An ECMO device fault diagnosis method
By collecting and converting vibration signals into angular domain signals in ECMO devices, and combining cyclic energy spectrum and entropy feature matrix, multi-scale features are extracted using CNN network, which solves the problem of low fault diagnosis accuracy in existing technologies and achieves highly sensitive identification and accurate diagnosis of early faults.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN ZHONGSHI INSTR TECH CO LTD
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-24
AI Technical Summary
Existing ECMO equipment fault diagnosis methods are unable to identify minute fault features inside the mechanical structure in the early stages, and are insufficient in extracting features of non-stationary vibration signals and complex coupled faults, resulting in low diagnostic accuracy.
Vibration signals from the ECMO device's blood pump were collected at three test speeds, converted into angular domain signals, and the cyclic energy spectrum was calculated. Harmonic order energy vectors and entropy feature matrices were constructed, and multi-scale features were extracted using a CNN network. The abnormal energy growth rate matrix was then fused to identify the blood pump failure type.
It improves the accuracy of fault identification, enabling early detection of minor abnormalities such as mechanical wear and bearing imbalance, reducing medical risks, and significantly enhancing the salience and stability of fault characteristics.
Smart Images
Figure CN121401525B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, and specifically to a fault diagnosis method for ECMO equipment. Background Technology
[0002] Currently, extracorporeal membrane oxygenation (ECMO) devices are widely used in critical cardiopulmonary support. The operational stability of one of its core components—the centrifugal blood pump—is directly related to patient safety. During prolonged high-speed operation, the blood pump is highly susceptible to mechanical wear, cavitation, bearing imbalance, and fluid disturbances, leading to abnormal vibration characteristics, which can cause flow fluctuations, blood contamination, and even equipment failure.
[0003] Existing ECMO equipment fault diagnosis methods mainly rely on the analysis of changes in electrical signals such as current, voltage, and rotational speed, or system parameters such as flow rate and pressure. However, these signals often reflect the macroscopic results of the fault and lack a sensitive response to the internal operating state of the mechanical structure, making it difficult to identify early fault characteristics in a timely manner. In addition, some studies have used frequency domain or time-frequency domain analysis methods to process vibration signals, such as Fast Fourier Transform (FFT), wavelet analysis, and Short-Time Fourier Transform (STFT). Although these methods can reveal the frequency characteristics of equipment vibration, they suffer from insufficient feature extraction and indistinct fault characteristics when dealing with non-stationary vibration signals and complex coupled faults, resulting in low accuracy in ECMO equipment fault diagnosis. Summary of the Invention
[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides an ECMO equipment fault diagnosis method that solves the problem of low accuracy in ECMO equipment fault diagnosis in the prior art.
[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a fault diagnosis method for ECMO equipment, comprising the following steps:
[0006] Vibration signals of the blood pump in the ECMO device during operation were collected using a vibration sensor at three test speeds.
[0007] Based on the fundamental frequency of the corresponding test rotation speed, the corresponding vibration signal is converted into an angle domain signal, and the cyclic energy spectrum of each segment of the angle domain signal is obtained.
[0008] In each cycle energy spectrum, the harmonic order energy under the same harmonic order is calculated, and the harmonic order energy vector is constructed.
[0009] Entropy feature values are obtained at different positions in the harmonic order energy vector, and an entropy feature matrix is constructed at the corresponding test speed.
[0010] Anomalies are obtained at different locations in the harmonic order energy vector, the abnormal energy is calculated, and the abnormal energy feature matrix at the corresponding test speed is constructed.
[0011] Based on the ratios of medium test speed to low test speed and high test speed to medium test speed on the abnormal energy characteristic matrix, the abnormal energy growth rate matrices of medium-low speed and high-medium speed are obtained.
[0012] A CNN network was used to extract and fuse features from three entropy feature matrices and two abnormal energy growth rate matrices. Multi-scale features were then extracted from the two fused features, and the features were enhanced and classified according to scale to obtain the blood pump failure types in ECMO devices.
[0013] Furthermore, the three test speeds include: low test speed, medium test speed, and high test speed. At the low test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the low test speed. At the medium test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the medium test speed. At the high test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the high test speed.
[0014] 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.
[0015] Furthermore, the process of obtaining the cyclic energy spectrum includes:
[0016] Based on the test rotation speed, obtain the fundamental frequency corresponding to the test rotation speed;
[0017] The instantaneous angle is obtained based on the fundamental frequency;
[0018] The vibration signal is resampled according to the instantaneous angle to obtain the angle domain signal;
[0019] The angle domain signal is divided into multiple segments, and a Fourier transform is performed on each segment to obtain the cyclic energy spectrum of each segment of the angle domain signal.
[0020] Furthermore, the formula for the cyclic energy spectrum of each segment of the angle domain signal is:
[0021] ,
[0022] Where E(α,ω) is the cyclic energy at (α,ω) in the cyclic energy spectrum of each segment of the angle domain signal. Let be the spectral value of each segment of the angle-domain signal at ω+α / 2 after Fourier transform. for The complex conjugate, Let be the spectral value of each segment of the angular domain signal at ω-α / 2 after Fourier transform, || be the modulus of the complex number, ω be the angular frequency, and α be the harmonic order. The value of α is a positive integer from 1 to 4.
[0023] Furthermore, the process of constructing the harmonic order energy vector includes:
[0024] In each cycle energy spectrum, the harmonic order energy under the same harmonic order is calculated, where the harmonic order energy is the result of adding the cycle energies belonging to the same harmonic order;
[0025] The harmonic order energies of the same harmonic order belonging to the same vibration signal are arranged in chronological order to obtain the harmonic order energy vector.
[0026] Furthermore, the process of constructing the entropy feature matrix includes:
[0027] Add up all the harmonic order energies in the harmonic order energy vector to get the total order energy;
[0028] Divide the energy of each harmonic order in the harmonic order energy vector by the total order energy to obtain the normalized harmonic order energy vector.
[0029] Set a sliding window to slide on the normalized harmonic order energy vector and calculate the Shannon entropy under each sliding window.
[0030] Normalize the Shannon entropy to the range of 0 to 1 to obtain the entropy eigenvalues;
[0031] Using the entropy eigenvalues corresponding to the same harmonic order as elements of each row, an entropy feature matrix is constructed for the corresponding test speed.
[0032] Furthermore, the process of constructing the anomalous energy feature matrix includes:
[0033] The mean value of the vector is obtained by taking the average value of each harmonic order energy in each harmonic order energy vector.
[0034] Anomalies are identified by marking the harmonic order energies in the harmonic order energy vector that are greater than the mean of the vector.
[0035] Set a sliding window to slide on the harmonic order energy vector and extract outliers under each sliding window;
[0036] The harmonic order energies of the anomaly points under the sliding window are added together to obtain the cumulative value of the anomaly energy;
[0037] The cumulative abnormal energy value corresponding to the same harmonic order is used as the element of each row to form the abnormal energy feature matrix at the corresponding test speed.
[0038] Furthermore, the process of obtaining the abnormal energy growth rate matrices for low-to-medium speed and high-to-medium speed includes:
[0039] The ratio of corresponding elements of the abnormal energy characteristic matrix at the medium test speed to the abnormal energy characteristic matrix at the low test speed is calculated to obtain the abnormal energy growth rate matrix at medium to low speeds.
[0040] The ratio of corresponding elements in the abnormal energy characteristic matrix at high test speed to that at medium test speed is calculated to obtain the abnormal energy growth rate matrix at high-medium speed.
[0041] Furthermore, the process of determining the blood pump failure types in ECMO devices includes the following:
[0042] A CNN network was used to process the entropy feature matrix under low test speed, medium test speed, and high test speed respectively, to obtain the first entropy feature tensor, the second entropy feature tensor, and the third entropy feature tensor.
[0043] The first entropy feature tensor, the second entropy feature tensor, and the third entropy feature tensor are added element by element to obtain the entropy feature fusion tensor;
[0044] A CNN network was used to process the abnormal energy growth rate matrix of medium-low speed and the abnormal energy growth rate matrix of high-medium speed, respectively, to obtain the first growth rate feature tensor and the second growth rate feature tensor.
[0045] The first growth rate feature tensor and the second growth rate feature tensor are added element by element to obtain the growth rate feature fusion tensor;
[0046] Feature extraction is performed on the entropy feature fusion tensor and the growth rate feature fusion tensor at two different scales. The entropy feature and the growth rate feature at the same scale are then multiplied to obtain the first enhanced feature and the second enhanced feature.
[0047] The first and second enhanced features are concatenated, and a classifier is used to classify the concatenated features.
[0048] Furthermore, the process of obtaining the first and second enhancement features includes:
[0049] The first-scale feature extraction module is used to extract first-scale entropy features from the entropy feature fusion tensor;
[0050] The second-scale feature extraction module is used to extract second-scale entropy features from the entropy feature fusion tensor;
[0051] The third-scale feature extraction module is used to extract the first-scale growth rate feature by fusing the growth rate feature tensor;
[0052] The fourth-scale feature extraction module is used to extract the second-scale growth rate feature by fusing the growth rate feature with tensors;
[0053] Multiply the first-scale entropy feature and the first-scale growth rate feature to obtain the first enhanced feature;
[0054] Multiplying the second-scale entropy feature and the second-scale growth rate feature yields the second enhanced feature.
[0055] The beneficial effects of this invention are as follows:
[0056] 1. This invention correlates vibration signals with the fundamental frequency of rotational speed through angle domain transformation, extracts harmonic order energy by combining cyclic energy spectrum, and then fuses multi-dimensional features such as entropy characteristics, abnormal energy, and abnormal energy growth rate between rotational speeds to comprehensively capture the essential information of the fault. Combined with the multi-scale feature extraction and fusion capabilities of neural networks, it solves the problem of insufficient feature extraction for non-stationary signals and complex coupled faults in existing methods, significantly improving the accuracy of fault identification.
[0057] 2. This invention converts the vibration signal to the angular domain and calculates the cyclic energy spectrum and harmonic order energy in the angular domain, thereby eliminating the influence of rotational speed fluctuations on feature extraction. This results in extracted features with higher periodic consistency and structural stability, enabling a more accurate reflection of the dynamic behavior of the blood pump under different operating conditions. By introducing entropy features and abnormal energy into the harmonic order energy vector, this invention not only characterizes the energy distribution pattern of the blood pump vibration signal but also enhances the sensitivity to abnormal disturbances and local instability, allowing for the effective identification of early minor faults (such as bearing wear and early cavitation).
[0058] 3. This invention focuses on the internal vibration response of the blood pump's mechanical structure. Angular domain signals and cyclic energy spectra are highly sensitive to early, minute anomalies such as mechanical wear and bearing imbalance. Compared to existing methods that rely on electrical signals and system parameters, this invention can capture characteristic changes in the early stages of a fault earlier, reducing the medical risks caused by equipment failure.
[0059] 4. This invention covers three test speed scenarios and achieves angular domain standardization of vibration signals through fundamental frequency adaptation, eliminating the interference of speed fluctuations on feature extraction. The abnormal energy growth rate matrix further reflects the evolution law of faults under different speeds, highlighting fault characteristics and making the fault characteristics significant. Attached Figure Description
[0060] Figure 1 A flowchart of a fault diagnosis method for ECMO equipment;
[0061] Figure 2 A flowchart for obtaining the entropy feature fusion tensor;
[0062] Figure 3A flowchart for obtaining the growth rate feature fusion tensor;
[0063] Figure 4 A flowchart for obtaining the first and second enhancement features;
[0064] Figure 5 This is a schematic diagram of the first-scale feature extraction module and the third-scale feature extraction module.
[0065] Figure 6 This is a schematic diagram of the second-scale feature extraction module and the fourth-scale feature extraction module. Detailed Implementation
[0066] 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.
[0067] like Figure 1 As shown, a fault diagnosis method for ECMO equipment includes the following steps:
[0068] Vibration signals of the blood pump in the ECMO device during operation were collected using a vibration sensor at three test speeds.
[0069] Based on the fundamental frequency of the corresponding test rotation speed, the corresponding vibration signal is converted into an angle domain signal, and the cyclic energy spectrum of each segment of the angle domain signal is obtained.
[0070] In each cycle energy spectrum, the harmonic order energy under the same harmonic order is calculated, and the harmonic order energy vector is constructed.
[0071] Entropy feature values are obtained at different positions in the harmonic order energy vector, and an entropy feature matrix is constructed at the corresponding test speed.
[0072] Anomalies are obtained at different locations in the harmonic order energy vector, the abnormal energy is calculated, and the abnormal energy feature matrix at the corresponding test speed is constructed.
[0073] Based on the ratios of medium test speed to low test speed and high test speed to medium test speed on the abnormal energy characteristic matrix, the abnormal energy growth rate matrices of medium-low speed and high-medium speed are obtained.
[0074] A CNN network was used to extract and fuse features from three entropy feature matrices and two abnormal energy growth rate matrices. Multi-scale features were then extracted from the two fused features, and the features were enhanced and classified according to scale to obtain the blood pump failure types in ECMO devices.
[0075] In this embodiment, the three test speeds include: low test speed, medium test speed, and high test speed. At the low test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the low test speed. At the medium test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the medium test speed. At the high test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the high test speed. This invention collects the vibration signals of the blood pump corresponding to the three fixed test speeds.
[0076] The low test speed range for the blood pump 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] In this embodiment, the vibration sensor is typically placed on the housing of the blood pump or on a clamp that fits tightly against the blood pump.
[0078] In this embodiment, the process of obtaining the cyclic energy spectrum includes:
[0079] Based on the test rotational speed, the fundamental frequency corresponding to the test rotational speed is obtained: f r,i =n i / 60, where f r,i Let n be the fundamental frequency of the i-th test rotational speed. i For the i-th test speed, the value of i is 1, 2, or 3. When i is 1, n1 is the low test speed; when i is 2, n2 is the medium test speed; and when i is 3, n3 is the high test speed.
[0080] Based on the fundamental frequency, the instantaneous angle is obtained: θ i (t)=2πf r,i t, where θ i (t) represents the instantaneous angle at time t under the i-th test rotational speed, where t is the time.
[0081] The vibration signal is resampled according to the instantaneous angle to obtain the angle domain signal;
[0082] The angle domain signal is divided into multiple segments, and a Fourier transform is performed on each segment to obtain the cyclic energy spectrum of each segment of the angle domain signal.
[0083] ,
[0084] Where E(α,ω) is the cyclic energy at (α,ω) in the cyclic energy spectrum of each segment of the angle domain signal. Let be the spectral value of each segment of the angle-domain signal at ω+α / 2 after Fourier transform. for The complex conjugate, Let be the spectral value of each segment of the angular domain signal at ω-α / 2 after Fourier transform, || be the modulus of the complex number, ω be the angular frequency, and α be the harmonic order. The value of α is a positive integer from 1 to 4.
[0085] In this embodiment, the length of each segment of the angle domain signal is set to include 2 to 5 complete rotor revolutions.
[0086] In this embodiment, for the vibration signal x(t), the signal value corresponding to each angle θ is found by interpolation (such as linear interpolation), and the angle domain signal x(θ) is obtained.
[0087] In this embodiment, the process of constructing the harmonic order energy vector includes:
[0088] In each cycle energy spectrum, the harmonic order energy at the same harmonic order is calculated, where the harmonic order energy is the sum of the cycle energies belonging to the same harmonic order: E i,j,α Let E be the harmonic order energy of the j-th segment of the angle domain signal at the i-th test rotational speed under harmonic order α. i,j (α,ω) is the cyclic energy at (α,ω) in the cyclic energy spectrum of the j-th segment of the angle domain signal at the i-th test speed. The value of i is 1, 2, or 3. When i is 1, it corresponds to a low test speed; when i is 2, it corresponds to a medium test speed; when i is 3, it corresponds to a high test speed. j is the segment number of the angle domain signal.
[0089] Arrange the harmonic order energies of the same harmonic order belonging to the same vibration signal in chronological order to obtain the harmonic order energy vector: E i,α ={E i,1,α ,…, E i,j,α ,…, E i,K,α}, E i,α Let E be the harmonic order energy vector of the angle domain signal at the i-th test rotational speed at harmonic order α. i,1,α Let E be the harmonic order energy of the first segment of the angle domain signal at the i-th test rotational speed at harmonic order α. i,K,α Let K be the harmonic order energy of the Kth segment of the angle domain signal at the i-th test speed, where K is the number of segments in the angle domain signal.
[0090] This invention constructs harmonic order energy vectors to accumulate and sequentially arrange the cyclic energy information of vibration signals in different angular domain segments according to harmonic order, achieving a joint feature expression from the frequency domain to the angular domain. This process not only enhances the energy distribution characteristics under the same order but also reflects the dynamic law of energy change with angle, thereby improving the sensitivity to non-stationary vibration signals and early weak faults, making fault characteristics more significant and more distinguishable.
[0091] In this embodiment, the process of constructing the entropy feature matrix includes:
[0092] Add up all the harmonic order energies in the harmonic order energy vector to get the total order energy;
[0093] Divide the energy of each harmonic order in the harmonic order energy vector by the corresponding total order energy to obtain the normalized harmonic order energy vector.
[0094] Set a sliding window to slide on the normalized harmonic order energy vector and calculate the Shannon entropy under each sliding window.
[0095] Normalize the Shannon entropy to the range of 0 to 1 to obtain the entropy eigenvalues;
[0096] Using the entropy eigenvalues corresponding to the same harmonic order as elements of each row, an entropy feature matrix is constructed for the corresponding test speed.
[0097] In this embodiment, the formula for Shannon entropy is: Where H is the Shannon entropy, W is the sliding window length, and p n Let H' be the energy of the nth normalized harmonic order under the sliding window, where n is a positive integer; the formula for obtaining the entropy eigenvalue is: H′=H / log2(W), where H′ is the entropy eigenvalue.
[0098] The entropy feature matrix is: ,in, , … These are the entropy characteristic values obtained from the energy vector of the harmonic order when the harmonic order is 1 at the i-th test speed. , … The entropy characteristic values are obtained from the energy vector of the harmonic order when the harmonic order is 2 at the i-th test speed. , … The entropy characteristic values are obtained from the energy vector of the harmonic order when the harmonic order is 3 at the i-th test speed. , … For each entropy feature value obtained from the harmonic order energy vector corresponding to the harmonic order of 4 at the i-th test speed, M is the number of columns. The row index of the entropy feature matrix represents the harmonic order, and the column index represents the sliding position of the sliding window.
[0099] This invention, by normalizing the harmonic order energy vector and calculating the Shannon entropy under a sliding window, can characterize the complexity and uncertainty of the energy distribution of different harmonic orders. Shannon entropy reflects the uniformity of signal energy distribution across harmonic orders. When the blood pump operates stably, the harmonic energy distribution is relatively concentrated, and the entropy value is low; however, when faults such as cavitation, bearing wear, or fluid disturbance occur, the harmonic energy distribution tends to be discrete, and the entropy value increases significantly.
[0100] In this embodiment, the process of constructing the abnormal energy feature matrix includes:
[0101] The mean value of the vector is obtained by taking the average value of each harmonic order energy in each harmonic order energy vector.
[0102] Anomalies are identified by marking the harmonic order energies in the harmonic order energy vector that are greater than the mean of the corresponding vector.
[0103] Set a sliding window to slide on the harmonic order energy vector and extract outliers under each sliding window;
[0104] The harmonic order energies of the outliers under the sliding window are added together to obtain the cumulative outlier energy value. When there are no outliers, the cumulative outlier energy value is 0.
[0105] The cumulative abnormal energy value corresponding to the same harmonic order is used as the element of each row to form the abnormal energy feature matrix at the corresponding test speed.
[0106] In this embodiment, the two sliding windows have the same length, set to 5 elements, and the sliding step is 3 elements.
[0107] The expression for the anomalous energy characteristic matrix is: ,in, , … The cumulative abnormal energy values are obtained from the energy vector of the harmonic order corresponding to the first harmonic order at the i-th test speed. , … The cumulative values of each abnormal energy obtained from the harmonic order energy vector corresponding to the harmonic order of 2 at the i-th test speed. , … The cumulative values of each abnormal energy obtained from the harmonic order energy vector corresponding to the third harmonic order at the i-th test speed. , … The cumulative values of each abnormal energy obtained from the harmonic order energy vector corresponding to the harmonic order of 4 at the i-th test speed, M is the number of columns, the row index of the abnormal energy feature matrix represents the harmonic order, and the column index represents the sliding position of the sliding window.
[0108] This invention highlights the regional characteristics of energy abrupt changes during operation by labeling the energy components above the mean in the harmonic order energy vector and using sliding statistics to calculate the accumulated abnormal energy. When the blood pump is in normal operation, the harmonic order energy distribution is stable and changes gradually, with a low accumulated abnormal energy value. However, when faults such as bearing wear, cavitation, or fluid disturbance occur, the energy at specific harmonic orders will show a significant surge, causing a significant increase in the accumulated abnormal energy value.
[0109] In this embodiment, the process of obtaining the abnormal energy growth rate matrix for low-to-medium speed and high-to-medium speed includes:
[0110] The ratio of corresponding elements of the abnormal energy characteristic matrix at the medium test speed to the abnormal energy characteristic matrix at the low test speed is calculated to obtain the abnormal energy growth rate matrix at medium to low speeds.
[0111] The ratio of corresponding elements in the abnormal energy characteristic matrix at high test speed to that at medium test speed is calculated to obtain the abnormal energy growth rate matrix at high-medium speed.
[0112] When calculating the ratio, a small constant is added to the denominator to prevent the denominator from being zero.
[0113] This invention takes the ratio of the accumulated abnormal energy values at different test speeds (i.e., element-by-element division of corresponding elements) to reflect the increase in speed. If the blood pump in an ECMO device experiences mechanical wear or imbalance, its abnormal vibration often increases non-linearly with increasing speed. Therefore, the abnormal energy growth rate matrix can more sensitively capture this "abnormal amplification effect caused by speed," making the fault characteristics more significant.
[0114] In this embodiment, obtaining the blood pump failure type in the ECMO device includes the following process:
[0115] A CNN network was used to process the entropy feature matrix under low test speed, medium test speed, and high test speed respectively, to obtain the first entropy feature tensor, the second entropy feature tensor, and the third entropy feature tensor.
[0116] The first entropy feature tensor, the second entropy feature tensor, and the third entropy feature tensor are added element-wise to obtain the entropy feature fusion tensor, such as... Figure 2 As shown;
[0117] A CNN network was used to process the abnormal energy growth rate matrix of medium-low speed and the abnormal energy growth rate matrix of high-medium speed, respectively, to obtain the first growth rate feature tensor and the second growth rate feature tensor.
[0118] Adding the first growth rate feature tensor and the second growth rate feature tensor element by element yields the fused growth rate feature tensor, such as... Figure 3 As shown;
[0119] Feature extraction is performed on the entropy feature fusion tensor and the growth rate feature fusion tensor at two different scales. The entropy feature and the growth rate feature at the same scale are then multiplied to obtain the first enhanced feature and the second enhanced feature.
[0120] The first and second enhanced features are concatenated, and a classifier is used to classify the concatenated features.
[0121] A first CNN network is used to process the entropy feature matrix at low test speeds to obtain a first entropy feature tensor; a second CNN network is used to process the entropy feature matrix at medium test speeds to obtain a second entropy feature tensor; and a third CNN network is used to process the entropy feature matrix at high test speeds to obtain a third entropy feature tensor, as follows. Figure 2 As shown.
[0122] A fourth CNN network is used to process the abnormal energy growth rate matrix for medium-low speed rotation to obtain the first growth rate feature tensor; a fifth CNN network is used to process the abnormal energy growth rate matrix for high-medium speed rotation to obtain the second growth rate feature tensor, such as... Figure 3 As shown.
[0123] This invention extracts entropy feature matrices at low, medium, and high test speeds, processes them through a convolutional neural network, and then fuses them. Simultaneously, it introduces abnormal energy growth rate matrices for medium-low and high-medium speeds to characterize the growth trend of fault energy with speed variation. Furthermore, it performs multi-scale feature extraction and product enhancement on the entropy feature fusion tensor and the growth rate feature fusion tensor, achieving information complementarity between different scales and features. This comprehensively reflects the dynamic changes of the ECMO blood pump under different speed conditions, effectively enhancing the robustness of features to speed fluctuations and operating condition changes, improving the discriminative power and stability of feature expression, and thus significantly improving the accuracy and reliability of blood pump fault type identification.
[0124] like Figure 4 As shown, the process of obtaining the first and second enhanced features includes:
[0125] The first-scale feature extraction module is used to extract first-scale entropy features from the entropy feature fusion tensor;
[0126] The second-scale feature extraction module is used to extract second-scale entropy features from the entropy feature fusion tensor;
[0127] The third-scale feature extraction module is used to extract the first-scale growth rate feature by fusing the growth rate feature tensor;
[0128] The fourth-scale feature extraction module is used to extract the second-scale growth rate feature by fusing the growth rate feature with tensors;
[0129] Multiply the first-scale entropy feature and the first-scale growth rate feature element by element to obtain the first enhanced feature;
[0130] The second enhanced feature is obtained by multiplying the second-scale entropy feature and the second-scale growth rate feature element by element.
[0131] This invention extracts entropy and growth rate features at different scales and enhances their fusion by element-wise multiplication, enabling deep interaction between the two types of features within the same scale space. Entropy features reflect the uncertainty and complexity of signal energy distribution, while growth rate features characterize the changing trend of abnormal energy under different speed conditions. Multiplying the two can simultaneously take into account both steady-state and dynamic change information of energy features, thereby effectively strengthening fault-related features and suppressing redundancy and noise information.
[0132] The first-scale feature extraction module and the third-scale feature extraction module have the same structure and parameters, such as Figure 5 As shown, it includes: a first average pooling layer, a first convolutional block, a second convolutional block, a third convolutional block, and a first concat layer;
[0133] The input of the first average pooling layer is connected to the input of the first convolutional block and serves as the input of the first scale feature extraction module and the third scale feature extraction module. The output of the first concat layer serves as the output of the first scale feature extraction module and the third scale feature extraction module. The kernel size of the first convolutional block, the second convolutional block and the third convolutional block is 1×1.
[0134] The second-scale feature extraction module and the fourth-scale feature extraction module have the same structure and parameters, such as Figure 6 As shown, it includes: a second average pooling layer, a fourth convolutional block, a fifth convolutional block, a sixth convolutional block, and a second concat layer; the input of the second average pooling layer is connected to the input of the fourth convolutional block and serves as the input of the second-scale feature extraction module and the fourth-scale feature extraction module, and the output of the second concat layer serves as the output of the second-scale feature extraction module and the fourth-scale feature extraction module; the kernel size of the fourth, fifth, and sixth convolutional blocks is 3×3.
[0135] In this embodiment, the classifier uses a fully connected layer, and the output fault types include: bearing wear fault, impeller imbalance fault, rotor eccentricity or misalignment fault, etc.
[0136] This invention correlates vibration signals with the fundamental frequency of rotational speed through angle domain transformation, extracts harmonic order energy by combining cyclic energy spectrum, and then fuses multi-dimensional features such as entropy characteristics, abnormal energy, and abnormal energy growth rate between rotational speeds to comprehensively capture the essential information of the fault. Combined with the multi-scale feature extraction and fusion capabilities of neural networks, it solves the problem of insufficient feature extraction for non-stationary signals and complex coupled faults in existing methods, significantly improving the accuracy of fault identification.
[0137] This invention converts vibration signals to the angular domain and calculates the cyclic energy spectrum and harmonic order energy in the angular domain. This eliminates the influence of rotational speed fluctuations on feature extraction, resulting in extracted features with higher periodic consistency and structural stability, and more accurately reflecting the dynamic behavior of the blood pump under different operating conditions. By introducing entropy features and abnormal energy into the harmonic order energy vector, this invention not only characterizes the energy distribution of the blood pump vibration signal but also enhances the sensitivity to abnormal disturbances and local instability, enabling the effective identification of early minor faults (such as bearing wear and early cavitation).
[0138] This invention focuses on the internal vibration response of the blood pump's mechanical structure. Angular domain signals and cyclic energy spectra are highly sensitive to early, minute anomalies such as mechanical wear and bearing imbalance. Compared to existing methods that rely on electrical signals and system parameters, this invention can capture characteristic changes in the early stages of a fault earlier, reducing the medical risks caused by equipment failure.
[0139] This invention covers three test speed scenarios and achieves angular domain standardization of vibration signals through fundamental frequency adaptation, eliminating the interference of speed fluctuations on feature extraction. The abnormal energy growth rate matrix further reflects the evolution law of faults under different speeds, highlighting fault characteristics and making them significant.
[0140] 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. A method for diagnosing faults in ECMO equipment, characterized in that, Includes the following steps: Vibration signals of the blood pump in the ECMO device during operation were collected using a vibration sensor at three test speeds. Based on the fundamental frequency of the corresponding test rotation speed, the corresponding vibration signal is converted into an angle domain signal, and the cyclic energy spectrum of each segment of the angle domain signal is obtained. In each cycle energy spectrum, the harmonic order energy under the same harmonic order is calculated, and the harmonic order energy vector is constructed. Entropy feature values are obtained at different positions in the harmonic order energy vector, and an entropy feature matrix is constructed at the corresponding test speed. Anomalies are obtained at different locations in the harmonic order energy vector, the abnormal energy is calculated, and the abnormal energy feature matrix at the corresponding test speed is constructed. Based on the ratios of medium test speed to low test speed and high test speed to medium test speed on the abnormal energy characteristic matrix, the abnormal energy growth rate matrices of medium-low speed and high-medium speed are obtained. A CNN network was used to extract and fuse features from three entropy feature matrices and two abnormal energy growth rate matrices. Multi-scale features were then extracted from the two fused features, and the features were enhanced and classified according to scale to obtain the blood pump failure types in ECMO devices.
2. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The three test speeds are: low test speed, medium test speed, and high test speed. At the low test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the low test speed. At the medium test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the medium test speed. At the high test speed, vibration signals from the blood pump in the ECMO device are collected using a vibration sensor to obtain the vibration signal at the high test speed. 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.
3. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process of obtaining the cyclic energy spectrum includes: Based on the test rotation speed, obtain the fundamental frequency corresponding to the test rotation speed; The instantaneous angle is obtained based on the fundamental frequency. The vibration signal is resampled according to the instantaneous angle to obtain the angle domain signal; The angle domain signal is divided into multiple segments, and a Fourier transform is performed on each segment to obtain the cyclic energy spectrum of each segment of the angle domain signal.
4. The ECMO equipment fault diagnosis method according to claim 3, characterized in that, The formula for the cyclic energy spectrum of each segment of the angle domain signal is: ， Where E(α,ω) is the cyclic energy at (α,ω) in the cyclic energy spectrum of each segment of the angle domain signal. Let be the spectral value of each segment of the angle-domain signal at ω+α / 2 after Fourier transform. for The complex conjugate, Let be the spectral value of each segment of the angle domain signal at ω-α / 2 after Fourier transform, || be the modulus of the complex number, ω be the angular frequency, and α be the harmonic order. The value of α is a positive integer from 1 to 4.
5. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process of constructing harmonic order energy vectors includes: In each cycle energy spectrum, the harmonic order energy under the same harmonic order is calculated, where the harmonic order energy is the result of adding the cycle energies belonging to the same harmonic order; The harmonic order energies of the same harmonic order belonging to the same vibration signal are arranged in chronological order to obtain the harmonic order energy vector.
6. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process of constructing the entropy feature matrix includes: Add up all the harmonic order energies in the harmonic order energy vector to get the total order energy; Divide the energy of each harmonic order in the harmonic order energy vector by the total order energy to obtain the normalized harmonic order energy vector. Set a sliding window to slide on the normalized harmonic order energy vector and calculate the Shannon entropy under each sliding window. Normalize the Shannon entropy to the range of 0 to 1 to obtain the entropy eigenvalues; Using the entropy eigenvalues corresponding to the same harmonic order as elements of each row, an entropy feature matrix is constructed for the corresponding test speed.
7. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process of constructing the anomaly energy feature matrix includes: The mean value of the vector is obtained by taking the average value of each harmonic order energy in each harmonic order energy vector. Anomalies are identified by marking the harmonic order energies in the harmonic order energy vector that are greater than the mean of the vector. Set a sliding window to slide on the harmonic order energy vector and extract outliers under each sliding window; The harmonic order energies of the anomaly points under the sliding window are added together to obtain the cumulative value of the anomaly energy; The cumulative abnormal energy value corresponding to the same harmonic order is used as the element of each row to form the abnormal energy feature matrix at the corresponding test speed.
8. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process of obtaining the abnormal energy growth rate matrices for low-to-medium speed and high-to-medium speed includes: The ratio of corresponding elements of the abnormal energy characteristic matrix at the medium test speed to the abnormal energy characteristic matrix at the low test speed is calculated to obtain the abnormal energy growth rate matrix at medium to low speeds. The ratio of corresponding elements in the abnormal energy characteristic matrix at high test speed to that at medium test speed is calculated to obtain the abnormal energy growth rate matrix at high-medium speed.
9. The ECMO equipment fault diagnosis method according to claim 1, characterized in that, The process for determining the type of blood pump failure in an ECMO device includes the following steps: A CNN network is used to process the entropy feature matrices at low test speeds, medium test speeds, and high test speeds, respectively, to obtain a first entropy feature tensor, a second entropy feature tensor, and a third entropy feature tensor. The first entropy feature tensor, the second entropy feature tensor, and the third entropy feature tensor are then added element by element to obtain an entropy feature fusion tensor. A CNN network is used to process the abnormal energy growth rate matrix of medium-low speed and the abnormal energy growth rate matrix of high-medium speed, respectively, to obtain the first growth rate feature tensor and the second growth rate feature tensor; the first growth rate feature tensor and the second growth rate feature tensor are added element by element to obtain the growth rate feature fusion tensor. Feature extraction is performed on the entropy feature fusion tensor and the growth rate feature fusion tensor at two different scales. The entropy feature and the growth rate feature at the same scale are then multiplied to obtain the first enhanced feature and the second enhanced feature. The first and second enhanced features are concatenated, and a classifier is used to classify the concatenated features.
10. The ECMO equipment fault diagnosis method according to claim 9, characterized in that, The process of obtaining the first and second enhancement features includes: The first-scale feature extraction module is used to extract the first-scale entropy feature by fusing tensors with the entropy feature; the second-scale feature extraction module is used to extract the second-scale entropy feature by fusing tensors with the entropy feature; the third-scale feature extraction module is used to extract the first-scale growth rate feature by fusing tensors with the growth rate feature; the fourth-scale feature extraction module is used to extract the second-scale growth rate feature by fusing tensors with the growth rate feature; the first-scale entropy feature and the first-scale growth rate feature are multiplied together to obtain the first enhanced feature; the second-scale entropy feature and the second-scale growth rate feature are multiplied together to obtain the second enhanced feature.
Citation Information
Patent Citations
ECMO equipment operation state monitoring method
CN121207602A