A Fault Prediction Method for a Flexible Hand Used in Underwater Robot Operations

Abstract

The present invention discloses a method for fault prediction of a flexible hand for underwater robot operation, belonging to the technical field of fault prediction. The steps are as follows: acquiring historical fault data collected by a sensor unit for preprocessing, performing feature screening, constructing a fault prediction model based on a time convolutional network - bidirectional gated recurrent unit neural network, introducing a chaos mechanism and cross mutation to improve the sand cat algorithm for hyperparameter optimization of the fault prediction model, obtaining an optimized fault prediction model, inputting the online operation data of the flexible hand into the model, and performing fault prediction. The present invention can improve the accuracy of fault diagnosis, and effective fault prediction and maintenance measures can reduce the long-term maintenance cost of equipment and improve the overall economic benefit.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault prediction, and particularly relates to a method for fault prediction of a flexible hand for underwater robot operation. Background Art

[0002] In modern underwater robot technology, the flexible hands for underwater robot operation are playing an increasingly important role. These flexible arms are designed to cope with the extreme environments in the deep sea - whether it is a narrow pipeline, complex seabed terrain, or high-pressure and highly corrosive environments, the flexible hands can respond flexibly. For example, in deep-sea exploration, the flexible arm can act like a "finger in the ocean" and precisely operate in the complex seabed environment, such as picking up rare minerals or performing repair tasks. Such capabilities are particularly prominent in the face of traditional rigid arms.

[0003] Internationally, some top research institutions have made remarkable progress in the research and development of flexible hands. For example, the "Soft Robotics" project developed by the Massachusetts Institute of Technology has demonstrated the great potential of flexible hands in underwater tasks. These robots can not only operate stably in high-pressure environments but also perform delicate grasping and operation. Domestically, various universities are also actively promoting the research of flexible hand technology. In particular, the team at Shanghai Jiao Tong University has developed a flexible hand that can perform high-precision tasks in the harsh deep-sea environment, demonstrating the great potential of flexible hands in practical applications.

[0004] However, flexible hands face many challenges in practical applications. The high pressure and corrosion in the underwater environment pose severe tests on the materials and structures of the manipulators. In addition, the operating environment of underwater robots is extremely complex, and traditional fault detection and repair methods are often difficult to implement, which makes fault prediction technology particularly important. Effective fault prediction can give early warnings before problems occur, thus avoiding potential losses and downtime.

[0005] In the field of fault prediction for flexible hands, although there have been certain technological advancements, existing methods still have certain limitations and are unable to comprehensively handle complex working environments and fault modes. Current fault prediction systems mainly rely on real-time monitoring of the working state of equipment and combine data analysis with machine learning algorithms for fault early warning. For example, a team at Cornell University developed a fault prediction system based on real-time data analysis that can relatively accurately predict the faults of flexible hands and ensure the stability of the equipment. However, when faced with different operating modes, noise interference, or unstable working environments, this method often faces the risk of reduced prediction accuracy, and there is still room for improvement in the adaptability and real-time performance of the system. In addition, a research team at Shanghai Jiao Tong University proposed a hybrid method that combines the advantages of data-driven and model-driven approaches, improving the accuracy and robustness of fault prediction. Although this method shows good adaptability in some complex environments, there are still deficiencies in high-dimensional data processing, long model training time, and resource consumption. Therefore, existing technologies generally have difficulty meeting the requirements of high efficiency, real-time performance, and high accuracy, especially in applications in dynamic environments. It is precisely in response to these technical deficiencies that the present invention proposes a new fault prediction method, which solves the shortcomings of existing technologies in terms of real-time performance, adaptability, and efficiency through innovative model and data processing means. This method can not only improve the prediction accuracy but also maintain high robustness in complex environments, thus providing a stronger guarantee for the reliability of flexible hands in underwater robot operations. Through the present invention, flexible hand fault prediction can be effectively implemented in a wider range of application scenarios, further promoting the popularization and application of underwater robots in deep-sea operations and opening up new development paths for future ocean exploration. Summary of the Invention

[0006] In view of the above problems existing in the prior art, the present invention proposes a fault prediction method for a flexible hand used in underwater robot operations, with reasonable design, which solves the deficiencies of the prior art and has good effects.

[0007] A fault prediction method for a flexible hand used in underwater robot operations, comprising the following steps:

[0008] S1. Obtain historical fault data collected by the sensor unit, including current, force, torque, position information, and vibration signals;

[0009] S2. Perform data preprocessing on the data collected in S1, including data augmentation, short-time Fourier transform, overlapping sampling, standardization, and normalization;

[0010] S3. Select a balanced sampling selector to increase the sampling proportion of abnormal samples, and use an improved random forest algorithm to screen important features to obtain a screened data set;

[0011] S4. Construct a fault prediction model based on a time convolutional network - bidirectional gated recurrent unit neural network, and use the filtered dataset for training to obtain a trained fault prediction model;

[0012] S5. Introduce a chaotic mechanism and cross - mutation to improve the sand cat algorithm for hyperparameter optimization of the fault prediction model;

[0013] S6. Obtain the optimized fault prediction model, input the on - line operation data of the flexible hand into this model for fault prediction.

[0014] Furthermore, in S1, the collected historical fault data is divided into a dataset according to the fault type ;

[0015] In S2, data augmentation is performed on the dataset based on a generative adversarial network, and the augmented dataset is ; Perform short - time Fourier transform on to obtain time - frequency data , further augment the data through overlapping sampling to obtain , and finally perform dimension - unification processing on each data sample through standardization and normalization.

[0016] Furthermore, in S3, sub - sets are generated from the normalized dataset through bootstrap sampling. Each sub - set contains the same number of samples as the original dataset. Use a balanced sampling selector to increase the sampling probability of fault data in the entire dataset. Continue with balanced sampling after sampling, and determine whether the fault data in the sample is balanced through a threshold. The expression is:

[0017] ; (2)

[0018] where is the threshold, is the number of fault data, is the number of normal data;

[0019] If the proportion of fault data is less than , it means that the proportion weight of fault data needs to be increased in the sample, trigger the second - stage balanced sampler, and increase the proportion weight of sample points;

[0020] After the data is selected through balanced sampling, a new dataset is obtained, where represents having sub - sets.

[0021] Further, in S3, since there is a problem of diversity degradation in the random forest, the kernel density reconstruction gradient boosting algorithm is used to solve this problem. According to the prediction of each tree, the residual between the true value and the predicted value is calculated:

[0022] ;(3)

[0023] where is the residual of the th variable value, is the measured value, is the predicted value;

[0024] The distribution of the residual is reconstructed using kernel density, and the expression is:

[0025] );(4)

[0026] where is the estimated probability density of the residual, is the kernel function, is the bandwidth parameter; is the center value of the kernel function, which is a constant;

[0027] According to the reconstructed residual distribution, the weights of each decision tree are updated. Combining the results of all previous trees, a new model is generated by means of weighted sum;

[0028] For the dataset the improved random forest algorithm is used for feature screening to obtain the screened dataset .

[0029] Further, in S4, in the temporal convolutional network - bidirectional gated recurrent unit neural network, the screened dataset is input into the dilated causal convolutional layer, batch weight normalization layer, activation function layer, LRN layer, Dropout layer, convolutional unit, fully connected layer, channel and spatial attention fusion unit;

[0030] The dilated causal convolutional layer calculates the feature data of each time step. The batch weight normalization layer normalizes the weights of the feature data. The activation function layer uses the ReLU activation function to activate the normalized feature data, and then inputs it into the LRN layer. Its expression is:

[0031] ;(5)

[0032] where: is the activation value of the th pixel in the corresponding fault image feature map under the th channel, is the th channel The activation value of the pixel, is the output of the pixel under the th channel after normalization, ; , are hyperparameters; is the size of the normalized neighborhood; is the total number of channels in the current feature layer;

[0033] The output after normalization is input into the Dropout layer. The convolutional unit includes a 1×1 convolutional layer, a 2×2 convolutional layer, and a 3×3 convolutional layer. The output of the Dropout layer is respectively input into the three convolutional layers, and the inputs of the three convolutional layers are jointly input into the fully connected layer and the channel and spatial attention fusion unit. The spatial attention applies global average pooling to the output data of the fully connected layer and performs splicing. After splicing, spatial attention weights are generated through convolution, and the spatial attention is applied to the feature map; the average pooled feature vector of each channel is obtained through global average pooling. The feature vector undergoes a non-linear transformation through the fully connected layer and is spliced. After splicing, channel attention weights are generated through the fully connected layer, and the channel attention is applied to the feature map. The feature map with spatial attention applied and the feature map with channel attention applied are concatenated and fused to obtain the fused feature map;

[0034] The time convolutional network - bidirectional gated recurrent unit neural network further includes a 1D fully convolutional layer and an average pooling layer. The filtered dataset is sequentially input into the 1D fully convolutional layer and the average pooling layer. The compressed feature output by the average pooling layer is multiplied by the feature map output by the channel and spatial attention unit to obtain the feature map after attention features;

[0035] Finally, the feature map after attention features is input into the fully connected layer and the softmax layer. The softmax layer converts the logits of each category into probability values through the softmax function and outputs the predicted probabilities of each fault category.

[0036] Furthermore, in S5, specifically:

[0037] S5.1. The improved sand cat algorithm is adopted. The independent variables in the sand cat algorithm are the hyperparameters of the neural network, and the fitness function is the residual output by the model;

[0038] S5.2. A chaos mechanism is introduced to initialize the population;

[0039] S5.3. Through crossover operations, new individuals are generated to enhance population diversity; a crossover point is randomly selected, and the genes of two parent individuals are cut from the crossover point and part of the genes are exchanged;

[0040] S5.4. Introduce new features through mutation operations to increase the diversity of the population and prevent the algorithm from falling into local optimal solutions;

[0041] First, determine the mutation probability, set the mutation probability and decide the mutation frequency, and then execute the mutation process through the following formula:

[0042] ;(8)

[0043] Among them, is the predicted position, is the current position, is the mutation amplitude, is a pseudo-random number generated from the chaotic sequence, ;

[0044] S5.5. In the stage of searching for prey, the expression is:

[0045] ;(9)

[0046] );(10)

[0047] Among them, represents the updated search radius, which gradually shrinks as the number of iterations increases to control the search range, represents the maximum search range, that is, the farthest distance that the sand cat can explore at the current iteration, represents the current number of iterations, which is used to dynamically adjust the algorithm parameters, is the maximum number of iterations, which represents the maximum number of iterations for the algorithm to run. It is used to gradually shrink the search range during the iteration process, is the updated position of the sand cat, is the scaling factor, which is used to adjust the amplitude of position update, is the current best position, that is, the best position found at the th iteration, is a randomly generated value within the range of [0,1], which introduces randomness to increase the diversity of the search, is the position of the current individual, that is, the current position of the sand cat at the th iteration;

[0048] S5.6. In the stage of attacking prey, calculate the fitness during the iteration process to obtain the direction of function value decrease to replace the random direction, and calculate the next position. The expression is:

[0049] ;(11)

[0050] ;(12)

[0051] Among them, is the middle position, which is used to adjust the randomness and directionality of the position. is the global best position, that is, the best position from the start to the current of the algorithm. is the direction factor, which adjusts the position update direction. is the updated position of the sand cat in the stage of attacking the prey, which is used to determine the next search direction.

[0052] The final parameter and the main parameter for controlling the transition between the exploration and exploitation stages are When , the sand cat searches for the prey. When , the sand cat attacks the prey.

[0053] S5.7: Obtain a new group of populations, substitute the independent variables represented by the populations into the fitness function, calculate the fitness values, and determine whether the iteration condition is reached. If not, continue the iteration until the iteration condition is reached to obtain the optimal parameters.

[0054] Furthermore, in S5.2, specifically:

[0055] Select different chaotic mapping functions, including the Logistic mapping and the Tent mapping.

[0056] Define the intervals, divide the number axis [0, 1] into manually set intervals. The interval division formula is:

[0057] , ; (15)

[0058] Select the initial value as a small random number. , and the Logistic mapping and the Tent mapping functions are alternately used in each interval in turn, so that the generated chaotic sequence has different non-linear behaviors in different intervals, increasing the diversity of the initial population. The expression is:

[0059] ; (16)

[0060] Among them, is the chaotic mapping function corresponding to each interval. is the th chaotic mapping value. is the th chaotic mapping value;

[0061] Perform normalization operation on the generated chaotic sequence:

[0062] ; (17)

[0063] Among them, and are the minimum and maximum values of the sequence respectively.

[0064] The beneficial technical effects brought by the present invention are as follows:

[0065] (1) Improve diagnostic accuracy: By comprehensively analyzing various data such as the position, current, and vibration of the flexible hand actuator, the present invention combines a temporal convolutional network and a bidirectional gated recurrent unit to achieve early prediction of potential fault points. This method can perform maintenance before a fault occurs, significantly reducing the need for emergency repairs, thereby improving the overall operating efficiency and stability of the equipment.

[0066] (2) Optimize feature screening: The kernel density reconstruction gradient boosting algorithm is used for feature screening. Compared with the traditional random forest voting mechanism, it can provide more accurate, stable, and interpretable feature selection results, and better adapt to the complexity and diversity of data.

[0067] (3) Improve model performance: The present invention combines a temporal convolutional network and a bidirectional gated recurrent unit, greatly enhancing the processing ability of time series data. This integrated model shows higher accuracy and efficiency in fault diagnosis, and is particularly suitable for complex underwater operation environments.

[0068] (4) Reduce maintenance costs: Through accurate fault prediction, this system can make maintenance work more precise, reduce unnecessary repairs and inspections, and lower labor costs and component replacement costs. This efficient maintenance strategy can significantly save resources and reduce the total maintenance cost.

[0069] (5) Improve the sand cat algorithm: Chaos mapping and crossover mutation techniques are introduced into the sand cat algorithm, enhancing its global search ability and improving the convergence speed. This improvement not only enhances the overall performance of the algorithm, but also optimizes the local-global balance of the model, making fault prediction more accurate and efficient.

[0070] (6) Extend the equipment life: By timely identifying and handling potential faults, it helps the equipment maintain the best operating state, thereby extending the service life of the equipment. Effective fault prediction and maintenance measures can reduce the long-term maintenance cost of the equipment and improve the overall economic benefits.

[0071] (7) Improve real-time performance and response speed: Real-time monitor and analyze data, quickly respond to potential faults, and ensure timely control of the equipment status. This improvement in real-time performance and response speed helps to quickly handle faults, reduce downtime, and improve the operating reliability of the system. Description of the Drawings

[0072] Figure 1 This is the structural diagram of the sensor positions of the flexible hand for the underwater working robot operation in the present invention.

[0073] Among them, 1 - position sensor, 2 - current sensor, 3 - vibration sensor, 4 - force and torque sensor.

[0074] Figure 2 This is the flowchart of data pre - processing in the present invention.

[0075] Figure 3 This is the flowchart of a fault prediction method for the flexible hand used in underwater robot operation. Specific implementation mode

[0076] The following further explains the specific implementation mode of the present invention in combination with specific embodiments:

[0077] A fault prediction method for the flexible hand used in underwater robot operation includes the following steps:

[0078] S1. Obtain the historical fault data collected by the sensor unit, including current, force, torque, position information, vibration signals, etc.;

[0079] As Figure 1 shown, the sensor unit includes a position sensor 1, a current sensor 2, a vibration sensor 3, and a force and torque sensor 4; the current sensor is used to monitor the current of the flexible hand for robot operation in real - time to help monitor the stability of the power system; the force and torque sensor is used to capture the force and torque applied by the manipulator to optimize the motion control and improve the operation stability; the position sensor is used to track the position of the end - effector of the manipulator to ensure that the system can adjust the position and motion trajectory of the end - effector in real - time; the vibration sensor is used to detect the vibration signal of the manipulator to identify potential mechanical faults.

[0080] S2. Perform data pre - processing on the data collected in S1, including data augmentation, short - time Fourier transform, overlapping sampling, standardization, and normalization, as Figure 2 shown;

[0081] Divide the collected historical fault data into data sets according to the fault types ;

[0082] Perform data augmentation on the data set based on the generative adversarial network, and use the real fault data Input into the discriminator, and random noise is input into the generator; train the discriminator. The training of the discriminator depends on the confrontation with the generator. The generator tries to generate more realistic samples, while the discriminator tries to distinguish between true and false. In this process, the "basis" of the discriminator is not fixed, but dynamically adjusted during training to adapt to the improvement of the generator. Through the above process, ultimately, the game between the two prompts the generator to generate more realistic data. The enhanced dataset is ;

[0083] Perform a short-time Fourier transform on to obtain time-frequency data ;

[0084] Further enhance the data through overlapping sampling. The overlapping length should be determined according to the characteristics of the signal and the diagnostic requirements. Through overlapping sampling, multiple different training samples can be extracted from the original signal. These samples are time-shifted but contain similar fault characteristics. This helps the model learn richer fault information. For training samples, when collecting training samples from the data, by delimiting the sampling bands, there is an overlap between the signals of each band and the adjacent bands. Assume the number of enhanced data points is and the sampling band is L + offset, L where offset is the size of the window for each sampling, and its value range is from 10 to 100 data points. Offset controls the overlap amount between adjacent sampling windows. It represents the overlapping part between the windows when moving the sampling window each time. L takes 50% of

[0085] Then the enhanced fault dataset after overlapping sampling can be obtained using Equation (1) :

[0086] ; (1)

[0087] Since the scales and dimensions of each data are inconsistent, finally, dimensionality unification processing is performed on each data sampling through standardization and normalization.

[0088] S3. Select a balanced sampling selector, increase the sampling proportion of abnormal samples, and use an improved random forest algorithm to screen important features to obtain the screened dataset;

[0089] From the normalized dataset generate A subset, each subset contains the same number of samples as the original dataset. Use a balanced sampling selector to increase the sampling probability of faulty data in the entire dataset. After sampling, continue with balanced sampling and determine whether the faulty data in the samples is balanced through a threshold. The expression is:

[0090] ;(2)

[0091] Among them, is the threshold, is the number of faulty data, is the number of normal data;

[0092] If the proportion of faulty data is less than , it means that the weight of the proportion of faulty data needs to be increased in the samples, triggering the second balanced sampler to increase the proportion weight of the sample points;

[0093] After the data is selected by balanced sampling, a new dataset is obtained, where represents having subsets;

[0094] Due to the problem of diversity degradation in the random forest, the kernel density reconstruction gradient boosting algorithm is used to solve this problem. According to the prediction of each tree, calculate the residual between the true value and the predicted value :

[0095] ;(3)

[0096] Among them, is the residual of the th variable value, is the measured value, is the predicted value;

[0097] Use kernel density to reconstruct the distribution of the residuals. The expression is:

[0098] );(4)

[0099] Among them, is the estimated probability density of the residuals, is the kernel function, is the bandwidth parameter; is the center value of the kernel function, which is a constant;

[0100] Update the weights of each decision tree according to the reconstructed residual distribution. Combine the results of all previous trees and use the weighted sum method to generate a new model;

[0101] For the dataset Use the improved random forest algorithm for feature screening to obtain the screened dataset 。

[0102] S4. Build a fault prediction model based on the Time Convolutional Network - Bidirectional Gated Recurrent Unit Neural Network (TCN - BiGRU), and use the screened dataset for training to obtain a trained fault prediction model;

[0103] In the Time Convolutional Network - Bidirectional Gated Recurrent Unit Neural Network, input the screened dataset into the dilated causal convolutional layer, batch weight normalization layer, activation function layer, LRN layer, Dropout layer, convolutional unit, fully connected layer, channel and spatial attention fusion unit;

[0104] The dilated causal convolutional layer calculates the feature data for each time step. The batch weight normalization layer normalizes the weights of the feature data. During the neural network training process, by normalizing the weights, the gradients are stabilized and the training process is accelerated. For the weights of each batch, calculate its mean and method, and normalize them to stabilize the training process, reduce gradient problems, and accelerate the convergence of the network;

[0105] The activation function layer uses the ReLU activation function to activate the normalized feature data, and then inputs it into the LRN layer. For the activation value of each neuron at each position, select its local neighborhood (such as n adjacent channels) in the channel dimension, and calculate the local normalization coefficient: calculate the sum of squares on these channels to obtain the normalization coefficient; normalize the activation value of the neuron at this position with the normalization coefficient, and its expression is:

[0106] ;(5)

[0107] Where: is the th activation value of the pixel in the corresponding fault image feature map under the th channel, is the activation value of the pixel under the th channel, , is the output of the th channel after normalization for the pixel, 、 are hyperparameters, usually set α = ,β = 0.75; m is the size of the normalization neighborhood, that is, the number of surrounding channels considered during local response normalization; is the total number of channels in the current feature layer.

[0108] The normalized output is fed into the Dropout layer, which reduces overfitting and improves the generalization ability of the model by randomly dropping some neurons in the neural network. During training, it sets the activation values of some neurons to zero, preventing the network from over-relying on the training data. A dropout probability p is defined, for example, 0.5, indicating that each neuron has a 50% probability of being dropped in each iteration. In each training step, a random mask of the same shape as the network is generated, and each element in the mask is 0 with probability p and 1 with probability 1. To keep the output consistent between the training and prediction phases, the average value of the activation values of the dropped neurons during training is reduced. Therefore, during training, the output activation values are usually scaled by 1 / (1−p) to keep the expected values consistent in the prediction phase.

[0109] The convolutional unit includes a 1×1 convolutional layer, a 2×2 convolutional layer, and a 3×3 convolutional layer. The output of the Dropout layer is respectively input into the three convolutional layers, and the inputs of the three convolutional layers are jointly input into the fully connected layer and the channel and spatial attention fusion unit. The spatial attention applies global average pooling to the output data of the fully connected layer and concatenates them. After concatenation, spatial attention weights are generated through convolution, and the spatial attention is applied to the feature map. The average pooled feature vector of each channel is obtained through global average pooling. The feature vector undergoes a non-linear transformation through the fully connected layer and is concatenated. After concatenation, channel attention weights are generated through the fully connected layer, and the channel attention is applied to the feature map. The feature map with spatial attention applied and the feature map with channel attention applied are concatenated and fused to obtain the fused feature map;

[0110] The Temporal Convolution Network - Bidirectional Gated Recurrent Unit neural network also includes a 1D fully convolutional layer and an average pooling layer. The filtered dataset is sequentially input into the 1D fully convolutional layer and the average pooling layer. The compressed features output by the average pooling layer are multiplied by the feature map output by the channel and spatial attention unit to obtain the feature map after attention features;

[0111] Finally, the input after attention features is fed into the fully connected layer and the softmax layer. The softmax layer converts the logits of each class into probability values through the softmax function and outputs the predicted probabilities of each fault class. The network output is compared with the expected output, the loss is calculated, and the weights and biases of the model parameters are updated along the negative gradient direction of the loss function to minimize the loss function and improve the prediction performance of the model. If the loss continues to decay compared to the previous training, the hyperparameters are reset and training starts again;

[0112] The Temporal Convolution Network - Bidirectional Gated Recurrent Unit model is predicted and evaluated using the Mean Squared Error (MSE) and the Mean Absolute Error (MAE);

[0113] Calculate the mean squared error (MSE):

[0114] ;(6)

[0115] where is the number of prediction points, is the true value of the th point, is the corresponding predicted value;

[0116] Calculate the mean absolute error (MAE):

[0117] (7).

[0118] S5. Introduce a chaos mechanism and crossover mutation to improve the sand cat algorithm for hyperparameter optimization of the fault prediction model, including the following sub-steps:

[0119] S5.1. Adopt the improved sand cat ISCSO algorithm, where the independent variable in the sand cat algorithm is the hyperparameter of the neural network, and the fitness function is the residual of the model output;

[0120] S5.2. Introduce a chaos mechanism to initialize the population;

[0121] S5.3. Generate new individuals through crossover operations to enhance population diversity; randomly select a crossover point, cut the genes of two parent individuals from the crossover point, and exchange part of the genes;

[0122] S5.4. Introduce new features through mutation operations to increase population diversity and avoid the algorithm falling into a local optimal solution;

[0123] First, determine the mutation probability, set the mutation probability and the frequency of mutation, and then execute the mutation process through the following formula:

[0124] ;(8)

[0125] where is the prediction position, is the current position, is the mutation amplitude, is the pseudo-random number generated from the chaos sequence, ;

[0126] S5.5. In the stage of searching for prey, the expression is:

[0127] ;(9)

[0128] );(10)

[0129] where To represent the updated search radius, also known as the algorithm sensitivity, which gradually shrinks as the number of iterations increases to control the search range. The sensitivity of the sand cat is not fixed. A non-linear variation method is used to adjust the sensitivity, which is adjusted according to the distance between the current position of the sand cat and the optimal solution. When the sand cat is far from the optimal solution, a larger sensitivity value is given to expand the search range; when the sand cat is close to the optimal solution, a smaller sensitivity value is given for fine search. To represent the maximum search range, that is, the farthest distance that the sand cat can explore during the current iteration, To represent the current number of iterations, which is used to dynamically adjust the algorithm parameters, To represent the maximum number of iterations, which indicates the maximum number of iterations for the algorithm to run. It is used to gradually shrink the search range during the iteration process, To represent the updated position of the sand cat, To represent the scaling factor, which is used to adjust the amplitude of position update, To represent the current best position, that is, the best position found at the th iteration, To represent a randomly generated value within the range of [0,1], which introduces randomness to increase the diversity of the search, To represent the position of the current individual, that is, the current position of the sand cat at the t-th iteration;

[0130] S5.6. Attacking the prey stage. Assume that the sensitivity range of the sand dune cat is a circle, so the moving direction can be determined by a random angle on the circle, and its value is between -1 and 1. The roulette selection algorithm is used to select a random angle for each sand dune cat to approach the hunting position. To avoid falling into local optimum, a random angle is adopted. The random angle in the formula will have a positive effect on the hunting and guiding of the sand cat. The random position can ensure that the involved cats can approach the prey. However, the random angle cannot guarantee that the function value will definitely decrease, and it has great randomness. Therefore, by calculating the fitness during the iteration process to obtain the angle at which the function value decreases to replace the random angle. And calculate the next position, the expression is:

[0131] ;(11)

[0132] ;(12)

[0133] Among them, To represent the intermediate position, which is used to adjust the randomness and directionality of the position, To represent the global best position, that is, the best position from the start of the algorithm to the current time, To represent the direction factor, which adjusts the position update direction, To represent the updated position of the sand cat in the prey attack stage, which is used to determine the next search direction;

[0134] The final parameters and main parameters for controlling the transition between the exploration and development stages are , when , the sand cat searches for prey, and when , the sand cat attacks the prey;

[0135] S5.7: Obtain a new set of populations, substitute the independent variables represented by the populations into the fitness function, calculate the fitness values, and determine whether the iteration condition is met, that is, reach the convergence threshold. If not, continue the iteration until the iteration condition is reached to obtain the optimal parameters.

[0136] Specifically, S5.2 is as follows:

[0137] Select different chaotic mapping functions, including the Logistic mapping and the Tent mapping;

[0138] Logistic mapping:

[0139] ;(13)

[0140] Among them, is the control parameter, usually taking values between (3.57, 4).

[0141] Tent mapping:

[0142] ;(14)

[0143] Among them is the control parameter, usually taking values between (0, 1);

[0144] Define the interval, divide the number axis [0, 1] into intervals set artificially, and the interval division expression is:

[0145] , ;(15)

[0146] Among them, is the number of independent variables in a single population in the sand cat algorithm;

[0147] Select the initial value as a small random number. To ensure that the chaotic mapping function can generate a sequence with good randomness and avoid the sequence falling into periodicity or fixed points, , the Logistic mapping and Tent mapping functions are alternately used in each interval in turn, so that the generated chaotic sequence has different non-linear behaviors in different intervals, increasing the diversity of the initial population. The expression is:

[0148] ; (16)

[0149] Among them, is the chaotic mapping function corresponding to each interval, is the th chaotic mapping value, is the th chaotic mapping value;

[0150] The generated chaotic sequence may not be within the range required by the algorithm. Therefore, a normalization operation is required to ensure that the sequence meets the requirements of the initial population. The expression is:

[0151] ; (17)

[0152] Among them, and are the minimum and maximum values of the sequence respectively.

[0153] S6. Obtain the optimized fault prediction model, input the online operation data of the flexible hand into this model, and conduct fault prediction.

[0154] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A fault prediction method for a flexible hand used in underwater robot operations, characterized in that, Including the following steps: S1. Obtain the historical fault data collected by the sensor unit, including current, force, torque, position information, and vibration signals; S2. Perform data preprocessing on the data collected in S1, including data augmentation, short-time Fourier transform, overlapping sampling, standardization, and normalization; S3. Select a balanced sampling selector to increase the sampling ratio of abnormal samples, and use an improved random forest algorithm to screen important features to obtain a screened dataset; S4. Build a fault prediction model based on a time convolutional network - bidirectional gated recurrent unit neural network, and use the screened dataset for training to obtain a trained fault prediction model; S5. Introduce a chaotic mechanism and cross-mutation to improve the sand cat algorithm to optimize the hyperparameters of the fault prediction model; S6. Obtain the optimized fault prediction model, input the on-line operation data of the flexible hand into this model for fault prediction; In S3, due to the problem of diversity degradation in the random forest, the kernel density reconstruction gradient boosting algorithm is used to solve this problem. According to the prediction of each tree, calculate the residual between the true value and the predicted value: ；(3) Among them, is the residual of the th variable value, is the measured value, is the predicted value; Use kernel density to reconstruct the distribution of the residuals, and the expression is: )；(4) Among them, is the estimated probability density of the residual, is the kernel function, is the bandwidth parameter; is the central value of the kernel function and is a constant; Update the weights of each decision tree according to the reconstructed residual distribution, and combine the results of all previous trees, and use the weighted sum method to generate a new model; For the dataset Use the improved random forest algorithm for feature screening to obtain the screened dataset ; In S5, specifically: S5.

1. Adopt an improved sand cat algorithm. The independent variable in the sand cat algorithm is the hyperparameter of the neural network, and the fitness function is the residual output by the model; S5.

2. Introduce a chaotic mechanism to initialize the population; S5.

3. Through the crossover operation, generate new individuals to enhance the population diversity; randomly select a crossover point, cut the genes of two parent individuals from the crossover point, and exchange part of the genes; S5.

4. Introduce new features through the mutation operation to increase the diversity of the population and avoid the algorithm falling into a local optimal solution; S5.

5. Search for prey stage; S5.

6. Attack the prey stage. Calculate the direction of the function value drop by calculating the fitness during the iterative process to replace the random direction, and calculate the next position; S5.

7. Obtain a new group of populations, bring the independent variables represented by the populations into the fitness function, calculate the fitness value, and judge whether the iterative condition is reached. If not, continue to iterate until the iterative condition is reached to obtain the optimal parameters.

2. The fault prediction method for a flexible hand used in underwater robot operation according to claim 1, characterized in that In S1, the collected historical fault data is divided into data sets according to fault types ; In S2, data augmentation is performed on the dataset based on a generative adversarial network, and the augmented dataset is ; perform short-time Fourier transform to obtain time-frequency data , and further augment the data through overlapping sampling to obtain . Finally, dimension unification processing is performed on each data sample through standardization and normalization.

3. A method for predicting the failure of a flexible hand for underwater robot operation according to claim 2, characterized in that, In S3, from the normalized dataset generate subsets through bootstrap sampling. Each subset contains the same number of samples as the original dataset. Use a balanced sampling selector to increase the sampling probability of faulty data in the entire dataset. After sampling, continue with balanced sampling. Determine whether the faulty data in the samples is balanced through a threshold. The expression is: ；(2) Among them, is the threshold value, is the number of fault data, is the number of normal data; If the proportion of fault data is less than , it indicates that it is necessary to increase the proportion weight of fault data in the samples, trigger the second balancing sampler, and then increase the proportion weight of the sample points; After the data is selected by balanced sampling, a new data set is obtained , where means there are subsets 4. A flexible hand fault prediction method for underwater robot operation according to claim 3, characterized in that In S4, in the time convolutional network - bidirectional gated recurrent unit neural network, input the screened dataset into the dilated causal convolutional layer, batch weight normalization layer, activation function layer, LRN layer, Dropout layer, convolutional unit, fully connected layer, channel and spatial attention fusion unit; The dilated causal convolutional layer calculates the feature data of each time step. The batch weight normalization layer normalizes the weights of the feature data. The activation function layer uses the ReLU activation function to activate the normalized feature data, and then inputs it into the LRN layer. The expression is: ；(5) Wherein: is the activation value of the pixel in the corresponding fault image feature map under the th channel, is the activation value of the pixel under the th channel, is the output of the pixel under the th channel after normalization, , are hyperparameters; is the size of the normalized neighborhood; is the total number of channels in the current feature layer;​ The normalized output is input into the Dropout layer. The convolutional unit includes a 1×1 convolutional layer, a 2×2 convolutional layer, and a 3×3 convolutional layer. The output of the Dropout layer is respectively input into the three convolutional layers, and the inputs of the three convolutional layers are jointly input into the fully connected layer and the channel and spatial attention fusion unit. The spatial attention applies global average pooling to the output data of the fully connected layer and performs concatenation. After concatenation, spatial attention weights are generated through convolution, and the spatial attention is applied to the feature map; the average pooled feature vector of each channel is obtained through global average pooling. The feature vector undergoes a non-linear transformation through the fully connected layer and performs concatenation. After concatenation, channel attention weights are generated through the fully connected layer, and the channel attention is applied to the feature map. The feature map applied with spatial attention and the feature map applied with channel attention are concatenated and fused to obtain the fused feature map; The time convolutional network - bidirectional gated recurrent unit neural network further includes a 1D fully convolutional layer and an average pooling layer. The screened dataset is sequentially input into the 1D fully convolutional layer and the average pooling layer. The compressed feature output by the average pooling layer is multiplied by the feature map output by the channel and spatial attention unit to obtain the feature map after attention features; Finally, the feature map after attention features is input into the fully connected layer and the softmax layer. The softmax layer converts the logits of each category into probability values through the softmax function and outputs the predicted probabilities of each fault category.

5. A method for predicting faults of a flexible hand for underwater robot operation according to claim 4, characterized in that, The specific content of S5.4 is as follows: First, determine the mutation probability, set the mutation probability and the frequency of determining mutation, and then execute the mutation process through the following formula: ； (8) Among them, is the predicted position, is the current position, is the mutation amplitude, is a pseudo-random number generated from the chaotic sequence, ; In S5.5, the expression in the prey search stage is: ；(9) )；(10) Among them, represents the updated search radius, which gradually shrinks as the number of iterations increases to control the search range. represents the maximum search range, that is, the farthest distance that the sand cat can explore at the current iteration. represents the current number of iterations, which is used to dynamically adjust the algorithm parameters. is the maximum number of iterations, indicating the maximum number of iterations for the algorithm to run, and is used to gradually shrink the search range during the iteration process. is the updated position of the sand cat. is the scaling factor, which is used to adjust the amplitude of the position update. is the current best position, that is, the best position found at the th iteration. is a randomly generated value in the range of [0, 1], which introduces randomness to increase the diversity of the search. is the position of the current individual, that is, the current position of the sand cat at the th iteration; The specific content of S5.6 is as follows: ；(11) ；(12) Among them, is the middle position, which is used to adjust the randomness and directionality of the position, is the global best position, that is, the best position from the start of the algorithm to the current time, is the direction factor, which adjusts the position update direction, is the updated position of the sand cat in the stage of attacking the prey, which is used to determine the next search direction; The final and main parameters that control the transition between the exploration and development phases are , when , the sand cat searches for prey, and when , the sand cat attacks the prey.

6. A method for predicting faults of a flexible hand for underwater robot operation according to claim 5, characterized in that, In S5.2, specifically: Select different chaotic mapping functions, including the Logistic mapping and the Tent mapping; Define an interval to divide the number line [0, 1] into manually set intervals. The interval division formula is: ， ；(15) Select the initial value is a small random number, , the Logistic mapping and Tent mapping functions are alternately used in each interval in turn, so that the generated chaotic sequence has different nonlinear behaviors in different intervals, increasing the diversity of the initial population. The expression is: ；(16) Among them, is the chaotic mapping function corresponding to each interval, is the th chaotic mapping value, is the th chaotic mapping value; Perform normalization operations on the generated chaotic sequence: ；(17) Among them, and are the minimum and maximum values of the sequence, respectively.

Images