Automatic adjusting method and system for insulin pump
By using wavelet scattering transform and a lightweight temporal convolutional classification network, the insulin pump can distinguish between real mechanical shock and electromagnetic coupling interference in a strong electromagnetic radiation environment, solving the problem of sensor misjudgment and achieving continuity and accuracy in the infusion process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-12
- Publication Date
- 2026-04-14
AI Technical Summary
In environments with strong electromagnetic radiation, the status monitoring sensors of existing insulin pumps are susceptible to electromagnetic coupling noise interference, which can misinterpret events as physical impacts, leading to unnecessary infusion interruptions and loss of infusion sequence integrity.
By employing wavelet scattering transform and a lightweight temporal convolutional classification network, and through multi-order scattering coefficient feature extraction and adaptive adjustment control, the insulin pump can distinguish between real mechanical shock and electromagnetic coupling interference, thereby achieving adaptive infusion regulation.
It significantly improves the ability to distinguish between electromagnetic coupling interference signals and real mechanical impact signals, ensuring the continuity and timing integrity of the infusion process and reducing the false alarm rate.
Smart Images
Figure CN121846424A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automated infusion control technology, and in particular to an automatic adjustment method and system for an insulin pump. Background Technology
[0002] An insulin pump is a medical device that precisely delivers insulin to a patient via continuous subcutaneous infusion. Its core function is to automatically drive the reservoir piston according to a preset infusion plan, achieving programmed control over the infusion rate and duration. To ensure safe operation while worn and in motion, existing insulin pumps typically integrate status monitoring units such as accelerometers, vibration sensors, or proximity sensors to detect in real time whether the device has been dropped, impacted, or subjected to abnormal displacement. When the monitored signal exceeds a preset threshold, the control system immediately triggers an infusion interruption mechanism and simultaneously activates an audible and visual alarm unit to alert the user, aiming to prevent device damage or decreased infusion accuracy due to physical impact.
[0003] To reduce the frequency of false alarms caused by unintentional touches, brief vibrations, and other non-hazardous events during daily use, existing technologies generally employ a threshold comparison combined with delayed confirmation. Specifically, the system only determines an event as a valid impact and executes protective actions after the amplitude of the sensor's output signal continuously exceeds a set threshold for a specific duration window.
[0004] However, in real-world applications, insulin pumps may frequently be located near strong electromagnetic radiation sources, such as airport security gates, RFID readers, or high-power audio equipment. The specific frequency electromagnetic waves radiated by these devices can spatially couple into the sensor and its signal conditioning circuitry, inducing parasitic voltages in the analog signal path. The amplitude and time-domain waveform of such interference signals are highly similar to the sensor output generated by real mechanical impacts. Existing software filtering algorithms based on amplitude thresholds and time windows struggle to effectively identify this type of electromagnetic coupling noise, causing the device to misinterpret non-physical electromagnetic disturbances as actual impacts or drops, thus incorrectly activating the infusion lockout procedure and issuing continuous alarms. When users are in environments such as security queues or crowded areas, they often cannot reset the device in time, causing unnecessary pauses in the infusion process and compromising the integrity of the infusion sequence. Summary of the Invention
[0005] The purpose of this invention is to provide an automatic adjustment method and system for insulin pumps, which solves the problem that the status monitoring sensor in the above-mentioned strong electromagnetic radiation environment misjudges physical impact events due to inductive coupling noise, thus causing the infusion interruption program to be triggered unnecessarily.
[0006] To achieve the above objectives, the present invention provides an automatic adjustment method for an insulin pump, comprising the following steps: Step 100: Acquire the signal output from the sensor in the motion monitoring unit of the insulin pump, perform preprocessing on the signal, and obtain a preprocessed signal frame; Step 200: Perform wavelet scattering transform processing on the preprocessed signal frame to generate a multi-order scattering coefficient set containing a set of zero-order scattering coefficients, a set of first-order scattering coefficients, and a set of second-order scattering coefficients. Step 300: Calculate statistical characteristic parameters for the set of multi-order scattering coefficients, and concatenate the statistical characteristic parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector; Step 400: Input the multidimensional scattering feature vector into the pre-deployed lightweight temporal convolutional classification network and output the event classification result. The event classification result includes the event category label and the corresponding confidence probability value. The event category label includes the real mechanical impact category and the electromagnetic coupling interference category. Step 500: Perform adaptive adjustment control on the insulin pump infusion state based on event category labels and confidence probability values.
[0007] Further, preprocessing is performed on the signal to obtain a preprocessed signal frame, including: The signal is converted from analog to digital. The converted digital signal is framed according to the preset sliding window length N and step size M to obtain the digital signal sequence, where N and M are both positive integers and M is less than N. The digital signal sequence is subjected to DC bias removal and amplitude normalization to obtain a preprocessed signal frame.
[0008] Furthermore, statistical characteristic parameters are calculated for the set of multi-order scattering coefficients, and the statistical characteristic parameters corresponding to each order of scattering coefficient are concatenated to construct a multi-dimensional scattering feature vector, including: Calculate the mean, variance, kurtosis, and L2 norm normalized energy values for the zero-order scattering coefficient set, the first-order scattering coefficient set, and the second-order scattering coefficient set, respectively. The mean, variance, kurtosis, and L2 norm normalized energy values corresponding to the zero-order scattering coefficient set, the first-order scattering coefficient set, and the second-order scattering coefficient set are concatenated according to order and scale number to construct a multidimensional scattering feature vector.
[0009] Furthermore, the lightweight temporal convolutional classifier network consists of a one-dimensional causal convolutional layer, a squeezed excitation attention module, and a fully connected output layer. The multidimensional scattering feature vector is processed by a one-dimensional causal convolutional layer to extract local multi-scale correlation features. The squeezed excitation attention module performs adaptive weight recalibration on the features of each channel. Finally, the fully connected output layer maps the results to the event classification.
[0010] Furthermore, adaptive adjustment control of the insulin pump infusion state is performed based on event category labels and confidence probability values, including: When the event category label is real mechanical impact and the confidence probability value is greater than or equal to the preset first confidence threshold, an infusion interruption command is sent to the infusion drive unit and the audible and visual alarm unit is activated. When the event category label is electromagnetic coupling interference and the confidence probability value is greater than or equal to the preset second confidence threshold, the current infusion rate of the infusion drive unit remains unchanged, and the timestamp of the preprocessed signal frame and the multidimensional scattering feature vector are written to the event log storage area. When the confidence probability value is less than the corresponding confidence threshold, the delayed confirmation process is initiated.
[0011] Furthermore, initiate the delayed confirmation process, including: Subsequent signal frames are continuously acquired at preset frame intervals within a preset delay confirmation time window T; For each subsequent signal frame, wavelet scattering transform processing, statistical feature parameter calculation, and lightweight temporal convolutional classification network inference are repeatedly performed to obtain the event classification results of multiple frames within the delayed confirmation time window T. A weighted cumulative voting decision is performed on the event classification results of multiple frames. The event category label with the highest cumulative voting score is determined as the final event category, and the corresponding infusion adjustment control operation is performed according to the final event category.
[0012] Furthermore, the weighted cumulative voting decision for the multi-frame event classification results includes: The formula for calculating voting weight is as follows: ; in, For the first Frame voting weights The frame sequence number within the delayed confirmation window. , For the first The larger confidence probability value in the frame classification results The preset time decay factor, For the first The frame interval between the last frame of the frame interval window; The cumulative voting scores for the real mechanical impact category and the electromagnetic coupling interference category are calculated based on the voting weights, using the following formulas: ; ; in, For the first The confidence probability value of the true mechanical impact category of the frame. For the first The confidence probability value of the electromagnetic coupling interference category of the frame. The cumulative voting score for the Real Mechanical Impact category. The cumulative voting score for the electromagnetic coupling interference category.
[0013] Furthermore, the event category label with the highest cumulative voting score is determined as the final event category, and the corresponding injection adjustment control operation is performed based on the final event category. Specifically, if... If the final event category is a real mechanical impact, the infusion will be interrupted and the audible and visual alarm unit will be activated; if If the event category is determined to be electromagnetic coupling interference, the current infusion state will remain unchanged.
[0014] An insulin pump automatic adjustment system for performing any of the above-described insulin pump automatic adjustment methods includes: The signal acquisition and preprocessing module is configured to acquire signals output by sensors in the motion monitoring unit, perform preprocessing on the signals, and obtain preprocessed signal frames. The wavelet scattering transform module is configured to perform wavelet scattering transform processing on the preprocessed signal frame to generate a set of multi-order scattering coefficients. The feature construction module is configured to calculate statistical feature parameters for a set of multi-order scattering coefficients and concatenate the statistical feature parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector. The event classification module is configured to input multidimensional scattering feature vectors into a pre-deployed lightweight temporal convolutional classification network and output event classification results. The infusion regulation control module is configured to perform adaptive regulation control on the infusion status of the insulin pump based on event classification results.
[0015] Therefore, the present invention employs the above-described automatic adjustment method and system for an insulin pump, which has the following beneficial effects: (1) By introducing wavelet scattering transform to decompose the sensor signal into multiple levels and scales, its translation invariance is used to eliminate the influence of random shift at the impact time on feature extraction, and its deformation stability is used to resist feature drift caused by waveform differences of different electromagnetic interference sources. Compared with the existing amplitude threshold and time window determination methods, it significantly improves the ability to distinguish between electromagnetic coupling interference signals and real mechanical impact signals. (2) Adaptive weight recalibration is performed on the scattering feature channels through the squeezing excitation attention mechanism in the lightweight temporal convolutional classification network, so that the network automatically focuses on the feature dimension with high discriminative power, and millisecond-level real-time inference is achieved on the embedded platform with limited model parameters and computational cost. (3) The compact multi-scale features generated by wavelet scattering transform reduce the input dimension of the classification network by an order of magnitude, enabling the lightweight network to perform high-precision classification on the insulin pump embedded processor. (4) By using a three-branch adaptive adjustment control strategy and a weighted cumulative voting decision mechanism under low confidence conditions, false alarms and unnecessary infusion interruptions in the electromagnetic environment are effectively avoided while ensuring a rapid response to real impact events, thus ensuring the continuity and timing integrity of the infusion process.
[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0017] Figure 1 This is a flowchart of an automatic adjustment method for an insulin pump according to the present invention. Detailed Implementation
[0018] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0019] Please see Figure 1 An automatic adjustment method for an insulin pump includes the following steps: Step 100: Acquire the signal output from the sensor in the motion monitoring unit of the insulin pump, perform preprocessing on the signal, and obtain a preprocessed signal frame; It is understood that the main body of this invention is the embedded controller in the insulin pump, which is connected to the motion monitoring unit, infusion drive unit, and audible and visual alarm unit via an internal bus. Specifically, the motion monitoring unit of the insulin pump has a built-in sensor for sensing physical impact (such as an accelerometer), which continuously outputs an analog voltage signal reflecting the device's motion state. The embedded controller performs analog-to-digital conversion on this analog signal and extracts a fixed-length signal frame according to a preset sliding window strategy. The signal frame is then subjected to DC bias removal and amplitude normalization to obtain a preprocessed signal frame. The infusion drive unit receives infusion control commands from the embedded controller and drives the micro-motor and transmission mechanism to accurately output the medication from the reservoir according to a preset flow rate and time period. The audible and visual alarm unit is used to simultaneously issue an audible alert and an optical flashing signal according to the control command when the embedded controller determines that an abnormal state has occurred (such as blockage of the infusion path, device drop or impact, or misjudgment caused by electromagnetic interference), to alert the user to pay attention to the device status and take appropriate action.
[0020] In an embodiment of the present invention, a patient wearing an insulin pump passes through a security checkpoint at the entrance of a public facility. The electromagnetic waves of the security gate's operating frequency band induce a parasitic voltage in the signal conditioning circuit of the sensor in the motion monitoring unit through spatial coupling. This parasitic voltage is superimposed on the normal output signal of the sensor to form a mixed signal, which is then preprocessed and used as the input for subsequent analysis.
[0021] Specifically, the signal preprocessing process is as follows: (1) Perform analog-to-digital conversion on the signal to obtain a digital signal sequence: The motion monitoring unit of the insulin pump continuously outputs an analog voltage signal at a fixed sampling frequency, and the analog-to-digital converter built into the embedded controller converts the analog signal into a digital signal sequence. ,in The sampling point number, , The preset sliding window length represents the total number of sampling points in each frame. The step size between adjacent frames is... ,in are positive integers and This achieves inter-frame overlap, ensuring the temporal continuity of the signal.
[0022] (2) For digital signal sequences Perform DC bias removal processing: remove DC bias from the entire digital signal sequence. The value of each sampling point The arithmetic mean of the digital signal sequence is calculated using the arithmetic mean method. ; the value of each sample point in the digital signal sequence and arithmetic mean The result after removing the DC bias is obtained by using point-by-point interpolation calculation. Signal value at each sampling point It should be noted that the purpose of DC bias removal is to eliminate the impact of sensor zero-point drift and analog-to-digital converter DC offset on subsequent frequency domain analysis.
[0023] (3) For the signal after removing DC bias Perform amplitude normalization: remove all values after removing DC bias. Signal value at each sampling point The maximum absolute value within the current frame is calculated by taking the maximum absolute value. ; Remove the DC signal value from each sampling point With maximum absolute amplitude The first frame in the preprocessed signal frame is obtained by using a division normalization operation. The value of each sampling point Amplitude normalization maps the signal amplitude to... The interval eliminates the dimensional differences in the sensor output amplitude under different operating conditions, making the output of subsequent wavelet scattering transform comparable.
[0024] In an embodiment of the present invention, the sampling frequency of the sensor is set to 1000Hz, and the sliding window length is... Set to 512 sampling points, corresponding to a signal duration of 512 milliseconds, step size The setting is 256 sampling points, corresponding to a 50% inter-frame overlap rate. This parameter setting can meet the real-time computing requirements of the embedded processor while ensuring the signal temporal resolution.
[0025] Step 200: Perform wavelet scattering transform processing on the preprocessed signal frame to generate a multi-order scattering coefficient set containing a set of zero-order scattering coefficients, a set of first-order scattering coefficients, and a set of second-order scattering coefficients. The embedded controller performs wavelet scattering transform on the preprocessed signal frame. This transform, through cascaded convolution and modulus operations using a multi-layer wavelet filter bank, decomposes the signal into scattering coefficients of multiple orders. The zero-order scattering coefficient reflects the low-frequency average energy characteristics of the signal, the set of first-order scattering coefficients captures the energy distribution characteristics at different frequency scales, and the set of second-order scattering coefficients reveals the modulation structure characteristics across scales. It should be noted that real mechanical impact signals exhibit a wideband uniform energy distribution in the scattering coefficient space, while electromagnetic coupling interference signals, due to their frequency band being concentrated in a specific frequency band, exhibit significant narrowband energy concentration characteristics in the first and second-order scattering coefficients. The two types of signals are highly separable in the scattering coefficient space.
[0026] Specifically, the embedded controller processes the preprocessed signal frames. Perform wavelet scattering transform processing. The wavelet scattering transform is performed through a preset... A layered wavelet filter bank is used for cascaded convolution and modulus calculation, where The preset decomposition scale is a positive integer. The wavelet filter bank includes bandpass wavelet filters. and low-pass average filter ,in For frequency scale parameters, , For convolution operators, This is the modulo operator.
[0027] Zero-order scattering coefficient The calculation formula is: ; in, For the preprocessed signal frame of the nth sampling point, It is a low-pass average filter. For convolution operators, Characterizes the low-frequency average energy of the preprocessed signal frame.
[0028] First-order scattering coefficient The calculation process is as follows: First, calculate the first-order wavelet modulus coefficients. Then, it is subjected to low-pass average filtering, and the calculation formula is: ; ; in, For scale The corresponding bandpass wavelet filter, The modulo operator, For scale The first-order wavelet modulus coefficients, For scale The first-order scattering coefficients below. The set of first-order scattering coefficients contains a total of Energy distribution information at various frequency scales.
[0029] Second-order scattering coefficient The calculation process is as follows: for the first-order wavelet modulus coefficients Perform wavelet convolution and modulus calculation again, followed by low-pass averaging filtering. The calculation formula is as follows: ; ; in, For scale The corresponding bandpass wavelet filter, The frequency scaling parameter of the second-order decomposition and , The modulo operator, For scale pairs The second-order wavelet modulus coefficients are as follows: For scale pairs The second-order scattering coefficients are listed below. The set of second-order scattering coefficients contains a total of Information on the cross-scale modulation structure.
[0030] It should be noted that wavelet scattering transform possesses two key characteristics: translation invariance and deformation stability. Translation invariance ensures that the scattering coefficients are unaffected by signal time shifts; even if the precise moment of a mechanical impact occurs randomly within a frame, the extracted features remain stable. Deformation stability makes the scattering coefficients robust to small waveform changes in the signal; although interference waveforms from different electromagnetic radiation sources may have slight differences, their scattering coefficients remain highly consistent.
[0031] In an embodiment of the present invention, the decomposition scale number is... When set to 4, the wavelet filter bank uses the Morlet wavelet as the bandpass filter. At this point, the first-order scattering coefficient set contains energy distribution information at four frequency scales, and the second-order scattering coefficient set contains... Information on the modulation structure across scales. In the security checkpoint scenario, the energy of electromagnetic interference signals is concentrated at a specific frequency scale (corresponding to the operating frequency band of the security gate), exhibiting a significant energy peak in the first-order scattering coefficient at that scale, while the energy is extremely low at other scales; real mechanical impact signals, on the other hand, show a relatively uniform energy distribution across all frequency scales.
[0032] Step 300: Calculate statistical characteristic parameters for the set of multi-order scattering coefficients, and concatenate the statistical characteristic parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector; Specifically, the embedded controller calculates statistical characteristic parameters for each order of scattering coefficient sequence, including mean, variance, kurtosis, and L2 norm normalized energy. The mean reflects the average amplitude level of the scattering coefficients, the variance measures the degree of fluctuation of the scattering coefficients, the kurtosis measures the sharpness of the scattering coefficient distribution, and the L2 norm normalized energy characterizes the proportion of each scattering coefficient in the total energy. The statistical characteristic parameters corresponding to all orders of scattering coefficients are concatenated according to order and scale number to construct a multidimensional scattering feature vector.
[0033] Specifically, the embedded controller calculates four statistical characteristic parameters for each order of scattering coefficients. Let a certain scattering coefficient sequence be... ,in is the element index of the scattering coefficient sequence. , The length of this scattering coefficient sequence is given by the zero-order scattering coefficients generated in step 200. First-order scattering coefficient or second-order scattering coefficient Any one of the groups.
[0034] scattering coefficient sequence All The mean of the scattering coefficient sequence is calculated using the arithmetic mean method for each element. mean This reflects the average amplitude level of the scattering coefficient.
[0035] scattering coefficient sequence All Elements and Mean The variance of the scattering coefficient sequence was calculated using the variance calculation method. ,variance This measures the degree of fluctuation in the scattering coefficient. Due to its transient characteristics, the variance of the scattering coefficient of a real mechanical impact signal is significantly greater than that of an electromagnetically coupled interference signal.
[0036] scattering coefficient sequence All Individual elements, mean and variance The kurtosis value of the scattering coefficient sequence was calculated using the kurtosis calculation method. kurtosis value This measures the sharpness of the scattering coefficient distribution relative to a normal distribution. Due to its quasi-periodic characteristics, the kurtosis value of the scattering coefficient of electromagnetic coupling interference signals approaches the baseline value of a normal distribution; while the pulse characteristics of mechanical shock signals cause their scattering coefficients to exhibit a super-Gaussian distribution, with significantly larger kurtosis values.
[0037] scattering coefficient sequence All The L2 norm squared value of the scattering coefficient sequence is calculated for each element using the L2 norm calculation method; all order scattering coefficients generated in step 200 (i.e., ,all and all The sum of the squared L2 norms of all scattering coefficients of all orders is obtained by summing the squared L2 norms of each individual L2 norm. The sum of the squared L2 norm values of the scattering coefficient sequence and the sum of the squared L2 norm values. The L2 norm normalized energy value was calculated using ratio normalization. , This characterizes the proportion of each scattering coefficient in the total signal energy. The energy of electromagnetic coupling interference signals is highly concentrated on a few specific scales, with the corresponding L2 norm normalized energy value approaching 1 on these scales and approaching 0 on other scales; the energy of real mechanical impact signals is more uniformly distributed across all scales, with smaller differences between the L2 norm normalized energy values.
[0038] The four statistical characteristic parameters corresponding to all scattering coefficients are concatenated according to their order and scale number to construct a dimensionless array. Multidimensional scattering eigenvectors : ; in, Zero-order scattering coefficient The mean, for variance for kurtosis value, for L2 norm normalized energy value; to For each first-order scattering coefficient ( The four statistical characteristic parameters of ); to For each second-order scattering coefficient ( The four statistical characteristic parameters of ). ,in The number of decomposition scales preset in step 200.
[0039] Step 400: Input the multidimensional scattering feature vector into a lightweight temporal convolutional classification network pre-deployed in the insulin pump embedded processor, and output the event classification result; The event classification results include event category labels and corresponding confidence probability values. Event category labels include real mechanical impact category and electromagnetic coupling interference category.
[0040] Specifically, the lightweight temporal convolutional classification network is pre-trained offline and deployed in the embedded processor of the insulin pump. The network consists of a one-dimensional causal convolutional layer, a squeezed activation attention module, and a fully connected output layer.
[0041] One-dimensional causal convolutional layers for multi-dimensional scattering feature vectors Perform a one-dimensional causal convolution operation to generate Feature mapping for each channel. The total number of output channels is and the kernel length is . , multidimensional scattering feature vector The elements and Group convolution kernel weights (in This is the output channel number. , This is the offset index within the convolution kernel. ) and the corresponding bias parameters The results were obtained by employing one-dimensional causal convolution operations and a linear rectified activation function. Feature mapping of each channel ,in This refers to the feature location index. Causal convolution ensures that each output location depends only on the input at the current and previous locations (when...). (When the input is zero), it satisfies the causal constraints of real-time processing.
[0042] The squeeze-activated attention module performs channel-level adaptive weight recalibration on the output of the causal convolutional layer, including a squeeze step and an activation step. In the squeeze step, all feature maps for each channel are... ( , The feature length of the output of the causal convolutional layer is calculated by using global average pooling. Global description value of each channel This forms the channel description vector. In the excitation step, the channel description vector is... Weight matrix of the first fully connected layer of the extrusion excitation module Bias vector The weight matrix of the second fully connected layer Bias vector The attention weight vector is calculated by employing a squeeze-incentive attention mechanism (sequentially mapping through the first fully connected layer, linear rectified activation, mapping through the second fully connected layer, and sigmoid activation). ,in For the first Attention weights for each channel. The feature values of each channel output from the causal convolutional layer. With the corresponding attention weights The recalibrated eigenvalues were calculated using channel-by-channel scalar multiplication. .
[0043] The fully connected output layer maps the recalibrated features to binary classification probability outputs. It then maps the recalibrated features of each channel... ( The result is obtained by using global average pooling. The convergence value of each channel This forms the convergence vector. The convergence vector A linear fractional vector is obtained through mapping via a fully connected layer. ,in The linear output score corresponds to the actual mechanical impact category. For the linear output score corresponding to the electromagnetic coupling interference category; convert the linear score vector The confidence probability value of the true mechanical impact category was calculated using Softmax normalization. Confidence probability value of electromagnetic coupling interference category Event category tags are taken from and The category corresponding to the larger value is the confidence probability value of that larger value.
[0044] In an embodiment of the present invention, the number of output channels Set to 16, kernel length Set to 3. (This refers to the number of decomposition scales.) At that time, the input feature dimension was 44, the number of parameters of the entire network was about 1200 floating-point numbers, and the inference time per frame was about 2 milliseconds, which met the real-time constraints of the insulin pump embedded processor.
[0045] The embedded controller inputs a multidimensional scattering feature vector into a pre-trained, lightweight temporal convolutional classification network deployed in on-chip memory. This network sequentially performs forward inference through a one-dimensional causal convolutional layer, a squeezed excitation attention module, and a fully connected output layer, outputting an event classification result. The event classification result includes an event category label and a corresponding confidence probability value; the event category label is either a real mechanical impact category or an electromagnetic coupling interference category. In the embodiments of this invention, in the aforementioned security checkpoint scenario, because the electromagnetic interference signal exhibits narrow-band clustering characteristics in the scattering coefficient space, the network can classify it as an electromagnetic coupling interference category with high confidence.
[0046] Step 500: Perform adaptive adjustment control on the insulin pump infusion state based on event category labels and confidence probability values.
[0047] Specifically, the embedded controller executes three control branches based on the event category label and the confidence probability value.
[0048] First control branch: When the event category label is real mechanical impact and the confidence probability value... Greater than or equal to the preset first confidence threshold When the infusion interruption command is detected, the embedded controller sends an infusion interruption command to the infusion drive unit. The drive motor stops pushing the drug reservoir piston, and at the same time, the audible and visual alarm unit is activated to emit sound and flashing indicator lights, prompting the user that the device has detected a real physical impact and that the user needs to check the device status and manually resume the infusion.
[0049] Second control branch: When the event category label is electromagnetic coupling interference and the confidence probability value is... Greater than or equal to the preset second confidence threshold At this time, the embedded controller maintains the current infusion rate of the infusion drive unit unchanged, does not perform any interrupt or alarm operation, and writes the timestamp and multidimensional scattering feature vector of the current preprocessed signal frame into the event log storage area for subsequent backtracking analysis.
[0050] The third control branch: When the confidence probability value is less than the corresponding confidence threshold (i.e., the event category label is a true mechanical impact category but...) Or the event category is labeled as electromagnetic coupling interference but Embedded controller startup delay confirmation process.
[0051] In an embodiment of the present invention, the first confidence threshold The default value is set to 0.85, the second confidence threshold. The default value is set to 0.80.
[0052] In an embodiment of the present invention, the delayed confirmation process occurs within a preset delayed confirmation time window. Internal execution. The embedded controller continuously acquires subsequent signal frames at preset frame intervals, and for each subsequent signal frame, it sequentially performs wavelet scattering transform processing, statistical feature parameter calculation, and lightweight temporal convolutional inference network inference (steps 200-400) to obtain the delay confirmation time window. The multi-frame event classification results within the frame. Define a delay confirmation time window. Internal acquisition Frame event classification results It is a positive integer.
[0053] A weighted cumulative voting decision is performed on the multi-frame event classification results. The voting weight of each frame is related to the classification confidence and temporal position of that frame. The voting weight of the frame is The calculation formula is: ; in, The frame sequence number within the delayed confirmation window. , For the first The larger confidence probability value in the frame classification results (i.e. , derived from the first The confidence probability value output after the frame signal is repeatedly executed from step 200 to 400. The preset time decay factor has a value range of [value range missing]. , For the first The frame interval is the distance between the last frame of the frame interval window. This weighting design gives higher confidence levels and frames that are closer to the current time in time greater voting weight.
[0054] Cumulative voting score for the Real Mechanical Impact category Cumulative voting score for the electromagnetic coupling interference category The calculation formulas are as follows: ; ; in, For the first The confidence probability value of the true mechanical impact category of the frame. For the first The confidence probability value of the electromagnetic coupling interference category of the frame. For the first Frame voting weights The cumulative voting score for the Real Mechanical Impact category. The cumulative voting score for the electromagnetic coupling interference category.
[0055] The event category label with the highest cumulative voting score is determined as the final event category, and the corresponding input adjustment control operation is performed based on the final event category: If If the final event category is a real mechanical impact, the infusion will be interrupted and the audible and visual alarm unit will be activated; if If the event category is determined to be electromagnetic coupling interference, the current infusion state will remain unchanged.
[0056] In an embodiment of the present invention, a delayed confirmation time window is defined. Set to 2 seconds, frame interval and step size The corresponding time is consistent (i.e., 256 milliseconds), and approximately 8 frames are acquired within the delay confirmation window. Time decay factor Set to 0.9. In the security checkpoint scenario, electromagnetic coupling interference dominated for eight consecutive frames of signal, with each frame's... The values are all higher than After weighted cumulative voting Significantly greater than The system ultimately determined it to be electromagnetic coupling interference and maintained the infusion rate unchanged.
[0057] In this embodiment of the invention, a three-branch adaptive adjustment control strategy combined with a weighted cumulative voting decision mechanism with delayed confirmation is used to achieve immediate response under high confidence conditions, and to improve decision reliability by fusing multi-frame time sequence information under low confidence conditions, thus effectively balancing the system's response speed and decision accuracy.
[0058] When the system determines that the interference is a real mechanical impact and the confidence level is higher than a preset threshold, the infusion is immediately interrupted and an alarm is triggered to protect patient safety. When the system determines that the interference is electromagnetic coupling and the confidence level is higher than a preset threshold, the current infusion rate is maintained to avoid unnecessary infusion interruptions. When the confidence level is lower than a preset threshold, a delayed confirmation process is initiated, continuously acquiring multiple frames of signals within a time window and repeating the above analysis steps. The final event category is determined through weighted cumulative voting. In the embodiments of the present invention, in the above security checkpoint scenario, the system maintains the infusion after correctly identifying electromagnetic interference, and the patient does not need to manually reset the device, ensuring the integrity of the infusion sequence.
[0059] An automatic adjustment system for an insulin pump includes: The signal acquisition and preprocessing module is configured to acquire the signal output by the sensor through the motion monitoring unit of the insulin pump, perform preprocessing on the signal, and obtain a preprocessed signal frame.
[0060] The wavelet scattering transform module is configured to perform wavelet scattering transform processing on the preprocessed signal frame to generate a multi-order scattering coefficient set.
[0061] The feature construction module is configured to calculate statistical feature parameters for a set of multi-order scattering coefficients and concatenate the statistical feature parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector.
[0062] The event classification module is configured to input multidimensional scattering feature vectors into a lightweight temporal convolutional classification network pre-deployed in an insulin pump embedded processor, and output event classification results.
[0063] The infusion regulation control module is configured to perform adaptive regulation control on the infusion status of the insulin pump based on event classification results.
[0064] In an embodiment of the invention, the motion monitoring unit of the insulin pump acquires signals output by sensors, performs preprocessing on the signals to obtain preprocessed signal frames, performs wavelet scattering transform on the preprocessed signal frames to generate a set of multi-order scattering coefficients, calculates statistical feature parameters on the set of multi-order scattering coefficients to construct a multi-dimensional scattering feature vector, inputs the multi-dimensional scattering feature vector into a lightweight temporal convolutional classification network to output event classification results, and performs adaptive adjustment control on the infusion state of the insulin pump based on the event classification results. By combining wavelet scattering transform with a lightweight temporal convolutional classification network, high-precision real-time differentiation between real mechanical impact signals and electromagnetic coupling interference signals is achieved on the embedded platform of the insulin pump, significantly reducing the false alarm rate in electromagnetic environments and ensuring the continuity and temporal integrity of the infusion process.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. An automatic adjustment method for an insulin pump, characterized in that, Includes the following steps: Step 100: Acquire the signal output from the sensor in the motion monitoring unit of the insulin pump, perform preprocessing on the signal, and obtain a preprocessed signal frame; Step 200: Perform wavelet scattering transform processing on the preprocessed signal frame to generate a multi-order scattering coefficient set containing a set of zero-order scattering coefficients, a set of first-order scattering coefficients, and a set of second-order scattering coefficients. Step 300: Calculate statistical characteristic parameters for the set of multi-order scattering coefficients, and concatenate the statistical characteristic parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector; Step 400: Input the multidimensional scattering feature vector into the pre-deployed lightweight temporal convolutional classification network and output the event classification result. The event classification result includes the event category label and the corresponding confidence probability value. The event category label includes the real mechanical impact category and the electromagnetic coupling interference category. Step 500: Perform adaptive adjustment control on the insulin pump infusion state based on event category labels and confidence probability values.
2. The method for automatic adjustment of an insulin pump according to claim 1, characterized in that, Preprocessing is performed on the signal to obtain a preprocessed signal frame, including: The signal is converted from analog to digital. The converted digital signal is framed according to the preset sliding window length N and step size M to obtain the digital signal sequence, where N and M are both positive integers and M is less than N. The digital signal sequence is subjected to DC bias removal and amplitude normalization to obtain a preprocessed signal frame.
3. The method for automatic adjustment of an insulin pump according to claim 2, characterized in that, Statistical characteristic parameters are calculated for the set of multi-order scattering coefficients. The statistical characteristic parameters corresponding to each order of scattering coefficient are then concatenated to construct a multi-dimensional scattering feature vector, including: Calculate the mean, variance, kurtosis, and L2 norm normalized energy values for the zero-order scattering coefficient set, the first-order scattering coefficient set, and the second-order scattering coefficient set, respectively. The mean, variance, kurtosis, and L2 norm normalized energy values corresponding to the zero-order scattering coefficient set, the first-order scattering coefficient set, and the second-order scattering coefficient set are concatenated according to order and scale number to construct a multidimensional scattering feature vector.
4. The method for automatic adjustment of an insulin pump according to claim 3, characterized in that, The lightweight temporal convolutional classifier network consists of a one-dimensional causal convolutional layer, a squeezed excitation attention module, and a fully connected output layer. The multidimensional scattering feature vector is processed by a one-dimensional causal convolutional layer to extract local multi-scale correlation features. The squeezed excitation attention module performs adaptive weight recalibration on the features of each channel. Finally, the fully connected output layer maps the results to the event classification.
5. The method for automatic adjustment of an insulin pump according to claim 4, characterized in that, Adaptive control of insulin pump infusion status is performed based on event category labels and confidence probability values, including: When the event category label is real mechanical impact and the confidence probability value is greater than or equal to the preset first confidence threshold, an infusion interruption command is sent to the infusion drive unit and the audible and visual alarm unit is activated. When the event category label is electromagnetic coupling interference and the confidence probability value is greater than or equal to the preset second confidence threshold, the current infusion rate of the infusion drive unit remains unchanged, and the timestamp of the preprocessed signal frame and the multidimensional scattering feature vector are written to the event log storage area. When the confidence probability value is less than the corresponding confidence threshold, the delayed confirmation process is initiated.
6. The method for automatic adjustment of an insulin pump according to claim 5, characterized in that, Initiate a delayed confirmation process, including: Subsequent signal frames are continuously acquired at preset frame intervals within a preset delay confirmation time window T; For each subsequent signal frame, wavelet scattering transform processing, statistical feature parameter calculation, and lightweight temporal convolutional classification network inference are repeatedly performed to obtain the event classification results of multiple frames within the delayed confirmation time window T. A weighted cumulative voting decision is performed on the event classification results of multiple frames. The event category label with the highest cumulative voting score is determined as the final event category, and the corresponding infusion adjustment control operation is performed according to the final event category.
7. The method for automatic adjustment of an insulin pump according to claim 6, characterized in that, The weighted cumulative voting decision for multi-frame event classification results includes: The formula for calculating voting weight is as follows: ； in, For the first Frame voting weights The frame sequence number within the delayed confirmation window. , For the first The larger confidence probability value in the frame classification results The preset time decay factor, For the first The frame interval between the last frame of the frame interval window; The cumulative voting scores for the real mechanical impact category and the electromagnetic coupling interference category are calculated based on the voting weights, using the following formulas: ； ； in, For the first The confidence probability value of the true mechanical impact category of the frame. For the first The confidence probability value of the electromagnetic coupling interference category of the frame. The cumulative voting score for the Real Mechanical Impact category. The cumulative voting score for the electromagnetic coupling interference category.
8. The method for automatic adjustment of an insulin pump according to claim 7, characterized in that, The event category label with the highest cumulative voting score is determined as the final event category, and the corresponding input adjustment control operation is performed based on the final event category. Specifically: If If the final event category is a real mechanical impact, the infusion will be interrupted and the audible and visual alarm unit will be activated; if If the event category is determined to be electromagnetic coupling interference, the current infusion state will remain unchanged.
9. An automatic adjustment system for an insulin pump, characterized in that, It is used to perform an automatic adjustment method for an insulin pump according to any one of claims 1-8, comprising: The signal acquisition and preprocessing module is configured to acquire signals output by sensors in the motion monitoring unit, perform preprocessing on the signals, and obtain preprocessed signal frames. The wavelet scattering transform module is configured to perform wavelet scattering transform processing on the preprocessed signal frame to generate a set of multi-order scattering coefficients. The feature construction module is configured to calculate statistical feature parameters for a set of multi-order scattering coefficients and concatenate the statistical feature parameters corresponding to each order of scattering coefficients to construct a multi-dimensional scattering feature vector. The event classification module is configured to input multidimensional scattering feature vectors into a pre-deployed lightweight temporal convolutional classification network and output event classification results. The infusion regulation control module is configured to perform adaptive regulation control on the infusion status of the insulin pump based on event classification results.