A temperature control method of infrared polarized light therapeutic instrument based on real-time feedback

Abstract

The present application relates to the technical field of infrared polarized light therapeutic instrument temperature control, and discloses an infrared polarized light therapeutic instrument temperature control method based on real-time feedback. The method first collects multi-source thermal radiation signals at the irradiation head, performs time-frequency domain joint analysis through band-pass filtering, short-time Fourier transform and wavelet packet decomposition, extracts a dynamic feature vector representing the thermal response state of the tissue in combination with a two-dimensional thermal diagram matrix, then inputs it into a nonlinear mapping model initialized by a chaotic sequence, outputs an energy regulation instruction sequence containing pulse width modulation parameters and polarization state switching parameters, and finally synchronously regulates the infrared light source driving current and the polarization filter rotation angle according to the instructions, eliminates the dynamic coupling by using feedforward decoupling control, compares the dynamic feature vector by using the difference ratio, iteratively updates the model weight by using the gradient descent algorithm with band matrix estimation, and realizes adaptive tracking of the thermal response state of the tissue, thereby improving the treatment safety and control accuracy.

Description

Technical Field

[0001] This invention relates to the field of temperature control technology for infrared polarized light therapy devices, and specifically to a temperature control method for infrared polarized light therapy devices based on real-time feedback. Background Technology

[0002] Infrared polarized light therapy devices are widely used in clinical rehabilitation treatment. They achieve deep tissue thermal effects through infrared light source irradiation combined with polarized light modulation. The treatment effect is closely related to the distribution of photothermal energy output by the irradiation head.

[0003] However, the temperature control of existing therapeutic devices mostly adopts open-loop or simple closed-loop control methods, which only output energy based on preset fixed parameters and cannot be dynamically adjusted according to the real-time thermal response of the tissue.

[0004] In actual treatment, the absorption characteristics of tissues in different parts of the human body vary greatly, and the continuous changes in tissue temperature during treatment cause the thermal response state to change constantly. Traditional control methods are difficult to track these changes in real time, which can easily lead to local overheating or insufficient energy, affecting the treatment effect or even causing tissue damage.

[0005] In addition, existing control methods do not fully utilize the time-frequency features in multi-source thermal radiation signals, feature extraction is limited to a single dimension, and the control model lacks online adaptive update capability, resulting in a significant decrease in control accuracy after long-term operation.

[0006] Meanwhile, there is a dynamic coupling between the infrared light source drive and the polarization filter rotation. Existing decoupling strategies are not perfect, which further restricts the system's fast response performance and control accuracy. Summary of the Invention

[0007] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a temperature control method for an infrared polarized light therapy device based on real-time feedback, thereby solving the problems of low temperature control accuracy, slow response, and lack of adaptive adjustment capability in existing infrared polarized light therapy devices.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a temperature control method for an infrared polarized light therapy device based on real-time feedback, the method comprising: Step S1: Collect multi-source thermal radiation signals at the irradiation head of the infrared polarized light therapy instrument, perform time-frequency domain joint analysis on the multi-source thermal radiation signals, and extract dynamic feature vectors characterizing the thermal response state of the tissue surface. Step S2: Input the dynamic feature vector into the pre-trained nonlinear mapping model and output an energy adjustment command sequence that matches the current thermal response state. The energy adjustment command sequence includes pulse width modulation parameters and polarization state switching parameters. Step S3: Synchronously adjust the driving current waveform of the infrared light source and the rotation angle of the polarization filter according to the energy adjustment command sequence, so that the distribution of photothermal energy output by the irradiation head changes dynamically with the tissue thermal response state. Step S4: After performing energy regulation, continuously collect new multi-source thermal radiation signals, perform differential comparison between the newly extracted dynamic feature vector and the feature vector of the previous period, and iteratively update the input weights of the nonlinear mapping model based on the comparison results.

[0009] Preferably, in one possible implementation of the first aspect, the time-frequency domain joint analysis process in step S1 includes: Perform bandpass filtering on multi-source thermal radiation signals to filter out baseline drift components and high-frequency electromagnetic interference components; The filtered signal is truncated by applying a window function, and the time-domain signal is mapped to the time-frequency plane by short-time Fourier transform to generate a time-frequency energy matrix. Integrate the time-frequency energy matrix along the frequency axis to extract the temporal envelope features within a preset time window; Integrate the time-frequency energy matrix along the time axis to extract the frequency domain power spectral density features; The wavelet packet decomposition algorithm is used to decompose the multi-source thermal radiation signal to obtain the energy entropy of each frequency band component and the energy ratio of each node. The time-domain envelope features, frequency-domain power spectral density features, energy entropy, and node energy percentage are combined into a time-frequency joint feature set.

[0010] Preferably, in one possible implementation of the first aspect, the dynamic feature vector extraction process includes: Construct a two-dimensional heat map matrix corresponding to the geometry of the irradiation head end face, and map the time-frequency joint feature set to the corresponding grid nodes of the two-dimensional heat map matrix; Calculate the magnitude and direction of the grayscale gradient in the horizontal and vertical directions of the two-dimensional heatmap matrix; Solve for the invariant moments and central moments of the two-dimensional thermogram matrix to extract the geometric features characterizing the symmetry of the thermal field distribution; Singular value decomposition is performed on the two-dimensional heatmap matrix, and the eigenvalues ​​of a predetermined proportion are extracted to form the principal component eigenvectors. The gray-level gradient magnitude, direction, invariant moment, central moment, and principal component eigenvectors are sequentially concatenated to generate dynamic eigenvectors.

[0011] Preferably, in one possible implementation of the first aspect, the nonlinear mapping model in step S2 includes an input layer, a delay layer, a hidden layer, and an output layer; The input layer receives dynamic feature vectors and the energy adjustment instruction sequence from the previous time step. The delay layer performs a preset delay on the data received from the input layer and then outputs it to the hidden layer; The hidden layer contains several neurons, and the hyperbolic tangent function is used as the transfer function. The output layer maps the output of the hidden layer to the target control space, generating the initial instruction matrix.

[0012] Preferably, in one possible implementation of the first aspect, the connection weights and bias parameters of neurons in the hidden layer are initialized using a chaotic sequence; The initial weight matrix and bias vector within the sensitive region of the activation function are generated using the Tent chaotic mapping. Calculate the partial derivative of the output of each neuron in the hidden layer with respect to the input layer variables, and solve for the global Lipschitz constant of the nonlinear mapping model; The learning rate parameter of the hidden layer is dynamically adjusted based on the overall Lipschitz constant. Apply Dropout regularization to the output of the hidden layer to randomly deactivate a predetermined proportion of neurons.

[0013] Preferably, in one possible implementation of the first aspect, the energy regulation command sequence output process includes: The initial instruction matrix is ​​normalized to constrain the values ​​of the matrix elements to a preset unit range; The time series resolution is improved by upsampling the normalized initial instruction matrix using radial basis function interpolation. The upsampled command matrix is ​​decoupled using an orthogonal triangular decomposition algorithm to separate the light intensity control component and the polarization control component. The light intensity control component is activated by the Sigmoid function and then converted into the duty cycle value of the pulse width modulation parameter. The polarization control component is activated by an inverse trigonometric function and then converted into the target angle value of the polarization state switching parameter.

[0014] Preferably, in one possible implementation of the first aspect, the process of synchronously adjusting the driving current waveform of the infrared light source and the rotation angle of the polarizing filter in step S3 includes: Establish the photothermal conversion transfer function of the infrared light source and the electromechanical servo transfer function of the polarization filter; Based on the duty cycle and target angle values ​​contained in the energy regulation command sequence, the desired trajectories of the current loop and position loop are calculated respectively. A feedforward decoupling control strategy is adopted to eliminate the dynamic coupling term between the driving current waveform adjustment and the polarization filter rotation. Based on the desired trajectory and the current feedback of the actual current and angle values, calculate the current loop error and the position loop error; The current loop error and position loop error are input into the corresponding proportional-integral regulator, which outputs the drive voltage control signal and the motor drive signal.

[0015] Preferably, in one possible implementation of the first aspect, step S3 further includes: A space vector pulse width modulation algorithm is used to generate the switching transistor trigger signal for the infrared light source driving circuit; The voltage space vector perpendicular to the stationary coordinate system is synthesized based on the conduction state of the three-phase bridge arms, and the target average voltage is obtained through zero-vector interleaving modulation. The rotation angle of the polarization filter is decoded by a sine and cosine encoder to obtain the absolute position and angular velocity values. The absolute position and angular velocity values ​​are fed forward to the output of the PI controller to form a composite control law. The pulse width and phase of the motor drive signal are corrected based on the composite control law.

[0016] Preferably, in one possible implementation of the first aspect, the differential comparison process in step S4 includes: Calculate the Euclidean distance and cosine similarity between the newly extracted dynamic feature vector and the dynamic feature vector of the previous period; Construct the Bach coefficient and KL divergence values ​​to characterize the difference in probability distributions of two periodic eigenvectors; The values ​​of Euclidean distance, cosine similarity, Bach coefficient, and KL divergence are combined into a set of comparison indicators. By comparing the pre-defined weights assigned to each indicator in the indicator set and performing a weighted summation, a comprehensive difference scalar is generated. The integrated differential scalar is compared with a preset threshold. If it exceeds the preset threshold, it is determined that the tissue thermal response state has changed abruptly, and a weight update trigger signal is generated.

[0017] Preferably, in one possible implementation of the first aspect, the input weight iterative update process includes: In response to the weight update trigger signal, the prediction error gradient of the nonlinear mapping model in the current period is calculated; By introducing a momentum term and an adaptive learning rate factor, a gradient descent update rule with moment estimation is constructed. The update step size of the connection weights is adaptively adjusted by using the first-order moment estimation and second-order moment estimation of the prediction error gradient. Based on the preprocessed prediction error gradient, the backpropagation algorithm is used to expand and update the connection weights between the hidden layer and the input layer in the nonlinear mapping model along the time axis. Apply maximum norm constraints to the updated connection weights to prevent weight overflow.

[0018] The beneficial effects of this invention are as follows: By collecting multi-source thermal radiation signals and performing joint analysis in the time and frequency domains, this invention constructs a multi-dimensional joint time and frequency feature set that includes time domain envelope, frequency domain power spectral density, energy entropy, and node energy ratio. Combined with the gradient, invariant moment, central moment, and principal component features of the two-dimensional thermogram matrix, it achieves a comprehensive and accurate characterization of the tissue thermal response state.

[0019] The nonlinear mapping model employs a chaotic initialization and adaptive learning rate strategy, combined with Dropout regularization to improve the model's generalization ability and convergence speed. The output energy adjustment command sequence, after orthogonal decoupling and activation function transformation, can synchronously regulate the driving current waveform and the rotation angle of the polarization filter. The feedforward decoupling control effectively eliminates the dynamic coupling between the two channels.

[0020] The differential alignment mechanism introduces Euclidean distance, cosine similarity, Bach coefficient, and KL divergence to construct multidimensional alignment indices. Combined with a gradient descent algorithm based on moment estimation, it enables online iterative updates of model weights, allowing the system to adaptively track abrupt changes in tissue thermal response. This method also integrates multimodal safety monitoring and adaptive calibration mechanisms, significantly improving the safety, effectiveness, and personalization of treatment. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This application provides a flowchart of a temperature control method for an infrared polarized light therapy device based on real-time feedback. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: As Figure 1 As shown, the present invention provides a temperature control method for an infrared polarized light therapy device based on real-time feedback, comprising: Step S1: Collect multi-source thermal radiation signals from the irradiation head of the infrared polarized light therapy instrument, perform time-frequency domain joint analysis on the multi-source thermal radiation signals, and extract dynamic feature vectors that characterize the thermal response state of the tissue surface.

[0025] In this embodiment, eight wavelengths are evenly distributed in a ring at the front end of the irradiation head. arrive Infrared thermopile sensor and a center wavelength The InGaAs photodiodes are used, and all sensor sampling frequencies are set to 1000Hz. The sampling data is transmitted through 24-bit... The quantized data is then fed into the FPGA buffer by the analog-to-digital converter.

[0026] A fourth-order Butterworth bandpass filter is performed on the buffered multi-source thermal radiation signal. The lower cutoff frequency of the filter passband is set to 0.1Hz and the upper cutoff frequency is set to 30Hz to filter out the baseline drift component caused by human breathing and the high-frequency electromagnetic interference component caused by the switching power supply.

[0027] The filtered signal is truncated by applying a Hanning window function with a window length of 256 sampling points and an overlap rate of 50% between adjacent windows. The time-domain signal is mapped to the time-frequency plane by a short-time Fourier transform with 512 points, generating a time-frequency energy matrix with a dimension of 256×257.

[0028] Integrating the time-frequency energy matrix along the frequency axis extracts the temporal envelope feature within a preset time window of 2 seconds. This envelope feature is a 256-dimensional vector formed by the sum of the elements in each column of the time-frequency energy matrix. Integrating the time-frequency energy matrix along the time axis extracts the frequency domain power spectral density feature, which is a 257-dimensional vector formed by the sum of the elements in each row of the time-frequency energy matrix.

[0029] The multi-source thermal radiation signal was decomposed into five levels using a wavelet packet decomposition algorithm. The db4 wavelet basis was employed to obtain the energy entropy and node energy proportions of each frequency band component. The formula for calculating the energy entropy is as follows: , For the first The percentage of energy at each node relative to the total energy is the ratio of the energy at each node to the total energy at the time and frequency.

[0030] The temporal envelope features, frequency power spectral density features, energy entropy, and node energy percentage are combined into a time-frequency joint feature set. A 16×16 two-dimensional heatmap matrix corresponding to the geometry of the irradiation head end face is constructed, with the matrix grid spacing corresponding to the actual physical size of 0.5 ms. The time-frequency joint feature set is mapped to the corresponding grid nodes of the two-dimensional heatmap matrix through bilinear interpolation.

[0031] The 3×3 Sobel operator is used to calculate the gray-level gradient magnitude and direction of the two-dimensional heatmap matrix in the horizontal and vertical directions. The formula for calculating the gray-level gradient magnitude is as follows: The formula for direction calculation is: .

[0032] Solving for the 7th-order Hu invariant moments and the 2nd-order central moments of the 2D thermogram matrix extracts geometric features characterizing the symmetry of the heat field distribution. The invariant moments are obtained by nonlinearly combining the normalized central moments. The formula for calculating the central moments is as follows: Singular value decomposition (SVD) is performed on the two-dimensional heatmap matrix, and the top 20% of eigenvalues ​​are extracted to form the principal component eigenvectors. SVD is implemented using the Jacobi iterative algorithm, with the iterative convergence threshold set to [value missing]. The gray-level gradient magnitude, direction, invariant moments, central moments, and principal component eigenvectors are sequentially concatenated to generate a dynamic feature vector with a total dimension of 1024.

[0033] Step S2: Input the dynamic feature vector into the pre-trained nonlinear mapping model and output an energy adjustment command sequence that matches the current thermal response state. The energy adjustment command sequence includes pulse width modulation parameters and polarization state switching parameters.

[0034] In this embodiment, the pre-trained nonlinear mapping model adopts a four-layer recurrent neural network architecture, including an input layer, a delay layer, a hidden layer, and an output layer. The input layer has 1024 neurons, receiving a 1024-dimensional dynamic feature vector, and simultaneously receiving the 12-dimensional energy adjustment command sequence output from the previous time step. A fully connected weight matrix with dimensions 1036×800 is set between the input layer and the delay layer. The delay layer performs fixed-step delay processing on the input data, with a delay step size of 5 sampling periods and a sampling period duration of 1ms. The delayed data is then mapped to the hidden layer through a linear transformation. The hidden layer has 800 neurons, and the neuron transfer function is the hyperbolic tangent function, expressed as follows: In the hidden layer, the input signals of each neuron are first weighted and summed, then activated by the hyperbolic tangent function before being output to the next layer. The output layer is configured with 12 neuron nodes, which are used to map the output of the hidden layer to the target control space and generate an initial instruction matrix with a dimension of 12×1.

[0035] The connection weights and bias parameters of the nonlinear mapping model are initialized using the Tent chaotic map, the expression of which is: ,in The value is 1.9999. The initial value is set to 0.30001, and values ​​within the sensitive region of the hyperbolic tangent function are generated iteratively. The initial weight matrix and bias vector are set to a value between 0.8 and 0.8. After initialization, the partial derivatives of the output of each neuron in the hidden layer with respect to the input layer variables are calculated. These partial derivatives are solved using an automatic differentiation algorithm, thereby deriving the overall Lipschitz constant of the nonlinear mapping model. The formula for calculating the Lipschitz constant is as follows: ,in For the model output function, The learning rate parameter of the hidden layer is dynamically adjusted based on the calculation results, with the initial value set to 0.001. The adjustment formula is as follows: Dropout regularization is applied to the output of the hidden layer, with a random deactivation ratio of 0.3. The deactivation operation is achieved by randomly setting the output of 30% of neurons to zero according to the Bernoulli distribution during the training phase. During the testing phase, all neuron outputs are retained and multiplied by 0.7 for scaling compensation.

[0036] The model training process uses historical clinical treatment data to construct a training set, which contains 5000 samples. Each sample contains a 1024-dimensional dynamic feature vector and a corresponding 12-dimensional energy adjustment instruction label. Before training, the input features are Z-score standardized, using the following formula: ,in The characteristic mean, The standard deviation of the feature is denoted as . The training uses an adaptive moment estimation algorithm with momentum term, the momentum coefficient is set to 0.9, the second moment estimation decay rate is set to 0.999, and the weight decay coefficient is set to 0.0001.

[0037] During training, each batch contains 64 samples, and the training lasts for 200 epochs. The loss function used is the mean squared error function, calculated using the following formula: ,in For the sample size, For the first The predicted output for each sample, For the first The true labels of each sample are then saved. After training, the model parameters are saved, including the connection weight matrix between the input layer and the time-delay layer, the connection weight matrix between the time-delay layer and the hidden layer, the connection weight matrix between the hidden layer and the output layer, and the bias vectors of each layer.

[0038] The energy regulation command sequence output process first involves normalizing the initial command matrix. The normalization formula is as follows: The matrix elements are constrained to a unit range of 0 to 1. The normalized initial instruction matrix is ​​then upsampled using radial basis function interpolation. A Gaussian kernel is used as the radial basis function, with a kernel width of 0.5 and an upsampling factor of 10, expanding the 12x1 matrix to 120x1, thus improving the time series resolution.

[0039] The upsampled command matrix is ​​decoupled using an orthogonal triangular decomposition algorithm. This decomposition is achieved through Householder transformation, decomposing the 120x1 matrix into intensity control components and polarization control components. Both the intensity control and polarization control components have a 60x1 dimension. The intensity control components are then activated by the Sigmoid function and converted into duty cycle values ​​for pulse width modulation parameters. The Sigmoid function expression is as follows: The duty cycle ranges from 0 to 100%. The polarization control component, activated by an inverse trigonometric function, is converted into the target angle value of the polarization state switching parameter. The inverse trigonometric function uses an arcsine function, and its expression is: The target angle ranges from 0 to 180°. The final generated energy regulation command sequence contains 60 pulse width modulation duty cycle parameters and 60 polarization state switching target angle parameters, with a sampling interval of 0.1ms. The command sequence is transmitted to the controller via the SPI communication interface.

[0040] Step S3: Synchronously adjust the driving current waveform of the infrared light source and the rotation angle of the polarization filter according to the energy adjustment command sequence, so that the distribution of photothermal energy output by the irradiation head changes dynamically with the tissue thermal response state.

[0041] In this embodiment, the photothermal transfer function of the infrared light source and the electromechanical servo transfer function of the polarization filter are established. A semiconductor laser array with a peak wavelength of 980nm is selected as the infrared light source, and its photothermal transfer function is determined experimentally as follows:

[0042] in For steady-state gain, the measured value in this embodiment is... , The time constant is 0.12 seconds, as measured in this embodiment. The polarizing filter is a liquid crystal polymer adjustable polarizer with an extinction ratio greater than 1000:1, and its electromechanical servo transfer function is determined by the system identification as follows:

[0043] in The position loop gain, measured in this embodiment, is 1.2 rad / V. The damping ratio is 0.7, as measured in this embodiment. The natural frequency is 85 rad / s, as measured in this embodiment.

[0044] Based on the duty cycle and target angle values ​​contained in the energy regulation command sequence, the desired trajectories of the current loop and position loop are calculated respectively. The desired trajectory of the current loop is the target current value corresponding to the duty cycle value. The calculation formula is:

[0045] in The maximum rated current of the infrared light source is set to 5A, and D is the duty cycle value, ranging from 0 to 1. The desired trajectory of the position loop is the encoder absolute position value corresponding to the target angle value, and the target position... The calculation formula is:

[0046] in This is the encoder's maximum count value. The target angle value ranges from 0 to... radian.

[0047] A feedforward decoupling control strategy is adopted to eliminate the dynamic coupling term between the drive current waveform adjustment and the polarization filter rotation. The coupling term mainly originates from the thermal deformation of the polarization filter caused by the infrared light source and the electromagnetic interference generated by the motor motion. The feedforward decoupling controller is designed as follows: The inverse matrix is ​​actually decoupled using a diagonal matrix. The decoupled current loop control quantity is:

[0048] The position loop control variable is:

[0049] in This is the actual current feedback value. This is the actual angle feedback value.

[0050] Based on the desired trajectory and the current feedback values ​​of the actual current and angle, calculate the current loop error and the position loop error. Current loop error Defined as:

[0051] Position loop error Defined as:

[0052] The current loop error and position loop error are input into the corresponding proportional-integral (PI) controller, which outputs the drive voltage control signal and the motor drive signal. The transfer function of the current loop PI controller is:

[0053] in The scaling factor is set to 0.8. The integral coefficient is set to 25. The transfer function of the position loop proportional-integral controller is:

[0054] in The scaling factor is set to 1.5. The integral coefficient is set to 40. Drive voltage control signal. for The output is limited to a range of 0 to 24V. Motor drive signal. for The output range is limited to -10 to 10V.

[0055] A space vector pulse width modulation (SVM) algorithm is used to generate the trigger signal for the switching transistors acting on the infrared light source driving circuit. The specific implementation steps of the SVM algorithm are as follows: First, the conduction state of the three-phase bridge arms is encoded into six non-zero vectors. to and 2 zero vectors , The duration of action of each vector is determined by sector determination. Sector determination is based on the reference voltage vector. Where Based on the coordinate system position, calculate the three variables X, Y, and Z:

[0056]

[0057]

[0058] in and For the reference voltage vector in Components in the coordinate system. The sector number N is determined based on the signs of X, Y, and Z, with the following correspondence: If but otherwise ;like but otherwise ;like but otherwise ; sector number .

[0059] Calculate the duration of action of two adjacent non-zero vectors within each sector. and The calculation formula is:

[0060]

[0061] in The switching cycle is set to [value]. , The DC bus voltage is set to 48V. The angle between the reference voltage vector and the starting edge of the current sector. Zero vector duration. The switching sequence of vectors is determined based on the sector number, and a seven-segment modulation method is used, with zero vectors inserted in each switching cycle. and Zero-vector interleaving positions are selected at the beginning and end of the switching cycle to reduce switching losses. The target average voltage is obtained through zero-vector interleaving modulation. for:

[0062] The rotation angle of the polarizing filter is decoded using a sine / cosine encoder feedback to obtain the absolute position and angular velocity values. A 17-bit absolute sine / cosine encoder is selected, whose output signals are two orthogonal sine wave signals, Sin and Cos. The decoding process includes: first, analog-to-digital conversion of the Sin and Cos signals, with a sampling frequency of 1MHz and a resolution of 12 bits. Then, arctangent calculation is performed on the sampled signals.

[0063] The original angle value is obtained. Then, the rotational speed and angular velocity are calculated from the original angle value. for:

[0064] The calculation is performed using the four-point central difference method, with a time interval of... The absolute position and angular velocity values ​​are fed forward to the output of the PI controller to form a composite control law. The expression for the composite control law is:

[0065] in The position feedforward coefficient is set to 0.95. The speed feedforward coefficient is set to 0.05. The pulse width and phase of the motor drive signal are corrected according to the composite control law. The corrected motor drive signal is then output to the motor drive chip via a digital signal processor. The motor drive chip uses a DRV8301, which outputs a three-phase pulse width modulation signal to drive the motor. A brushless DC motor is selected as the motor.

[0066] The infrared light source driving circuit adopts a synchronous buck topology, using a silicon carbide MOSFET (model C2M0080120D) as the switching transistor and a Si8233 as the driving chip. The space vector pulse width modulation signal is input to the IN pin of the Si8233, and after dead-time setting, outputs high-side and low-side driving signals. The dead-time is set to 500ns. The driving voltage control signal... The voltage is converted to analog voltage by a digital-to-analog converter and connected to the VREF pin of the Si8233 for adjusting the output voltage amplitude. The rotation angle of the polarization filter is fed back in real time by an encoder, forming a closed-loop control to ensure that the tracking error between the actual angle and the target angle is less than 0.1°. The dynamic changes in the distribution of photothermal energy are monitored in real time by an infrared thermal imager. The monitoring results show that the temperature control accuracy reaches ±0.3℃ and the response time is less than 0.5s.

[0067] Step S4: After performing energy regulation, continuously collect new multi-source thermal radiation signals, perform differential comparison between the newly extracted dynamic feature vector and the feature vector of the previous period, and iteratively update the input weights of the nonlinear mapping model based on the comparison results.

[0068] In this embodiment, after the synchronization control action is completed, the irradiation is conducted through eight wavelengths evenly distributed in a ring at the front end of the irradiation head. arrive An infrared thermopile sensor and an InGaAs photodiode with a center wavelength of 1550nm continuously acquire new multi-source thermal radiation signals at a sampling frequency of 1000Hz. The sampled data is processed through a 24-bit... After quantization by the analog-to-digital converter, the data is fed into the FPGA buffer. Following the joint time-frequency domain analysis process and the dynamic feature vector extraction process, a 1024-dimensional dynamic feature vector for the current period is generated. This generated dynamic feature vector is denoted as... Let the dynamic feature vector of the previous cycle be denoted as First calculate and The Euclidean distance between them is calculated using the following formula:

[0069] in express The dimensional elements, express The Dimensional element. Then calculate. and The cosine similarity between them is calculated using the following formula:

[0070] in This represents the dot product of the two. express L2 norm, express The L2 norm is then used. Subsequently, the Bartholomew's coefficient and KL divergence values, representing the difference in probability distributions between the two periodic eigenvectors, are constructed, specifically: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] and The data is mapped to 256 histogram bins. The width of each bin is equal to the maximum value minus the minimum value, divided by 256. The number of elements in each bin is counted, and the probability density is calculated. and The formula for calculating the Bach coefficient is:

[0071] The formula for calculating KL divergence is:

[0072] Euclidean distance Cosine similarity Bach coefficient and KL divergence The set of comparison indicators is combined, and four indicators in this set are assigned preset weights and summed using a weighted average. The preset weights are 0.35, 0.25, 0.2, and 0.2, respectively. The weighted summation formula is as follows:

[0073] Generate synthetic differential scalar .Will Compared with the preset threshold of 0.15, if If this occurs, it is determined that the tissue's thermal response state has undergone a sudden change, and a weight update trigger signal is generated.

[0074] In response to the weight update trigger signal, the prediction error gradient of the nonlinear mapping model in the current cycle is first calculated. The prediction error is defined as the difference between the actual output of the energy adjustment command sequence and the model's predicted output. The gradient is solved using an automatic differentiation algorithm, which adopts a reverse mode, calculating partial derivatives layer by layer from the output layer to the input layer. A momentum term and an adaptive learning rate factor are introduced to construct a gradient descent update rule with moment estimation. The momentum term coefficient is set to 0.9, and the adaptive learning rate factor uses the moment estimation method of the Adam optimizer. The formula for calculating the first-order moment estimation is as follows:

[0075] The formula for calculating the second-order moment estimate is:

[0076] in This represents the prediction error gradient at the current moment. This is an estimate of the first moment from the previous moment. This is the second-order moment estimate from the previous time step. Using the first-order and second-order moment estimates of the prediction error gradient, the update step size of the connection weights is adaptively adjusted. The adjusted learning rate is calculated using the following formula:

[0077] Based on the preprocessed prediction error gradient, the connection weights between the hidden layer and the input layer in the nonlinear mapping model are updated by expanding along the time axis using the backpropagation algorithm. The time expansion step size of the backpropagation is set to 5 sampling periods, consistent with the delay step size of the time delay layer. A maximum norm constraint is applied to the updated connection weights, with the maximum norm set to 10. The constraint method is to prune the absolute value of any weight element exceeding 10 to 10 or -10 to prevent weight overflow. After the update is completed, the new connection weights are written to the parameter storage area of ​​the nonlinear mapping model for model inference in the next cycle.

[0078] Example 2: The present invention provides a temperature control method for an infrared polarized light therapy device based on real-time feedback, and also includes a multimodal safety monitoring mechanism for the treatment process.

[0079] Specifically, during the energy regulation execution phase, environmental parameters inside the irradiation head are collected simultaneously, including cavity temperature rise, cooling fan speed, and harmonic components of the light source operating current. By establishing a three-level temperature warning threshold, when the cavity temperature exceeds 45°C, the PWM duty cycle is automatically reduced by 10%, and when it exceeds 55°C, the light source drive is cut off and the polarizer position is locked.

[0080] Simultaneously, an integrated contact impedance detection circuit is used. When poor contact between the treatment head and the patient's skin causes an impedance surge exceeding 30%, a soft shutdown procedure is immediately triggered. This replaces hard shutdown with a gradually decreasing power attenuation curve, preventing thermal shock damage. This safety mechanism shares the same clock domain with the main control system, and the response delay is controlled at the microsecond level, ensuring equipment and personnel safety under abnormal operating conditions.

[0081] Example 3: The present invention provides a temperature control method for an infrared polarized light therapy device based on real-time feedback, and also includes an adaptive calibration process based on the treatment course.

[0082] Specifically, after completing a single treatment cycle, the system automatically enters a sleep calibration mode, using a built-in standard blackbody radiation source to perform zero-point drift correction on the multi-source thermal radiation sensor. The aging curve of the sensor is fitted using a sliding window least squares method to dynamically compensate for sensitivity decay caused by long-term use, and the compensation coefficients are stored in non-volatile memory.

[0083] It also supports remote parameter configuration. Medical staff can connect to a handheld terminal via Bluetooth and upload customized temperature control curves based on individual patient differences. The system then converts these curves into bias parameters for a nonlinear mapping model, enabling personalized treatment. During calibration, the nodal energy thresholds of wavelet packet decomposition are updated simultaneously to ensure a control accuracy of ±0.3℃ under different ambient temperatures.

[0084] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A temperature control method for an infrared polarized light therapy device based on real-time feedback, characterized in that, The method includes: Step S1: Collect multi-source thermal radiation signals at the irradiation head of the infrared polarized light therapy instrument, perform time-frequency domain joint analysis on the multi-source thermal radiation signals, and extract dynamic feature vectors characterizing the thermal response state of the tissue surface. Step S2: Input the dynamic feature vector into the pre-trained nonlinear mapping model and output an energy adjustment command sequence that matches the current thermal response state. The energy adjustment command sequence includes pulse width modulation parameters and polarization state switching parameters. Step S3: Synchronously adjust the driving current waveform of the infrared light source and the rotation angle of the polarization filter according to the energy adjustment command sequence, so that the distribution of photothermal energy output by the irradiation head changes dynamically with the tissue thermal response state. Step S4: After performing energy regulation, continuously collect new multi-source thermal radiation signals, perform differential comparison between the newly extracted dynamic feature vector and the feature vector of the previous period, and iteratively update the input weights of the nonlinear mapping model based on the comparison results.

2. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 1, characterized in that, The time-frequency domain joint analysis process in step S1 includes: Perform bandpass filtering on multi-source thermal radiation signals to filter out baseline drift components and high-frequency electromagnetic interference components; The filtered signal is truncated by applying a window function, and the time-domain signal is mapped to the time-frequency plane by short-time Fourier transform to generate a time-frequency energy matrix. Integrate the time-frequency energy matrix along the frequency axis to extract the temporal envelope features within a preset time window; Integrate the time-frequency energy matrix along the time axis to extract the frequency domain power spectral density features; The wavelet packet decomposition algorithm is used to decompose the multi-source thermal radiation signal to obtain the energy entropy of each frequency band component and the energy ratio of each node. The time-domain envelope features, frequency-domain power spectral density features, energy entropy, and node energy percentage are combined into a time-frequency joint feature set.

3. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 2, characterized in that, The dynamic feature vector extraction process includes: Construct a two-dimensional heat map matrix corresponding to the geometry of the irradiation head end face, and map the time-frequency joint feature set to the corresponding grid nodes of the two-dimensional heat map matrix; Calculate the magnitude and direction of the grayscale gradient in the horizontal and vertical directions of the two-dimensional heatmap matrix; Solve for the invariant moments and central moments of the two-dimensional thermogram matrix to extract the geometric features characterizing the symmetry of the thermal field distribution; Singular value decomposition is performed on the two-dimensional heatmap matrix, and the eigenvalues ​​of a predetermined proportion are extracted to form the principal component eigenvectors. The gray-level gradient magnitude, direction, invariant moment, central moment, and principal component eigenvectors are sequentially concatenated to generate dynamic eigenvectors.

4. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 1, characterized in that, The nonlinear mapping model in step S2 includes an input layer, a delay layer, a hidden layer, and an output layer. The input layer receives dynamic feature vectors and the energy adjustment instruction sequence from the previous time step. The delay layer performs a preset delay on the data received from the input layer and then outputs it to the hidden layer; The hidden layer contains several neurons, and the hyperbolic tangent function is used as the transfer function. The output layer maps the output of the hidden layer to the target control space, generating the initial instruction matrix.

5. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 4, characterized in that, The connection weights and bias parameters of neurons in the hidden layer are initialized using a chaotic sequence; The initial weight matrix and bias vector within the sensitive region of the activation function are generated using the Tent chaotic mapping. Calculate the partial derivative of the output of each neuron in the hidden layer with respect to the input layer variables, and solve for the global Lipschitz constant of the nonlinear mapping model; The learning rate parameter of the hidden layer is dynamically adjusted based on the overall Lipschitz constant. Apply Dropout regularization to the output of the hidden layer to randomly deactivate a predetermined proportion of neurons.

6. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 5, characterized in that, The energy regulation command sequence output process includes: The initial instruction matrix is ​​normalized to constrain the values ​​of the matrix elements to a preset unit range; The time series resolution is improved by upsampling the normalized initial instruction matrix using radial basis function interpolation. The upsampled command matrix is ​​decoupled using an orthogonal triangular decomposition algorithm to separate the light intensity control component and the polarization control component. The light intensity control component is activated by the Sigmoid function and then converted into the duty cycle value of the pulse width modulation parameter. The polarization control component is activated by an inverse trigonometric function and then converted into the target angle value of the polarization state switching parameter.

7. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 1, characterized in that, The process of synchronously adjusting the driving current waveform of the infrared light source and the rotation angle of the polarization filter in step S3 includes: Establish the photothermal conversion transfer function of the infrared light source and the electromechanical servo transfer function of the polarization filter; Based on the duty cycle and target angle values ​​contained in the energy regulation command sequence, the desired trajectories of the current loop and position loop are calculated respectively. A feedforward decoupling control strategy is adopted to eliminate the dynamic coupling term between the driving current waveform adjustment and the polarization filter rotation. Based on the desired trajectory and the current feedback of the actual current and angle values, calculate the current loop error and the position loop error; The current loop error and position loop error are input into the corresponding proportional-integral regulator, which outputs the drive voltage control signal and the motor drive signal.

8. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 7, characterized in that, Step S3 further includes: A space vector pulse width modulation algorithm is used to generate the switching transistor trigger signal for the infrared light source driving circuit; The voltage space vector perpendicular to the stationary coordinate system is synthesized based on the conduction state of the three-phase bridge arms, and the target average voltage is obtained through zero-vector interleaving modulation. The rotation angle of the polarization filter is decoded by a sine and cosine encoder to obtain the absolute position and angular velocity values. The absolute position and angular velocity values ​​are fed forward to the output of the PI controller to form a composite control law. The pulse width and phase of the motor drive signal are corrected based on the composite control law.

9. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 1, characterized in that, The differential alignment process in step S4 includes: Calculate the Euclidean distance and cosine similarity between the newly extracted dynamic feature vector and the dynamic feature vector of the previous period; Construct the Bach coefficient and KL divergence values ​​to characterize the difference in probability distributions of two periodic eigenvectors; The values ​​of Euclidean distance, cosine similarity, Bach coefficient, and KL divergence are combined into a set of comparison indicators. By comparing the pre-defined weights assigned to each indicator in the indicator set and performing a weighted summation, a comprehensive difference scalar is generated. The integrated differential scalar is compared with a preset threshold. If it exceeds the preset threshold, it is determined that the tissue thermal response state has changed abruptly, and a weight update trigger signal is generated.

10. The temperature control method for an infrared polarized light therapy device based on real-time feedback as described in claim 9, characterized in that, The input weight iterative update process includes: In response to the weight update trigger signal, the prediction error gradient of the nonlinear mapping model in the current period is calculated; By introducing a momentum term and an adaptive learning rate factor, a gradient descent update rule with moment estimation is constructed. The update step size of the connection weights is adaptively adjusted by using the first-order moment estimation and second-order moment estimation of the prediction error gradient. Based on the preprocessed prediction error gradient, the backpropagation algorithm is used to expand and update the connection weights between the hidden layer and the input layer in the nonlinear mapping model along the time axis. Apply maximum norm constraints to the updated connection weights to prevent weight overflow.

Images