A method for monitoring the operating state of an ECMO device

Abstract

The application discloses an ECMO equipment operation state monitoring method, and belongs to the technical field of ECMO equipment operation state monitoring. First, the application collects vibration signals of a blood pump under low, medium and high test rotating speeds through a vibration sensor; then, the signals are subjected to mean value removal processing, and low-frequency, medium-frequency and high-frequency fault enhancement signals are synthesized based on frequency band energy weighting; subsequently, energy abnormality synthesis values and extreme value abnormality synthesis values are extracted for the fault enhancement signals and envelope signals thereof; then, the two types of abnormality synthesis values of different frequency bands are spliced respectively to obtain an energy abnormality splicing matrix and an extreme value abnormality splicing matrix; finally, the two matrices are processed by using an operation state evaluation neural network, and the operation state score of the blood pump is outputted, so that the precise monitoring of the operation state of the blood pump of the ECMO equipment is realized.

Description

Technical Field

[0001] This invention relates to the field of ECMO device operation status monitoring technology, and specifically to a method for monitoring the operation status of ECMO devices. Background Technology

[0002] ECMO (Extracorporeal Membrane Oxygenation) devices are critical life support equipment in the field of critical care medicine. Their operational stability directly affects the treatment effect and life safety of critically ill patients. Among them, the blood pump, as the core power component of ECMO devices, is particularly important for monitoring its operational status. Currently, existing ECMO device operational status monitoring technologies mostly rely on pressure sensors and flow sensors to collect parameters such as fluid pressure and flow rate during blood pump operation, or temperature sensors to monitor the temperature of key components of the equipment. Data processing often uses simple threshold judgment methods or basic signal filtering analysis, that is, setting fixed pressure, flow rate, and temperature thresholds, and judging that the equipment has an operational abnormality when the monitored parameters exceed the threshold range, or directly comparing the collected signals with standard signal waveforms after conventional filtering to determine the status. However, such existing technologies can only monitor changes in macroscopic parameters during blood pump operation, and cannot capture early abnormalities caused by subtle faults such as mechanical wear and loose parts. Moreover, they are easily interfered with by external factors such as changes in patient position and differences in blood viscosity, resulting in low accuracy in monitoring the operational status of ECMO device blood pumps and difficulty in accurately and timely identifying potential fault risks. Summary of the Invention

[0003] To address the aforementioned shortcomings in the existing technology, the present invention provides an ECMO device operation status monitoring method that solves the problem of low accuracy in monitoring the blood pump operation status of ECMO devices in the existing technology.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for monitoring the operating status of an ECMO device, comprising the following steps:

[0005] Vibration signals of the blood pump in the ECMO device were collected using vibration sensors at low, medium, and high test speeds.

[0006] The vibration signals from the three test speeds were averaged and weighted based on frequency band energy to synthesize low-frequency, mid-frequency, and high-frequency fault enhancement signals.

[0007] The envelope signal is obtained from the fault enhancement signal, and then the energy anomaly synthesis value and extreme value anomaly synthesis value are extracted from each segment of the fault enhancement signal and the envelope signal.

[0008] The energy anomaly composite values ​​of low-frequency, mid-frequency and high-frequency fault enhancement signals are spliced ​​together to obtain an energy anomaly splicing matrix.

[0009] The extreme anomaly composite values ​​of low-frequency, mid-frequency and high-frequency fault enhancement signals are spliced ​​together to obtain the extreme anomaly splicing matrix;

[0010] The operational status score of the blood pump in the ECMO device was obtained by processing the energy anomaly splicing matrix and the extreme value anomaly splicing matrix using a neural network for operational status assessment.

[0011] Furthermore, the process of synthesizing low-frequency, mid-frequency, and high-frequency fault enhancement signals includes:

[0012] The mean vibration amplitude is calculated for each test rotation speed, and the mean vibration amplitude is subtracted from each vibration amplitude to obtain the preprocessed vibration signal;

[0013] The preprocessed vibration signal is subjected to Fourier transform to extract the corresponding spectrum. The spectrum is divided into low-frequency band spectrum, mid-frequency band spectrum and high-frequency band spectrum. The frequency band spectrum is then subjected to inverse Fourier transform to obtain the low-frequency band signal, mid-frequency band signal and high-frequency band signal.

[0014] For the same frequency band signals corresponding to the three test speeds, the signals are weighted based on the frequency band energy to synthesize the corresponding frequency band fault enhancement signal.

[0015] Furthermore, the weighted formula is:

[0016] ,

[0017] in, Let x be the vibration amplitude of the i-th fault enhancement signal at time t. 1,i,t x represents the vibration amplitude at time t on the signal corresponding to the low test speed. 2,i,t Let x be the vibration amplitude at time t on the signal corresponding to the test rotational speed. 3,i,t e represents the vibration amplitude at time t on the signal corresponding to the high test speed. 1,i,t e represents the energy at time t on the signal corresponding to the low test speed. 2,i,t e represents the energy at time t on the signal corresponding to the test rotational speed. 3,i,t e represents the energy at time t on the signal corresponding to the high test speed. total,i,t For e 1,i,t e 2,i,t and e 3,i,t The sum of , where i takes the values ​​1, 2, and 3, corresponding to low frequency, mid frequency, and high frequency respectively, and t is the time number.

[0018] Furthermore, the process of extracting the composite values ​​of energy anomalies and the composite values ​​of extreme anomalies includes:

[0019] The fault enhancement signal is subjected to Hilbert transform to obtain the analytical signal;

[0020] Extract the envelope signal from the analytic signal;

[0021] Both the envelope signal and the fault enhancement signal are processed in segments.

[0022] The energy of each segment of the envelope signal and the energy of each segment of the fault enhancement signal are calculated to obtain the composite energy value of each segment.

[0023] Based on the extreme values ​​of each segment of the envelope signal and the extreme values ​​of each segment of the fault enhancement signal, the extreme value anomaly synthesis value of each segment is obtained, where the extreme values ​​include the maximum amplitude value.

[0024] Furthermore, the process of obtaining the energy anomaly composite value for each segment includes:

[0025] Calculate the energy for each segment of the envelope signal to obtain the envelope energy for each segment of the envelope signal;

[0026] The energy of each segment of the fault enhancement signal is calculated to obtain the enhancement energy of each segment of the fault enhancement signal;

[0027] The first energy anomaly coefficient is extracted based on the envelope energy of each segment of the envelope signal.

[0028] Based on the enhanced energy of each segment of the fault enhancement signal, the second energy anomaly coefficient is extracted;

[0029] The combined energy anomaly value is obtained by adding the first energy anomaly coefficient and the second energy anomaly coefficient.

[0030] Furthermore, the process of extracting the first energy anomaly coefficient includes: taking the average value of the envelope energy of each segment to obtain the average envelope energy; subtracting the average envelope energy from the envelope energy of each segment and taking the absolute value to obtain the envelope energy deviation; and using the ratio of the envelope energy deviation to the average envelope energy as the first energy anomaly coefficient.

[0031] The process of extracting the second energy anomaly coefficient includes: taking the average value of the enhanced energy of each segment to obtain the enhanced energy mean; subtracting the enhanced energy mean from the enhanced energy of each segment and taking the absolute value to obtain the enhanced energy deviation; and using the ratio of the enhanced energy deviation to the enhanced energy mean as the second energy anomaly coefficient.

[0032] Furthermore, the process of obtaining the composite extreme value of each segment includes:

[0033] Extract the first extreme value anomaly coefficient based on each extreme value of the envelope signal;

[0034] Extract the second extreme value anomaly coefficient based on each extreme value of the fault enhancement signal;

[0035] The first extreme value anomaly coefficient and the second extreme value anomaly coefficient are added together to obtain the extreme value anomaly composite value.

[0036] Furthermore, the process of extracting the first extreme value anomaly coefficient includes:

[0037] The mean value of the envelope amplitude is obtained by taking the average value of each amplitude of the envelope signal. The mean value of the envelope amplitude is obtained by subtracting the mean value of the envelope amplitude from each extreme value of the envelope signal and taking the absolute value. The ratio of the mean value of the envelope amplitude to the mean value of the envelope amplitude is used as the first extreme value anomaly coefficient.

[0038] The process of extracting the coefficient of the second extreme value anomaly includes:

[0039] The mean of the enhanced amplitude is obtained by taking the average value of each amplitude of the fault enhancement signal. The mean of the enhanced amplitude is obtained by subtracting the mean of the enhanced amplitude from each extreme value of the fault enhancement signal and taking the absolute value. The ratio of the extreme value deviation to the mean of the enhanced amplitude is used as the second extreme value anomaly coefficient.

[0040] Furthermore, the operating state evaluation neural network includes: a first shallow feature extraction unit, a second shallow feature extraction unit, a shared weight generation unit, a multiplier M1, a multiplier M2, a first deep feature extraction unit, a second deep feature extraction unit, an adder A1, and a fully connected layer;

[0041] The input of the first shallow feature extraction unit is used to input the energy anomaly splicing matrix, and its output is connected to the first input of the multiplier M1 and the first input of the shared weight generation unit, respectively.

[0042] The input of the second shallow feature extraction unit is used to input the extreme value anomaly concatenation matrix, and its output is connected to the first input of the multiplier M2 and the second input of the shared weight generation unit, respectively.

[0043] The output of the shared weight generation unit is connected to the second input of multiplier M1 and the second input of multiplier M2, respectively.

[0044] The output of multiplier M1 is connected to the input of the first deep feature extraction unit;

[0045] The output of multiplier M2 is connected to the input of the second deep feature extraction unit;

[0046] The input of adder A1 is connected to the output of the first deep feature extraction unit and the output of the second deep feature extraction unit, and its output is connected to the input of the fully connected layer.

[0047] The output of the fully connected layer serves as the output of the running state evaluation neural network.

[0048] Furthermore, the shared weight generation unit includes: a first convolutional layer, a second convolutional layer, an adder A2, and a first sigmoid layer;

[0049] The input of the first convolutional layer serves as the first input of the shared weight generation unit;

[0050] The input of the second convolutional layer serves as the second input of the shared weight generation unit.

[0051] The input of adder A2 is connected to the output of the first convolutional layer and the output of the second convolutional layer, and its output is connected to the input of the first sigmoid layer.

[0052] The output of the first Sigmoid layer serves as the output of the shared weight generation unit.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. This invention collects vibration signals of the blood pump at different speeds using a vibration sensor, and combines mean-reduction processing with signal weighting based on frequency band energy to synthesize low-frequency, mid-frequency, and high-frequency fault enhancement signals. This can keenly detect early anomalies caused by subtle faults such as mechanical wear and loose parts, making up for the shortcomings of existing technologies that rely on macroscopic parameter monitoring to detect early faults.

[0055] 2. Compared with existing technologies, which are easily affected by external factors such as changes in patient position and differences in blood viscosity, the vibration signal-based monitoring and processing method of this invention is less affected by such external factors, effectively reducing interference and improving the stability and reliability of monitoring.

[0056] 3. This invention acquires the envelope signal of the fault enhancement signal, and then combines the energy and extreme values ​​of the fault enhancement signal and the envelope signal to extract the composite values ​​of energy anomalies and extreme value anomalies, reflecting the energy anomalies and extreme value anomalies. This not only preserves the dynamic mechanical operation information contained in the fault enhancement signal, but also captures the amplitude modulation characteristics of the signal through envelope analysis. Then, the composite values ​​of energy anomalies and extreme value anomalies of low-frequency, mid-frequency and high-frequency fault enhancement signals are concatenated into a matrix. The energy anomaly concatenation matrix and extreme value anomaly concatenation matrix are processed by the operating state evaluation neural network, avoiding information loss caused by mutual interference of features in different frequency bands. It has a stronger ability to identify complex anomalies caused by the coordinated failure of multiple components in the blood pump.

[0057] 4. The operating status assessment neural network can establish a nonlinear mapping relationship between features and the health status of the blood pump by deeply mining the splicing matrix of energy anomalies and extreme value anomalies, breaking through the limitations of traditional threshold methods or simple waveform comparisons. Attached Figure Description

[0058] Figure 1 A flowchart of a method for monitoring the operational status of an ECMO device;

[0059] Figure 2 A schematic diagram of the structure of a neural network for evaluating operational status;

[0060] Figure 3 A schematic diagram of the structure of the shared weight generation unit;

[0061] Figure 4 This is a schematic diagram of the shallow feature extraction module;

[0062] Figure 5 This is a schematic diagram of the feature enhancement module.

[0063] Figure 6 This is a schematic diagram of the structure of the first deep feature extraction unit and the second deep feature extraction unit. Detailed Implementation

[0064] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0065] like Figure 1 As shown, a method for monitoring the operating status of an ECMO device includes the following steps:

[0066] Vibration signals of the blood pump in the ECMO device were collected using vibration sensors at low, medium, and high test speeds.

[0067] The vibration signals from the three test speeds were averaged and weighted based on frequency band energy to synthesize low-frequency, mid-frequency, and high-frequency fault enhancement signals.

[0068] The envelope signal is obtained from the fault enhancement signal, and then the energy anomaly synthesis value and extreme value anomaly synthesis value are extracted from each segment of the fault enhancement signal and the envelope signal.

[0069] The energy anomaly composite values ​​of low-frequency, mid-frequency and high-frequency fault enhancement signals are spliced ​​together to obtain an energy anomaly splicing matrix.

[0070] The extreme anomaly composite values ​​of low-frequency, mid-frequency and high-frequency fault enhancement signals are spliced ​​together to obtain the extreme anomaly splicing matrix;

[0071] The operational status score of the blood pump in the ECMO device was obtained by processing the energy anomaly splicing matrix and the extreme value anomaly splicing matrix using a neural network for operational status assessment.

[0072] In this embodiment, the process of synthesizing low-frequency, mid-frequency, and high-frequency fault enhancement signals includes:

[0073] The mean vibration amplitude is calculated for each test rotation speed, and the mean vibration amplitude is subtracted from each vibration amplitude to obtain the preprocessed vibration signal;

[0074] The preprocessed vibration signal is subjected to Fourier transform to extract the corresponding spectrum. The spectrum is divided into low-frequency, mid-frequency, and high-frequency bands. The frequency band spectrum is then subjected to inverse Fourier transform to obtain the low-frequency, mid-frequency, and high-frequency signals (inverse Fourier transform of the low-frequency spectrum yields the low-frequency signal, inverse Fourier transform of the mid-frequency spectrum yields the mid-frequency signal, and inverse Fourier transform of the high-frequency spectrum yields the high-frequency signal).

[0075] For the same frequency band signals corresponding to the three test speeds, the signals are weighted based on the frequency band energy to synthesize the corresponding frequency band fault enhancement signal.

[0076] In this embodiment, the low test speed range is 1000-2000 r / min, the medium test speed range is 2000-4500 r / min, and the high test speed range is >4500 r / min.

[0077] The low-frequency range is set to 0-500Hz, the mid-frequency range is set to 500Hz-1000Hz, and the high-frequency range is set to >1000Hz.

[0078] The low-frequency signals corresponding to the three test speeds are weighted based on the frequency band energy to synthesize a low-frequency fault enhancement signal.

[0079] The mid-frequency signals corresponding to the three test speeds are weighted based on the frequency band energy to synthesize a mid-frequency fault enhancement signal.

[0080] The high-frequency signals corresponding to the three test speeds are weighted based on the frequency band energy to synthesize a high-frequency fault enhancement signal.

[0081] In this embodiment, the weighting formula is:

[0082] ,

[0083] in, Let x be the vibration amplitude of the i-th fault enhancement signal at time t. 1,i,t x represents the vibration amplitude at time t on the signal corresponding to the low test speed. 2,i,t Let x be the vibration amplitude at time t on the signal corresponding to the test rotational speed. 3,i,t e represents the vibration amplitude at time t on the signal corresponding to the high test speed. 1,i,t e represents the energy at time t on the signal corresponding to the low test speed. 2,i,te represents the energy at time t on the signal corresponding to the test rotational speed. 3,i,t e represents the energy at time t on the signal corresponding to the high test speed. total,i,t For e 1,i,t e 2,i,t and e 3,i,t The sum of , where i takes the values ​​1, 2, and 3, corresponding to low frequency, mid frequency, and high frequency respectively, and t is the time number.

[0084] In this embodiment, the energy at time t on the signal is calculated by squaring the vibration amplitude at time t on the signal corresponding to a certain test rotation speed to obtain the energy at time t on the signal corresponding to the test rotation speed.

[0085] This invention calculates the mean of vibration signals at each test speed and performs mean-removal processing to eliminate DC component (baseline) interference in the signal. The preprocessed vibration signal is then subjected to Fourier transform to extract the spectrum and divide it into low-frequency, mid-frequency, and high-frequency bands. Inverse Fourier transform is then used to obtain the signals for each frequency band, separating the low-frequency, mid-frequency, and high-frequency fault characteristics generated by the blood pump under different operating conditions (corresponding to different speeds and fault types). This invention uses a weighted average of the signals in the same frequency band at the three test speeds to synthesize a fault enhancement signal for that frequency band based on the frequency band energy. The weighting formula adjusts the weight of the signals at each speed by utilizing the energy proportion of the signals at different speeds, ensuring that signals with higher energy (i.e., more pronounced fault characteristics) have a higher proportion in the synthesized fault enhancement signal. This effectively enhances fault characteristics, suppresses the influence of noise and irrelevant signals, improves the saliency of fault characteristics, and facilitates more accurate monitoring and evaluation of the blood pump's operating status.

[0086] In this embodiment, the process of extracting the energy anomaly composite value and the extreme value anomaly composite value includes:

[0087] The fault enhancement signal is subjected to Hilbert transform to obtain the analytical signal;

[0088] Extract the envelope signal from the analytic signal;

[0089] Both the envelope signal and the fault enhancement signal are processed in segments.

[0090] The energy of each segment of the envelope signal and the energy of each segment of the fault enhancement signal are calculated to obtain the composite energy value of each segment.

[0091] Based on the extreme values ​​of each segment of the envelope signal and the extreme values ​​of each segment of the fault enhancement signal, the extreme value anomaly synthesis value of each segment is obtained, where the extreme values ​​include the maximum amplitude value.

[0092] In this embodiment, the process of extracting the energy anomaly composite value and the extreme value anomaly composite value is for signals in the same frequency band. For example, when the fault enhancement signal is a low-frequency fault enhancement signal, the corresponding envelope signal is obtained by processing the low-frequency fault enhancement signal.

[0093] The formula for extracting the envelope signal is:

[0094] ,

[0095] Where z(t) is the envelope signal, x(t) is the fault enhancement signal, H[x(t)] is the analytic signal, and H is the Hilbert transform.

[0096] This invention obtains the analytical signal and extracts the envelope signal through Hilbert transform. The envelope signal can effectively reflect the amplitude change trend of the fault enhancement signal and highlight the amplitude envelope characteristics of the signal. The envelope signal and the fault enhancement signal are segmented. Considering that the vibration signal of the ECMO device blood pump may have non-stationary characteristics and the signal characteristics may differ in different time periods, after segmentation, energy and extreme value analysis can be performed on each segment of the signal, which can more meticulously explore the abnormal situation in different time periods.

[0097] In this embodiment, when processing in segments, the segment length can be set to 50 sampling points (corresponding to a duration of 10 milliseconds).

[0098] In this embodiment, the process of obtaining the energy anomaly synthesis value for each segment includes:

[0099] Calculate the energy for each segment of the envelope signal to obtain the envelope energy for each segment of the envelope signal;

[0100] The energy of each segment of the fault enhancement signal is calculated to obtain the enhancement energy of each segment of the fault enhancement signal;

[0101] The first energy anomaly coefficient is extracted based on the envelope energy of each segment of the envelope signal.

[0102] Based on the enhanced energy of each segment of the fault enhancement signal, the second energy anomaly coefficient is extracted;

[0103] The combined energy anomaly value is obtained by adding the first and second energy anomaly coefficients that belong to the same time period.

[0104] In this embodiment, the method for calculating the energy of each segment of the envelope signal and each segment of the fault enhancement signal is as follows: for each segment of the signal, the sum of the squares of all its sampling points is calculated, which is the energy.

[0105] In this embodiment, the process of extracting the first energy anomaly coefficient includes: taking the average value of the envelope energy of each segment to obtain the average envelope energy; subtracting the average envelope energy from the envelope energy of each segment and taking the absolute value to obtain the envelope energy deviation; and using the ratio of the envelope energy deviation to the average envelope energy as the first energy anomaly coefficient.

[0106] The process of extracting the second energy anomaly coefficient includes: taking the average value of the enhanced energy of each segment to obtain the enhanced energy mean; subtracting the enhanced energy mean from the enhanced energy of each segment and taking the absolute value to obtain the enhanced energy deviation; and using the ratio of the enhanced energy deviation to the enhanced energy mean as the second energy anomaly coefficient.

[0107] This invention considers energy anomalies in both the envelope signal and the fault enhancement signal. The envelope signal reflects the amplitude variation trend of the fault enhancement signal, while the fault enhancement signal contains vibration energy information at three test speeds. Extracting and synthesizing energy anomaly coefficients from these two dimensions enables a more comprehensive capture of energy anomalies caused by different faults (such as bearing wear, impeller imbalance, etc.) during blood pump operation. This avoids the anomaly omissions that may occur with single-signal-dimensional analysis and improves the comprehensiveness of energy anomaly identification.

[0108] In this embodiment, the process of obtaining the composite extreme value of each segment includes:

[0109] Extract the first extreme value anomaly coefficient based on each extreme value of the envelope signal;

[0110] Extract the second extreme value anomaly coefficient based on each extreme value of the fault enhancement signal;

[0111] The extreme anomaly coefficients belonging to the same time period are added together to obtain the composite extreme anomaly value.

[0112] In this embodiment, the process of extracting the first extreme value anomaly coefficient includes:

[0113] The mean value of the envelope amplitude is obtained by taking the average value of each amplitude of the envelope signal. The mean value of the envelope amplitude is obtained by subtracting the mean value of the envelope amplitude from each extreme value of the envelope signal and taking the absolute value. The ratio of the mean value of the envelope amplitude to the mean value of the envelope amplitude is used as the first extreme value anomaly coefficient.

[0114] The process of extracting the coefficient of the second extreme value anomaly includes:

[0115] The mean of the enhanced amplitude is obtained by taking the average value of each amplitude of the fault enhancement signal. The mean of the enhanced amplitude is obtained by subtracting the mean of the enhanced amplitude from each extreme value of the fault enhancement signal and taking the absolute value. The ratio of the extreme value deviation to the mean of the enhanced amplitude is used as the second extreme value anomaly coefficient.

[0116] This invention can simultaneously combine extreme value anomalies of envelope signals (reflecting amplitude change trends) and fault enhancement signals (preserving original vibration characteristics). By using the "ratio of extreme value deviation to mean value," the anomalies of the two types of signals are quantified into coefficients and added together. This not only accurately reflects the degree of deviation of each signal extreme value from the overall level, but also avoids fault information that may be missed by single signal analysis. It can also offset the influence of the overall amplitude difference of the signal under different operating conditions. In ECMO blood pump monitoring, it can more accurately and comprehensively identify extreme value anomalies caused by mechanical wear, loose parts, etc., and is not affected by external factors such as speed fluctuations and changes in blood viscosity.

[0117] Energy reflects the "sum of signal intensity" over a given range. For example, during ECMO blood pump operation, when the pump experiences mechanical wear, loose components, or other malfunctions, the energy of the vibration signal often changes. This could be a sudden increase in localized energy (such as additional energy generated by increased friction) or an abnormal change in energy distribution. By extracting the composite value of energy anomalies, we can determine from the perspective of overall energy whether there are abnormal energy fluctuations caused by malfunctions during blood pump operation. This is crucial for identifying malfunctions that cause continuous changes in signal energy (such as long-term bearing wear, which may cause vibration energy to gradually accumulate and increase).

[0118] Extreme values ​​focus on the "maximum amplitude" of a signal within a certain range. During blood pump operation, malfunctions such as impeller jamming or momentary seal failure may generate large-amplitude pulse vibrations within a short period. In such cases, the signal's extreme values ​​will exhibit obvious abnormal peaks or troughs. Extracting composite values ​​of extreme anomalies can accurately capture these instantaneous, large-amplitude abnormal changes.

[0119] In this embodiment, the size of the energy anomaly splicing matrix and the extreme value anomaly splicing matrix is ​​3×M. Since the energy anomaly composite value and the extreme value anomaly composite value are obtained for each signal segment, there are multiple energy anomaly composite values ​​and extreme value anomaly composite values. These are spliced ​​into a matrix of size 3×M, where M is the number of segments, i.e., the number of rows in the matrix, and 3 is the number of columns in the matrix. For example, the energy anomaly composite value of the first signal segment corresponding to the low-frequency, medium-frequency, and high-frequency fault enhancement signals is taken as the first row element of the energy anomaly splicing matrix, and so on.

[0120] like Figure 2 As shown, the running state evaluation neural network includes: a first shallow feature extraction unit, a second shallow feature extraction unit, a shared weight generation unit, a multiplier M1, a multiplier M2, a first deep feature extraction unit, a second deep feature extraction unit, an adder A1, and a fully connected layer;

[0121] The input of the first shallow feature extraction unit is used to input the energy anomaly splicing matrix, and its output is connected to the first input of the multiplier M1 and the first input of the shared weight generation unit, respectively.

[0122] The input of the second shallow feature extraction unit is used to input the extreme value anomaly concatenation matrix, and its output is connected to the first input of multiplier M2 and the second input of the shared weight generation unit, respectively. The output of the shared weight generation unit is connected to the second input of multiplier M1 and the second input of multiplier M2, respectively. The output of multiplier M1 is connected to the input of the first deep feature extraction unit. The output of multiplier M2 is connected to the input of the second deep feature extraction unit. The input of adder A1 is connected to the output of the first deep feature extraction unit and the output of the second deep feature extraction unit, respectively, and its output is connected to the input of the fully connected layer. The output of the fully connected layer serves as the output of the running state evaluation neural network.

[0123] This invention performs shallow feature extraction on the energy anomaly splicing matrix and the extreme value anomaly splicing matrix respectively, which can fully explore the basic features of the two different types of abnormal data. The shared weight generation unit allows the two branches to generate shared weights during the feature extraction process, which helps to increase the attention to "simultaneous energy and extreme value anomalies". Then, deep features are extracted by the first deep feature extraction unit and the second deep feature extraction unit respectively. Feature fusion is achieved by the adder A1. The operating status score of the blood pump in the ECMO device is output through the fully connected layer.

[0124] like Figure 3 As shown, the shared weight generation unit includes: a first convolutional layer, a second convolutional layer, an adder A2, and a first sigmoid layer;

[0125] The input of the first convolutional layer serves as the first input of the shared weight generation unit; the input of the second convolutional layer serves as the second input of the shared weight generation unit.

[0126] The input of adder A2 is connected to the output of the first convolutional layer and the output of the second convolutional layer, and its output is connected to the input of the first sigmoid layer: the output of the first sigmoid layer serves as the output of the shared weight generation unit.

[0127] The kernel size of both the first and second convolutional layers is 1×1.

[0128] The first and second convolutional layers perform convolution operations on different inputs (such as shallow features related to energy and extrema) to extract their respective feature representations. The outputs of the two layers are added together by adder A2, achieving preliminary fusion and interaction of features from different sources. Then, the first sigmoid layer generates weights to enhance features that exhibit "simultaneous anomalies in energy and extrema".

[0129] The first shallow feature extraction unit and the second shallow feature extraction unit have the same structure, both including: a shallow feature extraction module and a feature enhancement module. The input end of the shallow feature extraction module serves as the input end of the first shallow feature extraction unit and the second shallow feature extraction unit. The output end of the shallow feature extraction module is connected to the input end of the feature enhancement module, and the output end of the feature enhancement module serves as the output end of the first shallow feature extraction unit and the second shallow feature extraction unit.

[0130] like Figure 4 As shown, the shallow feature extraction module includes: a third convolutional layer, a fourth convolutional layer, a first concat layer, and a fifth convolutional layer. The kernel size of the third convolutional layer is 1×1, the kernel size of the fourth convolutional layer is 3×3, and the kernel size of the fifth convolutional layer is 3×3.

[0131] This invention employs parallel processing of the third and fourth convolutional layers to extract features at different scales, enriching the features. The features are then concatenated through the first concat layer, and further extracted and optimized through the fifth convolutional layer.

[0132] like Figure 5 As shown, the feature enhancement module includes: a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a softmax layer, a second sigmoid layer, multiplier M1, and multiplier M2.

[0133] The kernel size of the sixth, seventh, and eighth convolutional layers is 1×1. After the seventh convolutional layer extracts features, attention weights are generated through a softmax layer and then multiplied element-wise with the output of the sixth convolutional layer (M1 operation). This allows the model to focus on key features, suppress irrelevant information, and increase attention to important content. The second sigmoid layer compresses the output of the eighth convolutional layer to the 0-1 range, introducing non-linearity to make the features more discriminative. It can also be used to generate gating-like mechanisms to further adjust the transfer and utilization of features. This invention combines the processed features with the original input features through the M2 operation to achieve secondary feature enhancement and fully utilize feature information at different stages.

[0134] like Figure 6 As shown, the first deep feature extraction unit and the second deep feature extraction unit have the same structure, both including: a ninth convolutional layer, a tenth convolutional layer, an adder A3, a max pooling layer, an average pooling layer, and a second Concat layer.

[0135] The ninth and tenth convolutional layers extract deep features, which are then fused through the A3 operation (element-wise addition) to integrate multi-scale feature information, forming a residual structure to prevent gradient vanishing. The maximum value is then extracted through the max pooling layer and the mean value is extracted through the average pooling layer, reducing the amount of data.

[0136] The kernel size of the ninth and tenth convolutional layers is 3×3.

[0137] Normal operating condition (score range: 8-10 points): The blood pump has no mechanical faults, the impeller rotates smoothly without jamming, the bearings are not worn, and the seals are in good condition.

[0138] Mild abnormal condition (score range: 6-8 points): The blood pump has slight mechanical wear, such as slight insufficient lubrication of the bearings or a small amount of thrombus on the impeller surface (which does not affect blood flow).

[0139] Moderate abnormal condition (scoring range: 4-6 points): Obvious mechanical failure of blood pump, such as moderate wear of bearing (increased rotational resistance) and local wear of impeller (affecting blood flow uniformity).

[0140] Severe abnormal condition (scoring range: 0-4 points): The blood pump has a serious mechanical failure, such as a stuck bearing (the impeller cannot rotate normally), a severely deformed / broken impeller (leading to the risk of blood flow interruption), or a completely damaged seal (large amount of blood / liquid leakage).

[0141] In this embodiment, the running state evaluation neural network is trained using the existing gradient descent method.

[0142] This invention collects vibration signals from a blood pump at different speeds using a vibration sensor. By combining mean-reduction processing with frequency band energy-based signal weighting, it synthesizes low-frequency, mid-frequency, and high-frequency fault enhancement signals. This can keenly detect early anomalies caused by subtle faults such as mechanical wear and loose parts, thus overcoming the shortcomings of existing technologies that rely on macroscopic parameter monitoring to detect early faults.

[0143] Compared to existing technologies, which are easily affected by external factors such as changes in patient position and differences in blood viscosity, the vibration signal-based monitoring and processing method of this invention is less affected by such external factors, effectively reducing interference and improving the stability and reliability of monitoring.

[0144] This invention acquires the envelope signal of the fault enhancement signal, and then combines the energy and extreme values ​​of the fault enhancement signal and the envelope signal to extract the composite values ​​of energy anomalies and extreme value anomalies, reflecting the energy anomalies and extreme value anomalies. This not only preserves the dynamic mechanical operation information contained in the fault enhancement signal, but also captures the amplitude modulation characteristics of the signal through envelope analysis. The composite values ​​of energy anomalies and extreme value anomalies of low-frequency, mid-frequency and high-frequency fault enhancement signals are then concatenated into a matrix. The energy anomaly concatenation matrix and extreme value anomaly concatenation matrix are processed by a running state evaluation neural network, avoiding information loss caused by mutual interference of features in different frequency bands. It has a stronger ability to identify complex anomalies caused by the coordinated failure of multiple components in blood pumps.

[0145] The operational status assessment neural network can establish a nonlinear mapping relationship between features and the health status of the blood pump by deeply mining the splicing matrix of energy anomalies and extreme value anomalies, breaking through the limitations of traditional threshold methods or simple waveform comparisons.

[0146] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An ECMO device operating state monitoring method, characterized by, The method comprises the following steps: Collecting vibration signals of the blood pump in the ECMO device at low, medium and high test rotating speeds through a vibration sensor; Performing mean value processing on the vibration signals at the three test rotating speeds, and synthesizing low, medium and high frequency fault enhancement signals based on frequency band energy; Obtaining an envelope signal from the fault enhancement signal, and extracting an energy anomaly synthesis value and an extreme value anomaly synthesis value from each segment of the fault enhancement signal and the envelope signal; The process of obtaining the energy anomaly synthesis value of each segment comprises: Calculating the energy of each segment of the envelope signal to obtain the envelope energy of each segment of the envelope signal; Calculating the energy of each segment of the fault enhancement signal to obtain the enhancement energy of each segment of the fault enhancement signal; Extracting a first energy anomaly coefficient according to the envelope energy of each segment of the envelope signal; Extracting a second energy anomaly coefficient according to the enhancement energy of each segment of the fault enhancement signal; Adding the first energy anomaly coefficient and the second energy anomaly coefficient belonging to the same time segment to obtain the energy anomaly synthesis value; The process of extracting the first energy anomaly coefficient comprises: obtaining the mean value of the envelope energy of each segment to obtain the mean value of the envelope energy, subtracting the mean value of the envelope energy from the envelope energy of each segment, taking the absolute value to obtain the envelope energy deviation, and taking the ratio of the envelope energy deviation to the mean value of the envelope energy as the first energy anomaly coefficient; The process of extracting the second energy anomaly coefficient comprises: obtaining the mean value of the enhancement energy of each segment to obtain the mean value of the enhancement energy, subtracting the mean value of the enhancement energy from the enhancement energy of each segment, taking the absolute value to obtain the enhancement energy deviation, and taking the ratio of the enhancement energy deviation to the mean value of the enhancement energy as the second energy anomaly coefficient; The process of obtaining the extreme value anomaly synthesis value of each segment comprises: Extracting a first extreme value anomaly coefficient according to the extreme value of each segment of the envelope signal, wherein the extreme value comprises a maximum amplitude; Extracting a second extreme value anomaly coefficient according to the extreme value of each segment of the fault enhancement signal; Adding the first extreme value anomaly coefficient and the second extreme value anomaly coefficient belonging to the same time segment to obtain the extreme value anomaly synthesis value; The process of extracting the first extreme value anomaly coefficient comprises: Obtaining the mean value of each amplitude of the envelope signal to obtain the mean value of the envelope amplitude, subtracting the mean value of the envelope amplitude from the extreme value of each segment of the envelope signal, taking the absolute value to obtain the envelope extreme value deviation, and taking the ratio of the envelope extreme value deviation to the mean value of the envelope amplitude as the first extreme value anomaly coefficient; The process of extracting the second extreme value anomaly coefficient comprises: Obtaining the mean value of each amplitude of the fault enhancement signal to obtain the mean value of the enhancement amplitude, subtracting the mean value of the enhancement amplitude from the extreme value of each segment of the fault enhancement signal, taking the absolute value to obtain the enhancement extreme value deviation, and taking the ratio of the enhancement extreme value deviation to the mean value of the enhancement amplitude as the second extreme value anomaly coefficient; Splicing the energy anomaly synthesis values of the low, medium and high frequency fault enhancement signals to obtain an energy anomaly splicing matrix; Splicing the extreme value anomaly synthesis values of the low, medium and high frequency fault enhancement signals to obtain an extreme value anomaly splicing matrix; Processing the energy anomaly splicing matrix and the extreme value anomaly splicing matrix by using a running state evaluation neural network to obtain a running state score of the blood pump in the ECMO device.

2. The ECMO device operating state monitoring method according to claim 1, characterized in that, The process of synthesizing the low, medium and high frequency fault enhancement signals comprises: The mean value of the vibration amplitude is calculated for the vibration signal of each test speed, each vibration amplitude is subtracted from the mean value of the vibration amplitude to obtain a preprocessed vibration signal; The preprocessed vibration signal is subjected to Fourier transform, the corresponding spectrum is extracted, the spectrum is divided into low-frequency spectrum, medium-frequency spectrum and high-frequency spectrum, and the spectrum is subjected to inverse Fourier transform to obtain low-frequency signal, medium-frequency signal and high-frequency signal; The same frequency band signals corresponding to the three test speeds are weighted based on the frequency band energy to synthesize the fault enhanced signal of the corresponding frequency band.

3. The ECMO device operating state monitoring method according to claim 2, characterized in that, The weighting formula is: ， wherein, is the vibration amplitude of the ith fault enhanced signal at time t, x 1,i,t is the vibration amplitude of the corresponding signal at time t at low test speed, x 2,i,t is the vibration amplitude of the corresponding signal at time t at medium test speed, x 3,i,t is the vibration amplitude of the corresponding signal at time t at high test speed, e 1,i,t is the energy of the corresponding signal at time t at low test speed, e 2,i,t is the energy of the corresponding signal at time t at medium test speed, e 3,i,t is the energy of the corresponding signal at time t at high test speed, e total,i,t is the sum of e 1,i,t , e 2,i,t and e 3,i,t , i takes 1, 2, 3 respectively for low, medium and high frequencies, and t is the time index.

4. The ECMO device operating state monitoring method according to claim 1, characterized by, The process of extracting the energy anomaly synthesis value and the extreme value anomaly synthesis value includes: The Hilbert transform is performed on the fault enhanced signal to obtain an analytic signal; The envelope signal is extracted from the analytic signal; The envelope signal and the fault enhanced signal are both subjected to segmentation processing; The energy of each segment of the envelope signal and the energy of each segment of the fault enhanced signal are calculated to obtain the energy anomaly synthesis value of each segment; The extreme value anomaly synthesis value of each segment is obtained according to the extreme value of each segment of the envelope signal and the extreme value of each segment of the fault enhanced signal.

5. The ECMO device operating state monitoring method according to claim 1, characterized in that, The running state evaluation neural network comprises a first shallow feature extraction unit, a second shallow feature extraction unit, a shared weight generation unit, a multiplier M1, a multiplier M2, a first deep feature extraction unit, a second deep feature extraction unit, an adder A1 and a full connection layer; The input end of the first shallow feature extraction unit is used for inputting the energy anomaly splicing matrix, and the output end thereof is connected with the first input end of the multiplier M1 and the first input end of the shared weight generation unit respectively; The input end of the second shallow feature extraction unit is used for inputting the extreme value anomaly splicing matrix, and the output end thereof is connected with the first input end of the multiplier M2 and the second input end of the shared weight generation unit respectively; The output end of the shared weight generation unit is connected with the second input end of the multiplier M1 and the second input end of the multiplier M2 respectively; The output end of the multiplier M1 is connected with the input end of the first deep feature extraction unit; The output end of the multiplier M2 is connected with the input end of the second deep feature extraction unit; The input end of the adder A1 is connected with the output end of the first deep feature extraction unit and the output end of the second deep feature extraction unit respectively, and the output end thereof is connected with the input end of the full connection layer; The output end of the full connection layer serves as the output end of the running state evaluation neural network.

6. The ECMO device operating state monitoring method according to claim 5, characterized in that, The shared weight generation unit comprises a first convolutional layer, a second convolutional layer, an adder A2 and a first Sigmoid layer; The input end of the first convolutional layer serves as the first input end of the shared weight generation unit; The input end of the second convolutional layer serves as the second input end of the shared weight generation unit; The input end of the adder A2 is connected with the output end of the first convolutional layer and the output end of the second convolutional layer respectively, and the output end thereof is connected with the input end of the first Sigmoid layer; The output end of the first Sigmoid layer serves as the output end of the shared weight generation unit.

Images