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

Abstract

The invention relates to the field of data processing, in particular to an engineering patrol robot positioning method based on multi-sensor data fusion, and the method comprises the steps: collecting real sensor data through at least two different types of sensors, generating sensor data through a generative adversarial network algorithm based on meta-learning, a neural network algorithm based on sea wave energy capture optimization is adopted to train a feature extraction model, an auto-encoder algorithm based on a dynamic adjustment matrix is adopted to train a feature dimension reduction model, and an extreme learning machine algorithm based on an adaptive error correction term is adopted to train a classifier. New real sensor data are collected and input into the trained classifier, and the position classification of the engineering patrol robot in the position environment is obtained. An existing positioning method is not accurate in positioning accuracy due to the fact that the sample data size is small and the data processing capacity is weak, and the method has accurate positioning accuracy due to the fact that the sample data size is expanded and data processing is optimized.

Description

Technical Field

[0001] The present invention relates to the field of information monitoring of electrical digital data processing, and in particular to a positioning method of an engineering inspection robot based on multi-sensor data fusion. Background Art

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

[0003] Existing positioning methods usually include four steps: collecting data through sensors, extracting features from the collected data, reducing the dimension of the feature data, and classifying the reduced-dimensional data.

[0004] Collecting data is the basis for subsequent training data, but data acquisition, labeling and preprocessing are time-consuming and labor-intensive. Traditional engineering inspection robot positioning methods mostly rely on a single type of sensor for data collection, such as visual or radar signals, and use the real sensor data collected by the sensor as sample data. A single type of sensor may face problems such as signal interference under changing environmental conditions, resulting in low accuracy of collected data and a small amount of collected data. The problem of insufficient sample data will limit the training effect and generalization ability of the model, resulting in poor generalization ability of the model, thereby affecting the accuracy of the model and making it difficult for the training sample data to fully cover the data diversity of the application scenario.

[0005] With the increase in the number and types of sensors, the difficulty of data processing will increase, which will lead to problems in data processing and ultimately affect the positioning results. For example, the existing positioning methods for engineering inspection robots use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions. Therefore, the existing positioning methods often fail to fully and dynamically adjust the extraction strategy of key features, which may lead to insufficient learning of the model on specific data features, affecting the stability of training and the performance of the model, and thus affecting the final positioning accuracy. In terms of feature dimensionality reduction, the existing positioning methods fail to efficiently encode and reconstruct complex data, and the loss of information during the dimensionality reduction process may lead to a decline in the performance of the final model. In the process of classifying the reduced-dimensional data, the existing positioning methods may have difficulty capturing small data changes due to insufficient weight initialization and update strategies when processing complex or low-dimensional data, which will limit the classification accuracy and response sensitivity, thereby affecting the positioning accuracy of the model.

[0006] Therefore, the existing positioning method has the problem of relatively inaccurate positioning accuracy. Summary of the invention

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

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

[0009] S1. Collect real sensor data through at least two different types of sensors, classify and annotate the real sensor data, and obtain annotated real sensor data;

[0010] S2. Generate sensor data according to the annotated real sensor data by using a generative adversarial network algorithm based on meta-learning, wherein the annotated real sensor data and the generated sensor data constitute sample data;

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

[0012] S4, inputting the sample data after feature extraction into a feature dimensionality reduction model, and training the feature dimensionality reduction model using an autoencoder algorithm based on a dynamic adjustment matrix to obtain a trained feature dimensionality reduction model and sample data after dimensionality reduction;

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

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

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

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

[0017] S201, initialize generator parameters and the discriminator parameters

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

[0019] S203, according to the feedback provided by the discriminator, using a meta-learning optimizer 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 meta-learning and the discriminator parameters Update again;

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

[0022] As a further improvement of the present invention: the training steps of the feature extraction model in S3 are as follows:

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

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

[0025] S303, simulating the dynamic behavior of the waves to optimize the weight of the neural network, the calculation method of the update amount of the weight of the neural network 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 neural network in the last iteration;

[0028] S304, input data is passed through the neural network, using the forward propagation method, and output is generated after 6 layers of processing;

[0029] S305, repeating S302 to S304 until a preset condition for stopping iteration is met, which means that the feature extraction model training is completed.

[0030] As a further improvement of the present invention: the training steps of the feature dimensionality reduction model training in S4 are as follows:

[0031] S401, initializing the parameters of the autoencoder algorithm based on the dynamic adjustment matrix, including the weight and bias of the autoencoder;

[0032] S402, the data is transmitted layer by layer through the encoder layer by forward propagation, and the encoder output is expressed as:

[0033] h r =Re(W r ·X r +b r ),

[0034] In the formula, 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;

[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:

[0036]

[0037] In the formula, ∥∥ is the L2 norm, X r is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder, λ r is the regularization coefficient of the autoencoder, ∥W r ∥ F is the weight W r The Frobenius norm, ρ r is the coefficient that controls the strength of sparse regularization; n g is the number of autoencoder layers; is the indicator function, when h rj Greater than θ rj The value is 1 when it is, otherwise it is 0;

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

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

[0040] S501, extreme learning machine weight parameter W u Initialize;

[0041] S502, normalizing the low-dimensional feature vector after dimensionality reduction;

[0042] S503, inputting the normalized features into an extreme learning machine classifier;

[0043] S504: Use the loss function based on the adaptive error correction term to perform loss calculation and back propagation to ensure the generalization ability of the model. The loss function L u The calculation method is:

[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, ∥∥ is the L2 norm, W u is the extreme learning machine weight parameter, β u is the nonlinear regularization coefficient of the extreme learning machine, sin 2 (W u ) is the sine square term for the extreme learning machine weight parameter, and Rz(t) is the adaptive error correction term for the tth iteration;

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

[0047] The beneficial effects of the present invention are as follows: The positioning accuracy of the positioning method provided by the present invention is relatively accurate. The present invention uses multiple sensors for data collection, obtains multi-sensor data, and can collect more accurate and comprehensive location data. In addition to the more accurate real sensor data collected, the present invention also uses a generative adversarial network based on meta-learning for data expansion, optimizes the parameter update strategy of the generator and the discriminator, to generate a variety of generated sensor data to enrich the sample data, in order to cope with the problems of high data collection cost and insufficient sample data in practical applications, so the sample database of the present invention includes real sensor data and generated sensor data, the sample data volume is relatively high, and the collection cost is relatively unchanged, which improves the generalization ability of the model, thereby improving the accuracy of the model. At the same time, the present invention adopts a neural network algorithm based on wave energy capture optimization for feature extraction, combines an adaptive mask layer to dynamically adjust the mask layer in the network, focuses on features that have a greater impact on the positioning of the engineering inspection robot, so as to improve the effectiveness of feature extraction and the generalization ability of the model, and enhances the stability and accuracy of feature extraction; adopts an autoencoder based on a dynamically adjusted matrix for feature dimensionality reduction, strengthens the focus on important features, and weakens the learning of noise or irrelevant features, so that the model can better extract key information, reduce the interference of irrelevant factors on the results, and improve the efficiency and accuracy of the data dimensionality reduction process; adopts an extreme learning machine algorithm based on adaptive error correction terms for classification, which can optimize the weights and biases of the extreme learning machine model in real time during the training process, so that the model fits the training data more accurately, while avoiding overfitting problems, and improving classification accuracy. By optimizing data processing in three aspects: feature extraction, feature data dimensionality reduction, and classification, the purpose of making the positioning accuracy of the positioning method more accurate is finally achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0050] [Example 1]

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

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

[0053] Collect real sensor data through lidar, acoustic wave sensor, inertial measurement unit IMU, temperature and humidity sensor and GPS sensor, manually classify and label the real sensor data, and obtain the labeled real sensor data;

[0054] Since real sensor data includes data collected by multiple sensors, it is a kind of multi-sensor data.

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

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

[0057] The data collected by the LiDAR include: the distance data of the LiDAR, in meters; the angle data of the LiDAR, in degrees;

[0058] The data collected by the acoustic sensor include: the echo time of the acoustic sensor, in seconds; the intensity of the acoustic sensor;

[0059] The data collected by the inertial measurement unit include: IMU acceleration data, in meters per second2; IMU angular velocity data, in degrees per second;

[0060] The data collected by the temperature and humidity sensor include: temperature data, in degrees Celsius; humidity data, in %;

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

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

[0063] S2. Multi-sensor data expansion:

[0064] Generate sensor data according to the annotated real sensor data by using a generative adversarial network algorithm based on meta-learning, wherein the annotated real sensor data and the generated sensor data constitute sample data;

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

[0066] S3. Train feature extraction model:

[0067] Establishing a feature extraction model based on a neural network algorithm optimized for wave energy capture, inputting the sample data into the feature extraction model for extracting features from multi-sensor data, training the feature extraction model using a neural network algorithm optimized for wave energy capture, and obtaining a trained feature extraction model and sample data after feature extraction;

[0068] The neural network algorithm based on wave energy capture optimization dynamically optimizes the weights of the neural network by simulating the dynamic behavior of waves, and utilizes the momentum characteristics of waves, including changes in velocity and acceleration, combined with 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 momentum of historical updates. Specifically, the first-order damping term (the influence of velocity) and the second-order damping term (the influence of acceleration) are used to balance the weight update intensity, thereby improving the stability and accuracy of neural network training.

[0069] S4. Training feature dimensionality reduction model:

[0070] A feature dimensionality reduction model is established based on an autoencoder algorithm of a dynamic adjustment matrix, the sample data after feature extraction is input into the feature dimensionality reduction model for performing feature dimensionality reduction on multi-sensor data, and the feature dimensionality reduction model is trained using an autoencoder algorithm based on a dynamic adjustment matrix to obtain a trained feature dimensionality reduction model and sample data after dimensionality reduction;

[0071] The autoencoder algorithm based on the dynamic adjustment matrix uses the dynamic adjustment matrix to update the weights and biases by monitoring the activation of hidden layer features, thereby strengthening the focus on 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 classifier:

[0073] Establishing a classifier model based on an extreme learning machine algorithm with an adaptive error correction term, inputting the sample data after dimensionality reduction into a classifier for classifying multi-sensor data, and training the classifier with an extreme learning machine algorithm based on an adaptive error correction term to obtain a trained classifier;

[0074] The extreme learning machine algorithm based on adaptive error correction terms uses adaptive error correction terms to dynamically adjust the error between the model output and the true label. It can optimize the weights and biases of the extreme learning machine model in real time during the training process, making the model fit the training data more accurately while avoiding overfitting problems.

[0075] S6. Engineering inspection robot positioning:

[0076] New real sensor data is collected through at least two different types of sensors, and the new real sensor data 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 you want to make the classification result more accurate, you can input the new real sensor data into the trained feature extraction and feature dimensionality reduction model for feature processing, and input the processed features 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 2]

[0078] The overall steps are the same as those in Example 1, except that the training steps of the generative adversarial network algorithm based on meta-learning in S2 are as follows:

[0079] S201, initialize generator parameters and the discriminator parameters The initialization method is:

[0080]

[0081] In the formula, is the variance of the initialization parameters, The mean is 0 and the standard deviation is The normal distribution of .

[0082] Randomly select a batch of data from the labeled real sensor data set as the training basis, assuming that the selected batch of data is X c,batch , where x c,i is the i-th training sample data in the selected batch training data, D c represents the discriminator, N c The number of batches of data.

[0083] S202, the generator receives a random noise vector z c , generating the corresponding newly generated sensor data, expressed as:

[0084]

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

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

[0087]

[0088] Where Sig( ) is the Sigmoid activation function, Kr is the number of basis functions, and basis functions refer to mixed Gaussian kernel functions with different parameters; c is the 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 batches, and outputs a value between 0 and 1, indicating the probability that the input data belongs to the real sensor data. The evaluation method is expressed as:

[0090]

[0091] Where D c () is the discriminator function; x represents the independent variable of the function, which represents the real sensor data or the generated sensor data. is the discriminator parameter, Sig( ) is the 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 The calculation is performed by adaptive weight adjustment, and the calculation method is:

[0093]

[0094] In the formula, β c Adjust the parameters for the first discriminator, β c Set the value to 0.1; c Adjust the parameters for the second discriminator, γ c Set the value to 2, tanh() is the hyperbolic tangent activation function, V c is the pre-weight matrix, x represents the input of the function, u c is the pre-bias term.

[0095] S203. According to the feedback provided by the discriminator, the parameter update strategy of the generator and the discriminator is adjusted using the meta-learning optimizer. The adjustment method is expressed as:

[0096]

[0097] In the formula, are the parameters adjusted based on meta-learning, including the generator parameters adjusted based on meta-learning and the discriminator parameters θ c is the current parameters, including generator parameters and the discriminator parameters λ c is the meta-learning rate, is the loss function for generating adversarial networks About θ c The gradient of c is the discriminator, G c For the generator;

[0098] Loss Function The calculation method is:

[0099]

[0100] Where 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 index of the number of samples processed by the current batch of generative adversarial networks, λ c is the regularization parameter, λ c Adopt adaptive adjustment mode to set, and its initial value is set to 3; ∥θ c ∥ 2 is the parameter θ c The square of the L2 regularization term.

[0101] Regularization parameter λ c The adaptive adjustment method is expressed as:

[0102]

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

[0104] Hidden variable adjustment coefficient λ adjust The calculation method is:

[0105]

[0106] In the formula, Var(h c ) represents the variance of the latent variable, σ threshold is a preset threshold, σ threshold Set the value to 0.5.

[0107] A dynamic loss adjustment mechanism based on latent variable regression can be used to dynamically adjust the loss according to the intrinsic structure of the data in order to more finely control the model training process, thereby improving the quality of generated sensor data and the training efficiency of the model. Define the latent variable h c It is used to capture features or attributes that are not explicitly expressed in the training data. The latent variables are represented by an independent network layer H. c Generate, network layer H c Take the actual data x real and generate sensor data G c (z c ) is the input, and the hidden variable h c The calculation method is:

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

[0109] In the formula, H c () is the hidden variable network layer function, G c () is the generator function, z c is the noise vector input to the generator, φ c Represents the parameters of the hidden variable generation layer.

[0110] According to the latent variable h c Calculate the regression target c , which characterizes the intrinsic quality of the data sample data, then the latent variable loss function The calculation method is:

[0111]

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

[0113] The hidden variable network layer function uses the Sigmoid function as the activation function, so the label ye based on the hidden variable prediction ci The calculation method is:

[0114]

[0115] Where Sig( ) is the Sigmoid activation function, and are the weight and bias of the regression layer in the latent variable network layer, h ci Intermediate features of data generated by the i-th generator.

[0116] S204: According to the gradient descent strategy, the generator parameters adjusted based on meta-learning and the discriminator parameters Update again, the update method is expressed as:

[0117]

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

[0119] is the generator parameter for the next iteration, is the loss function About the generator parameters adjusted based on meta-learning The gradient of is the generator loss function; is the discriminator parameter for the next iteration, is the loss function About the discriminator parameters adjusted based on meta-learning The gradient of is the loss function of the discriminator;

[0120] S205, repeat the above steps until the preset stop iteration condition is met, which means that the feature model training is completed, and the preset maximum number of iterations is set to 1000 times.

[0121] [Example 3]

[0122] The overall steps are the same as those in the first embodiment, except that the training steps of the feature extraction model in S3 by 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:

[0124]

[0125] In the formula, is the weight of the lth layer of the neural network, and l is the layer index of the neural network; is the number of nodes in the l-1th layer of the neural network; Represents a normal distribution with mean 0 and standard deviation 1.

[0126] S302, set an adaptive mask layer in the first three layers of the network, dynamically adjust the mask layer in the network through the adaptive mask enhancement strategy, and pay more attention to the features that have a greater impact on the different positioning of the engineering inspection robot during 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] Where 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 an adaptive learning strategy is used to optimize β p Parameter update; X p is the input feature, b p is the bias term;

[0129] Dynamically adjust β using a gradient-based method p , parameter β p The update amount Δβ p The calculation method is:

[0130]

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

[0132] Cross entropy loss L based on one-hot encoding p β p The partial derivative of Calculated by the chain rule, it is expressed as:

[0133]

[0134] Where, L p is the cross entropy loss based on one-hot encoding, M p is the output of the mask layer, β p are the learnable parameters of the mask layer.

[0135] S303, simulating the dynamic behavior of the waves to optimize the weight of the neural network, 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 neural network weight; η p is the learning rate of the neural network, η p Set the value 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 Set the value to 0.1; is the update amount of the weight of the neural network in the last iteration;

[0138] Energy gradient The calculation of depends not only on the current weight position, but also on the historical speed and acceleration of weight change. Second-order dynamic elements are used to simulate this physical behavior. The calculation method is:

[0139]

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

[0141] The first and second order derivatives are calculated by discrete time difference method, and the calculation method is:

[0142]

[0143] Where Δt is the interval of iteration times, W p (t) is the neural network weight of the tth iteration, t is the number of iterations, W p (t-Δt) is the difference between the neural network weights at the tth iteration and the Δt interval iteration.

[0144] W p (t-2Δt) is the difference between the neural network weights at the tth iteration and the iteration with twice the Δt interval.

[0145] S304, input data is passed through the neural network, using the forward propagation method, and after 6 layers of processing, the output is generated; the loss is calculated based on the actual output and the expected output. This positioning method uses the cross entropy loss function based on one-hot encoding to measure the performance of the model. Suppose the output Y of the forward propagation is p Through the accumulation of inter-layer transfer calculation, the calculation method is:

[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 neural network in the lth layer, is the output of the neural network at layer l-1 of the neural network,

[0148] is the output of the adaptive masking layer of the neural network layer l; ⊙ represents the element-wise product, i.e., the Hadamard product, which is used to apply the adaptive masking; is the bias term of the lth layer.

[0149] Cross entropy loss function L based on one-hot encoding p The calculation method is:

[0150]

[0151] Where, 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, For the kth p One-hot encoding of the target labels of the samples, For the model pair k p The probability of sample prediction is calculated by the Softmax function.

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

[0153] [Example 4]

[0154] The overall steps are the same as those in Example 1, except that the training steps for training the feature dimensionality reduction model using the autoencoder algorithm based on the dynamic adjustment matrix in S4 are as follows:

[0155] S401, initializing the parameters of the autoencoder algorithm based on the dynamic adjustment matrix, including the weight and bias of the autoencoder, the initialization method is expressed as:

[0156]

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

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

[0159] h r =Re(W r ·X r +b r ),

[0160] In the formula, 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;

[0161] The low-dimensional features encoded by the encoder are input into the decoder to reconstruct the original data. The output of the decoder is expressed as:

[0162]

[0163] Where Xe is the reconstructed output data, represents the transpose of the weight parameters of the autoencoder, h r represents the encoded hidden layer features, b r Represents the bias parameter of the autoencoder.

[0164] Transpose of the weight parameters of the autoencoder It can be calculated by dynamically adjusting the strategy:

[0165]

[0166] Where W r is the weight parameter of the autoencoder, D r To dynamically adjust the matrix.

[0167] Dynamically adjust the matrix D r The element d rij According to the hidden layer feature h r The activation status is adjusted according to the calculation method:

[0168]

[0169] In the formula, To dynamically adjust the matrix D r The i c Row j c Elements of a column; and They are Adjust the intensity and threshold of the element, For the jth c The features of the hidden layer, i c To dynamically adjust the matrix row index; j c To dynamically adjust the matrix column index.

[0170] Threshold It is set by adaptive calculation, and the calculation method is:

[0171]

[0172] In the formula, ξ r is an adjustable coefficient, ξ r Set to 3; Median() is the intermediate value function, h rj Encoded features for the jth 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:

[0174]

[0175] In the formula, ∥∥ represents the L2 norm, X r is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder; r is the regularization coefficient of the autoencoder, λ r Set the value to 0.2; ∥W r ∥ F is the weight W r The Frobenius norm of r is the coefficient that controls the strength of the sparse regularization, ρ r Set to 2; n g is the number of autoencoder layers, which is the same as the number of columns of the dynamic adjustment matrix; is the indicator function, when Greater than The value is 1 when it is, otherwise it is 0;

[0176] Weight W r The Frobenius norm of Contains nonlinear penalties for each element of the weight matrix, calculated as:

[0177]

[0178] In the formula, W r The i-th c Row j c Column element; ∈ r is the nonlinear penalty coefficient, ∈ r Set the value to 1.5.

[0179] Dynamically adjust the matrix D r Elements The adjustment strength is adjusted through dynamic routing, and the adjustment method is expressed as:

[0180]

[0181] And, according to the adjusted Update the parameters of the autoencoder, and the update method is expressed as:

[0182]

[0183] In the formula, γ r is the routing sensitivity parameter, is the candidate weight from input to output, is the updated autoencoder weight parameter set, Softmax() is the Softmax function, for The adjusted strength of the element, W r The i-th c Row j c Column elements.

[0184] According to the updated autoencoder weight parameter set The bias parameter update method of the autoencoder is similar. The weight W of the autoencoder is updated again using the gradient descent method. r , the update method is expressed as:

[0185]

[0186] Where η r is the gradient descent learning rate, η r Set to 0.01; is the loss function relative to the weight gradient.

[0187] S404, repeat the above steps until the preset stop iteration condition is met, which means that the feature dimensionality reduction model training is completed, and the preset maximum number of iterations is set to 1000 times.

[0188] [Example 5]

[0189] The overall steps are the same as those in the first embodiment, except that the training steps of performing classifier training by using the extreme learning machine algorithm based on the adaptive error correction term in S5 are as follows:

[0190] S501, extreme learning machine weight parameter W u Initialize the extreme learning machine weight parameter W u The initialization method is expressed as:

[0191]

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

[0193] Angle matrix Θ u is randomly generated from a uniform distribution between 0 and π and is calculated as:

[0194] Θ u =π·Rand(0,1,size(W u )),

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

[0196] S502, normalize the low-dimensional feature vector after dimensionality reduction, and express it as:

[0197]

[0198] Where X' u is the normalized input feature, X u is the low-dimensional feature vector after dimensionality reduction, μ u With σ u X u The mean and standard deviation of

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

[0200] H u =Sig(W u ·X' u +b u ),

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

[0202] Extreme learning machine weight parameter W u During the training process, the adaptive gradient manifold exploration method is used for adjustment. The calculation method of the gradient adjustment increment is:

[0203]

[0204] In the formula, ΔW u is the update amount of the extreme learning machine weight; η u is the learning rate of the extreme learning machine, η u Set to 0.01; is the gradient of the extreme learning machine's loss function with respect to the weight; α u is the learning rate coefficient for the weight adjustment of the extreme learning machine, α u Set to 3; is the gradient of the manifold exploration function with respect to the extreme learning machine weight; R() is the manifold exploration function; κ u is the manifold exploration sensitivity, κ u Set to 1; is the extreme learning machine i d Layer neural network layer j d The weight parameter of each neuron, 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 neurons in the layer of the extreme learning machine, j d is the index of the number of neurons in the layer of the extreme learning machine, τ u Set to 1.5.

[0205] S504: Use the loss function based on the adaptive error correction term to perform loss calculation and back propagation to ensure the generalization ability of the model. The loss function L u The calculation method 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, For the i e The true label of sample data, i e is the sample index input to the extreme learning machine, For the i e The predicted label of sample data; u is the regularization parameter of the extreme learning machine, λ u Set to 0.4; ∥∥ 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, β u Set to 0.6; i is the i-th neural network layer, j is the j-th neuron, sin 2 (W u ) is the sine square term for the extreme learning machine weight parameter, and Rz(t) is the adaptive error correction term for the tth iteration;

[0208] Update the parameters of the extreme learning machine, and the update method is expressed as:

[0209]

[0210] Where W u represents the extreme learning machine weight parameter, b u is the bias parameter of the extreme learning machine; η u is the learning rate, η u Set to 0.01; and is the gradient of the loss function with respect to the weights and biases.

[0211] The adaptive error correction term with dynamic threshold adjustment is used to adjust the network output in real time to reduce the prediction error. The calculation method of the dynamic threshold of the tth iteration is defined as:

[0212]

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

[0214] At the tth iteration, the i-th e The calculation of the output label of the model of sample data is obtained by adaptive error correction. The adaptive error correction method is expressed as:

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

[0216] In the formula, 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 data distribution, ye u (t) and y u (t) are the true label and the output label of the model at the tth iteration, δ u (t) represents the dynamic threshold at the tth iteration.

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

[0218]

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

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

[0221] It should be noted that the present invention is not limited to the specific structures shown in the drawings in the above-mentioned embodiments, and various changes can be made thereto within the knowledge scope of ordinary technicians in this field.

Claims

1. The positioning method of the engineering inspection robot based on multi-sensor data fusion is characterized by: The following steps are involved: S1. Collect real sensor data through at least two different types of sensors, classify and annotate the real sensor data, and obtain annotated real sensor data; S2. Generate sensor data according to the annotated real sensor data by using a generative adversarial network algorithm based on meta-learning, wherein the annotated real sensor data and the generated sensor data constitute sample data; S3, inputting the sample data into a feature extraction model, and training the feature extraction model using a neural network algorithm based on wave energy capture optimization to obtain a trained feature extraction model and sample data after feature extraction; S4, inputting the sample data after feature extraction into a feature dimensionality reduction model, and training the feature dimensionality reduction model using an autoencoder algorithm based on a dynamic adjustment matrix to obtain a trained feature dimensionality reduction model and sample data after dimensionality reduction; S5, inputting the sample data after dimension reduction into a classifier, and training the classifier using an extreme learning machine algorithm based on an adaptive error correction term to obtain a trained classifier; S6. Collect new real sensor data through at least two different types of sensors, input the new real sensor data into the trained classifier, obtain the position classification of the engineering inspection robot in the position environment, and realize the position positioning of the engineering inspection robot.

2. The method for positioning an engineering inspection robot based on multi-sensor data fusion according to claim 1, characterized in that: The marking categories marked in S1 include: machine room area, office area, production workshop area and warehouse area.

3. The method for positioning an engineering inspection robot based on multi-sensor data fusion according to claim 1, characterized in that: The training steps of the generative adversarial network algorithm based on meta-learning described in S2 are as follows: S201, initialize generator parameters and the discriminator parameters S202, the generator receives a random noise vector z c , generating corresponding newly generated sensor data; S203, according to the feedback provided by the discriminator, using a meta-learning optimizer to adjust the parameter update strategy of the generator and the discriminator; S204: According to the gradient descent strategy, the generator parameters adjusted based on meta-learning and the discriminator parameters Update again; S205, repeating the above steps until a preset stop iteration condition is met, indicating that the feature model training is completed.

4. The method for positioning an engineering inspection robot based on multi-sensor data fusion according to claim 1, characterized in that: 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 three layers of the network, wherein the adaptive mask layer dynamically adjusts its transmittance according to the characteristics of the input data; S303, simulating the dynamic behavior of the waves to optimize the weight of the neural network, the calculation method of the update amount of the weight of the neural network is: 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 neural network in the last iteration; S304, input data is passed through the neural network, using the forward propagation method, and output is generated after 6 layers of processing; S305, repeating S302 to S304 until a preset condition for stopping iteration is met, which means that the feature extraction model training is completed.

5. The method for positioning an engineering inspection robot based on multi-sensor data fusion according to claim 1, characterized in that: The training steps of the feature dimensionality reduction model training in S4 are as follows: S401, initializing the parameters of the autoencoder algorithm based on the dynamic adjustment matrix, including the weight and bias of the autoencoder; S402, the data is transmitted layer by layer through the encoder layer by forward propagation, and the encoder output is expressed as: h r =Re(W r X r +b r ), In the formula, 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, 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: In the formula, ∥∥ is the L2 norm, X r is the input data of the autoencoder, Xe is the reconstructed output data of the autoencoder, λ r is the regularization coefficient of the autoencoder, ∥W r ∥ F is the weight W r The Frobenius norm, ρ r is the coefficient that controls the strength of sparse regularization; n g is the number of autoencoder layers; is the indicator function, when h rj Greater than θ rj The value is 1 when it is, otherwise it is 0; S404, repeating S402 to S403 until a preset condition for stopping iteration is met, which means that the feature dimensionality reduction model training is completed.

6. The method for positioning an engineering inspection robot based on multi-sensor data fusion according to claim 1, characterized in that: The training steps of the classifier training in S5 are as follows: S501, extreme learning machine weight parameter W u Initialize; S502, normalizing the low-dimensional feature vector after dimensionality reduction; S503, inputting the normalized features into an extreme learning machine classifier; S504: Use the loss function based on the adaptive error correction term to perform loss calculation and back propagation to ensure the generalization ability of the model. The loss function L u The calculation method is: 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, ∥∥ is the L2 norm, W u is the extreme learning machine weight parameter, β u is the nonlinear regularization coefficient of the extreme learning machine, sin 2 (W u ) is the sine square term for the extreme learning machine weight parameter, and Rz(t) is the adaptive error correction term for the tth iteration; S505, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed.

Images