Medical equipment maintenance device and system based on AI automatic learning

Abstract

This invention relates to the field of medical equipment maintenance technology, specifically disclosing an AI-based automatic learning-based medical equipment maintenance device and system. The device includes a maintenance probe body, which comprises a probe housing, a multi-dimensional sensor array, a signal conditioning circuit, a microprocessor unit, and a communication interface. This invention integrates the multi-dimensional sensor array at the front end of the probe housing and embeds a microprocessor unit with an integrated neural network acceleration engine within the probe housing. This allows the microprocessor unit to load and execute fault diagnosis algorithm firmware in normal detection mode and load and execute self-learning update algorithm firmware in self-learning update mode. This achieves dual-mode operation capability of the maintenance equipment under a single probe architecture. The maintenance probe body can complete the entire workflow from data acquisition and fault diagnosis to model update without requiring on-site equipment replacement or connection to an external computing unit, reducing the hardware cost and on-site operational complexity of the maintenance equipment.

Description

Technical Field

[0001] This invention relates to the field of medical equipment maintenance technology, and in particular to medical equipment maintenance equipment and systems based on AI automatic learning. Background Technology

[0002] As medical equipment plays an increasingly prominent role in clinical diagnosis and treatment, its operational stability and reliability are directly related to medical quality and patient safety. Accurate and efficient monitoring and fault diagnosis of the operational status of medical equipment has become a core component of equipment management in medical institutions.

[0003] Currently, various technical solutions have emerged in the field of medical equipment maintenance. A common approach is to use portable or handheld testing devices that collect equipment operating signals through a single sensor and compare them with preset fixed thresholds to determine if the equipment is malfunctioning. These devices are simple in structure and low in cost, meeting the needs of routine inspections to a certain extent. Furthermore, some existing technologies attempt to integrate multiple sensors into handheld probes to achieve simultaneous acquisition of multiple parameters. With the development of artificial intelligence technology, machine learning and deep learning methods are also being introduced into the field of medical equipment fault diagnosis, enabling automatic identification of fault types by learning from historical operating data.

[0004] Existing handheld maintenance equipment mostly uses general-purpose microcontrollers or microprocessors, which have limited computing power and cannot simultaneously meet the dual requirements of real-time fault reasoning and online model updates. The reasoning and training of deep learning models both require high computing power. Under the limited power consumption and heat dissipation conditions of handheld devices, how to achieve efficient neural network reasoning and incremental learning locally is a key technical challenge that restricts the level of intelligence of maintenance equipment.

[0005] Therefore, we propose an AI-based automatic learning medical device repair equipment and system to address the problems mentioned in the background section. Summary of the Invention

[0006] The purpose of this invention is to provide medical device repair equipment and system based on AI automatic learning to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a medical device repair device based on AI automatic learning, comprising a repair probe body, wherein the repair probe body includes a probe housing, a multi-dimensional sensor array, a signal conditioning circuit, a microprocessor unit, a storage module, a communication interface, and a display module.

[0008] The probe housing has an internal cavity, the multidimensional sensor array is integrated at the front end of the probe housing, and the signal conditioning circuit is electrically connected to the multidimensional sensor array for amplifying, filtering, and converting the acquired analog signal to digital.

[0009] The microprocessor unit has a built-in neural network acceleration engine and an integrated finite state machine. The finite state machine controls the microprocessor unit to switch between normal detection mode and self-learning update mode according to an external trigger signal. In normal detection mode, the neural network acceleration engine loads and executes the fault diagnosis algorithm firmware to perform feature extraction and fault mode classification on the signals collected by the multi-dimensional sensor array. In self-learning update mode, the neural network acceleration engine loads and executes the self-learning update algorithm firmware to perform gradient calculation and parameter update in the incremental learning process.

[0010] The microprocessor unit also executes a dynamic priority detection strategy: calculating a priority score based on the uncertainty of the fault diagnosis results. , The calculation formula is ,in Score the severity of the fault. Scoring based on the frequency of failure occurrence. Scoring for fault uncertainty, For the weighting coefficients, satisfying Scoring based on the stated priority The execution order of subsequent testing items can be dynamically adjusted.

[0011] Furthermore, the multidimensional sensing array includes a vibration sensor, a temperature sensor, and a current sensor. The vibration sensor is used to collect mechanical vibration signals during the operation of the medical device. The temperature sensor is used to collect the surface temperature or junction temperature of key components of the device. The current sensor is used to collect the operating current waveform of the device circuit. The multidimensional sensing array also includes a noise sensor and a speed sensor. The noise sensor is used to collect acoustic characteristic signals during the operation of the device. The speed sensor is used to collect the speed signals of the motor or rotating parts. Each sensor adopts a synchronous sampling mechanism, and the sampling frequency satisfies the Shannon sampling theorem. The sampling rate of the vibration sensor is not less than 10kHz, the resolution of the temperature sensor is not less than 0.1℃, and the bandwidth of the current sensor is not less than 5kHz.

[0012] Furthermore, the storage module is located inside the probe housing and is used to store historical maintenance data, feature libraries, and algorithm model parameters; the communication interface is located at the rear end of the probe housing and is used to establish a data and power connection with the maintenance control terminal; the display module is located on the side wall of the probe housing and is used to display detection values ​​and fault diagnosis results in real time. The storage module adopts a redundant storage architecture, including a main storage area and a backup storage area; the main storage area stores real-time detection data and current algorithm model parameters, and the backup storage area stores historical maintenance data; when an anomaly is detected in the main storage area, the system automatically switches to the backup storage area to read data.

[0013] Furthermore, the signal conditioning circuit includes a preamplifier, an anti-aliasing filter, and a 24-bit high-precision analog-to-digital converter. The sampling rate of the analog-to-digital converter is set to 25.6kHz, supporting multi-channel synchronous sampling, and the crosstalk suppression ratio of each channel is not less than 80dB.

[0014] Furthermore, the microprocessor unit is a dedicated maintenance processor SoC that integrates a multi-core CPU and an AI accelerator. This SoC has a computing power of ≥4 TOPS and a built-in FPGA coprocessor. Its internal partitions store fault diagnosis algorithm firmware, feature extraction algorithm firmware, and self-learning update algorithm firmware.

[0015] Furthermore, the microprocessor unit performs adaptive threshold calculation: it calculates the average value of each sensor signal based on historical maintenance data. and standard deviation The fault determination threshold is dynamically updated, and the threshold calculation formula is as follows: ,in The confidence coefficient ranges from 2.5 to 4.0 and is adaptively adjusted based on the equipment model and the operating conditions of the equipment. When the real-time acquired signal exceeds the threshold T, a fault alarm is triggered.

[0016] Furthermore, the main body of the inspection probe also includes a positioning module, which is electrically connected to the communication interface; in positioning mode, the main body of the inspection probe receives a position query signal sent by the inspection control terminal, and returns three-dimensional coordinates through the built-in UWB ultra-wideband positioning chip to achieve accurate tracking of the probe position.

[0017] Furthermore, the microprocessor unit performs a device health assessment based on the data collected by the multi-dimensional sensor array: calculating a comprehensive health index. , The calculation formula is ,in For the first The normalized value of the health status of the test indicators For the first The weight coefficients of the indicators satisfy the following conditions: The health index H ranges from 0 to 100. When the value falls below a preset threshold, a preventative maintenance reminder is generated.

[0018] The AI-based automatic learning medical equipment maintenance system includes the AI-based automatic learning medical equipment maintenance device and a maintenance control terminal, wherein the maintenance control terminal includes a main control chip, a graphics rendering engine and a wireless communication module.

[0019] The maintenance control terminal receives the detection data and fault diagnosis results sent by the maintenance equipment through the communication interface and communication line, and drives the graphics rendering engine to generate a visual maintenance report.

[0020] The maintenance control terminal also includes a touch screen, which is electrically connected to the main control chip. The touch screen supports two display modes. When it receives normal detection data sent by the maintenance equipment, the main control chip configures the display mode to data monitoring mode, displaying real-time data in waveform and numerical form. When it receives a fault alarm signal, the main control chip configures the display mode to fault warning mode and triggers an audible and visual alarm.

[0021] The maintenance control terminal also integrates a fault knowledge graph database. When the main control chip receives the fault diagnosis result, it automatically associates with the fault knowledge graph database to generate a maintenance plan that includes fault cause analysis, handling suggestions and similar historical cases, and presents it on the touch screen.

[0022] Furthermore, the maintenance control terminal communicates with the cloud server through the wireless communication module; in the self-learning update mode, the maintenance control terminal uploads the detection data stored locally on the maintenance equipment to the cloud server, receives the updated algorithm model parameters returned by the cloud server, and then transmits the algorithm model parameters downlink to the maintenance equipment.

[0023] The wireless communication module supports at least one of Wi-Fi, Bluetooth, or 5G communication protocols; the main control chip is an ARM architecture processor running a Linux or Android operating system; and the graphics rendering engine runs in software on the main control chip.

[0024] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention integrates a multi-dimensional sensing array into the front end of the probe housing and incorporates a microprocessor unit with an integrated neural network acceleration engine within the probe housing. This allows the microprocessor unit to load and execute fault diagnosis algorithm firmware in normal detection mode and self-learning update algorithm firmware in self-learning update mode, thus achieving dual-mode operation of the maintenance equipment under a single probe architecture. By reusing the computing resources of the neural network acceleration engine on the same hardware platform, the maintenance probe body can complete the entire workflow from data acquisition and fault diagnosis to model update without requiring equipment replacement or connection to external computing units at the maintenance site, reducing the hardware cost of the maintenance equipment and the complexity of on-site operation.

[0025] 2. This invention integrates a finite state machine within a microprocessor unit, which controls the microprocessor unit to switch between normal detection mode and self-learning update mode based on external trigger signals. This ensures that the transition between the two operating modes has definite trigger conditions and clear logical transition relationships. The introduction of this finite state machine avoids algorithm mis-execution or task scheduling conflicts caused by the uncertainty of mode states. The finite state machine performs mode switching based on the level of the external trigger signal or the command frame type, ensuring that the neural network acceleration engine only loads and executes the algorithm firmware corresponding to the current mode at any given time. This avoids storage resource competition and power consumption superposition caused by two algorithms residing in memory simultaneously, ensuring the atomicity and reliability of the mode switching process. This allows the inspection probe body to maintain a long working time under limited battery power conditions.

[0026] 3. This invention performs adaptive threshold calculation in the microprocessor unit, which calculates the mean and standard deviation of each sensor signal based on historical maintenance data, dynamically updates the fault judgment threshold, and adaptively adjusts it according to the equipment model and aging conditions. The real-time acquired signal is then compared with this dynamic threshold to trigger a fault alarm. By dynamically adjusting the judgment boundary based on the statistical distribution of historical data, the fault judgment threshold is updated to follow changes in the current state of the equipment. This avoids false alarms when the normal operating parameters of the equipment fluctuate within a reasonable range, and promptly detects anomalies when the parameters deviate from the normal statistical distribution, thus controlling the false alarm rate while ensuring the fault detection rate. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1This is a structural framework diagram of a medical equipment repair system based on AI automatic learning. Figure 2 This is a structural diagram of the main body of the inspection probe in an AI-based automatic learning medical equipment inspection device. Figure 3 This is a structural framework diagram of a multi-dimensional sensor array in a medical equipment maintenance device based on AI automatic learning. Figure 4 This is a structural framework diagram of the signal conditioning circuit in a medical equipment repair device based on AI automatic learning. Figure 5 This is a structural framework diagram of the microprocessor unit in a medical equipment repair device based on AI automatic learning. Figure 6 This is a structural framework diagram of the storage module in a medical equipment maintenance device based on AI automatic learning. Figure 7 This is a structural framework diagram of a medical equipment maintenance system based on AI automatic learning.

[0029] Explanation of reference numerals in the attached figures: 1. Inspection probe body; 101. Probe housing; 102. Multi-dimensional sensor array; 1021. Vibration sensor; 1022. Temperature sensor; 1023. Current sensor; 1024. Noise sensor; 1025. Speed ​​sensor; 103. Signal conditioning circuit; 1031. Preamplifier; 1032. Anti-aliasing filter; 1033. 24-bit high-precision analog-to-digital converter; 104. Microprocessor unit; 105. Storage module; 106. Communication interface; 107. Positioning module; 108. Display module; 2. Inspection and control terminal; 201. Main control chip; 202. Touch screen display; 203. Graphics rendering engine; 204. Fault knowledge graph database; 205. Wireless communication module; 206. Cloud server. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments provide the structural connection, data connection and operation process between the inspection probe body 1, the inspection control terminal 2 and the cloud server 206. Those skilled in the art can manufacture and operate the medical device inspection equipment and system of the present invention according to the structural dimensions, interface form, sampling parameters and processing steps.

[0031] Please see Figures 1 to 7The AI-based automatic learning medical device repair equipment and system in this embodiment includes a repair probe body 1. The repair probe body 1 is a cylindrical handheld structure extending in the front-to-back direction, with a length that can be set from 180mm to 240mm and a maximum outer diameter that can be set from 38mm to 55mm; in this embodiment, the length is 210mm and the maximum outer diameter is 46mm. A mounting frame extending along the axis is provided in the receiving cavity. The front end of the mounting frame carries a multi-dimensional sensor array 102 and a signal conditioning circuit 103, and the middle section of the mounting frame carries a microprocessor unit 104 and a storage module 105. The rear end carries the communication interface 106 and the positioning module 107; the display module 108 is embedded in the side wall opening of the probe housing 101, and its display surface is flush with or 0.2mm to 0.8mm lower than the outer surface of the probe housing 101. A transparent cover plate covers the display surface and is bonded to the periphery of the side wall opening; the communication interface 106 is located at the rear end of the probe housing 101, and the interface insertion direction is parallel to the axis of the inspection probe body 1. When the maintenance personnel hold the middle section of the probe housing 101, the communication line is led out from the back of the palm, and the multi-dimensional sensor array 102 is located on the front of the palm and faces the detection part of the medical device to be inspected.

[0032] like Figure 1 and Figure 2 As shown, the multidimensional sensor array 102 is mounted on a sensor mounting base at the front end of the probe housing 101. The sensor mounting base is connected to the front end of the mounting frame by four circumferentially distributed screws. An insulating gasket with a thickness of 1mm to 3mm is sandwiched between the sensor mounting base and the mounting frame. The vibration sensor 1021 is arranged at the center of the sensor mounting base. The vibration sensor 1021 can be a piezoelectric accelerometer or a triaxial MEMS accelerometer. In this embodiment, a triaxial MEMS accelerometer with a range of ±16g is used. One of its sensitive axes coincides with the axis of the inspection probe body 1, and the other two axes are parallel to the probe housing 101. 1. The horizontal and vertical directions; the front side of the vibration sensor 1021 is connected to a circular metal contact head. The metal contact head passes through the guide hole in the center of the front cover and extends 2mm to 5mm out of the front cover. A compression spring is sleeved on the rear side of the metal contact head. One end of the compression spring abuts against the limiting shoulder of the metal contact head, and the other end abuts against the sensor mounting base. During maintenance, the metal contact head abuts against the detection surface of the motor housing, bearing housing, pump body, or fan cover. The compression spring is axially compressed, and the vibration is transmitted to the vibration sensor 1021 through the metal contact head. An elastic anti-slip ring is sleeved on the outer periphery of the metal contact head. When the elastic anti-slip ring contacts the surface to be inspected, it restricts the probe from sliding laterally along the surface.

[0033] Temperature sensor 1022 is disposed outside vibration sensor 1021. The sensing end of temperature sensor 1022 is connected to an annular heat-conducting plate, which surrounds the metal contact head and protrudes from the annular opening of the front cover. An insulating spacer is provided between the annular heat-conducting plate and the metal contact head to separate the heat conduction path of temperature sensor 1022 from the mechanical conduction path of vibration sensor 1021. Temperature sensor 1022 can be a platinum resistance thermometer, thermocouple, or digital temperature chip; this embodiment uses a Pt1000 platinum resistance thermometer. The temperature range is -20℃ to 150℃, with a resolution of 0.1℃. After the annular heat-conducting sheet is attached to the surface of the device, heat is transferred to the platinum resistance along the heat-conducting sheet. The signal conditioning circuit 103 provides constant current excitation to the platinum resistance and reads the voltage across its terminals. In order to reduce the proportion of heat transferred from the handheld device to the temperature sensor 1022 along the probe housing 101, the insulating gasket between the sensor mounting base and the mounting frame is made of a material with a thermal conductivity of less than 0.3W / (m·K). After contact, the temperature channel first records the environmental baseline, and then records the continuous temperature value and the temperature rise per unit time.

[0034] The current sensor 1023 is arranged on one side of the sensor mounting base. The current sensor 1023 adopts a Hall closed-loop current detection structure. The measurement bandwidth of the current sensor 1023 is not less than 5kHz, and the range can be set to ±5A, ±20A or ±50A according to the object under maintenance. The signal conditioning circuit 103 selects the corresponding gain parameter according to the range. When the openable magnetic ring is not connected, the microprocessor unit 104 reads the interface identification resistor of the current channel, and the current channel status is marked as not connected. The fault diagnosis algorithm does not write the no-load value of the channel into the feature vector.

[0035] Noise sensor 1024 is located near the outer periphery of the front cover. Noise sensor 1024 uses a microelectromechanical microphone, and its acquisition direction faces the contact direction of vibration sensor 1021. After the inspection probe body 1 is against the equipment, the distance between the acoustic through-hole and the equipment surface is maintained at 10mm to 30mm. Speed ​​sensor 1025 is located on the other side of the front cover. Speed ​​sensor 1025 includes an infrared emitter and a photoelectric receiver. The optical axes of the infrared emitter and the photoelectric receiver form an overlapping area on the outer side of the front cover. After reflective marks are placed on the rotating shaft, fan impeller, or pulley, the reflective marks pass through once per revolution. In the overlapping area, the photodetector outputs a pulse; for rotating parts with magnetic markings, the speed sensor 1025 can also be a Hall switch, and the output terminal of the Hall switch is connected to the pulse input terminal of the signal conditioning circuit 103; the vibration sensor 1021, temperature sensor 1022, current sensor 1023, noise sensor 1024 and speed sensor 1025 are distributed around the sensor mounting base, and the leads of each sensor pass through the independent wire holes on the back of the sensor mounting base and are connected to the signal conditioning circuit 103. The vibration channel and the current channel use shielded wires, and the shielding layer is connected at a single point to the analog ground terminal of the signal conditioning circuit 103.

[0036] like Figure 3 and Figure 4 As shown, the signal conditioning circuit 103 is installed behind the multi-dimensional sensor array 102. The signal conditioning circuit 103 adopts a four-layer or six-layer printed circuit board. The analog signal area and the digital signal area are respectively located on both sides of the circuit board, and the two are connected at the reference ground position of the 24-bit high-precision analog-to-digital converter 1033. The preamplifier 1031 includes a differential amplification unit and a programmable gain unit corresponding to each sensor channel. The gain of the vibration channel can be switched between 1x, 2x, 4x and 8x, the gain of the noise channel can be switched between 10x and 40x, the gain of the current channel is switched according to the range of the current sensor 1023, the temperature channel adopts a constant current source and an instrumentation amplifier, and the speed channel adopts a Schmitt trigger shaping circuit. Each analog input terminal is connected in series with a current limiting resistor and in parallel with a transient suppression device. The transient voltage generated when the sensor is connected or disconnected is limited to the input range of the preamplifier 1031.

[0037] An anti-aliasing filter 1032 is connected between the preamplifier 1031 and the 24-bit high-precision analog-to-digital converter 1033. The vibration and noise channels employ fourth-order low-pass filters with cutoff frequencies set to 8kHz and 10kHz, respectively; the current channel's cutoff frequency is set to 5kHz; and the temperature channel's cutoff frequency is set to 5Hz. The rotational speed pulse channel is first shaped and then directly input to the counting input of the microprocessor unit 104, while simultaneously sending the pulse status to the 24-bit high-precision analog-to-digital converter 1033 via a low-speed channel. The 24-bit high-precision analog-to-digital converter 1033 employs a multi-channel synchronous sampling structure with a main sampling rate set to 25.6kHz. Vibration and noise... The current channel is output according to the main sampling rate, and the temperature channel is sampled and processed by the microprocessor unit 104 to form 10 to 100 temperature samples per second. The sampling clock of the 24-bit high-precision analog-to-digital converter 1033 is output by the FPGA coprocessor in the microprocessor unit 104. Each analog channel is latched simultaneously on the same sampling trigger edge, and the speed pulse counter reads the current count value on the trigger edge. The data frames of each channel are written to the same timestamp. The signal conditioning circuit 103 is connected to the microprocessor unit 104 via SPI, parallel data bus or low voltage differential bus. In this embodiment, four-wire SPI is used to transmit control commands and eight-bit parallel bus is used to transmit sampling data.

[0038] like Figure 5 As shown, the microprocessor unit 104 is mounted on the main circuit board in the middle section of the probe housing 101. The main circuit board and the signal conditioning circuit 103 are arranged along the probe axis and connected by a board-to-board connector. The microprocessor unit 104 is a maintenance processor SoC integrating a multi-core CPU, a neural network acceleration engine, and an FPGA coprocessor, with a neural network computing power of no less than 4 TOPS. The FPGA coprocessor receives the data output from the 24-bit high-precision analog-to-digital converter 1033, writes it into the circular buffer according to the channel number, and performs DC component removal, digital downsampling, window function multiplication, and data rearrangement before fast Fourier transform. The multi-core CPU manages the detection task, display interface, communication protocol, storage read / write, and state machine. The neural network acceleration engine receives the normalized feature matrix and performs convolution, matrix multiplication, activation, and classification output. The front end of the main circuit board is connected to the signal conditioning circuit 103, the side is connected to the display module 108 via a flexible ribbon cable, the rear end is connected to the communication interface 106 via a differential line, and the storage module 105 and the positioning module 107 are connected via an SPI bus.

[0039] The finite state machine inside the microprocessor unit 104 includes at least a normal detection state and a self-learning update state. After the maintenance probe body 1 is powered on, the finite state machine first enters the initialization process, reads the device configuration table, sensor calibration coefficients, current model version number and last power-off flag in the storage module 105, and calculates the cyclic redundancy check value for the fault diagnosis algorithm firmware, feature extraction algorithm firmware and self-learning update algorithm firmware respectively. When the current model check value is consistent with the version table, the finite state machine enters the normal detection state. When the maintenance control terminal 2 sends an update command and the maintenance probe body 1 is not currently performing a sampling task, the finite state machine enters the self-learning update state. When the update is completed or the update is terminated, the finite state machine returns to the normal detection state. During the state transition, the FPGA coprocessor stops writing new frames to the model input buffer, the data that has been collected is written to the storage module 105, and the display module 108 displays the current status and model version.

[0040] like Figure 6 As shown, the storage module 105 is located on the back or adjacent to the main circuit board where the microprocessor unit 104 is located. The storage module 105 can adopt a combination structure of embedded multimedia memory and serial flash memory. The storage module 105 is divided into a main storage area and a backup storage area. The main storage area stores the current fault diagnosis model, current feature extraction parameters, real-time detection data and task logs for the day. The backup storage area stores the previous version model, historical maintenance data, fault labels confirmed by maintenance personnel, and model update logs. Each detection record includes the device number, device model, detection point number, coordinates returned by the positioning module 107, sampling start and end time, sensor channel status, raw data index, feature vector, fault category probability, threshold comparison result, and health index. The record header is written with a length field and a verification field. The microprocessor unit 104 first writes the record content and then writes the record completion mark. Records without a completion mark will not be included in the training sample set during the next startup. When the verification of the main storage area is inconsistent after two consecutive reads, the microprocessor unit 104 reads the previous version model and the most recent complete task table in the backup storage area and displays the storage switching status on the display module 108.

[0041] Communication interface 106, etc. Figure 1 and Figure 2The interface 106 is located at the rear end of the probe housing 101. In this embodiment, a USB Type-C interface is used. The interface metal shell is connected to the shielding plate inside the probe housing 101. The power pin is connected to the power management circuit, and the differential data pin is connected to the USB physical layer of the microprocessor unit 104. The communication interface 106 is connected to the maintenance control terminal 2 via a communication line. The maintenance control terminal 2 provides 5V DC power to the maintenance probe body 1 and establishes a bidirectional data channel. The uplink data includes raw sampling data, feature data, fault classification results, health index, positioning coordinates, and operation log. The downlink data includes the detection task table, sensor parameters, model update package, location query signal, and time synchronization command. Each data frame includes a frame header, command word, task number, data length, timestamp, data body, and checksum. After receiving the data, the maintenance control terminal 2 returns an acknowledgment frame. If no acknowledgment frame is received within a specified time, the maintenance probe body 1 retransmits the corresponding data frame from the transmission buffer.

[0042] The positioning module 107 is installed at the rear end of the probe housing 101 near the non-metallic outer wall, and is axially spaced 15mm to 35mm from the communication interface 106. The positioning module 107 includes a UWB ultra-wideband positioning chip, a clock circuit, and an antenna. The antenna radiating surface faces the outside of the probe housing 101, and a clearance is maintained between the antenna and the metal shielding area of ​​the main circuit board. The positioning module 107 is connected to the microprocessor unit 104 via SPI or UART. After the maintenance control terminal 2 sends a location query signal, the microprocessor unit 104 controls the positioning module 107 to exchange ranging messages with the UWB base stations set in the maintenance area. Based on the ranging results of at least three base stations, the three-dimensional coordinates of the maintenance probe body 1 are calculated, and the coordinates are bound to the current detection point number. The detection point coordinates of the medical device are pre-stored in the task file of the maintenance control terminal 2. When the distance between the measured coordinates and the target coordinates is greater than the set tolerance, the touch screen 202 displays the measured position and the target position. After the maintenance personnel move the maintenance probe body 1, they initiate a location query again.

[0043] The display module 108 is located on the upper side of the probe housing 101 or on a side that is easy to observe. In this embodiment, a 2.4-inch organic light-emitting diode display screen is used, and its interface is connected to the microprocessor unit 104 via a flexible ribbon cable. Under normal detection conditions, the display module 108 sequentially displays the device number, detection point number, connected sensor channels, root mean square value of vibration, temperature value, effective current value, noise sound pressure level, rotational speed value, fault type and classification probability. Under self-learning update conditions, the display module 108 displays the model version to be updated, the number of samples read, parameter writing progress and verification results. A confirmation button and a return button without the attached icon can be set below the display module 108. The button contacts are connected to the general input / output terminals of the microprocessor unit 104. Maintenance personnel can start sampling, stop sampling or confirm the field label by pressing the button.

[0044] like Figure 7 As shown, the maintenance control terminal 2 includes a main control chip 201, a touch screen display 202, a graphics rendering engine 203, a fault knowledge graph database 204, and a wireless communication module 205. The main control chip 201 is installed on the internal motherboard of the maintenance control terminal 2. The touch screen display 202 is connected to the main control chip 201 through a display interface and a touch interface. The communication cable connected to the communication interface 106 is inserted into the host interface of the maintenance control terminal 2. The main control chip 201 reads the data sent by the maintenance probe body 1 through the host controller. The graphics rendering engine 203 runs on the main control chip 201 as a program module, processing the time domain data, frequency domain data, and trend data of each sensor. The classification probabilities are converted into waveforms, spectrum graphs, trend curves, and component status diagrams. The fault knowledge graph database 204 is stored in the solid-state memory of the maintenance control terminal 2. The database nodes include medical equipment models, component names, fault modes, characteristic intervals, maintenance actions, and historical cases. The nodes are connected through belonging, association, cause, and treatment relationships. The wireless communication module 205 is connected to the main control chip 201 through PCIe, USB, or serial bus. In this embodiment, a communication module that supports both Wi-Fi and 5G is used. The maintenance control terminal 2 exchanges maintenance samples, equipment dictionaries, and model update packages with the cloud server 206 through the wireless communication module 205.

[0045] Before performing the maintenance task, the maintenance personnel select the model and equipment number of the medical device to be inspected on the touch screen 202. The main control chip 201 reads the corresponding component list, detection point coordinates, sensor combination, sampling rate, sampling duration and initial threshold from the fault knowledge graph database 204, forms a test task table and sends it to the maintenance probe body 1 through the communication interface 106. After receiving the task table, the microprocessor unit 104 displays the task number and model version on the display module 108, and adjusts the gain and anti-aliasing of the preamplifier 1031 according to the task table. The channel parameters of the superimposed filter 1032 and the sampling channel of the 24-bit high-precision analog-to-digital converter 1033 are configured; the maintenance personnel move the maintenance probe body 1 to the target detection point, the positioning module 107 returns the coordinates, and after the coordinates fall within the tolerance range of the target detection point, the metal contact head on the front side of the vibration sensor 1021 and the annular heat-conducting plate on the front side of the temperature sensor 1022 are pressed onto the detection surface, the opening and closing magnetic ring of the current sensor 1023 is placed on the target wire, and the noise sensor 1024 and the speed sensor 1025 are oriented toward the corresponding sound source and rotation mark.

[0046] After sampling begins, the microprocessor unit 104 sends a synchronous trigger signal to the 24-bit high-precision analog-to-digital converter 1033 and the speed counter. Vibration, temperature, current, noise, and speed data enter the circular buffer according to the task number. For stable operation detection, the single sampling duration can be set to 2 to 10 seconds. For startup process detection, 1 second of pre-trigger data can be retained before the medical device starts, and data can be continuously collected until 10 to 30 seconds after startup. The graphics rendering engine 203 displays a scrolling waveform during sampling, and the microprocessor unit 104 calculates the corresponding feature after each analysis window is completed. Vibration channel meter The system calculates the mean, root mean square value, peak value, peak-to-peak value, peak factor, kurtosis, dominant frequency, harmonic amplitude, and band energy; the temperature channel calculates the contact temperature, relative ambient temperature rise, and temperature rise rate; the current channel calculates the effective value, peak value, waveform factor, fundamental amplitude, harmonic amplitude, and starting current duration; the noise channel calculates the sound pressure level, spectral centroid, narrowband peak value, and specified band energy; the speed channel calculates the instantaneous speed, average speed, and speed fluctuation rate based on the interval between adjacent pulses; each feature is normalized according to the dimensional range, mean, and standard deviation in storage module 105, and then arranged into a feature matrix in a fixed channel order.

[0047] The fault diagnosis algorithm can use a multi-branch one-dimensional convolutional neural network. Vibration, current and noise data are fed into the corresponding convolutional branches, while temperature and speed features are fed into fully connected branches. The outputs of each branch are connected in a fusion layer, and a classification layer is set after the fusion layer. The neural network acceleration engine performs inference on each analysis window according to the current model parameters and outputs the probabilities of categories such as normal, bearing wear, rotor imbalance, insufficient lubrication, abnormal heat dissipation, abnormal power supply, load blockage and abnormal sensor contact. The microprocessor unit 104 reads the category corresponding to the highest probability as the current candidate category.

[0048] Microprocessor unit 104 calculates a dynamic priority score for each detection point. , ,in Score the severity of the fault. Scoring is done on the frequency of failures in the same model of equipment or in the historical records of the same equipment. Scoring for fault uncertainty, and The weights are non-negative and satisfy the following conditions: ; , and All values ​​are normalized to 0 to 1. In this embodiment, the values ​​can be taken as follows: 0.5 0.2 The value is 0.3; after the initial scan is completed, the microprocessor unit 104 sends the P value of each detection point to the maintenance control terminal 2, and the main control chip 201 rearranges the unfinished tasks according to the P value from largest to smallest; when When the value is greater than 0.7, the task schedule will increase the sampling duration of that detection point to twice the original duration and require at least one retest; when When the value is greater than 0.5, the task table adds sensor channels related to candidate faults. For example, when vibration and noise classifications are close, a speed channel is added; when current anomalies and load stagnation classifications are close, a vibration channel and a startup process sampling channel are added.

[0049] The adaptive threshold is calculated from historical normal samples in storage module 105; for any feature The microprocessor unit 104 reads the most recent data within the same device model, the same detection point, and the same load range. The second feature value was confirmed to be normal. A range of 30 to 200 can be used. First, outliers that exceed the absolute deviation of the median are removed, and then the mean is calculated. and standard deviation Upper limit threshold lower threshold ,in Use a value between 2.5 and 4.0; when the equipment service time increases or the historical fluctuation range expands, the corresponding task configuration table can be selected. Value; features with only an upper limit are adopted. Features with only a lower limit are adopted Features subject to both upper and lower limits are compared simultaneously. and After a real-time feature exceeds a threshold, the microprocessor unit 104 records the direction of the boundary violation, the magnitude of the boundary violation, and the number of continuous windows, and sends them to the maintenance control terminal 2 along with the neural network classification results.

[0050] Equipment health assessment is performed after all testing points for the current equipment have been completed; for the first... The microprocessor unit 104 calculates a normalized health score based on the deviation of the test indicator from the normal range. , The range is 0 to 100. Values ​​within the normal range are taken as close to 100, while values ​​deviating to the maximum set deviation in the task schedule are taken as 0. (Overall Health Index) ,in For the first The normalized value of the health status of the test indicators For the first The weight coefficients of the indicators satisfy the following conditions: The main control chip 201 can set weights according to the importance level of equipment components. The weights of power supply, drive motor and safety interlock related indicators are greater than those of housing temperature rise or low-level noise indicators. Value, minimum The corresponding detection point and in multiple consecutive tasks The changes should be recorded in the maintenance report; When the value is below the task table threshold, the touch display 202 enters the fault warning mode and displays the corresponding component, out-of-bounds characteristics, and candidate fault categories.

[0051] After receiving the fault diagnosis result, the main control chip 201 accesses the fault knowledge graph database 204 using the equipment model, detection point number, candidate fault category, main out-of-bounds characteristics, and location coordinates as query conditions. The database returns the fault cause node, inspection action node, and historical case node connected to the query conditions. The graphics rendering engine 203 displays the returned content side by side with the real-time waveform on the touch screen 202. After the maintenance personnel complete the disassembly or replacement, they select the actual fault label on the touch screen 202 and write the label, processing action, and retest result back to the current detection record. The retest data and the data before processing are collected using the same task parameters. The main control chip 201 saves the two records together. When the processed characteristics return to the normal range, the fault label is marked as a confirmed sample. If the out-of-bounds condition still occurs after processing, the record is marked as a sample to be reviewed and is not directly entered into the local incremental learning dataset.

[0052] Self-learning updates can be initiated by maintenance personnel via commands on the touchscreen 202, or by commands issued by the maintenance control terminal 2 when the number of confirmed samples reaches a set value and the maintenance probe body 1 is in a non-sampling state. After the microprocessor unit 104 enters the self-learning update state, it reads the confirmed samples from the backup storage area, groups them according to the fault category, and performs time window segmentation, normalization, and class balancing on the samples. The parameters of the current model's feature extraction layer remain unchanged, while the parameters of the classification layer or incremental adaptation layer are updated. The training data and validation data can be divided in an 8:2 ratio, the batch size can be 8 to 32, and the learning rate can be 0.00001 to 0. 001, the number of update rounds can be 5 to 20; the neural network acceleration engine performs forward computation and gradient computation, and the multi-core CPU performs sample scheduling, loss computation and parameter writing; the updated model is first written to the inactive model partition of storage module 105, and inference is run on the retained validation set. The microprocessor unit 104 records the classification accuracy, recall, false positive rate and confusion matrix of each class; when the validation index reaches the switching condition in the task configuration table, the model version table points to the new model partition. When the validation index does not reach the switching condition, the model version table still points to the original model partition. The new model record is retained in the update log for the maintenance control terminal 2 to read.

[0053] During the cloud update process, the maintenance control terminal 2 reads the confirmed sample index and corresponding data from the maintenance probe body 1, replaces the equipment number with a random task identifier, and uploads it to the cloud server 206 via the wireless communication module 205. The cloud server 206 aggregates data from different maintenance tasks according to the equipment model and component type, performs version checks and category statistics on the samples, and trains a new model parameter package. The model parameter package includes the model structure identifier, parameter data, applicable equipment model, minimum firmware version, version number, hash value, and digital signature. After receiving the model parameter package, the maintenance control terminal 2 first compares the equipment model and firmware version, and then verifies the hash value and digital signature. After the verification is successful, the model parameter package is transmitted to the inactive model partition of the maintenance probe body 1 via the communication interface 106. After the data transmission is completed, the microprocessor unit 104 uses the reference input vector provided with the package to perform inference and compares the reference output. When the output is consistent, the model version table is updated. When communication is interrupted, verification is inconsistent, or the reference output is inconsistent, the current model pointer is not changed.

[0054] Taking the maintenance of the turbine motor inside a medical ventilator as an example, the maintenance control terminal 2 reads the coordinates and task parameters of the front bearing housing, rear bearing housing, power input line, and impeller speed detection point of the turbine motor from the fault knowledge graph database 204; the maintenance personnel press the vibration sensor 1021 and temperature sensor 1022 onto the surface of the front bearing housing, put the openable magnetic ring of the current sensor 1023 around the outer periphery of the motor power line, make the noise sensor 1024 face the air intake side of the turbine housing, and make the speed sensor 1025 align with the reflective mark on the impeller shaft end; after the coordinates returned by the positioning module 107 are bound to the coordinates of the front bearing housing detection point, the microprocessor unit 104 collects 10 seconds of steady-state data and one start-up process data; if the vibration spectrum shows a side band near the rotational frequency harmonic, the noise channel shows a peak in the same frequency band, and the effective value of the current and the speed fluctuation rate increase simultaneously, the fault diagnosis algorithm outputs the highest probability of the bearing wear category, and the microprocessor unit 104 calculates according to the probability distribution. The value is calculated based on the fault level and historical frequency. The value is set, and the test point is placed at the beginning of the retest task. After the maintenance personnel replace the bearing, the test is repeated at the same position and with the same parameters. The data before and after the processing, the fault labels and maintenance actions are written into the same case relationship. The confirmed sample can participate in the classification layer parameter update in the subsequent self-learning update.

[0055] The structural dimensions, sampling rate, filter cutoff frequency, model type, training parameters, and threshold range in this embodiment are executable configuration examples. Finally, it should be noted that the above embodiments are used to illustrate the technical solutions of the present invention, rather than limiting the scope of protection. Without changing the structural connection relationship and data processing relationship, any substitutions made to the sensor model, housing material, interface form, and algorithm parameters can fall within the implementation scope of the corresponding technical solutions.

Claims

1. A medical equipment repair device based on AI automatic learning, characterized in that: The device includes a maintenance probe body (1), which includes a probe housing (101), a multi-dimensional sensor array (102), a signal conditioning circuit (103), a microprocessor unit (104), a storage module (105), a communication interface (106), and a display module (108). The probe housing (101) has an internal cavity, the multidimensional sensor array (102) is integrated at the front end of the probe housing (101), and the signal conditioning circuit (103) is electrically connected to the multidimensional sensor array (102) for amplifying, filtering and converting the acquired analog signal into digital signal. The microprocessor unit (104) has a built-in neural network acceleration engine and an integrated finite state machine. The finite state machine controls the microprocessor unit (104) to switch between normal detection mode and self-learning update mode according to an external trigger signal. In normal detection mode, the neural network acceleration engine loads and executes the fault diagnosis algorithm firmware to perform feature extraction and fault mode classification on the signals collected by the multi-dimensional sensor array (102). In self-learning update mode, the neural network acceleration engine loads and executes the self-learning update algorithm firmware to perform gradient calculation and parameter update in the incremental learning process. The microprocessor unit (104) also executes a dynamic priority detection strategy: calculating a priority score based on the uncertainty of the fault diagnosis result. , The calculation formula is ,in Score the severity of the fault. Scoring based on the frequency of failure occurrence. Scoring for fault uncertainty, For the weighting coefficients, satisfying Scoring based on the stated priority The execution order of subsequent testing items can be dynamically adjusted.

2. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The multidimensional sensor array (102) includes a vibration sensor (1021), a temperature sensor (1022), and a current sensor (1023). The vibration sensor (1021) is used to collect mechanical vibration signals during the operation of the medical device. The temperature sensor (1022) is used to collect the surface temperature or junction temperature of key components of the device. The current sensor (1023) is used to collect the operating current waveform of the device circuit. The multidimensional sensor array (102) also includes a noise sensor (1024) and a speed sensor (1025). The noise sensor (1024) is used to collect acoustic characteristic signals during the operation of the device. The speed sensor (1025) is used to collect the speed signals of the motor or rotating parts. Each sensor adopts a synchronous sampling mechanism, and the sampling frequency satisfies the Shannon sampling theorem. The sampling rate of the vibration sensor (1021) is not less than 10kHz, the resolution of the temperature sensor (1022) is not less than 0.1℃, and the bandwidth of the current sensor (1023) is not less than 5kHz.

3. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The storage module (105) is located inside the probe housing (101) and is used to store historical maintenance data, feature library and algorithm model parameters; the communication interface (106) is located at the rear end of the probe housing (101) and is used to establish a data and power connection with the maintenance control terminal (2); the display module (108) is located on the side wall of the probe housing (101) and is used to display the detection values ​​and fault diagnosis results in real time. The storage module (105) adopts a redundant storage architecture, including a main storage area and a backup storage area; the main storage area stores real-time detection data and current algorithm model parameters, and the backup storage area stores historical maintenance data; when an abnormality is detected in the main storage area, the system automatically switches to the backup storage area to read the data.

4. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The signal conditioning circuit (103) includes a preamplifier (1031), an anti-aliasing filter (1032), and a 24-bit high-precision analog-to-digital converter (1033). The sampling rate of the analog-to-digital converter is set to 25.6kHz, supports multi-channel synchronous sampling, and the crosstalk suppression ratio of each channel is not less than 80dB.

5. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The microprocessor unit (104) is a dedicated maintenance processor SoC that integrates a multi-core CPU and an AI accelerator. The SoC has a computing power of ≥4 TOPS and a built-in FPGA coprocessor. Its internal partition stores fault diagnosis algorithm firmware, feature extraction algorithm firmware and self-learning update algorithm firmware.

6. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The microprocessor unit (104) performs adaptive threshold calculation: it calculates the average value of each sensor signal based on historical maintenance data. and standard deviation The fault determination threshold is dynamically updated, and the threshold calculation formula is as follows: ,in The confidence coefficient ranges from 2.5 to 4.0 and is adaptively adjusted based on the equipment model and the operating conditions of the equipment. When the real-time acquired signal exceeds the threshold T, a fault alarm is triggered.

7. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The main body (1) of the inspection probe also includes a positioning module (107), which is electrically connected to the communication interface (106). In positioning mode, the main body (1) of the inspection probe receives the position query signal sent by the inspection control terminal and returns the three-dimensional coordinates through the built-in UWB ultra-wideband positioning chip to achieve accurate tracking of the probe position.

8. The medical equipment repair equipment based on AI automatic learning according to claim 1, characterized in that: The microprocessor unit (104) performs a device health assessment based on the data collected by the multidimensional sensor array (102): calculating a comprehensive health index. , The calculation formula is ,in For the first The normalized value of the health status of the test indicators For the first The weight coefficients of the indicators satisfy the following conditions: The health index H ranges from 0 to 100. When the value falls below a preset threshold, a preventative maintenance reminder is generated.

9. A medical equipment maintenance system based on AI automatic learning, characterized in that: The device includes an AI-based automatic learning medical device repair equipment as described in any one of claims 1-8 and a repair control terminal (2), wherein the repair control terminal (2) includes a main control chip (201), a graphics rendering engine (203) and a wireless communication module (205). The maintenance control terminal (2) receives the detection data and fault diagnosis results sent by the maintenance equipment through the communication interface (106) and the communication line, and drives the graphics rendering engine (203) to generate a visual maintenance report. The maintenance control terminal (2) also includes a touch screen (202), which is electrically connected to the main control chip (201). The touch screen (202) supports two display modes. When it receives normal detection data sent by the maintenance equipment, the main control chip (201) configures the display mode to data monitoring mode and displays real-time data in waveform and numerical form. When it receives a fault alarm signal, the main control chip (201) configures the display mode to fault warning mode and triggers an audible and visual alarm. The maintenance control terminal (2) also integrates a fault knowledge graph database (204); when the main control chip (201) receives the fault diagnosis result, it automatically associates with the fault knowledge graph database (204) to generate a maintenance plan containing fault cause analysis, handling suggestions and historical similar cases, and presents it on the touch screen (202).

10. The AI-based automatic learning medical equipment maintenance system according to claim 9, characterized in that: The maintenance control terminal (2) communicates with the cloud server (206) through the wireless communication module (205); in the self-learning update mode, the maintenance control terminal (2) uploads the detection data stored locally on the maintenance equipment to the cloud server (206), receives the updated algorithm model parameters returned by the cloud server (206), and then transmits the algorithm model parameters downlink to the maintenance equipment; The wireless communication module (205) supports at least one of Wi-Fi, Bluetooth or 5G communication protocols; the main control chip (201) is an ARM architecture processor running Linux or Android operating system; the graphics rendering engine (203) runs on the main control chip (201) in software form.

Images