Rehabilitation assessment method based on autoencoder and softmax classifier

Abstract

This invention discloses a rehabilitation assessment method based on an autoencoder and a Softmax classifier, belonging to the field of flexible strain sensor measurement technology. Specifically, it involves: multiple trainees with known rehabilitation levels completing a Carroll hand function assessment task; finger motion monitoring gloves acquiring finger data; and processing the data through interpolation and sampling; constructing a rehabilitation level assessment model composed of an autoencoder and a Softmax classifier; training the rehabilitation level assessment model using the processed data as input and the known rehabilitation levels as training targets; having the test subject complete the task and obtain finger data to be assessed, which is then interpolated and sampled before being input into the trained rehabilitation level assessment model, outputting the rehabilitation level assessment result. This invention employs cubic Hermitian interpolation and sampling followed by equal-value interpolation, using fewer data points to more accurately reflect the characteristics of the finger data, and combining an autoencoder and a Softmax classifier to achieve high-accuracy assessment.

Description

Technical Field

[0001] This invention belongs to the field of flexible strain sensor measurement technology, specifically relating to a rehabilitation assessment method based on an autoencoder and a Softmax classifier. Background Technology

[0002] Hand injuries, spasms, and strokes can all cause hand dysfunction. Traditional hand rehabilitation level assessment requires professional doctors, which is time-consuming and laborious, and the assessment results may be affected by the doctor's subjectivity. Effective identification of hand rehabilitation levels through intelligent algorithms will promote the development of digital healthcare. Currently, the process of digitally identifying hand rehabilitation levels mainly includes: (1) extracting data features using principal component analysis; (2) using machine learning methods such as support vector machine, decision tree, random forest, and Naive Bayes to assess hand rehabilitation levels based on the extracted data features. However, for the original data of different hand rehabilitation levels collected by finger motion monitoring gloves (i.e., flexible strain sensors distributed on the five fingers), the data dimension is high after transformation using principal component analysis, resulting in poor recognition effect of subsequent machine learning. Summary of the Invention

[0003] To address the problems in the prior art, this invention provides a rehabilitation assessment method based on an autoencoder and a Softmax classifier, achieving a high-accuracy assessment of hand rehabilitation levels.

[0004] The technical solution adopted in this invention is as follows:

[0005] The rehabilitation assessment method based on autoencoders and Softmax classifiers is characterized by the following steps:

[0006] S1. After multiple trainees with known rehabilitation levels complete the Carroll hand function assessment task, the hand movement is sensed in real time through a finger movement monitoring glove to obtain the corresponding finger data.

[0007] S2. Interpolate and extract the data collected from the fingers to obtain the processed data of each trainee, so as to ensure that the processed data of each trainee contains the same number of data points.

[0008] S3. Construct a rehabilitation level assessment model consisting of an autoencoder and a Softmax classifier; wherein, the autoencoder is used to extract features from the processed data, and the Softmax classifier performs rehabilitation level assessment based on the extracted features.

[0009] S4. Using the processed data of each trainee as input and the known rehabilitation level as the training target, train the rehabilitation level assessment model to obtain the post-training rehabilitation level assessment model.

[0010] S5. The test subject completes the Carroll hand function assessment task. The data collected from the fingers to be assessed is obtained in real time through the finger movement monitoring glove. After interpolation and sampling, the data is input into the post-training rehabilitation level assessment model, and the rehabilitation level assessment result of the test subject is finally output.

[0011] Furthermore, the specific processes of interpolation and decrementing in S2 are as follows:

[0012] S21. Calculate the time Δt for each trainee to acquire finger data through the finger movement monitoring glove.

[0013] S22. Interpolate and sample data for each trainee within the longest time period T during which finger movement monitoring gloves are used to acquire finger data. The selected time point is 10. x x represents 10 equally spaced points within the range of [-1, 1.47]. First, the finger data collected within Δt is subjected to three Hermitian interpolations and decimations. Then, based on the finger data collected at time Δt, equal numerical interpolation is performed within the time range from time Δt to time T.

[0014] Furthermore, T should not exceed 30 seconds.

[0015] Furthermore, the structure of the autoencoder in S3 includes:

[0016] One input layer with 50 neurons;

[0017] There are 3 hidden layers, with 15, 3 and 15 neurons in each layer, respectively;

[0018] One output layer with 50 neurons;

[0019] The activation functions for the input layer, hidden layer, and output layer are all ReLU functions.

[0020] Furthermore, the structure of the Softmax classifier in S3 includes:

[0021] One input layer with 3 neurons;

[0022] One output layer with 4 neurons;

[0023] There are 12 structural parameters for hidden layers:

[0024] (1) One hidden layer with 10 neurons;

[0025] (2) One hidden layer with 5 neurons;

[0026] (3) One hidden layer with 20 neurons;

[0027] (4) Two hidden layers, with 5 and 10 neurons in each layer;

[0028] (5) Two hidden layers, with 10 and 5 neurons in each layer, respectively;

[0029] (6) Two hidden layers, with 10 neurons in each layer and 10 neurons in each layer;

[0030] (7) Two hidden layers, with 15 and 10 neurons in each layer, respectively;

[0031] (8) Two hidden layers, with 10 and 15 neurons in each layer, respectively;

[0032] (9) Two hidden layers, with 15 neurons in each layer;

[0033] (10) 3 hidden layers, with 10, 10 and 10 neurons in each layer respectively;

[0034] (11) There are 3 hidden layers, with 15, 10 and 10 neurons in each layer, respectively;

[0035] (12) There are 3 hidden layers, with 15, 15 and 10 neurons in each layer, respectively;

[0036] The activation function for the output layer is the Softmax function, while the activation functions for the input layer and the hidden layer are both the ReLU function.

[0037] The beneficial effects of this invention are as follows:

[0038] (1) For the time-varying output signal collected by the finger movement monitoring glove, the existing technology usually uses equal time intervals for interpolation and sampling, and the interpolation and sampling methods are the same throughout the time range. If the present invention uses the existing technology to process the finger collection data, there will be a large amount of data after processing, and the error of the processed data will be large in the time range from Δt to T. Based on this, the rehabilitation assessment method based on autoencoder and Softmax classifier proposed in this invention performs three Hermitian interpolations and samplings within the time range within Δt, and equal numerical interpolation is performed in the time range from Δt to T. The interpolation and sampling time points change exponentially, and the finger collection data characteristics are reflected more accurately with fewer data points.

[0039] (2) Since the dimension of the processed data is high, the present invention uses an autoencoder to reduce the dimension of the processed data. Compared with the dimension reduction process of the principal component analysis method (the dimension of the data after dimension reduction is still high), the autoencoder only uses three-dimensional data to reflect the characteristics of the finger-collected data, ensuring the high accuracy of subsequent machine learning.

[0040] (3) The present invention uses the Softmax classifier to evaluate the level of hand rehabilitation, which is more suitable for the features of the processed data extracted by the autoencoder and achieves high accuracy. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the rehabilitation assessment method based on an autoencoder and a Softmax classifier proposed in Embodiment 1 of the present invention.

[0042] Figure 2 This is a schematic diagram of the structure of the self-encoder in Embodiment 1 of the present invention;

[0043] Figure 3 This is a schematic diagram of the Softmax classifier in Embodiment 1 of the present invention;

[0044] Figure 4 This is a waveform comparison diagram of the finger-collected data and the data after interpolation and sampling in Embodiment 1 of the present invention;

[0045] Figure 5 This is a comparison of the loss between the training set and the test set of the rehabilitation level assessment model using the hidden layer with structural parameter (1) in Embodiment 1 of the present invention during the training process. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Example 1

[0048] This embodiment provides a rehabilitation assessment method based on an autoencoder and a Softmax classifier, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0049] S1. After multiple individuals with known rehabilitation levels complete the Carroll hand function assessment task (e.g., grasping a wooden block), the hand movement is sensed in real time through a finger movement monitoring glove, and corresponding finger data is collected.

[0050] It is worth noting that individuals have different known rehabilitation levels. During the Carroll hand function assessment task, the sensor of the finger movement monitoring glove will output a signal, and the data acquisition instrument may switch gears due to large changes in the sensor output signal, resulting in non-equal time intervals between adjacent data points. In addition, individuals with different hand rehabilitation levels take different times to complete the same task, resulting in different data collection times (finger data) collected by the finger movement monitoring glove, but the longest time T generally does not exceed 30 seconds.

[0051] 80% of the participants were randomly selected as trainers to form the training set, and the remaining 20% ​​were selected as testers to form the test set.

[0052] S2. Interpolate and sample the finger data collected from each trainee to obtain processed data for each trainee, ensuring that the processed data for each trainee contains the same number of data points; specifically including the following steps:

[0053] S21. Calculate the time Δt for each trainee to acquire finger data through the finger movement monitoring glove.

[0054] S22. Interpolation and sampling are performed on the longest time T = 30s for each trainee to acquire finger data using the finger movement monitoring glove, with the selected time point being 10. x x represents 10 equally spaced points within the range [-1, 1.47], with -1 as the first time point and 1.47 as the last time point; firstly, the finger data collected within Δt is subjected to three Hermite interpolations and decimations, and then, based on the finger data collected at time Δt, equal numerical interpolation is performed within the time range from time Δt to time T.

[0055] S3. Construct a rehabilitation level assessment model consisting of an autoencoder and a Softmax classifier; among which,

[0056] The autoencoder is used to extract features from the processed data, and its structure is as follows: Figure 2 As shown, it includes:

[0057] One input layer with 50 neurons;

[0058] There are 3 hidden layers, with 15, 3 and 15 neurons in each layer, respectively;

[0059] One output layer with 50 neurons;

[0060] The activation functions for the input layer, hidden layer, and output layer are all ReLU functions;

[0061] The Softmax classifier assesses rehabilitation level based on extracted features, and its structure is as follows: Figure 3 As shown, it includes:

[0062] One input layer with 3 neurons;

[0063] One output layer with 4 neurons;

[0064] There are 12 structural parameters for hidden layers:

[0065] (1) One hidden layer with 10 neurons;

[0066] (2) One hidden layer with 5 neurons;

[0067] (3) One hidden layer with 20 neurons;

[0068] (4) Two hidden layers, with 5 and 10 neurons in each layer;

[0069] (5) Two hidden layers, with 10 and 5 neurons in each layer, respectively;

[0070] (6) Two hidden layers, with 10 neurons in each layer and 10 neurons in each layer;

[0071] (7) Two hidden layers, with 15 and 10 neurons in each layer, respectively;

[0072] (8) Two hidden layers, with 10 and 15 neurons in each layer, respectively;

[0073] (9) Two hidden layers, with 15 neurons in each layer;

[0074] (10) 3 hidden layers, with 10, 10 and 10 neurons in each layer respectively;

[0075] (11) There are 3 hidden layers, with 15, 10 and 10 neurons in each layer, respectively;

[0076] (12) There are 3 hidden layers, with 15, 15 and 10 neurons in each layer, respectively;

[0077] The activation function of the output layer is the Softmax function, while the activation functions of the input layer and the hidden layer are both the ReLU function.

[0078] S4. Using the processed data of each trainee as input and the known rehabilitation level as the training target, train the rehabilitation level assessment model to obtain the post-training rehabilitation level assessment model.

[0079] S5. Perform the same interpolation and sampling on the finger data collected from each tester as on the trainer, input the results into the post-training rehabilitation level assessment model, and finally output the rehabilitation level assessment results of each tester, which are then compared with their corresponding known rehabilitation levels.

[0080] Figure 4 The diagram shows a waveform comparison between the finger-collected data obtained in S1 and the processed data obtained in S2 after interpolation and sampling. The diagram demonstrates that the waveform of the processed data accurately reflects the waveform characteristics of the finger-collected data. Furthermore, during the time period when the finger motion monitoring glove does not collect data, since the output signal of the glove theoretically does not change, fewer interpolated data points are needed to reflect the data characteristics during this period.

[0081] Figure 5 The results of the comparison between the training set and the test set loss of the rehabilitation level assessment model with hidden layer using structural parameter (1) during the training process are shown (the average of 10 training results is taken). As the number of training times increases, the loss of the training set and the test set gradually decreases and tends to stabilize, and the two curves are very close, indicating that the rehabilitation level assessment model has a small degree of overfitting.

[0082] In this embodiment, the rehabilitation level assessment models based on the above 12 hidden layer structural parameters were trained respectively. The comparison results of the loss between the training set and the test set (the average of 10 training results) are shown in Table 1. It can be seen that the rehabilitation level assessment models based on hidden layers with different structural parameters all show high accuracy.

[0083] Table 1

[0084]

[0085]

Claims

1. A rehabilitation assessment method based on an autoencoder and a Softmax classifier, characterized in that, Includes the following steps: S1. After multiple trainees with known rehabilitation levels complete the Carroll hand function assessment task, the hand movement is sensed in real time through a finger movement monitoring glove, and corresponding finger data is collected; wherein, the finger movement monitoring glove has flexible strain sensors distributed on five fingers. S2. Interpolate and sample the finger-collected data to obtain the processed data for each trainee. The specific process is as follows: S21. Calculate the time Δt for each trainee to acquire finger data through the finger movement monitoring glove. S22. Interpolate and sample data for each trainee within the longest time period T during which finger movement monitoring gloves are used to acquire finger data. The selected time point is 10. x x represents 10 equally spaced points within the range of [-1, 1.47]. First, the finger data collected within Δt is subjected to three Hermitian interpolations and decimations. Then, based on the finger data collected at time Δt, equal numerical interpolation is performed within the time range from time Δt to time T. S3. Construct a rehabilitation level assessment model consisting of an autoencoder and a Softmax classifier; wherein, the autoencoder is used to extract features from the processed data, and the Softmax classifier performs rehabilitation level assessment based on the extracted features. S4. Using the processed data of each trainee as input and the known rehabilitation level as the training target, train the rehabilitation level assessment model to obtain the post-training rehabilitation level assessment model. S5. The test subject completes the Carroll hand function assessment task. The data collected from the fingers to be assessed is obtained in real time through the finger movement monitoring glove. After interpolation and sampling, the data is input into the post-training rehabilitation level assessment model, and the rehabilitation level assessment result of the test subject is finally output.

2. The rehabilitation assessment method based on autoencoder and Softmax classifier according to claim 1, characterized in that, T should not exceed 30 seconds.

3. The rehabilitation assessment method based on autoencoder and Softmax classifier according to claim 1, characterized in that, The structure of the autoencoder in S3 includes: One input layer with 50 neurons; There are 3 hidden layers, with 15, 3 and 15 neurons in each layer, respectively; One output layer with 50 neurons; The activation functions for the input layer, hidden layer, and output layer are all ReLU functions.

4. The rehabilitation assessment method based on autoencoder and Softmax classifier according to claim 1, characterized in that, The structure of the Softmax classifier in S3 includes: One input layer with 3 neurons; One output layer with 4 neurons; There are 12 structural parameters for hidden layers: (1) One hidden layer with 10 neurons; (2) One hidden layer with 5 neurons; (3) One hidden layer with 20 neurons; (4) Two hidden layers, with 5 and 10 neurons in each layer; (5) Two hidden layers, with 10 and 5 neurons in each layer, respectively; (6) Two hidden layers, with 10 neurons in each layer and 10 neurons in each layer; (7) Two hidden layers, with 15 and 10 neurons in each layer, respectively; (8) Two hidden layers, with 10 and 15 neurons in each layer, respectively; (9) Two hidden layers, with 15 neurons in each layer; (10) 3 hidden layers, with 10, 10 and 10 neurons in each layer respectively; (11) There are 3 hidden layers, with 15, 10 and 10 neurons in each layer, respectively; (12) There are 3 hidden layers, with 15, 15 and 10 neurons in each layer, respectively; The activation function for the output layer is the Softmax function, while the activation functions for the input layer and the hidden layer are both the ReLU function.

Images