Deep learning based underwater robot state analysis system

Abstract

The application relates to the technical field of underwater robot control, in particular to an underwater robot state analysis system based on deep learning, which comprises a data acquisition module used for acquiring a sonar image sequence, a three-axis acceleration sequence, a three-axis velocity sequence, a thruster feedback current value sequence and a thruster instruction value sequence; a data preprocessing module used for obtaining a sonar image tensor matrix sequence, a first motion state vector sequence, a second motion state vector sequence, a first running state vector sequence and a second running state vector sequence; a deep learning module used for calculating a water flow disturbance characteristic value vector when it is judged that an underwater robot deviates from an expected motion state; and an identification module used for judging whether the underwater robot deviates from the expected motion state due to water flow disturbance according to the water flow disturbance characteristic value vector. The underwater robot external water flow disturbance state identification is realized through deep learning.

Description

Technical Field

[0001] This invention relates to the field of underwater robot control technology, specifically an underwater robot state analysis system based on deep learning. Background Technology

[0002] Underwater robots are widely used in marine resource exploration, underwater operations, and environmental monitoring. During actual operations, underwater robots are typically equipped with fiber optic inertial navigation systems, Doppler logs, sonar devices, and other detection equipment to acquire information about their motion and the surrounding environment. In complex waters, turbulent currents and other disturbances can cause the underwater robot's motion to deviate from its intended state due to external forces. In such cases, operators should steer the underwater robot away from the affected area. However, the reasons for this deviation can be varied. Not only can current disturbances cause deviations, but malfunctions in the underwater robot's control system and propulsion systems can also lead to deviations.

[0003] However, existing technologies can typically only detect whether an underwater robot deviates from its expected motion state, but cannot accurately identify the cause of this deviation. This makes it difficult for operators to distinguish whether the deviation is caused by external water flow disturbances or a malfunction in the equipment itself, hindering the implementation of targeted countermeasures. Blindly withdrawing the robot reduces operational efficiency, while ignoring the malfunction could lead to more serious safety accidents. With the development of deep learning technology, attribution and root cause analysis using multi-dimensional data has become possible, providing a feasible path to solving this technical problem.

[0004] Therefore, it is necessary to develop an underwater robot state analysis system based on deep learning. After determining that the underwater robot has deviated from the expected motion state, the system uses deep learning algorithms to analyze the data, thereby accurately determining whether the underwater robot has deviated from the expected motion state due to water flow disturbance. Summary of the Invention

[0005] (1) Technical problems to be solved The purpose of this invention is to provide a deep learning-based underwater robot state analysis system to accurately determine whether an underwater robot deviates from its expected motion state due to water flow disturbance.

[0006] (2) Technical solution To achieve the above objectives, the present invention provides a deep learning-based underwater robot state analysis system, the system comprising: The data acquisition module is used to acquire sonar image sequences, triaxial acceleration sequences, triaxial velocity sequences, thruster feedback current value sequences, and thruster command value sequences.

[0007] The data preprocessing module, connected to the data acquisition module, is used to map the sonar image sequence into a sonar image tensor matrix sequence; perform time alignment and normalization on the triaxial acceleration sequence and triaxial velocity sequence to obtain a first motion state vector sequence and a second motion state vector sequence; and perform time alignment and normalization on the thruster feedback current value sequence and thruster command value sequence to obtain a first operating state vector sequence and a second operating state vector sequence, respectively.

[0008] The deep learning module, connected to the data preprocessing module, is used to determine whether the underwater robot deviates from the expected motion state based on the first motion state vector sequence and the second motion state vector sequence. When it is determined that the underwater robot deviates from the expected motion state, visual features reflecting the water flow disturbance texture are extracted based on the sonar image tensor matrix sequence, and the water flow disturbance identification results reflecting the effect of the thruster motor working current control are extracted based on the first and second running state vector sequences, i.e., running features. The visual features and running features are fused to obtain the water flow disturbance feature value vector.

[0009] The identification module, connected to the deep learning module, is used to compare the water flow disturbance feature vector with the pre-set first identification threshold vector and second identification threshold vector to determine whether the underwater robot has deviated from the expected motion state due to water flow disturbance.

[0010] Furthermore, the data acquisition module includes: The sonar image sequence acquisition module is used to acquire sonar image sequences within a pre-set time window and at a pre-set first sampling frequency using a sonar device mounted on an underwater robot.

[0011] The three-axis acceleration sequence acquisition module is connected to the sonar image sequence acquisition module. It is used to acquire the three-axis acceleration sequence through the fiber optic inertial navigation equipment carried by the underwater robot within a preset time window and at a preset first sampling frequency. The three-axis acceleration sequence represents the acceleration component sequence in the X-axis, Y-axis and Z-axis directions in a preset three-dimensional spatial coordinate system. The three-dimensional spatial coordinate system is composed of mutually perpendicular X-axis, Y-axis and Z-axis.

[0012] The three-axis velocity sequence acquisition module, connected to the three-axis acceleration sequence acquisition module, is used to acquire the three-axis velocity sequence within a preset time window and at a preset first sampling frequency using a Doppler log on the underwater robot; the three-axis velocity sequence represents the velocity component sequence in the X-axis, Y-axis and Z-axis directions in a preset three-dimensional spatial coordinate system.

[0013] A thruster feedback current value sequence acquisition module, connected to a three-axis velocity sequence acquisition module, is used to acquire the thruster feedback current value sequence via a current sensor within a preset time window at a preset first sampling frequency. The thruster feedback current value sequence represents the effective value sequence of the actual operating current of the underwater robot's thruster motors. The number of thrusters in the underwater robot is denoted as... The thrusters were numbered from the first thruster to the second. Thrusters; respectively propelling the first thruster to the second thruster The sequence of thruster feedback current values ​​is denoted as the first thruster feedback current value sequence up to the second. Thruster feedback current value sequence.

[0014] A thruster command value sequence acquisition module, connected to a thruster feedback current value sequence acquisition module, is used to read the thruster command value sequence from the underwater robot control system within a preset time window at a preset first sampling frequency; the thruster command value sequence represents the command current value sent by the underwater robot control system to the thruster motor; the sequence is used to acquire the first thruster command value sequence to the second thruster command value sequence. The thruster command value sequence of the thrusters is denoted as the first thruster command value sequence to the second. Thruster command value sequence.

[0015] Furthermore, the data preprocessing module includes: The pixel extraction module is used to convert each frame of the sonar image in the sonar image sequence into a sonar image tensor matrix using a pixel aggregation method; and to stack the sonar image tensor matrices in the order of acquisition time to obtain a sonar image tensor matrix sequence.

[0016] The pixel aggregation method is as follows: extract the RGB color values ​​of all pixels in each frame of the sonar image sequence, and convert them into grayscale values ​​using an RGB color value weighted average algorithm; divide each frame of the sonar image sequence into pre-set number of rows. and number of split columns Divided into Each grid has an equal size; the average gray value of all pixels in each grid is extracted as the representative gray value of that grid; the... The representative gray values ​​of each raster cell are arranged in row and column order to form a raster. OK The column is a two-dimensional matrix, denoted as the sonar image tensor matrix.

[0017] Furthermore, the data preprocessing module also includes: The motion state data preprocessing module, connected to the pixel extraction module, is used to time-align the triaxial acceleration sequence and triaxial velocity sequence, and normalize them using the maximum-minimum normalization method to obtain normalized triaxial acceleration and velocity sequences. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial acceleration sequence are combined to obtain a first motion state vector. All first motion state vectors within a time window are combined to obtain a first motion state vector sequence. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial velocity sequence are combined to obtain a second motion state vector. All second motion state vectors within a time window are combined to obtain a second motion state vector sequence.

[0018] Furthermore, the data preprocessing module also includes: The operation status data preprocessing module, connected to the motion status data preprocessing module, is used to sequence the feedback current values ​​of the first thruster to the first... Thruster feedback current value sequence, first thruster command value sequence to the... The thruster command value sequence is time-aligned and normalized using the maximum-minimum normalization method to obtain the normalized first thruster feedback current value sequence up to the normalized second... The sequence of thruster feedback current values, the normalized first thruster command value sequence, and the normalized second thruster command value sequence. Thruster command value sequence; normalize the first thruster feedback current value corresponding to the same moment to the normalized second thruster command value. The thruster feedback current values ​​are combined to obtain a first operating state vector; all first operating state vectors within the time window are arranged in chronological order to obtain a first operating state vector sequence; the normalized first thruster command value corresponding to the same moment is then assigned to the normalized first operating state vector sequence. The thruster command values ​​are combined to obtain a second operating state vector; all the second operating state vectors within the time window are arranged in chronological order to obtain a second operating state vector sequence.

[0019] Furthermore, the deep learning module includes: A motion state deviation identification module is used to construct an acceleration reference vector. The acceleration reference vector consists of pre-defined X-axis acceleration reference values, Y-axis acceleration reference values, and Z-axis acceleration reference values. Each first motion state vector in the first motion state vector sequence is subtracted from the acceleration reference vector, and the absolute value is taken to obtain an acceleration deviation vector sequence. The modulus of all acceleration deviation vectors in the acceleration deviation vector sequence is taken to obtain an acceleration deviation modulus sequence. A velocity reference vector is constructed. The velocity reference vector consists of pre-defined X-axis velocity reference values, Y-axis velocity reference values, and Z-axis velocity reference values. Each second motion state vector in the second motion state vector sequence is subtracted from the velocity reference vector, and the absolute value is taken to obtain a velocity deviation vector sequence. The modulus of all velocity deviation vectors in the velocity deviation vector sequence is taken to obtain a velocity deviation modulus sequence. If any acceleration deviation modulus in the acceleration deviation modulus sequence is greater than a pre-set first trigger threshold or any velocity deviation modulus in the velocity deviation modulus sequence is greater than a pre-set second trigger threshold, then the underwater robot is determined to have deviated from the expected motion state.

[0020] Furthermore, the deep learning module also includes: The turbulent water flow texture recognition module, connected to the motion state deviation identification module, is used to input the tensor matrix sequence of the sonar image into a pre-trained three-dimensional convolutional neural network to extract visual features reflecting water flow disturbance texture in the sonar image, and map them into X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index; the X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index are combined to obtain a water flow disturbance texture index vector.

[0021] The training method for the three-dimensional convolutional neural network is as follows: A training sample set of water flow disturbance textures is obtained. Each training sample in the water flow disturbance texture training sample set includes a pre-acquired sequence of historical sonar image tensor matrices and a corresponding first real water flow disturbance vector. The first real water flow disturbance vector is contained within the time interval of the historical sonar image tensor matrix sequence acquisition, and includes real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions. The real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions respectively represent the underwater machine. The normalized water flow disturbance intensity at the location of the person is measured in the X, Y, and Z axes. This normalized water flow disturbance intensity is obtained by dimensionless processing of the components of the water flow velocity measured by the Doppler current meter in the X, Y, and Z axes. The sequence of tensor matrices of the historical sonar images is input into the three-dimensional convolutional neural network to be trained to obtain the predicted water flow disturbance texture index vector. The error between the predicted water flow disturbance texture index vector and the actual water flow disturbance vector is calculated. The parameters of the three-dimensional convolutional neural network are updated through the backpropagation algorithm until the error converges, resulting in a trained three-dimensional convolutional neural network.

[0022] Furthermore, the deep learning module also includes: The operation state deviation identification module, connected to the turbulent water flow texture identification module, is used to subtract the first and second operation state vectors corresponding to the same moment in the first and second operation state vector sequences to obtain an operation deviation vector sequence; each operation deviation vector in the operation deviation vector sequence contains elements from the first thruster operation deviation value to the... Thruster operating deviation value; respectively, the operating deviation value of the first thruster up to the... Statistical analysis was performed on the sequence of thruster operating deviation values ​​within a time window to extract the mean operating deviation, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding a preset deviation threshold for each thruster; the first thruster to the second thruster were then analyzed. The mean operating deviation of the thruster, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding the preset deviation threshold are combined in order of thruster number to obtain the operating state deviation feature vector. The operating state deviation feature vector is input into a pre-trained support vector regression model to extract the water flow disturbance identification results reflected by the thruster motor operating current control effect, and mapped to X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index. The X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index are combined to obtain the water flow disturbance current index vector.

[0023] The training method for the support vector regression model is as follows: A training sample set of water flow disturbance current is obtained. Each training sample in the water flow disturbance current training sample set includes: a pre-acquired historical operating state deviation feature vector and a corresponding second true water flow disturbance vector. The second true water flow disturbance vector contains the true values ​​of water flow disturbance in the X-axis direction, Y-axis direction, and Z-axis direction within the time interval from which the historical operating state deviation feature vector was collected. The historical operating state deviation feature vector is used as input, and the second true water flow disturbance vector is used as output. The support vector regression algorithm is used to train the model. The parameters of the support vector regression model are optimized by minimizing the prediction error to obtain the trained support vector regression model.

[0024] Furthermore, the deep learning module also includes: The weighting module, connected to the operating state deviation identification module, is used to calculate the water flow disturbance feature value vector using a three-dimensional weighting formula; the three-dimensional weighting formula is: ; in, This represents the eigenvalue vector of water flow disturbance. , , These represent the characteristic values ​​of water flow disturbance along the X-axis, Y-axis, and Z-axis, respectively. This represents the texture index vector for water flow disturbance. , , These represent the X-axis water flow disturbance texture index, the Y-axis water flow disturbance texture index, and the Z-axis water flow disturbance texture index, respectively. This represents the index vector of water flow disturbance current. , , These represent the X-axis water flow disturbance current index, the Y-axis water flow disturbance current index, and the Z-axis water flow disturbance current index, respectively. This represents the pre-set weighting, with a value ranging from 0 to 1.

[0025] Furthermore, the identification module includes: The comparison module is used to compare the water flow disturbance feature value vector with the preset first identification threshold vector and the second identification threshold vector. If any element in the water flow disturbance feature value vector is greater than the corresponding element in the preset first identification threshold vector or less than the corresponding element in the preset second identification threshold vector, it is determined that the underwater robot has deviated from the expected motion state due to water flow disturbance.

[0026] (3) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: By combining visual features from sonar images with operational features from thruster current data for water flow disturbance attribution analysis, the system can distinguish whether the underwater robot's deviation from the expected motion state is caused by external water flow disturbance. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the module composition of the underwater robot state analysis system based on deep learning according to Embodiment 1 of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Before giving examples, it is necessary to describe the application scenario of the present invention. The present invention is used to determine whether the deviation of an underwater robot from its expected motion state in complex open water is caused by external water flow disturbance.

[0030] Example 1: As Figure 1 As shown, this embodiment provides a deep learning-based underwater robot state analysis system, the system comprising: The data acquisition module is used to acquire sonar image sequences, triaxial acceleration sequences, triaxial velocity sequences, thruster feedback current value sequences, and thruster command value sequences.

[0031] The data preprocessing module, connected to the data acquisition module, is used to map the sonar image sequence into a sonar image tensor matrix sequence; perform time alignment and normalization on the triaxial acceleration sequence and triaxial velocity sequence to obtain a first motion state vector sequence and a second motion state vector sequence; and perform time alignment and normalization on the thruster feedback current value sequence and thruster command value sequence to obtain a first operating state vector sequence and a second operating state vector sequence, respectively.

[0032] The deep learning module, connected to the data preprocessing module, is used to determine whether the underwater robot deviates from the expected motion state based on the first motion state vector sequence and the second motion state vector sequence. When it is determined that the underwater robot deviates from the expected motion state, visual features reflecting the water flow disturbance texture are extracted based on the sonar image tensor matrix sequence, and the water flow disturbance identification results reflecting the effect of the thruster motor working current control are extracted based on the first and second running state vector sequences, i.e., running features. The visual features and running features are fused to obtain the water flow disturbance feature value vector.

[0033] The identification module, connected to the deep learning module, is used to compare the water flow disturbance feature vector with the pre-set first identification threshold vector and second identification threshold vector to determine whether the underwater robot has deviated from the expected motion state due to water flow disturbance.

[0034] For example, the data acquisition module is used to acquire multi-source sensor data of the underwater robot during operation. In practical applications, within a pre-set time window of 50 seconds, the underwater robot acquires 500 frames of sonar images using its onboard sonar equipment at a pre-set first sampling frequency of 10Hz. Simultaneously, it acquires a three-axis acceleration sequence using a fiber optic inertial navigation system at the same first sampling frequency, a three-axis velocity sequence using a Doppler log at the same first sampling frequency, a thruster feedback current value sequence using a current sensor at the same first sampling frequency, and a thruster command value sequence read from the underwater robot control system at the same first sampling frequency. The thruster feedback current value sequence reflects the actual effective value of the operating current of each thruster motor of the underwater robot, and the thruster command value sequence reflects the command current value sent by the control system to each thruster motor. For an underwater robot equipped with four thrusters, the thruster feedback current value sequence includes the first thruster feedback current value sequence to the fourth thruster feedback current value sequence, and the thruster command value sequence includes the first thruster command value sequence to the fourth thruster command value sequence.

[0035] The data preprocessing module is connected to the data acquisition module and is used to convert the acquired raw data into a data format suitable for processing by the deep learning module. For the sonar image sequence, a pixel aggregation method is used to convert each frame of the sonar image into a sonar image tensor matrix. Specifically, taking the first frame of the acquired sonar image as an example, the RGB color values ​​of all pixels in the image are extracted and converted into grayscale values ​​using an RGB color value weighted average algorithm. The frame of the sonar image is divided into 50×50 equally sized grids according to a pre-set number of rows and columns. The average grayscale value of all pixels in each grid is extracted as the representative grayscale value of that grid. The representative grayscale values ​​of 2500 grids are arranged in the row and column order of the grids to form a 50x50 two-dimensional matrix, which is the sonar image tensor matrix corresponding to the frame of the sonar image. The sonar image tensor matrices corresponding to 500 frames of sonar images are stacked in the order of acquisition time to obtain the sonar image tensor matrix sequence. For the triaxial acceleration and velocity sequences, time alignment is first performed to ensure that acceleration and velocity data at the same moment correspond. Then, the maximum-minimum normalization method is used for normalization, resulting in normalized triaxial acceleration and velocity sequences. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial acceleration sequence are combined to obtain a first motion state vector. For example, if the normalized acceleration components at a certain moment are 0.05, 0, and 0, then the first motion state vector is formed. The first motion state vectors at all moments within a 50-second time window are arranged chronologically to obtain a sequence of 500 first motion state vectors. Similarly, the X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized three-axis velocity sequence are combined to obtain the second motion state vector, and thus the second motion state vector sequence. For the sequences of first thruster feedback current values ​​to fourth thruster feedback current values, and the sequences of first thruster command values ​​to fourth thruster command values, time alignment and normalization are also performed to obtain normalized sequences of first thruster feedback current values ​​to fourth thruster feedback current values, and normalized sequences of first thruster command values ​​to fourth thruster command values. The normalized first thruster feedback current values ​​to fourth thruster feedback current values ​​corresponding to the same moment are combined to obtain a first operating state vector. The first operating state vectors at all moments within the time window are arranged chronologically to obtain the first operating state vector sequence. Similarly, the normalized first thruster command values ​​to the fourth thruster command values ​​corresponding to the same moment are combined to obtain the second operating state vector, and then the second operating state vector sequence is obtained. To ensure the consistency of the calculation, the maximum and minimum value parameters in the maximum-minimum value normalization method are fixed values ​​given in advance.

[0036] The deep learning module is connected to the data preprocessing module. First, it determines whether the underwater robot deviates from the expected motion state based on the first and second motion state vector sequences. Specifically, X-axis acceleration reference values, Y-axis acceleration reference values, and Z-axis acceleration reference values ​​are pre-set to be 0, 0, and 0, respectively. Each first motion state vector in the first motion state vector sequence is subtracted from the acceleration reference vector, and the absolute value is taken to obtain an acceleration deviation vector sequence. Then, the modulus of each acceleration deviation vector is taken to obtain an acceleration deviation modulus sequence. Similarly, X-axis velocity reference values, Y-axis velocity reference values, and Z-axis velocity reference values ​​are pre-set, for example, to be 0.2, 0, and 0, respectively. Each second motion state vector in the second motion state vector sequence is subtracted from the velocity reference vector, and the absolute value is taken to obtain a velocity deviation vector sequence. Then, the modulus of each velocity deviation vector is taken to obtain a velocity deviation modulus sequence. If any acceleration deviation magnitude in the acceleration deviation magnitude sequence exceeds a pre-set first trigger threshold (e.g., 0.1), or any velocity deviation magnitude in the velocity deviation magnitude sequence exceeds a pre-set second trigger threshold (e.g., 0.3), the underwater robot is determined to have deviated from its expected motion state. The first and second trigger thresholds are positively correlated. Once the underwater robot is determined to have deviated from its expected motion state, the deep learning module initiates subsequent water flow disturbance attribution analysis. Based on the sonar image tensor matrix sequence, a pre-trained deep learning model extracts visual features reflecting the water flow disturbance texture. These visual features are represented by X-axis water flow disturbance texture indices, Y-axis water flow disturbance texture indices, and Z-axis water flow disturbance texture indices, which are combined to obtain a water flow disturbance texture index vector. The X-axis, Y-axis, and Z-axis water flow disturbance texture indices are all dimensionless values ​​ranging from -1 to 1. Simultaneously, based on the first and second operating state vector sequences, the identification results of water flow disturbances reflected by the propeller motor's operating current control effect are extracted, i.e., operating features. These operating features are manifested as X-axis water flow disturbance current indices, Y-axis water flow disturbance current indices, and Z-axis water flow disturbance current indices, which are combined to obtain a water flow disturbance current index vector. The X-axis, Y-axis, and Z-axis water flow disturbance current indices are all dimensionless values ​​ranging from -1 to 1. Then, a weighted fusion method is used to weight and sum the water flow disturbance texture index vector and the water flow disturbance current index vector to obtain a three-dimensional water flow disturbance feature value vector, with each dimension corresponding to the water flow disturbance feature value in the X-axis, Y-axis, and Z-axis directions, respectively.

[0037] The identification module is connected to the deep learning module and is used to ultimately determine whether the underwater robot's deviation from the expected motion state is caused by water flow disturbance. The water flow disturbance feature value vector calculated by the deep learning module is compared with a pre-set identification threshold vector. The identification threshold vector includes X-axis identification threshold, Y-axis threshold, and Z-axis threshold; the first identification threshold vector is set to... The second identification threshold vector is If any element in the water flow disturbance feature vector is greater than 0.5 or less than -0.5, then the underwater robot is determined to have deviated from its expected motion state due to water flow disturbance.

[0038] Furthermore, the data acquisition module includes: The sonar image sequence acquisition module is used to acquire sonar image sequences within a pre-set time window and at a pre-set first sampling frequency using a sonar device mounted on an underwater robot.

[0039] The three-axis acceleration sequence acquisition module is connected to the sonar image sequence acquisition module. It is used to acquire the three-axis acceleration sequence through the fiber optic inertial navigation equipment carried by the underwater robot within a preset time window and at a preset first sampling frequency. The three-axis acceleration sequence represents the acceleration component sequence in the X-axis, Y-axis and Z-axis directions in a preset three-dimensional spatial coordinate system. The three-dimensional spatial coordinate system is composed of mutually perpendicular X-axis, Y-axis and Z-axis.

[0040] The three-axis velocity sequence acquisition module, connected to the three-axis acceleration sequence acquisition module, is used to acquire the three-axis velocity sequence within a preset time window and at a preset first sampling frequency using a Doppler log on the underwater robot; the three-axis velocity sequence represents the velocity component sequence in the X-axis, Y-axis and Z-axis directions in a preset three-dimensional spatial coordinate system.

[0041] A thruster feedback current value sequence acquisition module, connected to a three-axis velocity sequence acquisition module, is used to acquire the thruster feedback current value sequence via a current sensor within a preset time window at a preset first sampling frequency. The thruster feedback current value sequence represents the effective value sequence of the actual operating current of the underwater robot's thruster motors. The number of thrusters in the underwater robot is denoted as... The thrusters were numbered from the first thruster to the second. Thrusters; respectively propelling the first thruster to the second thruster The sequence of thruster feedback current values ​​is denoted as the first thruster feedback current value sequence up to the second. Thruster feedback current value sequence.

[0042] A thruster command value sequence acquisition module, connected to a thruster feedback current value sequence acquisition module, is used to read the thruster command value sequence from the underwater robot control system within a preset time window at a preset first sampling frequency; the thruster command value sequence represents the command current value sent by the underwater robot control system to the thruster motor; the sequence is used to acquire the first thruster command value sequence to the second thruster command value sequence. The thruster command value sequence of the thrusters is denoted as the first thruster command value sequence to the second. Thruster command value sequence.

[0043] Furthermore, the data preprocessing module includes: The pixel extraction module is used to convert each frame of the sonar image in the sonar image sequence into a sonar image tensor matrix using a pixel aggregation method; and to stack the sonar image tensor matrices in the order of acquisition time to obtain a sonar image tensor matrix sequence.

[0044] The pixel aggregation method is as follows: extract the RGB color values ​​of all pixels in each frame of the sonar image sequence, and convert them into grayscale values ​​using an RGB color value weighted average algorithm; divide each frame of the sonar image sequence into pre-set number of rows. and number of split columns Divided into Each grid has an equal size; the average gray value of all pixels in each grid is extracted as the representative gray value of that grid; the... The representative gray values ​​of each raster cell are arranged in row and column order to form a raster. OK The column is a two-dimensional matrix, denoted as the sonar image tensor matrix.

[0045] For example, the sonar images in the sonar image sequence can be either color images or grayscale images. For color images, the RGB color values ​​of the pixels are converted to grayscale values ​​using an RGB color value weighted averaging algorithm. For grayscale images, the RGB color values ​​of the pixels are also converted to grayscale values ​​using an RGB color value weighted averaging algorithm, because the grayscale standard may differ for different grayscale images, and converting to grayscale values ​​using an RGB color value weighted averaging algorithm can improve the algorithm's versatility. In the RGB color value weighted averaging algorithm, the weight of the R channel is 0.3, the weight of the G channel is 0.59, and the weight of the B channel is 0.11.

[0046] Furthermore, the data preprocessing module also includes: The motion state data preprocessing module, connected to the pixel extraction module, is used to time-align the triaxial acceleration sequence and triaxial velocity sequence, and normalize them using the maximum-minimum normalization method to obtain normalized triaxial acceleration and velocity sequences. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial acceleration sequence are combined to obtain a first motion state vector. All first motion state vectors within a time window are combined to obtain a first motion state vector sequence. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial velocity sequence are combined to obtain a second motion state vector. All second motion state vectors within a time window are combined to obtain a second motion state vector sequence.

[0047] Furthermore, the data preprocessing module also includes: The operation status data preprocessing module, connected to the motion status data preprocessing module, is used to sequence the feedback current values ​​of the first thruster to the first... Thruster feedback current value sequence, first thruster command value sequence to the... The thruster command value sequence is time-aligned and normalized using the maximum-minimum normalization method to obtain the normalized first thruster feedback current value sequence up to the normalized second... The sequence of thruster feedback current values, the normalized first thruster command value sequence, and the normalized second thruster command value sequence. Thruster command value sequence; normalize the first thruster feedback current value corresponding to the same moment to the normalized second thruster command value. The thruster feedback current values ​​are combined to obtain a first operating state vector; all first operating state vectors within the time window are arranged in chronological order to obtain a first operating state vector sequence; the normalized first thruster command value corresponding to the same moment is then assigned to the normalized first operating state vector sequence. The thruster command values ​​are combined to obtain a second operating state vector; all the second operating state vectors within the time window are arranged in chronological order to obtain a second operating state vector sequence.

[0048] Furthermore, the deep learning module includes: A motion state deviation identification module is used to construct an acceleration reference vector. The acceleration reference vector consists of pre-defined X-axis acceleration reference values, Y-axis acceleration reference values, and Z-axis acceleration reference values. Each first motion state vector in the first motion state vector sequence is subtracted from the acceleration reference vector, and the absolute value is taken to obtain an acceleration deviation vector sequence. The modulus of all acceleration deviation vectors in the acceleration deviation vector sequence is taken to obtain an acceleration deviation modulus sequence. A velocity reference vector is constructed. The velocity reference vector consists of pre-defined X-axis velocity reference values, Y-axis velocity reference values, and Z-axis velocity reference values. Each second motion state vector in the second motion state vector sequence is subtracted from the velocity reference vector, and the absolute value is taken to obtain a velocity deviation vector sequence. The modulus of all velocity deviation vectors in the velocity deviation vector sequence is taken to obtain a velocity deviation modulus sequence. If any acceleration deviation modulus in the acceleration deviation modulus sequence is greater than a pre-set first trigger threshold or any velocity deviation modulus in the velocity deviation modulus sequence is greater than a pre-set second trigger threshold, then the underwater robot is determined to have deviated from the expected motion state.

[0049] Furthermore, the deep learning module also includes: The turbulent water flow texture recognition module, connected to the motion state deviation identification module, is used to input the tensor matrix sequence of the sonar image into a pre-trained three-dimensional convolutional neural network to extract visual features reflecting water flow disturbance texture in the sonar image, and map them into X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index; the X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index are combined to obtain a water flow disturbance texture index vector.

[0050] The training method for the three-dimensional convolutional neural network is as follows: A training sample set of water flow disturbance textures is obtained. Each training sample in the water flow disturbance texture training sample set includes a pre-acquired sequence of historical sonar image tensor matrices and a corresponding first real water flow disturbance vector. The first real water flow disturbance vector is contained within the time interval of the historical sonar image tensor matrix sequence acquisition, and includes real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions. The real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions respectively represent the underwater machine. The normalized water flow disturbance intensity at the location of the person is measured in the X, Y, and Z axes. This normalized water flow disturbance intensity is obtained by dimensionless processing of the components of the water flow velocity measured by the Doppler current meter in the X, Y, and Z axes. The sequence of tensor matrices of the historical sonar images is input into the three-dimensional convolutional neural network to be trained to obtain the predicted water flow disturbance texture index vector. The error between the predicted water flow disturbance texture index vector and the actual water flow disturbance vector is calculated. The parameters of the three-dimensional convolutional neural network are updated through the backpropagation algorithm until the error converges, resulting in a trained three-dimensional convolutional neural network.

[0051] For example, the turbulent water flow texture recognition module is used to extract visual features reflecting water flow disturbances from the tensor matrix sequence of sonar images. The basic principle of sonar imaging is to emit sound waves and receive target echoes; the pixel brightness in the image reflects the echo intensity. When there is water flow disturbance in the water area where the underwater robot is located, phenomena such as turbulence and eddies in the water flow will cause changes in the local density and velocity of the water body, and may carry air bubbles. These changes cause the acoustic impedance characteristics of the water medium to exhibit a non-uniform distribution in space. When sound waves pass through such non-uniform water bodies, they will produce effects such as scattering and refraction, causing the intensity of the echo signal to exhibit a specific fluctuation pattern in the spatial distribution, thus forming a unique texture pattern in the sonar image that is different from that of still water. For example, the wake disturbance caused by a ship passing by will present a specific texture pattern in the sonar image, and this texture feature is intrinsically related to the intensity and direction of the water flow disturbance. The turbulent water flow texture recognition module employs a 3D convolutional neural network (CNN) to process the tensor matrix sequence of sonar images. This CNN can simultaneously extract both spatial and temporal features from the sonar images. Spatial features capture the texture patterns formed by water flow disturbances in a single frame of the sonar image, while temporal features capture the dynamic characteristics of the water flow disturbances over time. Through the joint extraction of spatiotemporal features, the CNN maps the tensor matrix sequence of the sonar images into X-axis, Y-axis, and Z-axis water flow disturbance texture indices. These three indices represent the visual intensity of water flow disturbances along the X, Y, and Z axes, respectively, and are combined to obtain a water flow disturbance texture index vector. The calculated water flow disturbance texture index vector is as follows: .

[0052] Furthermore, the deep learning module also includes: The operation state deviation identification module, connected to the turbulent water flow texture identification module, is used to subtract the first and second operation state vectors corresponding to the same moment in the first and second operation state vector sequences to obtain an operation deviation vector sequence; each operation deviation vector in the operation deviation vector sequence contains elements from the first thruster operation deviation value to the... Thruster operating deviation value; respectively, the operating deviation value of the first thruster up to the... Statistical analysis was performed on the sequence of thruster operating deviation values ​​within a time window to extract the mean operating deviation, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding a preset deviation threshold for each thruster; the first thruster to the second thruster were then analyzed. The mean operating deviation of the thruster, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding the preset deviation threshold are combined in order of thruster number to obtain the operating state deviation feature vector. The operating state deviation feature vector is input into a pre-trained support vector regression model to extract the water flow disturbance identification results reflected by the thruster motor operating current control effect, and mapped to X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index. The X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index are combined to obtain the water flow disturbance current index vector.

[0053] The training method for the support vector regression model is as follows: A training sample set of water flow disturbance current is obtained. Each training sample in the water flow disturbance current training sample set includes: a pre-acquired historical operating state deviation feature vector and a corresponding second true water flow disturbance vector. The second true water flow disturbance vector contains the true values ​​of water flow disturbance in the X-axis direction, Y-axis direction, and Z-axis direction within the time interval from which the historical operating state deviation feature vector was collected. The historical operating state deviation feature vector is used as input, and the second true water flow disturbance vector is used as output. The support vector regression algorithm is used to train the model. The parameters of the support vector regression model are optimized by minimizing the prediction error to obtain the trained support vector regression model.

[0054] For example, the operational state deviation identification module is used to extract operational characteristics reflecting water flow disturbances from the thruster current data. The motion control of an underwater robot relies on the thruster system. The control system sends command current values ​​to each thruster according to the desired motion state, and the thruster motors generate thrust accordingly. Current sensors collect the actual operating current values ​​of the thruster motors in real time as feedback. In an ideal situation without external interference, the thruster feedback current value should be consistent with the command current value, with minimal deviation. When the underwater robot is subjected to water flow disturbances, the additional force generated by the water flow acts on the robot body, which is equivalent to introducing an external disturbance into the thruster control loop. To resist external disturbances, the control system needs to adjust the output thrust of the thrusters. However, due to limitations in the control system's response speed, a deviation will occur between the command current value and the feedback current value, which typically manifests as a step-like abrupt change in the time series. Furthermore, the statistical characteristics of the deviation, such as its amplitude and duration, are correlated with the intensity of the water flow disturbance. It is worth noting that if traditional analytical algorithms are used to map the thruster motor operating current control effect to X-axis, Y-axis, and Z-axis water flow disturbance identification results, the installation positions of the first to fourth thrusters on the underwater robot need to be obtained. However, using the support vector regression model described in this patent, it is not necessary to obtain the installation positions of each thruster beforehand; the model can autonomously learn the implicit mapping relationship between the installation positions and water flow disturbances during training. The calculated water flow disturbance current index vector is as follows: .

[0055] Furthermore, the deep learning module also includes: The weighting module, connected to the operating state deviation identification module, is used to calculate the water flow disturbance feature value vector using a three-dimensional weighting formula; the three-dimensional weighting formula is: ; in, This represents the eigenvalue vector of water flow disturbance. , , These represent the characteristic values ​​of water flow disturbance along the X-axis, Y-axis, and Z-axis, respectively. This represents the texture index vector for water flow disturbance. , , These represent the X-axis water flow disturbance texture index, the Y-axis water flow disturbance texture index, and the Z-axis water flow disturbance texture index, respectively. This represents the index vector of water flow disturbance current. , , These represent the X-axis water flow disturbance current index, the Y-axis water flow disturbance current index, and the Z-axis water flow disturbance current index, respectively. This represents the pre-set weighting, with a value ranging from 0 to 1.

[0056] For example, A value of 0.5 indicates that the identification results of the operational state deviation identification module and the turbulent water flow texture identification module are given the same level of trust. If the identification results of the operational state deviation identification module are given more trust, then... The value should be less than 0.5. If you have more confidence in the recognition results of the turbulent water flow texture recognition module, then... The value is greater than 0.5.

[0057] Furthermore, the identification module includes: The comparison module is used to compare the water flow disturbance feature value vector with the preset first identification threshold vector and the second identification threshold vector. If any element in the water flow disturbance feature value vector is greater than the corresponding element in the preset first identification threshold vector or less than the corresponding element in the preset second identification threshold vector, it is determined that the underwater robot has deviated from the expected motion state due to water flow disturbance.

[0058] For example, the calculated eigenvector of the water flow disturbance is as follows: This indicates that the underwater robot deviated from its expected motion state due to water flow disturbance. Furthermore, the water flow disturbance feature vector shows that the disturbance primarily originates from the X and Y axes.

[0059] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based underwater robot state analysis system, characterized in that, The system includes: The data acquisition module is used to acquire sonar image sequences, triaxial acceleration sequences, triaxial velocity sequences, thruster feedback current value sequences, and thruster command value sequences. The data preprocessing module, connected to the data acquisition module, is used to map the sonar image sequence into a sonar image tensor matrix sequence; perform time alignment and normalization on the triaxial acceleration sequence and triaxial velocity sequence to obtain a first motion state vector sequence and a second motion state vector sequence; and perform time alignment and normalization on the thruster feedback current value sequence and thruster command value sequence to obtain a first operating state vector sequence and a second operating state vector sequence, respectively. The deep learning module, connected to the data preprocessing module, is used to determine whether the underwater robot deviates from the expected motion state based on the first motion state vector sequence and the second motion state vector sequence. When it is determined that the underwater robot deviates from the expected motion state, visual features reflecting the water flow disturbance texture are extracted based on the sonar image tensor matrix sequence, and the water flow disturbance identification results reflecting the control effect of the thruster motor working current are extracted based on the first and second running state vector sequences, i.e., running features. The visual features and running features are fused to obtain the water flow disturbance feature value vector. The identification module, connected to the deep learning module, is used to compare the water flow disturbance feature vector with the pre-set first identification threshold vector and second identification threshold vector to determine whether the underwater robot has deviated from the expected motion state due to water flow disturbance.

2. The underwater robot state analysis system based on deep learning as described in claim 1, characterized in that, The data acquisition module includes: The sonar image sequence acquisition module is used to acquire sonar image sequences within a preset time window and at a preset first sampling frequency using the sonar equipment carried by the underwater robot. The three-axis acceleration sequence acquisition module, connected to the sonar image sequence acquisition module, is used to acquire a three-axis acceleration sequence within a pre-set time window and at a pre-set first sampling frequency using a fiber optic inertial navigation device mounted on an underwater robot. The three-axis acceleration sequence represents the acceleration component sequence in the X-axis, Y-axis, and Z-axis directions in a pre-set three-dimensional spatial coordinate system. The three-dimensional spatial coordinate system is composed of mutually perpendicular X-axis, Y-axis, and Z-axis. The three-axis velocity sequence acquisition module, connected to the three-axis acceleration sequence acquisition module, is used to acquire the three-axis velocity sequence within a preset time window and at a preset first sampling frequency using a Doppler log on the underwater robot; the three-axis velocity sequence represents the velocity component sequence in the X-axis, Y-axis and Z-axis directions in a preset three-dimensional spatial coordinate system. A thruster feedback current value sequence acquisition module, connected to a three-axis velocity sequence acquisition module, is used to acquire the thruster feedback current value sequence via a current sensor within a preset time window at a preset first sampling frequency. The thruster feedback current value sequence represents the effective value sequence of the actual operating current of the underwater robot's thruster motors. The number of thrusters in the underwater robot is denoted as... The thrusters were numbered from the first thruster to the second. Thrusters; respectively propelling the first thruster to the second thruster The sequence of thruster feedback current values ​​is denoted as the first thruster feedback current value sequence up to the second. Thruster feedback current value sequence; A thruster command value sequence acquisition module, connected to a thruster feedback current value sequence acquisition module, is used to read the thruster command value sequence from the underwater robot control system within a preset time window at a preset first sampling frequency; the thruster command value sequence represents the command current value sent by the underwater robot control system to the thruster motor; the sequence is used to acquire the first thruster command value sequence to the second thruster command value sequence. The thruster command value sequence of the thrusters is denoted as the first thruster command value sequence to the second. Thruster command value sequence.

3. The underwater robot state analysis system based on deep learning as described in claim 2, characterized in that, The data preprocessing module includes: The pixel extraction module is used to convert each frame of sonar image in the sonar image sequence into a sonar image tensor matrix using a pixel aggregation method; and to stack the sonar image tensor matrices in the order of acquisition time to obtain a sonar image tensor matrix sequence. The pixel aggregation method is as follows: extract the RGB color values ​​of all pixels in each frame of the sonar image sequence, and convert them into grayscale values ​​using an RGB color value weighted average algorithm; divide each frame of the sonar image sequence into pre-set number of rows. and number of split columns Divided into Each grid has an equal size; the average gray value of all pixels in each grid is extracted as the representative gray value of that grid; the... The representative gray values ​​of each raster cell are arranged in row and column order to form a raster. OK The column is a two-dimensional matrix, denoted as the sonar image tensor matrix.

4. The underwater robot state analysis system based on deep learning as described in claim 3, characterized in that, The data preprocessing module further includes: The motion state data preprocessing module, connected to the pixel extraction module, is used to time-align the triaxial acceleration sequence and triaxial velocity sequence, and normalize them using the maximum-minimum normalization method to obtain normalized triaxial acceleration and velocity sequences. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial acceleration sequence are combined to obtain a first motion state vector. All first motion state vectors within a time window are combined to obtain a first motion state vector sequence. The X-axis, Y-axis, and Z-axis components corresponding to the same moment in the normalized triaxial velocity sequence are combined to obtain a second motion state vector. All second motion state vectors within a time window are combined to obtain a second motion state vector sequence.

5. The underwater robot state analysis system based on deep learning as described in claim 4, characterized in that, The data preprocessing module further includes: The operation status data preprocessing module, connected to the motion status data preprocessing module, is used to sequence the feedback current values ​​of the first thruster to the first... Thruster feedback current value sequence, first thruster command value sequence to the... The thruster command value sequence is time-aligned and normalized using the maximum-minimum normalization method to obtain the normalized first thruster feedback current value sequence up to the normalized second... The sequence of thruster feedback current values, the normalized first thruster command value sequence, and the normalized second thruster command value sequence. Thruster command value sequence; normalize the first thruster feedback current value corresponding to the same moment to the normalized second thruster command value. The thruster feedback current values ​​are combined to obtain a first operating state vector; all first operating state vectors within the time window are arranged in chronological order to obtain a first operating state vector sequence; the normalized first thruster command value corresponding to the same moment is then assigned to the normalized first operating state vector sequence. The thruster command values ​​are combined to obtain a second operating state vector; all the second operating state vectors within the time window are arranged in chronological order to obtain a second operating state vector sequence.

6. The underwater robot state analysis system based on deep learning as described in claim 5, characterized in that, The deep learning module includes: A motion state deviation identification module is used to construct an acceleration reference vector. The acceleration reference vector consists of pre-defined X-axis acceleration reference values, Y-axis acceleration reference values, and Z-axis acceleration reference values. Each first motion state vector in the first motion state vector sequence is subtracted from the acceleration reference vector, and the absolute value is taken to obtain an acceleration deviation vector sequence. The modulus of all acceleration deviation vectors in the acceleration deviation vector sequence is taken to obtain an acceleration deviation modulus sequence. A velocity reference vector is constructed. The velocity reference vector consists of pre-defined X-axis velocity reference values, Y-axis velocity reference values, and Z-axis velocity reference values. Each second motion state vector in the second motion state vector sequence is subtracted from the velocity reference vector, and the absolute value is taken to obtain a velocity deviation vector sequence. The modulus of all velocity deviation vectors in the velocity deviation vector sequence is taken to obtain a velocity deviation modulus sequence. If any acceleration deviation modulus in the acceleration deviation modulus sequence is greater than a pre-set first trigger threshold or any velocity deviation modulus in the velocity deviation modulus sequence is greater than a pre-set second trigger threshold, then the underwater robot is determined to have deviated from the expected motion state.

7. The underwater robot state analysis system based on deep learning as described in claim 6, characterized in that, The deep learning module also includes: The turbulent water flow texture recognition module, connected to the motion state deviation identification module, is used to input the tensor matrix sequence of the sonar image into a pre-trained three-dimensional convolutional neural network to extract visual features reflecting water flow disturbance texture in the sonar image, and map them into X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index; the X-axis water flow disturbance texture index, Y-axis water flow disturbance texture index, and Z-axis water flow disturbance texture index are combined to obtain a water flow disturbance texture index vector; The training method for the three-dimensional convolutional neural network is as follows: A training sample set of water flow disturbance textures is obtained. Each training sample in the water flow disturbance texture training sample set includes a pre-acquired sequence of historical sonar image tensor matrices and a corresponding first real water flow disturbance vector. The first real water flow disturbance vector is contained within the time interval of the historical sonar image tensor matrix sequence acquisition, and includes real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions. The real values ​​of water flow disturbance in the X-axis, Y-axis, and Z-axis directions respectively represent the underwater machine. The normalized water flow disturbance intensity at the location of the person is measured in the X, Y, and Z axes. This normalized water flow disturbance intensity is obtained by dimensionless processing of the components of the water flow velocity measured by the Doppler current meter in the X, Y, and Z axes. The sequence of tensor matrices of the historical sonar images is input into the three-dimensional convolutional neural network to be trained to obtain the predicted water flow disturbance texture index vector. The error between the predicted water flow disturbance texture index vector and the actual water flow disturbance vector is calculated. The parameters of the three-dimensional convolutional neural network are updated through the backpropagation algorithm until the error converges, resulting in a trained three-dimensional convolutional neural network.

8. The underwater robot state analysis system based on deep learning as described in claim 7, characterized in that, The deep learning module also includes: The operation state deviation identification module, connected to the turbulent water flow texture identification module, is used to subtract the first and second operation state vectors corresponding to the same moment in the first and second operation state vector sequences to obtain an operation deviation vector sequence; each operation deviation vector in the operation deviation vector sequence contains elements from the first thruster operation deviation value to the... Thruster operating deviation value; respectively, the operating deviation value of the first thruster up to the... Statistical analysis was performed on the sequence of thruster operating deviation values ​​within a time window to extract the mean operating deviation, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding a preset deviation threshold for each thruster; the first thruster to the second thruster were then analyzed. The mean operating deviation of the thruster, the absolute value of the maximum operating deviation, and the duration of the operating deviation exceeding the preset deviation threshold are combined in order of thruster number to obtain the operating state deviation feature vector. The operating state deviation feature vector is input into a pre-trained support vector regression model to extract the water flow disturbance identification result reflected by the thruster motor operating current control effect, and mapped to X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index. The X-axis water flow disturbance current index, Y-axis water flow disturbance current index, and Z-axis water flow disturbance current index are combined to obtain the water flow disturbance current index vector. The training method for the support vector regression model is as follows: A training sample set of water flow disturbance current is obtained. Each training sample in the water flow disturbance current training sample set includes: a pre-acquired historical operating state deviation feature vector and a corresponding second true water flow disturbance vector. The second true water flow disturbance vector contains the true values ​​of water flow disturbance in the X-axis direction, Y-axis direction, and Z-axis direction within the time interval from which the historical operating state deviation feature vector was collected. The historical operating state deviation feature vector is used as input, and the second true water flow disturbance vector is used as output. The support vector regression algorithm is used to train the model. The parameters of the support vector regression model are optimized by minimizing the prediction error to obtain the trained support vector regression model.

9. The underwater robot state analysis system based on deep learning as described in claim 8, characterized in that, The deep learning module also includes: The weighting module, connected to the operating state deviation identification module, is used to calculate the water flow disturbance feature value vector using a three-dimensional weighting formula; the three-dimensional weighting formula is: ； in, This represents the eigenvalue vector of water flow disturbance. , , These represent the characteristic values ​​of water flow disturbance along the X-axis, Y-axis, and Z-axis, respectively. This represents the texture index vector for water flow disturbance. , , These represent the X-axis water flow disturbance texture index, the Y-axis water flow disturbance texture index, and the Z-axis water flow disturbance texture index, respectively. This represents the index vector of water flow disturbance current. , , These represent the X-axis water flow disturbance current index, the Y-axis water flow disturbance current index, and the Z-axis water flow disturbance current index, respectively. This represents the pre-set weighting, with a value ranging from 0 to 1.

10. The underwater robot state analysis system based on deep learning as described in claim 9, characterized in that, The identification module includes: The comparison module is used to compare the water flow disturbance feature value vector with the preset first identification threshold vector and the second identification threshold vector. If any element in the water flow disturbance feature value vector is greater than the corresponding element in the preset first identification threshold vector or less than the corresponding element in the preset second identification threshold vector, it is determined that the underwater robot has deviated from the expected motion state due to water flow disturbance.

Images