Intelligent control method and system for internal cooling wet preparation nickel-titanium root canal treatment file

Abstract

The invention relates to an intelligent control method and system for an internal cooling wet preparation nickel-titanium root canal treatment file, and solves the problems that traditional control is difficult to adapt to treatment requirements of root canals in different forms, excessive or insufficient cutting is likely to occur, and complications such as root canal wall lateral penetration and instrument breakage are caused. The method comprises the following steps: iteratively optimizing a root canal preparation strategy through a multi-target reward function by taking pressure, temperature, cleaning effect and anatomical characteristics as state spaces and taking a filing tool rotating speed, a motion mode and flushing fluid flow as action spaces, and generating a parameter combination comprising the optimal rotating speed, the optimal motion mode and the optimal flushing fluid flow; the parameter combination is output to a control system to drive the nickel-titanium root canal treatment file to execute corresponding operation, a safety threshold value is dynamically corrected according to the treatment stage and anatomical characteristics, and when the parameters deviate from the threshold value, early warning is triggered and compensation is executed. The method has the advantages that multi-parameter intelligent regulation and control are achieved, and the safety and effectiveness of root canal therapy are improved.

Description

Technical Field

[0001] The present invention relates to the field of oral medical instruments, and in particular to an intelligent control method and system for an internally cooled wet-prepared nickel-titanium root canal treatment file. Background Art

[0002] In the field of oral disease treatment, root canal therapy is a key means of treating pulp disease and periapical disease. Its core goal is to completely remove the infected substances in the root canal and tightly fill the root canal to promote the healing of periapical lesions. The internally cooled wet-prepared nickel-titanium root canal treatment file plays an important role in the root canal preparation process due to its good flexibility and cutting ability. The file delivers irrigation fluid through the internal cooling system, which promptly removes debris and reduces the temperature during the rotary cutting process. At the same time, it flushes the inner wall of the root canal, assists in removing infected tissue and smear layer, and lays the foundation for subsequent treatment. However, the structure of the root canal system is complex and changeable, and individual differences are significant, which places extremely high demands on the operation and control of the treatment file.

[0003] Currently, traditional root canal file control relies primarily on the physician's clinical experience and manual manipulation, assessing root canal preparation through visual observation and simple instrument measurement. With technological advancements, some studies have attempted to incorporate sensor technology to monitor parameters such as pressure and temperature within the root canal, enabling simple feedback control based on preset thresholds. Other studies have used CBCT scans to obtain information about root canal anatomy to assist physicians in formulating treatment plans. However, in practice, this information is not closely integrated with real-time data from the treatment process, hindering efficient dynamic control.

[0004] Existing control methods have many drawbacks. Due to the complex anatomical structure of the root canal, traditional control is difficult to adapt to the treatment needs of root canals of different shapes, and is prone to over- or under-cutting, leading to complications such as lateral penetration of the root canal wall and instrument breakage. Feedback control based solely on preset thresholds cannot fully consider the dynamic changes in the environment inside the root canal during treatment, and the ability to comprehensively regulate pressure, temperature fluctuations, and cleaning effects is insufficient. At the same time, there is a lack of effective coupling control between the various parameters, making it difficult to ensure cutting efficiency while taking into account the safe range of temperature and pressure, affecting the treatment effect and patient comfort. Therefore, a more intelligent and precise control method is urgently needed to improve the safety and effectiveness of root canal treatment. Summary of the Invention

[0005] In order to achieve multi-parameter intelligent control and improve the safety and effectiveness of root canal treatment, the present application provides an intelligent control method and system for an internally cooled wet-prepared nickel-titanium root canal treatment file.

[0006] In a first aspect, the present application provides an intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file, which adopts the following technical solutions:

[0007] An intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file, comprising:

[0008] Obtain root canal 3D structural parameters through CBCT scanning, import them into the control system, initialize multiple sensors, and perform self-test and zero-point calibration;

[0009] The system simultaneously collects raw data on root canal pressure, temperature, and rotation speed, integrates CBCT anatomical features to build a normalized model, calculates pressure and temperature deviations, and obtains cleaning effect indicators through irrigant reflux turbidity detection and pressure fluctuation analysis.

[0010] Taking pressure deviation, temperature deviation and rotation speed as input, the flushing fluid flow and file rotation speed are dynamically adjusted through the fuzzy rule base to achieve preliminary coupled control of cutting resistance and temperature.

[0011] A deep reinforcement learning model was constructed, with pressure, temperature, cleaning effect, and anatomical features as the state space, and file speed, motion pattern, and irrigation fluid flow rate as the action space. The root canal preparation strategy was iteratively optimized using a multi-objective reward function to generate a parameter combination containing the optimal speed, motion pattern, and irrigation fluid flow rate. This parameter combination was then output to the control system to drive the nickel-titanium root canal treatment file to perform the corresponding operation. The motion modes included continuous rotation, reciprocating rotation, and pulsed rotation.

[0012] The safety threshold is dynamically modified according to the treatment stage and anatomical characteristics. When the parameters deviate from the threshold, an early warning is triggered and compensation is performed. At the same time, the control effect is fed back to the reinforcement learning model for strategy update.

[0013] By adopting the above technical solution, this method integrates CBCT data and multi-sensor information, and dynamically adjusts the parameters of the treatment file through fuzzy rules and deep reinforcement learning to achieve coupled control of cutting resistance and temperature; it corrects the threshold according to the treatment stage and anatomical characteristics, improves the accuracy of root canal treatment, reduces the risk of complications, and improves the safety and effectiveness of treatment.

[0014] Optionally, the raw data of pressure, temperature, and rotation speed in the root canal are collected synchronously, and the normalized model is constructed by integrating the CBCT anatomical features. The pressure deviation and temperature deviation are calculated, including:

[0015] Synchronously collect the raw data of pressure, temperature and speed in the root canal, and use hardware clock synchronization technology and cubic spline interpolation algorithm to achieve spatiotemporal registration and preprocessing of multi-source data;

[0016] Root canal anatomical features are extracted from CBCT images using 3D-CNN. Sensor temporal features are enhanced using a bidirectional LSTM combined with an attention mechanism. Cross-modal feature adaptive fusion is achieved through a gated fusion unit.

[0017] A priori probability model of root canal pressure-temperature distribution is established based on the Gaussian mixture model. The real-time collected pressure and temperature data are combined with the priori model using the preset Bayesian update formula to obtain the posterior probability distribution.

[0018] The fused features are normalized through nonlinear mapping functions to calculate pressure and temperature deviations;

[0019] The covariance matrix of process noise and observation noise of Kalman filter is dynamically adjusted according to the fusion features, and the extended state space model is constructed by taking anatomical features as additional state variables.

[0020] By employing these technical solutions, the technology achieves precise registration and fusion of multi-source data, extracts root canal anatomy and sensor timing characteristics, builds a probabilistic model, and dynamically updates it. Through normalization and model expansion, it accurately calculates pressure and temperature deviations, providing a reliable data foundation for intelligent control of root canal treatment parameters and improving treatment accuracy.

[0021] In a second aspect, the present application provides an intelligent control system for an internally cooled wet-prepared nickel-titanium root canal treatment file, which adopts the following technical solutions:

[0022] An intelligent control system for an internally cooled wet-prepared nickel-titanium root canal treatment file comprises a memory, a processor, and a program stored in the memory and executable on the processor. The program, when loaded and executed by the processor, can implement the intelligent control method for the internally cooled wet-prepared nickel-titanium root canal treatment file as described in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of an intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The present application is further described in detail below with reference to the accompanying drawings.

[0025] Reference Figure 1 , is an intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file disclosed in this application, comprising:

[0026] Step S100 , obtaining the three-dimensional structural parameters of the root canal through CBCT scanning, importing them into the control system and initializing the multi-sensor, and performing self-test and zero-point calibration.

[0027] CBCT scanning, also known as cone-beam CT scanning, is an imaging technology used in oral and maxillofacial medicine. It scans the target area with a cone-shaped X-ray beam to obtain three-dimensional structural information of the root canal. To obtain three-dimensional root canal structural parameters, a CBCT scanner places the patient's teeth and surrounding maxillofacial tissues in the device. The device scans with X-rays according to set parameters. The detector receives the signal and converts it into digital image data. A computer then processes the data using three-dimensional reconstruction algorithms (such as filtered back projection) to generate a three-dimensional image of the root canal and its associated parameters.

[0028] Root canal 3D structural parameters: Parameters of the root canal's physical characteristics in three dimensions, including length, diameter, and curvature, obtained using CBCT scanning technology. Control system: A device or system that receives and processes various sensor data and, based on pre-set algorithms and control logic, sends commands to adjust the operation of the treatment device. Multiple sensors: These include pressure sensors, temperature sensors, and speed sensors, which are used to detect pressure changes and temperature conditions within the root canal, as well as the rotational speed of the nickel-titanium root canal file.

[0029] The general process is described as follows:

[0030] 1. Data processing and import: Preliminary processing of the three-dimensional structural parameters of the root canal obtained from the CBCT scan is performed. A three-dimensional digital model of the root canal (a set of coordinate points) and related parameters (cross-sectional area, etc.) are obtained using a three-dimensional reconstruction algorithm. The processed data is then imported into the control system. This may involve data format conversion (e.g., DICOM to STL format).

[0031] 2. Multi-sensor initialization: Determine the range, resolution, and installation location of each sensor and perform calibration. For example, a pressure sensor with a range of 0-100kPa and a resolution of 0.1kPa is installed at the tip of the nickel-titanium endodontic file; a temperature sensor with a range of 20°C-60°C and a resolution of 0.01°C is installed near the nickel-titanium endodontic file; and a speed sensor with a measurement range of 0-1000 rpm and an accuracy of ±1 rpm is installed on the nickel-titanium endodontic file motor or drive shaft.

[0032] 3. System Self-Test and Zero-Point Calibration: The control system performs a self-test to check the communication connections, operating voltage and current of each sensor, and whether any fault codes are present. Zero-point calibration is then performed. The pressure sensor is set to output zero pressure when the root canal is at normal atmospheric pressure. The temperature sensor is recorded and calibrated at a standard temperature. The speed sensor records the speed output signal when the nickel-titanium root canal file is stationary and sets it to zero speed.

[0033] In step S200 , the raw data of pressure, temperature, and rotation speed in the root canal are synchronously collected, the anatomical features of CBCT are integrated to construct a normalized model, the pressure deviation and temperature deviation are calculated, and the cleaning effect index is obtained through the turbidity detection of the irrigating fluid reflux and the pressure fluctuation analysis.

[0034] Among them, pressure deviation refers to the difference between the actual pressure in the root canal and the set target pressure, reflecting the degree of deviation between the current pressure state and the ideal state. Temperature deviation refers to the gap between the actual temperature in the root canal and the set target temperature, which is used to measure whether the temperature is within the expected range. Normalization model is a model that standardizes data, converting data of different dimensions and orders of magnitude to the same scale range for subsequent processing and analysis. Cleaning effect index is a quantitative indicator that comprehensively reflects the degree of cleanliness in the root canal, which is derived based on information such as the turbidity of the reflux of the irrigation fluid and pressure fluctuations.

[0035] Synchronous raw data collection: Initialized and calibrated pressure, temperature, and speed sensors are used to acquire real-time root canal pressure, temperature, and speed data. The sampling frequency can be set as needed, such as 10 samples per second. Testing and analysis to obtain cleaning effect indicators: A turbidity sensor is used to detect the turbidity of the irrigant return flow and analyze pressure fluctuations. Through specific weighting and comprehensive calculation, the cleaning effect index is obtained. For example, the turbidity weight can be set to 0.6 and the pressure fluctuation weight can be set to 0.4. The calculation formula is: Cleaning effect index = turbidity test result × 0.6 + pressure fluctuation analysis result × 0.4.

[0036] Construct a normalized model: Based on the anatomical features of CBCT, use the normalization formula (such as the linear normalization formula: , where x is the original data, and are the maximum and minimum values ​​of the data, The collected raw data is processed and a normalized model is constructed.

[0037] In step S300 , the pressure deviation, temperature deviation and rotation speed are used as inputs, and the flushing liquid flow rate and the file rotation speed are dynamically adjusted through the fuzzy rule base to achieve preliminary coupled control of cutting resistance and temperature.

[0038] Among them, the fuzzy rule base is a set of rules established based on fuzzy logic theory, which is used to convert the control rules described by fuzzy language variables into specific control actions.

[0039] Coupling control: refers to the comprehensive control of multiple interrelated variables (such as cutting resistance and temperature) to achieve stable operation of the system.

[0040] The general process is described as follows:

[0041] 1. Input variable processing: The pressure deviation, temperature deviation, and speed obtained in step S200 are used as input variables and fuzzified. For example, triangular fuzzy numbers are used to fuzzify each input variable, converting the precise numerical value into a fuzzy linguistic variable, such as "small," "medium," or "large."

[0042] 2. Fuzzy rule matching: Match the fuzzified input variables according to the rules in the fuzzy rule base. For example, for the rule "If the pressure deviation is large and the temperature deviation is high, increase the flushing fluid flow and reduce the file speed", check whether the current input variables meet the conditions of "large pressure deviation and high temperature deviation".

[0043] 3. Output variable determination: Based on the matching fuzzy rules, the fuzzy values ​​of the output variables are determined. For example, based on the above rules, the fuzzy outputs of "increasing the flushing fluid flow" and "reducing the file speed" are obtained.

[0044] 4. Defuzzification: Convert the fuzzy output into an accurate value. For example, the centroid method is used for defuzzification, and the calculation formula is: ,in, is the possible value of the output variable, is the corresponding fuzzy membership degree, thereby obtaining the specific values ​​of flushing fluid flow and file speed, and realizing the preliminary coupling control of cutting resistance and temperature.

[0045] Step S400: Construct a deep reinforcement learning model, using pressure, temperature, cleaning effect, and anatomical features as the state space, and file speed, motion mode, and irrigation fluid flow rate as the action space. The root canal preparation strategy is iteratively optimized through a multi-objective reward function to generate a parameter combination including the optimal speed, motion mode, and irrigation fluid flow rate. The parameter combination is then output to a control system to drive the nickel-titanium root canal treatment file to perform the corresponding operation, wherein the motion modes include continuous rotation, reciprocating rotation, and pulsed rotation.

[0046] Among them, the deep reinforcement learning model is a model that combines deep learning and reinforcement learning, which learns the optimal behavior strategy through the interaction between the intelligent agent and the environment. The deep reinforcement learning model of this application is constructed using the deep Q-network (DQN) architecture. Its core is the Q-network, which is used to approximate the Q-value function. The input of the Q-network is the state space, and the output is the Q-value corresponding to each action. The network weights are continuously updated through interaction with the environment. For example, a convolutional neural network (CNN) is used as the main structure of the Q-network, which contains multiple convolutional layers and fully connected layers to extract state features and calculate Q values.

[0047] State space: The set of states that describe the environment, in this case, the combination of pressure, temperature, cleaning effect, and anatomical features. Action space: The set of actions that the agent can take, in this case, the combination of file speed, movement pattern, and irrigation fluid flow rate.

[0048] Multi-objective reward function: A function used to measure the quality of an agent's behavior, taking into account multiple objective factors.

[0049] According to the requirements of treatment effect and safety, the reward function is set. For example, the reward function ,in, Reward for cleaning effect, For security rewards, For the preparation efficiency reward, w1, w2, and w3 are the weight coefficients of each reward item, and w1+w2+w3=1.

[0050] The general process is described as follows:

[0051] 1. Model construction and initialization: Build a deep reinforcement learning model based on DQN, initialize the weight parameters of the Q network, and set hyperparameters such as learning rate and discount factor.

[0052] 2. Definition of state space and action space: Quantify pressure, temperature, cleaning effect and anatomical characteristics as vector representations of the state space; define the action space as a combination of file speed (e.g., 0-500 rpm), motion mode (continuous rotation, reciprocating rotation, pulsed rotation) and flushing fluid flow rate (e.g., 0-10 ml / min).

[0053] 3. Multi-objective reward function setting: According to factors such as cleaning effect, safety and efficiency during the treatment process, set the reward function. For example, if the cleaning effect is good, then Take a higher positive value, if the parameter is close to the safety threshold Take a negative value, if the preparation efficiency is high Take a positive value.

[0054] 4. Iterative Optimization and Policy Output: By interacting with the simulated root canal preparation environment, the agent executes actions, receives reward signals, and uses experience replay to store interaction data. Stochastic gradient descent is then used to update the Q network weights, continuously optimizing the policy. After multiple rounds of iteration, a parameter combination containing the optimal rotational speed, motion pattern, and irrigation flow rate is obtained and output to the control system to drive the nickel-titanium root canal file to perform the corresponding operation.

[0055] Step S500: Dynamically modify the safety threshold according to the treatment stage and anatomical characteristics. When the parameter deviates from the threshold, an early warning is triggered and compensation is performed. At the same time, the control effect is fed back to the reinforcement learning model for strategy update.

[0056] Safety thresholds: The limits of parameters such as pressure and temperature set during root canal treatment to ensure treatment safety. Compensation mechanisms: When parameters deviate from safety thresholds, the system takes adjustments to return them to their normal range. Feedback control: The process of feeding control effects back to the reinforcement learning model to update and optimize the strategy.

[0057] The general process is described:

[0058] Dynamically adjust the safety threshold based on the treatment phase (e.g., initial preparation, detailed preparation, etc.) and anatomical features (e.g., root canal curvature, diameter, etc.). For example, during the initial preparation phase, the root canal pressure safety threshold can be relaxed to 85kPa; during the detailed preparation phase, it can be tightened to 75kPa.

[0059] Early Warning and Compensation: Root canal parameters are monitored in real time. When parameters deviate from safety thresholds, an early warning is triggered. Based on the compensation mechanism, appropriate compensation actions are taken. For example, if the pressure reaches 90kPa (exceeding the initial preparation threshold of 85kPa), the irrigation fluid flow rate is reduced by 10%.

[0060] Feedback and policy updates: Feedback the control effect to the reinforcement learning model to update the weight coefficients in the reward function. For example, if the cleaning effect is significantly improved after a certain compensation, the value of w1 can be appropriately increased, while w2 and w3 can be adjusted to maintain the sum of 1, thereby optimizing the subsequent policy.

[0061] The original data of pressure, temperature, and speed in the root canal are collected simultaneously, and the normalized model is constructed by integrating the CBCT anatomical features. The pressure deviation and temperature deviation are calculated, including:

[0062] Step S210 , synchronously collect the raw data of pressure, temperature, and rotation speed in the root canal, and use hardware clock synchronization technology and cubic spline interpolation algorithm to achieve spatiotemporal registration and preprocessing of multi-source data.

[0063] Hardware clock synchronization technology utilizes the clock signals of hardware devices to align the times of multiple data acquisition sources, ensuring synchronous data collection. Cubic spline interpolation is a mathematical interpolation method used to generate smooth curves between known data points and estimate the values ​​of unknown points. Spatiotemporal registration aligns data from different sources in time and space, making them comparable and fusible.

[0064] The general process is as follows: 1. Synchronous acquisition: Utilizing hardware clock synchronization technology, we ensure that pressure, temperature, and speed sensors collect data on the same time scale. For example, a trigger signal is used to cause all sensors to begin collecting data on the rising edge of each clock cycle. 2. Spatiotemporal registration and preprocessing: The collected data is aligned in time and space. For example, data from pressure and temperature sensors at different locations is unified into the same spatial coordinate system through coordinate transformation. Then, a cubic spline interpolation algorithm is applied to the data to fill in the gaps between data collections and generate a continuous data sequence.

[0065] In step S220, 3D-CNN is used to extract root canal anatomical features from the CBCT image, bidirectional LSTM is combined with an attention mechanism to enhance sensor temporal features, and cross-modal feature adaptive fusion is achieved through a gated fusion unit.

[0066] Among them, 3D-CNN (three-dimensional convolutional neural network): a deep learning model for processing three-dimensional data, which extracts features in three-dimensional space through convolution operations. Bidirectional LSTM (long short-term memory network): a recurrent neural network capable of processing sequence data. Bidirectional LSTM can simultaneously utilize past and future contextual information. Attention mechanism: a technology used to enhance the model's ability to focus on important features, enabling the model to dynamically assign weights to different time steps or features. Gated Fusion Unit (GFU): a structure for multimodal data fusion that adaptively fuses features of different modalities through a gating mechanism.

[0067] The general process is described:

[0068] Feature extraction: 3D-CNN is used to perform convolution operations on CBCT images to extract the anatomical features of the root canal. For example, if the CBCT image size is 64×64×64, the feature map obtained after 3D-CNN extraction is 16×16×16 with 64 channels.

[0069] Time series feature enhancement: The time series data collected by the sensor is input into a bidirectional LSTM, and the attention mechanism is used to dynamically assign weights to highlight the features of key time steps. For example, for time series data of length 100, the bidirectional LSTM outputs the hidden state of each time step, and the attention mechanism calculates the weight of each time step, ultimately obtaining weighted time series features.

[0070] Feature fusion: Adaptively fuse image features and temporal features through the gated fusion unit. For example, the formula of the gated fusion unit is: ;

[0071] in, is the image feature, is a temporal feature, is the sigmoid activation function, is the weight matrix, is the bias term, Indicates element-by-element multiplication. Finally, the fusion feature is obtained .

[0072] Step S230 , establishing a root canal pressure-temperature distribution prior probability model based on the Gaussian mixture model, and combining the real-time collected pressure and temperature data with the prior model using a preset Bayesian update formula to obtain a posterior probability distribution.

[0073] Among them, Gaussian mixture model: a tool based on probability generation model, which assumes that data is a mixture of multiple Gaussian distributions and is used to represent complex data distribution.

[0074] Bayesian update formula: Based on Bayesian theorem, prior knowledge and new observation data are used to update the belief of unknown parameters. The formula is ,in is the posterior probability, It seems that is the prior probability, is the marginal likelihood.

[0075] The general process is described:

[0076] Establish a priori model: Assume that the root canal pressure and temperature data are a mixture of two Gaussian distributions in the form of .

[0077] in, is the mixing coefficient, and The mean and covariance matrices are used to fit the historical data using the EM algorithm to obtain the model parameters. .

[0078] Bayesian update: After collecting pressure p and temperature t data in real time, calculate the likelihood , assuming the prior probability is , marginal likelihood Obtained by data normalization. Substituted into the Bayesian formula, the posterior probability distribution is obtained , reflecting the updated probability estimates of root canal pressure and temperature after observing new data.

[0079] Step S240 , normalizing the fused features through a nonlinear mapping function, and calculating the pressure and temperature deviations.

[0080] Nonlinear mapping functions: Functions used to transform data from its original space to another space for better processing and analysis. Normalization: The process of scaling data to a specific range (such as [0, 1]) to ensure that different features have the same scale. Deviation calculation: Calculating the difference between actual and target values ​​to evaluate system performance.

[0081] The general process is described:

[0082] 1. Processing after feature fusion: Assume that after processing in steps S220 and S230, a fused feature vector f is obtained, whose dimension is n.

[0083] Non-linear mapping: Input the fused feature vector f into the non-linear mapping function. For example, using the Sigmoid function:

[0084] ;

[0085] in, is the Sigmoid function, is the eigenvector after mapping.

[0086] Normalization: the mapped feature vector Perform normalization. For example, use Min-Max normalization:

[0087] ;

[0088] in, is the normalized eigenvector, and are the minimum and maximum values ​​of f′ respectively.

[0089] Deviation calculation: Calculate the deviation between the actual and target values ​​of pressure and temperature. Assume the target pressure is , the target temperature is , the actual pressure is , the actual temperature is .

[0090] The deviation is: ; ;in, is the pressure deviation, These deviations will be used in the subsequent control and optimization process.

[0091] Step S250 , dynamically adjusting the covariance matrix of the process noise and observation noise of the Kalman filter according to the fusion features, and constructing an extended state space model using the anatomical features as additional state variables.

[0092] Among them, Kalman filtering is a recursive algorithm for estimating the state of a dynamic system, capable of estimating the state of the system using incomplete and noisy measurement data. Extended state-space model: Based on the traditional state-space model, additional features or variables are incorporated into the model as additional state variables to enhance the model's representation capabilities.

[0093] The general process is described as follows:

[0094] Initialize the Kalman filter: Assume that the system state vector is x, the process noise covariance matrix is ​​Q, the observation noise covariance matrix is ​​R, and the initial state estimate is , the initial estimation error covariance matrix is .

[0095] Dynamically adjust the noise covariance matrix: Dynamically adjust Q and R based on the reliability of the fused features and environmental changes. For example, if the sensor data noise increases, increase the value of R.

[0096] Construct an extended state space model: incorporate anatomical features as additional state variables into the state vector to form a new state vector ,in, is the anatomical feature vector. Accordingly, the state transfer matrix F and observation matrix H of the system are updated to reflect the influence of anatomical features on the system state.

[0097] Kalman filter iteration: performs the prediction and update steps of the Kalman filter. The prediction step includes state prediction and error covariance prediction The update step consists of calculating the Kalman gain , status update and error covariance update ,in, is the forecast estimation error covariance matrix, is the observation matrix, which is used to map the state vector to the observation space.

[0098] The root canal anatomical features in CBCT images are extracted through 3D-CNN, and the sensor temporal features are enhanced using a bidirectional LSTM combined with an attention mechanism. The gated fusion unit is used to achieve cross-modal feature adaptive fusion, including:

[0099] In step S221, multi-scale feature extraction is performed on the CBCT image through a 3D convolutional neural network. A three-dimensional convolutional attention module is used in combination with spatial pyramid pooling to perform semantic-level positioning of the apical stenosis area, root canal bifurcation, root canal curvature, and calcified area, generating a three-dimensional anatomical feature map containing spatial position weights.

[0100] Among them, the 3D Convolutional Neural Network (3D-CNN) is a deep learning model for processing 3D data, extracting features in 3D space through 3D convolution operations. The 3D Convolutional Attention Module combines convolution operations with an attention mechanism to enhance the model's ability to focus on important features. Spatial Pyramid Pooling is a spatial pooling technique for processing features of different scales, capable of extracting multi-scale features.

[0101] The general process is described:

[0102] Multi-scale feature extraction: 3D-CNN is used to perform convolution operations on CBCT images to extract features at different scales. Assuming the CBCT image size is 64×64×64, the feature map obtained after 3D-CNN extraction is 16×16×16 with 64 channels.

[0103] Attention module enhancement: A 3D convolutional attention module is added after the convolution layer, combined with spatial pyramid pooling, to enhance the feature extraction capabilities of the apical narrow area, root canal bifurcation, root canal curvature, and calcified areas. For example, the formula of the 3D convolutional attention module is:

[0104] ;

[0105] ;

[0106] Among them, F is the input feature map, is the convolution kernel, is the bias term, σ is the sigmoid activation function, It is the feature map after attention enhancement.

[0107] Spatial Pyramid Pooling: Perform spatial pyramid pooling on the feature map after attention enhancement to generate multi-scale feature representations. For example, the feature map is downsampled to 8×8×8, 4×4×4, and 2×2×2 scales, and then pooled to obtain multi-scale feature vectors.

[0108] Semantic-level localization: Multi-scale features are classified using fully connected or convolutional layers to generate a 3D anatomical feature map containing spatial position weights. For example, using a fully connected layer to classify multi-scale features, semantic-level localization results for apical stenosis, root canal bifurcations, root canal curvatures, and calcified areas are obtained, and corresponding feature maps are generated.

[0109] In step S222, a bidirectional long short-term memory network is used to capture the time series features of the pressure sensor data and the temperature sensor data. A temporal self-attention mechanism is introduced and combined with the treatment stage embedding vector. By calculating the attention weight of the query-key-value triple, the importance of the temporal features of different treatment stages is dynamically adjusted to generate temporal enhanced features. The embedding vector includes crown preparation, mid-section preparation, and apical preparation.

[0110] Bidirectional Long Short-Term Memory (BiLSTM): A recurrent neural network that leverages both past and future information, suitable for processing time series data. Temporal Self-Attention: A mechanism that focuses on key time steps in a time series, dynamically adjusting the importance of features at different time steps by calculating attention weights for query-key-value triples. Embedding Vector: A low-dimensional vector used to represent categorical variables (such as treatment stage).

[0111] The general process is described:

[0112] Time series feature extraction: The time series data collected by the pressure and temperature sensors is fed into a bidirectional LSTM to capture time series features. For example, for a time series of length 100, the bidirectional LSTM outputs the hidden state for each time step.

[0113] Enhanced attention mechanism: Introducing the temporal self-attention mechanism, combined with the treatment stage embedding vector (such as crown preparation, mid-preparation, apical preparation), dynamically adjusts the importance of temporal features of different treatment stages. For example, calculating the attention weight of the query-key-value triple:

[0114] ;

[0115] Among them, Q, K, and V are query, key, and value matrices respectively. is the dimension of the key. In this way, the features of the key time steps are highlighted.

[0116] Time-enhanced feature generation: The features adjusted by the attention mechanism are used to generate time-enhanced features for subsequent feature fusion.

[0117] Step S223: construct a dynamic routing feature interaction network based on the capsule network, generate a joint representation matrix of anatomical features and temporal features through bilinear pooling, use the cross-attention mechanism to establish bidirectional semantic associations of cross-modal features, optimize the feature interaction weights through the iterative routing algorithm, and generate cross-modal interaction features.

[0118] Capsule Network: A novel deep learning network architecture that uses capsules to better capture the spatial hierarchical relationships between features. Bilinear Pooling: A pooling method for generating interactions between features, capable of capturing high-order relationships between features. Cross-Attention: A mechanism for enhancing the interaction between features from two different modalities.

[0119] The general process is described:

[0120] Feature interaction: A dynamic routing feature interaction network based on a capsule network is constructed, and anatomical features and temporal features are input into the capsule network. For example, anatomical features and temporal features are represented as two capsules, each containing multiple neurons.

[0121] Bilinear pooling: Bilinear pooling is used to generate a joint representation matrix of anatomical features and temporal features. For example, if the dimension of the anatomical features is da and the dimension of the temporal features is dt, the dimension of the joint representation matrix after bilinear pooling is da×dt.

[0122] Cross-attention mechanism: Use the cross-attention mechanism to establish bidirectional semantic associations between cross-modal features. For example, calculate the attention weights of anatomical features on temporal features, and vice versa:

[0123] ;

[0124] ;

[0125] in, and is the weight matrix, and They are anatomical features and temporal features respectively.

[0126] Iterative routing optimization: This algorithm optimizes feature interaction weights through iterative routing to generate cross-modal interaction features. For example, after r rounds of iterative routing, the capsule weights are updated to ultimately generate cross-modal interaction features.

[0127] In step S224, the root canal curvature, root canal calcification rate, and root canal diameter change rate parameters are calculated based on the three-dimensional anatomical features, and a complexity index is generated through the anatomical complexity evaluation module. This index is used as the input of the gated fusion unit, and the complexity index is mapped to the [0,1] interval through the Sigmoid activation function to generate a fusion weight coefficient and construct a fusion function.

[0128] The specific fusion function is as follows:

[0129] ;

[0130] in, The final output fusion feature contains information about anatomical structure and sensor data. is the Sigmoid activation function, is the anatomical complexity index, are learnable weights and biases, For anatomical features, is the sensor characteristic, This is element-wise multiplication.

[0131] The anatomical complexity assessment module is used to assess the complexity of root canal anatomy. The gated fusion unit is a structure used to adaptively fuse multimodal data, dynamically adjusting the weights of different features through learning. The sigmoid activation function is a commonly used activation function that maps inputs to the [0, 1] interval and is used to generate weight coefficients.

[0132] The general process is as follows: 1. Calculate anatomical feature parameters: Key parameters such as root canal curvature, calcification rate, and diameter change rate are calculated from 3D anatomical features. 2. Generate complexity index: These parameters are input into the anatomical complexity assessment module and weighted summed to obtain the complexity index C. 3. Map fusion weight coefficient: The complexity index is mapped to the [0, 1] interval using the Sigmoid function to generate the fusion weight coefficient. 4. Based on the fusion weight coefficient, the weights of the anatomical and sensor features are dynamically adjusted to generate the final fused feature.

[0133] Taking pressure deviation, temperature deviation and rotation speed as input, the flushing fluid flow and file speed are dynamically adjusted through the fuzzy rule base to achieve preliminary coupled control of cutting resistance and temperature.

[0134] Step S310 , calculating the Kullback-Leibler divergence between the real-time pressure and the prior distribution, constructing a two-dimensional feature vector in combination with the temperature change rate, mapping the file rotation speed to a preset fuzzy state space, and generating a membership function matrix.

[0135] Among them, the Kullback-Leibler divergence (KL divergence) is a measure of the asymmetry between two probability distributions. A two-dimensional eigenvector combines two related features into a single vector, describing the current state of a system. A fuzzy state space defines the range of possible values ​​of a fuzzy variable and its corresponding fuzzy sets. A membership function matrix represents the degree of membership of an input variable with respect to each fuzzy set.

[0136] The general process is described:

[0137] 1. Calculate KL divergence: Calculate real-time pressure distribution With the prior pressure distribution KL divergence of:

[0138] ;

[0139] 2. Construct a two-dimensional feature vector: Construct a two-dimensional feature vector by combining KL divergence and temperature change rate .

[0140] 3. Mapping Speed ​​to Fuzzy State Space: Map the file speed r to a predefined fuzzy state space (e.g., low, medium, high). Assuming the fuzzy state space is defined as {low, medium, high}, and the speed range is 0–1000 rpm, a membership function (e.g., a triangular membership function) can be set to determine the degree of membership of the speed to each fuzzy state.

[0141] 4. Generate the membership function matrix: Generate the membership function matrix based on the membership of the speed. For example, for a speed of r = 500 rpm, if its membership to "low speed" is 0.3, its membership to "medium speed" is 0.7, and its membership to "high speed" is 0, then the membership function matrix is ​​[0.3, 0.7, 0].

[0142] In step S320, based on the three-dimensional anatomical feature map and the anatomical complexity index, a preset morphologically specific fuzzy rule subset is matched, and the rule weights are modulated time-varyingly in combination with the treatment stage identification results to form a two-dimensional decision logic based on anatomical features and treatment stages.

[0143] The morphology-specific fuzzy rule subset is a set of fuzzy rules pre-defined based on the different morphological characteristics of the root canal. Treatment stage identification results are the results of determining the current stage of root canal treatment (e.g., coronal preparation, mid-section preparation, apical preparation). Time-varying rule weight modulation is the process of dynamically adjusting fuzzy rule weights based on the treatment stage. Bidimensional decision logic is a logical framework that integrates anatomical features and treatment stage information for decision-making.

[0144] The general process is described:

[0145] 1. Matching rule subsets: Based on the three-dimensional anatomical feature map and anatomical complexity index, a rule subset suitable for the current root canal morphology is matched in the preset morphologically specific fuzzy rule library.

[0146] 2. Modulate the rule weights: Combined with the treatment stage identification results, the weights of each rule in the matched rule subset are modulated in a time-varying manner. For example, in the apical preparation stage, the weights of the rules related to the apical region are increased. Assuming the initial rule weight is w0, in the apical preparation stage, the weight is adjusted to:

[0147] ;

[0148] in, is the modulation coefficient, which can be set based on clinical experience or experimental data.

[0149] 3. Forming a dual-dimensional decision logic: By integrating anatomical features and modulated rule weights, a dual-dimensional decision logic is formed to guide subsequent control decisions. For example, a lookup table method is used to determine the rule weights for different anatomical features and treatment stages, and then decisions are made based on these weights.

[0150] In step S330, a Takagi-Sugeno fuzzy inference system is used to process the feature vector, a control quantity function expression is output through a neural network, and the fuzzy rule parameters are optimized using a preset meta-learning algorithm to improve the system's adaptability to new root canal morphologies.

[0151] Among them, the Takagi-Sugeno fuzzy inference system is an inference system based on fuzzy rules, whose output is a function of the input variables, commonly used in control and decision-making systems. A neural network is a computational model that simulates the neural network in the human brain and is used for pattern recognition and function approximation. A meta-learning algorithm is an algorithm designed to improve the generalization ability of a learning algorithm, which can learn how to learn better from experience.

[0152] The general process is described:

[0153] Eigenvector processing: The eigenvectors generated in steps S310 and S320 are input into the Takagi-Sugeno type fuzzy inference system. For example, assuming that the eigenvector is , the fuzzy rules are:

[0154] ;

[0155] in, is a fuzzy set, is the output function, which can be linear or nonlinear.

[0156] Neural network output control quantity function expression: The output function of the fuzzy inference system is approximated by a neural network. For example, using an MLP neural network, the activation function of its output layer is a linear function, and the output control quantity y is:

[0157] ;

[0158] Among them, w is the weight vector and b is the bias term.

[0159] Meta-learning algorithm optimization: Meta-learning algorithms are used to optimize fuzzy rule parameters and neural network weights to improve the system's adaptability to new root canal morphologies. For example, the MAML algorithm is used to train on multiple root canal morphology datasets to optimize the model's initial parameters so that it can quickly adapt to new tasks. The specific steps are as follows: 1. In the meta-training phase, tasks are sampled from multiple root canal morphology datasets, and the model is trained to quickly adapt on a small amount of data. 2. In the meta-testing phase, the model parameters are updated through gradient descent using the dataset of the new task to verify the model's rapid adaptability. 3. Through iterative optimization, the model's performance on new tasks is improved.

[0160] In step S340, the flushing liquid flow adjustment coefficient and the file speed correction coefficient are mapped to a preset safety threshold space to form a final control instruction; the H∞ control theory is introduced to design an anti-interference compensator to automatically correct the control instruction when the external environmental parameters suddenly change.

[0161] The safety threshold space defines the safe operating range for irrigation fluid flow and file speed, ensuring the safety of the treatment process. H∞ control theory is a robust control theory that aims to design controllers to minimize the system's sensitivity to disturbances. The anti-interference compensator is a device used to automatically correct control commands to ensure stable system operation even when external environmental parameters suddenly change.

[0162] The general process is described:

[0163] Generate the final control instruction: adjust the flushing fluid flow rate coefficient and file speed correction factor Mapped to the preset safety threshold space. For example, the safety threshold space is defined as:

[0164] ;

[0165] According to the output of the fuzzy inference system, the final control instructions are generated:

[0166] ;

[0167] in, and is the basic flow rate and speed.

[0168] Introducing anti-interference compensator: The anti-interference compensator is designed by H∞ control theory, and its design formula is:

[0169] ;

[0170] in, is the transfer function from disturbance d to error e, and K is the controller. Once designed, the compensator automatically adjusts the control instructions when external environmental parameters suddenly change. For example, if the external pressure suddenly changes, the compensator adjusts the control instructions to ensure stable system operation.

[0171] Generating a parameter combination including the optimal speed, motion pattern, and flushing fluid flow rate includes the following steps:

[0172] Step S410: Perform wavelet transform on the real-time pressure and temperature data to extract time domain and frequency domain features.

[0173] Wavelet transform is a mathematical tool used to decompose a signal into different frequency and time components, suitable for analyzing non-stationary signals. Time domain features describe the statistical properties of a signal along the time axis, such as mean, variance, and peak value. Frequency domain features describe the distribution characteristics of a signal in the frequency domain, such as power spectral density, dominant frequency, and frequency band energy.

[0174] Wavelet transform: Perform wavelet transform on real-time pressure and temperature data, and use Daubechies wavelet (such as db4) for multi-scale decomposition. For example, for pressure signal Perform wavelet decomposition:

[0175] ;

[0176] Among them, A represents the approximation coefficient and D represents the detail coefficient.

[0177] Time domain feature extraction: Calculate the time domain features of each scale after wavelet decomposition, such as mean, variance, and peak. For example, calculate the mean of the approximation coefficient:

[0178] ;

[0179] in, is the number of data points, is the approximate coefficient.

[0180] Frequency feature extraction: Calculate the frequency domain features of each scale after wavelet decomposition, such as power spectrum density and dominant frequency. For example, calculate the power spectrum density of detail coefficients:

[0181] ;

[0182] in, represents the Fourier transform, is the detail factor.

[0183] Step S420: Process the CBCT image through a graph neural network to generate an anatomical feature vector including curvature and calcification rate.

[0184] Among them, graph neural network (GNN): a neural network used to process graph-structured data, capable of capturing the relationship between nodes and their neighbors. Anatomical feature vector: a vector representation containing root canal anatomical information (such as curvature and calcification rate).

[0185] The general process is described as follows:

[0186] Constructing a graph structure: The graph structure of the root canal is extracted from the CBCT image, and the different parts of the root canal are represented as nodes, and the connection relationship between adjacent parts is represented as edges.

[0187] Feature extraction: Extract initial features for each node, such as grayscale value, gradient, etc.

[0188] Graph Neural Network Processing: Use GNN to perform convolution operations on the graph and update node features. After multi-layer GNN processing, node features containing anatomical information are obtained.

[0189] Generate anatomical feature vector: Pool the node features (such as global average pooling) to generate the anatomical feature vector. For example, use the formula:

[0190] ;

[0191] Where N is the number of nodes, is the eigenvector of the i-th node, is the anatomical feature vector.

[0192] Step S430: Using a gated recurrent unit to fuse the temporal sensor data with the spatial anatomical features to form a spatiotemporal state representation.

[0193] Among them, the Gated Recurrent Unit (GRU) is a recurrent neural network used to process sequence data and effectively captures temporal information. The spatiotemporal state representation is a state representation that integrates temporal and spatial information to comprehensively describe the system's operating status.

[0194] The general process is described:

[0195] Data preprocessing: Time series sensor data (such as pressure and temperature) and spatial anatomical features (such as root canal curvature and calcification rate) are preprocessed to have the same dimension and scale.

[0196] Feature fusion: Use the GRU model to fuse the preprocessed data. The input of GRU is the splicing vector of time series sensor data and spatial anatomical features. For example, assuming the dimension of time series sensor data is , the dimension of the spatial anatomical feature is , then the dimension of the input vector is .

[0197] State update: Update the hidden state through the GRU gating mechanism. The update formula of GRU is:

[0198] ;

[0199] ;

[0200] ;

[0201] ;

[0202] in, is the input vector, is the hidden state at the previous moment, and are the outputs of the update gate and reset gate respectively, is a candidate hidden state, is the hidden state at the current moment.

[0203] Generate spatiotemporal state representation: After GRU processing, the hidden state obtained That is, the spatiotemporal state representation, which can be used for subsequent decision-making and control.

[0204] Step S440 : Dynamically generate motion constraints based on the curvature and calcification rate in the anatomical feature vector to construct a safe motion space.

[0205] The anatomical feature vector is a vector representation of root canal anatomical information (e.g., curvature and calcification). The action constraints are dynamically generated based on anatomical features to ensure the safety and effectiveness of the operation. The safe action space defines the range of permissible operational parameters under the current anatomical conditions.

[0206] The general process is described as follows:

[0207] Anatomical feature extraction: Extract curvature C and calcification rate R from the anatomical feature vector.

[0208] Motion constraint generation: Generate motion constraints based on curvature and calcification rate. For example, set the upper limit of the rotation speed. and flushing fluid flow lower limit :

[0209] ;

[0210] ;

[0211] in, and is the base value, and is the adjustment factor.

[0212] The safe action space is constructed as follows: Based on the action constraints, a safe action space is constructed. For example, the safe action space can be expressed as:

[0213] ; Among them, r is the rotation speed, f is the flushing fluid flow, and m is the movement mode.

[0214] In step S450, the meta-strategy network selects a basic parameter combination from the preset strategy library based on the anatomical feature vector, executes the strategy network with the spatiotemporal state representation as input, calculates the parameter offset through the Actor-Critic architecture, and optimizes the basic parameter combination to form preliminary parameters.

[0215] The meta-policy network is a neural network used to select and optimize policies. It can choose a policy appropriate for the current anatomical features from a library of preset policies. The preset policy library contains a collection of various basic parameter combinations, each corresponding to different anatomical features and treatment phases. The actor-critic architecture is a reinforcement learning architecture that uses an actor network to select actions and a critic network to evaluate their value.

[0216] The general process is described as follows:

[0217] Select basic parameter combination: The meta-strategy network selects basic parameter combination from the preset strategy library according to the anatomical feature vector. For example, assuming the anatomical feature vector is , where C is the curvature, R is the calcification rate, and the output of the meta-policy network is the policy index i.

[0218] Execution Policy Network: Representing with Spatiotemporal State The Actor-Critic architecture calculates the parameter offset as input. The Actor network outputs the parameter offset Δθ, and the Critic network evaluates the value of the current strategy. For example, the output layer of the Actor network uses the tanh activation function to limit the offset to a certain range: ;in, is the weight matrix, is the bias term.

[0219] Optimize the basic parameter combination: Combine the basic parameter combination with the offset to form a preliminary parameter combination. For example, the basic speed r0 and the offset Δr are combined to obtain the preliminary speed rpre:

[0220] .

[0221] In step S460 , the preliminary parameter combination is projected into the safety constraint space, and the compensation amount of external disturbances such as tissue resistance fluctuation is calculated in combination with H∞ control theory to generate a robust parameter combination.

[0222] The general process is described as follows:

[0223] Projection to safety constraint space: Map the preliminary parameter combination (rpre, fpre, mpre) to the safety constraint space Asafe. For example, check whether the preliminary parameters satisfy:

[0224] ;

[0225] Compensation calculation: Combine H∞ control theory to calculate the compensation for external disturbances such as tissue resistance fluctuations. For example, the compensator can be designed by solving the Riccati equation:

[0226] ;

[0227] Generate robust parameter combination: Combine the preliminary parameter combination with the compensation amount to generate the final robust parameter combination. For example, the final speed and traffic for:

[0228] ;

[0229] in, and is the compensation amount.

[0230] Step S470: Calculate the expected value R of the robust parameter combination under the multi-objective reward function.

[0231] The multi-objective reward function is used to measure the overall performance of different objectives (such as cleaning effectiveness, safe operation, and preparation efficiency). The expected value R is calculated based on the reward function, reflecting the overall performance of the parameter combination under multiple objectives.

[0232] The general process is described:

[0233] Calculate the reward function: Based on the cleaning effect index , safety operation indicators and preparation efficiency index , and the corresponding weight coefficients (satisfy ), calculate the expected value R of the multi-objective reward function:

[0234] ;

[0235] Evaluate parameter combinations: Use the calculated expected value R to evaluate the comprehensive performance of the robust parameter combination under multiple objectives to determine whether it meets the treatment requirements.

[0236] In step S480, if the expected value R exceeds the preset threshold, the parameter combination output value is executed by the system.

[0237] The execution system is responsible for implementing control instructions and ensuring that the parameter combination is executed in actual operation. The preset threshold is the minimum expected value set according to treatment requirements and safety standards, which is used to determine whether the parameter combination is qualified.

[0238] The general process is described as follows:

[0239] Compare the expected value with the threshold: Compare the expected value R calculated in step S470 with the preset threshold For example, if the preset threshold is 0.8, the judgment condition is:

[0240] ;

[0241] Output parameter combination: If , then the robust parameter combination is output to the execution system to perform the corresponding operation. For example, the output parameter combination includes the final speed , flushing fluid flow and sport mode .

[0242] Step S490: If the expected value R does not reach the preset threshold, jump to the meta-learning mechanism to adjust the policy network parameters and regenerate the parameters.

[0243] An intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file further includes steps after the parameter combination output value execution system, specifically as follows:

[0244] Step S4A0: Start the online verification mechanism and monitor the deviation between the actual reward and the expected reward in real time through the Kalman filter.

[0245] Among them, the online verification mechanism: a mechanism that monitors the system operating status in real time and verifies whether the actual effect meets the expectations. Kalman filter: a recursive algorithm for estimating the state of a dynamic system, which can estimate the state of the system using incomplete and noisy measurement data.

[0246] The general process is described as follows:

[0247] Initialize the Kalman filter: define the state vector x, the process noise covariance matrix Q, the observation noise covariance matrix R, and the initial state estimate and the initial estimation error covariance matrix .

[0248] Prediction step: Predict the current state and estimate the error covariance based on the system dynamics model. For example:

[0249] ; ;

[0250] in, is the state transition matrix.

[0251] Update step: Use the actual observation data to update the state estimate and the estimated error covariance. For example:

[0252] ;

[0253] ;

[0254] ;

[0255] in, is the observation matrix, is the observation vector, is the Kalman gain.

[0256] Calculate the deviation: Calculate the deviation between the actual reward and the expected reward through the Kalman filter. For example, the actual reward is , the expected reward is , then the deviation is:

[0257] .

[0258] Step S4B0: analyzing whether the deviation exceeds a preset deviation.

[0259] Preset deviation: The deviation threshold set according to treatment requirements and system performance is used to judge the acceptable difference between the actual effect and the expected effect.

[0260] Step S4C0: If yes, start the adaptive calibration module and fine-tune the parameters based on the real-time data using the preset rules.

[0261] The adaptive calibration module automatically adjusts system parameters based on real-time data to maintain system performance stability. Preset rules guide parameter fine-tuning based on system characteristics and treatment requirements.

[0262] The general process is as follows: 1. Receive Deviation Signal: Receive a signal from step S4B0 indicating that the deviation exceeds the preset range. 2. Invoke Preset Rules: Based on the type and severity of the deviation, invoke the corresponding preset rule. For example, if the deviation is positive and large, apply the rule: "Increase the flushing fluid flow rate by 5%, and reduce the file speed by 3%." 3. Fine-tune Parameters: Fine-tune the parameters according to the preset rules. For example, adjust the flushing fluid flow rate from f to f × 1.05, and the file speed from r to r × 0.97. 4. Update System Settings: Update the fine-tuned parameters to the execution system to ensure that the system operates according to the new parameters.

[0263] In step S4D0, if the answer is no, the original setting is maintained unchanged.

[0264] Among them, maintaining the original settings unchanged means that when the deviation between the actual reward and the expected reward is within the preset range, the current treatment parameters and operation settings are maintained to ensure the stability and consistency of the treatment process.

[0265] Furthermore, the preset rules mentioned in step S4C0 can be executed according to the following steps:

[0266] Step 1: Perform a multi-layer decomposition of the pressure signal using a preset wavelet basis function. A preset threshold function is used to filter out high-frequency turbulence noise, retaining the primary characteristic signal. A preset state-space model is constructed, and multi-sensor data is fused using the Kalman gain. The state estimate and its covariance matrix are output as an uncertainty measure. A preset wear coefficient is defined and updated in real time based on the change in vibration frequency, dynamically modifying the efficiency reward function.

[0267] Wavelet basis functions: The basic functions used in wavelet transforms, used to decompose signals. Threshold function: A function used to filter out noise, setting a threshold to retain the main characteristic signals. State-space model: A mathematical model used to describe the dynamic behavior of a system. Kalman gain: A weighting factor used to fuse multi-sensor data. Wear coefficient: A coefficient used to dynamically modify the efficiency reward function, reflecting the degree of system wear.

[0268] The general process is described as follows:

[0269] Wavelet decomposition: Use preset wavelet basis functions (such as db4) to perform multi-layer decomposition on the pressure signal. For example, Perform 3-layer decomposition:

[0270] ;

[0271] Noise Filtering: Filter out high-frequency turbulence noise using a preset threshold function (such as a soft threshold function). For example, apply a soft threshold function to the detail coefficient D3:

[0272] ;

[0273] in, is the threshold.

[0274] State estimation: Construct a preset state space model and fuse multi-sensor data through Kalman gain. For example, the state space model is:

[0275] ;

[0276] ;

[0277] in, is the state vector, is the input vector, is the output vector, and the Kalman gain K is used to update the state estimate: ;

[0278] Dynamic correction: define preset wear coefficients , through the vibration frequency change Real-time update, dynamic correction of efficiency reward function. For example, the efficiency reward function is:

[0279] ;

[0280] in, It is the basic efficiency bonus value.

[0281] Step 2: Calculate the gradient matrix of each parameter combination for each sub-goal, quantifying the impact of parameter adjustments on safety, efficiency, quality, and adaptability. Parameter contributions are calculated using a pre-set variance decomposition method, and the optimization priority order is determined (e.g., speed > flow > motion mode). A pre-set dimensionality reduction process is performed on the gradient matrix, extracting the main gradient directions whose cumulative contribution exceeds a pre-set threshold to form an optimization search space.

[0282] Gradient matrix: A matrix used to quantify the impact of parameter adjustments on each target. Variance decomposition method: A method used to calculate parameter contributions. Optimization priority: An optimization order determined by parameter contributions. Dimensionality reduction: A method used to extract the main gradient directions.

[0283] The general process is described as follows:

[0284] Calculate the gradient matrix: Calculate the gradient of the parameter combination for each sub-goal (such as safety, efficiency, quality, and adaptability). For example, for the parameter combination , calculate the gradient matrix G:

[0285] ;

[0286] Calculate parameter contribution: Calculate the contribution of each parameter to the target using the variance decomposition method. For example, use principal component analysis (PCA) to calculate the contribution of each parameter:

[0287] ;

[0288] Determine the optimization priority order: Sort by contribution to obtain the optimization priority order, such as speed > flow > sports mode.

[0289] Dimensionality reduction: Perform dimensionality reduction on the gradient matrix and extract the main gradient directions whose cumulative contribution rate exceeds a preset threshold (such as 95%) to form an optimized search space. For example, after using PCA dimensionality reduction, retain the first two principal components to form a new gradient matrix. .

[0290] Step 3: Define a preset safe distance function to ensure that parameter adjustments meet the safety threshold constraints of claim 1. Ensure system convergence through stability parameters. Design a preset projection operator to project parameter updates into the safe action space to prevent out-of-bounds operations. Adopt a preset adaptive strategy to dynamically scale the learning rate based on the gradient norm, balancing optimization speed and stability.

[0291] The safety distance function ensures that parameter adjustments meet safety threshold constraints. The stability parameter ensures system convergence. The projection operator projects parameter updates onto the safe action space. The adaptive strategy dynamically scales the learning rate based on the gradient norm.

[0292] The general process is described as follows:

[0293] Define the safety distance function: Construct a safety distance function based on the safety threshold and ensure that the parameter adjustment meets the safety constraints. For example, the safety distance function is:

[0294] ;

[0295] in, is the current parameter, is the safety threshold.

[0296] Design a projection operator: Project the parameter update into the safe action space to prevent out-of-bounds operations. For example, the projection operator is:

[0297] ;

[0298] in, is the projection operation, It is a safe action space, is the parameter update amount.

[0299] Adopt an adaptive strategy: dynamically scale the learning rate based on the gradient norm to balance optimization speed and stability. For example, the learning rate is:

[0300] ;

[0301] in, is the base learning rate, ϵ is a small constant, is the gradient of the reward function.

[0302] Step 4: Build a pre-set Gaussian process model and output the predicted mean and variance of the adjusted parameters to quantify uncertainty. Define a pre-set risk indicator. When the risk exceeds the pre-set threshold, trigger a conservative adjustment strategy to reduce the impact of uncertainty. Add a pre-set entropy regularization term to balance strategy exploration and exploitation to avoid falling into local optima.

[0303] The Gaussian process model is a nonparametric Bayesian model used for regression and classification that provides a predicted mean and variance. The risk indicator is a metric used to quantify the risk of parameter adjustments. The conservative adjustment strategy is a cautious adjustment strategy adopted when risk is high. The entropy regularization term is a regularization term used to balance strategy exploration and exploitation.

[0304] The general process is described as follows:

[0305] 1. Construct Gaussian process model: Based on historical data, train the Gaussian process model to predict the system state after parameter adjustment. For example, for the parameter combination p, the predicted mean and variance :

[0306] ;

[0307] Here, GPR stands for Gaussian process regression.

[0308] 2. Define risk indicators: Define risk indicators based on the prediction variance. For example, the risk indicator is:

[0309] ; where ϵ is a small constant that prevents the denominator from being zero.

[0310] 3. Trigger conservative adjustment strategy: when the risk indicator exceeds the preset threshold When , the conservative adjustment strategy is triggered. For example, if , then reduce the parameter adjustment range.

[0311] Adding an entropy regularization term: Adding an entropy regularization term to the reinforcement learning algorithm balances policy exploration and exploitation. For example, the optimization objective function is: ;

[0312] in, are policy network parameters, is the expected reward, is the policy entropy, is the regularization coefficient.

[0313] Step 5: Build a pre-set index structure based on the anatomical feature vectors to quickly retrieve similar historical cases. If the preset number of adjustments fails to produce satisfactory results, the internal parameters are updated using a pre-set meta-learning algorithm to accelerate adaptation to the new environment. Define a pre-set migration coefficient. When case similarity exceeds a threshold, reuse the historically optimal parameters and weightedly integrate them with the current parameters.

[0314] The following examples include: Index structure: a data structure used to quickly retrieve similar historical cases; Meta-learning algorithm: a learning algorithm used to accelerate adaptation to new environments; and Transfer coefficient: a coefficient used to measure case similarity and control the degree of parameter fusion.

[0315] The general process is described:

[0316] Build index structure: Build a preset index structure based on the anatomical feature vector, such as using a kd tree index structure, to quickly retrieve similar historical cases. For example, for the anatomical feature vector , construct a kd tree:

[0317] ;

[0318] Retrieve similar cases: Search the index structure for historical cases similar to the current anatomical feature vector. For example, use the Euclidean distance to measure similarity and find the nearest neighbor case:

[0319] ;

[0320] Meta-learning algorithm: When the effect of continuous preset adjustments is not good, the internal parameters are updated through the meta-learning algorithm. For example, the Reptile algorithm is used to update the policy network parameters:

[0321] ;

[0322] in, are the parameters fine-tuned on the current task, is the learning rate.

[0323] Parameter reuse and fusion: defining preset migration coefficients , when the case similarity exceeds the threshold When , the historical optimal parameters are reused and weighted fused with the current parameters. For example, the parameter fusion formula is:

[0324] ;

[0325] in, is the historically optimal parameter. is the current parameter.

[0326] Step 6: Calculate the state change rate within the preset time window after parameter adjustment. If it exceeds the preset threshold, a rollback mechanism is triggered. Compare the cumulative reward curve within the preset time period with the expected curve. If the error exceeds the preset range, re-optimization is initiated. The verification results are stored in the preset knowledge graph. Graph neural networks are used to identify morphological patterns of difficult root canals and update the risk assessment model.

[0327] Among them, the rollback mechanism: a mechanism to restore the system to its previous stable state when the system state deteriorates after parameter adjustment. The cumulative reward curve: a curve that reflects the changes in the system's cumulative rewards over a period of time.

[0328] Knowledge graph: A graph structure used to store and represent knowledge, with nodes representing entities and edges representing relationships. Graph neural network (GNN): A neural network used to process graph-structured data, capable of capturing the relationships between nodes and their neighbors.

[0329] The general process is described:

[0330] Calculate the state change rate: Calculate the state change rate within the preset time window after the parameters are adjusted. For example, the state change rate is:

[0331] ;in, and They are the state after adjustment and the state before adjustment respectively.

[0332] Compare the cumulative reward curve: Compare the cumulative reward curve within a preset time period with the expected curve. For example, the error of the cumulative reward curve is:

[0333] ;

[0334] in, and are actual rewards and expected rewards respectively.

[0335] Triggering a rollback or re-optimization: If the state change rate or cumulative reward error exceeds a preset threshold, a rollback mechanism or re-optimization is triggered. For example, if the error exceeds the threshold ϵ=0.1, the parameters are re-optimized.

[0336] Update the knowledge graph: Store the verification results in the preset knowledge graph and use the graph neural network to find the morphological feature patterns of difficult-to-treat root canals. For example, use GNN to update the node features in the knowledge graph:

[0337] ;

[0338] in, is the node feature, Finally, the risk assessment model is updated through GNN to improve the adaptability and robustness of the system.

[0339] Dynamically modify safety thresholds based on treatment phase and anatomical characteristics, including:

[0340] In step S510 , based on the CBCT anatomical features and the real-time preparation progress, the treatment is divided into three stages: coronal, middle, and apical. Each stage corresponds to a different safety control priority.

[0341] CBCT anatomical features: The three-dimensional structural features of the root canal acquired through CBCT scanning, including the root canal's shape, position, curvature, etc. Safety control priority: The control priority level set according to the treatment stage and anatomical features is used to dynamically adjust the safety threshold.

[0342] The general process is as follows: 1. Treatment stage division: Based on CBCT anatomical features and real-time preparation progress, the treatment is divided into three stages: coronal, middle, and apical. For example, it is divided according to one-third of the root canal length. 2. Priority setting: Different safety control priorities are assigned to each stage. For example, temperature deviation is controlled first in the coronal stage, and temperature deviation is controlled first in the middle stage.

[0343] Balance temperature and pressure control in the periapical phase, and prioritize pressure deviation control in the apical phase. 3. Dynamic Adjustment: During treatment, safety control priorities are dynamically adjusted based on the current phase and real-time preparation progress. For example, a rule from the rule library can be used: "When entering the apical phase, increase the priority of pressure control."

[0344] Step S520: Extract root canal curvature, calcification rate, and diameter variation coefficient through graph neural network to construct an anatomical risk index. , quantify the operational risk level, where is the weight coefficient.

[0345] Among them, Graph Neural Network (GNN): a neural network used to process graph structured data, which can capture the relationship between nodes and their neighbors.

[0346] Anatomical risk index: An indicator used to quantify the level of procedural risk, taking into account the anatomical characteristics of the root canal.

[0347] Step S530 : Initializing the safety threshold through the fuzzy rule base according to the stage and the risk index.

[0348] The fuzzy rule base is a set of rules based on expert experience and fuzzy logic, used to convert fuzzy inputs into specific outputs. Safety thresholds define the safe range of operating parameters (such as pressure and temperature) to ensure the safety of the treatment process.

[0349] The general process is described as follows:

[0350] 1. Determine fuzzy rules: Based on the treatment stage and anatomical risk index, formulate fuzzy rules. For example, “If the treatment stage is apical and the anatomical risk index is high, then reduce the pressure threshold by 10%.”

[0351] 2. Initialize the safety threshold: Initialize the safety threshold based on the fuzzy rule base, combined with the current treatment stage and anatomical risk index. For example, if the basic pressure threshold is 80kPa, the pressure threshold adjusted according to the rule is:

[0352] ;

[0353] The basic temperature threshold is 45°C. The temperature threshold adjusted according to the rules is:

[0354] .

[0355] In step S540 , in combination with the multi-objective reward function of reinforcement learning, when the pressure / temperature approaches the threshold and the reward value decreases, the threshold range is dynamically contracted by controlling the preset barrier function.

[0356] The barrier function is a function used to dynamically adjust the safety threshold range to ensure that the operating parameters are within the safe range. The multi-objective reward function of reinforcement learning is a function used to evaluate the comprehensive performance of operating parameters, combining multiple objectives (such as cleaning effect, safe operation, and preparation efficiency).

[0357] The general process is as follows: 1. Monitoring pressure / temperature: Real-time monitoring of pressure and temperature. When they approach the threshold and the reward value decreases, the barrier function is triggered. 2. Dynamically shrinking threshold: Dynamically adjust the safety threshold range through a preset barrier function. For example, using an exponential barrier function: ;in, is the adjustment coefficient, is the amount by which the reward value decreases.

[0358] Update safety thresholds: Update safety control parameters based on the adjusted thresholds to ensure that operations are performed within the new safety range.

[0359] Based on the same inventive concept, an embodiment of the present invention provides an intelligent control system for an internally cooled wet-prepared nickel-titanium root canal treatment file, comprising a memory and a processor, wherein the memory stores a program that can be executed on the processor to implement the following Figure 1 program.

[0360] The embodiments of this specific implementation method are all preferred embodiments of the present application and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. An intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file, characterized in that: include: Obtain root canal 3D structural parameters through CBCT scanning, import them into the control system, initialize multiple sensors, and perform self-test and zero-point calibration; The system simultaneously collects raw data on root canal pressure, temperature, and rotation speed, integrates CBCT anatomical features to build a normalized model, calculates pressure and temperature deviations, and obtains cleaning effect indicators through irrigant reflux turbidity detection and pressure fluctuation analysis. Taking pressure deviation, temperature deviation and rotation speed as input, the flushing fluid flow and file rotation speed are dynamically adjusted through the fuzzy rule base to achieve preliminary coupled control of cutting resistance and temperature. A deep reinforcement learning model was constructed, with pressure, temperature, cleaning effect, and anatomical features as the state space, and file speed, motion pattern, and irrigation fluid flow rate as the action space. The root canal preparation strategy was iteratively optimized using a multi-objective reward function to generate a parameter combination containing the optimal speed, motion pattern, and irrigation fluid flow rate. This parameter combination was then output to the control system to drive the nickel-titanium root canal treatment file to perform the corresponding operation. The motion modes included continuous rotation, reciprocating rotation, and pulsed rotation. The safety threshold is dynamically modified according to the treatment stage and anatomical characteristics. When the parameters deviate from the threshold, an early warning is triggered and compensation is performed. At the same time, the control effect is fed back to the reinforcement learning model for strategy update.

2. The intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to claim 1, characterized in that: The original data of pressure, temperature, and speed in the root canal are collected simultaneously, and the normalized model is constructed by integrating the CBCT anatomical features. The pressure deviation and temperature deviation are calculated, including: Synchronously collect the raw data of pressure, temperature and speed in the root canal, and use hardware clock synchronization technology and cubic spline interpolation algorithm to achieve spatiotemporal registration and preprocessing of multi-source data; Root canal anatomical features are extracted from CBCT images using 3D-CNN. Sensor temporal features are enhanced using a bidirectional LSTM combined with an attention mechanism. Cross-modal feature adaptive fusion is achieved through a gated fusion unit. A priori probability model of root canal pressure-temperature distribution is established based on the Gaussian mixture model. The real-time collected pressure and temperature data are combined with the priori model using the preset Bayesian update formula to obtain the posterior probability distribution. The fused features are normalized through nonlinear mapping functions to calculate pressure and temperature deviations; The covariance matrix of process noise and observation noise of Kalman filter is dynamically adjusted according to the fusion features, and the extended state space model is constructed by taking anatomical features as additional state variables.

3. The intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to claim 2, characterized in that: The root canal anatomical features in CBCT images are extracted through 3D-CNN, and the sensor temporal features are enhanced using a bidirectional LSTM combined with an attention mechanism. The gated fusion unit is used to achieve cross-modal feature adaptive fusion, including: Multi-scale feature extraction is performed on CBCT images using a 3D convolutional neural network. A 3D convolutional attention module combined with spatial pyramid pooling is used to perform semantic-level localization of apical stenosis, root canal bifurcation, root canal curvature, and calcified areas, generating a 3D anatomical feature map containing spatial position weights. A bidirectional long short-term memory network is used to capture the time series features of pressure and temperature sensor data. A temporal self-attention mechanism is introduced, combined with treatment stage embedding vectors. By calculating the attention weights of query-key-value triples, the importance of temporal features at different treatment stages is dynamically adjusted to generate temporal enhancement features. The embedding vectors include crown preparation, mid-root preparation, and apical preparation. A dynamic routing feature interaction network based on capsule networks is constructed. A joint representation matrix of anatomical and temporal features is generated through bilinear pooling. A cross-attention mechanism is used to establish bidirectional semantic associations between cross-modal features. The feature interaction weights are optimized through an iterative routing algorithm to generate cross-modal interaction features. The root canal curvature, root canal calcification rate, and root canal diameter change rate parameters are calculated based on three-dimensional anatomical features. The complexity index is generated through the anatomical complexity assessment module. This index is used as the input of the gated fusion unit. The complexity index is mapped to the [0, 1] interval through the Sigmoid activation function to generate a fusion weight coefficient and construct a fusion function. The specific fusion function is as follows: ； in, The final output fusion feature contains information about anatomical structure and sensor data. is the Sigmoid activation function, is the anatomical complexity index, are learnable weights and biases, For anatomical features, is the sensor characteristic, This is element-wise multiplication.

4. The intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to claim 3, characterized in that: Taking pressure deviation, temperature deviation and rotation speed as input, the flushing fluid flow and file speed are dynamically adjusted through the fuzzy rule base to achieve preliminary coupled control of cutting resistance and temperature. The Kullback-Leibler divergence between the real-time pressure and the prior distribution is calculated, and a two-dimensional feature vector is constructed based on the temperature change rate. The file speed is mapped to the preset fuzzy state space to generate a membership function matrix. Based on the three-dimensional anatomical feature map and anatomical complexity index, a preset morphologically specific fuzzy rule subset is matched. The rule weights are modulated time-varyingly in combination with the treatment stage recognition results to form a dual-dimensional decision logic based on anatomical features and treatment stage. A Takagi-Sugeno fuzzy inference system is used to process feature vectors, a neural network is used to output the control function expression, and a preset meta-learning algorithm is used to optimize the fuzzy rule parameters to improve the system's adaptability to new root canal morphologies. The flushing fluid flow adjustment coefficient and the file speed correction coefficient are mapped to the preset safety threshold space to form the final control instruction; the H∞ control theory is introduced to design an anti-interference compensator, which automatically corrects the control instruction when the external environmental parameters mutate.

5. The intelligent control method for an internally cooled wet preparation nickel-titanium root canal treatment file according to any one of claims 1 to 4, characterized in that: Generating a parameter combination including the optimal speed, motion pattern, and flushing fluid flow rate includes the following steps: Perform wavelet transform on real-time pressure and temperature data to extract time domain and frequency domain features; CBCT images are processed through graph neural networks to generate anatomical feature vectors including curvature and calcification rate; Use gated recurrent units to fuse temporal sensor data with spatial anatomical features to form a spatiotemporal state representation; Dynamically generate motion constraints based on the curvature and calcification rate in the anatomical feature vector to construct a safe motion space; The meta-policy network selects basic parameter combinations from a preset policy library based on anatomical feature vectors. The execution policy network takes the spatiotemporal state representation as input, calculates parameter offsets through the actor-critic architecture, and optimizes the basic parameter combination to form preliminary parameters. The preliminary parameter combination is projected into the safety constraint space, and the compensation amount of external disturbances such as tissue resistance fluctuation is calculated by combining H∞ control theory to generate a robust parameter combination. Calculate the expected value R of the robust parameter combination under the multi-objective reward function; If the expected value R exceeds the preset threshold, the parameter combination output value is executed by the system; If the expected value R does not reach the preset threshold, it jumps to the meta-learning mechanism to adjust the policy network parameters and regenerate the parameters.

6. The intelligent control method for an internally cooled wet preparation nickel-titanium root canal treatment file according to claim 5, characterized in that: The steps after executing the system for combining the parameters and outputting the values ​​are also included, as follows: Start the online verification mechanism and monitor the deviation between actual rewards and expected rewards in real time through the Kalman filter; Analyze whether the deviation exceeds the preset deviation; If yes, the adaptive calibration module is started to fine-tune the parameters based on real-time data using preset rules. If not, keep the original settings unchanged.

7. The intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to claim 1, characterized in that: Dynamically modify safety thresholds based on treatment phase and anatomical characteristics, including: Based on CBCT anatomical features and real-time preparation progress, treatment is divided into three stages: coronal, mid-root, and apical. Each stage corresponds to a different safety control priority. The graph neural network is used to extract root canal curvature, calcification rate, and diameter variation coefficient to construct an anatomical risk index, as follows: , quantify the operational risk level, where is the weight coefficient; Initialize the safety threshold through the fuzzy rule base according to the stage and risk index; Combined with the multi-objective reward function of reinforcement learning, when the pressure / temperature approaches the threshold and the reward value decreases, the threshold range is dynamically contracted by controlling the preset barrier function.

8. An intelligent control system for an internally cooled wet-prepared nickel-titanium root canal treatment file, characterized in that: The invention comprises a memory, a processor, and a program stored in the memory and executable on the processor, wherein the program can be loaded and executed by the processor to implement an intelligent control method for an internally cooled wet-prepared nickel-titanium root canal treatment file according to any one of claims 1 to 7.

Images