Medical wire tension adaptive adjustment device and method based on artificial intelligence

Abstract

The present invention relates to the technical field of medical wires, and discloses an artificial intelligence-based medical wire tension adaptive adjustment device and method, the method comprising: collecting historical tension data of the medical wire and wire-environment interface impedance spectrum data; performing multi-scale feature extraction to obtain multi-scale feature representation; performing cross-channel multi-head attention analysis and gated depth separable convolution operations to obtain target tension sensitive features; constructing a first tension distribution model and an intelligent impedance matching model; acquiring real-time tension data and inputting the data into the first tension distribution model for tension adjustment to obtain an initial tension adjustment strategy, and performing strategy optimization through the intelligent impedance matching model to obtain a target tension adjustment strategy; executing the target tension adjustment strategy and performing feedback update to obtain a second tension distribution model, the present invention realizes more intelligent, accurate and safe medical wire tension adaptive adjustment.

Description

Technical Field

[0001] The present invention relates to the technical field of medical wires, and in particular to an artificial intelligence-based medical wire tension adaptive adjustment device and method. Background Art

[0002] The tension control of medical wires directly affects the accuracy of the operation. Traditional tension adjustment methods mainly rely on experience and subjective judgment, which is difficult to adapt to complex and changing surgical environments and individual needs. However, existing tension adjustment technologies often ignore the dynamic interaction process between wires and surrounding tissues, and cannot effectively deal with the uncertainty and nonlinear factors in the surgical process.

[0003] In addition, medical cables face challenges of multiple sources of interference and environmental changes during use, such as tissue deformation, body fluid effects, and electromagnetic interference, which can lead to instability and unpredictability of tension signals. Traditional single models or fixed parameter control strategies have difficulty capturing these complex dynamic characteristics, thus affecting the accuracy and robustness of tension regulation. At the same time, different types of medical cables and surgical scenarios have their own unique requirements for tension control, and how to implement an adaptive tension regulation strategy has become an urgent problem to be solved. Summary of the invention

[0004] The present invention provides an artificial intelligence-based medical wire tension adaptive adjustment device and method, which realizes more intelligent, accurate and safe medical wire tension adaptive adjustment.

[0005] In a first aspect, the present invention provides a method for self-adapting tension of a medical wire based on artificial intelligence, the method comprising:

[0006] Collect historical tension data of medical wires and wire-environment interface impedance spectrum data;

[0007] Inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation;

[0008] Performing cross-channel multi-head attention analysis and gated depth-separable convolution operations on the multi-scale feature representation to obtain target tension-sensitive features;

[0009] Constructing a first tension distribution model according to the target tension-sensitive characteristics, and constructing an intelligent impedance matching model according to the wire-environment interface impedance spectrum data;

[0010] Acquire real-time tension data and input it into the first tension distribution model to perform tension adjustment, obtain an initial tension adjustment strategy, and optimize the strategy through the intelligent impedance matching model to obtain a target tension adjustment strategy;

[0011] The target tension adjustment strategy is executed and the first tension distribution model is updated through feedback to obtain a second tension distribution model.

[0012] In a second aspect, the present invention provides an artificial intelligence-based medical wire tension adaptive adjustment device, the artificial intelligence-based medical wire tension adaptive adjustment device comprising:

[0013] The acquisition module is used to collect historical tension data of medical wires and wire-environment interface impedance spectrum data;

[0014] An extraction module, used for inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation;

[0015] An analysis module, configured to perform cross-channel multi-head attention analysis and gated depth-separable convolution operations on the multi-scale feature representation to obtain target tension-sensitive features;

[0016] A construction module, used to construct a first tension distribution model according to the target tension-sensitive characteristics, and to construct an intelligent impedance matching model according to the wire-environment interface impedance spectrum data;

[0017] An adjustment module, used for acquiring real-time tension data and inputting the first tension distribution model to perform tension adjustment, obtaining an initial tension adjustment strategy, and optimizing the strategy through the intelligent impedance matching model to obtain a target tension adjustment strategy;

[0018] A feedback module is used to execute the target tension adjustment strategy and perform feedback update on the first tension distribution model to obtain a second tension distribution model.

[0019] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned artificial intelligence-based adaptive tension adjustment method for medical wires.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned artificial intelligence-based adaptive tension adjustment method for medical wires.

[0021] In the technical solution provided by the present invention, multi-scale feature extraction is performed on historical tension data through a multi-scale Transformer neural network, which can capture the tension change pattern on different time scales and improve the comprehensiveness and robustness of feature representation. Cross-channel multi-head attention analysis of multi-scale feature representation is helpful to capture the relationship between different feature channels and enhance the model's ability to understand complex tension patterns. The gated deep separable convolution operation is adopted to reduce the number of parameters while maintaining the model's expressiveness, thereby improving the computational efficiency and generalization ability of the model. An intelligent impedance matching model is constructed, which can dynamically adapt to changes in the wire-environment interface, optimize the tension adjustment strategy, and improve the system's adaptability to environmental changes. By executing the target tension adjustment strategy and performing feedback updates, the dynamic optimization of the tension distribution model is realized, and fractional-order calculus operations and fractional-order transfer functions are introduced to more accurately describe complex impedance characteristics and improve the accuracy of impedance matching. The initial tension adjustment strategy is optimized by multiple objectives, taking into account multiple performance indicators of tension control, and achieving a more comprehensive and balanced adjustment effect. The adaptive neural fuzzy inference system is used for parameter adjustment, combining the learning ability of the neural network and the interpretability of fuzzy logic, thereby improving the interpretability and reliability of the model. The second tension distribution model is generated by conditional variational inference, which can better capture the uncertainty of tension distribution. The present invention realizes more intelligent, accurate and safer adaptive adjustment of the tension of medical wires. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 Schematic diagram of the steps of a method for adaptively adjusting tension of a medical wire based on artificial intelligence in an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of the structure of an artificial intelligence-based adaptive tension adjustment device for medical wires in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] Embodiments of the present invention provide an artificial intelligence-based adaptive tension adjustment device and method for medical wires. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, an embodiment of a method for adaptively adjusting tension of a medical wire based on artificial intelligence includes:

[0027] Step S1, collecting historical tension data of medical wires and wire-environment interface impedance spectrum data;

[0028] It is understandable that the execution subject of the present invention can be an artificial intelligence-based medical wire tension adaptive adjustment device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0029] Specifically, a multi-frequency and multi-amplitude AC excitation signal is applied to the medical wire. Through the preset excitation scheme, the excitation signal covers different frequency ranges from low frequency to high frequency, and a variety of different amplitudes are applied at each frequency to fully capture the response characteristics of the medical wire under various electrical conditions. In this process, the tension sensor and the electrochemical impedance spectrometer will work simultaneously to monitor and record the tension and electrochemical characteristics of the medical wire in real time. The tension sensor will capture the tension changes of the wire under different excitation conditions to obtain the original tension time series data, while the electrochemical impedance spectrometer will record the impedance response of the wire under these excitation conditions to obtain the corresponding impedance spectrum response data. The original tension time series is filtered. The filtering process can adopt a variety of filtering technologies such as low-pass filtering, high-pass filtering or band-pass filtering, and the specific selection depends on the frequency range of the tension signal to be retained. Through filtering, the noise component can be effectively removed to obtain purer historical tension data. At the same time, the impedance spectrum response data is aligned in time and frequency to ensure that the impedance data collected at different time points can correctly correspond to the corresponding frequency conditions to avoid data distortion caused by time or frequency misalignment. Time-frequency alignment can be achieved through cross-correlation analysis, short-time Fourier transform and other technologies. By aligning the impedance data to a unified time-frequency coordinate system, the accuracy and consistency of the analysis results can be ensured, and the wire-environment interface impedance spectrum data can be obtained.

[0030] Step S2, inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation;

[0031] Specifically, the historical tension data is subjected to wavelet multi-resolution decomposition, and the original tension signal is decomposed into a set of subsequences with different frequency components, each of which corresponds to different scale features of the tension signal. The subsequence sets are positionally encoded, and position information is added to the features of each time step to ensure that the model can identify the positional relationship of the features in the sequence in the subsequent processing process. By embedding the position information into the subsequence, the network can better understand the temporal characteristics of the data and obtain the encoded subsequence. The encoded subsequence is input into the multi-head self-attention module in the multi-scale Transformer neural network for feature weighting operation. In this module, the network performs weighted operation on the input features through the multi-head self-attention mechanism. The multi-head self-attention mechanism can capture features in parallel in different feature subspaces and assign appropriate weights to each input feature to obtain attention-weighted feature representation. The attention-weighted feature representation is input into the feedforward neural network in the multi-scale Transformer neural network. The feedforward neural network consists of two layers of linear transformation and ReLU activation function. The input features are mapped to high-dimensional space through linear transformation. The ReLU activation function introduces nonlinearity, so that the model can capture more complex feature relationships and obtain the output of the single-scale Transformer block. The output of the single-scale Transformer block is input into the cross-scale fusion module in the multi-scale Transformer neural network. The features of different scales are processed uniformly through adaptive pooling and convolution operations. Adaptive pooling can automatically adjust the pooling parameters according to the characteristics of the data, while the convolution operation can extract common features between different scales and realize the unified representation of features of different scales. The unified representation of features of different scales is analyzed by multi-scale attention mechanism. By calculating the correlation matrix between scales, the relationship between features of different scales is identified, and corresponding weights are assigned to the features of each scale to obtain the scale-weighted feature representation. The scale-weighted feature representation is input into the residual connection module. In the residual connection module, the scale-weighted feature representation is combined with the original input features through addition operation to form a residual enhanced feature representation. By introducing residual connection, the risk of information loss in deep network is reduced, and the stability and effectiveness of the model are ensured. The layer normalization operation is performed on the residual enhanced feature representation. Layer normalization calculates the mean and variance of the feature and performs a linear transformation to make the feature distribution more uniform, which helps to speed up the convergence of the network and improve the performance of the model, and obtain a multi-scale feature representation.

[0032] Pyramid pooling is performed on the output feature map of the single-scale Transformer block. Pyramid pooling is a multi-level downsampling technique that generates a multi-scale feature pyramid by pooling feature maps at different spatial resolutions. Each scale represents the information of the feature at different spatial resolutions, allowing the model to capture richer spatial context information. A 1x1 convolution operation is performed on each scale feature in the feature pyramid, and the number of channels is adjusted to ensure that the features of all scales have the same dimension, forming a multi-scale feature set with unified dimensions. The 1x1 convolution can not only effectively reduce the computational complexity, but also realize the redistribution of channel dimensions to ensure the consistency of multi-scale features in subsequent processing. The unified multi-scale feature set is input into the self-attention mechanism module for cross-scale attention weight calculation. In this module, the model generates inter-scale correlation features by calculating the mutual correlation between features of different scales. Bilinear interpolation is performed on the inter-scale correlation features, and all features of different scales are adjusted to the same spatial resolution through upsampling. Bilinear interpolation can smoothly unify the resolution of the feature map while maintaining the feature details, ensuring that features of different scales can be effectively fused in the subsequent processing. The upsampled features are input into the channel attention module for a squeeze-excitation operation. The squeeze-excitation operation is a technique that emphasizes important features and suppresses irrelevant features by adaptively adjusting channel weights, so that the model can weight different feature channels according to their importance and obtain a more accurate channel-weighted feature representation. The channel-weighted feature representation is subjected to a cross-scale fusion operation to organically combine features from different scales to form a comprehensive cross-scale feature representation. The fusion not only retains the advantages of features at different scales, but also enhances the robustness of the model through information sharing and complementarity. The cross-scale fused features are input into the residual dense block for processing. The residual dense block uses multiple layers of dense connections and short-circuit connections to efficiently transfer and fuse features between multiple network layers, thereby enhancing the expressive power of multi-scale features. The enhanced multi-scale features are subjected to a feature recalibration operation. By re-adjusting the weights of the features, the contributions of features at different scales are ensured to be properly balanced in the final representation. Feature recalibration can be achieved through operations such as linear transformation and normalization, and finally a unified representation of features at different scales is obtained.

[0033] Step S3, performing cross-channel multi-head attention analysis and gated depth-separable convolution operations on the multi-scale feature representation to obtain target tension-sensitive features;

[0034] Specifically, the multi-scale feature representation is subjected to channel grouping operation. The channels of the multi-scale feature representation are evenly divided and decomposed into multiple feature subsets. Each feature subset represents a part of the original feature. The grouping operation helps to reduce the computational complexity and make subsequent processing more efficient. The feature subsets are input into the cross-channel linear projection module. In the linear projection, the query, key and value matrices are calculated respectively through three independent fully connected layers. The query matrix is ​​used to represent the features that need to be paid attention to at present, the key matrix is ​​used to represent the correlation of all features, and the value matrix contains the actual content information of the features. The scaled dot product attention calculation is performed based on the query, key and value matrices. By calculating the dot product of the query and the key and scaling them, the model can identify the importance of each feature and generate multi-head attention output. The advantage of the multi-head attention mechanism is that it can capture the complex relationship between features in parallel in different subspaces, which helps to extract key features related to tension regulation. The multi-head attention output is subjected to channel reorganization and splicing operations, and the cross-channel attention features are generated through linear transformation. The generated cross-channel attention features are input into the gating unit for processing. In the gated unit, the cross-channel attention features are operated by the sigmoid activation function to obtain the channel importance weights. The sigmoid function can compress the attention features into the range of 0 to 1, so that the model can effectively evaluate the importance of each channel. For the original multi-scale feature representation, channel-by-channel convolution and point-by-point convolution are performed in sequence to generate preliminary convolution features. The channel-by-channel convolution operation can capture the detailed features in each channel, while the point-by-point convolution can integrate information between channels to obtain more comprehensive preliminary convolution features. The channel importance weights are multiplied element-by-element with the preliminary convolution features. By weighting each convolution feature, the model can enhance or suppress the features according to the importance of the channel to generate weighted convolution features. The weighted convolution features are subjected to residual connection and layer normalization operations. The residual connection can retain the important information in the input features by adding the weighted convolution features to the original input features, while avoiding the gradient vanishing problem that may occur in deep network training. The layer normalization operation standardizes the mean and variance of the features to ensure that the features maintain a stable distribution in subsequent processing. After residual connection and layer normalization, the target tension-sensitive features are obtained.

[0035] Channel dimension analysis is performed on the multi-scale feature representation to evaluate the importance of each channel in the feature representation. Through a series of statistical or gradient-based methods, the model can rank the importance of each channel and generate a channel importance ranking. Based on the ranking results, the threshold for feature channel grouping is calculated and the grouping parameters are determined. Channel rearrangement and grouping operations are performed on the multi-scale feature representation based on the grouping parameters. Through channel rearrangement, the channels in the feature representation are rearranged according to their importance so that more critical channels are concentrated in specific subsets. According to the preset grouping parameters, the arranged channels are divided into several initial subsets. The initial subsets are adaptively normalized. Adaptive normalization eliminates possible distribution differences between different subsets by standardizing the mean and variance of the features of each subset. The normalized multiple feature subsets are input into the fully connected layer with shared weights for linear transformation and nonlinear activation function operations. The fully connected layer with shared weights ensures that all feature subsets are processed under the same conditions to generate intermediate feature representations. Linear transformation allows features to be remapped in high-dimensional space, while nonlinear activation functions introduce nonlinear characteristics, allowing the model to capture complex feature relationships. The intermediate feature representation obtained is a multi-scale feature after preliminary processing, which contains the potential patterns and relationships between different subsets. The intermediate feature representation is subjected to channel shuffling. Channel shuffling can better fuse the information between different channels by rearranging the channel order of features randomly or according to specific rules, generating a hybrid feature representation. The hybrid feature representation breaks the limitations of the original feature subsets, allowing the model to find associations in a wider feature space. The hybrid feature representation is input into three independent fully connected layers, and linear transformations of different weight matrices are performed to generate query, key, and value matrices respectively. These matrices represent the role of different feature subsets in the attention mechanism, respectively. The query matrix is ​​used to locate important features, the key matrix is ​​used to match related features, and the value matrix carries the actual feature information. These three matrices work together in the subsequent scaled dot product attention calculation to achieve a comprehensive understanding and capture of multi-scale features.

[0036] Step S4, constructing a first tension distribution model according to the target tension sensitivity characteristics, and constructing an intelligent impedance matching model according to the wire-environment interface impedance spectrum data;

[0037] Specifically, the target tension-sensitive feature is subjected to principal component dimensionality reduction processing. Principal component analysis is a commonly used dimensionality reduction technique. By finding the direction with the largest variance in the data, the high-dimensional tension-sensitive feature is mapped to a low-dimensional space to obtain the tension feature after dimensionality reduction. Gaussian mixture clustering is performed on the tension feature after dimensionality reduction. Gaussian mixture model is a probability density estimation method. By assigning data points to different Gaussian distributions, a probability density function of tension distribution is generated, which reflects the probability distribution under different tension conditions and helps to capture the intrinsic structure and pattern of tension features. Based on this probability density function, a variational autoencoder is constructed. The variational autoencoder is a generative model that optimizes its parameters by maximizing the lower bound of evidence so that the model can effectively learn the potential structure of tension distribution. After training and optimization, the first tension distribution model is obtained. At the same time, the time-frequency characteristics and instantaneous frequency characteristics of the impedance spectrum data of the wire-environment interface are extracted. The time-frequency characteristics can reveal the distribution of impedance spectrum data at different times and frequencies, while the instantaneous frequency characteristics reflect the frequency changes of the impedance spectrum at an instant. Based on these characteristics, a complex-valued neural network is constructed. The neural network can process data in the complex domain. Through the training of the complex domain back propagation algorithm, the network can learn the complex impedance spectrum characteristics and generate an impedance spectrum fitting model. Perform fractional calculus operations on the impedance spectrum fitting model. Fractional calculus can handle the nonlinearity and memory effects in the system. By calculating the fractional impedance parameters, the frequency response characteristics of the impedance spectrum are described. Based on the fractional impedance parameters, a fractional transfer function is constructed to fully reflect the dynamic response of the wire-environment interface. The fractional transfer function is input into the adaptive neural fuzzy inference system, and the parameters are adjusted through the hybrid learning algorithm to obtain an intelligent impedance matching model. The adaptive neural fuzzy inference system combines the learning ability of the neural network and the reasoning ability of fuzzy logic. Through learning and reasoning, the model can automatically adjust the matching strategy according to different environmental impedance conditions to achieve precise control and adaptive adjustment of the tension of medical wires.

[0038] Step S5, acquiring real-time tension data and inputting the first tension distribution model to perform tension adjustment, obtaining an initial tension adjustment strategy, and optimizing the strategy through an intelligent impedance matching model to obtain a target tension adjustment strategy;

[0039] Specifically, real-time tension data is obtained, and the real-time tension data is subjected to wavelet packet decomposition and empirical mode decomposition to obtain a first multi-scale tension characteristic component and a first intrinsic mode function. Wavelet packet decomposition can decompose a signal into sub-signals of different frequency ranges, and empirical mode decomposition can decompose a signal into multiple intrinsic mode functions. Based on the multi-scale tension characteristic components and the intrinsic mode functions, the real-time tension data is reconstructed to eliminate noise and unstable factors in the signal, and denoised tension data is obtained. The denoised tension data is input into the first tension distribution model, and the posterior probability of the tension distribution is calculated to obtain the distribution parameters of the tension, which describe the distribution of the real-time tension under different conditions. Gaussian process regression analysis is performed on the tension distribution parameters. Gaussian process regression is a non-parametric Bayesian method that can establish a tension prediction function. This function can not only predict the future tension change trend, but also reflect the relationship between the tension characteristics and its distribution parameters. Based on the tension prediction function, multiple candidate tension adjustment strategies are generated, which represent the adjustment measures that may be taken in different situations. The candidate tension adjustment strategies are evaluated using the Bayesian information criterion. The Bayesian Information Criterion is a criterion for model selection. It can select the best one among multiple candidate strategies and obtain the initial tension adjustment strategy after evaluation. The strategy has the highest adaptability and effectiveness in the current situation. The initial tension adjustment strategy is input into the intelligent impedance matching model. The intelligent impedance matching model performs fuzzy rule reasoning through an adaptive neural fuzzy inference system to generate impedance matching parameters that best match the current situation. The adaptive neural fuzzy inference system combines the learning ability of neural networks and the reasoning ability of fuzzy logic, and can adjust the initial tension adjustment strategy according to real-time impedance data. Based on the generated impedance matching parameters, the initial tension adjustment strategy is optimized for multiple objectives. The optimization process aims to meet multiple adjustment objectives at the same time, such as stability, response speed, and energy efficiency. Through multi-objective optimization, the target tension adjustment strategy is obtained.

[0040] Step S6: Execute the target tension adjustment strategy and perform feedback update on the first tension distribution model to obtain a second tension distribution model.

[0041] Specifically, the target tension adjustment strategy is discretized. By converting the continuous adjustment strategy into a discrete tension adjustment sequence, the specific tension adjustment operation is performed in discrete time steps. According to the discretized tension adjustment sequence, the tension of the medical wire is adjusted, and feedback data is collected in real time during the process. The real-time feedback data records the tension response of the wire after the tension adjustment operation is performed, thereby obtaining tension response data. Wavelet packet transform and empirical mode decomposition are performed on the tension response data. Through wavelet packet transform, the tension response data is decomposed into multiple frequency components, revealing the tension characteristics on different time scales. At the same time, empirical mode decomposition decomposes the data into intrinsic mode functions (IMFs) to reflect the inherent dynamic characteristics of the tension signal. After the decomposition operation, the second multi-scale tension characteristic component and the second intrinsic mode function are obtained. Based on the second multi-scale tension characteristic component and the second intrinsic mode function, time series modeling is performed to establish a dynamic tension prediction model. The dynamic tension prediction model aims to capture the changing trend and pattern of the tension of the medical wire during the adjustment process and predict the future tension state. The established dynamic tension prediction model is integrated with the first tension distribution model. By fusing these two models and combining their respective advantages, a more comprehensive and accurate tension prediction and regulation model is formed. Hyperparameter tuning is performed on the fused tension model. By adjusting the key parameters in the model, the performance of the model is optimized so that it can more accurately reflect the tension distribution and prediction situation, and the optimized tension model is obtained. The latent variable representation of the tension distribution is generated according to the optimized tension model. The latent variable representation is a compression and abstraction of the tension distribution, which can describe the complex distribution characteristics of tension with fewer dimensions. Conditional variational inference is performed on the latent variable representation to estimate the optimal solution of the tension distribution under given conditions and obtain the second tension distribution model.

[0042] In the embodiment of the present invention, multi-scale feature extraction is performed on historical tension data through a multi-scale Transformer neural network, which can capture the tension change pattern on different time scales and improve the comprehensiveness and robustness of feature representation. Cross-channel multi-head attention analysis of multi-scale feature representation is helpful to capture the relationship between different feature channels and enhance the model's ability to understand complex tension patterns. The gated deep separable convolution operation is adopted to reduce the number of parameters while maintaining the model's expressiveness, thereby improving the computational efficiency and generalization ability of the model. An intelligent impedance matching model is constructed, which can dynamically adapt to changes in the wire-environment interface, optimize the tension adjustment strategy, and improve the system's adaptability to environmental changes. By executing the target tension adjustment strategy and performing feedback updates, the dynamic optimization of the tension distribution model is realized, and the introduction of fractional calculus operations and fractional transfer functions more accurately describes the complex impedance characteristics and improves the accuracy of impedance matching. The initial tension adjustment strategy is optimized by multiple objectives, taking into account multiple performance indicators of tension control, and achieving a more comprehensive and balanced adjustment effect. The adaptive neural fuzzy inference system is used for parameter adjustment, which combines the learning ability of the neural network and the interpretability of fuzzy logic to improve the interpretability and reliability of the model. The second tension distribution model is generated by conditional variational inference, which can better capture the uncertainty of tension distribution. The present invention realizes more intelligent, accurate and safer adaptive adjustment of the tension of medical wires.

[0043] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0044] Apply multi-frequency and multi-amplitude AC excitation signals to the medical wires, perform parallel sampling through tension sensors and electrochemical impedance spectrometers, and obtain original tension time series and impedance spectrum response data;

[0045] The original tension time series is filtered to obtain historical tension data, and the impedance spectrum response data is aligned in time and frequency to obtain the wire-environment interface impedance spectrum data.

[0046] Specifically, a multi-frequency and multi-amplitude AC excitation signal is applied to the medical wire. The excitation signal is usually in the form of an AC signal and includes a combination of multiple frequencies and amplitudes to ensure that the response characteristics of the wire under different working conditions are covered. The frequency range of the excitation signal can be from tens of hertz to several thousand hertz, and the amplitude can be dynamically adjusted according to the physical properties of the wire. After the excitation signal is applied, the tension sensor and the electrochemical impedance spectrometer start working synchronously. The tension sensor is used to monitor the tension changes of the wire under different excitation conditions and record the tension change curve over time, that is, the original tension time series. At the same time, the electrochemical impedance spectrometer records the impedance changes between the wire and the environmental interface. The impedance spectrometer measures the current and voltage responses of the wire at different frequencies, calculates the amplitude and phase of the impedance, and forms the impedance spectrum response data. The original tension time series is filtered to remove high-frequency noise and low-frequency drift in the data and retain the true tension change characteristics. Filtering can be achieved by applying a low-pass filter (such as a Butterworth filter), and the design of the filter needs to be adjusted according to the frequency characteristics of the tension signal. Assume that the original tension signal is , the filtered historical tension data can be expressed as:

[0047] ;

[0048] in, is the filtered tension data, is the impulse response function of the filter, is a time variable. At the same time, the impedance spectrum response data is time-frequency aligned. The impedance spectrum response data is usually presented in a mixed form of time domain and frequency domain, that is, the impedance values ​​at different frequencies are recorded at different time points. Time-frequency alignment is used to ensure that the impedance data can accurately correspond to the frequency of the excitation signal at all time points. The impedance data is rearranged on the time-frequency plane so that different frequency data at the same time point can form a unified sequence. This process can be achieved through short-time Fourier transform (STFT), and the STFT formula is:

[0049] ;

[0050] in, is the time-frequency representation of the impedance spectroscopy data, is the original impedance signal, is a window function used to limit the integration range. is a frequency variable. This conversion converts the time series into a time-frequency representation, so that the different frequency components at each time point can correspond one to one. After time-frequency alignment, the obtained wire-environment interface impedance spectrum data includes impedance values ​​at different frequencies and can accurately reflect the dynamic characteristics of these impedance changes over time.

[0051] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0052] Perform wavelet multi-resolution decomposition on the historical tension data to obtain subsequence sets of multiple scales, and perform position encoding on the subsequence sets of multiple scales to obtain encoded subsequences;

[0053] Input the encoded subsequence into the multi-head self-attention module in the multi-scale Transformer neural network for feature weighting operation to obtain the attention-weighted feature representation;

[0054] The attention-weighted feature representation is input into the feed-forward neural network in the multi-scale Transformer neural network, and the output of the single-scale Transformer block is obtained through two layers of linear transformation and ReLU activation function;

[0055] The output of the single-scale Transformer block is input into the cross-scale fusion module in the multi-scale Transformer neural network, and a unified representation of features at different scales is obtained through adaptive pooling and convolution operations;

[0056] Perform multi-scale attention mechanism analysis on the unified representation of features at different scales, and obtain scale-weighted feature representation by calculating the correlation matrix between scales;

[0057] The scale-weighted feature representation is input into the residual connection module in the multi-scale Transformer neural network and combined with the original input through addition operation to obtain the residual enhanced feature representation;

[0058] The residual enhanced feature representation is layer normalized, and the multi-scale feature representation is obtained by calculating the mean and variance and performing linear transformation.

[0059] Specifically, the historical tension data is subjected to wavelet multi-resolution decomposition. Wavelet multi-resolution decomposition is an effective tool for signal processing, which can decompose the signal into sub-signals of different scales, thereby capturing different characteristics of the signal in the time and frequency domains. Assume that the original historical tension data is a time series , through wavelet decomposition, Decompose into a set of subsequences with different resolutions, each subsequence corresponds to the characteristics of the signal in a certain frequency range. Position encoding is performed on the subsequence set to add position information to the characteristics of each time step, because the Transformer structure itself does not have the ability to perceive sequential information. By position encoding the elements in each subsequence, the timing information is embedded into the model, allowing the model to understand the relative position of each element in the sequence. Position encoding is usually implemented through sine and cosine functions. For a given time step and feature dimensions , the position encoding formula can be expressed as:

[0060] ;

[0061] ;

[0062] in, and Represent the encoding values ​​of even and odd positions respectively, is the index of the feature dimension, is the total dimension of the feature. This encoding method ensures that the relative position of the feature at different time steps can be recognized by the model to obtain the encoded subsequence. The encoded subsequence is input into the multi-head self-attention module in the multi-scale Transformer neural network for feature weighting operation. The multi-head self-attention mechanism achieves weighting of important features by calculating the attention weights between different features. For each subsequence, an attention weight matrix is ​​obtained by performing dot product calculations with the features of other subsequences to reflect the relationship between different features. The attention weight matrix is ​​applied to the original features to generate an attention-weighted feature representation. Due to the use of a multi-head mechanism, this process is carried out in parallel in multiple subspaces, thereby capturing different aspects of the features. The attention-weighted feature representation is input into the feedforward neural network in the multi-scale Transformer neural network. The feedforward neural network consists of two layers of linear transformation and ReLU activation function. The features are mapped to a higher dimensional space through linear transformation, and then nonlinearity is introduced through the ReLU activation function to enhance the expressive power of the model. Assume that the attention-weighted feature representation is , then the output of the feedforward neural network can be expressed as:

[0063] ;

[0064] in, and is the weight matrix of the linear transformation, and is the bias term. Through this process, the output of the single-scale Transformer block is generated, which represents the processing result of the features at a single scale. The output of the single-scale Transformer block is input into the cross-scale fusion module in the multi-scale Transformer neural network, and the features from different scales are uniformly represented through adaptive pooling and convolution operations. The role of adaptive pooling is to map features of different scales to the same dimension so that they can be effectively fused in subsequent processing; while the convolution operation is used to extract the common features between features of different scales to ensure that the fused representation can comprehensively reflect the information of each scale. Through these operations, a unified representation of features of different scales is obtained. The unified representation is analyzed by the multi-scale attention mechanism. The multi-scale attention mechanism identifies the intrinsic connection between each scale by calculating the correlation matrix between different scales, and obtains the scale-weighted feature representation. The correlation matrix can be obtained by calculating the dot product between the features of each scale, and each element in the matrix reflects the correlation between the features of two scales. The matrix is ​​applied to the unified representation to generate a scale-weighted feature representation. The scale-weighted feature representation is input into the residual connection module in the multi-scale Transformer neural network. Residual connections preserve important information in the original features by adding scale-weighted features to the original input, thus avoiding information loss in deep networks. The residual enhanced feature representation is combined with the original input through the addition operation. The residual enhanced feature representation is subjected to layer normalization. Layer normalization calculates the mean and variance of the feature and transforms it linearly to ensure the stability and consistency of the feature in subsequent processing. Assume that the residual enhanced feature representation is , then the normalized output of the layer can be expressed as:

[0065] ;

[0066] in, and are the mean and standard deviation of the features, and is a learnable parameter used to adjust the normalization result. After layer normalization, the obtained multi-scale feature representation not only has a stable distribution, but also can better adapt to the subsequent deep network processing.

[0067] In a specific embodiment, the step of inputting the output of the single-scale Transformer block into the cross-scale fusion module in the multi-scale Transformer neural network, and obtaining a unified representation of features of different scales through adaptive pooling and convolution operations may specifically include the following steps:

[0068] Perform pyramid pooling on the output feature map of the single-scale Transformer block, and obtain feature pyramids of different spatial resolutions through multi-level downsampling;

[0069] Perform a 1x1 convolution operation on each scale feature in the feature pyramid, obtain a multi-scale feature set of uniform dimension by adjusting the number of channels, and input the multi-scale feature set of uniform dimension into the self-attention mechanism module to calculate the cross-scale attention weights to obtain the inter-scale correlation features;

[0070] Perform bilinear interpolation on inter-scale correlation features and adjust different scale features to the same spatial resolution through upsampling;

[0071] The adjusted features are input into the channel attention module for squeezing-excitation operation to obtain channel-weighted feature representation, and the channel-weighted feature representation is cross-scale fused to obtain cross-scale fused features;

[0072] The cross-scale fusion features are input into the residual dense block, and enhanced multi-scale features are obtained through multi-layer dense connections and short-circuit connections. The enhanced multi-scale features are recalibrated to obtain a unified representation of features of different scales.

[0073] Specifically, pyramid pooling is a technique for capturing information of different scales in an image or feature map. By downsampling the input feature map at multiple scales, feature pyramids of different spatial resolutions are obtained. Each scale of the feature map corresponds to a different field of view. Pyramid pooling is performed on the output feature map of a single-scale Transformer block. Assume that the dimension of the input feature map is ,in is the number of channels, and are the height and width of the feature map respectively. Through pyramid pooling, the feature map is downsampled at multiple levels, for example The pooling operation is performed on the feature map with a window size of , and the feature pyramids at different resolutions are obtained. A 1x1 convolution operation is performed on each scale feature in the feature pyramid. 1x1 convolution is an important convolution operation, which is used to adjust the number of channels of the feature map without changing its spatial resolution. By performing a 1x1 convolution operation on the feature map of each scale, the number of channels of these feature maps is unified to a fixed dimension. Assume that the number of channels of the input feature is , which can be adjusted to , this process can be expressed as:

[0074] ;

[0075] in, is the output feature map, is the input feature map, is the weight matrix of the 1x1 convolution, is a bias term. Through the convolution operation, a multi-scale feature set of unified dimension is obtained. The multi-scale feature set of unified dimension is input into the self-attention mechanism module to calculate the cross-scale attention weight. The self-attention mechanism captures the relationship between features of different scales by calculating the attention weights between features. For each feature scale, the dot product between it and other feature scales is calculated to obtain the attention weight matrix, which reflects the mutual correlation between different scales. The attention weight is applied to the feature set to generate inter-scale correlation features. The attention mechanism can be described by the following formula:

[0076] ;

[0077] in, Representation scale and scale The attention weights between and The scale and Through this process, a cross-scale attention-weighted feature representation is generated to capture the important relationship between different scales. Bilinear interpolation is performed on the inter-scale correlation features, and the features of different scales are adjusted to the same spatial resolution through upsampling. Bilinear interpolation is a commonly used upsampling method that estimates new pixel values ​​by linearly interpolating neighboring pixels. Assume that the size of the input low-resolution feature map is , the target resolution is , through bilinear interpolation operation, the low-resolution feature map is adjusted to the target resolution. The adjusted features are input into the channel attention module for squeeze-excitation operation. The squeeze-excitation operation is an attention mechanism for the channel dimension, which emphasizes important features and suppresses irrelevant features by weighting each channel. The squeeze operation compresses the features of each channel into a scalar through global average pooling, and then performs a nonlinear transformation through two fully connected layers and ReLU activation function, and finally generates channel weights through the sigmoid function. The excitation operation applies these weights to the original channels to obtain channel-weighted feature representations. The squeeze-excitation operation can be described by the following formula:

[0078] ;

[0079] in, It is a channel The weight of and are the weight matrices of the fully connected layer, It is a channel The global average pooling value of is the sigmoid activation function. In this way, the channel attention module can dynamically adjust the contribution of each channel and improve the accuracy of feature representation. Perform cross-scale fusion operations on the channel-weighted feature representation. Features of different scales are organically combined to form a comprehensive multi-scale feature representation. By adding or concatenating the weighted features, it is ensured that the fused features can comprehensively consider the information of each scale and occupy a more advantageous position in the feature space. The cross-scale fused features are input into the residual dense block for processing. The residual dense block combines the advantages of residual connections and dense connections. Through multiple layers of dense connections and short-circuit connections, features can be efficiently transferred between multiple layers. Assume that the input feature is ,pass After convolutional layers, the output features are obtained , then the residual connection can be expressed as:

[0080] ;

[0081] The dense connection connects the outputs of all previous layers to the input of the current layer. The residual dense block can effectively fuse features at different levels to obtain enhanced multi-scale features. The enhanced multi-scale features are recalibrated. By normalizing and adjusting the features again, the contributions of features of different scales are reasonably balanced in the final representation.

[0082] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0083] Perform channel grouping operation on the multi-scale feature representation, obtain multiple feature subsets by uniform division, perform cross-channel linear projection on multiple feature subsets, and obtain query, key and value matrices through three independent fully connected layers;

[0084] Perform scaled dot product attention calculations on the query, key, and value matrices to obtain multi-head attention outputs, and perform channel reorganization and splicing operations on the multi-head attention outputs to obtain cross-channel attention features through linear transformation;

[0085] The cross-channel attention features are input into the gated unit for sigmoid activation function operation to obtain the channel importance weight, and the original multi-scale feature representation is processed by channel-by-channel convolution and point-by-point convolution to obtain the preliminary convolution features;

[0086] The channel importance weight is element-wise multiplied with the preliminary convolutional features to obtain weighted convolutional features, and the weighted convolutional features are subjected to residual connection and layer normalization operations to obtain the target tension-sensitive features.

[0087] Specifically, the multi-scale feature representation is grouped into channels, and the channel dimension of the feature representation is divided into several subsets, each of which contains a part of the features. The feature subsets are linearly projected across channels. The features of each subset are mapped to different feature spaces to enhance the network's expressive power. Through three independent fully connected layers, each feature subset is mapped to query, key, and value matrices. Assume that the feature subset is , then the query, key, and value matrices after linear projection can be expressed as:

[0088] ;

[0089] in, , , is the weight matrix of the fully connected layer, , , is the corresponding bias term. Through these operations, the query, key, and value matrices are represented as , and . Scaled dot product attention is calculated on the query, key, and value matrices. Scaled dot product attention is a commonly used attention mechanism that determines the attention weights by calculating the dot product between the query and the key, and then applies these weights to the value matrix to generate the attention output. The calculation formula for the attention weight is:

[0090] ;

[0091] in, is the dimension of the key matrix, is a scaling factor used to prevent the dot product value from being too large. The dot product result is converted into a probability distribution through the softmax function, which is then applied to the value matrix , and obtain the multi-head attention output. The multi-head attention output can calculate attention in parallel in multiple subspaces, allowing the model to capture the complex relationship between different features. Channel reorganization and splicing operations are performed on the multi-head attention output. The purpose of channel reorganization is to rearrange the feature channels to increase the interaction between features, while the splicing operation merges the outputs of multiple attention heads together to form a new feature representation. The spliced ​​feature representation is converted into a cross-channel attention feature through a linear transformation, and the outputs of different attention heads are integrated to generate a unified feature representation. The cross-channel attention feature is input into the gating unit for processing. The gating unit generates channel importance weights through the sigmoid activation function operation. The sigmoid function maps the features to a range between 0 and 1, thereby determining the importance weight of each channel. These weights are used to emphasize important features and suppress irrelevant features. Assume that the cross-channel attention feature is , then the channel importance weight It can be expressed as:

[0092] ;

[0093] in, is the weight matrix, is the bias term, is the sigmoid activation function. Through these operations, the model can dynamically adjust the contribution of each channel and generate a weighted feature representation. The original multi-scale feature representation is processed by channel-by-channel convolution and point-by-point convolution. Channel-by-channel convolution refers to the independent application of convolution operations in each channel to capture local features within the channel, while point-by-point convolution is to integrate information between channels by applying 1x1 convolution at each pixel. Assume that the original multi-scale feature representation is , the results of channel-by-channel convolution and point-by-point convolution can be expressed as:

[0094] ;

[0095] in, and Represent channel-by-channel convolution and point-by-point convolution operations respectively. Through the convolution operation, the preliminary convolution features are obtained, which contain local and global information represented by multi-scale features. The channel importance weight is multiplied element-by-element with the preliminary convolution features to generate weighted convolution features. Element-by-element multiplication adjusts the amplitude of the convolution features through the channel weights, highlighting important features and weakening irrelevant features. Assume that the channel importance weight is , the initial convolution feature is , then the weighted convolution feature can be expressed as:

[0096] ;

[0097] Among them, © represents element-by-element multiplication. The weighted convolution features are subjected to residual connection and layer normalization. Residual connection preserves the key information in the original features by adding the weighted convolution features to the original features, and helps to avoid the gradient vanishing problem in deep networks. Layer normalization calculates the mean and variance of the features and standardizes them to ensure that the distribution of the features is more stable and consistent in subsequent processing. Assume that the weighted convolution feature is , the result of residual connection and layer normalization can be expressed as:

[0098] ;

[0099] in, and are the mean and standard deviation of the features, and are learnable scaling and offset parameters. The target tension-sensitive features are obtained.

[0100] In a specific embodiment, the execution step performs a channel grouping operation on the multi-scale feature representation, obtains multiple feature subsets by uniform division, and performs cross-channel linear projection on the multiple feature subsets, and obtains the query, key and value matrices respectively through three independent fully connected layers. The process can specifically include the following steps:

[0101] Perform channel dimension analysis on the multi-scale feature representation to obtain the channel importance ranking, and calculate the feature channel grouping threshold based on the channel importance ranking to obtain the grouping parameter;

[0102] Rearrange and group channels of multi-scale feature representation based on grouping parameters to obtain initial subsets, and perform adaptive normalization operations on the initial subsets to obtain multiple feature subsets;

[0103] Input multiple feature subsets into the fully connected layer with shared weights for linear transformation and nonlinear activation function operation to obtain intermediate feature representation, and perform channel shuffling operation on the intermediate feature representation to obtain mixed feature representation;

[0104] The mixed feature representation is input into three independent fully connected layers for linear transformation with different weight matrices to obtain query, key and value matrices.

[0105] Specifically, the multi-scale feature representation is analyzed in channel dimension to evaluate the importance of each channel in the entire feature representation. The importance of the channel is measured by calculating the average activation value, variance of each channel or by the gradient during model training. Assume that the multi-scale feature representation is ,in The dimension is , is the number of channels, and are the height and width of the feature map respectively. By analyzing the activation strength of each channel, the importance ranking is obtained. According to the importance ranking, the feature channel grouping threshold is calculated to determine how to divide the channels into different groups for more fine-grained processing. Assume that the channels are divided into The sorted channels are divided into groups according to their cumulative importance, and the cumulative values ​​of the number of channels and importance weights in each group are as close as possible. The calculation of the grouping parameters can be expressed as:

[0106] ;

[0107] in, Indicates The cumulative importance of the groups, Indicates channel The importance rating of Indicates that the The number of channels in each group. Through this formula, the number of channels and boundaries of each group are obtained to form the final grouping parameters. Based on the grouping parameters, the multi-scale feature representation is rearranged and grouped. The channels are rearranged according to importance so that important channels are concentrated in the front. The grouping operation divides the channels into several initial subsets according to the calculated grouping parameters. The initial subsets are adaptively normalized, and the mean and variance within the subsets are calculated and standardized so that the features of each subset can better adapt to different scales and distributions in subsequent processing. The normalized feature subsets are input into the fully connected layer with shared weights for linear transformation and nonlinear activation function operations. The role of the fully connected layer with shared weights is to ensure that all feature subsets are processed under the same conditions, so as to ensure the consistency of the feature processing results. The role of linear transformation is to map features to a higher dimensional space, while the nonlinear activation function introduces nonlinear characteristics to enhance the complexity and expression ability of feature representation. Assume that the normalized feature subset is , the intermediate feature after linear transformation and nonlinear activation function operation is expressed as:

[0108] ;

[0109] in, is the weight matrix of the fully connected layer, is the bias term, and ReLU is a commonly used nonlinear activation function. Through this operation, an intermediate feature representation is generated . Perform channel shuffling on the intermediate feature representation. Channel shuffling is a technique for rearranging feature channels, aiming to break the original channel order and increase the interaction between features. The channel shuffling operation can rearrange feature channels randomly or according to a specific rule so that the information between different channels can be better integrated. The shuffled feature representation is a mixed feature representation. The mixed feature representation is input into three independent fully connected layers for linear transformation to obtain query, key, and value matrices. Each feature representation is projected through different weight matrices to generate query (Q), key (K), and value (V) matrices. The specific calculation formula is:

[0110] ;

[0111] in, , and are the weight matrices of the three fully connected layers, , and is the corresponding bias term. Through the projection operation, the query, key and value matrices generated , and It is used for subsequent scaled dot product attention calculations and is ultimately used to capture important relationships in multi-scale feature representations.

[0112] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0113] Perform principal component dimensionality reduction processing on the target tension-sensitive features to obtain the tension features after dimensionality reduction, and perform Gaussian mixture clustering on the tension features after dimensionality reduction to obtain the probability density function of tension distribution;

[0114] A variational autoencoder is constructed according to the probability density function of the tension distribution, and the parameters are optimized by maximizing the lower bound of the evidence to obtain the first tension distribution model.

[0115] The time-frequency characteristics and instantaneous frequency characteristics of the wire-environment interface impedance spectrum data are extracted, and a complex-valued neural network is constructed based on the time-frequency characteristics and instantaneous frequency characteristics. The impedance spectrum fitting model is obtained by training the neural network through the complex domain back propagation algorithm.

[0116] Perform fractional calculus operation on the impedance spectrum fitting model to obtain fractional impedance parameters, and construct a fractional transfer function based on the fractional impedance parameters;

[0117] The fractional-order transfer function is input into the adaptive neuro-fuzzy inference system, and the parameters are adjusted through the hybrid learning algorithm to obtain the intelligent impedance matching model.

[0118] Specifically, principal component analysis is performed on the target tension-sensitive features to reduce the dimension of the data by projecting the data to the direction with the largest variance while retaining as much information as possible. Assume that the target tension-sensitive features are the matrix , whose dimensions are ,in is the number of samples, is the feature dimension. By calculating The covariance matrix of the data is calculated and its eigenvalues ​​and eigenvectors are solved to find the main direction of the data, i.e., the principal component. The data is projected onto the first few principal components to obtain the tension feature after dimensionality reduction. , whose dimensions are ,in is the dimension after dimensionality reduction (usually ). Gaussian mixture clustering is performed on the tension features after dimensionality reduction. Gaussian mixture model is a commonly used clustering method. It assumes that the data is composed of multiple Gaussian distributions and estimates the parameters of each Gaussian distribution through the expectation maximization algorithm. Assume Represents the tension feature after dimensionality reduction, then the Gaussian mixture model can be expressed as:

[0119] ;

[0120] in, It is The mixing coefficient of the Gaussian distribution satisfies and They are The mean and covariance matrix of a Gaussian distribution, is the number of Gaussian distributions. By training the Gaussian mixture model, we get the probability density function of the tension distribution , describing the distribution of tension features under different states. Based on the probability density function of tension distribution, a variational autoencoder is constructed. A variational autoencoder is a generative model that generates new data samples by probabilistically modeling the latent variables of the data. When constructing a variational autoencoder, the probability density function of the tension feature is used as the prior distribution. As input, the encoder maps it to the latent space , the decoder uses The original features are reconstructed from . In order to optimize the parameters of the variational autoencoder, training is performed by maximizing the lower bound of evidence. The formula for maximizing the lower bound of evidence is:

[0121] ;

[0122] in, is the approximate posterior distribution, is the reconstructed distribution of the decoder, and KL represents the Kullback-Leibler divergence. By maximizing the lower bound of the evidence, the parameters of the variational autoencoder can be effectively optimized to obtain the first tension distribution model, which can generate samples that conform to the tension distribution characteristics for prediction and adjustment. At the same time, the time-frequency characteristics and instantaneous frequency characteristics of the wire-environment interface impedance spectrum data are extracted. Impedance spectrum data is usually a complex signal that varies with time. The time-frequency characteristics can be extracted by short-time Fourier transform, while the instantaneous frequency characteristics can be calculated by Hilbert transform. Assume that the impedance spectrum signal is , its short-time Fourier transform is expressed as:

[0123] ;

[0124] in, is the time-frequency representation, is a window function, is the frequency variable. The instantaneous frequency can be obtained by solving the derivative of the phase with time. After extracting these features, they are used to construct a complex-valued neural network. A complex-valued neural network is a neural network model that can process complex data and is trained using a complex-domain back-propagation algorithm. The complex-valued neural network not only updates weights in the real domain, but also adjusts phase and amplitude in the complex domain. Assume that the input complex signal is , the weight is , then the output of the complex convolution can be expressed as:

[0125] ;

[0126] in, and are the real and imaginary parts of the complex signal, respectively. and are the real and imaginary parts of the weight. Through the complex domain back propagation algorithm, the complex-valued neural network can learn the complex characteristics of the impedance spectrum and obtain the impedance spectrum fitting model. Perform fractional calculus on the impedance spectrum fitting model. Fractional calculus is an extended form of calculus that can better describe nonlinear systems and systems with memory effects. Assume that the output of the impedance spectrum model is , then the fractional calculus operation can be expressed as:

[0127] ;

[0128] in, is the fractional derivative operator, is the fractional order, is the gamma function. Through this operation, the fractional impedance parameters are obtained, which reflect the dynamic characteristics of the impedance spectrum. Based on the fractional impedance parameters, the fractional transfer function is constructed. The fractional transfer function is used to describe the relationship between the input and output of the system. Assume that the fractional transfer function is , it can be expressed as:

[0129] ;

[0130] in, and are the numerator and denominator polynomials, respectively. and is the coefficient, and is a fractional order. This function is used to describe the impedance characteristics between the wire and the environment interface. The fractional transfer function is input into the adaptive neural fuzzy inference system, and the parameters are adjusted through the hybrid learning algorithm to obtain an intelligent impedance matching model. The adaptive neural fuzzy inference system combines the advantages of fuzzy logic and neural networks and can handle nonlinear and uncertainty problems. By inputting the fractional transfer function, the adaptive neural fuzzy inference system can learn the optimal impedance matching strategy and achieve precise adjustment of the wire tension.

[0131] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0132] Acquire real-time tension data, and perform wavelet packet decomposition and empirical mode decomposition on the real-time tension data to obtain a first multi-scale tension characteristic component and a first intrinsic mode function, and reconstruct the real-time tension data according to the first multi-scale tension characteristic component and the first intrinsic mode function to obtain denoised tension data;

[0133] The denoised tension data is input into the first tension distribution model to calculate the posterior probability of tension distribution, and obtain tension distribution parameters;

[0134] Gaussian process regression analysis is performed on the tension distribution parameters to obtain a tension prediction function, and multiple candidate tension adjustment strategies are generated according to the tension prediction function;

[0135] According to the Bayesian information criterion, multiple candidate tension adjustment strategies are evaluated to obtain an initial tension adjustment strategy, and the initial tension adjustment strategy is input into the intelligent impedance matching model, and fuzzy rule reasoning is performed through the adaptive neural fuzzy inference system to obtain impedance matching parameters;

[0136] The initial tension adjustment strategy is optimized by multiple objectives according to the impedance matching parameters to obtain the target tension adjustment strategy.

[0137] Specifically, real-time tension data is obtained from the real-time monitoring system. Wavelet packet decomposition and empirical mode decomposition are performed on the real-time tension data. Wavelet packet decomposition is a multi-resolution analysis technology that can decompose the signal into sub-signals in different frequency bands and capture the multi-scale features in the signal. Assume that the real-time tension data is , it can be represented as a series of components of different scales through wavelet packet decomposition:

[0138] ;

[0139] in, is the decomposition coefficient, is the wavelet basis function, and Represent the scale and translation parameters respectively. Through the decomposition process, the first multi-scale tension characteristic component is obtained, which represents the characteristics of the tension data in different frequency bands. At the same time, the real-time tension data is subjected to empirical mode decomposition. Empirical mode decomposition is an adaptive signal decomposition method that decomposes the signal into a set of intrinsic mode functions (IMFs) through repeated iterations. Each IMF represents an oscillation mode of the signal. Assume that the decomposed IMF is , then the original signal can be expressed as the sum of these IMFs:

[0140] ;

[0141] in, is the residue after decomposition. Through empirical mode decomposition, the first intrinsic mode function is obtained to capture the main oscillation characteristics in the tension signal. The real-time tension data is reconstructed according to the first multi-scale tension characteristic component and the first intrinsic mode function to eliminate noise and retain the main characteristics of the signal to obtain denoised tension data. This process enhances the clarity and reliability of the signal by combining all valid components and mode functions and ignoring noise components. The denoised tension data is input into the first tension distribution model to calculate the posterior probability of tension distribution. Assume that the denoised tension data is , then its posterior probability It can be calculated by Bayes formula:

[0142] ;

[0143] in, is the parameter of the tension distribution, is the likelihood function, is the prior distribution, is the marginal likelihood. Through calculation, the parameters of tension distribution are obtained to describe the probability distribution of tension data in the current state. Gaussian process regression analysis is performed on the tension distribution parameters. Gaussian process regression is a non-parametric Bayesian regression method used to construct a probability model between input data and output data. Assume that the tension distribution parameters are , the Gaussian process regression model can be expressed as:

[0144] ;

[0145] in, is the mean function, is the covariance function. Through Gaussian process regression, we get the tension prediction function , which is used to predict the future tension change trend. According to the tension prediction function, multiple candidate tension adjustment strategies are generated. Each strategy corresponds to a possible tension adjustment scheme, which is used to adjust the tension of the wire under different tension conditions. Since the effects of different strategies may be different, they need to be evaluated. When evaluating these candidate tension adjustment strategies, the Bayesian Information Criterion (BIC) is used for selection. BIC is a model selection criterion used to select the optimal model from the candidate models. The formula of BIC is:

[0146] ;

[0147] in, is the maximum likelihood estimate of the model, is the number of model parameters, is the number of samples. By calculating the BIC value of each strategy, the strategy with the smallest BIC value is selected as the initial tension adjustment strategy. The initial tension adjustment strategy is input into the intelligent impedance matching model, and fuzzy rule reasoning is performed through the adaptive neuro-fuzzy inference system. The adaptive neuro-fuzzy inference system combines the advantages of neural networks and fuzzy logic and can handle complex nonlinear relationships and uncertainties. Through fuzzy rule reasoning, the adaptive neuro-fuzzy inference system generates impedance matching parameters, which are used to adjust the details of the initial tension adjustment strategy to better adapt to the current impedance characteristics. The initial tension adjustment strategy is multi-objective optimized according to the impedance matching parameters to obtain the target tension adjustment strategy. Multi-objective optimization aims to meet multiple adjustment objectives at the same time, such as tension stability, response speed, and energy efficiency. Through optimization algorithms (such as genetic algorithms or particle swarm optimization), an optimal solution is found in the multi-dimensional target space to ensure that the target tension adjustment strategy can achieve the best performance in a complex environment.

[0148] In a specific embodiment, the process of executing step S6 may specifically include the following steps:

[0149] Discretize the target tension adjustment strategy to obtain a tension adjustment sequence, and perform tension adjustment operations and real-time feedback on the medical wire according to the tension adjustment sequence to obtain tension response data;

[0150] Performing wavelet packet transform and empirical mode decomposition on the tension response data to obtain a second multi-scale tension characteristic component and a second intrinsic mode function;

[0151] Based on the second multi-scale tension characteristic component and the second intrinsic mode function, time series modeling is performed to obtain a dynamic tension prediction model;

[0152] The dynamic tension prediction model and the first tension distribution model are integrated to obtain a fused tension model, and hyperparameter tuning is performed on the fused tension model to obtain an optimized tension model;

[0153] A latent variable representation of tension distribution is generated according to the optimized tension model, and conditional variational inference is performed on the latent variable representation to obtain a second tension distribution model.

[0154] Specifically, the self-standard tension adjustment strategy is discretized. The continuous tension adjustment strategy is converted into a series of operation sequences that can be executed at discrete time points. Assume that the target tension adjustment strategy is a continuous function ,in Indicates time by Discretize to get Tension value These discretized tension values ​​constitute the tension adjustment sequence ,in is the number of time points. Perform tension adjustment operations on the medical wire according to the tension adjustment sequence. Each discrete time point corresponds to a specific operation, and the control system adjusts the tension of the wire according to the tension value in the sequence. At the same time, the tension response of the wire at each time point is monitored and recorded in real time to generate tension response data, reflecting the actual response of the wire under different tension adjustments. Wavelet packet transform and empirical mode decomposition are performed on the obtained tension response data. Wavelet packet transform is a signal processing technology that can decompose the signal into sub-signals in multiple frequency bands and extract multi-scale features in the signal. Assume that the tension response data is , through wavelet packet transform, it is represented as a superposition of a series of subcomponents:

[0155] ;

[0156] in, is the wavelet packet transform coefficient, is the corresponding wavelet packet basis function, and Represent the index of scale and position respectively. Through this process, the second multi-scale tension characteristic component is obtained to describe the characteristics of the tension response data at different frequencies and scales. At the same time, the tension response data is processed using empirical mode decomposition. Empirical mode decomposition is an adaptive signal decomposition method that decomposes the signal into several intrinsic mode functions (IMFs), each of which represents an oscillation mode in the signal. Assume It is The tension response data can be expressed as the superposition of all IMFs:

[0157] ;

[0158] in, is the residual component, is the number of IMFs. Through empirical mode decomposition, the second intrinsic mode function is extracted to reflect the intrinsic oscillation characteristics of the tension response data. Time series modeling is performed based on the second multi-scale tension characteristic component and the second intrinsic mode function to predict future tension changes and build a dynamic tension prediction model. Assumptions Indicates time The tension characteristics at this time, the dynamic tension prediction model can be expressed as:

[0159] ;

[0160] in, is a nonlinear function describing the relationship between current and past tension characteristics, is the lag order of the model. By training the dynamic tension prediction model, the future tension changes are predicted. The dynamic tension prediction model is integrated with the first tension distribution model. The purpose of model integration is to integrate the prediction results of multiple models to improve the accuracy and robustness of the prediction. Assume that the output of the dynamic tension prediction model is , the output of the first tension distribution model is , the fused tension model can be expressed as:

[0161] ;

[0162] in, is the fusion coefficient, which is used to balance the contribution of the two models. Through fusion, the timing advantage of the dynamic tension prediction model and the probability advantage of the first tension distribution model are combined to obtain a more stable and accurate tension prediction result. The fused tension model is hyperparameter tuned to optimize the performance of the model so that it can perform better under different environments and data conditions. Common hyperparameter tuning methods include grid search, random search, and Bayesian optimization. Through tuning, the optimal combination of model parameters is found to obtain the optimized tension model. According to the optimized tension model, a latent variable representation of the tension distribution is generated. Latent variable representation is a technique that compresses high-dimensional data into a low-dimensional space in order to capture the main change patterns of the data. Assume that the latent variable representation of the tension data is , then we can use conditional variational inference to infer it, which approximates the posterior distribution by maximizing the variational lower bound. Assume that the approximate posterior distribution of the hidden variable is , the goal of conditional variational inference is to maximize the following formula:

[0163] ;

[0164] in, is the likelihood function of the observed data, is the prior distribution of latent variables, and KL represents the Kullback-Leibler divergence. Through conditional variational inference, the second tension distribution model is obtained, which not only considers the current tension state, but also combines the potential information of latent variables to achieve more accurate tension prediction and control.

[0165] The above describes the method for self-adapting tension adjustment of medical wires based on artificial intelligence in an embodiment of the present invention. The following describes the device for self-adapting tension adjustment of medical wires based on artificial intelligence in an embodiment of the present invention. Figure 2 In one embodiment of the present invention, an artificial intelligence-based medical wire tension adaptive adjustment device includes:

[0166] The acquisition module is used to collect historical tension data of medical wires and wire-environment interface impedance spectrum data;

[0167] An extraction module is used to input historical tension data into a multi-scale Transformer neural network for multi-scale feature extraction to obtain a multi-scale feature representation;

[0168] The analysis module is used to perform cross-channel multi-head attention analysis and gated depth-separable convolution operations on multi-scale feature representations to obtain target tension-sensitive features;

[0169] A construction module is used to construct a first tension distribution model according to target tension sensitivity characteristics, and to construct an intelligent impedance matching model according to wire-environment interface impedance spectrum data;

[0170] An adjustment module is used to obtain real-time tension data and input the data into a first tension distribution model to adjust the tension, obtain an initial tension adjustment strategy, and optimize the strategy through an intelligent impedance matching model to obtain a target tension adjustment strategy;

[0171] The feedback module is used to execute the target tension adjustment strategy and perform feedback update on the first tension distribution model to obtain a second tension distribution model.

[0172] Through the collaboration of the above components, the multi-scale Transformer neural network is used to extract multi-scale features of historical tension data, which can capture the tension change patterns at different time scales and improve the comprehensiveness and robustness of feature representation. Cross-channel multi-head attention analysis of multi-scale feature representation helps to capture the relationship between different feature channels and enhance the model's ability to understand complex tension patterns. The gated deep separable convolution operation is used to reduce the number of parameters while maintaining the model's expressiveness, improving the computational efficiency and generalization of the model. An intelligent impedance matching model is constructed, which can dynamically adapt to changes in the wire-environment interface, optimize the tension adjustment strategy, and improve the system's adaptability to environmental changes. By executing the target tension adjustment strategy and performing feedback updates, the dynamic optimization of the tension distribution model is achieved. The introduction of fractional-order calculus operations and fractional-order transfer functions can more accurately describe complex impedance characteristics and improve the accuracy of impedance matching. The initial tension adjustment strategy is optimized with multiple objectives, taking into account multiple performance indicators of tension control, and achieving a more comprehensive and balanced adjustment effect. The adaptive neural fuzzy inference system is used for parameter adjustment, which combines the learning ability of neural networks and the interpretability of fuzzy logic, improving the interpretability and reliability of the model. The second tension distribution model is generated by conditional variational inference, which can better capture the uncertainty of tension distribution. The present invention realizes more intelligent, accurate and safe adaptive tension adjustment of medical wires.

[0173] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the artificial intelligence-based medical wire tension adaptive adjustment method in the above-mentioned embodiments.

[0174] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the artificial intelligence-based medical wire tension adaptive adjustment method.

[0175] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0176] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0177] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adaptively adjusting tension of medical wires based on artificial intelligence, characterized in that: The method comprises: Collect historical tension data of medical wires and wire-environment interface impedance spectrum data; Inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation; Performing cross-channel multi-head attention analysis and gated depth-separable convolution operations on the multi-scale feature representation to obtain target tension-sensitive features; A first tension distribution model is constructed according to the target tension-sensitive feature, and an intelligent impedance matching model is constructed according to the wire-environment interface impedance spectrum data; specifically comprising: performing principal component dimensionality reduction processing on the target tension-sensitive feature to obtain the tension feature after dimensionality reduction, and performing Gaussian mixture clustering on the tension feature after dimensionality reduction to obtain the probability density function of the tension distribution; constructing a variational autoencoder according to the probability density function of the tension distribution, and optimizing the parameters by maximizing the lower bound of evidence to obtain the first tension distribution model; extracting the time-frequency characteristics and instantaneous frequency characteristics of the wire-environment interface impedance spectrum data, and constructing a complex-valued neural network according to the time-frequency characteristics and the instantaneous frequency characteristics, and training through a complex domain back propagation algorithm to obtain an impedance spectrum fitting model; performing fractional-order calculus operations on the impedance spectrum fitting model to obtain fractional-order impedance parameters, and constructing a fractional-order transfer function according to the fractional-order impedance parameters; inputting the fractional-order transfer function into an adaptive neural-fuzzy inference system, and adjusting the parameters through a hybrid learning algorithm to obtain an intelligent impedance matching model; Acquire real-time tension data and input it into the first tension distribution model for tension adjustment to obtain an initial tension adjustment strategy, and optimize the strategy through the intelligent impedance matching model to obtain a target tension adjustment strategy; specifically including: acquire real-time tension data, and perform wavelet packet decomposition and empirical mode decomposition on the real-time tension data to obtain a first multi-scale tension characteristic component and a first intrinsic mode function, and reconstruct the real-time tension data according to the first multi-scale tension characteristic component and the first intrinsic mode function to obtain denoised tension data; input the denoised tension data into the first tension distribution model to calculate the posterior probability of tension distribution to obtain tension distribution parameters; perform Gaussian process regression analysis on the tension distribution parameters to obtain a tension prediction function, and generate multiple candidate tension adjustment strategies according to the tension prediction function; evaluate the multiple candidate tension adjustment strategies according to the Bayesian information criterion to obtain an initial tension adjustment strategy, and input the initial tension adjustment strategy into the intelligent impedance matching model, perform fuzzy rule reasoning through the adaptive neural fuzzy inference system to obtain impedance matching parameters; perform multi-objective optimization on the initial tension adjustment strategy according to the impedance matching parameters to obtain a target tension adjustment strategy; The target tension adjustment strategy is executed and the first tension distribution model is updated through feedback to obtain a second tension distribution model.

2. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 1, characterized in that: The collecting of historical tension data of the medical wire and wire-environment interface impedance spectrum data includes: Apply multi-frequency and multi-amplitude AC excitation signals to the medical wires, perform parallel sampling through tension sensors and electrochemical impedance spectrometers, and obtain original tension time series and impedance spectrum response data; The original tension time series is filtered to obtain historical tension data, and the impedance spectrum response data is time-frequency aligned to obtain wire-environment interface impedance spectrum data.

3. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 1, characterized in that: The step of inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation includes: Performing wavelet multi-resolution decomposition on the historical tension data to obtain subsequence sets of multiple scales, and performing position encoding on the subsequence sets of multiple scales to obtain encoded subsequences; Inputting the encoded subsequence into a multi-head self-attention module in a multi-scale Transformer neural network to perform feature weighting operation to obtain an attention-weighted feature representation; Inputting the attention-weighted feature representation into the feedforward neural network in the multi-scale Transformer neural network, and obtaining the output of a single-scale Transformer block through two layers of linear transformation and ReLU activation function; Inputting the output of the single-scale Transformer block into the cross-scale fusion module in the multi-scale Transformer neural network, and obtaining a unified representation of features of different scales through adaptive pooling and convolution operations; Performing a multi-scale attention mechanism analysis on the unified representation of the features of different scales, and obtaining a scale-weighted feature representation by calculating a correlation matrix between scales; Inputting the scale-weighted feature representation into a residual connection module in the multi-scale Transformer neural network, and combining it with the original input through an addition operation to obtain a residual enhanced feature representation; A layer normalization operation is performed on the residual enhanced feature representation, and a multi-scale feature representation is obtained by calculating the mean and variance and performing a linear transformation.

4. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 3, characterized in that: The step of inputting the output of the single-scale Transformer block into the cross-scale fusion module in the multi-scale Transformer neural network to obtain a unified representation of features of different scales through adaptive pooling and convolution operations includes: Perform pyramid pooling on the output feature map of the single-scale Transformer block, and obtain feature pyramids of different spatial resolutions through multi-level downsampling; Performing a 1x1 convolution operation on each scale feature in the feature pyramid, obtaining a multi-scale feature set of uniform dimension by adjusting the number of channels, and inputting the multi-scale feature set of uniform dimension into a self-attention mechanism module to calculate cross-scale attention weights, thereby obtaining inter-scale correlation features; Performing a bilinear interpolation operation on the inter-scale correlation features, and adjusting the features of different scales to the same spatial resolution by upsampling; Inputting the adjusted features into the channel attention module for a squeeze-excitation operation to obtain a channel-weighted feature representation, and performing cross-scale fusion on the channel-weighted feature representation to obtain a cross-scale fused feature; The cross-scale fusion features are input into the residual dense block, and enhanced multi-scale features are obtained through multi-layer dense connections and short-circuit connections. The enhanced multi-scale features are subjected to feature recalibration to obtain a unified representation of features of different scales.

5. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 1, characterized in that: The multi-scale feature representation is subjected to cross-channel multi-head attention analysis and gated depth-separable convolution operation to obtain target tension-sensitive features, including: Performing a channel grouping operation on the multi-scale feature representation to obtain a plurality of feature subsets by uniform division, performing cross-channel linear projection on the plurality of feature subsets, and obtaining query, key, and value matrices respectively through three independent fully connected layers; Performing scaled dot product attention calculation on the query, key and value matrices to obtain multi-head attention output, and performing channel reorganization and splicing operations on the multi-head attention output to obtain cross-channel attention features through linear transformation; The cross-channel attention feature is input into the gated unit for sigmoid activation function operation to obtain the channel importance weight, and the original multi-scale feature representation is subjected to channel-by-channel convolution and point-by-point convolution processing to obtain preliminary convolution features; The channel importance weight is element-wise multiplied by the preliminary convolution feature to obtain a weighted convolution feature, and the weighted convolution feature is subjected to residual connection and layer normalization operations to obtain a target tension-sensitive feature.

6. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 5, characterized in that: The multi-scale feature representation is subjected to channel grouping operation to obtain multiple feature subsets by uniform division, and the multiple feature subsets are subjected to cross-channel linear projection to obtain query, key and value matrices respectively through three independent fully connected layers, including: Performing channel dimension analysis on the multi-scale feature representation to obtain channel importance ranking, and calculating feature channel grouping thresholds according to the channel importance ranking to obtain grouping parameters; Rearranging and grouping the multi-scale feature representation channels based on the grouping parameters to obtain an initial subset, and performing an adaptive normalization operation on the initial subset to obtain a plurality of feature subsets; Input the multiple feature subsets into a fully connected layer with shared weights for linear transformation and nonlinear activation function operation to obtain an intermediate feature representation, and perform a channel shuffle operation on the intermediate feature representation to obtain a mixed feature representation; The mixed feature representation is input into three independent fully connected layers respectively for linear transformation of different weight matrices to obtain query, key and value matrices.

7. The method for adaptively adjusting the tension of medical wires based on artificial intelligence according to claim 1, characterized in that: The executing the target tension adjustment strategy and performing feedback update on the first tension distribution model to obtain a second tension distribution model includes: Discretizing the target tension adjustment strategy to obtain a tension adjustment sequence, and performing tension adjustment operations and real-time feedback on the medical wire according to the tension adjustment sequence to obtain tension response data; Performing wavelet packet transform and empirical mode decomposition on the tension response data to obtain a second multi-scale tension characteristic component and a second intrinsic mode function; Performing time series modeling based on the second multi-scale tension characteristic component and the second intrinsic mode function to obtain a dynamic tension prediction model; The dynamic tension prediction model is integrated with the first tension distribution model to obtain a fused tension model, and hyperparameter tuning is performed on the fused tension model to obtain an optimized tension model; A latent variable representation of tension distribution is generated according to the optimized tension model, and conditional variational inference is performed on the latent variable representation to obtain a second tension distribution model.

8. An artificial intelligence-based adaptive tension adjustment device for medical wires, characterized in that: For executing the method for adaptively adjusting the tension of a medical wire based on artificial intelligence according to any one of claims 1 to 7, the device comprises: The acquisition module is used to collect historical tension data of medical wires and impedance spectrum data of the wire-environment interface; An extraction module, used for inputting the historical tension data into a multi-scale Transformer neural network to extract multi-scale features and obtain a multi-scale feature representation; An analysis module, configured to perform cross-channel multi-head attention analysis and gated depth-separable convolution operations on the multi-scale feature representation to obtain target tension-sensitive features; A construction module, used to construct a first tension distribution model according to the target tension-sensitive characteristics, and to construct an intelligent impedance matching model according to the wire-environment interface impedance spectrum data; An adjustment module, used for acquiring real-time tension data and inputting the first tension distribution model to perform tension adjustment, obtaining an initial tension adjustment strategy, and optimizing the strategy through the intelligent impedance matching model to obtain a target tension adjustment strategy; A feedback module is used to execute the target tension adjustment strategy and perform feedback update on the first tension distribution model to obtain a second tension distribution model.

Images