encoding, M p is the output of the mask layer, and β p is the learnable parameter of the mask layer.

[0135] S303, the dynamic behavior of the sea wave is simulated to optimize the weight of the neural network, and the calculation method of the update amount of the weight of the neural network is:

[0136]

[0137] In the formula, ΔW p is the update amount of the weight of the neural network; η p is the learning rate of the neural network, η p is set to 0.01; is the energy gradient corresponding to the current weight of the neural network; α p is the momentum factor of the neural network, α p is set to 0.1; is the update amount of the weight of the neural network in the last iteration;

[0138] The calculation of the energy gradient not only depends on the current weight position, but also depends on the historical speed and acceleration of the weight change, and a second-order dynamic element is used to simulate this physical behavior, and the calculation method is:

[0139]

[0140] In the formula, γ p represents a first-order damping coefficient, representing the direct influence of the speed of weight change on energy; μ p represents a second-order damping coefficient, representing the acceleration of weight change; and respectively represent the first and second derivatives of the weight; dt is the time interval for derivation, and the time interval is the interval of the iteration number.

[0141] The first and second derivatives are calculated in the form of discrete time difference, and the calculation method is:

[0142]

[0143] In the formula, Δt is the interval of the iteration number, W p (t) is the weight of the neural network in the tth iteration, t is the iteration number, W p (t-Δt) is the difference value of the weight of the neural network between the tth iteration and the iteration interval of Δt,

[0144] W p (t-2Δt) is the difference value of the weight of the neural network between the tth iteration and the iteration interval of 2Δt.

[0145] S304, the input data passes through the neural network, adopts the forward propagation mode, and generates an output after 6 layers of processing; the loss is calculated according to the actual output and the expected output, and the positioning method adopts a cross-entropy loss function based on one-hot encoding to measure the performance of the model, and the output Y of the forward propagation is set as p The cumulative interlayer transmission calculation is calculated in the following manner:

[0146]

[0147] In the formula, is the output of the lth layer of the neural network, Re() is the ReLU activation function, is the weight of the lth layer of the neural network, is the output of the (l-1)th layer of the neural network,

[0148] is the output of the adaptive mask layer of the lth layer of the neural network; ⊙ represents element multiplication, that is, Hadamard multiplication, which is used to apply the adaptive mask; is the bias term of the lth layer.

[0149] The cross-entropy loss function L based on one-hot encoding p is calculated in the following manner:

[0150]

[0151] In the formula, L p is the cross-entropy loss function based on one-hot encoding, K pcv is the number of samples input to the neural network in the current batch, is the one-hot encoding of the target label of the k p th sample, is the probability predicted by the model for the k p th sample, which is calculated by the Softmax function.

[0152] S305, repeat the above steps until the preset stopping iteration condition is met, that is, the feature extraction model training is completed, and the preset maximum iteration number is set to 1000 times.

[0153]

Example Four

[0154] The overall steps are the same as in Example One, except that in S4, the feature dimension reduction model training is performed by the autoencoder algorithm based on the dynamically adjusted matrix, and the training steps are as follows:

[0155] S401, initialize the parameters of the autoencoder algorithm based on the dynamically adjusted matrix, including the weights and biases of the autoencoder, and the initialization mode is represented as:

[0156]

[0157] wherein, is the initial value of the weight of the autoencoder, n r is the number of input layer nodes of the autoencoder, μ r and are the mean and variance of the batch data input to the autoencoder, respectively, is the initial value of the bias of the autoencoder, initialized to zero.

[0158] S402, the data is transmitted layer by layer through the encoder layer by means of forward propagation, and the output of the encoder is represented as:

[0159] h r = Re(W r ·X r +b r ),

[0160] wherein, h r is the encoded hidden layer feature, Re() is the ReLU activation function, W r is the weight parameter of the autoencoder, X r is the input data, and b r is the bias parameter of the autoencoder;

[0161] The low-dimensional feature obtained by the encoder is input into the decoder to reconstruct the original data, and the output of the decoder is represented as:

[0162]

[0163] wherein, Xe is the reconstructed output data, represents the transpose of the weight parameter of the autoencoder, h r represents the encoded hidden layer feature, and b r represents the bias parameter of the autoencoder.

[0164] The transpose of the weight parameter of the autoencoder can be calculated by a dynamic adjustment strategy:

[0165]

[0166] wherein, W r is the weight parameter of the autoencoder, and D r is the dynamic adjustment matrix.

[0167] The elements d r of the dynamic adjustment matrix D rij are adjusted according to the activation of the hidden layer feature h r , and the calculation method is:

[0168]

[0169] In the formula, For dynamically adjusting matrix D r The i-th c Line j c Column elements; and They are respectively The adjustment intensity and threshold of the element, For the j-th c Features of hidden layers, i c To dynamically adjust the matrix row indices; j c To dynamically adjust the matrix column indexes.

[0170] threshold The settings are configured using adaptive calculation, and the calculation method is as follows:

[0171]

[0172] In the formula, ξ r ξ is an adjustable coefficient. r Set to 3; Median() is the median function, h rj This is the feature encoded after the j-th hidden layer.

[0173] S403. Calculate the composite loss function based on data fidelity and coding sparsity control to evaluate model performance. The loss function L of the autoencoder... r The calculation method is as follows:

[0174]

[0175] In the formula, ∥∥ represents the L2 norm, X r λ is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder; r λ is the regularization coefficient of the autoencoder. r Set the value to 0.2; ||W r ∥ F Weight W r The Frobenius norm; ρ r The coefficient ρ is used to control the sparsity regularization strength. r Set to 2; n g The number of autoencoder layers is the same as the number of columns in the dynamic adjustment matrix; For indicator functions, when Greater than The value is 1 if it is true, and 0 otherwise.

[0176] Weight W r Frobenius norm It includes a non-linear penalty for each element of the weight matrix, calculated as follows:

[0177]

[0178] In the formula, For W r The i-th c Line j c Column elements; ∈ r For nonlinear penalty coefficients, ∈ r Set the value to 1.5.

[0179] Dynamic adjustment matrix D r elements The adjustment intensity is adjusted through dynamic routing, and the adjustment method is expressed as follows:

[0180]

[0181] Furthermore, according to the adjusted Update the parameters of the autoencoder, expressed as follows:

[0182]

[0183] In the formula, γ r For routing sensitivity parameters, Candidate weights from input to output, The updated set of autoencoder weight parameters; Softmax() is the Softmax function. for Adjustment intensity of elements, For W r The i-th c Line j c Column elements.

[0184] Based on the updated autoencoder weight parameter set The bias parameters of the autoencoder are updated in the same way. The weights W of the autoencoder are updated again using gradient descent. r The update method is represented as follows:

[0185]

[0186] In the formula, η r Let η be the learning rate for gradient descent. r Set to 0.01; The loss function is relative to the weights. The gradient.

[0187] S404. Repeat the above steps until the preset stopping iteration condition is met, which means that the feature dimensionality reduction model training is complete. The preset maximum number of iterations is set to 1000.

[0188] Example Five

[0189] The overall steps are the same as Example One, except that the training step of the classifier training in S5 is performed by an extreme learning machine algorithm based on adaptive error correction terms, as follows:

[0190] S501, initialize the extreme learning machine weight parameter W u S502, normalize the low-dimensional feature vector after dimension reduction, denoted as: u The initialization method is represented as:

[0191]

[0192] In the formula, n u is the number of neurons in the input layer, V u is a random matrix drawn from a standard normal distribution, and Θ u is an angle matrix.

[0193] The angle matrix Θ u is randomly generated from a uniform distribution from 0 to π, and its calculation method is:

[0194] Θ u = π·Rand(0,1,size(W u )),

[0195] In the formula, π is the ratio of a circle, Rand() is a random generation function, size() is a size calculation function, and W u is the extreme learning machine weight parameter.

[0196] S502, normalize the low-dimensional feature vector after dimension reduction, denoted as:

[0197]

[0198] In the formula, X' u is the normalized input feature, X u is the low-dimensional feature vector after dimension reduction, μ u and σ u are the mean and standard deviation of X u , respectively;

[0199] S503, input the normalized feature into the extreme learning machine classifier, and the process of the data forward propagation is represented as:

[0200] H u = Sig(W u ·X' u +b u ),

[0201] In the formula, H uis the output of the activated extreme learning machine, Sig() is the Sigmoid activation function, W u represents the weight parameter of the extreme learning machine, X u is the normalized input feature, b u is the bias parameter of the extreme learning machine;

[0202] The weight parameter W of the extreme learning machine u is adjusted in the training process by using adaptive gradient manifold exploration. The calculation method of the increment of gradient adjustment is:

[0203]

[0204] In the formula, ΔW u is the update amount of the weight of the extreme learning machine; η u is the learning rate of the extreme learning machine, η u is set to 0.01; is the gradient of the loss function of the extreme learning machine with respect to the weight; α u is the learning rate coefficient of the weight adjustment of the extreme learning machine, α u is set to 3; is the gradient of the manifold exploration function with respect to the weight of the extreme learning machine; R() is the manifold exploration function; κ u is the manifold exploration sensitivity, κ u is set to 1; is the weight parameter of the i d th neuron of the j d th neural network layer, τ u is the manifold exploration threshold, Lay is the number of layers of the extreme learning machine, i d is the layer index of the extreme learning machine, Lec is the number of layer neurons of the extreme learning machine, j d is the layer neuron index of the extreme learning machine, τ u is set to 1.5.

[0205] S504, a loss function based on an adaptive error correction term is used for loss calculation and back propagation to ensure the generalization ability of the model. The calculation method of the loss function L u is:

[0206]

[0207] In the formula, m u is the number of sample data to be classified, L u is the loss function of the extreme learning machine, is the true label of the i e th sample data, i e is the sample index input to the extreme learning machine, For the i-th e Predicted labels for each sample data; λ u Let λ be the regularization parameter of the extreme learning machine. u Set to 0.4; ∥∥ is the L2 norm, W u For the weight parameters of the Extreme Learning Machine; β u β is the nonlinear regularization coefficient of the extreme learning machine. u Set to 0.6; i is the i-th neural network layer, j is the j-th neuron, sin 2 (W u Rz(t) is the squared sine term for the weight parameters of the extreme learning machine, and Rz(t) is the adaptive error correction term for the t-th iteration.

[0208] The parameters of the Extreme Learning Machine are updated, and the update method is expressed as follows:

[0209]

[0210] In the formula, W u b represents the weight parameters of the Extreme Learning Machine. u η represents the bias parameters of the extreme learning machine. u Let η be the learning rate. u Set to 0.01; and This is the gradient of the loss function with respect to the weights and biases.

[0211] An adaptive error correction term with dynamic threshold adjustment is used to adjust the network output in real time to reduce prediction error. The calculation method for the dynamic threshold in the t-th iteration is defined as follows:

[0212]

[0213] In the formula, δ u (t) represents the dynamic threshold in the t-th iteration. and They are respectively the i-th time in the t-th iteration e The true labels of n sample data and the output labels of the model, where || represents the absolute value, n u This represents the number of neurons in the input layer.

[0214] At the t-th iteration, the i-th e The output label of the model for each sample data is calculated using adaptive error correction, which is expressed as follows:

[0215] ye' u (t)=ye u (t)+γ u (t)·(δ u (t)-|ye u (t)-yu (t),

[0216] wherein ye u (t) is the output label of the model after adaptive error correction, γ u (t) is the learning rate adaptively adjusted based on the data distribution, ye u (t) is the output label of the model at the tth iteration, δ u (t) is the real label at the tth iteration, δ u (t) represents the dynamic threshold value at the tth iteration.

[0217] The adaptive error correction term Rz(t) at the tth iteration is calculated as follows:

[0218]

[0219] wherein ρ u is the weight of the adaptive error correction term, ρ u is set to 0.3; max() is the maximum value function, δ u is δ u (t) represents the dynamic threshold value at the tth iteration.

[0220] S505, repeat the above steps until the preset stopping iteration condition is met, that is, the model training is completed, and the preset maximum iteration number is set to 1000 times.

[0221] It should be noted that the present application is not limited to the specific structure shown in the above embodiments and the accompanying drawings, and various changes can be made within the knowledge possessed by those skilled in the art.

Claims

1. A method for positioning an engineering inspection robot based on multi-sensor data fusion, characterized in that, The method comprises the following steps: S1, collecting real sensor data through at least two different types of sensors, and classifying and labeling the real sensor data to obtain labeled real sensor data; S2, generating generated sensor data through a meta-learning-based generative adversarial network algorithm according to the labeled real sensor data, wherein the labeled real sensor data and the generated sensor data constitute sample data; The training steps of the meta-learning-based generative adversarial network algorithm are as follows: S201, initialize generator parameters with discriminator parameters S202、The generator receives a random noise vector z c , generates corresponding new generated sensor data; S203, adjusting the parameter update strategy of the generator and the discriminator by using a meta-learning optimizer according to the feedback provided by the discriminator, and the adjustment mode is represented as: In the formula, are parameters adjusted based on meta-learning. θ c is the current parameter; λ c is the meta learning rate, D c is the discriminator, G c is the generator; Loss function is calculated as follows: In the formula, N c is the number of batches of data, y ci is the i-th real sensor data label, D c is the discriminator, x ci is the i-th sample data input to the discriminator, G c is the generator, z ci is the i-th noise vector input to the generator; i is the sample number index processed by the current batch of generative adversarial networks, λ c is a regularization parameter; Regularization parameter λ c The adaptive adjustment mode of λ is represented as: where, is the regularization parameter after adaptive adjustment, grad prev is the gradient of the previous step, ||| is the L2 norm, λ adjust is the hidden variable adjustment coefficient, is the hidden variable loss function, loss function with respect to the parameters θ c is the gradient; ∈r is a small amount to prevent the denominator from being zero; Hidden variable adjustment coefficient λ adjust The calculation method is as follows: In the formula, Var(h c ) represents the variance of the hidden variable, σ threshold is a preset threshold value. S204, updating the generator parameters based on the meta-learning adjustment according to the gradient descent strategy with the discriminator parameters for further updating; S205, repeating the above steps until a preset stop iteration condition is met, that is, the feature model training is completed; S3, inputting the sample data into a feature extraction model, training the feature extraction model by using a neural network algorithm based on sea wave energy capture optimization, and obtaining a trained feature extraction model and feature-extracted sample data; S4, inputting the feature-extracted sample data into a feature dimension reduction model, training the feature dimension reduction model by using a self-encoder algorithm based on a dynamically adjusted matrix, and obtaining a trained feature dimension reduction model and dimension-reduced sample data; S5, inputting the dimension-reduced sample data into a classifier, training the classifier by using a limit learning machine algorithm based on an adaptive error correction term, and obtaining a trained classifier; S6, collecting new real sensor data through at least two different types of sensors, inputting the new real sensor data into the trained classifier, obtaining the position classification of the engineering patrol robot in the position environment, and realizing the position positioning of the engineering patrol robot.

2. The multi-sensor data fusion based engineering patrol robot positioning method according to claim 1, wherein, The labeled categories in S1 include machine room area, office area, production workshop area, and warehouse area. 3.The multi-sensor data fusion based engineering patrol robot positioning method according to claim 1, wherein, The training steps of the feature extraction model in S3 are as follows: S301, initializing the weight parameters of the neural network; S302, setting an adaptive mask layer in the first 3 layers of the network, wherein the adaptive mask layer dynamically adjusts its transmittance according to the characteristics of the input data; S303, optimizing the weights of the neural network by simulating the dynamic behavior of sea waves, and the calculation method of the update amount of the weights of the neural network is: where ΔW p is the update amount of the neural network weight; η p is the learning rate of the neural network; is the energy gradient corresponding to the current weight of the neural network; α p is the momentum factor of the neural network; is the update amount of the weight of the neural network in the last iteration. S304, input data passes through the neural network in a forward propagation manner and generates output after 6 layers of processing; S305, repeating S302-S304 until a preset stop iteration condition is met, that is, the feature extraction model training is completed.

4. The multi-sensor data fusion based engineering patrol robot positioning method according to claim 1, wherein, The training steps of the feature dimension reduction model in S4 are as follows: S401, initializing the parameters of the self-encoder algorithm based on the dynamically adjusted matrix, including the weights and biases of the self-encoder; S402, the data passes through the encoder layer layer by layer through forward propagation, and the encoder output is represented as: h r = Re(W r ·X r + b r ), where h r is the encoded hidden layer feature, Re() is the ReLU activation function, W r is the weight parameter of the autoencoder, X r is the input data, b r is the bias parameter of the autoencoder; S403、According to the data fidelity and the encoding sparsity control calculation compound loss function to evaluate the model performance, the loss function L of the autoencoder r The calculation method is: In the formula, |||| is the L2 norm, X r Xe is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder, and λ is the input data. r Let W be the regularization coefficient of the autoencoder. r || F Weight W r The Frobenius norm, ρ r The coefficient n is used to control the strength of sparsity regularization. g This represents the number of autoencoder layers. As an indicator function, when h rj Greater than θ rj The value is 1 if it is true, and 0 otherwise. S404, repeating S402-S403 until a preset stop iteration condition is met, that is, the feature dimension reduction model training is completed.

5. The multi-sensor data fusion based engineering tour robot positioning method according to claim 1, wherein, The training steps of the classifier in S5 are as follows: S501, limiting learning machine weight parameter W u Initialization is performed; S502, normalizing the low-dimensional feature vector after dimension reduction; S503, input the normalized features into the extreme learning machine classifier; S504, loss calculation and back propagation are performed using a loss function based on an adaptive error correction term to ensure the generalization ability of the model, and the loss function L u The calculation method is as follows: wherein m u is the number of sample data to be classified, L u is the loss function of extreme learning machine, T ui is the true label of the i-th sample data, Te ui is the predicted label of the i-th sample data, λ u is the regularization parameter of extreme learning machine, ||| u is the weight parameter of extreme learning machine, β u is the nonlinear regularization coefficient of extreme learning machine, sin 2 (W u ) is the sinusoidal square term for the weight parameter of extreme learning machine, Rz(t) is the adaptive error correction term of the t-th iteration; S505, repeat the iteration of the above steps until the preset stop iteration condition is met, that is, the model training is completed.

Images