Multi-modal sensing-based equipment adaptive regulation and control method and system

Abstract

The invention discloses an equipment self-adaptive regulation and control method and system based on multi-mode perception, and relates to the technical field of equipment intelligent control. The method comprises the following steps: acquiring real-time operation parameters of medical equipment, and generating multi-modal equipment state data; performing depth state estimation based on the data and a pre-stored historical operation database, identifying a current equipment operation mode and a performance degradation trend, and generating a dynamic target set point interval; matching the dynamic target set point interval with a preset control strategy library to generate a self-adaptive control instruction set for driving an actuator group so as to achieve a target parameter; issuing an instruction set to drive an actuator group to execute regulation and control actions, and continuously collecting operation feedback data; and the feedback data is compared with the dynamic target set point interval to obtain the operation deviation, and the control instruction set is dynamically optimized by adopting a control algorithm based on the deviation, so that the self-adaptive monitoring and cooperative control of the operation state of the medical equipment are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment control technology, specifically to an adaptive control method and system for equipment based on multimodal perception. Background Technology

[0002] Intelligent control of medical devices based on multimodal sensing is an important development direction for improving the automation and safety of medical devices. It aims to achieve stable and efficient operation of equipment under complex conditions by monitoring the operating status of the equipment in real time and dynamically adjusting control parameters. In recent years, with the advancement of sensing technology and data analysis algorithms, the control strategies for medical devices are evolving from fixed thresholds and single-parameter judgments towards dynamic and adaptive approaches.

[0003] Existing technologies monitor key operating parameters of medical equipment, such as motor current or core temperature, and trigger alarms or perform protective operations like shutdown when these parameters exceed preset safety thresholds. These methods provide basic safety protection and have a certain degree of reliability in typical scenarios where equipment operation is relatively stable and load changes are gradual. However, for precision equipment operating in complex medical scenarios such as surgery with high intensity and dynamic changes, the deterioration of its operating condition is often a gradual process involving the coupling of multiple physical quantities. Existing technologies struggle to provide early warnings of equipment failures or performance degradation, and cannot offer proactive parameter adaptive adjustments when equipment operating modes switch (e.g., from no-load to high-speed operation). These technological bottlenecks are particularly prominent in the intelligent control scenarios of medical and nursing equipment. For example, the operating state of an infusion pump is not only affected by motor current but also closely related to pipeline vibration, blockage, and ambient temperature. Traditional single-variable control strategies cannot fully perceive these multimodal state changes, leading to slow response times, insufficient precision, and even potential safety risks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a device adaptive control method and system based on multimodal sensing.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention discloses a device adaptive control method based on multimodal sensing, which is used to monitor and coordinate the operating status of medical devices, including the following steps:

[0007] The real-time operating parameters of the medical equipment are acquired to generate multimodal equipment status data, which includes at least current, vibration, noise, and temperature signals.

[0008] Based on the multimodal device status data, by performing deep status estimation with the pre-stored historical operation database, the current device operation mode and performance degradation trend are identified, and a dynamic target setpoint interval is generated;

[0009] The dynamic target setpoint range is matched with a preset control strategy library to generate a corresponding adaptive control instruction set; wherein, the adaptive control instruction set is used to drive the actuator group to reach the target parameters;

[0010] Based on the equipment's operational safety priority, the adaptive control instruction set is sent to the actuator group to drive the actuator group to perform control actions and continuously collect operational feedback data of the medical equipment;

[0011] The operational feedback data is compared with the dynamic target setpoint range to obtain the equipment operational deviation, and based on the operational deviation, the adaptive control instruction set is dynamically optimized using an adaptive control algorithm.

[0012] Secondly, the present invention discloses a device adaptive control system based on multimodal sensing, which uses the above-mentioned device adaptive control method based on multimodal sensing, including:

[0013] The multimodal sensing module is used to collect real-time operating parameters of medical equipment and generate multimodal equipment status data.

[0014] The status recognition module is used to identify the current operating mode and performance degradation trend of the device based on the multimodal device status data, by performing deep status estimation with a pre-stored historical operation database, and to generate a dynamic target setpoint range.

[0015] The control instruction generation module is used to match the dynamic target setpoint range with a preset control strategy library to generate a corresponding adaptive control instruction set.

[0016] The instruction issuance and execution module is used to issue the adaptive control instruction set to the actuator group, drive the actuator group to perform control actions, and continuously collect the operation feedback data of the medical device;

[0017] The operation feedback optimization module is used to compare the operation feedback data with the dynamic target setpoint range to obtain the equipment operation deviation, and to dynamically optimize the adaptive control instruction set based on the operation deviation using a control algorithm.

[0018] An actuator group is used to receive the optimized adaptive control instruction set and drive one or more actuators to adjust the operating parameters of the medical device.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. This invention achieves comprehensive perception and in-depth monitoring of the operating status of medical devices by acquiring multimodal real-time operating parameters. Based on the fusion analysis of multi-dimensional data, it can more accurately capture subtle changes in the device's operating conditions, overcoming the limitations of monitoring from a single data source, thus providing a more reliable and comprehensive data foundation for subsequent control decisions.

[0021] 2. By matching real-time data with a pre-stored historical operating database, this invention can identify the equipment operating mode and generate a dynamic target setpoint range. It can dynamically adjust the control target according to the actual state of the equipment, avoiding the rigidity problem of fixed threshold control strategies and improving the adaptability and reliability of the system under different operating conditions.

[0022] 3. This invention can continuously correct control commands to ensure that the operating parameters of the equipment are stably close to the dynamic target setpoint range, thereby effectively improving the accuracy, stability and safety of medical equipment operation, while enhancing the system's self-optimization and anti-interference capabilities. Attached Figure Description

[0023] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0024] Figure 1 This is a flowchart of the steps of the present invention;

[0025] Figure 2 This is a flowchart illustrating the working principle of the present invention;

[0026] Figure 3 This is a system module architecture diagram of the present invention. Detailed Implementation

[0027] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0028] In existing technologies, the monitoring of medical equipment operation status largely relies on parameter threshold alarms or predictive maintenance using fixed models, making it difficult to accurately capture and adaptively compensate for early signs of equipment performance degradation. Traditional methods fail to dynamically adjust control models when equipment faces complex and variable operating conditions or when components experience slow performance decline, leading to decreased control accuracy and lag. Existing systems lack in-depth fusion analysis of multimodal operating parameters such as current, vibration, noise, and temperature. Especially when equipment loads change abruptly or mechanical wear intensifies, single-parameter control strategies become insufficiently adaptable, failing to meet the demands of high-reliability medical scenarios.

[0029] To address the aforementioned issues, this study discovered a deep coupling relationship between equipment performance degradation trends and the changing patterns of multimodal operating parameters. A forward-looking control mechanism based on deep state estimation and a dynamic target setpoint interval generation mechanism can be established. The research revealed that Long Short-Term Memory (LSTM) networks can effectively learn the temporal dependencies of multimodal parameters and accurately estimate equipment health status; while reinforcement learning algorithms possess the ability to optimize strategies online in uncertain environments. Therefore, this study proposes combining deep state estimation with adaptive control algorithms and introducing safety boundary constraints based on Lyapunov stability to form a closed-loop control architecture that balances intelligence and safety. Further experimental verification demonstrates that the mapping relationship between equipment operating deviations and the contribution of multimodal parameters is incorporated into the online optimization mechanism of the control law, forming a dynamically adjusted closed loop.

[0030] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Example 1:

[0032] like Figure 1 As shown, the device adaptive control method based on multimodal sensing includes the following steps:

[0033] Acquire real-time operating parameters of medical equipment and generate multimodal equipment status data, which includes at least current, vibration, noise and temperature signals;

[0034] Based on multimodal equipment status data, the system performs deep status estimation with a pre-stored historical operation database to identify the current equipment operating mode and performance degradation trend, and generates a dynamic target setpoint range.

[0035] The dynamic target setpoint range is matched with the preset control strategy library to generate the corresponding adaptive control instruction set; the adaptive control instruction set is used to drive the actuator group to achieve the target parameters.

[0036] Based on the equipment's operational safety priority, the adaptive control instruction set is sent to the actuator group to drive the actuator group to perform control actions and continuously collect operational feedback data from the medical equipment.

[0037] The operational feedback data is compared with the dynamic target setpoint range to obtain the equipment operation deviation. Based on the operation deviation, the adaptive control instruction set is dynamically optimized using an adaptive control algorithm.

[0038] like Figure 2 As shown, the working principle of this application is as follows: by integrating multiple sensors (such as current transformers, vibration accelerometers, acoustic microphones, and temperature sensors) deployed on key components of the equipment, multimodal raw signals that comprehensively reflect the mechanical, electrical, and thermodynamic states of the equipment are collected synchronously and in real time, including current, vibration, noise, and temperature. After preprocessing such as filtering, amplification, and analog-to-digital conversion, the raw data is integrated into structured multimodal equipment status data.

[0039] The preprocessed multimodal equipment status data is compared and analyzed with a pre-stored historical operation database, which contains historical operational data characteristics of the equipment under normal, various abnormal, and different degradation stages. By executing a specific estimation algorithm, key features characterizing the current health status of the equipment are extracted from the multimodal equipment status data. This allows for a comprehensive judgment of the equipment's current operating mode (e.g., no-load, full-load, abnormal disturbance, etc.) and an estimation of its performance degradation trend over time. Based on this judgment, the system dynamically calculates and generates a suitable dynamic target setpoint range. This dynamic target setpoint range defines the range of safe and efficient operating parameters that the equipment needs to maintain, rather than fixed values, thus providing flexibility for control.

[0040] The generated dynamic target setpoint range is sent to a pre-defined control strategy library for matching and querying. This library contains a knowledge base of various control logics and rules, with each strategy corresponding to one or more device states and operational goals. The system uses a matching algorithm (such as similarity calculation) to find the most suitable control strategy from the library for the current state and the dynamic target setpoint range, and interprets it into a series of specific, executable adaptive control instructions. These instructions specify which actuators (such as motors, valves, heaters, etc.) need to be driven and the specific target parameters of the actions, guiding the equipment to enter and stabilize within the dynamic target setpoint range.

[0041] Before issuing instructions, the system performs a risk assessment and prioritizes the instruction set according to a set of built-in device operation safety priority rules. For high-priority, critical instructions that concern device or patient safety, the system immediately issues them to the actuator group through a high-priority communication channel; for lower-priority, optimized instructions, they can be temporarily stored in a queue awaiting confirmation or sent in sequence.

[0042] Upon receiving a command, the actuator assembly drives the corresponding mechanical or electrical components to perform control actions (such as adjusting speed, opening degree, or power). Simultaneously, the system continuously collects operational feedback data from the equipment in parallel, forming the basis for closed-loop control.

[0043] Finally, the collected operational feedback data is compared in real time with the dynamic target setpoint range to calculate the deviation between the current operating state and the desired state. This deviation value is input into an adaptive control algorithm (such as an adaptive PID controller). The adaptive control algorithm can dynamically adjust its internal control parameters or logic based on the magnitude, direction, and trend of the deviation, and accordingly optimize and correct the previously generated adaptive control instruction set online. This allows the control instructions to better adapt to the dynamic changes in the equipment state, ultimately causing the equipment operation to converge within the dynamic target setpoint range.

[0044] This application improves the comprehensiveness of equipment condition monitoring and early fault warning capabilities, enhances the system's adaptability and robustness in the face of different operating conditions and equipment performance degradation, ensures the safety and accuracy of the control process, and ultimately improves the reliability, efficiency and safety of medical equipment operation while reducing the manual burden of medical staff on equipment condition monitoring.

[0045] This application further proposes that the steps for depth state estimation based on multimodal device state data include:

[0046] Real-time data window for synchronously acquiring multimodal device status data Short-term historical time window and long-term historical time window The system captures the dynamic evolution of equipment status by simultaneously opening three time windows of different scales: real-time window... Reflects the current instantaneous device status; short-term historical window Used to observe recent trends; long-term historical window Used to identify patterns associated with long-term, slow performance degradation.

[0047] Based on a preset feature extraction algorithm (e.g., calculating the root mean square value of vibration signals, the spectral characteristics of noise signals, the effective value and harmonic components of current, the slope of temperature change, etc.), the feature vectors of the data in each time window are calculated respectively.

[0048] Among them, vibration signals: the root mean square (RMS), kurtosis, and peak factor in the time domain are calculated based on mutual information or Pearson correlation coefficient; the frequency domain amplitudes of the first, second, and third harmonics are extracted by FFT.

[0049] Current signal: Calculate the effective value, total harmonic distortion, and fundamental amplitude;

[0050] Noise signal: Calculate sound pressure level and frequency band energy (125Hz~8kHz divided into 8 segments);

[0051] Temperature signal: Calculate the current value, moving average (window 10s), and rate of change (°C / min).

[0052] Finally, a 20-dimensional feature vector is generated (vibration 6-dimensional, current 4-dimensional, noise 8-dimensional, temperature 2-dimensional).

[0053] The feature vectors are input into a trained Long Short-Term Memory (LSTM) network model, which outputs a deep state estimate representing the device's health status. and the rate of change of the state corresponding to the depth state estimate Depth state estimate It is a comprehensive scalar or vector that precisely quantifies the device's current "health score"; corresponding to the rate of change of the depth state estimate. It reveals whether the health status of the equipment is improving, stabilizing, or deteriorating, and the extent of its changes.

[0054] The Long Short-Term Memory (LSTM) network model is a three-layer stacked LSTM structure, with each layer containing 128 neurons, followed by a fully connected layer that outputs a deep state estimate. and the rate of change of the corresponding depth state estimate The training data comes from over 10,000 hours of multimodal equipment status data in the historical operation database, covering various operating conditions such as normal, abnormal, and fault. The input is for each time window (…). The 20-dimensional feature vector extracted from within is output. The depth state estimate is between 0 and 1. The rate of change of the corresponding depth state estimate (depth state estimate) For a dimensionless quantity, the rate of change The unit is min -1 (or / min), such as This indicates that the health score decreases by 1% per minute. LSTM networks, as state observers in soft measurement, enable indirect and accurate estimation of the internal health status of devices.

[0055] The training data for the Long Short-Term Memory (LSTM) network model comes from a pre-stored historical operating database. This database should contain time-series data collected under normal operation, various known abnormal modes (such as bearing wear, blade imbalance, and coil insulation aging), and different performance degradation stages. It should contain at least 10,000 hours of valid operating data.

[0056] Training samples are constructed from the original time-series data using a sliding window method. Each sample contains data at continuous time steps, and the window length should cover long-term historical windows. (e.g., 1 hour). The feature vector for each time step is obtained by feature extraction from multimodal equipment status data (current, vibration, noise, temperature) (e.g., a 20-dimensional feature vector). The sample label y is the actual health status label of the equipment at the end of the corresponding window (determined by post-maintenance records or experts based on operating parameters, e.g., normal = 1.0, slight degradation = 0.7, severe failure = 0.3).

[0057] Z-score standardization (i.e., subtracting the mean and dividing by the standard deviation) is applied to each feature dimension to make its mean 0 and standard deviation 1, thereby accelerating model convergence. The constructed sample set is randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0058] Mean Squared Error (MSE) is used as the loss function L to measure the health status estimate predicted by the model. Differences from the true label y:

[0059] ;

[0060] Where N is the number of samples in the batch.

[0061] Update network weights using the Adam optimizer (Adaptive Moment Estimation), with an initial learning rate set to 1. e-4 During training, a learning rate decay strategy is employed. For example, when the validation set loss no longer decreases within 5 consecutive epochs, the learning rate is halved to fine-tune the model parameters.

[0062] Training batches are set according to computing resources, typically 32, 64, or 128, with a maximum training epoch of 100, and an early stopping strategy is employed. Training is terminated early when the validation set loss no longer decreases within 10 consecutive epochs, and the model parameters that minimize the validation set loss are restored to prevent overfitting.

[0063] After training, evaluate the model's final performance on an independent test set (e.g., MSE should be below 0.01). The rate of change of the corresponding deep state estimates after model training is complete. By calculating the depth state estimate at two consecutive time steps Obtained by the difference:

[0064]

[0065] Where Δt is the data sampling interval, t is the current time step, and t-1 is the previous time step.

[0066] Based on depth state estimation and rate of change of state Calculate the performance degradation gradient of computing devices The formula is as follows:

[0067] ;

[0068] in, and The preset weighting coefficients, and ,For example = 0.7, = 0.3, used to balance the contributions of the current absolute level of health status and the trend of deterioration, respectively. Characterizing the current health status and ideal state ( The greater the gap, the greater the degradation gradient. Ensure that only when the rate of change of state It only contributes positively to the gradient when the value is negative (i.e., the health status deteriorates). If the result is positive (state improvement), then this contribution is zero.

[0069] Therefore, positive A negative value indicates that the device's performance is deteriorating rapidly, and its magnitude indicates the severity of the degradation rate; a negative or zero value indicates that the device is stable or recovering.

[0070] Depth state estimate and performance degradation gradient By comparing the modal features with those in the historical operating database, the current operating mode and performance degradation trend of the equipment can be identified.

[0071] The modal features in the historical operation database specifically refer to the feature templates of known operating modes (such as "normal operation", "minor wear", "early bearing failure" etc.). Through this pattern recognition, the system can accurately determine which operating mode the equipment is currently in and qualitatively describe its performance degradation trend (such as "slow linear degradation" or "accelerated degradation").

[0072] Through the above technical solutions, this application enhances the depth and foresight of state estimation. By providing data-driven and quantitative insights, it ensures that the system can maintain high robustness and reliability even when facing slow performance degradation or sudden modal switching, thereby fundamentally improving the intelligence level and safety assurance capabilities of medical equipment management.

[0073] This application further proposes that the steps of identifying the current equipment operating mode and performance degradation trend, and generating a dynamic target setpoint range, also include:

[0074] The system obtains the security status parameter vector of external objects interacting with the device through a dedicated data interface. Safety state parameter vector It contains key physiological status information, such as blood oxygen saturation (SpO2), heart rate (HR), and respiratory rate (RR) provided by the monitor. Simultaneously, the system transmits its own multimodal device status data vector. (Including current, vibration, etc.) and The data is then integrated to form a comprehensive vector that reflects the overall state of the "human-machine system".

[0075] Based on the aforementioned comprehensive state vector, a dynamic safety boundary function is constructed according to predefined safety constraint rules. When the value of the dynamic safety boundary function is less than or equal to zero, it indicates that the system is within the safety domain; where This is a multimodal device state data vector, where t is the acquisition time;

[0076] Safety constraints are specified as a series of threshold inequalities based on clinical safety standards and equipment specifications. For example, patient oxygen saturation (SpO2) must not be lower than 90%, equipment motor temperature must not exceed 45°C, and vibration RMS value must not exceed 2.5g. Safety boundary functions. The value is determined by the most severe overlimit condition, i.e. Once the safety boundary function is established... If the value is greater than 0, the safety constraint correction mechanism will be triggered immediately.

[0077] To ensure the stability and convergence of the safety boundary itself, the safety boundary function is updated using an online learning algorithm based on Lyapunov stability.

[0078] If the control instruction set causes the safety boundary function value to be greater than 0, an alternative target setpoint interval is generated for safety constraint correction.

[0079] Define a Lyapunov function using an online learning algorithm. The function is positive definite and its time derivative is negative definite. By monitoring the system's state trajectory in real time, the algorithm continuously adjusts the safety boundary function. The parameters are set to ensure that, starting from any state point, the system's control actions are constrained within the safety boundary and eventually converge stably to the desired safe equilibrium point.

[0080] Lyapunov functions can be designed as quadratic forms. , where the state deviation vector , and These represent the desired balance points between patient safety and equipment status, respectively. The initial value of the positive definite matrix P can be set as the identity matrix and updated using an online gradient descent algorithm:

[0081] ;

[0082] in, For learning rate (e.g.) =0.01), for The gradient with respect to P. The update rule is expressed as: finding the parameter e such that... It always holds true near the boundary, among which... The rate of change of the system state is predicted by the equipment dynamics model.

[0083] Before issuing the instruction, the system will use the initially generated adaptive control instruction set as input to predictively simulate the new state of the "human-machine system" after executing the instruction. And calculate the new state. Safety boundary function The function value. When the function value is predicted. If the value is >0, meaning the control command will cause the system to exceed the safety boundary, the command set will be immediately intercepted, and a safety constraint correction will be triggered. Based on the gradient information of the safety boundary function, an alternative, conservative target setpoint range will be generated, and the control strategy will be rematched. This will generate a set of alternative control commands that improve safety, ensuring that any adjustment action of the device is based on the absolute premise of not harming the safety of the patient and the device itself.

[0084] Simulation verification of the strategy self-evolution was achieved through a high-fidelity digital twin environment. This environment, built on MATLAB / Simulink and Simscape Multibody, includes the following modules:

[0085] 1. Equipment physical model: Import the CAD model of the medical equipment and generate a multibody dynamics model (such as motor, gearbox, fluid pipeline).

[0086] 2. Sensor Model: Simulates the output of current, vibration, noise, and temperature sensors (with white noise and drift error added);

[0087] 3. Patient physiological model: Integrates the HumMod physiological model to simulate the response of parameters such as blood oxygenation and heart rate to device operation;

[0088] 4. Security Boundary Verification Module: Real-time calculation of security boundary functions. To ensure the policy satisfies the safety boundary function ≤ 0, the agent explores the environment for 10,000 rounds during training, with each round containing 1,000 decision steps. The reward function r is designed to integrate device efficiency (the reciprocal of the target tracking error) and energy consumption (current integral).

[0089] This application elevates safety control from a passive "post-event alarm" to an active "pre-event prevention," intercepting potentially dangerous commands through a priori simulation to prevent problems before they occur. It constructs a closed-loop control system by strongly coupling the patient's physiological state with the equipment's operating state, mathematically guaranteeing the system's stability and convergence under safety boundary constraints and avoiding oscillations or divergences in the control process.

[0090] This application further proposes to match the dynamic target setpoint range with a preset control strategy library to generate a corresponding adaptive control instruction set. The control strategy library includes a self-evolution mechanism, which specifically includes:

[0091] Set a matching threshold θ (range 0.7~0.9, for example 0.8);

[0092] When the dynamic target setpoint interval matches the control strategy library using the cosine similarity calculation When the score is below the matching threshold θ, the control strategy library is triggered to enter the strategy self-evolution process.

[0093] Specifically, the system represents the "dynamic target setpoint interval" that needs to be achieved as a feature vector, and calculates the matching degree by performing cosine similarity calculation with the feature vector of the dynamic target setpoint interval corresponding to each control strategy in the control strategy library. Its formula is:

[0094]

[0095] Where A represents the feature vector of the current dynamic target setpoint interval, B represents the feature vector of the dynamic target setpoint interval corresponding to the existing control strategies in the control strategy library, and Mc's value range is [-1, 1], the closer to 1, the more similar. After traversing the strategy library, if the highest matching degree max( of all existing strategies with the current target) is... If the value is still below the threshold θ, it is determined that there is no directly effective strategy in the current library, and the strategy self-evolution process is immediately triggered.

[0096] Based on multimodal device status data (including device status and patient physiological status) and historical operation data, new control strategies are generated online using reinforcement learning algorithms (such as TD3 or DDPG). New control strategies For optimization under new operating conditions, the current system's multimodal state data is used as the state s in reinforcement learning, and possible combinations of control commands are used as actions a. A new control strategy is proposed. Predicting the long-term cumulative reward that can be obtained by performing action a in a specific state s can clearly indicate which optimal action should be taken in which state.

[0097] New control strategies Simulation verification and safety checks are performed in a high-fidelity digital twin simulation environment. The predicted control trajectory is substituted into the dynamic safety boundary function Ψ for verification, ensuring that the entire control process does not exceed safety constraints. After the new strategy is validated through effectiveness simulation and safety boundary checks, it is updated to the control strategy library, completing the update and evolution. Before the strategy self-evolution process is complete, the execution of currently low-matching instructions is temporarily suspended, and a backup safety strategy is activated.

[0098] This application addresses the bottleneck problem of strategy failure in traditional control systems when faced with new or rare operating conditions. By using cosine similarity judgment and online generation through reinforcement learning, it ensures that the system creates feasible strategies. Through simulation and safety double-layer verification, it guarantees the reliability and safety of the system evolution process, avoids the risks introduced by blind learning, and can continuously accumulate experience and optimize performance with the long-term operation of the equipment, providing adaptive control services throughout the entire life cycle of medical equipment.

[0099] This application further proposes that the reinforcement learning algorithm uses the dual-delay deep deterministic policy gradient algorithm (TD3 algorithm) to update the new control policy. New control strategy The objective function is as follows:

[0100] ;

[0101] in, This indicates the control policy before the update;

[0102] s and These represent the current device state and the next device state, respectively, which are characterized by the fusion of multimodal device state data (current, vibration, noise, temperature) and patient physiological data.

[0103] 'a' represents the control action selected from the preset action space (such as "increase the pump speed by 5%)).

[0104] r represents the instant reward derived from a comprehensive evaluation of equipment operating efficiency and energy consumption. This reward is derived from a comprehensive evaluation of multiple performance indicators, such as equipment operating efficiency (e.g., target achievement speed, control accuracy) and system energy consumption (e.g., motor power consumption, heat generation), which guides the algorithm to learn high-performance, low-energy-consumption control strategies.

[0105] η is the learning rate, which controls the speed at which new information is incorporated into the strategy and is a key hyperparameter for ensuring a smooth learning process.

[0106] This is a discount factor with a value less than 1 to ensure the convergence of cumulative rewards; for example, a value of 0.99.

[0107] By using two independent target criticism networks and The minimum of the two values ​​is used to update the value function, which effectively overcomes the overestimation bias of the value function and makes the learning more stable.

[0108] ϵ represents the exploration noise, which follows a Gaussian distribution N(0,0.1) to prevent the policy from overfitting to a local optimum.

[0109] Criticizing the internet in the name of the target The target policy network is updated every 2 steps, and its parameters are updated slowly, which further increases the stability of the learning process.

[0110] By repeatedly executing the above steps, a new control strategy is developed. Through continuous optimization, it eventually converges to a high-performance strategy that can provide the highest long-term cumulative reward (i.e., optimally balance efficiency and energy consumption) for a specific device state s.

[0111] This application solves the stability problem of policy learning in a continuous, high-dimensional control action space by introducing the TD3 algorithm, ensuring that the learning process can converge reliably and avoiding training collapse. It transforms complex engineering goals (such as efficiency and energy consumption) into quantitative indicators that the algorithm can directly optimize, enabling the system to automatically discover and execute fine control strategies.

[0112] This application further proposes that, based on the equipment's operational safety priority, an adaptive control instruction set be issued to the actuator group, the specific steps of which include:

[0113] The dynamic target setpoint range is matched with the preset equipment operation safety threshold range to determine the equipment operation priority; the safety threshold range is a multi-dimensional space that defines the absolute safe operation boundary of the equipment on various key parameters (such as maximum allowable temperature, maximum speed, pressure limit, etc.).

[0114] The equipment's operating priority is determined by a priority scoring function. To perform quantitative grading, an example is as follows:

[0115] ;

[0116] in, Represents the current device state vector With the nearest security boundary The Euclidean distance or Mahalanobis distance, the smaller the value, the closer to the dangerous state, and the higher the priority; (rate of change of state) and The larger the value of (performance degradation gradient), the faster the state is deteriorating, and the higher the priority. Based on the score range of P_rank, the instruction set is precisely determined into three safety priorities: high, medium, and low.

[0117] If a command is determined to be of high priority, indicating that its execution is related to the immediate safety of the equipment or the patient (such as emergency deceleration or overheat protection shutdown), the adaptive control command set is placed in the high-priority queue and immediately sent to the actuator group for execution. The high-priority queue is directly managed by the system's real-time operating system (RTOS) kernel, bypassing all intermediate buffers and waiting links, and is immediately sent to the actuator group for execution via a high-bandwidth, low-latency fieldbus (such as CANopen or EtherCAT), ensuring microsecond-level response of critical control actions.

[0118] If the instruction is determined to be of medium or low priority, it is typically used for performance optimization, energy-saving adjustment, or preventive maintenance (such as fine-tuning operating parameters to improve energy efficiency). In this case, the adaptive control instruction set is placed in the confirmation queue, and a verification request is sent to the system monitoring terminal (such as a central nurse station or engineer station) according to a predefined industrial communication protocol (such as Modbus-TCP, OPCUA, or DDS). The verification request not only includes the instruction content, but also the reason for its generation (such as "to improve energy efficiency"), the expected effect, and related status data, providing sufficient basis for the decision-making of the monitoring personnel.

[0119] Upon receiving a verification request, monitoring personnel (such as nurses or clinical engineers) can make a judgment based on the actual clinical situation. Only after receiving a verification confirmation instruction that conforms to the protocol specifications from the system monitoring terminal will the corresponding adaptive control instruction set in the confirmation queue be sent to the actuator group for execution. If a "reject" instruction is received or no response is received within a timeout period, the system will automatically discard the instruction set, triggering log recording or alarms.

[0120] The safety priority-based hierarchical instruction issuance and confirmation mechanism introduced in this application enables refined and differentiated scheduling of control instructions. It ensures rapid response to emergency safety instructions, balances control efficiency and operational safety, and constructs an information closed loop from the equipment layer to the monitoring layer. This makes the entire control process transparent and traceable, enhancing the practicality of the project.

[0121] This application further proposes that, based on operational deviations, an adaptive control algorithm is used to dynamically optimize the adaptive control instruction set, specifically including:

[0122] The adaptive control algorithm uses an adaptive PID control algorithm;

[0123] Based on the equipment operating deviation and its rate of change, the proportional coefficient of the PID controller is dynamically adjusted according to fuzzy logic rules. Integral time constant and differential time constant One or more of them;

[0124] The inputs to the fuzzy logic rules are the deviation *e* and the rate of change of deviation *ec*. Based on the membership degrees of the current inputs *e* and *ec*, all relevant fuzzy rules are activated, and the output fuzzy set is obtained through fuzzy inference methods (such as the Mamdani method). The output is the parameter adjustment amount: the proportional coefficient adjustment amount Δ. Adjustment amount of integral time constant Δ Adjustment amount of differential time constant The deviation *e* and the rate of change of deviation *ec* need to undergo a fuzzification process, that is, by using a preset membership function, they are converted into fuzzy sets represented by linguistic values ​​such as "negative large (NB)", "negative medium (NM)", "zero (ZO)", "positive small (PS)", and "positive large (PB)". For example, for the deviation *e*, its universe of discourse can be defined as [-10, 10], and five fuzzy subsets NB, NM, ZO, PM, and PB can be set, calculating the membership degree of the current value of *e* to each subset. Subsequently, the fuzzy inference engine makes decisions based on a pre-designed fuzzy rule base. This rule base is the culmination of control expert experience and takes the form of a series of "IF...THEN..." fuzzy conditional statements.

[0125] Parameter adjustment amount Δ The calculation formula is as follows:

[0126] ;

[0127] Where z is the universe of discourse value of the output variable. The adjustment amount Δ is the proportional coefficient. Adjustment amount of integral time constant Δ Adjustment amount of differential time constant , This represents the membership degree corresponding to z. This allows us to obtain the precise parameter adjustment amount.

[0128] The adjusted PID parameters include proportionality coefficient Used to adjust the system's response speed, integral time constant Used to eliminate the steady-state error of the system, the differential time constant Used to improve system stability and predictability.

[0129] Calculate the output adjustment amount based on the adjusted PID parameters. Adjust the amount based on the output. Generate an optimized adaptive control instruction set and adjust the target parameters.

[0130] This application solves the inherent shortcomings of traditional PID parameters, such as difficulty in tuning and inability to adapt to dynamic changes in the system. It realizes automatic optimization of controller parameters under different operating states (such as large deviation catch-up and small deviation fine-tuning), so that the system always maintains a near-optimal control state, effectively eliminating steady-state error and enhancing the system's stability and predictability of changing trends.

[0131] This application further proposes that, based on operational deviations, the step of dynamically optimizing the adaptive control instruction set using an adaptive control algorithm also includes:

[0132] The system first accesses the control decision function U, which can be understood as converting the multimodal state vector... The mathematical relationship mapped to control commands. This is for evaluating each state parameter. The system calculates the partial derivative of the parameter ∂U / with respect to the decision function, representing the magnitude of the influence of factors such as current, vibration, and temperature on the current decision. , The larger the value, the better U is for The more sensitive it is to changes, that is... The greater the potential influence, the greater the sensitivity will be relative to the instantaneous change in this parameter. By combining these parameters, the contribution of each state parameter in the multimodal device state data to the decision-making process is calculated:

[0133] ;

[0134] The denominator is the sum of the instantaneous influences of all n state parameters, normalized. This represents the sum of the instantaneous influences of all state parameters, and its value is always positive. When the system is first started or all parameters remain unchanged, this sum may be zero. In this case, a default uniform weight (such as the contribution of each parameter) can be set. =1 / n), until a parameter change is detected.

[0135] This represents the contribution of the m-th state parameter. ∈(0,1), representing the percentage contribution of the m-th parameter to the total change in the control decision at the current time; U represents the control decision function. This represents the m-th state parameter. This represents the change in the m-th state parameter;

[0136] Based on contribution Calculated contribution vector Adjusting the weights of the corresponding state parameters in the adaptive control algorithm Optimize the adaptive control instruction set. Parameters with higher contribution have greater weight in the control algorithm. It will also be enhanced accordingly.

[0137] This application realizes data-driven and dynamic adaptive allocation of control algorithm weights, overcoming the shortcomings of fixed weights and inability to respond to changes in operating conditions in traditional control, improving the system's control accuracy, response speed, and robustness to changes in internal parameters, and ensuring the optimal operating performance of medical equipment under various dynamic operating conditions.

[0138] This application further proposes that the device adaptive control method based on multimodal perception can also be used for multiple medical devices working collaboratively:

[0139] The historical operation database and control policy library contain a collaborative policy mapping table among multiple devices; this mapping table defines a knowledge graph of the coupling relationships between control logics of the devices. Each record can be formally represented as follows: This means "when device A is in a certain state or triggers a certain event, what coordinated action should device B perform?" For example, the mapping table may contain the rule: "IF ventilator (device A) triggers 'high airway pressure alarm' (state / event) THEN notify infusion pump (device B) to perform 'pause infusion' (action)."

[0140] When a change in the operating status of the first device in the collaborative group is detected, the status information of the first device is shared in real time with the second device in the collaborative group through the industrial IoT gateway; the collaborative group is dynamically determined based on the network of device functional coupling relationships.

[0141] The second device matches and executes corresponding cooperative adaptive control commands based on the received state change information.

[0142] The triggering of collaborative functions relies on a dynamic and intelligent networking mechanism. Devices determine their Coordination Groups in real time based on a dynamically constructed network of device functional coupling relationships. This network is modeled as a graph G=(V,E), where nodes V represent individual medical devices, and edges E represent functional coupling relationships (such as physiological connections or treatment process connections) between devices.

[0143] The system dynamically calculates and determines which devices belong to the same group that should be coordinated, based on real-time treatment tasks (such as "performing anesthesia for a patient in a specific bed") and the physical / logical connections of the equipment. For example, anesthesia machines, monitors, and infusion pumps serving a patient will be automatically and dynamically grouped. When the system detects a significant change in the operating status of the first device (such as a monitor) within the coordination group (e.g., a continuous decrease in the patient's SpO2), the coordination mechanism is immediately activated. The status information of the first device (such as SpO2 value and its trend) is then displayed. The data is encapsulated into a standard structured data message and shared in real-time with low latency to a second device (such as an anesthesia machine) within the collaborative group via a highly reliable industrial IoT gateway (such as a gateway using the OPCUA or MQTT protocol). This gateway ensures seamless interconnection between medical devices of different brands and protocols. Upon receiving status change information from the first device, the intelligent controller of the second device (such as the anesthesia machine) immediately performs pattern matching in its local collaborative policy mapping table. Once a corresponding collaborative policy is matched, the second device automatically generates or invokes the corresponding collaborative adaptive control instruction set and executes it immediately, without waiting for centralized commands from the central controller.

[0144] This application builds a control simulation model of an intelligent infusion pump in the MATLAB / Simulink 2022a environment and compares it with the traditional single-parameter threshold control method. The simulation conditions are as follows:

[0145] The transfer function model of a certain type of infusion pump is used, including modules for motor dynamics, fluid resistance, and pipeline vibration. Current (±5mA accuracy), vibration (±0.1g accelerometer), noise (50dB background noise), and temperature (±0.5℃) are simulated. Gearbox wear is also simulated (vibration signal amplitude increases by 5% per month).

[0146] The experimental results are shown in the table below:

[0147] Performance indicators Traditional threshold control This system Increase Speed ​​control error ±2.5% ±0.8% 68% Fault warning accuracy 72% 95% 23% Response time (step) 120ms 40ms 67% Energy consumption (24-hour operation) 1.2kW·h 0.9kW·h 25%

[0148] This application realizes the intelligent and flexible definition of collaborative groups, gets rid of the rigid mode of traditional hard-coded networking, can flexibly adapt to the changing clinical scenarios, realizes intelligent linkage and rapid response between devices, minimizes the latency of cross-device control, and greatly improves the system response speed and collaborative processing capability in emergency situations (such as when the patient's condition deteriorates).

[0149] The following is a specific embodiment of a multimodal sensing-based adaptive control method for medical devices (taking intelligent ventilator pressure control as an example):

[0150] A patient is receiving mechanical ventilation using a smart ventilator (model XYZ-9000). The system collects data in real time via a multimodal sensor network:

[0151] Equipment operating parameters: A Hall current sensor (sampling rate 1kHz) is used to monitor the main motor current fluctuation (range ±0.2A), a MEMS vibration sensor (sampling rate 500Hz) is used to detect the fan bearing vibration (frequency bandwidth 20-1000Hz), an infrared temperature sensor (accuracy ±0.1℃) is used to monitor the motor winding temperature in real time (dynamic range 25-80℃), and a microphone array is used to collect air path noise (sound pressure level 50-80dB).

[0152] Patient physiological parameters: Airway pressure (range 0-40 cmH2O) was recorded using a piezoresistive sensor in the breathing tubing (sampling rate 200 Hz), end-tidal CO2 partial pressure was monitored by a carbon dioxide sensor (PETCO2 range 30-45 mmHg), and blood oxygen saturation (SpO2) was simultaneously acquired by a pulse oximeter probe (MasimoRadical-7). All data were synchronized at the hardware level (time error <0.5 ms) via the IEEE 11073 SDC protocol, generating a multimodal device status data vector. and patient safety state vector .

[0153] The system will display real-time data in the window ( =2s), short-term historical time window ( =30s) and long-term historical time window ( Multimodal data within 8 hours were input into a pre-trained three-layer LSTM deep state estimation model (128 hidden neurons, 12,000 hours of training data).

[0154] Model output depth state estimate =0.72 (1.0 represents a completely new state); corresponding to the rate of change of the depth state estimate. =−0.05 / min (slow performance degradation trend); performance degradation gradient =0.72×(−0.05)=−0.036 (linear decay mode). Based on comparison with historical operating database, the current operating mode is identified as "slight motor wear accompanied by increased airway resistance". Combined with patient safety parameters (SpO2=92%, PETCO2=38mmHg), a dynamic target setpoint range is generated: pressure support level 12-14cmH2O, respiratory rate 14-16 breaths / min, peak flow rate 60-65L / min.

[0155] Constructing dynamic safety boundary functions Specific constraints include:

[0156] =45− ≤0 (motor temperature ≤45℃);

[0157] = -35≤0 (airway pressure≤35cmH2O);

[0158] =90−SpO2≤0 (blood oxygen saturation≥90%).

[0159] Boundary parameters are updated using the Lyapunov online learning algorithm to ensure system stability. Current monitored values ​​( =42℃, =33cmH2O, SpO2=92) both satisfy Ψ≤0. Match the dynamic target interval with the control strategy library (calculate the matching degree using cosine similarity). =0.85>θ=0.8), directly generate an adaptive control instruction set: adjust the motor speed to 2,200RPM, the valve opening ratio to 68%, and the heating wire power to 3.5W.

[0160] Based on safety priority: Pressure control command ( =0.91) is placed in the high-priority queue and immediately sent to the actuator group; heating wire control command ( =0.45) is placed in the pending confirmation queue and sent to the nurse station for verification before execution.

[0161] The actuator assembly drives the brushless motor and proportional valve to operate, continuously collecting feedback data. The operational feedback shows that the actual pressure is 13.2 cmH2O (target range center value 13 cmH2O), the deviation is e = 0.2 cmH2O, and the deviation change rate is ec = 0.05 cmH2O / s.

[0162] The instruction is dynamically optimized using an adaptive PID algorithm:

[0163] Fuzzy logic rules adjust parameters based on e and ec: Δ =+0.2, Δ =−0.1, =+0.05;

[0164] Updated parameters =1.2 + 0.2 = 1.4, =0.8−0.1=0.7, =0.3 + 0.05 = 0.35;

[0165] Calculate output adjustment amount ;

[0166] Generate optimization instructions: fine-tune the motor speed to 2230 RPM.

[0167] A collaborative monitor alarm was detected (patient's heart rate suddenly increased to 125 beats / min), and the event was shared to the ventilator via the industrial IoT gateway (OPCUA protocol). The collaborative strategy mapping table was matched (rule: IF heart rate > 120 THEN reduce PEEP level), and a collaborative instruction was generated: reduce PEEP from 8 cmH2O to 6 cmH2O. Simultaneously, based on the performance degradation gradient... =−0.036, predicting an 8% decrease in motor performance over the next 24 hours, prompting an early maintenance warning: it is recommended to replace the fan filter and calibrate the flow sensor.

[0168] This embodiment utilizes a multimodal sensing network to dynamically adjust control targets based on the device's health status and correct execution commands in real time based on patient safety needs. When the monitor indicates an abnormal heart rate, the ventilator can immediately and autonomously reduce its pressure support level without waiting for central dispatch. The system ensures stable and accurate ventilation pressure, transforming traditional passive response maintenance into proactive preventative control and improving the reliability, safety, and intelligent collaboration of medical equipment in dynamic and complex environments.

[0169] Example 2:

[0170] like Figure 3 As shown, the device adaptive control system based on multimodal sensing, using the aforementioned device adaptive control method based on multimodal sensing, includes:

[0171] The multimodal sensing module is used to collect real-time operating parameters of medical equipment and generate multimodal equipment status data.

[0172] The status recognition module is used to identify the current operating mode and performance degradation trend of the equipment based on multimodal equipment status data and to perform deep status estimation with a pre-stored historical operation database, and to generate a dynamic target setpoint range.

[0173] The control instruction generation module is used to match the dynamic target setpoint range with the preset control strategy library to generate the corresponding adaptive control instruction set.

[0174] The instruction issuance and execution module is used to issue adaptive control instruction sets to the actuator group, drive the actuator group to perform control actions, and continuously collect operational feedback data of the medical equipment;

[0175] The operation feedback optimization module is used to compare the operation feedback data with the dynamic target setpoint range to obtain the equipment operation deviation, and to dynamically optimize the adaptive control instruction set based on the operation deviation using a control algorithm.

[0176] The actuator group is used to receive the optimized adaptive control instruction set and drive one or more actuators to adjust the operating parameters of the medical device.

[0177] The multimodal sensing module consists of a physical sensing layer and a data acquisition layer. The physical sensing layer includes various sensors attached to or embedded in key components of the medical device, such as Hall current sensors (for acquiring motor current), MEMS vibration accelerometers (for acquiring body vibration), digital microphones (for acquiring operating noise), and PT100 temperature sensors (for acquiring temperature rise). The data acquisition layer typically consists of one or more high-precision data acquisition cards (DAQ) or distributed I / O modules, responsible for conditioning, filtering, and converting the analog signals from the sensors to digital data, and uploading the digitized multimodal device status data via a fieldbus (such as a CAN bus).

[0178] The core hardware of the state recognition module is an embedded industrial computer or a high-performance edge computing gateway (such as one based on an NVIDIA Jetson or Intel Atom platform). This computer is equipped with sufficient computing power (CPU / GPU), memory, and storage space to load and run a pre-stored historical running database and to execute deep state estimation algorithms (such as LSTM model inference).

[0179] The control instruction generation module is typically implemented by a programmable automation controller (PAC) or a high-end PLC (such as Beckhoff CX series or Siemens S7-1500). It internally stores a control strategy library and runs a matching algorithm, responsible for generating adaptive control instruction sets in real time based on the input status information.

[0180] The hardware for the instruction issuance and execution module includes a communication interface and an execution driver layer. The communication interface typically uses an industrial Ethernet switch and a protocol gateway (such as a Profinet to Modbus gateway), responsible for issuing instruction sets according to security priorities. The execution driver layer consists of a series of servo drivers, stepper motor drivers, and relay output modules, which directly receive instructions and drive the actuator assembly.

[0181] The operation feedback optimization module is typically completed collaboratively by the aforementioned PAC / PLC and edge computing gateway. The PAC is responsible for fast closed-loop control (such as PID calculation), while the edge computer is responsible for running more complex optimization algorithms (such as fuzzy inference and contribution calculation). The two exchange data at high speed through an internal bus.

[0182] The actuator group consists of the original actuators on the medical device and newly added drive components, such as servo motors (for precise pump speed control), voice coil motors (for precision valve control), solid-state relays (for heating control), alarm indicator lights, and human-machine interfaces (HMIs).

[0183] Taking an intelligent infusion pump integrating this system as an example, the multimodal sensing module monitors the current of the main drive motor through a current sensor installed inside the pump; a vibration sensor monitors the mechanical vibration of the pump body; a microphone listens to the operating noise of the gearbox; and a temperature sensor monitors the temperature rise of the motor and drive board. This data is collected in real time by an STM32 microcontroller as the core acquisition board. The data collected by the status recognition module is uploaded to a built-in Intel NUC edge computing unit via the EtherCAT bus. This unit runs a pre-trained LSTM model to analyze the data. The model identifies a slight fluctuation in the current of the motor, along with a slight increase in the high-frequency vibration signal, indicating early signs of slight wear in the gearbox (performance degradation trend). Based on this "slight wear" status, the control command generation module, through the edge computing unit, generates a dynamic target setpoint range, requiring the motor speed fluctuation to be controlled within ±1% (more stringent than the normal ±2%). This range is sent to the main control PLC (Beckham CX9020) of the infusion pump. The control strategy library in the PLC is matched with the "high-precision speed stabilization" strategy to generate a corresponding adaptive control instruction set, which aims to compensate for transmission errors caused by gear wear through more frequent fine-tuning.

[0184] The instruction issuance and execution module determines the instruction as high priority (related to infusion accuracy and patient safety) via the PLC and immediately sends the instruction to the servo driver via the EtherCAT bus. The actuator group drives the main drive servo motor through the servo driver to perform precise acceleration and deceleration actions to maintain a stable flow rate.

[0185] The system continuously monitors the deviation between the actual rotational speed and the dynamic target setpoint range through a feedback optimization module, and feeds this deviation back to the optimization algorithm in the PLC's built-in PID controller and edge computing unit. The algorithm determines the contribution of the "vibration" parameter to the current control decision. At its highest level, the control law is automatically adjusted, increasing the weight of vibration feedback, making the control commands more sensitive to vibration signals, thereby further optimizing the control effect and ensuring high-precision infusion even under slight mechanical wear.

[0186] This system achieves full-chain hardware collaboration from data acquisition to intelligent decision-making and precise execution, ensuring the system's real-time performance and reliability, and improving the accuracy, reliability, safety, and intelligence level of equipment adaptive control operation.

[0187] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A device adaptive control method based on multimodal sensing, used for monitoring and collaborative control of the operating status of medical equipment, characterized in that: Includes the following steps: The real-time operating parameters of the medical device are acquired to generate multimodal device status data, which includes at least current, vibration, noise, and temperature signals. Based on the multimodal device status data, by performing deep status estimation with the pre-stored historical operation database, the current device operation mode and performance degradation trend are identified, and a dynamic target setpoint interval is generated; The dynamic target setpoint range is matched with a preset control strategy library to generate a corresponding adaptive control instruction set; wherein, the adaptive control instruction set is used to drive the actuator group to reach the target parameters; Based on the equipment's operational safety priority, the adaptive control instruction set is sent to the actuator group to drive the actuator group to perform control actions and continuously collect operational feedback data of the medical equipment; The operational feedback data is compared with the dynamic target setpoint range to obtain the equipment operational deviation, and based on the operational deviation, the adaptive control instruction set is dynamically optimized using an adaptive control algorithm.

2. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: The steps for depth state estimation based on the multimodal device state data include: Real-time data window for synchronously acquiring the status data of the multimodal device Short-term historical time window and long-term historical time window Data within; Based on the preset feature extraction algorithm, the feature vectors of the data in each time window are calculated respectively. The feature vector is input into a trained Long Short-Term Memory (LSTM) network model, which outputs a deep state estimate representing the device's health status. and the rate of change of the state corresponding to the depth state estimate ; Based on the depth state estimate and the rate of change of state Calculate the performance degradation gradient of computing devices ; The depth state estimate and the performance degradation gradient By comparing the modal features with those in the historical operation database, the current operating mode and performance degradation trend of the device can be identified.

3. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: The steps of identifying the current equipment operating mode and performance degradation trend, and generating dynamic target setpoint intervals, also include: Obtain the security state parameter vector of the external object interacting with the device. ; Construct a dynamic safety boundary function based on predefined safety constraint rules. ,in This is a multimodal device state data vector, where t is the acquisition time; The safety boundary function is updated using an online learning algorithm based on Lyapunov stability; If the control instruction set causes the value of the safety boundary function to be greater than 0, then an alternative target setpoint interval is generated for safety constraint correction.

4. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: The dynamic target setpoint range is matched with a preset control strategy library to generate a corresponding adaptive control instruction set. The control strategy library includes a self-evolution mechanism, which specifically includes: Set the matching threshold θ; When the matching degree between the dynamic target setpoint interval and the control strategy library is calculated using cosine similarity... When the degree of matching falls below the matching threshold θ, the control strategy library is triggered to enter the strategy self-evolution process. Based on the multimodal device status data and historical operating data, a new control strategy is generated online using a reinforcement learning algorithm. ; For the new control strategy After simulation verification and security checks, the control strategy library is updated.

5. The device adaptive control method based on multimodal sensing according to claim 4, characterized in that: The reinforcement learning algorithm uses a two-delay deep deterministic policy gradient algorithm to update the new control policy. New control strategy The objective function is as follows: ； in, Indicates the control policy before the update, s and These represent the current device state and the next device state, respectively, as characterized by the multimodal device state data; 'a' represents the control action selected from the preset action space; 'r' represents the instantaneous reward derived from a comprehensive evaluation of device operating efficiency and energy consumption; and 'η' is the learning rate. ϵ is the discount factor; ϵ is the exploration noise. Criticizing the internet in the name of the target For the target policy network.

6. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: Based on the equipment's operational safety priority, the adaptive control instruction set is sent to the actuator group. Specific steps include: The dynamic target setpoint range is matched with the preset equipment operation safety threshold range to determine the equipment operation priority; If it is determined to be of high priority, the adaptive control instruction set is placed in the high priority queue and immediately sent to the actuator group for execution; If the priority is determined to be medium or low, the adaptive control instruction set is placed in the confirmation queue, and a verification request is sent to the system monitoring terminal according to the predefined industrial communication protocol. Upon receiving the verification confirmation instruction fed back through the system monitoring terminal, the corresponding adaptive control instruction set in the confirmation queue is sent to the actuator group for execution.

7. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: Based on the aforementioned operational deviation, an adaptive control algorithm is used to dynamically optimize the adaptive control instruction set, specifically including: The adaptive control algorithm employs an adaptive PID control algorithm. Based on the equipment operating deviation and its rate of change, the proportional coefficient of the PID controller is dynamically adjusted according to fuzzy logic rules. Integral time constant and differential time constant One or more of them; The inputs to the fuzzy logic rule are the deviation e and the rate of change of deviation ec, and the output is the parameter adjustment amount Δ. Δ , ; Calculate the output adjustment amount based on the adjusted PID parameters. And adjust the amount according to the output. Generate an optimized adaptive control instruction set and adjust the target parameters; The adjusted PID parameters include The proportionality coefficient The integral time constant is used to adjust the system's response speed. The differential time constant is used to eliminate the steady-state error of the system. Used to improve system stability and predictability.

8. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: Based on the aforementioned operational deviation, the step of dynamically optimizing the adaptive control instruction set using an adaptive control algorithm further includes: Calculate the contribution of each state parameter in the multimodal device state data to the decision-making process: ； in, Let represent the contribution of the m-th state parameter, and U represent the control decision function. This represents the m-th state parameter. This represents the change in the m-th state parameter; Based on the contribution Adjust the weights of the corresponding state parameters in the adaptive control algorithm to optimize the adaptive control instruction set.

9. The device adaptive control method based on multimodal sensing according to claim 1, characterized in that: The device adaptive control method based on multimodal sensing is also used for multiple medical devices working in tandem: The historical operation database and the control strategy library contain a collaborative strategy mapping table between multiple devices; When a change in the operating status of the first device in the collaborative group is detected, the status information of the first device is shared in real time with the second device in the collaborative group through the industrial IoT gateway. The collaborative group is dynamically determined based on the network of device functional coupling relationships; The second device matches and executes corresponding cooperative adaptive control commands based on the received state change information.

10. A device adaptive control system based on multimodal sensing, characterized in that: The device adaptive control method based on multimodal sensing as described in any one of claims 1 to 9 includes: The multimodal sensing module is used to collect real-time operating parameters of medical equipment and generate multimodal equipment status data. The status recognition module is used to identify the current operating mode and performance degradation trend of the device based on the multimodal device status data, by performing deep status estimation with a pre-stored historical operation database, and to generate a dynamic target setpoint range. The control instruction generation module is used to match the dynamic target setpoint range with a preset control strategy library to generate a corresponding adaptive control instruction set. The instruction issuance and execution module is used to issue the adaptive control instruction set to the actuator group, drive the actuator group to perform control actions, and continuously collect the operation feedback data of the medical device; The operation feedback optimization module is used to compare the operation feedback data with the dynamic target setpoint range to obtain the equipment operation deviation, and to dynamically optimize the adaptive control instruction set based on the operation deviation using a control algorithm. An actuator group is used to receive the optimized adaptive control instruction set and drive one or more actuators to adjust the operating parameters of the medical device.

Images