Fault prediction method for flexible hand for underwater robot operation

Abstract

The invention discloses a fault prediction method for a flexible hand for underwater robot operation, and belongs to the technical field of fault prediction, and the method comprises the steps: obtaining historical fault data collected by a sensor unit, carrying out the preprocessing, carrying out the feature screening, building a fault prediction model based on a time convolution network-bidirectional gating recursive unit neural network, and carrying out the fault prediction. And introducing a chaos mechanism and a crossover mutation improved samson algorithm to carry out hyper-parameter optimization on the fault prediction model to obtain an optimized fault prediction model, and inputting the online operation data of the flexible hand into the model to carry out fault prediction. The fault diagnosis precision can be improved, the long-term maintenance cost of equipment can be reduced through effective fault prediction and maintenance measures, and the overall economic benefit is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of fault prediction, and in particular relates to a method for predicting the fault of a flexible hand used for underwater robot operations. Background Art

[0002] In modern underwater robot technology, flexible hands for underwater robot operations are playing an increasingly important role. These flexible arms are designed to cope with extreme environments in the deep sea - whether it is narrow pipes, complex seabed terrain, or high-pressure, highly corrosive environments, flexible hands can cope with them flexibly. For example, in deep-sea exploration, flexible arms can act like "fingers in the ocean" to accurately operate in complex seabed environments, such as picking up rare minerals or performing maintenance tasks. Such capabilities are particularly prominent in front of traditional rigid arms.

[0003] Internationally, some top scientific research institutions have made significant progress in the research and development of flexible hands. For example, the "soft robot" project developed by MIT 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 fine grasping and manipulation. In China, various universities are also actively promoting the research of flexible hand technology. In particular, the team of Shanghai Jiaotong University has developed a flexible hand that can perform high-precision tasks in harsh deep-sea environments, demonstrating the great potential of flexible hands in practical applications.

[0004] However, flexible hands face many challenges in practical applications. The high pressure and corrosiveness in the underwater environment pose a severe test to the materials and structures of the manipulator. In addition, the operating environment of underwater robots is extremely complex, and traditional fault detection and maintenance methods are often difficult to implement, which makes fault prediction technology particularly important. Effective fault prediction can provide early warning before problems occur, thereby avoiding potential losses and downtime.

[0005] In the field of fault prediction of flexible hands, although there have been certain technological advances, existing methods still have certain limitations and cannot fully cope with complex working environments and fault modes. The current fault prediction system mainly relies on real-time monitoring of the working status of the equipment, and combines data analysis with machine learning algorithms for fault warning. For example, a team at Cornell University has developed a fault prediction system based on real-time data analysis, which can accurately predict the faults of flexible hands and ensure the stability of the equipment. However, this method often faces the risk of reduced prediction accuracy when facing different operating modes, noise interference or unstable working environments, and the adaptability and real-time performance of the system still need to be improved. In addition, the research team of Shanghai Jiaotong University proposed a hybrid method that combines the advantages of data-driven and model-driven to improve the accuracy and robustness of fault prediction. Although this method shows good adaptability in some complex environments, it still has shortcomings in high-dimensional data processing, long model training time and resource consumption. Therefore, it is generally difficult for existing technologies to take into account the requirements of high efficiency, real-time performance and high accuracy, especially in dynamic environments. In response to these technical deficiencies, the present invention proposes a new fault prediction method, which solves the shortcomings of existing technologies in terms of real-time, adaptability and efficiency through innovative models and data processing methods. This method can not only improve the prediction accuracy, but also maintain high robustness in complex environments, thereby 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 a new development path 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 for underwater robot operation, which has a reasonable design, solves the shortcomings of the prior art and has good effects.

[0007] A method for predicting a fault of a flexible hand for underwater robot operation comprises the following steps: S1. Obtain historical fault data collected by the sensor unit, including current, force, torque, position information and vibration signals; S2, preprocessing the data collected by S1, including data enhancement, short-time Fourier transform, overlapping sampling, standardization and normalization; S3, select a balanced sampling selector, increase the sampling proportion of abnormal samples, use the improved random forest algorithm to screen important features, and obtain the screened data set; S4. Building a fault prediction model based on a temporal convolutional network-bidirectional gated recurrent unit neural network, and using the screened data set for training to obtain a trained fault prediction model; S5. Chaos mechanism and cross-mutation are introduced to improve the Sand Cat algorithm to optimize the hyperparameters of the fault prediction model; S6. Obtain an optimized fault prediction model, input the online operation data of the flexible hand into the model, and perform fault prediction.

[0008] Furthermore, in S1, the collected historical fault data is divided into data sets according to the fault type. ; In S2, the data set is enhanced based on the generative adversarial network, and the enhanced data set is ;right Perform short-time Fourier transform to obtain time-frequency data , the data is further enhanced by overlapping sampling, and we get Finally, the data samples are standardized and normalized to unify the dimensions.

[0009] Furthermore, in S3, from the normalized data set Generated by self-service sampling Subsets, each of which contains the same number of samples as the original data set, use a balanced sampling selector to expand the sampling probability of faulty data in the entire data set, and continue to perform balanced sampling after sampling. The threshold is used to determine whether the faulty data in the sample is balanced. The expression is: ; (2) in, 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 , indicating that it is necessary to increase the weight of the fault data ratio in the sample, trigger the second balanced sampler, and then increase the weight of the sample point ratio; After balanced sampling, the data is obtained. ,in Indicates that there is subsets.

[0010] Furthermore, in S3, due to the problem of diversity degradation in random forests, 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: ; (3) in, For the The residuals for the variable values, is the measured value, is the predicted value; Use kernel density to reconstruct the distribution of residuals, the expression is: ); (4) in, is the estimated probability density of the residuals, is the kernel function, is the broadband parameter; is the center value of the kernel function, which is a constant; 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; For the dataset Use the improved random forest algorithm to perform feature screening and obtain the screened data set .

[0011] Further, in S4, in the temporal convolutional network-bidirectional gated recurrent unit neural network, the filtered data set is input into the dilated causal convolutional layer, the batch weight normalization layer, the activation function layer, the LRN layer, the Dropout layer, the convolutional unit, the fully connected layer, the channel and spatial attention fusion unit; The dilated causal convolution layer calculates the feature data of each time step, the batch weight normalization layer normalizes the weights of the feature data, and the activation function layer activates the normalized feature data using the ReLU activation function, which is then input into the LRN layer. The expression is: ; (5) in: It is The corresponding fault image feature map under each channel The activation value of the pixel, It is Channel The activation value of the pixel, After normalization, Channel The output of the pixel, , is a hyperparameter; is the normalized neighborhood size; is the total number of channels in the current feature layer; The normalized output is input into the Dropout layer. The convolution units include 1×1 convolution layer, 2×2 convolution layer and 3×3 convolution layer. The output of the Dropout layer is input into three convolution layers respectively. The input of the three convolution layers is 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, the spatial attention weight is 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 is nonlinearly transformed and concatenated through the fully connected layer. After concatenation, the channel attention weight is generated through the fully connected layer. The channel attention is applied to the feature map. The feature map to which the spatial attention is applied and the feature map to which the channel attention is applied are fused in series to obtain the fused feature map. The temporal convolutional network-bidirectional gated recurrent unit neural network also includes a 1D full convolution layer and an average pooling layer. The filtered data set is sequentially input into the 1D full convolution layer and the average pooling layer. The compressed features output by the average pooling layer are multiplied by the feature maps output by the channel and spatial attention units to obtain the feature maps after the 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 probability of each fault category.

[0012] Furthermore, in said S5, specifically: S5.1, using the improved sand cat algorithm, the independent variables in the sand cat algorithm are the hyperparameters of the neural network, and the fitness function is the residual output of the model; S5.2, introduce chaos mechanism to initialize the population; S5.3. Generate new individuals through crossover operation to enhance population diversity; randomly select a crossover point, cut the genes of the two parent individuals from the crossover point, and exchange some genes; S5.4, introduce new features through mutation operations to increase the diversity of the population and avoid the algorithm falling into the local optimal solution; First, determine the mutation probability, set the mutation probability and determine the frequency of mutation, and then execute the mutation process through the following formula: ; (8) in, To predict the location, is the current location, is the amplitude of variation, is a pseudo-random number generated from a chaotic sequence, ; S5.5, prey search phase, the expression is: ; (9) );(10) in, To represent the updated search radius, it gradually shrinks as the number of iterations increases, controlling the search range. represents the maximum search range, that is, the farthest distance that the sand cat can explore in the current iteration. To indicate the current number of iterations, it 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. It is used to gradually narrow the search range during the iteration process. For the updated sand cat location, is the scaling factor used to adjust the magnitude of the position update, is the current best position, that is, The best position found at iteration , A randomly generated value in the range [0,1] is introduced to increase the diversity of the search. is the current individual's position, i.e. the sand cat is in The current position of the iteration; S5.6, in the attack prey phase, the direction of function value decrease is obtained by calculating the fitness in the iterative process to replace the random direction, and the next step position is calculated. The expression is: ; (11) ; (12) in, It is the middle position, used to adjust the randomness and direction of the position. is the global best position, that is, the best position from the beginning of the algorithm to the current time, is the direction factor, which adjusts the position update direction. The updated position of the sand cat during the attack on prey phase is used to determine the next search direction; The final and main parameters controlling the transition between exploration and exploitation phases are ,when When the sand cat searches for prey, When the sand cat attacks the prey; S5.7. Get a new population, bring the independent variables represented by the population into the fitness function, calculate the fitness value, and determine whether the iteration conditions are met. If not, continue to iterate until the iteration conditions are met to obtain the optimal parameters.

[0013] Furthermore, in S5.2, specifically: Choose different chaotic mapping functions, including Logistic mapping and Tent mapping; Define the interval and divide the number axis [0,1] into artificially set The interval division formula is: , ; (15) Select Initial Value is a small random number, , Logistic mapping and Tent mapping functions are used alternately in each interval, so that the generated chaotic sequence has different nonlinear behaviors in different intervals, increasing the diversity of the initial population. The expression is: ; (16) in, is the chaotic mapping function corresponding to each interval, For the Chaotic map value, For the Chaotic map values; The generated chaotic sequence is normalized: ; (17) in, and are the minimum and maximum values ​​of the sequence respectively.

[0014] Beneficial technical effects brought by the present invention: (1) Improved diagnostic accuracy: By comprehensively analyzing the position, current, vibration and other data of the flexible hand actuator, the present invention combines a temporal convolutional network with a bidirectional gated recurrent unit to achieve early prediction of potential fault points. This method enables maintenance before a fault occurs, significantly reducing the need for emergency repairs, thereby improving the overall operating efficiency and stability of the equipment.

[0015] (2) Optimizing 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, while better adapting to the complexity and diversity of the data.

[0016] (3) Improved model performance: This paper combines a temporal convolutional network and a bidirectional gated recurrent unit to greatly enhance the processing capability of time series data. This comprehensive model demonstrates higher accuracy and efficiency in fault diagnosis and is particularly suitable for complex underwater operating environments.

[0017] (4) Reduce maintenance costs: Through accurate fault prediction, the system can make maintenance work more accurate, reduce unnecessary repairs and inspections, and reduce labor costs and parts replacement costs. This efficient maintenance strategy can significantly save resources and reduce total maintenance costs.

[0018] (5) Improved Sand Cat Algorithm: Chaotic mapping and crossover mutation techniques are introduced into the Sand Cat Algorithm to enhance its global search capability and improve its convergence speed. This improvement not only improves the overall performance of the algorithm, but also optimizes the local and global balance of the model, making fault prediction more accurate and efficient.

[0019] (6) Extending equipment life: By timely identifying and handling potential faults, the equipment can be kept in optimal operating condition, 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.

[0020] (7) Improve real-time performance and response speed: Monitor and analyze data in real time, respond quickly to potential faults, and ensure timely control of equipment status. This improvement in real-time performance and response speed helps to quickly handle faults, reduce downtime, and improve system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a structural diagram of the sensor positions of the flexible hand used for underwater robot operations in the present invention.

[0022] Among them, 1-position sensor, 2-current sensor, 3-vibration sensor, 4-force and torque sensor.

[0023] Figure 2 The figure is a flow chart of data preprocessing in the present invention.

[0024] Figure 3 This is a flow chart of a fault prediction method for a flexible hand used in underwater robot operations. DETAILED DESCRIPTION

[0025] The specific implementation of the present invention is further described below in conjunction with specific embodiments: A method for predicting a fault of a flexible hand for underwater robot operation comprises the following steps: S1. Obtain historical fault data collected by the sensor unit, including current, force, torque, position information and vibration signals; like Figure 1As 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 used for robot operation in real time, and help monitor the stability of the power system; the force and torque sensor is used to capture the force and torque applied by the robot arm, so as to optimize motion control and improve operational stability; the position sensor is used to track the position of the end effector of the robot arm, ensuring 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 robot arm in order to identify potential mechanical failures.

[0026] S2, preprocess the data collected by S1, including data enhancement, short-time Fourier transform, overlapping sampling, standardization and normalization, such as Figure 2 As shown; Divide the collected historical fault data into data sets according to fault type ; Based on the generative adversarial network, the data set is enhanced to transform the real fault data Input to the discriminator, random noise input to 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, and the discriminator tries to distinguish true from false. In this process, the "basis" of the discriminator is not fixed, but dynamically adjusted with training to adapt to the improvement of the generator. Through the above process, in the end, the game between the two prompts the generator to generate more realistic data. The enhanced data set is ; right Perform short-time Fourier transform to obtain time-frequency data ; The data is further enhanced by sampling with overlap. The overlap length should be determined according to the characteristics of the signal and the diagnostic requirements. Through overlapping sampling, a variety of different training samples can be extracted from the original signal. These samples are offset in time but contain similar fault characteristics. This helps the model learn more abundant fault information. For training samples, when collecting training samples from the data, the sampling bands are delineated so that each band signal overlaps with the adjacent band signal. Assume that the enhanced data point is The sampling band is L+offset. L The window size for each sampling is between 10 and 100 data points. offset Controls the amount of overlap between adjacent sampling windows. It indicates the overlap between windows each time the sampling window is moved. Offset The value of L 50% of.

[0027] Then, the corresponding enhanced fault data set after overlapping sampling can be obtained using formula (1): : ; (1) Since the scales and dimensions of the data are inconsistent, the data are finally sampled, standardized and normalized to unify the dimensions.

[0028] S3, select a balanced sampling selector, increase the sampling proportion of abnormal samples, use the improved random forest algorithm to screen important features, and obtain the screened data set; From the normalized dataset Generated by self-service sampling Subsets, each of which contains the same number of samples as the original data set, use a balanced sampling selector to expand the sampling probability of faulty data in the entire data set, and continue to perform balanced sampling after sampling. The threshold is used to determine whether the faulty data in the sample is balanced. The expression is: ; (2) in, 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 , indicating that it is necessary to increase the weight of the fault data ratio in the sample, trigger the second balanced sampler, and then increase the weight of the sample point ratio; After balanced sampling, the data is obtained. ,in Indicates that there is subsets; Since the problem of diversity degradation occurs in random forests, 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. : ; (3) in, For the The residuals for the variable values, is the measured value, is the predicted value; Use kernel density to reconstruct the distribution of residuals, the expression is: ); (4) in, is the estimated probability density of the residuals, is the kernel function, is the broadband parameter; is the center value of the kernel function, which is a constant; 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; For the dataset Use the improved random forest algorithm to perform feature screening and obtain the screened data set .

[0029] S4. Building a fault prediction model based on a temporal convolutional network-bidirectional gated recurrent unit neural network TCN-BiGRU, and training it with the screened data set to obtain a trained fault prediction model; In the temporal convolutional network-bidirectional gated recurrent unit neural network, the filtered data set is input into the dilated causal convolution layer, batch weight normalization layer, activation function layer, LRN layer, Dropout layer, convolution unit, fully connected layer, channel and spatial attention fusion unit; The dilated causal convolution layer calculates the feature data for each time step, and the batch weight normalization layer normalizes the feature data. During the neural network training process, the weights are normalized to stabilize the gradient and accelerate the training process. For each batch of weights, its mean and method are calculated and normalized to stabilize the training process, reduce gradient problems, and accelerate the convergence of the network; 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 neuron activation value at each position, select its local neighborhood (such as n adjacent channels) in the channel dimension and calculate the local normalization coefficient: calculate the square sum on these channels to obtain the normalization coefficient; use the normalization coefficient to normalize the neuron activation value at this position, and its expression is: ; (5) in: It is Channel No. The corresponding fault image feature map under each channel The activation value of the pixel, It is Channel The activation value of the pixel, , After normalization, Channel The output of the pixel, , is a hyperparameter, usually set α= , β=0.75; m is the normalized neighborhood size, that is, the number of surrounding channels considered when normalizing the local response; is the total number of channels in the current feature layer.

[0030] The normalized output is input to the Dropout layer, which randomly discards some neurons in the neural network to reduce overfitting and improve the generalization ability of the model. It sets the activation values ​​of some neurons to zero during training, which prevents the network from over-relying on training data. Define a drop probability p, for example 0.5, which means that each neuron is dropped with a probability of 50% 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 0 with probability p. is 1. In order to keep the outputs consistent during training and prediction, the average value of the activations of neurons discarded during training is reduced. Therefore, during training, the output activations are usually scaled by 1 / (1−p) to keep the expected values ​​consistent during prediction.

[0031] The convolution unit includes a 1×1 convolution layer, a 2×2 convolution layer, and a 3×3 convolution layer. The outputs of the Dropout layer are respectively input into the three convolution layers. The inputs of the three convolution 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, the spatial attention weight is 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 is nonlinearly transformed and concatenated through the fully connected layer. After concatenation, the channel attention weight is generated through the fully connected layer. The channel attention is applied to the feature map. The feature map to which the spatial attention is applied and the feature map to which the channel attention is applied are fused in series to obtain the fused feature map. The temporal convolutional network-bidirectional gated recurrent unit neural network also includes a 1D full convolution layer and an average pooling layer. The filtered data set is input into the 1D full convolution layer and the average pooling layer in sequence. The compressed features output by the average pooling layer are multiplied by the feature maps output by the channel and spatial attention units to obtain the feature map after the attention features. Finally, the input after the attention feature is sent to the fully connected layer and the softmax layer. The softmax layer converts the logits of each category into a probability value through the softmax function and outputs the predicted probability of each fault category. Compare the network output with the expected output, calculate the loss, and update the weights and biases of the model parameters 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 last training, re-set the hyperparameters and start training; The temporal convolutional network-bidirectional gated recurrent unit model is predicted and evaluated using mean squared error (MSE) and mean absolute error (MAE); Calculate the mean squared error (MSE): ; (6) in, is the number of prediction points, It is The true value of a point, is the corresponding predicted value; Calculate the mean absolute error (MAE): (7).

[0032] S5. Introducing chaos mechanism and cross-mutation to improve the Sand Cat algorithm to optimize the hyperparameters of the fault prediction model, including the following sub-steps: S5.1, using the improved Sand Cat ISCSO algorithm, the independent variables in the Sand Cat algorithm are the hyperparameters of the neural network, and the fitness function is the residual output of the model; S5.2, introduce chaos mechanism to initialize the population; S5.3. Generate new individuals through crossover operation to enhance population diversity; randomly select a crossover point, cut the genes of the two parent individuals from the crossover point, and exchange some genes; S5.4, introduce new features through mutation operations to increase the diversity of the population and avoid the algorithm falling into the local optimal solution; First, determine the mutation probability, set the mutation probability and determine the frequency of mutation, and then execute the mutation process through the following formula: ; (8) in, To predict the location, is the current location, is the amplitude of variation, is a pseudo-random number generated from a chaotic sequence, ; S5.5, prey search phase, the expression is: ; (9) );(10) in, To represent the updated search radius, also known as algorithm sensitivity, it gradually decreases as the number of iterations increases. To control the search range, the sensitivity of the sand cat is not fixed. The sensitivity is adjusted using a nonlinear change method 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 a fine search. represents the maximum search range, that is, the farthest distance that the sand cat can explore in the current iteration. To indicate the current number of iterations, it 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. It is used to gradually narrow the search range during the iteration process. For the updated sand cat location, is the scaling factor used to adjust the magnitude of the position update, is the current best position, that is, The best position found at iteration , A randomly generated value in the range [0,1] is introduced to increase the diversity of the search. is the position of the current individual, i.e. the current position of the sand cat at the tth iteration; S5.6, attack prey stage, assume that the sensitivity range of the sand dune cat is a circle, so the direction of movement can be determined by a random angle on the circle, whose value is between -1 and 1. Use the roulette wheel selection algorithm to select a random angle for each sand dune cat to approach the hunting position. In order to avoid falling into the local optimum, a random angle is used. 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 cat involved can get close to the prey. However, the random angle cannot guarantee that the function value will drop, and it has a large randomness. Therefore, the angle at which the function value drops is obtained by calculating the fitness in the iterative process to replace the random angle. And calculate the next step position, the expression is: ; (11) ; (12) in, It is the middle position, used to adjust the randomness and direction of the position. is the global best position, that is, the best position from the beginning of the algorithm to the current time, is the direction factor, which adjusts the position update direction. The updated position of the sand cat during the attack on prey phase is used to determine the next search direction; The final and main parameters controlling the transition between exploration and exploitation phases are ,when When the sand cat searches for prey, When the sand cat attacks the prey; S5.7. Get a new group of populations, bring the independent variables represented by the population into the fitness function, calculate the fitness value, and determine whether the iteration condition is met, that is, the convergence threshold is reached. If not, continue to iterate until the iteration condition is met and the optimal parameters are obtained.

[0033] S5.2 specifically states: Choose different chaotic mapping functions, including Logistic mapping and Tent mapping; Logistic Mapping: ; (13) in, is a control parameter, and its value is usually between (3.57, 4).

[0034] Tent Mapping: ; (14) in is a control parameter, usually with a value between (0, 1); Define the interval and divide the number axis [0,1] into artificially set The expression for dividing the intervals is: , ; (15) in, is the number of independent variables in a single population in the sand cat algorithm; Select Initial Value is a small random number. In order to ensure that the chaotic mapping function can generate a sequence with good randomness and avoid the sequence falling into periodicity or fixed points, , Logistic mapping and Tent mapping functions are used alternately in each interval, so that the generated chaotic sequence has different nonlinear behaviors in different intervals, increasing the diversity of the initial population. The expression is: ; (16) in, is the chaotic mapping function corresponding to each interval, For the Chaotic map value, For the Chaotic map values; The generated chaotic sequence may not be within the range required by the algorithm, so a normalization operation is required to ensure that the sequence meets the requirements of the initialization population. The expression is: ; (17) in, and are the minimum and maximum values ​​of the sequence respectively.

[0035] S6. Obtain an optimized fault prediction model, input the online operation data of the flexible hand into the model, and perform fault prediction.

[0036] 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 technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting faults of a flexible hand for underwater robot operation, characterized in that: The following steps are involved: S1. Obtain historical fault data collected by the sensor unit, including current, force, torque, position information and vibration signals; S2, preprocessing the data collected by S1, including data enhancement, short-time Fourier transform, overlapping sampling, standardization and normalization; S3, select a balanced sampling selector, increase the sampling proportion of abnormal samples, use the improved random forest algorithm to screen important features, and obtain the screened data set; S4. Building a fault prediction model based on a temporal convolutional network-bidirectional gated recurrent unit neural network, and using the screened data set for training to obtain a trained fault prediction model; S5. Chaos mechanism and cross-mutation are introduced to improve the Sand Cat algorithm to optimize the hyperparameters of the fault prediction model; S6. Obtain an optimized fault prediction model, input the online operation data of the flexible hand into the model, and perform fault prediction.

2. A method for predicting faults of a flexible hand for underwater robot operation according to claim 1, characterized in that: In S1, the collected historical fault data is divided into data sets according to the fault type. ; In S2, the data set is enhanced based on the generative adversarial network, and the enhanced data set is ;right Perform short-time Fourier transform to obtain time-frequency data , the data is further enhanced by overlapping sampling, and we get Finally, the data samples are standardized and normalized to unify the dimensions.

3. A method for predicting faults of a flexible hand for underwater robot operation according to claim 2, characterized in that: In S3, the normalized dataset Generated by self-service sampling There are subsets, each of which contains the same number of samples as the original data set. A balanced sampling selector is used to expand the sampling probability of faulty data in the entire data set. Balanced sampling is continued after sampling, and the threshold is used to determine whether the faulty data in the sample is balanced. The expression is: ； (2) in, 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 , indicating that it is necessary to increase the weight of the fault data ratio in the sample, trigger the second balanced sampler, and then increase the weight of the sample point ratio; After balanced sampling, the data is obtained. ,in Indicates that there is subsets.

4. A method for predicting faults of a flexible hand for underwater robot operation according to claim 3, characterized in that: In S3, due to the problem of diversity degradation in random forests, 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: ；(3) in, For the The residuals for the variable values, is the measured value, is the predicted value; Use kernel density to reconstruct the distribution of residuals, the expression is: )；(4) in, is the estimated probability density of the residuals, is the kernel function, is the broadband parameter; is the center value of the kernel function, which is a constant; 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; For the dataset Use the improved random forest algorithm to perform feature screening and obtain the screened data set .

5. A method for predicting faults of a flexible hand for underwater robot operation according to claim 4, characterized in that: In S4, in the temporal convolutional network-bidirectional gated recurrent unit neural network, the filtered data set is input into the dilated causal convolution layer, the batch weight normalization layer, the activation function layer, the LRN layer, the Dropout layer, the convolution unit, the fully connected layer, the channel and spatial attention fusion unit; The dilated causal convolution layer calculates the feature data of each time step, the batch weight normalization layer normalizes the weights of the feature data, and the activation function layer activates the normalized feature data using the ReLU activation function, which is then input into the LRN layer. The expression is: ；(5) in: It is The corresponding fault image feature map under each channel The activation value of the pixel, It is Channel The activation value of the pixel, After normalization, Channel The output of the pixel, , is a hyperparameter; is the normalized neighborhood size; is the total number of channels in the current feature layer; The normalized output is input into the Dropout layer. The convolution units include 1×1 convolution layer, 2×2 convolution layer and 3×3 convolution layer. The output of the Dropout layer is input into three convolution layers respectively. The input of the three convolution layers is 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, the spatial attention weight is 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 is nonlinearly transformed and concatenated through the fully connected layer. After concatenation, the channel attention weight is generated through the fully connected layer. The channel attention is applied to the feature map. The feature map to which the spatial attention is applied and the feature map to which the channel attention is applied are fused in series to obtain the fused feature map. The temporal convolutional network-bidirectional gated recurrent unit neural network also includes a 1D full convolution layer and an average pooling layer. The filtered data set is sequentially input into the 1D full convolution layer and the average pooling layer. The compressed features output by the average pooling layer are multiplied by the feature maps output by the channel and spatial attention units to obtain the feature maps after the 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 probability of each fault category.

6. A method for predicting faults of a flexible hand for underwater robot operation according to claim 5, characterized in that: In the S5, specifically: S5.1, using the improved sand cat algorithm, the independent variables in the sand cat algorithm are the hyperparameters of the neural network, and the fitness function is the residual output of the model; S5.2, introduce chaos mechanism to initialize the population; S5.

3. Generate new individuals through crossover operation to enhance population diversity; randomly select a crossover point, cut the genes of the two parent individuals from the crossover point, and exchange some genes; S5.4, introduce new features through mutation operations to increase the diversity of the population and avoid the algorithm falling into the local optimal solution; First, determine the mutation probability, set the mutation probability and determine the frequency of mutation, and then execute the mutation process through the following formula: ； (8) in, To predict the location, is the current location, is the amplitude of variation, is a pseudo-random number generated from a chaotic sequence, ; S5.5, prey search phase, the expression is: ；(9) )；(10) in, To represent the updated search radius, it gradually shrinks as the number of iterations increases, controlling the search range. represents the maximum search range, that is, the farthest distance that the sand cat can explore in the current iteration. To indicate the current number of iterations, it is used to dynamically adjust the algorithm parameters. is the maximum number of iterations, indicating the maximum number of iterations of the algorithm, which is used to gradually narrow the search range during the iteration process. For the updated sand cat location, is the scaling factor used to adjust the magnitude of the position update, is the current best position, that is, The best position found at iteration , A randomly generated value in the range [0,1] is introduced to increase the diversity of the search. is the current individual's position, i.e. the sand cat is in The current position of the iteration; S5.6, in the attack prey phase, the direction of function value decrease is obtained by calculating the fitness in the iterative process to replace the random direction, and the next step position is calculated. The expression is: ；(11) ；(12) in, It is the middle position, used to adjust the randomness and direction of the position. is the global best position, that is, the best position from the beginning of the algorithm to the current time, is the direction factor, which adjusts the position update direction. The updated position of the sand cat during the attack on prey phase is used to determine the next search direction; The final and main parameters controlling the transition between exploration and exploitation phases are ,when When the sand cat searches for prey, When the sand cat attacks the prey; S5.

7. Get a new population, bring the independent variables represented by the population into the fitness function, calculate the fitness value, and determine whether the iteration conditions are met. If not, continue to iterate until the iteration conditions are met to obtain the optimal parameters.

7. A method for predicting faults of a flexible hand for underwater robot operation according to claim 6, characterized in that: In S5.2, specifically: Choose different chaotic mapping functions, including Logistic mapping and Tent mapping; Define the interval and divide the number axis [0,1] into artificially set The interval division formula is: ， ；(15) Select Initial Value is a small random number, , Logistic mapping and Tent mapping functions are used alternately in each interval, so that the generated chaotic sequence has different nonlinear behaviors in different intervals, increasing the diversity of the initial population. The expression is: ；(16) in, is the chaotic mapping function corresponding to each interval, For the Chaotic map value, For the Chaotic map values; The generated chaotic sequence is normalized: ；(17) in, and are the minimum and maximum values ​​of the sequence respectively.

Images