A construction method for three-dimensional pulse signals integrating time and space domains

Abstract

The present invention belongs to the technical field of biomedical signal processing, and in particular relates to a method for constructing a three-dimensional pulse signal by spatio-temporal domain fusion, including: S1: collecting pulse signals at the cun, guan, and chi positions using an intelligent bionic pulse diagnosis hand equipped with an array pressure sensor; S2: detecting the wave valleys of the pulse signals using an improved adaptive wave valley detection algorithm; S3: performing period segmentation and normalization processing according to the marked wave valleys; S4: using an improved Lagrange interpolation method, adding weight assignment, and adaptively interpolating and fitting the pressure pulse signals perpendicular to the pulse width direction to form two-dimensional pulse signals; S5: performing color mapping on the two-dimensional pulse signals, and synthesizing three two-dimensional pulse signals of the three positions after mapping to form a three-dimensional pulse signal. The present invention improves the accuracy through an improved adaptive wave valley detection algorithm, and constructs a three-dimensional pulse signal through a method of weight assignment, so that the three-dimensional pulse signal fuses multiple information in the time domain and the spatial domain.

Description

Technical Field

[0001] The present invention relates to the field of biomedical signal processing, and specifically to a method for constructing a three-dimensional pulse signal by fusing the time domain and the space domain. Background Technique

[0002] The pulse signal is one of the important biomedical signals and is widely used in health monitoring, disease diagnosis, and vital sign detection. Since the pulse signal can reflect the state of the cardiovascular system, its research and application in the medical field are of great significance. At the same time, in traditional Chinese medicine, the pulse is also regarded as an important basis for diagnosing human health. The combined diagnosis of different positions (inch, cusp, and cubit) of the pulse has a long application history in traditional Chinese medicine and can provide more detailed information about the physical condition. However, there are few studies on the combined analysis of the time domain and the space domain, and there are certain deficiencies.

[0003] Traditional methods for detecting the wave troughs of pulse signals mainly rely on threshold methods and first- and second-order derivative methods. These methods usually require manual setting of thresholds and lack the ability to adaptively detect wave troughs. When dealing with signals with multiple wave troughs or wave troughs with similar heights, it is easy to have misidentifications or missed detections. Especially for patients with cardiovascular diseases, their pulse signals will show irregular fluctuations, resulting in the appearance of abnormal peaks and troughs. In this case, it is very difficult for traditional wave trough detection methods to accurately identify all wave troughs, which will in turn affect subsequent cycle segmentation and other analysis processes.

[0004] For disease identification that requires combining pulse signals at three different positions of inch, cusp, and cubit, traditional methods need to input multiple pulse signal images. This not only increases the consumption of computing resources, improves the complexity and computational amount of training, but also may lead to an inhibitory effect between pulse features. The learning efficiency and training speed of the final model are greatly reduced, affecting the accuracy and real-time performance of disease identification. Therefore, we propose an adaptive wave trough detection algorithm and a method for constructing a three-dimensional pulse signal by fusing the time domain and the space domain to solve the above problems. Summary of the Invention

[0005] (I) Technical Problems to be Solved

[0006] Aiming at the problems that the accuracy of existing wave trough detection methods is relatively low and it is difficult to meet the requirements of pulse signal processing, and the current research mainly focuses on one-dimensional or two-dimensional pulse signals and lacks the exploration of a method for constructing a three-dimensional pulse signal by fusing the time domain and the space domain, the present invention provides a method for constructing a three-dimensional pulse signal by fusing the time domain and the space domain. The proposed adaptive wave trough detection algorithm and the method for constructing a three-dimensional pulse signal solve the problems proposed in the above background technique.

[0007] (II) Technical Solutions

[0008] The present invention specifically adopts the following technical solutions to achieve the above object:

[0009] A method for constructing a three-dimensional pulse signal by spatio-temporal domain fusion, comprising the following steps:

[0010] S1: Use an intelligent bionic pulse diagnosis hand equipped with 3 array pressure sensors to collect the pressure pulse signals and static pressures at the cun position, guan position, and chi position;

[0011] S2: Use an improved adaptive wave valley detection algorithm to detect the wave valleys of the pressure pulse signals of the pressure pulse signals;

[0012] S3: Perform cycle segmentation and normalization processing on the pulse signals according to the marked wave valleys of the pressure pulse signals;

[0013] S4: Use an improved Lagrange interpolation method, add weight distribution, and perform adaptive interpolation fitting on the pressure pulse signals after the normalization processing perpendicular to the pulse width direction to form a two-dimensional pulse signal;

[0014] S5: Perform color mapping on the two-dimensional pulse signals, add weights to the three two-dimensional pulse signals of the three positions after mapping, and perform synthesis of the three-dimensional pulse signals;

[0015] Further, the specific steps of S1 are as follows: The 3 1×4 array pressure sensors synchronously collect the pressure pulse signals and static pressures at the cun position, guan position, and chi position. Specifically, the array pressure sensor is composed of 4 independent sensing units, each unit has a size of 1.2 mm×3.25 mm, the adjacent unit spacing is 1 mm, the 3 array pressure sensors are installed at the fingertips of the intelligent bionic pulse diagnosis hand, and the three array pressure sensors are connected to a signal processing module and a collection module; the signal processing module is composed of an amplification circuit, a zero adjustment circuit, a power frequency limiting wave circuit, and a band-pass filter circuit, removes baseband drift, power frequency interference, and amplifies the pulse signal, and sends the processed pulse signal to the collection module. The collection module uses two 16-bit high-precision AD7616 conversion chips for data collection to obtain the data of 12 sensing units. The control chip selects FPGA, which is responsible for collecting the digital signals after AD conversion to ensure the efficient processing and accurate storage of data.

[0016] Further, the specific steps of S2 are as follows: Use the empirical mode decomposition algorithm (EMD) to decompose the original pulse signal to obtain N IMF components:

[0017]

[0018] Among them, is the One IMF component is the residual term.

[0019] The correlation between each IMF component and the original pulse signal is important. By calculating the Pearson correlation coefficient between each IMF component and the original pulse signal among them , the characteristic components (IMF) with strong correlation with the original pulse signal are selected. The calculation formula of the Pearson correlation coefficient is as follows :

[0020]

[0021] where and are the means of the IMF component and the original pulse signal respectively.

[0022] Furthermore, in order to dynamically screen the IMF components and select the IMF component most relevant to the original pulse signal, a dynamic threshold based on the signal-to-noise level is defined :

[0023]

[0024] where is the standard deviation of the original pulse signal is a tuning parameter, and its initial value is set to 0.56.

[0025] Furthermore, in order to more precisely weight each IMF component, the similarity between the IMF components with Pearson correlation coefficient greater than the threshold and the original pulse signal is measured by dynamic time warping (DTW) :

[0026]

[0027] where is the offset of time alignment.

[0028] Furthermore, according to the DTW similarity, weights are assigned to each of the IMF components

[0029]

[0030] where The initial value is set to 0.23.

[0031] Subsequently, the weighted IMF components are recombined into a new pulse signal ;

[0032]

[0033] Among them, M is the number of screened IMF components.

[0034] Furthermore, perform continuous wavelet transform (CWT) on the reconstructed pulse signal to obtain a time-frequency diagram , and the amplitude spectrum of the wavelet transform shows the local characteristics of the signal at different frequencies and times, and can find the position of the wave trough in the time-frequency domain. By detecting the local minimum of the CWT amplitude spectrum, the position of the wave trough is extracted:

[0035]

[0036] Furthermore, perform minimum value detection on the reconstructed pulse signal as well:

[0037]

[0038] Therefore, the positions of the wave troughs and are extracted.

[0039] Furthermore, weight distribution is performed for the wave trough points obtained by the EMD and CWT methods. The weight distribution is based on the standard deviation of the original pulse signal and the frequency spectrum , the standard deviation of the reconstructed pulse signal and the periodicity evaluation score of each IMF component ; integrating the standard deviation (noise sensitivity) of the signal, the periodicity evaluation (the periodic consistency between the IMF component and the original signal), and the frequency characteristics (the low-frequency characteristics of the signal), the required EMD and CWT weight equations are obtained:

[0040]

[0041]

[0042] In the formula, is set to 0.3, 0.2, 0.5 for the initial values; among them, determines the weight distribution of the EMD and CWT fusion for the noise level. The larger it is, the greater the influence of the signal noise, and the weight of CWT will increase accordingly; if the signal is relatively stable, is smaller, and the weight of EMD will increase; Control the influence of periodic evaluation on the weights. When the periodicity of the signal is strong, EMD should have a greater weight. Therefore, adjust the contribution degrees of EMD and CWT; Control the influence of the dominant signal frequency on the weights to balance the low-frequency and high-frequency characteristics of the signal. If the low-frequency part is dominant (strong periodicity), the weight of EMD is large; if the high-frequency part of the signal is dominant (large detail changes), the weight of CWT is large.

[0043] The calculation formula of

[0044]

[0045] is as follows: where k is the time-delay parameter and T is the length of the signal.

[0046] The standard deviation of the original pulse signal and the frequency spectrum , and the standard deviation of the reconstructed pulse signal are calculated as follows:

[0047]

[0048]

[0049]

[0050] Furthermore, according to the trough positions obtained by EMD and the trough positions and their weights obtained by CWT, perform adaptive weighted fusion to obtain the trough point:

[0051]

[0052] Through backward search for the trough of the original pulse signal.

[0053] Furthermore, the specific steps of S3 are as follows: According to divide the original pulse signal into multiple periodic signals , and use min-max normalization to adjust its amplitude. The normalized signal is calculated by the following equation:

[0054]

[0055] Further, the specific steps of S4 are as follows: using the improved Lagrange interpolation method and adding weight assignment to enable adaptive interpolation fitting. The equation is as follows:

[0056]

[0057] In the formula, is the smoothed curve obtained by adaptive interpolation fitting, is the weight of each sensor i at time t, corresponds to the pressure signal collected by sensor i at time t;

[0058] The weight is affected by the signal change rate . If the signal change between every two sensor units is fast, then this point is crucial for the fitting process and is automatically adjusted to be assigned a higher weight. Therefore, the weight can be obtained by the following formula:

[0059]

[0060] Where is the time interval between two adjacent sampling points;

[0061] Further, the specific steps of S5 are as follows: map the two-dimensional pulse diagram after fitting at the Cun position to the red channel (R), map the pulse diagram after fitting at the Guan position to the green channel (G), and map the Chi position to the blue channel (B). The mapping range is all , normalize the static pressure values (at the Cun position), (at the Guan position) and (at the Chi position) in step S1, so that the range is within . Use the normalized pressure values , , as the weights of the RGB color channels. Multiply the mapped red channel value (R) by the weight value of the Cun position, that is , multiply the mapped green channel value (G) by the weight value of the Guan position, that is , multiply the mapped blue channel value (B) by the weight value of the Chi position, that is ; by synthesizing the adjusted color channels, a three-dimensional RGB pulse image containing time-domain and space-domain information is obtained.

[0062] (III) Beneficial Effects

[0063] Compared with the prior art, the present invention provides a method for constructing a three-dimensional pulse signal with spatio-temporal domain fusion, having the following beneficial effects:

[0064] Through the improved adaptive wave trough detection algorithm, the present invention improves the accuracy of wave trough detection, significantly enhances the wave trough detection accuracy of irregular pulse signals. At the same time, in combination with the time-frequency relationship, the weight distribution of EMD and CWT is dynamically adjusted, thereby further ensuring the accuracy of the wave trough position.

[0065] By fusing the pulse signals in the time domain and the spatial domain, compared with the traditional method that needs to combine the pulse signals at three different positions of cun, guan, and chi, this method greatly reduces the size of the data set. We also assign weights to the red, green, and blue channels, and dynamically adjust the weights of each color channel according to the pressing force of the array pressure sensor. In this way, a three-dimensional pulse signal integrating the three positions of cun, guan, and chi is constructed. This not only reduces the computational amount and complexity of the later model, but also reduces the inhibitory effect between data, improves the training speed and efficiency of the model, and thus optimizes the construction of the later disease classification and prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic flow chart of the method for constructing a three-dimensional pulse signal with spatio-temporal domain fusion according to an embodiment of the present invention;

[0067] Figure 2 It is a schematic diagram of the working principle and architecture according to an embodiment of the present invention;

[0068] Figure 3 It is a performance test data graph of the 1×4 array pressure sensor according to an embodiment of the present invention;

[0069] Figure 4 It is a flow chart of the improved adaptive wave trough detection algorithm according to an embodiment of the present invention;

[0070] Figure 5 It is a result graph of the wave trough positions extracted by EMD and CWT according to an embodiment of the present invention;

[0071] Figure 6 It is an effect graph of using the improved adaptive wave trough detection algorithm in the embodiment of the present invention to detect multi-wave troughs or pulse signals with the same wave trough amplitude;

[0072] Figure 7 It is a result graph of inversely searching for the wave trough positions of the original pulse signal according to the wave trough positions after weighted fusion according to an embodiment of the present invention;

[0073] Figure 8 It is a result graph of the period segmentation and normalization processing of the pulse signal according to an embodiment of the present invention;

[0074] Figure 9 The 3D pulse image constructed by using the 3D pulse signal construction method in the embodiments of the present invention. The figure shows the 3D pulse diagrams of two different volunteers.

[0075] Figure 10 The color mapping result diagram of the 2D pulse signal in the embodiments of the present invention.

[0076] Figure 11 The accuracy rate of identifying heart failure patients for the dataset constructed by the 3D pulse diagram and the traditional 2D pulse signal diagram constructed in the embodiments of the present invention. Specific embodiments

[0077] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0078] Embodiment

[0079] As Figures 1-11 shown, a method for constructing a 3D pulse signal by spatio-temporal domain fusion, the specific steps are as follows:

[0080] S1: Use an intelligent bionic pulse diagnosis hand equipped with 3 array pressure sensors to collect the pressure pulse signals and static pressures at the cun position, guan position, and chi position.

[0081] In this embodiment, 3 1×4 array pressure sensors are respectively installed at the fingertips of the intelligent bionic pulse diagnosis hand to synchronously collect the pressure pulse signals and static pressures at the cun position, guan position, and chi position. The sensor uses flexible polyimide (PI) as the base material and combines a polydimethylsiloxane nanocomposite with a high signal-to-noise ratio (SNR = 690:1), which not only ensures excellent mechanical flexibility but also has good biocompatibility and a comfortable wearing experience. The sensor consists of 4 independent sensing units, each unit has a size of 1.2 mm × 3.25 mm, and the adjacent unit spacing is 1 mm. The sensitivity of the sensor , and the goodness-of-fit coefficient reaches . Under the same pressure, the output signal curves of different sensing units are consistent, and the difference between the maximum and minimum output values is only 0.0094. In addition, the output errors of 12 sensing units are all controlled within 0.4%, as Figure 3 shown.

[0082] To achieve high-precision signal acquisition, three array pressure sensors are connected to a self-made signal processing module and an acquisition module. The signal processing module consists of an amplification circuit, a zero-adjustment circuit, a power frequency notch filter circuit, and a band-pass filter circuit. It removes baseband drift, power frequency interference, and amplifies the pulse signal, and sends the processed pulse signal to the acquisition module. The acquisition module uses two 16-bit high-precision AD7616 conversion chips for data acquisition to obtain data of 12 sensing units. The control chip selects FPGA, which is responsible for acquiring the digital signal after AD conversion to ensure efficient processing and accurate storage of data. During the pulse signal acquisition process, multiple sensing units work together to effectively improve the measurement accuracy and data stability.

[0083] S2: Detect the troughs of the original pressure pulse signal using an improved adaptive trough detection algorithm of the original pressure pulse signal;

[0084] A1: As Figure 4 shown, first use the empirical mode decomposition algorithm (EMD) to decompose the original pulse signal to obtain N IMF components:

[0085]

[0086] Among them, is the th IMF component, is the residue term.

[0087] The correlation between each IMF component and the original pulse signal is very important. By calculating the Pearson correlation coefficient between each IMF component and the original pulse signal , select the characteristic components (IMF) with strong correlation with the original pulse signal . The calculation formula of the Pearson correlation coefficient is:

[0088]

[0089] Among them, and are the means of the IMF component and the original pulse signal respectively.

[0090] In this embodiment, in order to dynamically screen IMF components, select the IMF component most relevant to the original pulse signal , and define a dynamic threshold based on the signal noise level:

[0091]

[0092] Among them, is the standard deviation of the original pulse signal, is a regulation parameter, and its initial value is set to 0.56.

[0093] A2: To more precisely weight each IMF component, the Pearson correlation coefficient greater than the threshold is measured by dynamic time warping (DTW) for the IMF component and the original pulse signal similarity :

[0094]

[0095] wherein, is the offset of time alignment. According to the DTW similarity, weights are assigned to each of the IMF components:

[0096]

[0097] wherein, the initial value is set to 0.23.

[0098] Subsequently, the weighted IMF components are recombined into a new pulse signal :

[0099]

[0100] wherein, M is the number of IMF components after screening.

[0101] A3: The recombined pulse signal is subjected to continuous wavelet transform (CWT) to obtain a time-frequency diagram , and the amplitude spectrum of the wavelet transform shows the local characteristics of the signal at different frequencies and times, and can find the position of the wave valley in the time-frequency domain. By performing local minimum detection on the CWT amplitude spectrum, the position of the wave valley is extracted:

[0102]

[0103] Local minimum detection is also performed on the recombined pulse signal :

[0104]

[0105] Therefore, the position of the wave valley is extracted and . In this embodiment, local minimum detection is performed on the recombined pulse signal and the time-frequency diagram respectively to extract the position of the wave valley, and the results are as Figure 5 shown.

[0106] A4: Weight assignment is performed for the trough points obtained by the EMD and CWT methods. The weight assignment is based on the standard deviation and frequency spectrum of the original pulse signal , the reconstructed pulse signal 's standard deviation and the periodicity evaluation scores of each IMF component ; By synthesizing the standard deviation (noise sensitivity) of the signal, periodicity evaluation (periodic consistency between the IMF component and the original signal), and frequency characteristics (low-frequency characteristics of the signal), the required EMD and CWT weight equations are obtained:

[0107]

[0108]

[0109] In the formula, 's initial values are set to 0.3, 0.2, 0.5, where determines the weight assignment of the noise level to the EMD and CWT fusion. The larger is, the greater the influence of signal noise, and the weight of CWT will increase accordingly; if the signal is relatively stable, is smaller, and the weight of EMD will increase; controls the influence of periodicity evaluation on the weight. When the periodicity of the signal is strong, EMD should have a greater weight. Therefore, adjust 's contribution to EMD and CWT; dominates the influence of the signal frequency on the weight, and is used to balance the low-frequency and high-frequency characteristics of the signal. If the low-frequency part dominates (strong periodicity), the weight of EMD is larger. If the high-frequency part of the signal dominates (large detail changes), the weight of CWT is larger.

[0110] The calculation formula of

[0111]

[0112] is as follows: where k is the time delay parameter and T is the length of the signal.

[0113] The standard deviation and frequency spectrum of the original pulse signal , the reconstructed pulse signal 's standard deviation The calculation formula is as follows:

[0114]

[0115]

[0116]

[0117] The trough positions obtained according to EMD and the trough positions obtained by CWT , and their weights and , perform adaptive weighted fusion to obtain the trough point:

[0118]

[0119] By backward searching for the trough of the original pulse signal , the searched trough positions are as Figure 6 shown.

[0120] In addition, using the improved adaptive trough detection algorithm in this embodiment also has a good application effect on pulse signals with multiple troughs or the same trough amplitude, as Figure 7 shown.

[0121] S3: Perform cycle segmentation and normalization processing on the pulse signal according to the marked trough of the pressure pulse signal , and the result graph is as Figure 8 shown.

[0122] A1: Through the trough points of the original pulse signal , the pulse signal is segmented into multiple cycles. Each two consecutive trough points represent one cycle, that is, each cycle is the interval defined by two consecutive trough points and . The pulse cycle signal is:

[0123]

[0124] Therefore, the original pulse signal is divided into multiple cycles .

[0125] A2: The amplitudes of the pulse signals are different between different cycles, which will affect the subsequent interpolation fitting and the construction of the three-dimensional pulse signal. It is necessary to perform normalization processing on the signals of each cycle to adjust the signals of all cycles to the same size. The minimum-maximum normalization is selected to compress the amplitudes of the pulse signals to the specified range. The signal after minimum-maximum normalization can be calculated by the following formula:

[0126]

[0127] S4: Use the improved Lagrange interpolation method, add weight assignment, and perform adaptive interpolation fitting on the pressure pulse signal after the normalization process perpendicular to the pulse width direction to form a two-dimensional pulse signal.

[0128] For the normalized pulse signal, use the improved Lagrange interpolation method, add weight assignment, so that it can perform adaptive interpolation fitting. The equation is as follows:

[0129]

[0130] In the formula, is the smooth curve obtained by adaptive interpolation fitting, is the weight of each sensing unit i at time t, corresponds to the pressure signal collected by the sensing unit i at time t.

[0131] Weight is based on the spatial weight and the signal change rate weight. In this embodiment, the distance between every two sensing units is 1 mm. Therefore, the spatial weight is a constant and is not considered. Assume that the distance between the sensing unit and the target point is , then the spatial weight is:

[0132]

[0133] If the signals of every two sensor units change rapidly, then this point is crucial for the fitting process and is automatically adjusted to give a higher weight. The signal change rate can be obtained by calculating the difference of this point in time:

[0134]

[0135] where is the time interval between two adjacent sampling points;

[0136] Taking these two factors into comprehensive consideration, the calculation equation of the final weight can be written as:

[0137]

[0138] S5: Perform color mapping on the two-dimensional pulse signal, add weights to the three two-dimensional pulse signals of the three mapped parts respectively, and perform the synthesis of the three-dimensional pulse signal;

[0139] To make this embodiment more persuasive, the pulse signals of two volunteers were collected for comparison after the construction of the three-dimensional pulse signal, as Figure 9 shown.

[0140] A1: Map the three two-dimensional pulse signals obtained from the pulse signal after S4 fitting to the three channels of the RGB color space, as Figure 10 shown:

[0141] Map the fitted pulse signal at the Cun position to the red channel (R).

[0142] Map the fitted pulse signal at the Guan position to the green channel (G).

[0143] Map the fitted pulse signal at the Chi position to the blue channel (B).

[0144] The color range of each channel is (the pixel range of the standard RGB image), so that the time variation of each pulse signal will be reflected on the corresponding channel through the color intensity.

[0145] A2: Dynamically adjust the weights of each color channel according to the static pressure values (at the Cun position), (at the Guan position) and (at the Chi position). Normalize the static pressure so that the range is within , which helps to map them to the effective range of the RGB color space during the processing:

[0146]

[0147] The normalized pressure values are , , , and use them as the weights of the RGB color channels. Multiply the mapped red channel value (R) by the weight value at the Cun position, that is , multiply the mapped green channel value (G) by the weight value at the Guan position, that is , multiply the mapped blue channel value (B) by the weight value at the Chi position, that is .

[0148] A3: Finally, by synthesizing the adjusted color channels, a three-dimensional RGB pulse image containing time-domain and spatial-domain information is obtained, that is:

[0149]

[0150] Using the method for constructing three-dimensional pulse signals of this embodiment, a dataset for heart failure identification was constructed. For comparison, a corresponding dataset was also constructed based on traditional two-dimensional pulse signals. After classifying and training the two groups of data using a convolutional neural network, it was found that the dataset based on three-dimensional pulse signals had a higher accuracy in identifying heart failure than two-dimensional signals, as Figure 11 shown.

[0151] The three-dimensional pulse signals constructed in this embodiment need to be further clarified. Currently, many three-dimensional pulse signals proposed by researchers are actually considered two-dimensional pulse signals within the framework of this patent. Specifically, for the three-dimensional pulse signals defined by other researchers, their three-dimensional space is constructed in the following way: the x-axis represents time, the y-axis represents the direction perpendicular to the wrist, which is usually used to describe the width of the pulse, and the z-axis represents the intensity of the pulse. However, in this patent, the definition of the three-dimensional pulse signal is different. The three-dimensional pulse signals we defined still use time (x-axis) and the direction perpendicular to the wrist (y-axis) to describe the width of the pulse, and the z-axis still represents the intensity of the pulse. On this basis, different from other studies, this patent fuses the pulse signals at three different positions of cun, guan, and chi, and represents the pulse signals at each position with different colors in the RGB color gamut. Compared with the traditional representation of three-dimensional pulse signals, this method adds an additional dimension to the pulse signals. In other words, if the pulse signals defined by other researchers are three-dimensional pulse signals, then according to the definition of this patent, the constructed pulse signals should be regarded as four-dimensional pulse signals, and can be understood further according to Figure 2 this concept.

[0152] Aiming at the problems of low accuracy of existing trough detection methods and difficulty in meeting the requirements of pulse signal processing, and the current research mainly focuses on one-dimensional or two-dimensional pulse signals, lacking the exploration of the construction method of three-dimensional pulse signals that fuses the time domain and the spatial domain, the present invention provides a construction method of three-dimensional pulse signals that fuses the time and space domains, and the proposed adaptive trough detection algorithm and three-dimensional pulse signal construction method. And the present invention also assigns weights to the red, green, and blue channels, and dynamically adjusts the weights of each color channel according to the pressing force of the array pressure sensor, thereby constructing a three-dimensional pulse signal that fuses the three different parts of cun, guan, and chi. This not only reduces the computational amount and complexity of the later model, but also reduces the inhibitory effect between data, improves the training speed and efficiency of the model, and thus optimizes the construction of the later disease classification and prediction model. Therefore, the method proposed in this paper can be used as an effective and feasible way to construct three-dimensional pulse signals.

[0153] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for constructing a three-dimensional pulse signal fused in time and space domains, characterized in that: The specific steps are as follows: S1: Use the intelligent bionic pulse diagnosis hand equipped with three array pressure sensors to collect pressure pulse signals and static pressure at the Cun, Guan and Chi positions; S2: Detect the trough of the pressure pulse signal S(t) using an improved adaptive trough detection algorithm; The specific step of S2 is to use the empirical mode decomposition algorithm EMD to decompose the original pulse signal S(t) and weightedly reorganize its IMF components with strong correlation to obtain the recombined pulse signal S new (t); for the recombined pulse signal S new (t) Perform continuous wavelet transform CWT to obtain the time-frequency diagram The reconstructed pulse signal S is extracted using a local minimum detection algorithm new The trough position of (t) EMD and the time-frequency diagram The trough position t CWT ; The trough position t obtained according to the EMD EMD and the trough position t obtained by the CWT CWT , and their weights ω EMD and ω CWT , perform adaptive weighted fusion, and obtain the trough t after weighted fusion valley ; Finally, through t valley ={t1,t2,…,t m } Reversely search for the trough of the original pulse signal S(t); The specific performance of the weighted reorganization is to measure the similarity D between the characteristic component with strong correlation and the original pulse signal S(t) by dynamic time warping DTW. DTW (IMF i , S(t)), assign weights to each of the IMF components according to the DTW similarity: Then the weighted IMF components are recombined into a new pulse signal S new (t); The specific performance of the adaptive weighted fusion of the EMD and CWT trough positions is to calculate the standard deviation σ of the original pulse signal S(t) S and frequency spectrum F S (f) Reconstructed pulse signal S new The standard deviation of (t) EMD And the cyclical assessment score of each IMF component i , and obtain the required EMD and CWT weight equations: oh CWT =1-h EMD In the formula, the initial values ​​of α1, α2, and α3 are set to 0.3, 0.2, and 0.5; The trough position t obtained according to the EMD EMD and the trough position t obtained by the CWT CWT , and their weights ω EMD and ω CWT , perform adaptive weighted fusion to obtain the valley point: t valley =ω EMD ·t EMD +ω CWT ·t CWT ; S3: performing period segmentation and normalization processing of the pulse signal according to the marked troughs of the pressure pulse signal; S4: using an improved Lagrange interpolation method and adding weight distribution, the normalized pressure pulse signal perpendicular to the pulse width direction is adaptively interpolated and fitted to form a two-dimensional pulse signal; S5: performing color mapping on the two-dimensional pulse signal, and respectively adding weights to the three two-dimensional pulse signals of the three parts after the mapping to synthesize the three-dimensional pulse signal; The specific steps of S5 are: mapping the two-dimensional pulse map after fitting at the Cun position to the red channel R, mapping the pulse map after fitting at the Guan position to the green channel G, and mapping the Chi position to the blue channel B, wherein the mapping range is [0, 255]; normalizing the static pressure value F1 at the Cun position, the static pressure value F2 at the Guan position, and the static pressure value F3 at the Chi position in step S1 so that the range is within [0, 1]; and normalizing the normalized pressure value F' normalized,1 , F' normalized,2 , F' normalized,3 As the weights ω of the RGB color channels R =F' normalized,1 ,ω G =F' normalized,2 ,ω B =F' normalized,3 , the mapped red channel value R and the weight value ω of the inch position R Multiply, that is, R'=ω R ×R, the green channel value G after mapping and the weight value of the relevant part ω G Multiply, that is, G'=ω G ×G, the mapped blue channel value B and the weight value ω of the scale part B Multiply, that is, B'=ω B ×B; by synthesizing the adjusted color channels, a three-dimensional RGB pulse image containing time domain and space domain information is obtained.

2. The method for constructing a three-dimensional pulse signal fused in time and space domains according to claim 1, characterized in that: The specific steps of S1 are: three 1×4 array pressure sensors synchronously collect pressure pulse signals and static pressure at Cun, Guan and Chi positions, which are specifically manifested as follows: the array pressure sensor is composed of four independent sensing units, each unit has a size of 1.2mm×3.25mm, and the spacing between adjacent units is 1mm. The three array pressure sensors are installed at the fingertips of the intelligent bionic pulse diagnosis hand, and the three array pressure sensors are connected to a signal processing module and an acquisition module; the signal processing module is composed of an amplifier circuit, a zero adjustment circuit, an industrial frequency limiting circuit and a bandpass filter circuit, which removes baseband drift, industrial frequency interference and amplification of pulse signals, and sends the processed pulse signal to the acquisition module. The acquisition module uses two 16-bit high-precision AD7616 conversion chips for data acquisition to obtain data from 12 sensing units. The control chip uses FPGA, which is responsible for collecting digital signals after AD conversion to ensure efficient processing and accurate storage of data.

3. The method for constructing a three-dimensional pulse signal fused in time and space domains according to claim 1, characterized in that: The specific steps of S3 are: valley ={t1,t2,…,t m } Divide the original pulse signal S(t) into multiple periodic signals P i , use minimum-maximum normalization to adjust its amplitude, the normalized signal S norm,i (t) is calculated by the following equation:

4. The method for constructing a three-dimensional pulse signal fused in time and space domains according to claim 1, characterized in that: The specific steps of S4 are to use the improved Lagrange interpolation method and add weight distribution to enable adaptive interpolation fitting. The equation is as follows: In the formula, is the smooth curve obtained by adaptive interpolation fitting, ω i (t) is the weight of each sensor i at time t, S norm,i (t) corresponds to the pressure signal collected by sensor i at time t; The weight ω i (t) Signal change rate δS norm,i (t) If the signal of every two sensor units changes faster, then this point is crucial to the fitting process and is automatically adjusted to give a higher weight, so the weight ω i (t) is obtained by the following formula: Where Δt is the time interval between two adjacent sampling points.

Images