Multi-task motion pattern recognition system and method for underwater exoskeleton robot

Abstract

This invention relates to a multi-task motion pattern recognition system and method for underwater exoskeleton robots. Through innovative data preprocessing and network architecture design, it utilizes a small amount of IMU data to synchronously, in real-time, and with high precision complete motion pattern classification and continuous motion phase prediction, thereby driving the exoskeleton to achieve precise and synchronous assisted control. The multi-task motion pattern recognition system for underwater exoskeleton robots includes a sensing module for acquiring raw posture data of a diver; a data processing module for receiving and calculating relative IMU posture data within the human coordinate system to construct an input tensor; an intent recognition module, including a multi-task deep neural network based on an inverted Transformer architecture, for receiving the input tensor and synchronously outputting motion pattern classification results and continuous motion phase prediction results; and a control module for generating control signals for the underwater exoskeleton based on the motion pattern classification results and continuous motion phase prediction results.

Description

Technical Field

[0001] This invention relates to the field of underwater wearable robots and human-computer interaction technology, and in particular to a real-time motion intention perception system and method based on an inverted Transformer architecture for underwater exoskeletons, which aims to accurately identify the lower limb movement patterns of divers and estimate continuous motion phases to achieve synchronous assistance. Background Technology

[0002] Human divers experience tremendous physical exertion when performing underwater missions. Underwater assisted exoskeletons are key equipment for extending operational time, and their effectiveness lies in the precise and real-time perception of the diver's movement intentions. Due to the limitations of the underwater environment, sensing methods such as vision and electromyography are restricted, making inertial measurement units (IMUs) the mainstream method for acquiring motion data.

[0003] Existing technologies mostly focus on simple underwater motion phase detection, lacking effective integration with exoskeleton control systems; and among methods for underwater exoskeletons, there are deficiencies in algorithm adaptability and real-time performance. First, regarding sensing and generalization capabilities, although reference [1] proposes using the deformation signal of a flexible capacitive sensor to identify swimming stroke and phase, its dependence on specific sensor design increases the complexity of hardware implementation. More importantly, this method does not fully consider individual differences, and the inconsistency between muscle and skin deformation severely restricts the generalization robustness of the system. Reference [3] directly uses IMU data with directional differences for calculation, which may pose a risk of insufficient generalization capability in real-world scenarios.

[0004] Secondly, in terms of adaptability to non-steady-state motion, the traditional threshold detection method represented by reference [2] has obvious lag and misjudgment risks. Such algorithms rely on weighted moving averages for period estimation. Although they perform reasonably well in steady-state cruise, when divers perform variable-frequency motions such as rapid acceleration and deceleration (e.g., against currents, hovering), the historical average cannot quickly converge to the current state, resulting in severe phase lag. In addition, when facing non-periodic maneuvers such as attitude adjustment and obstacle avoidance, rigid thresholds are very likely to trigger misjudgments, apply incorrect torques, and thus disrupt human balance.

[0005] Finally, regarding the real-time architecture of the system, references [1] and [3] adopt a serial cascaded architecture of "pattern classification-phase prediction", which introduces unavoidable cumulative delay. In particular, the BiLSTM network in reference [3], although its bidirectional structure fits the extraction of temporal features, its serial inference mechanism leads to a significant increase in computational overhead, making it difficult to achieve a balance between high-precision prediction and low-latency control.

[0006] Reference [1]: Publication number CN116785673A, A real-time feedback training system and method based on swimming phase monitoring; Literature [2]: Wu X, Xu M, Zhou Z, et al. An Underwater Exoskeleton forScuba Diving: Reducing Air Consumption and Muscle Activation Through KneeAssistance[J]. IEEE Transactions on Robotics, 2025; Literature[3]: Chen L, Hu D. Study of a Cable-Driven Hip Swimming-AssistedExoskeleton Utilizing Adaptive Active Control Strategy[J]. Journal of FieldRobotics, 2025, 42(5): 1874-1886.

[0007] Therefore, there is an urgent need for an underwater motion intent recognition technology that can achieve high accuracy, strong robustness, and low latency under limited IMU sensing conditions. Summary of the Invention

[0008] This invention aims to overcome the shortcomings of existing technologies and provide a multi-task motion pattern recognition system and method for underwater exoskeleton robots. Its purpose is to achieve motion pattern classification and continuous motion phase prediction synchronously, in real time and with high accuracy by using a small amount of IMU data through innovative data preprocessing and network architecture design, thereby driving the exoskeleton to achieve precise and synchronous assisted control.

[0009] To address the aforementioned problems, in one aspect, embodiments of the present invention provide a multi-task motion pattern recognition system for underwater exoskeleton robots, characterized in that it includes: The sensing module is used to collect the diver's raw attitude data, including at least one reference inertial measurement unit (IMU) mounted on the torso and at least one target IMU mounted on the lower limbs. The data processing module is used to receive the raw attitude data, calculate the relative attitude data of the target IMU relative to the reference IMU, and construct an input tensor based on the relative attitude data, wherein the basic unit of the input tensor is the time-series data of the sensor channel; The intent recognition module includes a multi-task deep neural network based on an inverted Transformer architecture, used to receive the input tensor and simultaneously output motion pattern classification results and continuous motion phase prediction results; The multi-task deep neural network includes: An inverted embedding layer is used to encode the timing data of each sensor channel into independent variable tokens; The encoder is composed of at least one stacked inverted Transformer block, each of which includes a multivariate self-attention layer that performs cross-channel interaction of the variable tokens and a feedforward network layer that independently extracts temporal features from each variable token. A motion pattern classification head is used to generate a probability distribution of motion pattern categories based on the output of the encoder; A motion phase prediction head is used to generate continuous periodic motion phase values ​​based on the output of the encoder; The control module is used to generate control signals for the underwater exoskeleton based on the motion pattern classification results and the continuous motion phase prediction results.

[0010] Furthermore, the data processing module calculates the relative attitude data in the following specific way: for each target IMU, it uses its original attitude quaternion q target The original attitude quaternion q of the reference IMU ref Calculate relative rotation quaternions ,in This represents quaternion multiplication.

[0011] Furthermore, each of the inverted Transformer blocks, before the multivariate self-attention layer, includes a layer normalization layer that independently normalizes each variable token, with the mean and standard deviation on which the normalization is based derived from the statistical characteristics of the sensor channel time-series data represented by the variable token within the time window.

[0012] Furthermore, the motion pattern classification head includes a fully connected layer and a Softmax activation function to output probability distributions for at least three categories, including freestyle leg kicks, breaststroke leg kicks, and random movements.

[0013] Furthermore, the control module is configured to: when the probability of the random action category in the motion pattern classification result exceeds a preset threshold or is the maximum probability, control the underwater exoskeleton to enter a transparent mode with zero torque output; otherwise, based on the identified specific swimming stroke pattern and its corresponding continuous motion phase value, call the preset auxiliary torque curve to generate an auxiliary torque control signal.

[0014] Furthermore, the motion phase prediction head outputs a continuous value in the range of [0, N], where N is a real number greater than 0, and this value linearly corresponds to the phase of the complete motion cycle from 0% to 100%.

[0015] Secondly, this invention proposes a recognition method based on the aforementioned multi-task motion pattern recognition system for underwater exoskeleton robots, characterized by the following steps: Acquisition steps: Acquire raw attitude data from a reference IMU mounted on the diver's torso and at least one target IMU mounted on the lower limbs; Preprocessing steps: Calculate the relative attitude data of each target IMU relative to the reference IMU, and construct an input tensor based on all relative attitude data, with the sensor channel as the basic unit; Recognition Steps: The input tensor is fed into a multi-task deep neural network trained on an inverted Transformer architecture. The multi-task deep neural network performs inference synchronously through the following sub-steps: a) Mapping the temporal data of each sensor channel into independent variable tokens through an inverted embedding layer; b) Extracting features from the variable tokens through an encoder, which performs self-attention calculation across variable tokens and independent temporal feature transformation for each variable token; c) Based on the encoded features, outputting the probability distribution of motion pattern categories through a parallel motion pattern classification branch, and outputting continuous periodic motion phase values ​​through a parallel motion phase prediction branch. Control steps: Based on the probability distribution of the motion pattern category and the continuous periodic motion phase value, generate instructions for controlling the underwater exoskeleton actuator.

[0016] Furthermore, in step b) of the identification process, before performing self-attention calculation on the variable tokens, each variable token is first subjected to independent layer normalization processing. The normalization parameters are derived from the statistical characteristics of the time-series data of the sensor channel corresponding to the token within the time window.

[0017] Furthermore, the movement pattern category includes at least periodic swimming strokes and non-periodic random movements; the control steps include: when non-periodic random movements are identified as the dominant pattern, suppressing or stopping the assist output of the underwater exoskeleton.

[0018] Secondly, the present invention proposes a computer program including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned underwater exoskeleton motion intention recognition method.

[0019] Compared with the prior art, the beneficial effects of the present invention are: (1) High computational efficiency and low inference latency This invention is based on the iTransformer architecture, which inverts the attention dimension to sensor channels. Since the number of channels is much smaller than the time step, the computational complexity is reduced from... Significantly reduced It fully supports parallel computing. While improving the model's generalization ability, it can meet the latency requirements for exoskeleton control.

[0020] (2) Dynamic tracking without lag, adapting to unsteady motion (solving the frequency conversion following problem) This invention utilizes a multivariate self-attention mechanism to dynamically capture the coupling characteristics of different sensor channels at the current moment. It calculates the phase in real time, avoiding the risk of lag introduced by weighting with historical phases. Even under aperiodic or variable-frequency motion, the continuous phase predicted by the model closely follows the actual limb movements.

[0021] (3) Eliminates attitude interference and has strong generalization ability Existing technologies typically use raw IMU data (acceleration / angular velocity) or absolute angles directly, which are susceptible to the diver's overall posture (such as pitch and lateral movement) and sensor drift. This invention employs "relative quaternions" based on a torso coordinate system as input features, combined with channel-independent layer normalization. This approach physically eliminates interference from changes in diving orientation and effectively suppresses sensor zero-bias drift, ensuring accurate model recognition under various complex diving postures.

[0022] (4) Integrated architecture Existing technologies employ a serial cascaded architecture of "classification first, regression later," requiring separate maintenance of classification and regression networks. This results in parameter redundancy, and errors in the preceding classification stage can lead to control failures in the subsequent stage. This invention constructs a multi-task learning architecture with a shared encoder, enabling a single model to output motion patterns and phase information in parallel. This end-to-end design reduces system computational overhead.

[0023] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a structural diagram of the multi-task motion pattern recognition system for underwater exoskeleton robots proposed in this invention; Figure 2 This is a flowchart of the multi-task motion pattern recognition method for underwater exoskeleton robots proposed in this invention; Figure 3 The graphs show the efficiency-performance analysis of different models during the training and inference phases. The left graph represents the training efficiency of different models, with evaluation metrics including training time, mean squared error (MSE), and peak GPU memory usage. The right graph represents the inference efficiency of different models, with metrics including inference latency, mean absolute error (MAE), and GPU memory usage (represented by bubble area). Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. 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. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention.

[0027] See Figure 1 The present invention provides a multi-task motion pattern recognition system for an underwater exoskeleton robot, comprising: Sensing module: includes at least one reference inertial measurement unit (IMU) installed on the diver's torso and at least one target IMU installed on the lower limbs, for acquiring raw attitude data.

[0028] Data processing module: used to receive the raw attitude data, calculate the relative attitude data of the target IMU with respect to the reference IMU to eliminate the interference of the overall attitude change of the diver, and construct an input tensor with the sensor channel time series data as the basic unit.

[0029] The intent recognition module's core is a multi-task deep neural network trained on an inverted Transformer architecture. This network receives the input tensor and encodes the temporal data of each sensor channel into independent variable tokens through an inverted embedding layer. Subsequently, the encoder performs cross-channel interactive multivariate self-attention calculations on the variable token sequence and extracts independent temporal features from each token. Finally, through parallel motion pattern classification and motion phase prediction heads, it synchronously outputs the probability distribution of the current motion pattern category and continuous periodic motion phase values.

[0030] The control module is used to generate and output control signals to drive the underwater exoskeleton actuator based on the motion pattern classification results and the continuous motion phase prediction results. When a random motion is identified, the exoskeleton is controlled to enter a transparent mode with zero torque output; when a periodic swimming stroke is identified, the corresponding auxiliary torque template is invoked to apply assistance based on the phase value.

[0031] The multi-task motion pattern recognition system for underwater exoskeleton robots described above, when in operation, includes the following steps: Acquire raw attitude data from a reference IMU mounted on the diver's torso and a target IMU mounted on the lower limbs; Calculate the relative attitude data of each target IMU with respect to the reference IMU, and construct an input tensor with sensor channels as the basic unit; The input tensor is fed into a multi-task deep neural network based on the inverted Transformer architecture; the network processes the motion pattern classification results and continuous motion phase prediction results through inverted embedding, layer normalization, cross-channel self-attention and feedforward network; Based on the classification results and phase prediction results, control commands for the underwater exoskeleton are generated.

[0032] Example 1: A multi-task motion pattern recognition system for underwater exoskeleton robots, the system's hardware relying on a flexible underwater exoskeleton system (PEAKED), such as... Figure 1 The lower left section illustrates the system's sensing module, which includes three waterproof inertial measurement units (IMUs). One lumbar IMU serves as a reference IMU, installed inside the diver's lower back buoyancy vest, acting as a baseline coordinate system for monitoring torso posture. The other two, as target IMUs, are fixed to flexible straps on the outer sides of the left and right thighs, respectively, to capture real-time lower limb swing signals. All sensor data is transmitted to the central controller via a CAN bus or EtherCAT bus, with a sampling rate set between 100Hz and 1000Hz. Based on the calculation results, an execution module consisting of two waterproof motor drive units pulls the thigh brace via Bowden cables, precisely contracting the cable according to the phase signal, thereby providing effective auxiliary torque for the diver's hip flexion and extension.

[0033] This embodiment uses three IMU sensors and their corresponding locations. However, it should be noted that adjusting the number and location of the sensors does not affect the applicability of the invention's technical architecture. For example, adding an IMU at the lower leg or webbed feet can obtain more granular kinematic data, or reducing it to using only a single leg + waist IMU.

[0034] The data processing module is deployed in the central controller. Its core preprocessing step is relative attitude calculation, which is implemented through the following methods: S100: Acquire raw IMU data, which contains quaternions in the global coordinate system.

[0035] S101: Calculate relative attitude data This invention calculates the relative quaternions of the limb end sensors relative to the torso reference sensor, mapping all sensor data to a local coordinate system centered on the torso. This eliminates the overall directionality of the diver in the water and avoids interference from changes in heading, roll, and pitch (such as head-down descent) on the recognition of local limb movements. Taking the calculation process of the left leg IMU as an example, let the quaternion of the waist IMU be... The quaternion of the left thigh IMU is Then the quaternion of the relative rotation of the left thigh with respect to the waist. The calculation formula is:

[0036] in To represent quaternion multiplication, This represents the conjugate (inverse) of a quaternion. This step transforms the input features from the global space to the human anatomy space, significantly improving the model's generalization ability across different diving postures.

[0037] S102: Constructing the input tensor In this embodiment, quaternion sequence data of the left and right thighs relative to the waist are selected as input, totaling... There are [number] channels, and the construction time window length is [length]. (For example Input tensor of sampling points Each column represents a sensor channel in the past Data at any given moment.

[0038] The core algorithm flow of the intent recognition module in this embodiment is as follows: Figure 1 , Figure 2 As shown: This invention employs a multi-task learning network built upon the iTransformer (InvertedTransformer) architecture to simultaneously perform motion phase recognition (MPP) and motion pattern classification (MMC) tasks for an underwater exoskeleton. The core of iTransformer lies in its dimension inversion processing. The core algorithm flow of the intent recognition module includes the following steps: S103: Inverted Embedding Unlike traditional Transformer-based time series models, (i.e., all sensor values ​​at a given moment) are mapped to a single token. This invention uses iTransformer to map the input tensor... (The entire historical sequence or data from a single channel of a sensor) is mapped to a single token (also called a variable token). For the first... Data sequence of each channel Embedded through a multilayer perceptron or a linear layer:

[0039] in This is the embedding dimension. At this point, the model obtains... Each token represents channel data from a physical sensor. Each token contains the complete temporal dynamic characteristics of the sequence.

[0040] S104: Layer Normalization For each variable token hc Perform layer normalization independently: ,in , The mean and standard deviation of the time series data for the channel corresponding to the token.

[0041] Thanks to the iTransformer architecture's inverted approach to token construction, the processing object of layer normalization shifts from traditional time slices to variate tokens representing the temporal characteristics of a single sensor channel. This normalization operation for individual channels can utilize the statistical characteristics within the time window to dynamically correct the non-stationary distribution of each sensor caused by changes in the underwater environment, thereby effectively suppressing the impact of sensor zero-bias drift and amplitude scaling on model feature extraction.

[0042] S105: Multivariate Self-Attention The normalized token sequence is input into an encoder consisting of N stacked iTransformer blocks. At the core of each block is a multivariate self-attention layer that computes attention weights between the variable tokens.

[0043]

[0044] iTransformer's self-attention mechanism calculates cross-channel variable dependencies. By treating sensor channels as tokens, the model-generated attention map dynamically quantifies the coupling strength of data from different channels within the current motion pattern. Given the number of sensor channels... (like Much smaller than the time window (like This mechanism optimizes computational complexity to This significantly reduces computational load while laying a solid foundation for real-time inference.

[0045] S106: Temporal Feature Feedforward Network (FFN) Following the attention layer, each token independently enters the feedforward network (FFN). Since each token represents a time series, the FFN is actually responsible for extracting the complex features of the sensor channel in the time dimension (such as periodicity, amplitude variation, waveform distortion, etc.).

[0046] S107: Multitasking Output go through After processing by the layer encoder, the feature matrix is ​​obtained through the mapping layer. After being flattened, the matrix is ​​split into two parallel task heads: S108a: Motion Pattern Classification Header (MMCHead): Composed of fully connected layers and a Softmax activation function, it maps feature vectors into a three-dimensional probability distribution covering freestyle leg kicks, breaststroke leg kicks, and random movements. Based on this, the exoskeleton system executes a confidence-based safety control strategy. When the predicted probability of random movements is dominant, the exoskeleton immediately switches to a transparent mode with zero torque output to avoid false triggering. Otherwise, it calls the corresponding auxiliary torque curve template to apply cooperative assistance according to the specific swimming stroke category identified.

[0047] S108b: Motion Phase Prediction Head (MPPHead): Continuous phase variables in intervals are obtained by using fully connected layers. Its range is [0,10] and corresponds to a phase of 0~100%, so as to accurately represent the relative progress in the action cycle at the current moment.

[0048] In addition to outputting a linear phase, the output here can also output sine / cosine phase components to address the numerical jumps at 0% and 100%. The classification head can be extended to recognize more patterns, such as "underwater walking" or "emergency surfacing."

[0049] S109: Generate exoskeleton control commands.

[0050] The training objective of the multi-task deep neural network is optimized through a composite loss function designed to balance the learning of motion pattern classification and motion phase prediction tasks. In one specific embodiment, the composite loss function is a weighted sum of the classification task loss and the regression task loss, and its expression is:

[0051] in, Use a loss for motion pattern classification (e.g., cross-entropy loss). The loss for motion phase prediction can be, for example, mean square error loss. and These are adjustable weights used to balance the learning rates of the two tasks. By optimizing this composite loss function, the network can simultaneously learn high-precision pattern discrimination and phase estimation capabilities.

[0052] Experimental verification The underwater exoskeleton motion pattern recognition system and method based on the iTransformer architecture proposed in this invention have been fully trained and tested on a real underwater exoskeleton sensor dataset, and have been compared in detail with current mainstream time-series processing methods. The experimental environment was trained on an NVIDIA RTX 4080 / 5090 and inference tested on a Jetson AGX Orin embedded platform.

[0053] Experimental subjects and evaluation indicators (1) To verify the advancement of this invention, the following three types of models were selected as benchmarks for comparison: Advanced Transformer-based models: Informer, Crossformer, PatchTST.

[0054] Models based on multilayer perceptrons (MLP): TimeMixer (representing an efficient and lightweight model).

[0055] Traditional machine learning method: Support Vector Machine (SVM).

[0056] (2) The evaluation indicators cover two aspects: computational efficiency and task accuracy. Computational efficiency metrics: training time, training memory usage, inference latency, and inference memory usage. Mission accuracy metrics: mean square error (MSE) and mean absolute error (MAE) of phase prediction; accuracy (Acc) of motion pattern classification.

[0057] The experimental results of the model performance are shown in Table 1 and Figure 3Among them, the Crossformer-256, iTransformer-1024 and iTransformer-512 models have the best experimental results, but considering the balance between computational efficiency and task accuracy, the iTransformer-512 model is selected as the target model.

[0058] Table 1 Overall Experimental Results

[0059] — These tests were all conducted on an RTX 4080. Due to memory limitations at the current hidden layer dimension, a batch size of 512 cannot be used for training directly, so a suitable batch size was used instead.

[0060] The model achieved a classification accuracy of 99.98% across the three modes. This demonstrates that the architecture can almost perfectly identify the diver's movement intentions, effectively preventing the exoskeleton from making erroneous movements in incorrect modes. Motion phase prediction accuracy: MSE: 0.0816 MAE: 0.1435 (at an output scale of 0-10 phases).

[0061] The device inference efficiency results are shown in Table 2. On the NVIDIA RTX 4080 platform, the inference latency is approximately 3.0-3.2ms. On embedded edge devices (such as the Jetson platform), the inference speed also meets the real-time control requirements and is significantly better than the comparison model (Crossformer latency >20ms, which cannot meet the real-time inference requirements of the current exoskeleton control frequency).

[0062] Table 2. Inference speed (ms) and power consumption of the model on different hardware platforms

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-task motion pattern recognition system for underwater exoskeleton robots, characterized in that, include: The sensing module is used to collect the diver's raw attitude data, including at least one reference inertial measurement unit (IMU) mounted on the torso and at least one target IMU mounted on the lower limbs. The data processing module is used to receive the raw attitude data, calculate the relative attitude data of the target IMU relative to the reference IMU, and construct an input tensor based on the relative attitude data, wherein the basic unit of the input tensor is the time-series data of the sensor channel; The intent recognition module includes a multi-task deep neural network based on an inverted Transformer architecture, used to receive the input tensor and simultaneously output motion pattern classification results and continuous motion phase prediction results; The multi-task deep neural network includes: An inverted embedding layer is used to encode the timing data of each sensor channel into independent variable tokens; The encoder is composed of at least one stacked inverted Transformer block, each of which includes a multivariate self-attention layer that performs cross-channel interaction of the variable tokens and a feedforward network layer that independently extracts temporal features from each variable token. A motion pattern classification head is used to generate a probability distribution of motion pattern categories based on the output of the encoder; A motion phase prediction head is used to generate continuous periodic motion phase values ​​based on the output of the encoder; The control module is used to generate control signals for the underwater exoskeleton based on the motion pattern classification results and the continuous motion phase prediction results.

2. The multi-task motion pattern recognition system for underwater exoskeleton robots according to claim 1, characterized in that, The data processing module calculates the relative attitude data in the following way: for each target IMU, it uses its original attitude quaternion q. target The original attitude quaternion q of the reference IMU ref Calculate relative rotation quaternions ,in This represents quaternion multiplication.

3. The multi-task motion pattern recognition system for underwater exoskeleton robots according to claim 1, characterized in that, Each of the inverted Transformer blocks includes a layer normalization layer before the multivariate self-attention layer, which normallys each variable token independently. The mean and standard deviation of the normalization are derived from the statistical characteristics of the sensor channel time-series data represented by the variable token within a time window.

4. The multi-task motion pattern recognition system for underwater exoskeleton robots according to claim 1, characterized in that, The motion pattern classification head includes a fully connected layer and a Softmax activation function, used to output probability distributions for at least three categories, including freestyle leg kicks, breaststroke leg kicks, and random movements.

5. The multi-task motion pattern recognition system for underwater exoskeleton robots according to claim 4, characterized in that, The control module is configured to: when the probability of random action category in the motion pattern classification result exceeds a preset threshold or is the maximum probability, control the underwater exoskeleton to enter a transparent mode with zero torque output; otherwise, based on the identified specific swimming stroke pattern and its corresponding continuous motion phase value, call the preset auxiliary torque curve to generate an auxiliary torque control signal.

6. The multi-task motion pattern recognition system for underwater exoskeleton robots according to claim 1, characterized in that, The motion phase prediction head outputs a continuous value in the range of [0, N], where N is a real number greater than 0, and this value linearly corresponds to the phase of the complete motion cycle from 0% to 100%.

7. A method for recognizing the movement intent of an underwater exoskeleton, characterized in that, Including the following steps: Acquisition steps: Acquire raw attitude data from a reference IMU mounted on the diver's torso and at least one target IMU mounted on the lower limbs; Preprocessing steps: Calculate the relative attitude data of each target IMU relative to the reference IMU, and construct an input tensor based on all relative attitude data, with the sensor channel as the basic unit; Recognition Steps: The input tensor is fed into a multi-task deep neural network trained based on an inverted Transformer architecture; the multi-task deep neural network performs inference synchronously through the following sub-steps: a) Map the timing data of each sensor channel to independent variable tokens by inverting the embedding layer; b) The variable tokens are feature extracted by an encoder, which performs self-attention computation across variable tokens and independent temporal feature transformation for each variable token; c) Based on the encoded features, the probability distribution of motion pattern categories is output through a parallel motion pattern classification branch, and continuous periodic motion phase values ​​are output through a parallel motion phase prediction branch. Control steps: Based on the probability distribution of the motion pattern category and the continuous periodic motion phase value, generate instructions for controlling the underwater exoskeleton actuator.

8. The method according to claim 7, characterized in that, In step b) of the identification process, before performing self-attention calculation on the variable tokens, each variable token is first subjected to independent layer normalization. The normalization parameters are derived from the statistical characteristics of the time-series data of the sensor channel corresponding to the token within the time window.

9. The method according to claim 7, characterized in that, The movement pattern category includes at least periodic swimming strokes and non-periodic random movements; the control steps include: when non-periodic random movements are identified as the dominant pattern, suppressing or stopping the assist output of the underwater exoskeleton.

10. A computer program comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the underwater exoskeleton motion intent recognition method as described in any one of claims 7 to 9.

Images