Flexible pulse data detection method and system based on three-cell neural network

Abstract

The invention discloses a flexible pulse data detection method and system based on a three-cell neural network. The method comprises the following steps that S1, a pulse data set is collected; s2, generating a triple sample according to the pulse data set; s3, based on the ternary sample, an optimal solution is solved for the triplet residual neural network, and an optimal pulse regression prediction model is obtained; and S4, performing pulse rate prediction by using the optimal pulse regression prediction model. According to the method, the complexity and the cost of artificial feature engineering are reduced, a mathematical model of the pulse and the physiological signals is established according to the extracted features through the predictive ability of the regression algorithm, and accurate estimation of the pulse is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of data detection, and in particular relates to a flexible pulse data detection method and system based on a three-cell neural network. Background Art

[0002] The human pulse contains a wealth of cardiovascular system parameters and has always been an important source of information for clinical health assessment. Therefore, accurately and effectively monitoring the human pulse and analyzing the physiological health parameters it contains is an effective way to prevent and treat cardiovascular diseases. With the development of sensor technology, electronic devices and artificial intelligence, wearable smart health monitoring devices have attracted much attention. In order to achieve autonomous health monitoring anytime and anywhere, more and more wearable devices have emerged to achieve surface pulse monitoring to assess the health of the human cardiovascular system.

[0003] A pulse sensor is a sensor used to detect pulse-related signals. Pulse refers to arterial pulsation. A pulse sensor is used to detect the pressure changes generated when the artery pulsates and convert them into electrical signals that can be observed and detected more intuitively. Pulse sensors have two output modes: analog output and digital output. According to the way of collecting signals, they can be mainly divided into three types: piezoelectric, piezoresistive and photoelectric. Among them, piezoelectric and piezoresistive types convert the pressure process of pulse beating into signal output through micro-pressure materials (piezoelectric sheets, bridges, etc.). Photoelectric pulse sensors convert the changes in the transmittance of blood vessels during the pulse beating process into signal output through reflection or radiation. Pulse sensors are mainly used in medical equipment and teaching equipment, teaching training and other fields, such as blood oxygen measurement, heart rate monitoring and Chinese medicine pulse diagnosis, etc.

[0004] However, most of the current pulse sensors use a single sensor unit. The data detected by this method is not accurate enough. Although multiple measurements can ensure the accuracy of the data, they are inefficient and costly. In addition, in most devices, when the skin temporarily leaves part of the contact surface of the sensor due to movement of the body during daily activities, the skin will be separated from the sensor due to the pressure depression part, causing mechanical wave transmission loss, which reduces the accuracy of detection. Generally speaking, the intensity of the sensor signal received by the device is weak, which is not conducive to direct analysis. At the same time, the device will also generate various random noises during operation, and light noise will also be generated during the propagation of light in the optical fiber. Therefore, the weak signal itself is easily submerged by these noise factors, which is not conducive to data processing. When predicting pulse, artificial feature engineering is more complex, more expensive, and not accurate enough.

[0005] Deep learning is a machine learning technology based on neural networks. In 2006, deep learning began to emerge. The biggest advantage of deep learning lies in its versatility. It can automatically extract the features of input data, thus avoiding the tedious process of manually designing features. In addition, deep learning is also interpretable, and the decision-making process within the model can be explained in a visual way, which improves the credibility of the model. The deep twin network consists of two sub-networks with the same structure and shared parameters. It can process two different inputs at the same time and calculate the similarity or distance between the two. The triplet neural network (TNN) is an extension of the deep twin network.

[0006] At present, in order to overcome the above problems, it is very important to study a method that can accurately detect pulse without being restricted by physical activities. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a flexible pulse data detection method and system based on a three-cell neural network.

[0008] The technical solution of the present invention is: a flexible pulse data detection method based on a three-cell neural network comprises the following steps:

[0009] S1, collect pulse data set;

[0010] S2, generating triplet samples based on the pulse data set;

[0011] S3, based on the three-element sample, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model;

[0012] S4. Use the optimal pulse regression prediction model to predict pulse rate.

[0013] Furthermore, S2 includes the following sub-steps:

[0014] S21, randomly selecting a pulse data from the pulse data set as a plotting point label value, and generating a range of labels corresponding to positive samples and a range of labels corresponding to negative samples according to a first threshold and a second threshold;

[0015] S22. Extract positive samples from the range of labels corresponding to positive samples, and extract negative samples from the range of labels corresponding to negative samples to generate triplet samples.

[0016] Furthermore, in S21, the expression of the range of labels corresponding to the positive samples is: [t Ah -P,t Ah +P]; where t Ah represents the anchor point label value, and P represents the first threshold;

[0017] In S21, the expression of the range of labels corresponding to negative samples is: [-∞,t Ah -N]∪[t Ah +N, +∞]; where N represents the second threshold;

[0018] In S23, the expression of the triple sample is: [Y A ,Y P ,Y N ]; where Y A Indicates the anchor point, Y P represents a positive example, Y N Indicates negative examples.

[0019] Furthermore, S3 includes the following sub-steps:

[0020] S31, inputting the triplet sample into the shared network, extracting the features of the anchor sample, the features of the positive sample and the features of the negative sample;

[0021] S32, constructing a triplet loss function;

[0022] S33, constructing a ternary contrast loss function based on the features of the anchor samples, the features of the positive samples, and the features of the negative samples;

[0023] S34. Based on the ternary contrast loss function, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model.

[0024] Furthermore, in S31, the feature F of the anchor sample a The expression is:

[0025]

[0026] In the formula, θ L represents the parameter, φ L represents the transformation function of the Lth layer of the shared network, represents the output of the L-1th layer anchor sample of the shared network, σ(·) represents the activation function, and L represents the number of layers of the shared network;

[0027] In S31, the feature F of the positive sample p The expression is:

[0028]

[0029] In the formula, Represents the output of the positive sample of the L-1 layer of the shared network;

[0030] In S31, the feature F of the negative sample n The expression is:

[0031]

[0032] In the formula, Represents the output of the negative sample of the shared network layer L-1.

[0033] Furthermore, in S32, the expression of the triplet loss function Loss is:

[0034] Loss=max(d(a,p)-d(a,n)+margin,0);

[0035] Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, d(·) represents the distance measure between the feature vectors of two samples, max(·) represents the maximum value, and margin represents the hyperparameter.

[0036] Furthermore, in S33, the ternary contrast loss function L TC (Y A ,Y P ,Y N ) is:

[0037]

[0038] l Ap =|l YA -l YP |;

[0039] l AN =|l YA -l YN |;

[0040] C AP = cos(φ ResNet (Y A ), φ ResNet (Y P ));

[0041] C AN = cos(φ ResNet (Y A ), φ ResNet (Y N ));

[0042] Where Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C ANrepresents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, l YN represents the anchor point, l YP represents the positive label value, l YN represents the negative example label value, φ ResNet (·) represents the feature data obtained after processing through the shared network structure.

[0043] Furthermore, in S34, the expression for solving the optimization problem of the optimal solution is:

[0044]

[0045] Where, L TC (Y A ,Y P ,Y N ) represents the ternary contrast loss function, Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C AN represents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, and log(·) represents the logarithmic function.

[0046] The beneficial effects of the present invention are as follows: the pulse prediction method based on the three-cell neural network proposed in the present invention can improve the performance and stability of pulse prediction and reduce errors and fluctuations; automatically learn pulse-related features from a variety of physiological signals, reducing the complexity and cost of artificial feature engineering, and then through the prediction ability of the regression algorithm, a mathematical model of the pulse and physiological signals is established according to the extracted features to achieve accurate estimation of the pulse.

[0047] Based on the above method, the present invention also proposes a flexible pulse data detection system based on a three-cell neural network, including a flexible sensor module, a wireless communication module, a data acquisition module, a three-cell neural network deep learning module, an APP display module and an alarm module;

[0048] The flexible sensor module is used to measure environmental parameters;

[0049] The wireless communication module is used for wirelessly transmitting environmental parameters;

[0050] The data acquisition module is used for collecting and processing pulse data according to the perceived environmental parameters;

[0051] A three-cell neural network deep learning module is used to make predictions based on the processed pulse data;

[0052] The display module is used to display the predicted pulse data;

[0053] The alarm module is used to perform alarm processing when the pulse data is abnormal.

[0054] The beneficial effects of the present invention are: the flexible pulse data detection system proposed in the present invention can completely transmit the tiny pressure and weak mechanical wave changes under the skin to the sensing layer, effectively improving the signal-to-noise ratio, reducing signal loss, and can further reduce the mechanical wave transmission loss caused by the skin's depressed part being separated from the sensor due to the pressure. The accuracy of detection can still be guaranteed when the human body is moving, making the application range of the detection device wider and less restricted by conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the flexible pulse data detection method based on three-cell neural network;

[0056] Figure 2 It is a structural diagram of a flexible pulse data detection system based on a three-cell neural network;

[0057] Figure 3 It is a structural schematic diagram of the flexible sensor;

[0058] In the figure, 1-1 is an upper packaging layer; 1-2 is a lower packaging layer; 2 is an auxiliary layer; 3 is a flexible substrate; 4 is a sensitive layer; 5 is an electrode layer; 6 is a flexible circuit board; and 7 is a micro-column structure. DETAILED DESCRIPTION

[0059] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0060] like Figure 1 As shown, the present invention provides a flexible pulse data detection method based on a three-cell neural network, comprising the following steps:

[0061] S1, collect pulse data set;

[0062] S2, generating triplet samples based on the pulse data set;

[0063] S3, based on the three-element sample, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model;

[0064] S4. Use the optimal pulse regression prediction model to predict pulse rate.

[0065] In this embodiment of the present invention, S2 includes the following sub-steps:

[0066] S21, randomly selecting a pulse data from the pulse data set as a plotting point label value, and generating a range of labels corresponding to positive samples and a range of labels corresponding to negative samples according to a first threshold and a second threshold;

[0067] S22. Extract positive samples from the range of labels corresponding to positive samples, and extract negative samples from the range of labels corresponding to negative samples to generate triplet samples.

[0068] In the embodiment of the present invention, in S21, the expression of the range to which the positive sample corresponds to the label is: [t Ah -P,t Ah +P]; where t Ah represents the anchor point label value, and P represents the first threshold;

[0069] In S21, the expression of the range of labels corresponding to negative samples is: [-∞,t Ah -N]∪[t Ah +N, +∞]; where N represents the second threshold;

[0070] Set P=0.1, N=0.5.

[0071] In S23, the expression of the triple sample is: [Y A ,Y P ,Y N ]; where Y A Indicates the anchor point, Y P represents a positive example, Y N Indicates negative examples.

[0072] In this embodiment of the present invention, S3 includes the following sub-steps:

[0073] S31, inputting the triplet sample into the shared network, extracting the features of the anchor sample, the features of the positive sample and the features of the negative sample;

[0074] S32, constructing a triplet loss function;

[0075] S33, constructing a ternary contrast loss function based on the features of the anchor samples, the features of the positive samples, and the features of the negative samples;

[0076] S34. Based on the ternary contrast loss function, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model.

[0077] The training goal of the triplet neural network is to minimize the high-dimensional representation distance between the anchor point and the positive example, and to maximize the high-dimensional representation distance between the anchor point and the negative example.

[0078] In the embodiment of the present invention, in S31, the feature F of the anchor point sample a The expression is:

[0079]

[0080] In the formula, θ L represents the parameter, φ L represents the transformation function of the Lth layer of the shared network, represents the output of the L-1th layer anchor sample of the shared network, σ(·) represents the activation function, and L represents the number of layers of the shared network;

[0081] In S31, the feature F of the positive sample p The expression is:

[0082]

[0083] In the formula, Represents the output of the positive sample of the L-1 layer of the shared network;

[0084] In S31, the feature F of the negative sample n The expression is:

[0085]

[0086] In the formula, Represents the output of the negative sample of the shared network layer L-1.

[0087] In the embodiment of the present invention, in S32, the expression of the triple loss function Loss is:

[0088] Loss=max(d(a,p)-d(a,n)+margin,0);

[0089] Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, d(·) represents the distance measure between the feature vectors of two samples, max(·) represents the maximum value, and margin represents the hyperparameter.

[0090] In the embodiment of the present invention, in S33, the ternary contrast loss function L TC (Y A ,Y P ,Y N ) is:

[0091]

[0092] l AP =|l YA -l YP |;

[0093] l AN =|l YA -l YN |;

[0094] C AP = cos(φ ResNet (Y A ), φ ResNet (Y P ));

[0095] C AN = cos(φ ResNet (Y A ), φ ResNet (Y N ));

[0096] Where Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C AN represents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, l YA represents the anchor point, l YP represents the positive label value, l YN represents the negative example label value, φ ResNet (·) represents the feature data obtained after processing through the shared network structure.

[0097] In the embodiment of the present invention, in S34, the expression for solving the optimization problem of the optimal solution is:

[0098]

[0099] Where, L TC (Y A ,Y P ,Y N ) represents the ternary contrast loss function, Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C AN represents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, and log(·) represents the logarithmic function.

[0100] The triplet residual neural network is used for feature extraction, and the Adam optimizer is used to calculate the gradient of the ternary contrast loss function TC Loss with respect to the network variables to optimize the network. The batch size is set to 128, and the learning rate is set to 0.001, so that the anchor point output by the network is close to the positive feature data, and the anchor point is far from the negative feature data. The Adam formula is as follows:

[0101] m t =μ*m t-1 +g t -μg t ;

[0102] n t =v*m t-1 +g t 2 -vg t 2 ;

[0103]

[0104] In the formula, m t represents the first-order moment estimate, μ represents a hyperparameter close to 1, and m t-1 represents the first-order moment estimate of the previous time step, g t represents the gradient, n t represents the second-order moment estimate, v represents a hyperparameter close to 1, represents the value after the deviation correction of the first-order moment, represents the value after the deviation correction of the second-order moment, Δθ t represents the update parameter, ∈ represents a very small number, usually 10 -8 , used to prevent division by zero, and η represents the learning rate.

[0105] Let θ t represents the parameter of the tth iteration, t=1,2,...,T, then l TC (θ t ) represents the loss value of the tth iteration, η t represents the learning rate at the tth iteration.

[0106] When the learning rate Then when the maximum number of iterations T→∞, the loss value of the deep triplet residual network is l TC (θ t ) converges to the optimal solution Get the optimal model.

[0107] The residual blocks of the deep residual neural network ResNet-50 are Model, and finally predict the pulse rate based on the optimal pulse regression prediction model.

[0108] Based on the above method, the present invention also proposes a flexible pulse data detection system based on a three-cell neural network, such as Figure 2 As shown, it includes a flexible sensor module, a wireless communication module, a data acquisition module, a three-cell neural network deep learning module, an APP display module and an alarm module;

[0109] The flexible sensor module is used to measure environmental parameters;

[0110] The wireless communication module is used for wirelessly transmitting environmental parameters;

[0111] The data acquisition module is used for collecting and processing pulse data according to the perceived environmental parameters;

[0112] A three-cell neural network deep learning module is used to make predictions based on the processed pulse data;

[0113] The display module is used to display the predicted pulse data;

[0114] The alarm module is used to perform alarm processing when the pulse data is abnormal.

[0115] Set multi-level thresholds to detect anomalies in real-time data, and make threshold judgments on monitoring data. When the data exceeds the set threshold, an alarm is triggered; a set of rules is predefined to match the data. The normal range of human pulse rate is 60-100 beats / minute. When the pulse rate is less than 60 beats / minute or greater than 100 beats / minute, it is set to abnormal mode. When a specific abnormal mode is met, an alarm is triggered, and an alarm notification is sent according to the alarm level and the configured alarm mode; it is displayed on the mobile phone APP, and an information alarm or a sound alarm is sounded.

[0116] like Figure 3As shown, the flexible pulse sensor in the flexible sensor module includes an upper packaging layer 1-1, an auxiliary layer 2, a sensing layer including (flexible substrate 3, sensitive layer 4, electrode layer 5), a flexible circuit board 6 and a lower packaging layer 1-2. The flexible substrate 3 is made of thermoplastic polyurethane, and the upper packaging layer 1-1 and the lower packaging layer 1-2 are made of silicone. The thickness of the auxiliary layer 2 is 100um-200um. The addition of thermoplastic polyurethane to the material of the auxiliary layer 2 can ensure that the elastic modulus of the auxiliary layer 2 meets the range of 0.1-1MPa on the one hand, and can improve the flexibility of the flexible pulse sensor on the other hand. The electrode layer 5 is provided with multiple groups, each electrode is independently provided and connected to the signal acquisition circuit through a separate signal line; a micro-column structure 7 is arranged on the sensitive layer 4, wherein the side where the micro-column structure 7 is located faces the electrode; the number of sensitive layers 4 is the same as the number of electrode groups, and corresponds to the electrode position one by one, and each group of sensitive layers 4 is connected to the corresponding electrode through an elastic adhesive. The elastic adhesive precursor is a thermosetting elastomer, including a combination of thermosetting polydimethylsiloxane prepolymer and thermosetting polyurethane. It can improve the firmness of the connection between the layers, reduce the movement and dislocation between the layers, and thus improve the accuracy and sensitivity of the detection. Adjacent electrodes are connected by serpentine curves. The outer contour of the electrode is elliptical, and the electrodes are arranged in three groups, and the three groups of electrodes are arranged in parallel and coplanar.

[0117] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A flexible pulse data detection method based on a three-cell neural network, characterized in that: The following steps are involved: S1, collect pulse data set; S2, generating triplet samples based on the pulse data set; S3, based on the three-element sample, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model; S4. Use the optimal pulse regression prediction model to predict pulse rate.

2. The flexible pulse data detection method based on three-cell neural network according to claim 1 is characterized in that: The S2 comprises the following sub-steps: S21, randomly selecting a pulse data from the pulse data set as a plotting point label value, and generating a range of labels corresponding to positive samples and a range of labels corresponding to negative samples according to a first threshold and a second threshold; S22. Extract positive samples from the range of labels corresponding to positive samples, and extract negative samples from the range of labels corresponding to negative samples to generate triplet samples.

3. The flexible pulse data detection method based on three-cell neural network according to claim 2 is characterized in that: In S21, the expression of the range to which the positive sample corresponds to the label is: Ah -P,t Ah +P]; where t Ah represents the anchor point label value, and P represents the first threshold; In S21, the expression of the range of labels corresponding to negative samples is: [-∞,t Ah -N]∪[t Ah +N, +∞]; where N represents the second threshold; In S23, the expression of the triple sample is: [Y A ,Y P ,Y N ]; where Y A Indicates the anchor point, Y P represents a positive example, Y N Indicates negative examples.

4. The flexible pulse data detection method based on three-cell neural network according to claim 1 is characterized in that: The S3 comprises the following sub-steps: S31, inputting the triplet sample into the shared network, extracting the features of the anchor sample, the features of the positive sample and the features of the negative sample; S32, constructing a triplet loss function; S33, constructing a ternary contrast loss function based on the features of the anchor samples, the features of the positive samples, and the features of the negative samples; S34. Based on the ternary contrast loss function, the optimal solution of the triplet residual neural network is solved to obtain the optimal pulse regression prediction model.

5. The flexible pulse data detection method based on three-cell neural network according to claim 4 is characterized in that: In S31, the feature F of the anchor point sample a The expression is: In the formula, θ L represents the parameter, φ L represents the transformation function of the Lth layer of the shared network, represents the output of the L-1th layer anchor sample of the shared network, σ(·) represents the activation function, and L represents the number of layers of the shared network; In S31, the feature F of the positive sample p The expression is: In the formula, Represents the output of the positive sample of the L-1 layer of the shared network; In S31, the feature F of the negative sample n The expression is: In the formula, Represents the output of the negative sample of the shared network layer L-1.

6. The flexible pulse data detection method based on three-cell neural network according to claim 4 is characterized in that: In S32, the expression of the triplet loss function Loss is: Loss=max(d(a,p)-d(a,n)+margin,0); Where a represents the anchor sample, p represents the positive sample, n represents the negative sample, d(·) represents the distance measure between the feature vectors of two samples, max(·) represents the maximum value, and margin represents the hyperparameter.

7. The flexible pulse data detection method based on three-cell neural network according to claim 4 is characterized in that: In S33, the ternary contrast loss function L TC (Y A ,Y P ,Y N ) is: l AP ＝|l YA -l YP |； l AN ＝|l YA -l YN |； Where Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C AN represents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, l YA represents the anchor point, l YP represents the positive label value, l YN represents the negative example label value, φ ResNet (·) represents the feature data obtained after processing through the shared network structure.

8. The flexible pulse data detection method based on three-cell neural network according to claim 4 is characterized in that: In S34, the expression for solving the optimization problem of the optimal solution is: Where, L TC (Y A ,Y P ,Y N ) represents the ternary contrast loss function, Y A Indicates the anchor point, Y P represents a positive example, Y N represents a negative example, l AP Represents the absolute value of the difference between the anchor point and the positive example label value, C AP Indicates the distance between the metric anchor point and the positive example, l AN Represents the absolute value of the difference between the anchor point and the negative example label value, C AN represents the distance between the metric anchor point and the negative example, α2 represents the marginal parameter, exp(·) represents the exponential function, and log(·) represents the logarithmic function.

9. A flexible pulse data detection system based on a three-cell neural network, characterized in that: It includes flexible sensor module, wireless communication module, data acquisition module, three-cell neural network deep learning module, APP display module and alarm module; The flexible sensor module is used to measure environmental parameters; The wireless communication module is used for wirelessly transmitting environmental parameters; The data acquisition module is used to collect and process pulse data according to the perceived environmental parameters; The three-cell neural network deep learning module is used to make predictions based on the processed pulse data; The display module is used to display the predicted pulse data; The alarm module is used to perform alarm processing when the pulse data is abnormal.

Images