Frequency domain analysis method and system of exoskeleton auxiliary equipment based on concept learning

Abstract

The invention discloses a frequency domain analysis method and system for exoskeleton auxiliary equipment based on concept learning, and relates to the technical field of reinforcement learning, and the method comprises the steps: receiving user gait data, carrying out the state representation and KMeans clustering of the user gait data, obtaining a user gait state, carrying out the fitness evaluation based on the user gait state, obtaining a fitness score, and carrying out the frequency domain analysis of the exoskeleton auxiliary equipment based on the fitness score. Carrying out concept evolution based on the fitness score to obtain a gait concept description value; the user gait data and the gait concept description value are subjected to frequency domain and concept space fusion optimization and feature fusion based on Fourier transform, an aggregated time sequence is obtained, the gait concept description value is input into a pre-established concept knowledge base to be stored, and the user experience and the training effect can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of reinforcement learning, and in particular to a frequency domain analysis method and system for an exoskeleton assistive device based on concept learning. Background Art

[0002] In the field of reinforcement learning, model training efficiency is low and generalization performance is insufficient in tasks involving periodic movements, such as rehabilitation training, motion assistance, and robotic gait control. In intelligent rehabilitation devices and motion-assisting robots such as the HAL powered suit and the ReWalk exoskeleton system, patients or users need to repeat specific limb movements to increase muscle strength or restore walking ability. Traditional reinforcement learning methods often have difficulty effectively capturing the periodic characteristics of these movements, making it difficult for the system to accurately adapt to individual needs. In periodic tasks, many combinations of actions and states will recur. This recurring pattern is important for reinforcement learning strategy optimization. For example, the patient's gait pattern, the limb swing frequency in rehabilitation training, and the runner's step cycle can all be optimized through periodic data modeling.

[0003] Traditional methods often increase data sampling or model complexity to improve performance. However, these methods fail to effectively capture and utilize the periodic nature of the task, limiting the model's performance and efficiency. Recent research has made progress in various fields, such as computer vision and signal processing, using Fourier transforms to analyze the periodic nature of signals in the frequency domain. The Fourier neural operator model proposed by Li et al. learns the evolution of complex dynamic systems in the frequency domain, but it cannot effectively balance high-frequency and low-frequency features.

[0004] The CLIP model has demonstrated excellent performance in state feature extraction and classification. Radford et al. first proposed the CLIP model, which embeds visual and conceptual descriptions into a unified space, achieving a match between vision and natural language. However, it still lacks the ability to dynamically evolve and adapt to tasks, preventing it from effectively learning from constantly changing inputs in complex tasks.

[0005] The concept of context-action consistency proposed by Zhang et al. clarifies that similar contexts in reinforcement learning should correspond to similar action states. This constraint improves the effectiveness of policy learning and reduces ineffective exploration. The Bisimulation theory of Ferns et al. provides theoretical support from the perspective of state similarity. By introducing distance constraints in the state space, the model compresses the state space, improving the model's learning efficiency. However, in practical applications such as bionic robot gait training and rehabilitation training systems, state representation is often affected by factors such as sensor noise and individual differences. Traditional state distance-based optimization methods have difficulty coping with changes in complex environments. Summary of the Invention

[0006] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a frequency domain analysis method and system for exoskeleton assistive devices based on concept learning.

[0007] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a frequency domain analysis method for an exoskeleton assistive device based on concept learning, the method comprising the following steps:

[0008] Receive user gait data, perform state representation and KMeans clustering on the user gait data to obtain the user gait state, perform fitness evaluation based on the user gait state to obtain a fitness score, and perform concept evolution based on the fitness score to obtain a gait concept description value;

[0009] The user gait data and gait concept description values are subjected to frequency domain and concept space fusion optimization and feature fusion based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

[0010] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: performing state representation and KMeans clustering on the user gait data to obtain the user gait state:

[0011] Perform KMeans clustering on the state representation of the gait data to generate initial concepts for different gait states, and cluster to obtain n concepts categories, and use GPT-2 to generate n variants A natural language description, let s i is the state feature of the exoskeleton system, S={s1,s2,...,s N} is the set of all states, and KMeans clustering is used to obtain C k represents the kth cluster, μ k is the mean of the cluster, each C k Generate multiple descriptions via GPT-2 :D k ={G GPT-2 (P k )|P k ∈P},P k It is a task-related prompt word, G GPT-2 Represents text generated by GPT-2.

[0012] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: performing fitness evaluation based on the user's gait state to obtain a fitness score includes:

[0013] The evaluation of concept quality includes CLIP similarity, state internal consistency, and action internal consistency, that is, gait state s i With concept d k The matching degree, the state s within the same concept category i ,s j Similarity between actions in the same concept category i ,a j The similarity between them is calculated by concept description d k With state s i Similarity:

[0014]

[0015] f CLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept d k CLIP representation;

[0016] The internal consistency of the state is expressed as:

[0017]

[0018] Sim cos (s i ,s j ) is the state s i and s j The cosine similarity between

[0019] The internal consistency AC of an action is expressed as:

[0020]

[0021] Final fitness F k Rating:

[0022] F k =Sim CLIP (s i , d k )×SC×AC.

[0023] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: a process of performing concept evolution based on the fitness score to obtain a gait concept description value:

[0024] In the exoskeleton system, the user's gait will change with time or task, so the exoskeleton system dynamically adapts to the changes in the user's gait during the concept evolution process. threshold is the fitness threshold, and we have:

[0025]

[0026] G GPT-2 (D k ) is a conceptual description regenerated by GPT-2, allowing the exoskeleton system to adapt to the needs of different gaits and movement patterns.

[0027] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the process of performing frequency domain and concept space fusion optimization on the user gait data and the gait concept description value based on Fourier transform includes:

[0028] Frequency domain feature conversion:

[0029] The gait data in the exoskeleton system contains periodic features, which are extracted using Fourier transform. At this time, Fourier transform and CLIP concept representation are combined to optimize feature learning;

[0030] Assume that the input gait data x is a tensor of shape (B, C, H, W), where B is the batch size, C is the number of channels, and H and W are the spatial dimensions. Perform Fourier transform on the gait data x:

[0031]

[0032] Frequency domain features are mapped to concept space:

[0033] Fourier transformed data X FFT It is a complex matrix, so we take its mean representation and map it to the CLIP concept space through the freq_to_semantic network to obtain the conceptual representation of the gait state:

[0034]

[0035] S sem is the conceptual representation of Fourier characteristics, f freq2sem It is a two-layer MLP used to convert Fourier features into clip dim Conceptual representation of dimension;

[0036] Calculate concept similarity:

[0037] Calculate the frequency domain feature S sem With CLIP Concept C clip The cosine similarity between them is used to evaluate the matching degree between the current state of the exoskeleton system and the task-related concepts:

[0038]

[0039] Cosine similarity indicates the matching degree between the input data and each concept. The concept C that is most similar to the current gait is selected. k, and perform gait optimization based on the concept:

[0040]

[0041] Generate concept attention weights:

[0042] Select the most similar concept C best , enter semantic attention network:

[0043] W att =f att (C best )

[0044] where f att It is an MLP that generates truncation size ×truncation size The attention weight matrix of the shape, ; normalized: W att =σ(W att ), σ represents the sigmoid activation function, and the weight is set between [0,1];

[0045] Concept-guided Fourier features:

[0046] The concept attention weight is generated by the similarity of the concept, and the frequency component of the gait data is modulated, and W is used att Perform frequency domain modulation:

[0047] X′ FFT =X FFT ×(1+λW att )

[0048] λ is a hyperparameter that controls the influence of semantic concepts on frequency domain features. Concept-related frequency components will be enhanced: 1+λW att >1; conceptually irrelevant frequency components will be suppressed: 1+λW att <1;

[0049] Inverse Fourier transform to restore data

[0050] The frequency domain data X′ after conceptual modulation FFT Returning to the time domain, the gait task is performed in real time: Finally, the concept-optimized time-domain gait signal x' is obtained, which aligns the frequency-domain features with the concept information.

[0051] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: after the frequency domain and concept space fusion optimization based on Fourier transform, entering the multi-scale convolution residual Fourier layer, including:

[0052] Multi-scale convolution module:

[0053] The multi-scale convolution module captures gait features of different scales. The input gait data x is fed into the multi-scale convolution module. The multi-scale convolution module captures the local features in the gait data through convolution operations of multiple scales. B is the batch size, Cin is the number of input channels, D=(D1, D2, ..., D n ) is the size of the spatial dimension. This module uses multiple convolutional layers Conv(x) to process the input data and obtain multi-level features:

[0054] x multi-scale =MultiScaleConv(x)

[0055] Block module and position encoding:

[0056] The input gait data is divided into blocks of the specified size (M1, M2, ..., M n ), segmentation is performed, and local features are extracted independently within each block. The data after segmentation is represented as: x blocks =Block(x,(M1,M2,...,M n )), then, position encoding is added to the features of each position, so that the exoskeleton system can use the spatiotemporal position information of the gait data. The position encoding is generated by the function f(q), where q represents the coordinates of each position in the input data:

[0057] x encoded =x blocks +f(q)

[0058] Convolutional Residual Fourier Layer:

[0059] Fourier transform converts gait signals from the time domain to the frequency domain to capture periodic structures. Different frequency components are weighted through the CLIP concept space, so that the frequency components related to the gait concept are enhanced.

[0060] CLIP calculates the similarity S of state X in the concept space CLIP (X, D), convert it into frequency domain weighting:

[0061] W semantic =σ(W att ·S CLIP (X, D)

[0062] X enhanced =X FFT ×(1+λ·W semantic )

[0063] W semanticis the spectrum weight guided by the CLIP concept space, W att Calculated by the attention mechanism, λ controls the strength of concept enhancement;

[0064] The output of the convolution residual Fourier layer yout is the Fourier transform output of the Fourier layer yfft after spectrum enhancement and the high-frequency features yconv extracted by the convolution layer are fused through the residual connection:

[0065] y out =y fft +y conv +b feature fusion:

[0066] The processed multi-scale features and Fourier domain convolution features are fed into the multi-head self-attention module, which establishes the dependency between different features through the multi-head attention mechanism and learns the interaction between gait features. Given the feature matrix The multi-head attention mechanism is calculated by the following formula:

[0067] Y attn =MultiHeadAttention(X features )

[0068] Cross-Attention:

[0069] The crisscross attention mechanism calculates the crisscross attention output using the following formula:

[0070] Y cross =CrossAttention(X features , Z latent )

[0071] where Z latent is the representation of the latent variable.

[0072] Time series aggregation and output generation:

[0073] After gait feature extraction and cross-attention mechanism processing, the time series information of the input data is aggregated in the time series aggregation module. The time series aggregation module aggregates the processed gait features according to the time dimension and extracts the temporal dependency in the gait data. Let the processed feature sequence be Y features =[y1, y2, ..., y T ], the time series aggregation operation can be expressed as:

[0074] Y aggregated =TimeAggregation(Y features ).

[0075] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: a concept storage process of the pre-established concept knowledge base:

[0076] The concept set of a task T is represented as: C T ={D1, D2, ..., D N}, D i is the concept description in task T, C T is the set of all concepts of task T, C T Store in warehouse Record concept fitness: in It's Concept D i The fitness score of

[0077] Concept similarity calculation process:

[0078] According to the current task T B and historical missions T A Calculate the similarity of concepts and select the most similar concept to transfer. Let D A，i It is task T A The concept of D B，j It is a new task T B The query concept in , uses the CLIP model to calculate the cosine similarity:

[0079]

[0080] f CLIP (D A，i ) and f CLIP (D B，j ) is the eigenvector extracted by CLIP, and T is calculated. B Each D B，j With T A All D A，i Similarity:

[0081]

[0082] Select the most similar top k D A，i ;

[0083] The process of gradual concept transfer:

[0084] In T A to T B During the migration process, dynamically adjust T A and T B The mixing ratio is such that T B Gradually adapt to the new task, let α be the adaptation measure, range [0,1], when α = 0, fully use T AThe concept of α=1 is fully used. B The existing concept is that when α is between 0 and 1, T is used. A and T B The new task uses the concept of historical tasks and adds new features, namely:

[0085]

[0086] C A is the source task T A The most similar concept set selected, C B is the target task T B existing concepts;

[0087] Controlling the mixing quantity can be formalized as:

[0088]

[0089] Include A target task concept, Source task concept;

[0090] Process for generating hybrid concepts:

[0091] based on Generate a T B The concept of task characteristics, let D B，i and D A，i yes Two concepts in, use GPT-2 to generate the fused concept: D hybrid =G GPT-2 (D B，i , D A，j ), indicating input D B，i and D A，j After that, the synthetic concept description generated by GPT-2, the resulting mixed concept set is

[0092] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the concept-guided behavior consistency optimization process:

[0093] Concept intra-class consistency loss L SC :

[0094] Assume s i and s j Same category C k The two states under a i and a j are their corresponding actions. The means of state and action are respectively feature and mean action, gait category C k The state features of are close to each other, the action features are consistent, and the state feature consistency loss is expressed as:

[0095]

[0096] The action feature consistency loss is expressed as:

[0097]

[0098] The final concept intra-class consistency loss is expressed as:

[0099] SC total =SC state +SC action

[0100] Since the class size |C k |There are different situations. The loss of each category is normalized so that each category contributes fairly to the loss:

[0101]

[0102] Concept class consistency loss L clip :

[0103] Assume s i and s j It is the state feature, calculate the state feature state features The Euclidean distance between them gives the distance matrix of state features:

[0104] D state (s i , s j )=||s i -s j ||2

[0105] Assume a i and a j is the action feature, calculate the Euclidean distance between the action features, and get the distance matrix of the action features:

[0106] D action (a i , a j )=||a i -a j ||2

[0107] Calculate the similarity of the CLIP model output to obtain the CLIP similarity of each pair of states:

[0108]

[0109] fCLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept description d k CLIP representation;

[0110] Normalize the similarity so that the distance between states conforms to the similarity in the concept space:

[0111]

[0112] Convert similarity to distance:

[0113] D clip =1-Sim CLIP

[0114] The error between the distance between the state feature and the CLIP distance, and the error between the distance between the action feature and the CLIP distance is used as the loss:

[0115]

[0116] The total loss is the weighted sum of the intra-concept class consistency loss and the CLIP-guided inter-concept class consistency loss:

[0117] L total =λ SC ×L SC +λ clip ×L clip .

[0118] In a second aspect, in order to achieve the above-mentioned object, the present invention discloses a frequency domain analysis system for an exoskeleton assistive device based on concept learning, comprising:

[0119] A data processing module is used to receive user gait data, perform state representation and KMeans clustering on the user gait data to obtain the user gait state, perform fitness evaluation based on the user gait state to obtain a fitness score, and perform concept evolution based on the fitness score to obtain a gait concept description value;

[0120] The frequency domain analysis module is used to perform frequency domain and concept space fusion optimization and feature fusion on the user gait data and gait concept description values based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

[0121] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, the frequency domain analysis method of the exoskeleton assistive device based on concept learning as described above is adopted.

[0122] Beneficial effects of the present invention:

[0123] The present invention aims to overcome the shortcomings of existing reinforcement learning methods, such as large state-action space, high redundancy, and unstable training, improve the training efficiency and generalization ability of the model, and enhance the interpretability of the reinforcement learning method. In motion-assistance devices, this method can adaptively learn the user's gait pattern and generate personalized training feedback. It is suitable for a variety of task scenarios, such as intelligent rehabilitation equipment, motion-assistance robots, and automatic gait training systems. This allows the device to more accurately adapt to the exercise needs of different users, improving the user experience and training results. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0125] Figure 1 It is a schematic flow chart of the method of the present invention;

[0126] Figure 2 It is a schematic diagram of the overall framework of the present invention;

[0127] Figure 3 It is a schematic diagram of the adaptive evolution of the concept of the present invention;

[0128] Figure 4 This is a schematic diagram of the frequency domain and concept space fusion optimization of the present invention;

[0129] Figure 5 This is a schematic diagram of the cross-task concept transfer of the present invention;

[0130] Figure 6 Schematic diagram of the CLIP-guided constraint loss of the present invention;

[0131] Figure 7 Schematic diagram of the multi-scale convolution residual Fourier layer of the present invention;

[0132] Figure 8 Schematic diagram of multi-scale convolution of the present invention;

[0133] Figure 9It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0134] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0135] Example 1:

[0136] like Figure 1 As shown, a frequency domain analysis method for an exoskeleton assistive device based on concept learning comprises the following steps:

[0137] S101: receiving user gait data, performing state representation and KMeans clustering on the user gait data to obtain the user gait state, performing fitness evaluation based on the user gait state to obtain a fitness score, and performing concept evolution based on the fitness score to obtain a gait concept description value;

[0138] The process of performing state representation and KMeans clustering on user gait data to obtain the user gait state is as follows:

[0139] In the ReWalk exoskeleton system, the user's gait characteristics are the core elements of the training process. The gait data is clustered by KMeans to generate initial concepts for different gait states. These concepts can be used to guide the motion control strategy of the exoskeleton system. concepts categories, and use GPT-2 to generate n variants A natural language description. Let s i is the state feature of the exoskeleton system, S={s1,s2,...,s N} is the set of all states, and KMeans clustering is used to obtain C k represents the kth cluster, μ k is the mean of the cluster. Each C k Generate multiple concept descriptions through GPT-2: D k ={G GPT-2 (P k )|P k ∈P},P k is a task-related prompt word (such as "a humanoid robot walking with......"), G GPT-2 Represents text generated by GPT-2 to guide the exoskeleton system to optimize gait.

[0140] The process of performing fitness evaluation based on the user's gait state to obtain a fitness score includes:

[0141] Under the adaptive learning of gait, the evaluation of concept quality includes CLIP similarity, state internal consistency, and action internal consistency, that is, gait state s i With concept d k The matching degree, the state s within the same concept category i ,s j Similarity between actions in the same concept category i ,a j These evaluation indicators help the exoskeleton system determine the degree of match between each gait state and its corresponding concept, thereby adjusting its motion control strategy to better meet user needs. k With state s i The similarity can be understood as the semantic consistency between the exoskeleton in the current gait state and the preset concept:

[0142]

[0143] f CLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept d k CLIP representation.

[0144] Evaluating the similarity between different states within the same category actually means that the exoskeleton system should be able to maintain consistent motion patterns under the same category of gait states. For example, assuming state s i and state s j Belong to the same category C k , the internal consistency of their states is expressed as:

[0145]

[0146] Sim cos (s i , s j ) is the state s i and s j The cosine similarity between .

[0147] Calculate and evaluate the similarity between different actions within the same category. In gait recognition, similar gait categories should have similar motion characteristics. This requires the exoskeleton system to take consistent actions based on the gait category, and within the same category, the changes in motion should conform to physical laws to ensure a natural and stable gait. The internal consistency of the action is expressed as:

[0148]

[0149] The final fitness score is:

[0150] F k =Sim CLIP (s i , d k )×SC×AC

[0151] The process of obtaining the gait concept description value by performing concept evolution based on fitness score:

[0152] In an exoskeleton system, the user's gait will vary over time or with different tasks. Therefore, during the concept evolution process, the system dynamically adapts to the changes in the user's gait. That is, it needs to continuously adjust its gait control strategy in different task environments to adapt to changes in user behavior and needs. When the fitness of certain concepts is low, new variants will be regenerated through GPT-2 to optimize the current gait control strategy. Set it as the fitness threshold F threshold When the fitness of a concept is lower than this threshold, the concept will be automatically updated to ensure that gait control is accurate and natural:

[0153]

[0154] G GPT-2 (D k ) is a conceptual description regenerated by GPT-2, enabling the exoskeleton system to adapt to the needs of different gaits and movement patterns.

[0155] S102: Perform frequency domain and concept space fusion optimization and feature fusion on the user gait data and gait concept description values based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

[0156] The process of performing frequency domain and concept space fusion optimization on the user gait data and the gait concept description value based on Fourier transform includes:

[0157] Frequency domain feature conversion

[0158] Gait data from exoskeleton systems often contains periodic features, but these features are not obvious in the time domain. The Fourier transform can effectively extract these periodic features, converting time-domain data into a frequency-domain representation, and better capturing the periodic changes in gait. Specifically, the Fourier transform converts the input gait data into a set of frequency components, discovering the periodic patterns hidden in the data. Physically, this process simulates how to extract frequency-domain features from complex motion posture changes, namely, recurring similar motion postures, to achieve efficient representation of gait characteristics in the frequency domain.

[0159] Assume that the input gait data x is a tensor of shape (B, C, H, W), where B is the batch size, C is the number of channels, H, W

[0160] Is the spatial dimension. Perform Fourier transform on the gait data x:

[0161]

[0162] After Fourier transform, the exoskeleton system can extract periodic features from the time domain, providing a basis for the subsequent fusion of frequency domain and concept space.

[0163] Mapping frequency domain features to concept space

[0164] Fourier transformed data X FFT Is a complex matrix, so take its mean representation (for channel dimension

[0165] The mean is then mapped to the CLIP concept space via the freq_to_semantic network, yielding a conceptual representation of the gait state. This process is similar to linking frequency domain information with the exoskeleton system's concepts in the semantic space. This optimizes the gait feature representation, ensuring that gait characteristics are not only correlated with their own frequency information but also aligned with concepts in the semantic concept space. This allows the exoskeleton system to understand the relationship between gait characteristics and specific task requirements, enabling more precise movement control.

[0166]

[0167] S sem is the conceptual representation of Fourier characteristics, f freq2sem It is a two-layer MLP used to convert Fourier features into clip dim In this process, the system can not only capture the frequency domain features of gait, but also use the concept space of the CLIP model to optimize the representation of gait.

[0168] Calculating concept similarity

[0169] Calculate the frequency domain feature S sem With CLIP Concept Cclip The cosine similarity between them is used to evaluate the matching degree between the current state of the exoskeleton system and the task-related concepts:

[0170]

[0171] The similarity indicates the matching degree between the input data and each concept. The concept C that is most similar to the current gait is selected. k , and perform gait optimization based on this concept:

[0172]

[0173] Generating concept attention weights

[0174] Select the most similar concept C best , the exoskeleton system will input into the semantic attention The network generates the corresponding attention weights. The network generates the importance weights for each frequency component based on the semantic similarity of the concepts. Physically, this process reflects the adjustment of the relative importance of different frequency components in the frequency domain based on the matching degree between the concept and the gait state. The calculated attention weight matrix is used to weight different frequency components to match the gait requirements of the current task:

[0175] W att =f att (C best )

[0176] where f att It is an MLP that generates truncation size ×truncation size The attention weight matrix of shape indicates the importance of different frequency components. Normalized: W att =σ(W att ), σ represents the sigmoid activation function, and the weight is set between [0,1].

[0177] Concept-guided Fourier features

[0178] According to the calculated concept similarity and attention weight, the exoskeleton system will modulate the frequency components of the gait data. Specifically, the generated attention weight W att The features in the frequency domain are weighted to enhance the frequency components related to the current task and suppress the frequency components irrelevant to the task. This illustrates how to filter effective movements and postures during movement in the frequency domain, so that the gait system can respond to task requirements more accurately and avoid redundant and irrelevant information affecting the task:

[0179] X′ FFT =X FFT ×(1+λWatt )

[0180] λ is a hyperparameter that controls the influence of semantic concepts on frequency domain features. Concept-related frequency components will be enhanced (1+λW att >1). Conceptually irrelevant frequency components will be suppressed (1+λW att <1).

[0181] Inverse Fourier transform to restore data

[0182] The frequency domain data after conceptual modulation needs to be converted back to the time domain, that is, the modulated frequency domain data is restored to the time domain gait signal through inverse Fourier transform. Physically, this is to remap the optimized features in the frequency domain to the actions that the exoskeleton control system can perform. The frequency domain data X′ after conceptual modulation FFT Back to the time domain: Finally, a concept-optimized time-domain gait signal x' is obtained, whose frequency-domain features are aligned with the concept information.

[0183] Cross-task concept transfer

[0184] To enhance the adaptability of the exoskeleton system, a concept knowledge base is established to store gait concepts related to different tasks. This cross-task concept transfer can store and manage task-related concepts, forming a concept knowledge base. This allows for the search for concepts most similar to the new task, enabling cross-task knowledge sharing and progressive adaptation to new tasks. Concepts from the source and target tasks are blended based on the level of adaptation. When generating hybrid concepts, the concept representation of the new task incorporates both historical knowledge and the characteristics of the new task.

[0185] Conceptual Storage

[0186] The concept set of a task T is represented as: C T ={D1, D2, ..., D N}, D i is the concept description in task T, C T is the set of all concepts of task T. T Store in warehouse Record concept fitness: in It's Concept D i The fitness score of a concept reflects its effectiveness in task T. The exoskeleton system can assess which historical concepts have high adaptability to the new task and then decide whether to transfer them. The fitness of a concept is a quantitative evaluation of the system's actual performance of the gait concept in the task.

[0187] Concept similarity calculation

[0188] In order to achieve cross-task concept transfer, the exoskeleton system needs toB and historical missions T A Calculate the similarity of concepts and select the most similar concepts for migration, that is, in cross-task concept migration, it is necessary to find the new task T B With T in the library A The most similar concept. Let D A，i It is task T A The concept of D B，j It is a new task T B The query concept in B is calculated using the CLIP model to calculate its cosine similarity:

[0189]

[0190] f CLIP (D A，i ) and f CLIP (D B，j ) is the feature vector extracted by CLIP. Calculate T B Each D B，j With T A All D Ai的 Similarity:

[0191]

[0192] Select the most similar top k D A，i , used for transfer. This process calculates the similarity between tasks to help the exoskeleton system transfer important gait features from historical tasks, avoiding learning from scratch.

[0193] Gradual concept transfer

[0194] In cross-task transfer, the exoskeleton system will gradually adapt to the new task. A to T B During the migration process, dynamically adjust T A and T B The mixing ratio is such that T B Gradually adapt to the new task, let α be the adaptation measure, range [0,1], when α = 0, fully use T A The concept of α=1 is fully used. B The existing concept is that when α is between 0 and 1, T is used. A and T B The new task can effectively utilize the concept of historical tasks and add new features, namely:

[0195]

[0196] C A is the source task T A The most similar concept set selected, CB is the target task T B Current concepts of controlling the mixing ratio

[0197] α balances historical knowledge and new task requirements to adapt to new tasks more efficiently.

[0198] Controlling the mixing quantity can be formalized as:

[0199]

[0200] final Include A target task concept, The concept of a source task.

[0201] Generate hybrid concepts

[0202] based on Generate a more consistent T B The concept of task characteristics. Let DB, i and DA ，j yes Use GPT-2 to generate the fused concept: D hybrid =G GPT-2 (D B，i , D A，j ), indicating input D B，i and D A，j Finally, the synthetic concept description generated by GPT-2, the final mixed concept set is

[0203] The GPT-2 model generates a new synthetic concept description based on the descriptions of the source and target tasks, aiming to integrate the characteristics of both tasks. In practice, this process uses natural language processing models to transfer knowledge between tasks. The resulting hybrid concept reflects the characteristics of the historical tasks while adapting to the needs of the new task, enabling the exoskeleton system to handle complex and dynamic tasks.

[0204] Concept-guided behavioral consistency optimization

[0205] Concept class consistency loss (L SC )

[0206] During the gait optimization process for the ReWalk exoskeleton system, the states and actions of each gait category should be consistent. The exoskeleton system's responses should remain consistent under similar motion states. The SC loss ensures consistency between states and actions within the same category by classifying the state data into multiple categories, with the states and actions within each category sharing similar characteristics.

[0207] Assume s i and sj Same category C k The two states under a i and a j are their corresponding actions. The means of state and action are respectively feature and mean action The goal is to make the mean of state and action distance as small as possible. k The state features should be close to each other, and the motion features should also be consistent to ensure the smoothness and consistency of the exoskeleton under different gaits. The state feature consistency loss is expressed as:

[0208]

[0209] This loss function requires that there should be a small distance between states of the same category, and the response of the exoskeleton should be uniform under similar motion states.

[0210] The action feature consistency loss is expressed as:

[0211]

[0212] This loss function requires that actions in similar states should be similar, and the actions performed by the exoskeleton should have a high degree of matching with the gait characteristics.

[0213] The final concept intra-class consistency loss is expressed as:

[0214] SC total =SC state +SC action

[0215] Since the class size |C k | may be different, the loss of each category will be normalized, and each category contributes fairly to the loss:

[0216]

[0217] Concept class consistency loss (L cilp )

[0218] In the exoskeleton system, in order to improve the adaptability of different gait patterns, it is necessary to calculate the consistency loss between concept classes. i and s j It is the state feature, calculate the state feature state features The Euclidean distance between them gives the distance matrix of state features:

[0219] D state (s i , s j )=‖s i -sj ‖2

[0220] Assume a i and a j is the action feature, calculate the Euclidean distance between the action features, and get the distance matrix of the action features:

[0221] D action (a i , a j )=‖a i -a j ‖2

[0222] Calculate the similarity of the CLIP model output to obtain the CLIP similarity of each pair of states:

[0223]

[0224] f CLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept description d k CLIP representation.

[0225] Normalize the similarity so that the distance between states is more consistent with the similarity in the concept space:

[0226]

[0227] Converting the similarity between states into distance can quantify the differences between different states. The larger the distance value, the greater the difference between the states:

[0228] D clip =1-Sim CLIP

[0229] The error between the distance between the state feature and the CLIP distance, and the error between the distance between the action feature and the CLIP distance is used as the loss:

[0230]

[0231] The total loss is the weighted sum of the intra-concept class consistency loss and the CLIP-guided inter-concept class consistency loss:

[0232] L total =λ SC ×L SC +λ clip ×L clip

[0233] This loss function enables the exoskeleton system to better adapt to different gait patterns and improve performance on new tasks by minimizing the distance between state and action features and the concept space.

[0234] Multi-scale convolutional residual Fourier layer

[0235] Multi-scale convolution module

[0236] In the gait analysis of exoskeleton systems, the multi-scale convolution module can capture gait features at different scales. Gait data usually has complex periodicity and local patterns. The multi-scale convolution operation can simultaneously model the data at multiple scales and capture the details and global features in the gait signal. Specifically, the input data x is fed into the multi-scale convolution module, which captures the local features in the gait data through convolution operations at multiple scales. Specifically, for the input B is the batch size, Cin is the number of input channels, D=(D1, D2, ..., D n ) is the size of the spatial dimension. This module uses multiple convolutional layers Conv(x) to process the input data and obtain multi-level features:

[0237] x multi-scale =MultiScaleConv(x)

[0238] Convolution simulates the effect of a sliding window in space on gait data. The motion information of adjacent frames affects the features extracted by the convolution kernel. In exoskeleton systems, these features help understand the motion patterns of gait over time, such as stride width and gait period, allowing the exoskeleton to effectively optimize movements under different gaits.

[0239] Block Module and Position Encoding

[0240] In order to further enhance the representation of gait information, this method uses a block module to split the input gait data into specified block sizes. Each block is independently subjected to local feature extraction, that is, the input gait data is divided into specified block sizes (M1, M2, ..., M n ), segmentation. Local features are extracted independently within each block, and the data after segmentation is represented as: x blocks =Block(x,(M1,M2,...,M n )), then, position encoding is added to the features of each position, allowing the exoskeleton system to effectively utilize the spatiotemporal position information of the gait data. The position encoding is generated by a function f(q), where q represents the coordinates of each position in the input data:

[0241] x encoded =x blocks +f(q)

[0242] Convolutional Residual Fourier Layer

[0243] Based on the convolution operation, this method uses a convolutional residual Fourier layer to model frequency domain information and combines it with concept-guided spectral enhancement. The Fourier transform converts the gait signal from the time domain to the frequency domain, capturing the periodic structure. Different frequency components are weighted using the CLIP concept space, thereby enhancing frequency components related to the gait concept and suppressing irrelevant or noisy frequency features.

[0244] CLIP calculates the similarity S of state X in the concept space CLIP (X, D), convert it into frequency domain weighting:

[0245] W semantic =σ(W att ·S CLIP (X, D)

[0246] X enhanced =X FFT ×(1+λ·W semantic )

[0247] W semantic is the spectrum weight guided by the CLIP concept space, W att Calculated by the attention mechanism.

[0248] λ controls the strength of the concept enhancement.

[0249] The output of the convolution residual Fourier layer yout is the Fourier transform output of the Fourier layer yfft after spectrum enhancement and the high-frequency features yconv extracted by the convolution layer are fused through the residual connection:

[0250] y out =y fft +y conv +b

[0251] The Fourier transform provides the system with the spectral characteristics of the gait signal, which helps analyze and optimize gait patterns. By guiding these frequency characteristics, the system can more precisely adjust gait movements, making the exoskeleton more efficient and stable when performing tasks.

[0252] Feature fusion

[0253] Self-attention

[0254] The processed multi-scale features and Fourier domain convolution features are fed into the multi-head self-attention module. This module establishes the dependency between different features through the multi-head attention mechanism, learns the interaction between gait features, learns the interdependence between gait features in parallel from multiple subspaces, and comprehensively considers the features of multiple scales to more comprehensively understand the complex patterns in gait data. Specifically, given the feature matrix The multi-head attention mechanism is calculated by the following formula:

[0255] Y attn =MultiHeadAttention(X features ) The exoskeleton system finds connections between related features in multidimensional space, and when processing gait data, the system is able to notice the dependencies between gait details and overall patterns.

[0256] Cross-Attention

[0257] On this basis, this method also combines the cross attention mechanism, and the interaction between it and other latent variables further improves the ability of feature representation. The cross attention mechanism calculates the cross attention output through the following formula:

[0258] Y cross =CrossAttention(X features , Z latent )

[0259] where Z latent is the representation of latent variables. The cross-attention mechanism captures the relationship between gait features and other potential factors (such as environmental changes, control strategies, etc.), improving the robustness and adaptability of the system.

[0260] Time series aggregation and output generation

[0261] After gait feature extraction and attention mechanism processing, the time series information of the input data is aggregated in the time series aggregation module. This module aggregates the processed gait features according to the time dimension and extracts the temporal dependencies in the gait data. Assume that the processed feature sequence is Y features =[y1, y2, ..., y T ], the time series aggregation operation can be expressed as:

[0262] Y aggregated =TieAggregation(Y features )

[0263] Through these optimization steps, the exoskeleton system can understand and optimize different gait patterns, providing users with personalized assisted motion control and precise and efficient movement. Physically, time series aggregation ensures the continuity and consistency of gait, allowing the exoskeleton system to maintain natural movement and improve efficiency in complex tasks.

[0264] Example 2: The second aspect, as Figure 9 As shown, in order to achieve the above-mentioned purpose, the present invention discloses a frequency domain analysis system for an exoskeleton assistive device based on concept learning, comprising:

[0265] The data processing module 11 is used to receive user gait data, perform state representation and KMeans clustering on the user gait data to obtain the user gait state, perform fitness evaluation based on the user gait state to obtain a fitness score, and perform concept evolution based on the fitness score to obtain a gait concept description value;

[0266] The frequency domain analysis module 12 is used to perform frequency domain and concept space fusion optimization and feature fusion on the user gait data and gait concept description values based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

[0267] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0268] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium having a computer program stored thereon, which executes the above method when executed by a processor. The storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.

[0269] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0270] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements shall fall within the scope of the present disclosure.

Claims

1. A frequency domain analysis method for exoskeleton assistive devices based on concept learning, characterized in that: The method comprises the following steps: Receive user gait data, perform state representation and KMeans clustering on the user gait data to obtain the user gait state, perform fitness evaluation based on the user gait state to obtain a fitness score, and perform concept evolution based on the fitness score to obtain a gait concept description value; The user gait data and gait concept description values are subjected to frequency domain and concept space fusion optimization and feature fusion based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

2. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 1, characterized in that: The process of performing state representation and KMeans clustering on user gait data to obtain the user gait state is as follows: Perform KMeans clustering on the state representation of the gait data to generate initial concepts for different gait states, and cluster to obtain n concepts categories, and use GPT-2 to generate n variants A natural language description, let s i is the state feature of the exoskeleton system, S={s1,s2,...,s N } is the set of all states, and KMeans clustering is used to obtain C k represents the kth cluster, μ k is the mean of the cluster, each C k Generate multiple descriptions via GPT-2 :D k ={G GPT-2 (P k )|P k ∈P},P k It is a task-related prompt word, G GPT-2 Represents text generated by GPT-2.

3. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 2, characterized in that: The process of performing fitness evaluation based on the user's gait state to obtain a fitness score includes: The evaluation of concept quality includes CLIP similarity, state internal consistency, and action internal consistency, that is, gait state s i With concept d k The matching degree, the state s within the same concept category i ,s j Similarity between actions in the same concept category i ,a j The similarity between them is calculated by concept d k With state s i Similarity: f CLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept d k CLIP representation; The internal consistency of the state is expressed as: Sim cos (s i , s j ) is the state s i and s j The cosine similarity between The internal consistency AC of an action is expressed as: Final fitness F k Rating: F k =Sim CLIP (s i ,d k )×SC×AC。 4. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 3 is characterized in that: The process of obtaining the gait concept description value by performing concept evolution based on fitness score: In the exoskeleton system, the user's gait will change with time or task, so the exoskeleton system dynamically adapts to the changes in the user's gait during the concept evolution process. threshold is the fitness threshold, and we have: G GPT-2 (D k ) is a conceptual description regenerated by GPT-2, allowing the exoskeleton system to adapt to the needs of different gaits and movement patterns.

5. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 1, characterized in that: The process of performing frequency domain and concept space fusion optimization on the user gait data and the gait concept description value based on Fourier transform includes: Frequency domain feature conversion: The gait data in the exoskeleton system contains periodic features, which are extracted using Fourier transform. At this time, Fourier transform and CLIP concept representation are combined to optimize feature learning; Assume that the input gait data x is a tensor of shape (B, C, H, W), where B is the batch size, C is the number of channels, and H and W are the spatial dimensions. Perform Fourier transform on the gait data x: Frequency domain features are mapped to concept space: Fourier transformed data X FFT It is a complex matrix, so we take its mean representation and map it to the CLIP concept space through the freq_to_semantic network to obtain the conceptual representation of the gait state: S sem is the conceptual representation of Fourier characteristics, f freq2sem It is a two-layer MLP used to convert Fourier features into clip dim Conceptual representation of dimension; Calculate concept similarity: Calculate the frequency domain feature S sem With CLIP Concept C clip The cosine similarity between them is used to evaluate the matching degree between the current state of the exoskeleton system and the task-related concepts: Cosine similarity indicates the matching degree between the input data and each concept. The concept C that is most similar to the current gait is selected. k , and perform gait optimization based on the concept: Generate concept attention weights: Select the most similar concept C best , enter semantic attention network: W att =f att (C best ) where f att It is an MLP that generates truncation size ×truncation size The attention weight matrix of the shape, ; normalized: W att =σ(W att ), σ represents the sigmoid activation function, and the weight is set between [0,1]; Concept-guided Fourier features: The concept attention weight is generated by the similarity of the concept, and the frequency component of the gait data is modulated, and W is used att Perform frequency domain modulation: X′ FFT =X FFT ×(1+λW att ) λ is a hyperparameter that controls the influence of semantic concepts on frequency domain features. Concept-related frequency components will be enhanced: 1+λW att >1; conceptually irrelevant frequency components will be suppressed: 1+λW att <1; Inverse Fourier transform to restore data The frequency domain data X′ after conceptual modulation FFT Returning to the time domain, the gait task is performed in real time: Finally, the concept-optimized time-domain gait signal x' is obtained, which aligns the frequency-domain features with the concept information.

6. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 5, characterized in that: After the frequency domain and concept space fusion optimization based on Fourier transform, the multi-scale convolution residual Fourier layer is entered, including: Multi-scale convolution module: The multi-scale convolution module captures gait features of different scales. The input gait data x is fed into the multi-scale convolution module. The multi-scale convolution module captures the local features in the gait data through convolution operations of multiple scales. B is the batch size, Cin is the number of input channels, D=(D1, D2, ..., D n ) is the size of the spatial dimension. This module uses multiple convolutional layers Conv(x) to process the input data and obtain multi-level features: x multi-scale =MultiScaleConv(x) Block module and position encoding: The input gait data is divided into blocks of the specified size (M1, M2, ..., M n ), segmentation is performed, and local features are extracted independently within each block. The data after segmentation is represented as: x blocks =Block(x,(M1,M2,...,M n )), then, position encoding is added to the features of each position, so that the exoskeleton system can use the spatiotemporal position information of the gait data. The position encoding is generated by the function f(q), where q represents the coordinates of each position in the input data: x encoded =x blocks +f(q) Convolutional Residual Fourier Layer: Fourier transform converts gait signals from the time domain to the frequency domain to capture periodic structures. Different frequency components are weighted through the CLIP concept space, so that the frequency components related to the gait concept are enhanced. CLIP calculates the similarity S of state X in the concept space CLIP (X, D), convert it into frequency domain weighting: W semantic =σ(W att ·S CLIP (X,D)) X enhanced =X FFT ×(1+λ·W semantic ) W semantic is the spectrum weight guided by the CLIP concept space, W att Calculated by the attention mechanism, λ controls the strength of concept enhancement; The output of the convolution residual Fourier layer yout is the Fourier transform output of the Fourier layer yfft after spectrum enhancement and the high-frequency features yconv extracted by the convolution layer are fused through the residual connection: and out =and fft +y conv +b Feature fusion: The processed multi-scale features and Fourier domain convolution features are fed into the multi-head self-attention module, which establishes the dependency between different features through the multi-head attention mechanism and learns the interaction between gait features. Given the feature matrix The multi-head attention mechanism is calculated by the following formula: Y attn =MultiHeadAttention(X features ) Cross-Attention: The crisscross attention mechanism calculates the crisscross attention output using the following formula: Y cross =CrossAttention(X features ,Z latent ) where Z latent is the representation of the latent variable. Time series aggregation and output generation: After gait feature extraction and cross-attention mechanism processing, the time series information of the input data is aggregated in the time series aggregation module. The time series aggregation module aggregates the processed gait features according to the time dimension and extracts the temporal dependency in the gait data. Let the processed feature sequence be Y features =[y1, y2, ..., y T ], the time series aggregation operation can be expressed as: Y aggregated =TimeAggregation(Y features )。 7. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 6, characterized in that: The concept storage process of the pre-established concept knowledge base: The concept set of a task T is represented as: C T ={D1, D2, ..., D N }, D i is the concept description in task T, C T is the set of all concepts of task T, C T Store in warehouse Record concept fitness: in It's Concept D i The fitness score of Concept similarity calculation process: According to the current task T B and historical missions T A Calculate the similarity of concepts and select the most similar concept to transfer. Let D A，i It is task T A The concept of D B，j It is a new task T B The query concept in , uses the CLIP model to calculate the cosine similarity: f CLIP (D A，i ) and f CLIP (D B，j ) is the feature vector extracted by CLIP, calculate T B Each D B，j With T A All D A，i Similarity: Select the most similar top k D A，i ; The process of gradual concept transfer: In T A to T B During the migration process, dynamically adjust T A and T B The mixing ratio is such that T B Gradually adapt to the new task, let α be the adaptation measure, range [0,1], when α = 0, fully use T A The concept of α=1 is fully used. B The existing concept is that when α is between 0 and 1, T is used. A and T B The new task uses the concept of historical tasks and adds new features, namely: C A is the source task T A The most similar concept set selected, C B is the target task T B existing concepts; Controlling the mixing quantity can be formalized as: Include A target task concept, Source task concept; Process for generating hybrid concepts: based on Generate a T B The concept of task characteristics, let D B，i and D A，j yes Two concepts in, use GPT-2 to generate the fused concept: D hybrid =G GPT-2 (D B，i , D A，j ), indicating input D B，i and D A，j After that, the synthetic concept description generated by GPT-2, the resulting mixed concept set is 8. The frequency domain analysis method for an exoskeleton assistive device based on concept learning according to claim 7, characterized in that: The concept guides the behavioral consistency optimization process: Concept intra-class consistency loss L SC : Assume s i and s j Same category C k The two states under a i and a j are their corresponding actions. The means of state and action are respectively feature and mean action , gait category C k The state features of are close to each other, the action features are consistent, and the state feature consistency loss is expressed as: The action feature consistency loss is expressed as: The final concept intra-class consistency loss is expressed as: SC total =SC state +SC action Since the class size |C k |There are different situations. The loss of each category is normalized so that each category contributes fairly to the loss: Concept class consistency loss L clip : Assume s i and s j It is the state feature, calculate the state feature state features The Euclidean distance between them gives the distance matrix of state features: D state (s i ,s j )=‖s i -s j ‖2 Assume a i and a j is the action feature, calculate the Euclidean distance between the action features, and get the distance matrix of the action features: D action (a i ,a j )=‖a i -a j ‖2 Calculate the similarity of the CLIP model output to obtain the CLIP similarity of each pair of states: f CLIP (s i ) is the state s i CLIP indicates that g CLIP (d k ) is the concept description d k CLIP representation; Normalize the similarity so that the distance between states conforms to the similarity in the concept space: Convert similarity to distance: D clip =1-Yes CLIP The error between the distance between the state feature and the CLIP distance, and the error between the distance between the action feature and the CLIP distance is used as the loss: The total loss is the weighted sum of the intra-concept class consistency loss and the CLIP-guided inter-concept class consistency loss: L total =λ SC ×L SC +λ clip ×L clip 。 9. A frequency domain analysis system for exoskeleton assistive devices based on concept learning, characterized in that: include: A data processing module is used to receive user gait data, perform state representation and KMeans clustering on the user gait data to obtain the user gait state, perform fitness evaluation based on the user gait state to obtain a fitness score, and perform concept evolution based on the fitness score to obtain a gait concept description value; The frequency domain analysis module is used to perform frequency domain and concept space fusion optimization and feature fusion on the user gait data and gait concept description values based on Fourier transform to obtain an aggregated time series, wherein the gait concept description values are input into a pre-established concept knowledge base for storage.

10. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the frequency domain analysis method of the exoskeleton assistive device based on concept learning according to any one of claims 1 to 8 is adopted.

Images