Temporomandibular joint vibration mechanical wave acquisition and identification method
By collecting and processing temporomandibular joint vibration mechanical waves using sensor components, generating feature parameters, and inputting them into an optimized random forest model, the problem of poor recognition performance in existing technologies is solved, and efficient temporomandibular joint vibration mechanical wave recognition is achieved.
Patent Information
- Application Number
- CN202410919702.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2026-01-13
AI Technical Summary
In existing technologies, the recognition effect of temporomandibular joint vibration mechanical waves is not ideal, mainly due to the lack of artificial intelligence model training for the characteristics of temporomandibular joint vibration mechanical waves, resulting in poor recognition performance.
Sensor components are used to collect mechanical vibration waves of the temporomandibular joint, generate simulated piezoelectric signals, amplify and convert them from analog to digital, extract feature parameters, and input them into a trained random forest model for recognition. Feature parameters include root mean square of signal amplitude, cumulative value of signal waveform length, median frequency, and mean of power frequency distribution. By optimizing the hyperparameters and training process of the random forest model, the recognition accuracy is improved.
It achieves efficient acquisition and recognition of mechanical vibration waves of the temporomandibular joint, improves the recognition effect, solves the problem of unsatisfactory recognition effect in the existing technology, and the structural design of the sensor component ensures reliable signal acquisition.
Smart Images
Figure CN121313152A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical devices, and particularly relates to a temporomandibular joint vibration mechanical wave acquisition and identification method. BACKGROUND
[0002] Traditional temporomandibular joint examination methods mostly rely on the clinical experience of doctors and the subjective feelings of patients, and have certain limitations in accuracy and objectivity. In the research of temporomandibular joint disorders, by collecting mechanical wave signals generated by vibration during temporomandibular joint activity, and identifying the condition of temporomandibular joint disorders based on the collected signals, the vibration characteristics, stress distribution and pathological changes of the joint can be more comprehensively understood, thereby providing an important basis for early detection, diagnosis and treatment of diseases. The industry has tried to use artificial intelligence technology to automatically identify the collected temporomandibular joint vibration signals, but the results are often unsatisfactory. The main reason is that most of the conventional artificial intelligence algorithms are used, and there is a lack of artificial intelligence model training for the characteristics of temporomandibular joint vibration mechanical waves, so the recognition effect is not good. SUMMARY
[0003] In view of the above analysis, the embodiments of the present application aim to provide a temporomandibular joint vibration mechanical wave acquisition and identification method to solve the problem of unsatisfactory recognition effect of the existing temporomandibular joint vibration mechanical wave.
[0004] In one aspect, the embodiments of the present application provide a temporomandibular joint vibration mechanical wave acquisition and identification method, which specifically comprises:
[0005] Collecting temporomandibular joint vibration mechanical waves by using a sensor assembly and generating analog piezoelectric signals;
[0006] After amplifying and analog-to-digital converting the analog piezoelectric signals, digital piezoelectric signals are obtained;
[0007] Feature extraction is performed on the digital piezoelectric signals to obtain feature parameters of the digital piezoelectric signals;
[0008] The feature parameters are input into a trained random forest model to obtain an identification result of whether the temporomandibular joint vibration mechanical wave is abnormal.
[0009] The above technical solution has the following beneficial effects: the collection and processing of temporomandibular joint vibration mechanical waves are realized to generate digital piezoelectric signals, and the automatic identification of whether the temporomandibular joint vibration mechanical wave is abnormal is realized based on the automatic digital piezoelectric signals.
[0010] Based on the further improvement of the above method, the feature parameters of the digital piezoelectric signals specifically include signal amplitude root mean square, signal waveform length cumulative value, frequency median, and power frequency distribution mean value, wherein,
[0011] The median frequency is calculated using the following formula, which is specifically expressed as follows:
[0012] In the formula, MF is the median frequency, X(f) is the frequency domain signal obtained after the Fourier transform of the signal, and df is the differential of the frequency.
[0013] The mean value of the power frequency distribution is calculated using the following formula, which is specifically expressed as follows:
[0014] In the formula, MPF is the mean power frequency distribution; PSD(f) = |X(f)| 2 .
[0015] The beneficial effect of the above-mentioned further improvement scheme is that the four feature parameters have a better recognition effect than other feature parameters when used to identify temporomandibular joint vibration mechanical waves.
[0016] Based on further improvements to the above method, a trained random forest model is obtained through the following method, which specifically includes:
[0017] Feature parameters are extracted from the original signal samples used for model training, and the original signal samples are labeled to construct training and test sets;
[0018] The number of decision trees and the minimum number of leaf samples are used as model hyperparameters, and the range of values for the decision trees and the minimum leaf samples, as well as the maximum number of search times, are set respectively.
[0019] SA1: Randomly select one value from the decision tree and the minimum leaf sample value range, and generate a random forest model to be trained based on the random value;
[0020] SA2: The trained random forest model is generated by training the random forest model to be trained using the training set.
[0021] SA3: Calculate and save the fitness value of the trained random forest model;
[0022] Repeat SA1-SA3 until the maximum number of search rounds is reached;
[0023] The fitness values of each trained random forest model are calculated based on the training and test sets, and the random forest model with the best fitness value is selected as the trained random forest model.
[0024] The beneficial effect of the above-mentioned further improvement scheme is that the method can obtain a random forest model with optimal identification ability for whether the mechanical waves of temporomandibular joint vibration are abnormal.
[0025] Based on a further improvement of the above method, the training set is used to train the random forest model to be trained, and then each trained random forest model is generated. The method specifically includes:
[0026] In the training set, feature parameters of 75% of the categories of each sample are randomly selected and input into the root node of a decision tree in the random forest model to be trained.
[0027] The sample entropy before splitting is calculated based on the sample feature parameters of the input training set, and a decision tree is established through the following steps, which specifically include:
[0028] SS1: Sort the feature parameters of each category for each sample in descending order;
[0029] SS2: Calculate the split value sequentially based on the comparison results of adjacent feature parameters in each queue;
[0030] SS3: Splits from the current root node into two leaf nodes;
[0031] SS4: Based on the comparison results of each split value and the corresponding category feature parameter, each sample is divided into two groups. The samples whose corresponding category feature parameter is less than the split value are divided into one group, and the remaining samples are divided into another group. The sample entropy after grouping is calculated for each group.
[0032] SS5: Save the difference between the entropy before splitting and the entropy after grouping of each sample feature parameter as the gain, associate the corresponding category feature with the largest gain with the current root node, and use the splitting value corresponding to the feature parameter with the largest gain as the splitting value of the current root node;
[0033] SS6: Based on the split value grouping result of the current root node, assign the two groups of samples to a leaf node of the current root node respectively, and remove the feature parameter corresponding to the feature category with the largest gain from each sample;
[0034] SS1-SS6 are executed iteratively on the samples of each leaf node. In each iteration, the feature parameters of the class features of the associated nodes are ignored until all class features are associated with a certain node, at which point the decision tree is completed.
[0035] Once all decision trees in the random forest model to be trained are built, the trained random forest model is obtained.
[0036] The beneficial effects of the above-mentioned further improvement scheme are: 75% of the feature parameters of the type features are randomly selected from the samples in each training set, ensuring that the depth of the decision tree is appropriate during the construction of the decision tree, and that the samples used to build each decision tree are different to prevent the forest model from overfitting at any time; after grouping each sample based on the split value, the optimal split value is obtained by calculating the gain, so that each node of the constructed decision tree obtains the best recognition ability for a certain feature parameter of the input sample, which is conducive to improving the overall recognition ability of the random forest model.
[0037] A further improvement to the above method involves calculating the splitting value sequentially based on the comparison results of adjacent feature parameters in each queue, specifically including:
[0038] Each queue checks whether adjacent feature parameters are the same from the beginning. If they are the same, the comparison continues until adjacent feature parameters are different. The sum of 2 / 3 of the previous feature parameter and 1 / 3 of the next feature parameter is rounded to the nearest integer and saved as the split value.
[0039] Otherwise, the current feature parameter and the mean of adjacent feature parameters are rounded to the nearest integer and saved as the split value until each queue has completed the judgment of all adjacent feature parameters.
[0040] The beneficial effects of the above-mentioned further improvement plan are:
[0041] Further improvements to the above method, specifically including calculating the fitness values of each trained random forest model based on the training and test sets, and selecting the random forest model with the best fitness value as the trained random forest model, include:
[0042] The training set and the test set are respectively input into the trained random forest model, and the recognition results are output respectively.
[0043] Obtain the error between the recognition result and the labeled value of each sample, and obtain the error of the training set based on the error of all samples in the training set, and obtain the error of the test set based on the error of all samples in the test set; use the sum of the errors of the training set and the test set as the recognition error value of the random forest model.
[0044] The ratio of the recognition error value of each trained random forest model to the total number of samples in the input training and test sets is calculated as the fitness value of each trained random forest model.
[0045] The trained random forest model with the smallest fitness value is used as the trained random forest model.
[0046] The beneficial effect of the above-mentioned further improvement scheme is that the fitness value calculation helps to select the random forest model with the best recognition ability from all trained random forest models as the trained random forest model, thereby obtaining the best recognition effect for the mechanical waves of temporomandibular joint vibration.
[0047] Based on a further improvement of the above method, the sensor assembly includes a transmission component, a piezoelectric film, and a fixing component stacked together, wherein,
[0048] The transmission component is used to transmit the mechanical vibration waves of the temporomandibular joint to the piezoelectric film.
[0049] The piezoelectric film is used to detect the vibrational mechanical waves transmitted by the transmission component and trigger the generation of a simulated piezoelectric signal;
[0050] The piezoelectric film and the transmission component are stacked sequentially on the fixed component.
[0051] The beneficial effect of the above-mentioned further improvement scheme is that the sensor component structure can transmit the mechanical waves of temporomandibular joint vibration to the piezoelectric film to generate an analog electrical signal for abnormal detection.
[0052] Based on a further improvement of the above method, the transmission component is specifically:
[0053] Made of skin-friendly insulating material with good mechanical wave conduction capability;
[0054] The upper and lower surfaces are parallel, with the upper surface in contact with the human body and the lower surface in contact with the piezoelectric film.
[0055] The lower surface has the same shape as the piezoelectric film, but its cross-sectional area is slightly smaller than that of the piezoelectric film.
[0056] The edges are bonded and fixed to the piezoelectric film by center alignment with the piezoelectric film.
[0057] The beneficial effect of the above-mentioned further improvement scheme is that the structure and arrangement of the transmission component are conducive to better acquisition of the mechanical vibration waves of the temporomandibular joint and transmission to the piezoelectric film.
[0058] Based on a further improvement of the above method, the sensor assembly also includes a metal contact, the upper surface of which is at the same height as the upper surface of the transmission component, and is configured with a smooth surface and edge structure.
[0059] The beneficial effects of the above-mentioned further improvement scheme are that the structure and arrangement of the metal contact points are conducive to the sensor assembly fitting better with the human body and alleviating human discomfort.
[0060] Based on a further improvement of the above method, the fixing component is specifically: the upper and lower surfaces are parallel, wherein the groove or protrusion is provided on one side of the upper surface, the cross-sectional shape of the groove or protrusion is the same as that of the piezoelectric film, the area is smaller than that of the piezoelectric film, the piezoelectric film covers the groove or protrusion, and the edge is bonded to the fixing component; the metal contact is fixedly provided on the other side of the upper surface.
[0061] The beneficial effects of the above-mentioned further improvement scheme are: the structure and arrangement of the fixed component ensure the overall structural stability of the sensor assembly, which is conducive to the piezoelectric film generating simulated piezoelectric signals better based on vibration mechanical waves.
[0062] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from the description and drawings, which are particularly pointed out. Attached Figure Description
[0063] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0064] Figure 1 This is a schematic diagram of the temporomandibular joint vibration mechanical wave acquisition and identification method according to an embodiment of the present invention.
[0065] Figure 2 This is a schematic diagram of the sensor assembly structure according to an embodiment of the present invention.
[0066] Figure label:
[0067] 1-Transmission component; 2-Piezoelectric film; 3-Fixing component; 4-Metal contact. Detailed Implementation
[0068] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0069] A specific embodiment of the present invention discloses a method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint, such as... Figure 1 As shown.
[0070] The method specifically includes:
[0071] Sensor components are used to collect vibrational mechanical waves of the temporomandibular joint and generate simulated piezoelectric signals;
[0072] A digital piezoelectric signal is obtained by amplifying and converting the analog piezoelectric signal to a digital one.
[0073] Feature extraction is performed on the digital piezoelectric signal to obtain its characteristic parameters;
[0074] The feature parameters are input into a trained random forest model to obtain the identification results of whether the temporomandibular joint vibration mechanical wave is abnormal.
[0075] Specifically, the sensor assembly is fitted to the temporomandibular joint (TMJ) of the human body to collect mechanical vibration waves, and then processes the collected TMJ mechanical vibration waves into analog piezoelectric signals. To achieve this technical objective, the sensor assembly further includes a stacked transmission component, a piezoelectric film, and a fixing component, wherein...
[0076] The transmission component is used to transmit the mechanical vibration waves of the temporomandibular joint to the piezoelectric film.
[0077] The piezoelectric film is used to detect the vibrational mechanical waves transmitted by the transmission component and trigger the generation of a simulated piezoelectric signal;
[0078] The piezoelectric film and the transmission component are stacked sequentially on the fixed component.
[0079] like Figure 2 As shown, the sensor assembly includes a transmission component, a piezoelectric film, and a fixing component stacked together, wherein...
[0080] The transmission component is used to transmit the mechanical vibration waves of the temporomandibular joint to the piezoelectric film.
[0081] The piezoelectric film is used to detect the vibrational mechanical waves transmitted by the transmission component and trigger the generation of a simulated piezoelectric signal;
[0082] The piezoelectric film and the transmission component are stacked sequentially on the fixed component.
[0083] Specifically, the sensor assembly collects mechanical vibration waves generated by temporomandibular joint movement, causing a piezoelectric film to vibrate. The piezoelectric film then generates a simulated piezoelectric signal by precisely sensing the vibration, thus achieving signal acquisition of the mechanical vibration waves. The transmission component specifically comprises:
[0084] Made of skin-friendly insulating material with good mechanical wave conduction capability;
[0085] The upper and lower surfaces are parallel, with the upper surface in contact with the human body and the lower surface in contact with the piezoelectric film.
[0086] The lower surface has the same shape as the piezoelectric film, but its cross-sectional area is slightly smaller than that of the piezoelectric film.
[0087] The edges are bonded and fixed to the piezoelectric film by center alignment with the piezoelectric film.
[0088] The fixing component specifically comprises: upper and lower surfaces parallel to each other, wherein the groove or protrusion is provided on one side of the upper surface, the cross-sectional shape of the groove or protrusion is the same as that of the piezoelectric film, the area is smaller than that of the piezoelectric film, the piezoelectric film covers the groove or protrusion, and the edge is bonded to the fixing component.
[0089] Specifically, the upper and lower surfaces of the transmission component are parallel. The upper surface, which contacts the human body, is made of a skin-friendly insulating material to reduce discomfort and prevent noise signals from being transmitted from the human body's bioelectricity to the piezoelectric film. The lower surface contacts the piezoelectric film, has the same shape as the film, and has a slightly smaller cross-sectional area. It is aligned with the piezoelectric film at the center, and its edges are bonded to the upper surface of the film. The upper and lower surfaces of the fixing component are parallel, preferably a cube made of a regular and rigid material. The upper and lower surfaces of the fixing component have the same area. One side of the upper surface has a groove or protrusion with the same shape as the piezoelectric film, and the area of the groove or protrusion is smaller than that of the film. The lower surface of the piezoelectric film covers the groove or protrusion, and its lower surface edge is bonded to the upper surface of the fixing component. This ensures a stable structure with the piezoelectric film and facilitates the formation of a good cavity effect, which is beneficial for the piezoelectric film to better generate simulated piezoelectric signals based on vibrational mechanical waves.
[0090] Figure 2 As shown, the sensor assembly further includes a metal contact for conforming to the skin and reducing signal interference; furthermore, the upper surface of the metal contact is at the same height as the upper surface of the transmission component and is configured with a smooth surface and edges; the metal contact is also connected to the signal processing component via the connecting component.
[0091] Specifically, the metal contact is fixed to the other side of the upper surface of the fixing component and does not contact the piezoelectric film. By setting the metal contact, the sensor assembly forms a structure with parallel upper and lower surfaces and similar areas, which allows for a more stable fit to the human body during application.
[0092] To further amplify and convert the analog piezoelectric signal to digital piezoelectric signal, the system specifically includes a signal processing component and a signal recognition component. The sensor component outputs the analog piezoelectric signal to the signal processing component, where it is amplified and converted to digital signal to generate a digital piezoelectric signal, which is then output to the signal recognition component. The signal recognition component first performs noise reduction processing on the input digital piezoelectric signal. Preferably, the noise reduction method is mean filtering, which involves summing the digital piezoelectric signals within a certain time window and then dividing by the time window length to obtain the mean, thus reducing the complexity of signal processing and eliminating noise interference. Optionally, the noise-reduced signal can also be obtained through median filtering, low-pass filtering, or weighted mean filtering, but these are not preferred methods in this embodiment.
[0093] Based on the denoised digital piezoelectric signal, feature parameters are extracted. In this embodiment, the root mean square value and waveform length sum of the denoised signal are extracted as time-domain feature parameters, and the median frequency and mean power frequency distribution are extracted as frequency-domain feature parameters to determine whether the temporomandibular joint vibration mechanical wave is abnormal.
[0094] The root mean square value refers to the square root of the average signal amplitude within a certain time window, and is specifically expressed by the following formula:
[0095] In the formula, RMS is the root mean square of the signal amplitude, x i The signal is denoised, and N is the set time window.
[0096] The waveform length summation refers to the summation of signal waveform lengths within a certain time window N, specifically expressed by the following formula:
[0097] The larger the sum of the waveform lengths, the more drastic the change in the signal waveform.
[0098] The median frequency is obtained by analyzing the power spectral density of the signal. Power spectral density is a measurement index used to describe the distribution of signal power at different frequencies, representing the average power per unit frequency band (e.g., per Hertz). It is calculated by performing a Fourier transform on the signal, specifically as follows:
[0099] PSD(f)=|X(f)| 2 In the formula, X(f) refers to the amplitude value obtained after the noise-reduced signal undergoes Fourier transform, and f represents the frequency;
[0100] The median frequency is specifically expressed by the following formula:
[0101] In the formula, MF represents the median frequency, and df represents the derivative of the frequency;
[0102] The mean value of the signal power frequency distribution is specifically expressed by the following formula:
[0103] In the formula, MPF is the mean power frequency distribution.
[0104] Extensive experimental verification shows that selecting the root mean square value of the denoised signal, the cumulative sum of waveform lengths, the median frequency, and the mean power frequency distribution as characteristic parameters to determine whether the mechanical vibration wave of the temporomandibular joint is abnormal can yield very ideal results.
[0105] Furthermore, the feature parameters are normalized, preferably using the minimax normalization method. The feature parameters are mapped to the [0, 1] interval using the following formula, which aims to reduce the complexity of data processing. The formula is specifically expressed as follows:
[0106] In the formula, x is the characteristic parameter, x new The normalized feature parameters, x min x is the minimum value among the feature parameter categories. max The maximum value among the aforementioned feature parameter categories. Optionally, Z-score normalization and Sigmoid normalization methods can also be used for data normalization in this embodiment, but they are not preferred after experimental verification.
[0107] The signal recognition component also includes a trained random forest model. By inputting the preprocessed digital piezoelectric signal into the trained random forest model, the system can output a recognition result indicating whether the corresponding temporomandibular joint vibration mechanical wave is abnormal.
[0108] Furthermore, in this embodiment, the trained random forest model is obtained through the following method, which specifically includes:
[0109] Feature parameters are extracted from the original signal samples used for model training, and the original signal samples are labeled to construct training and test sets;
[0110] The number of decision trees and the minimum number of leaf samples are used as model hyperparameters, and the range of values for the decision trees and the minimum leaf samples, as well as the maximum number of search times, are set respectively.
[0111] SA1: Randomly select one value from the decision tree and the minimum leaf sample value range, and generate a random forest model to be trained based on the random value;
[0112] SA2: The trained random forest model is generated by training the random forest model to be trained using the training set.
[0113] SA3: Calculate and save the fitness value of the trained random forest model;
[0114] Repeat SA1-SA3 until the maximum number of search rounds is reached;
[0115] The fitness values of each trained random forest model are calculated based on the training and test sets, and the random forest model with the best fitness value is selected as the trained random forest model.
[0116] The structure and performance of a random forest model are closely related to the selected hyperparameters. In this embodiment, the number of decision trees and the minimum number of leaf samples are selected as the hyperparameters of the random forest model. In order to obtain the optimal hyperparameter values of the random forest model, i.e. the optimal number of decision trees and the minimum number of leaf samples, this embodiment adopts a random search method. Specifically, based on experience, a decision tree value range and a minimum leaf sample value range are preset. Then, a value is randomly selected from the preset decision tree value range and the preset minimum leaf sample value range each time to verify the performance of the constructed random forest model. The random values are not repeated until the maximum number of searches is reached. Then, the random forest model with the best performance is selected as the trained random forest model.
[0117] First, the signal recognition component performs the same noise reduction, feature parameter extraction, and normalization preprocessing on the original signal samples used for model training as on the input digital piezoelectric signal. Then, the original signal samples are labeled. Specifically, each original signal sample is labeled with a tag indicating whether it is abnormal. For example, a normal signal sample is labeled as 0, and an abnormal signal sample is labeled as 1. The labeling method is consistent with the final recognition result format. To ensure the correctness of the labeling of the original signal samples, expert manual labeling is preferred. The labeled original signal samples are used to construct training and test sets. Specifically, the construction of training and test sets involves randomly sampling the original signal samples the same number of times as the number of decision trees. Each time, 70% of the original signal samples are sampled to create a training set, and the remaining 30% is used to create a corresponding test set, resulting in multiple training and test sets. The number of decision trees selected here is one of the alternative values for decision trees. The purpose of randomly sampling 70% of the original signal samples with replacement to create the training set is to enhance the randomness of the samples in the constructed training set, which is beneficial to improving the robustness of the random forest model after training. For each decision tree candidate value and minimum leaf sample number combination, a corresponding training set and test set must be established, and the number of the training set and test set is the same as the number of the corresponding decision tree candidate values.
[0118] Next, the training set is used to train the random forest model to be trained, and each trained random forest model is generated by the following method, which specifically includes:
[0119] In the training set, feature parameters of 75% of the categories of each sample are randomly selected and input into the root node of a decision tree in the random forest model to be trained.
[0120] The sample entropy before splitting is calculated based on the sample feature parameters of the input training set, and a decision tree is established through the following steps, which specifically include:
[0121] SS1: Sort the feature parameters of each category for each sample in descending order;
[0122] SS2: Calculate the split value sequentially based on the comparison results of adjacent feature parameters in each queue;
[0123] SS3: Splits from the current root node into two leaf nodes;
[0124] SS4: Based on the comparison results of each split value and the corresponding category feature parameter, each sample is divided into two groups. The samples whose corresponding category feature parameter is less than the split value are divided into one group, and the remaining samples are divided into another group. The sample entropy after grouping is calculated for each group.
[0125] SS5: Save the difference between the entropy before splitting and the entropy after grouping of each sample feature parameter as the gain, associate the corresponding category feature with the largest gain with the current root node, and use the splitting value corresponding to the feature parameter with the largest gain as the splitting value of the current root node;
[0126] SS6: Based on the split value grouping result of the current root node, assign the two groups of samples to a leaf node of the current root node respectively, and remove the feature parameter corresponding to the feature category with the largest gain from each sample;
[0127] SS1-SS6 are executed iteratively on the samples of each leaf node. In each iteration, the feature parameters of the class features of the associated nodes are ignored until all class features are associated with a certain node, at which point the decision tree is completed.
[0128] Once all decision trees in the random forest model to be trained are built, the trained random forest model is obtained.
[0129] Specifically, for each decision tree candidate value and the minimum number of leaf samples candidate value to build a random forest model, the feature parameters of 75% of the corresponding training sets are randomly selected and input into the random forest model. In this embodiment, each original signal sample includes 4 types of feature parameters, and each training set randomly selects 3 types of feature parameters. The purpose is to limit the depth of the decision tree during the model training process while ensuring the difference in the feature parameters used by each decision tree, and to prevent overfitting.
[0130] In each random forest model, each training set randomly corresponds to the root node of a decision tree. Based on the characteristics of the random forest model, each node in the decision tree binary splits into two leaf nodes, and each node records and is associated with a feature parameter. In this embodiment, the entropy of each sample before and after the split is calculated using the following formula, and then the association between the node and the feature parameter is determined by calculating the gain of the entropy. The formula is specifically expressed as follows:
[0131] In the formula,
[0132] x i Represents characteristic parameters,
[0133] H(X) represents entropy.
[0134] p(x i ) represents the probability of a feature parameter value occurring, specifically x i The ratio of the number of times the value appears in the current node to the total number of samples in the current node.
[0135] The gain of entropy refers to the difference between the entropy after splitting the same sample feature parameters and the entropy before splitting.
[0136] Calculate and save the entropy of all sample parameters input to the root node;
[0137] Then, the sample parameters are sorted in descending order for each category of feature parameters. Specifically, in this embodiment, by sorting the feature parameters in descending order for each category, four kinds of sample descending order results are obtained. The purpose is to establish a sort for each category of feature parameters to facilitate pairwise verification and further enhance randomness.
[0138] The splitting value is calculated sequentially based on the comparison results of adjacent feature parameters in each queue, specifically including:
[0139] Each queue checks whether adjacent feature parameters are the same from the beginning. If they are the same, the comparison continues until adjacent feature parameters are different. The sum of 2 / 3 of the previous feature parameter and 1 / 3 of the next feature parameter is rounded to the nearest integer and saved as the split value.
[0140] Otherwise, the current feature parameter and the mean of adjacent feature parameters are rounded to the nearest integer and saved as the split value until each queue has completed the judgment of all adjacent feature parameters.
[0141] For example, if one of the characteristic parameters of the descending sequence is [7,4,4,3,3,2,1], then the calculated split value is [6,4,3,2].
[0142] After processing, we can obtain the split values based on the weighted average of the feature parameters of each pair of adjacent samples under the four types of sample descending order.
[0143] The next step is to find the optimal splitting value from among the various splitting values, specifically including:
[0144] Based on the comparison between each split value and the corresponding category feature parameter of each sample, the samples are divided into two groups. Specifically, only the category feature parameter of each sample with the calculated split value is compared with the corresponding split value. Samples with a category feature parameter smaller than the split value are divided into one group, and the remaining samples are divided into the other group. The entropy of each group is calculated. From the entropy calculation formula, it can be seen that when each sample is divided into two groups based on a feature, the probability p(x) of the feature parameter value of each sample before and after being divided into two groups is... i The entropy after grouping will change, resulting in a difference between the entropy before and after grouping. Therefore, the entropy of each sample before and after grouping based on the split values is calculated, and then each gain is calculated separately. The split value corresponding to the feature parameter with the largest gain is used as the split value of the current root node, thus completing the association between the current root node and a certain feature. At this time, the decision tree completes a split based on the split value of the current root node. The grouping results of each sample based on the split value of the current root node are assigned to a leaf node of the root node, and the feature parameter corresponding to the split value of the root node is removed from each sample. Then, the split value of each leaf node is recalculated, and the entropy before and after grouping of the corresponding samples is calculated and the maximum gain is found. The associated feature type of each leaf node is found, and then the split is performed downwards. This process is repeated until all feature types are associated with a certain node, and then the decision tree is built.
[0145] Once all decision trees are built, the random forest model training is complete, and the trained random forest models are obtained.
[0146] Generally, the performance of a random forest model can be evaluated by calculating its fitness value. In this embodiment, the random forest model with the smallest error after training is considered the optimal one.
[0147] Specifically, each of the trained random forest models has been trained on its respective input training set, learning and recording the features of the samples in the training set. Due to the randomness of the training set samples, as well as the randomness of the decision tree generation process and training of each trained random forest model, the performance of each trained random forest model is different. Therefore, the corresponding training set and test set are input again into each trained random forest model to output the recognition result. Specifically, in this embodiment, the random forest model outputs a qualitative recognition result on whether the input temporomandibular joint vibration mechanical wave signal is abnormal. The recognition result is then compared with the labeled value of the corresponding sample. If the recognition result is different from the labeled value of the corresponding sample, it is recorded as an error. The number of times the training set error and test set error occur for each trained random forest model are summarized and counted. The ratios are then divided by the number of samples in the training set and test set, and summed to obtain the error value for each trained random forest model. The random forest model with the smallest error value is the trained random forest model.
[0148] This embodiment discloses a method for acquiring and identifying temporomandibular joint (TMJ) vibration mechanical waves. The method includes acquiring vibration mechanical wave signals using a sensor assembly to generate analog piezoelectric signals. These analog piezoelectric signals are amplified and converted from analog to digital piezoelectric signals. Feature extraction is performed on the digital piezoelectric signals to obtain feature parameters. These feature parameters are then input into a trained random forest model to obtain an identification result indicating whether the TMJ vibration mechanical waves are abnormal. This embodiment extracts four unique feature parameters after denoising the digital piezoelectric signals by mean, and selects the number of decision trees and the minimum number of leaf samples as hyperparameters. A random forest model is constructed using random search, and the random forest model with the best fitness value is found for TMJ vibration mechanical wave identification, achieving the best identification effect. Furthermore, this embodiment uses a unique sensor assembly structure to acquire and generate TMJ vibration mechanical wave signals. Compared to existing technologies, the random forest model used in this embodiment is specifically designed for identifying whether TMJ vibration mechanical waves are abnormal. Through optimized algorithm training, it solves the problem of unsatisfactory TMJ vibration mechanical wave identification results in existing technologies. The unique sensor assembly structure also solves the problem of poor reliability in acquiring TMJ mechanical vibration signals.
[0149] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint, characterized in that, The method specifically includes: Sensor components are used to collect vibrational mechanical waves of the temporomandibular joint and generate simulated piezoelectric signals; A digital piezoelectric signal is obtained by amplifying and converting the analog piezoelectric signal to a digital one. Feature extraction is performed on the digital piezoelectric signal to obtain its characteristic parameters; The feature parameters are input into a trained random forest model to obtain the identification results of whether the temporomandibular joint vibration mechanical wave is abnormal.
2. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 1, characterized in that, The characteristic parameters of the digital piezoelectric signal specifically include the root mean square of the signal amplitude, the cumulative value of the signal waveform length, the median frequency, and the mean of the power frequency distribution, wherein... The median frequency is calculated using the following formula, which is specifically expressed as follows: In the formula, MF is the median frequency, X(f) is the frequency domain signal obtained after the Fourier transform of the signal, and df is the differential of the frequency. The mean value of the power frequency distribution is calculated using the following formula, which is specifically expressed as follows: In the formula, MPF is the mean power frequency distribution; PSD(f) = |X(f)| 2 .
3. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 2, characterized in that, The trained random forest model is obtained through the following method, which specifically includes: Feature parameters are extracted from the original signal samples used for model training, and the original signal samples are labeled to construct training and test sets; The number of decision trees and the minimum number of leaf samples are used as model hyperparameters, and the range of values for the decision trees and the minimum leaf samples, as well as the maximum number of search times, are set respectively. SA1: Randomly select one value from the decision tree and the minimum leaf sample value range, and generate a random forest model to be trained based on the random value; SA2: The trained random forest model is generated by training the random forest model to be trained using the training set. SA3: Calculate and save the fitness value of the trained random forest model; Repeat SA1-SA3 until the maximum number of search rounds is reached; The fitness values of each trained random forest model are calculated based on the training and test sets, and the random forest model with the best fitness value is selected as the trained random forest model.
4. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 3, characterized in that, The trained random forest models are generated by training the random forest model to be trained using the training set through the following method, the method specifically including: In the training set, feature parameters of 75% of the categories of each sample are randomly selected and input into the root node of a decision tree in the random forest model to be trained. The sample entropy before splitting is calculated based on the sample feature parameters of the input training set, and a decision tree is established through the following steps, which specifically include: SS1: Sort the feature parameters of each category for each sample in descending order; SS2: Calculate the split value sequentially based on the comparison results of adjacent feature parameters in each queue; SS3: Splits from the current root node into two leaf nodes; SS4: Based on the comparison results of each split value and the corresponding category feature parameter, each sample is divided into two groups. The samples whose corresponding category feature parameter is less than the split value are divided into one group, and the remaining samples are divided into another group. The sample entropy after grouping is calculated for each group. SS5: Save the difference between the entropy before splitting and the entropy after grouping of each sample feature parameter as the gain, associate the corresponding category feature with the largest gain with the current root node, and use the splitting value corresponding to the feature parameter with the largest gain as the splitting value of the current root node; SS6: Based on the split value grouping result of the current root node, assign the two groups of samples to a leaf node of the current root node respectively, and remove the feature parameter corresponding to the feature category with the largest gain from each sample; SS1-SS6 are executed iteratively on the samples of each leaf node. In each iteration, the feature parameters of the class features of the associated nodes are ignored until all class features are associated with a certain node, at which point the decision tree is completed. Once all decision trees in the random forest model to be trained are built, the trained random forest model is obtained.
5. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 4, characterized in that, The splitting value is calculated sequentially based on the comparison results of adjacent feature parameters in each queue, specifically including: Each queue checks whether adjacent feature parameters are the same from the beginning. If they are the same, the comparison continues until adjacent feature parameters are different. The sum of 2 / 3 of the previous feature parameter and 1 / 3 of the next feature parameter is rounded to the nearest integer and saved as the split value. Otherwise, the current feature parameter and the mean of adjacent feature parameters are rounded to the nearest integer and saved as the split value until each queue has completed the judgment of all adjacent feature parameters.
6. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 5, characterized in that, The process of calculating the fitness value of each trained random forest model based on the training and test sets, and selecting the random forest model with the best fitness value as the trained random forest model, specifically includes: The training set and the test set are respectively input into the trained random forest model, and the recognition results are output respectively. Obtain the error between the recognition result and the labeled value of each sample, and obtain the error of the training set based on the error of all samples in the training set, and obtain the error of the test set based on the error of all samples in the test set; use the sum of the errors of the training set and the test set as the recognition error value of the random forest model. The ratio of the recognition error value of each trained random forest model to the total number of samples in the input training and test sets is calculated as the fitness value of each trained random forest model. The trained random forest model with the smallest fitness value is used as the trained random forest model.
7. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 1, characterized in that, The sensor assembly includes a transmission component, a piezoelectric film, and a fixing component stacked together. The transmission component is used to transmit the mechanical vibration waves of the temporomandibular joint to the piezoelectric film. The piezoelectric film is used to detect the vibrational mechanical waves transmitted by the transmission component and trigger the generation of a simulated piezoelectric signal; The piezoelectric film and the transmission component are stacked sequentially on the fixed component.
8. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 7, characterized in that, The transmission component is specifically: Made of skin-friendly insulating material with good mechanical wave conduction capability; The upper and lower surfaces are parallel, with the upper surface in contact with the human body and the lower surface in contact with the piezoelectric film. The lower surface has the same shape as the piezoelectric film, but its cross-sectional area is slightly smaller than that of the piezoelectric film. The edges are bonded and fixed to the piezoelectric film by center alignment with the piezoelectric film.
9. The method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 8, characterized in that, The sensor assembly also includes a metal contact, the upper surface of which is at the same height as the upper surface of the transmission component, and is configured with a smooth surface and edge structure.
10. A method for acquiring and identifying vibrational mechanical waves of the temporomandibular joint according to claim 9, characterized in that, The fixing component specifically comprises: upper and lower surfaces parallel to each other, wherein a groove or protrusion is provided on one side of the upper surface, the cross-sectional shape of the groove or protrusion is the same as that of the piezoelectric film, and the area is smaller than that of the piezoelectric film, the piezoelectric film covers the groove or protrusion, and the edge is bonded to the fixing component; the metal contact is fixedly provided on the other side of the upper surface.