Multi-mode invasive pressure sensor based on embedded artificial intelligence induction

Abstract

The invention discloses a multi-mode invasive pressure sensor based on embedded artificial intelligence induction, and relates to the technical field of medical instruments, the multi-mode invasive pressure sensor comprises a sensor shell, a multi-mode sensing array and a fluid channel communicated with the multi-mode sensing array are integrated in the sensor shell, and a micro pump is arranged at the inlet end of the fluid channel; the system further comprises an embedded manual processing system. By providing an integrated architecture of a multi-modal sensing array and embedded AI processing, an embedded AI algorithm recognizes and corrects waveform distortion caused by over-damping and under-damping in real time, it is ensured that reading of systolic pressure and diastolic pressure accurately reflects real physiological values, the pressure measurement error is reduced, the precision is improved, and the measurement accuracy is improved. According to the present invention, the clinical high precision requirement on the vasoactive drug titration is met, the automatic diagnosis can be performed, the optimal dynamic response state can be rapidly adjusted, the sensor is particularly suitable for the time-sensitive emergency and operation environment, the sensor has the high overall waterproof grade, the clinical routine disinfection can be tolerated, and the service life meets the disposable use requirement.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to a multimodal invasive pressure sensor based on embedded artificial intelligence sensing. Background Technology

[0002] Invasive pressure monitoring is the gold standard for assessing circulatory function in critical care medicine and surgical anesthesia. By directly connecting a pressure sensor to an intravascular catheter, it acquires key parameters such as arterial pressure (ABP), central venous pressure (CVP), and pulmonary artery pressure (PAP) in real time, providing crucial information for fluid resuscitation and the administration of vasoactive drugs. However, the mainstream disposable invasive pressure sensors currently used in clinical practice are essentially passive pressure-to-electrical signal conversion devices, which have significant bottlenecks in terms of intelligence and functional reliability, specifically: 1. High reliance on manual calibration and uncontrollable accuracy drift: The zero-point drift of existing sensors requires medical staff to manually zero and calibrate them every 2-4 hours, and real-time compensation is not possible. Moreover, after 6 hours of continuous monitoring, the pressure measurement error of the uncalibrated sensor can reach 4.2±1.5 mmHg, far exceeding the clinically acceptable range, which can easily lead to misadjustment of vasoactive drug dosage and increase the risk of circulatory fluctuations. 2. Lagging signal quality assessment and strong subjective dependence: The dynamic response performance of the sensor needs to be tested by manually triggering square waves and relies on medical staff to interpret the waveform by eye. It cannot be quantitatively assessed in real time. Clinical surveys show that some abnormal pressure signals during surgery are not detected in time, leading to delays in diagnosis and treatment decisions. Third, lack of fault diagnosis capability and low troubleshooting efficiency: common clinical physical faults such as catheter adsorption to the blood vessel wall, bubble damping, and tubing kinking often require medical staff to observe the waveform for several minutes before they can be detected, resulting in a serious lag in signal quality assessment and even leading to incorrect clinical decisions. To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0003] The purpose of this invention is to provide an integrated architecture of multimodal sensing array and embedded AI processing. The embedded AI algorithm identifies and corrects waveform distortion caused by overdamping and underdamping in real time, ensuring that the readings of systolic and diastolic blood pressure accurately reflect the true physiological values. Pressure measurement errors are reduced and accuracy is improved, meeting the high-precision requirements of clinical titration of vasoactive drugs. It can also automatically diagnose and quickly adjust to the optimal dynamic response state, making it particularly suitable for time-sensitive emergency and surgical environments. At the same time, the sensor has a high overall waterproof rating, can withstand routine clinical disinfection, and its service life meets the requirements for single use.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a multimodal invasive pressure sensor based on embedded artificial intelligence sensing, including a sensor housing, wherein a multimodal sensing array and a fluid channel connected to the multimodal sensing array are integrated inside the sensor housing, a micro pump is provided at the inlet end of the fluid channel, and an embedded artificial processing system is also included. The multimodal sensing array includes a main pressure sensing unit, an optical reference sensing unit, a temperature sensing unit, and an impedance detection module, wherein: The main pressure sensing unit uses a MEMS piezoresistive sensor. The pressure-sensing surface of the MEMS piezoresistive sensor is in direct contact with the fluid channel to output a pressure electrical signal with drift. The optical reference sensing unit adopts a Fabry-Perot interferometer-type optical microcavity sensor. The optical microcavity sensor is pressure-coupled to the fluid channel through a flexible isolation membrane. The optical cavity of the optical microcavity sensor is isolated from the fluid in the fluid channel, and outputs a drift-free optical pressure reference signal. The impedance detection module includes a pair of miniature platinum ring electrodes embedded in the inner wall of the fluid channel for measuring fluid conductivity.

[0005] The embedded artificial intelligence processing system includes a data processing unit, an intelligent diagnostic unit, and a microfluidic execution unit, wherein: The data processing unit is used to acquire the main pressure signal, optical reference signal and temperature signal through the multimodal sensing array, and after fusing them based on the extended Kalman filter algorithm, estimate and compensate the zero drift of the main pressure sensor in real time, so as to output a high-precision calibration pressure value to the intelligent diagnostic unit. The intelligent diagnostic unit is used to construct a lightweight one-dimensional convolutional neural network model, obtain high-precision calibration pressure values, extract features, analyze the time-domain features of the pressure waveform time series, input the time-domain features into the convolutional neural network model to obtain intelligent diagnostic results, and send them to the microfluidic execution unit. The microfluidic actuator is used to acquire and analyze intelligent diagnostic results, obtain fault classification and corresponding feature vectors, compare the feature vectors with the preset optimal damping range, determine the current test state, generate microfluidic adjustment parameters, and control the micro pump to perform actions according to the microfluidic adjustment parameters to achieve fluid adjustment, thereby increasing resistance and suppressing waveform oscillation.

[0006] Furthermore, the specific process for outputting high-precision calibration pressure values ​​is as follows: S101. Obtain the detected pressure value and define the state vector. = ,in: The actual intravascular pressure value at time k; The zero-point drift of the main pressure sensor at time k; = - , representing the temperature difference between time k and the initial time.

[0007] S102. Establish the state transition equation: =A + ; Where, the state transition matrix A = , where α is a constant, and is the temperature-drift coupling coefficient, calibrated through isothermal experiments; Obtain process noise ~ (0, ), and the noise covariance matrix = The noise characteristics of the sensor were determined by actual measurement.

[0008] S103. Construct the observation model and define the observation vectors. = ,in: The output value of the main pressure sensing unit at time k; The output value of the optical reference sensing unit at time k; This is the output value of the temperature sensing unit at time k.

[0009] S103. Establish the observation equation: =H + ; Wherein, the observation matrix H = ; Obtain observation noise ~ Noise covariance matrix = Calculated from the accuracy class of each sensor; S104, The iterative calibration process is as follows: S1041. Initialization: During the quantitative period after power-on, the initial temperature is collected. Optical reference signal Initialize the state vector = covariance matrix = ; S1042. Prediction Steps: For each set of observation data collected, calculate the prior state. = Prior covariance = +Q; S1043, Update Steps: Calculate Kalman Gain = posterior state = Posterior covariance ; S10424, Output calibration value: Extract the true pressure estimate from the posterior state. = The output frequency is synchronized with the sampling rate of the main pressure sensor.

[0010] Furthermore, the specific process of constructing a lightweight one-dimensional convolutional neural network model and its training optimization is as follows: S201. A preset sampling frequency is used to obtain several pressure waveform segments based on the sampling frequency. The pressure waveform segments cover one or more cardiac cycles and contain complete waveform features, including systolic peak, diastolic trough, and dicrotic notch. S202. The pressure waveform segment is preprocessed using the Z-score standardization formula to eliminate the influence of baseline drift and amplitude difference. After standardization, the data range is controlled within [-3,3] and used as training samples. The training samples are divided into training set, validation set and test set in a ratio of 7:2:1. The weight file is downloaded and loaded onto the corresponding network to initialize the transfer network parameters. S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the intelligent diagnostic result, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain a lightweight one-dimensional convolutional neural network model. S204. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after several cycles, and a lightweight one-dimensional convolutional neural network model is obtained. S205. Deploy the trained lightweight one-dimensional convolutional neural network model in the intelligent diagnostic unit, input the temporal features into the lightweight one-dimensional convolutional neural network model, and obtain the intelligent diagnostic results.

[0011] Furthermore, the intelligent diagnostic results include fault classification and early warning logic, as well as feature vectors corresponding to the fault classifications: Obtain the preset fault probability distribution interval [ , , , Take the maximum probability (max()). , , , The diagnostic results are as follows, with fault classification and early warning thresholds as follows: Fault category one, i.e., normal state: ≥90%, no warning; Fault category two, namely, air bubbles in the pipeline: ≥85% triggers a yellow alert, with a buzzer at 1Hz and a flashing yellow light; Fault category three, namely, catheter adhesion to the wall: ≥80% triggers an orange alert, with a buzzer at 2Hz and an orange light flashing; Fault category four, namely partial coagulation: ≥75% triggers a red alert, the buzzer frequency is 3Hz, and the red light stays on.

[0012] Furthermore, the specific process for generating microfluidic adjustment parameters is as follows: S301. The time series of high-precision calibration pressure values ​​is analyzed by a lightweight one-dimensional convolutional neural network model to extract time-domain features, including the natural frequency of the current waveform, the rise slope and fall rate of the waveform, the overshoot amplitude and duration of the waveform. S302. Generate a feature vector based on time-domain features. The core indicator of the feature vector is the damping coefficient Rt during the pressure measurement process. Obtain the preset optimal damping range (Rmin, Rmax). The comparison process is as follows: If the damping coefficient Rt is less than Rmin, it indicates that the pressure measurement process is in an underdamped state, and the natural frequency of the multimodal sensing array is high at this time. S303. For the underdamped state, obtain standard parameters and generate adjustment parameters based on the difference between the actual parameters and the standard parameters. The adjustment parameters include the operating mode, flow rate, and duration of the micro pump.

[0013] Furthermore, the sensor housing is injection molded from medical-grade polycarbonate with a mold precision of ±0.05mm. The medical-grade polycarbonate is Bayer Makrolon 2805, which meets the ISO10993 biocompatibility standard. The surface of the sensor housing is plasma-treated, with a roughness Ra≤0.4μm to prevent bacterial adhesion.

[0014] Furthermore, the fluid channel is manufactured using a precision drilling process, with an inner diameter of 1.5 mm and a roundness error of ≤0.02 mm. The inner wall of the fluid channel is coated with a polytetrafluoroethylene coating with a thickness of 5 μm to reduce the risk of thrombosis.

[0015] Furthermore, it also includes a temperature sensing unit, which is attached to the substrate surface of the main pressure sensing unit and is ≤0.3mm away from the core area of ​​the main pressure sensing unit, and is used to monitor the operating temperature of the main pressure sensing unit.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This multimodal invasive pressure sensor, based on embedded artificial intelligence sensing, provides an integrated architecture of multimodal sensing array and embedded AI processing. The embedded AI algorithm identifies and corrects waveform distortion caused by overdamping and underdamping in real time, ensuring that the readings of systolic and diastolic blood pressure accurately reflect the true physiological values. The pressure measurement error is reduced and the accuracy is improved, meeting the high-precision requirements of clinical titration of vasoactive drugs. It can also automatically diagnose and quickly adjust to the optimal dynamic response state, making it particularly suitable for time-sensitive emergency and surgical environments. At the same time, the sensor has a high overall waterproof rating, can withstand routine clinical disinfection, and its service life meets the requirements for single use. Attached Figure Description

[0017] Figure 1 A schematic diagram of the embedded manual processing system structure of the present invention is shown; Figure 2 A schematic diagram of the sensor structure of the present invention is shown; Figure 3 A schematic diagram of the sensor structure from another angle of the present invention is shown; Legend: 1. Sensor housing; 2. Multimodal sensing array; 3. Fluid channel; 4. Micro pump. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: like Figure 1 As shown, a multimodal invasive pressure sensor based on embedded artificial intelligence sensing includes a sensor housing. The sensor housing 1 integrates a multimodal sensing array 2 and a fluid channel 3 connected to the multimodal sensing array 2. A micro pump 4 is provided at the inlet end of the fluid channel 3. It also includes an embedded artificial processing system. The embedded artificial intelligence processing system includes a data processing unit, an intelligent diagnostic unit, and a microfluidic actuator unit, wherein: The data processing unit is used to acquire the main pressure signal, optical reference signal and temperature signal through the multimodal sensing array 2, and after fusing them based on the extended Kalman filter algorithm, estimate and compensate the zero drift of the main pressure sensor in real time, so as to output a high-precision calibration pressure value to the intelligent diagnostic unit. The specific process for outputting high-precision calibration pressure values ​​is as follows: S101. Obtain the detected pressure value and define the state vector. = ,in: The actual intravascular pressure value at time k; The zero-point drift of the main pressure sensor at time k; = - , representing the temperature difference between time k and the initial time.

[0020] S102. Establish the state transition equation: =A + ; Where, the state transition matrix A = , where α is a constant, and is the temperature-drift coupling coefficient, calibrated through isothermal experiments; Obtain process noise ~ (0, ), and the noise covariance matrix = The noise characteristics of the sensor were determined by actual measurement.

[0021] S103. Construct the observation model and define the observation vectors. = ,in: The output value of the main pressure sensing unit at time k; The output value of the optical reference sensing unit at time k; This is the output value of the temperature sensing unit at time k.

[0022] S103. Establish the observation equation: =H + ; Wherein, the observation matrix H = ; Obtain observation noise ~ Noise covariance matrix = Calculated from the accuracy class of each sensor; S104, The iterative calibration process is as follows: S1041. Initialization: During the quantitative period after power-on, the initial temperature is collected. Optical reference signal Initialize the state vector = covariance matrix = ; S1042. Prediction Steps: For each set of observation data collected, calculate the prior state. = Prior covariance = +Q; S1043, Update Steps: Calculate Kalman Gain = posterior state = Posterior covariance ; S10424, Output calibration value: Extract the true pressure estimate from the posterior state. = The output frequency is synchronized with the sampling rate of the main pressure sensor.

[0023] The intelligent diagnostic unit is used to construct a lightweight one-dimensional convolutional neural network model, obtain high-precision calibration pressure values, extract features, analyze the time-domain features of the pressure waveform time series, input the time-domain features into the convolutional neural network model to obtain intelligent diagnostic results, and send them to the microfluidic execution unit. The specific process of constructing a lightweight one-dimensional convolutional neural network model and its training and optimization is as follows: S201. Preset sampling frequency, acquire several pressure waveform segments according to the sampling frequency, the pressure waveform segments cover one or more cardiac cycles and contain complete waveform features, including systolic peak, diastolic trough and dicrotic notch; S202. The pressure waveform segment is preprocessed using the Z-score standardization formula to eliminate the influence of baseline drift and amplitude difference. After standardization, the data range is controlled within [-3,3] and used as training samples. The training samples are divided into training set, validation set and test set in a ratio of 7:2:1. The weight file is downloaded and loaded onto the corresponding network to initialize the transfer network parameters. S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the intelligent diagnostic result, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain a lightweight one-dimensional convolutional neural network model. S204. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after several cycles, and a lightweight one-dimensional convolutional neural network model is obtained. S205. Deploy the trained lightweight one-dimensional convolutional neural network model in the intelligent diagnostic unit, input the temporal features into the lightweight one-dimensional convolutional neural network model, and obtain the intelligent diagnostic results.

[0024] The intelligent diagnostic results include fault classification and early warning logic, as well as the feature vectors corresponding to the fault classifications: Obtain the preset fault probability distribution interval [ , , , Take the maximum probability (max()). , , , The diagnostic results are as follows, with fault classification and early warning thresholds as follows: Fault category one, i.e., normal state: ≥90%, no warning; Fault category two, namely, air bubbles in the pipeline: ≥85% triggers a yellow alert, with a buzzer at 1Hz and a flashing yellow light; Fault category three, namely, catheter adhesion to the wall: ≥80% triggers an orange alert, with a buzzer at 2Hz and an orange light flashing; Fault category four, namely partial coagulation: ≥75% triggers a red alert, the buzzer frequency is 3Hz, and the red light stays on.

[0025] The microfluidic actuator is used to acquire and analyze intelligent diagnostic results, obtain fault classification and corresponding feature vectors, compare the feature vectors with the preset optimal damping range, determine the current test state, generate microfluidic adjustment parameters, and control the micro pump 4 to perform actions to achieve fluid adjustment, thereby increasing resistance and suppressing waveform oscillation.

[0026] The specific process for generating microfluidic adjustment parameters is as follows: S301. The time series of high-precision calibration pressure values ​​is analyzed by a lightweight one-dimensional convolutional neural network model to extract time-domain features, including the natural frequency of the current waveform, the rise slope and fall rate of the waveform, the overshoot amplitude and duration of the waveform. S302. Generate a feature vector based on time-domain features. The core indicator of the feature vector is the damping coefficient Rt during the pressure measurement process. Obtain the preset optimal damping range (Rmin, Rmax). The comparison process is as follows: If the damping coefficient Rt is less than Rmin, it indicates that the pressure measurement process is in an underdamped state, and the natural frequency of the multimodal sensing array is high at this time. S303. For underdamped conditions, standard parameters are obtained, and adjustment parameters are generated based on the difference between the actual parameters and the standard parameters. The adjustment parameters include the operating mode, flow rate, and duration of the micro pump 4.

[0027] This invention provides an integrated architecture of a multimodal sensing array 2 and embedded AI processing. The embedded AI algorithm identifies and corrects waveform distortion caused by overdamping and underdamping in real time, ensuring that the readings of systolic and diastolic blood pressure accurately reflect the true physiological values. Pressure measurement errors are reduced and accuracy is improved, meeting the high-precision requirements of clinical titration of vasoactive drugs. It can also automatically diagnose and quickly adjust to the optimal dynamic response state, making it particularly suitable for time-sensitive emergency and surgical environments. At the same time, the sensor has a high overall waterproof rating, can withstand routine clinical disinfection, and its service life meets the requirements for single use.

[0028] Example 2: like Figure 2-3 As shown, a multimodal invasive pressure sensor based on embedded artificial intelligence sensing includes a sensor housing 1. The sensor housing 1 integrates a multimodal sensing array 2 and a fluid channel 3 connected to the multimodal sensing array 2. A micro pump 4 is provided at the inlet end of the fluid channel 3. The multimodal sensing array includes a main pressure sensing unit, an optical reference sensing unit, a temperature sensing unit, and an impedance detection module, wherein: The main pressure sensing unit uses a MEMS piezoresistive sensor. The pressure-sensing surface of the MEMS piezoresistive sensor is in direct contact with the fluid channel to output a pressure electrical signal with drift. The optical reference sensing unit adopts a Fabry-Perot interferometer optical microcavity sensor. The optical microcavity sensor is pressure-coupled to the fluid channel 3 through a flexible isolation membrane. The optical cavity of the optical microcavity sensor is isolated from the fluid in the fluid channel 3, and outputs a drift-free optical pressure reference signal. The impedance detection module includes a pair of miniature platinum ring electrodes embedded in the inner wall of the fluid channel 3 for measuring fluid conductivity.

[0029] The sensor housing 1 is injection molded from medical-grade polycarbonate with a mold accuracy of ±0.05mm. The medical-grade polycarbonate is Bayer Makrolon 2805, which meets the ISO10993 biocompatibility standard. The surface of the sensor housing 1 is plasma treated, and its roughness Ra≤0.4μm prevents bacterial adhesion.

[0030] Fluid channel 3 is manufactured using a precision drilling process. The inner diameter of fluid channel 3 is 1.5 mm, and the roundness error is ≤0.02 mm. The inner wall of fluid channel 3 is coated with polytetrafluoroethylene with a coating thickness of 5 μm to reduce the risk of thrombosis.

[0031] It also includes a temperature sensing unit, which is mounted on the substrate surface of the main pressure sensing unit and is ≤0.3mm away from the core area of ​​the main pressure sensing unit. It is used to monitor the operating temperature of the main pressure sensing unit.

[0032] The main pressure sensing unit is mounted on the alumina sensing base with silver paste conductive adhesive. The pressure sensing surface is coated with a 1μm thick parylene as a biocompatible coating. The biocompatible coating meets the ISO10993-10 skin irritation standard. The leads are bonded to the signal conditioning circuit of the PCB through gold wire bonding. The signal conditioning circuit includes an instrumentation amplifier AD8221 to input the offset voltage, which is ≤10μV. The optical reference sensing unit is fabricated using deep reactive ion etching (DRIE) on silicon wafers. It has a cavity length of 100 μm and a cavity wall perpendicularity of ≥89.9°. It is fixed to the sensing base by laser welding. The flexible isolation membrane is attached to the fluid channel 3. The optical output end is connected to the photoelectric converter via a single-mode optical fiber. The photoelectric converter is a Thorlabs PDA100A with a response time of ≤10 ns and a dark current of ≤1 nA. The temperature sensing unit is mounted on the chip substrate of the MEMS piezoresistive sensor via thermally conductive silicon, and the leads are soldered to the Wheatstone bridge circuit on the PCB. The Wheatstone bridge circuit is powered by 3.3V and has a measurement resolution of 0.001℃. The impedance detection electrode is embedded in the inner wall of the fluid channel using laser engraving technology. The lead wire is connected to the impedance measurement circuit on the PCB through polyimide wire. The impedance measurement circuit includes a constant current source AD590, which outputs a constant current of 1mA with an accuracy of ±0.5%.

[0033] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0034] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, for example, the division of modules is merely a logical functional division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the apparatus or module can be electrical, mechanical or other forms. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multimodal invasive pressure sensor based on embedded artificial intelligence sensing, characterized in that, The system includes a sensor housing (1), which integrates a multimodal sensing array (2) and a fluid channel (3) connected to the multimodal sensing array (2). A micro pump (4) is provided at the inlet end of the fluid channel (3). The system also includes an embedded artificial processing system. The multimodal sensing array includes a main pressure sensing unit, an optical reference sensing unit, a temperature sensing unit, and an impedance detection module, wherein: The main pressure sensing unit adopts a MEMS piezoresistive sensor. The pressure sensing surface of the MEMS piezoresistive sensor is in direct contact with the fluid channel (3) to output a pressure electrical signal with drift. The optical reference sensing unit adopts a Fabry-Perot interferometer type optical microcavity sensor. The optical microcavity sensor is pressure-coupled to the fluid channel (3) through a flexible isolation membrane. The optical cavity of the optical microcavity sensor is isolated from the fluid in the fluid channel (3) and outputs a drift-free optical pressure reference signal. The impedance detection module includes a pair of miniature platinum ring electrodes embedded in the inner wall of the fluid channel (3) for measuring fluid conductivity; The embedded artificial intelligence processing system includes a data processing unit, an intelligent diagnostic unit, and a microfluidic execution unit, wherein: The data processing unit is used to acquire the main pressure signal, optical reference signal and temperature signal through the multimodal sensing array, and after fusing them based on the extended Kalman filter algorithm, estimate and compensate the zero drift of the main pressure sensor in real time, so as to output a high-precision calibration pressure value to the intelligent diagnostic unit. The intelligent diagnostic unit is used to construct a lightweight one-dimensional convolutional neural network model, obtain high-precision calibration pressure values, extract features, analyze the time-domain features of the pressure waveform time series, input the time-domain features into the convolutional neural network model to obtain intelligent diagnostic results, and send them to the microfluidic execution unit. The microfluidic actuator is used to acquire and analyze intelligent diagnostic results, obtain fault classification and corresponding feature vectors, compare the feature vectors with the preset optimal damping range, determine the current test state, generate microfluidic adjustment parameters, and control the micro pump (4) to perform actions to achieve fluid adjustment in order to increase resistance and suppress waveform oscillation.

2. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The specific process for outputting high-precision calibration pressure values ​​is as follows: S101. Obtain the detected pressure value and define the state vector. = ,in: The actual intravascular pressure value at time k; The zero-point drift of the main pressure sensor at time k; = - , representing the temperature difference between time k and the initial time; S102. Establish the state transition equation: =A + ; Where, the state transition matrix A = , where α is a constant, and is the temperature-drift coupling coefficient, calibrated through isothermal experiments; Obtain process noise ~ (0, ), and the noise covariance matrix = The noise characteristics of the sensor were determined by actual measurement. S103. Construct the observation model and define the observation vectors. = ,in: The output value of the main pressure sensing unit at time k; The output value of the optical reference sensing unit at time k; The output value of the temperature sensing unit at time k; S103. Establish the observation equation: =H + ; Wherein, the observation matrix H= ; Obtain observation noise ~ Noise covariance matrix = Calculated from the accuracy class of each sensor; S104, The iterative calibration process is as follows: S1041. Initialization: During the quantitative period after power-on, the initial temperature is collected. Optical reference signal Initialize the state vector = covariance matrix = ; S1042. Prediction Steps: For each set of observation data collected, calculate the prior state. = Prior covariance = +Q; S1043, Update Steps: Calculate Kalman Gain = posterior state = Posterior covariance ; S10424, Output calibration value: Extract the true pressure estimate from the posterior state. = The output frequency is synchronized with the sampling rate of the main pressure sensor.

3. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The specific process of constructing a lightweight one-dimensional convolutional neural network model and its training and optimization is as follows: S201. A preset sampling frequency is used to obtain several pressure waveform segments based on the sampling frequency. The pressure waveform segments cover one or more cardiac cycles and contain complete waveform features, including systolic peak, diastolic trough, and dicrotic notch. S202. The pressure waveform segment is preprocessed using the Z-score standardization formula to eliminate the influence of baseline drift and amplitude difference. After standardization, the data range is controlled within [-3,3] and used as training samples. The training samples are divided into training set, validation set and test set in a ratio of 7:2:

1. The weight file is downloaded and loaded onto the corresponding network to initialize the transfer network parameters. S203. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the intelligent diagnostic result, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain a lightweight one-dimensional convolutional neural network model. S204. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after several cycles, and a lightweight one-dimensional convolutional neural network model is obtained. S205. Deploy the trained lightweight one-dimensional convolutional neural network model in the intelligent diagnostic unit, input the temporal features into the lightweight one-dimensional convolutional neural network model, and obtain the intelligent diagnostic results.

4. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The intelligent diagnostic results include fault classification and early warning logic, as well as feature vectors corresponding to the fault classification: Obtain the preset fault probability distribution interval [ , , , ], take the maximum probability max( , , , The diagnostic results are as follows, with fault classification and early warning thresholds as follows: Fault category one, i.e., normal state: ≥90%, no warning; Fault category two, namely, air bubbles in the pipeline: ≥85% triggers a yellow alert, with a buzzer at 1Hz and a flashing yellow light; Fault category three, namely, catheter adhesion to the wall: ≥80% triggers an orange alert, with a buzzer at 2Hz and an orange light flashing; Fault category four, namely partial coagulation: ≥75% triggers a red alert, the buzzer sounds at 3Hz, and the red light stays on.

5. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The specific process for generating microfluidic adjustment parameters is as follows: S301. The time series of high-precision calibration pressure values ​​is analyzed by a lightweight one-dimensional convolutional neural network model to extract time-domain features, including the natural frequency of the current waveform, the rise slope and fall rate of the waveform, the overshoot amplitude and duration of the waveform. S302. Generate a feature vector based on time-domain features. The core indicator of the feature vector is the damping coefficient Rt during the pressure measurement process. Obtain the preset optimal damping range (Rmin, Rmax). The comparison process is as follows: If the damping coefficient Rt is less than Rmin, it indicates that the pressure measurement process is in an underdamped state, and the natural frequency of the multimodal sensing array is high at this time. S303. For the underdamped state, obtain standard parameters and generate adjustment parameters based on the difference between the actual parameters and the standard parameters. The adjustment parameters include the operating mode, flow rate and duration of the micro pump (4).

6. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The sensor housing (1) is made of medical grade polycarbonate injection molding, and the surface of the sensor housing (1) is plasma treated, with a roughness Ra≤0.4μm.

7. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, The fluid channel (3) is made using a precision drilling process, and the inner wall of the fluid channel (3) is coated with a polytetrafluoroethylene coating.

8. The multimodal invasive pressure sensor based on embedded artificial intelligence sensing according to claim 1, characterized in that, It also includes a temperature sensing unit, which is attached to the substrate surface of the main pressure sensing unit and is ≤0.3mm away from the core area of ​​the main pressure sensing unit, and is used to monitor the operating temperature of the main pressure sensing unit.

Images