Knee joint data classification method based on knowledge graph-improved Bayesian network

Abstract

The present invention provides a method for classifying knee joint data using a Bayesian network improved based on a knowledge graph, comprising the following steps: S1, obtaining a knee joint dataset; S2, performing feature extraction on the knee joint dataset; S3, dividing the knee joint dataset into a training set, a validation set, and a test set; S4, establishing a Bayesian network; S5, modifying the Bayesian network based on expert knowledge; S6, modifying the Bayesian network based on the common causal strength of signal features; S7, modifying the Bayesian network based on the common causal strength of related vibration sensors; S8, performing parameter learning to obtain quantitative relationships between nodes; S9, performing local search optimization on the order of signal feature nodes using the validation set; S10, testing the improved Bayesian network structure classification model; and S11, classifying new data. By repeatedly modifying the Bayesian network, the present invention improves the robustness of prediction under small sample sizes.

Description

Technical Field

[0001] The present invention relates to the field of data mining in computers, and specifically provides a method for classifying knee joint data based on a Bayesian network improved by a knowledge graph. Background Art

[0002] Sensors and data processing are two key areas in implementing real-time monitoring systems using wearable devices. Despite progress in sensor development to date, limitations remain in the accuracy of collected data and its precise analysis. This necessitates AI-powered wearable sensors to extract target signals for accurate data analysis.

[0003] Current AI approaches for processing wearable device data can be categorized as classical machine learning and deep learning. Classical machine learning approaches, by combining feature extraction based on prior knowledge, have been widely used in the detection and processing of heart, blood sugar, and mental health data. AI approaches with deep learning algorithms, such as deep neural networks, convolutional neural networks, and spiking neural networks, aim to extract features from raw data and perform classification within a single framework. However, existing machine learning algorithms require a large amount of data as a baseline, and the accuracy of the model depends largely on the quality of the data. Furthermore, existing AI methods, such as deep learning, achieve efficient classification by constructing a massive parameter space consisting of hundreds of layers and millions of parameters. This complex hierarchical structure, combined with the sheer volume of data, obscures the physical meaning of the parameters and the causal relationships between them, essentially reducing the ability to explain their internal workings and decision-making mechanisms.

[0004] Research and practice in recent years have shown that causal relationships reveal the true impact of one variable on another, and models based on causal relationships are inherently more explanatory and robust. Bayesian networks can effectively discover and apply causal relationships between variables and are widely used in fields such as bioinformatics computing, medicine, information retrieval, and image processing. Traditional Bayesian network learning mainly includes two construction methods: expert knowledge-based and data-based learning. However, controlling the Bayesian network structure based on expert knowledge is difficult to handle causal judgments under multiple feature dimensions. Data-driven causal judgments require a huge amount of data support, otherwise it is impossible to obtain an accurate network structure. Therefore, in view of the shortcomings of existing technologies, it is necessary to propose a technical solution to solve the problems existing in existing technologies. Summary of the Invention

[0005] In order to overcome the defects of the existing technology, the present invention provides a method for classifying knee joint data based on an improved Bayesian network based on a knowledge graph, combining data-driven and expert-controlled methods. On the other hand, a common causal strength judgment method based on information gain ratio and coefficient of variation is proposed, and the Bayesian network structure is optimized to improve the robustness of prediction under small sample sizes, so as to obtain more accurate evaluation results.

[0006] The present invention discloses a method for classifying knee joint data based on a Bayesian network improved by a knowledge graph. The specific technical solution includes the following steps:

[0007] S1, obtain the labeled knee joint dataset;

[0008] Select knee joint data from the knee joint database and label the selected knee joint data to form a knee joint dataset;

[0009] S2, feature extraction of knee joint dataset;

[0010] The features include signal features and functional features, wherein the signal features are obtained based on the vibration sensor waveform data, and the functional features are obtained from the knee joint function evaluation table data;

[0011] S3, divide the knee joint dataset into training set, validation set and test set;

[0012] S4, building a Bayesian network;

[0013] The features, fields, and labels of the indicator data in step S2 are selected as nodes of the Bayesian network, the node order for Bayesian network training is given, and the initial Bayesian network structure is obtained through learning the training set, wherein the search strategy is the hill climbing algorithm, and the scoring function is the K2 scoring function;

[0014] S5, modifying the Bayesian network of step S4 based on expert knowledge;

[0015] S6, judging the common causal strength based on the signal features and modifying the Bayesian network in step S5;

[0016] The specific method for judging the strength of common causality is as follows:

[0017]

[0018] I(A,B)=H(B)-H(B|A) (4)

[0019]

[0020] Among them, H(A) is the information entropy, i is the value of feature A, n is the value range of A, p iis the probability of the i-th value, H(B|A) is the conditional entropy of B with respect to A, p(B|A) represents the probability of B taking the value under condition A, I(A,B) is the information gain, IGR is the information gain ratio, C v is the coefficient of variation, where σ is the standard deviation of the information gain ratio and μ is the mean;

[0021] After obtaining the information gain ratio of the two signal features of each channel of the vibration sensor waveform data according to formulas (2)-(5), the standard deviation and mean of the information gain ratio of the same two signals in all channels of the vibration sensor waveform data are calculated. Finally, the coefficient of variation between the two signals is obtained according to formula (6), and the coefficient of variation is used as an indicator of the common causal strength. The coefficients of variation are sorted and selected from small to large according to a preset number, and the selected causal relationship is added to the Bayesian network obtained in step S5.

[0022] S7, selecting the signal indicator with the strongest causal relationship based on the common causal strength for the relevant vibration sensors, and modifying the Bayesian network obtained in step S6;

[0023] The signal index with the strongest causal relationship is selected by the common causal strength, and the signal index that characterizes the obvious common causal relationship between known related flexible piezoelectric sensors is determined. The specific method is to select two known related flexible piezoelectric sensors, calculate the information gain ratio of the same signal features of the vibration sensor waveform data of the corresponding channels of the two related flexible piezoelectric sensors, and obtain the coefficient of variation C of the corresponding signal features based on the information gain ratio of the same signal features of all related flexible piezoelectric sensors. v ,The signal feature with the smallest coefficient of variation is selected as the causal relationship and added to the corresponding signal features of the known related flexible piezoelectric sensors;

[0024] S8, performing parameter learning on the Bayesian network obtained in step S7 through Laplace smoothing and maximum likelihood estimation to obtain quantitative relationships between nodes;

[0025] S9, performing local search optimization on the signal feature node sequence given in step S4 using the validation set, further optimizing the Bayesian network classification model structure, and obtaining an improved Bayesian network classification model;

[0026] S10, using the test set to test the accuracy of the improved Bayesian network classification model;

[0027] S11, classification is performed using the improved Bayesian network classification model.

[0028] Preferably, each knee joint data in S1 consists of three types of data: vibration sensor waveform data, index data, and knee joint function evaluation table data; the labels are divided into three types: first state, second state and third state.

[0029] Preferably, the vibration sensor waveform data includes data from six channels, and the vibration sensor waveform data of the six channels are respectively derived from six flexible piezoelectric sensors on the same wearable knee pad. The six flexible piezoelectric sensors are evenly distributed on the wearable knee pad for collecting vibration data and saving them in the knee joint database; the knee joint function evaluation table data is obtained based on the processed Lysholm table, and at least includes height, weight, and graded lameness, support, locking, joint instability, pain, swelling, difficulty climbing stairs, and difficulty squatting data; the index data is data calculated based on the vibration sensor waveform data or the knee joint function evaluation table data.

[0030] Preferably, the S2 specifically includes:

[0031] After preprocessing the vibration sensor waveform data, five signal features are extracted from each channel of the vibration sensor waveform data: period, energy, kurtosis, correlation and number of inflection points;

[0032] Ten functional characteristics were extracted from the knee function evaluation form data: height, weight, limp, support, locking, joint instability, pain, swelling, difficulty climbing stairs, and difficulty squatting.

[0033] Preferably, the causal relationships determined by the expert knowledge in S5 include: BMI and label, lysholm table score and label, each functional feature and lysholm table score, height and BMI, weight and BMI, swelling and joint instability, pain and difficulty squatting, pain and difficulty climbing stairs, pain and swelling, lameness and support, height and weight. The above causal relationships are all the former as the cause and the latter as the effect. If the Bayesian network fails to train these causal relationships using data-driven training in step S4, the missing causal relationships need to be supplemented based on expert knowledge.

[0034] Preferably, the S11 uses an improved Bayesian network classification model for classification, specifically:

[0035] New knee joint data is input, and the new knee joint data only includes the vibration sensor waveform data and the height and weight in the knee joint function evaluation table data; signal features are obtained from the vibration sensor waveform data; the signal features and height and weight are used as the input of the final Bayesian network structure classification model, and the functional feature values ​​in the knee joint function evaluation table are first derived. Then, the functional feature values ​​in the joint function evaluation table and the height and weight are again used as the input of the improved Bayesian network classification model, and finally the value of the label node is output to classify the new knee joint data.

[0036] Preferably, the local search optimization in S9 adopts simulated annealing or gradient descent algorithm.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. This paper proposes a method for determining the strength of common causal relationships, identifying common causal relationships arising from the physical meaning of indicators. The strength of the causal relationship is measured using the information gain ratio. Furthermore, the coefficient of variation is used to measure the uniformity of the causal strength between two types of indicators, describing the common patterns of causal strength between the two types of data.

[0039] 2. The present invention combines data-driven and expert-controlled methods to construct a Bayesian network. First, the Bayesian network is constructed through data-driven methods. Then, the Bayesian network is corrected for the first time through expert knowledge. Secondly, the Bayesian network is corrected for the second time by using the common causal strength discrimination method for signal features. Then, the Bayesian network is corrected for the third time by using the common causal strength discrimination method between vibration sensors. Finally, the Bayesian network model structure is optimized by performing local search optimization on the order of signal feature nodes. By correcting the Bayesian network multiple times, the robustness of the Bayesian network prediction under small sample sizes is improved.

[0040] 3. When the present invention trains the Bayesian network, the knee joint data includes vibration sensor waveform data, index data and knee joint function evaluation table data. The vibration sensor waveform data as objective data and the knee joint function evaluation table data as subjective data are combined to improve the accuracy of knee joint data classification.

[0041] 4. When using the improved Bayesian network classification model, the present invention does not require knee joint function evaluation table data. Instead, it directly uses the vibration sensor waveform data and height and weight to obtain the knee joint function evaluation table data, and then uses the knee joint function evaluation table data and height and weight to obtain the label value. Therefore, the present invention only needs to input the objective vibration sensor waveform data when performing a new classification, but uses the subjective knee joint function evaluation table data when obtaining the conclusion, thereby effectively combining and utilizing the subjective data and the objective data in the improved Bayesian network classification model.

[0042] 5. The present invention proposes to divide the node sequence into functional feature node sequence and signal feature node sequence, and in subsequent steps only perform local search optimization on the signal feature node sequence, further optimize the Bayesian network model structure, and improve the network optimization speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the method for classifying knee joint data based on the improved Bayesian network of the knowledge graph of the present invention;

[0044] Figure 2 is the Bayesian network structure obtained using K2;

[0045] Figure 3 This is the improved Bayesian network model structure obtained by the present invention. DETAILED DESCRIPTION

[0046] To better understand the technical solutions of the present invention, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0047] The present invention is a method for classifying knee joint data based on a Bayesian network improved by knowledge graph, such as Figure 1 The specific implementation steps are as follows:

[0048] S1, obtain the labeled knee joint dataset.

[0049] Knee joint data are selected from the knee joint database, and the selected knee joint data are labeled to form a knee joint dataset.

[0050] Each knee joint data is made up of the following 3 kinds of data: vibration sensor waveform data, index data and knee joint function evaluation table data.In the present embodiment, vibration sensor waveform data includes the data of six channels, and the vibration sensor waveform data of six channels come from six flexible piezoelectric sensors on same wearable kneepad respectively, and six flexible piezoelectric sensors are evenly distributed on the wearable kneepad for collecting vibration data and are saved in the knee joint database, and vibration data is preferably data when flexing the knee joint with a fixed frequency. Knee joint function evaluation table data are obtained according to the lysholm table after processing, including height, weight, and the data such as lameness, support, interlocking, joint instability, pain, swelling, difficulty climbing stairs and difficulty squatting after grading, in the knee joint function evaluation table, the features of descriptive language adopted in the lysholm table are all graded, such as for feature lameness, needing to be grade 1 when using crutches, occasionally lame behavior grade 2, not lame behavior grade 3. The index data is data calculated based on the vibration sensor waveform data or the knee joint function evaluation table data. For example, in this embodiment, the index data includes the body mass index (BMI) and the lysholm table score.

[0051] In this example, 150 knee joint data sets were selected and labeled. The labels were categorized into three types: first state, second state, and third state. 110 of these data sets were in the first state, 30 were in the second state, and 10 were in the third state. The first through third states correspond to normal, mild, and severe, respectively. These 150 knee joint data sets constitute the knee joint dataset.

[0052] S2, feature extraction for knee joint dataset.

[0053] The features include signal features and functional features, wherein the signal features are obtained based on the vibration sensor waveform data, and the functional features are obtained from the knee joint function evaluation table data.

[0054] The vibration sensor waveform data was preprocessed using filtering, normalization, and kmeans discretization to eliminate the effects of noise and sensor performance variations. Five signal features were extracted from each channel of the preprocessed vibration sensor waveform data: period, energy, kurtosis, correlation, and number of inflection points. The period refers to the length of a single signal cycle; energy refers to the sum of squared amplitudes within a single signal cycle; kurtosis refers to the fourth-order moment of the normalized vibration amplitude, reflecting the signal's impact; correlation refers to the direct correlation between signals in adjacent cycles; and the number of inflection points refers to the number of inflection points within a single signal cycle. The vibration sensor waveform data consists of six channels, and the five signal features described above were extracted for each of the six channels, yielding a total of 30 signal features.

[0055] Ten functional characteristics were extracted from the knee function evaluation form data: height, weight, limp, support, locking, joint instability, pain, swelling, difficulty climbing stairs, and difficulty squatting.

[0056] It should be pointed out here that the signal characteristics and functional characteristics in this embodiment are schematic, and different signal characteristics and characteristic functions can be selected according to actual needs.

[0057] S3, the knee joint dataset is divided into training set, validation set and test set.

[0058] The knee joint dataset was divided into training, validation, and test sets in a ratio of 8:1:1. For example, if there were 150 data points, the training, validation, and test sets would be 120, 15, and 15, respectively. The three types of knee joint data were distributed proportionally across the training, validation, and test sets.

[0059] S4, establish a Bayesian network.

[0060] Determine the nodes of the Bayesian network and give the node order in advance, and obtain the initial Bayesian network structure through the training set.

[0061] In this embodiment, the features, fields of indicator data, and labels in step S2 are selected as nodes of the Bayesian network. Specifically, 30 signal features, 10 functional features, BMI, lysholm table scores, and labels are used as nodes, for a total of 43 nodes. The labels have three values, representing three types of labels.

[0062] Before training, a key parameter in the Bayesian network training process is pre-determined: node order. This node order is derived from expert knowledge and the physical relationship between features. It includes both functional feature node order and signal feature node order. The functional feature node order is based on the order given by the functional feature nodes and cannot be changed. The signal feature node order is based on the order given by the signal feature nodes and is only an approximate order that will be adjusted during subsequent training.

[0063] After inputting the node sequence, the initial Bayesian network structure is obtained through training set learning, such as Figure 2 As shown, the search strategy selects the hill climbing algorithm and the scoring function selects the K2 scoring function.

[0064] S5, Bayesian network correction based on expert knowledge.

[0065] The causal relationships not trained using data-driven training in step S4 are supplemented based on expert knowledge. Causal relationships determined by expert knowledge include: BMI and label, lysholm scale score and label, 10 functional characteristics and lysholm scale score, height and BMI, weight and BMI, swelling and joint instability, pain and difficulty squatting, pain and difficulty climbing stairs, pain and swelling, limp and support, and height and weight. The above causal relationships are all caused by the former and the latter by the latter. If the Bayesian network does not train these causal relationships using data-driven training in step S4, the missing causal relationships need to be supplemented based on expert knowledge to achieve the first correction of the Bayesian network.

[0066] S6, determines the common causal strength based on the signal characteristics and modifies the Bayesian network.

[0067] The specific method for determining the strength of common causality is to use the information gain ratio as an indicator of common causal strength and measure the strength of the causal relationship. Furthermore, the coefficient of variation is used to measure the uniformity of the causal strength between the two types of indicators, describing the common laws of causal strength between the two types of data. The smaller the coefficient of variation, the stronger the common causal relationship between the two groups of indicators. The specific method is as follows:

[0068]

[0069] I(A,B)=H(B)-H(B|A) (4)

[0070]

[0071] Among them, formula (2) is the calculation formula of information entropy H(A), where i is the value of feature A, n is the value range of A (number of types), and p iis the probability of the i-th value. Formula (3) is the calculation formula for the conditional entropy H(B|A) of B with respect to A, and p(B|A) represents the probability of B taking the value under condition A. Formula (4) is the calculation formula for information gain I(A,B), and formula (5) is the calculation formula for information gain ratio IGR. Formula (6) is the coefficient of variation C v The calculation formula is, where σ is the standard deviation of the information gain ratio and μ is the mean.

[0072] After obtaining the information gain ratio of the two signal features of each channel of the vibration sensor waveform data according to formulas (2)-(5), the standard deviation and mean of the information gain ratio of the same two signals in all channels of the vibration sensor waveform data are calculated. Finally, the coefficient of variation between the two signals is obtained according to formula (6), and the coefficient of variation is used as the common causal strength index. The coefficient of variation is sorted and selected from small to large according to the preset value. The selected causal relationship is added to the Bayesian network obtained in step S5 to achieve further correction of the Bayesian network. This is the second correction of the Bayesian network proposed by the present invention. In this embodiment, the common causal strength index between the five characteristic values ​​is shown in Table 1. The preset value is to select 1. Therefore, it is determined that there is an obvious common causal relationship between kurtosis and energy. According to the common causal strength index, it can be seen that kurtosis and energy are mutually causal. Therefore, the causal relationship between kurtosis and energy in each channel is added to the Bayesian network.

[0073] Table 1

[0074]

[0075] S7, in the case where the vibration sensors are known to be related, the signal indicator with the strongest causal relationship is selected by the common causal strength, and the Bayesian network obtained in step S6 is modified.

[0076] According to the position of the flexible piezoelectric sensors, it can be known that there is correlation between some flexible piezoelectric sensors and they will be affected at the same time, such as flexible piezoelectric sensors in symmetrical or adjacent positions. Therefore, the signal indicator with the strongest causal relationship can be selected through the common causal strength to determine the signal indicator that characterizes the obvious common causal relationship between known related flexible piezoelectric sensors. The specific method is to select two known related flexible piezoelectric sensors, calculate the information gain ratio of the same signal features of the vibration sensor waveform data of the corresponding channels of the two related flexible piezoelectric sensors, and obtain the coefficient of variation C of the corresponding signal features based on the information gain ratio of the same signal features of all related flexible piezoelectric sensors. v ,The signal feature with the smallest coefficient of variation is selected as the causal relationship and added between the corresponding signal features of the known related flexible piezoelectric sensors.

[0077] For example, if it is known that there is correlation between the flexible piezoelectric sensors 1-6, 2-5, and 3-4, the information gain ratios of the same signal features are calculated between the flexible piezoelectric sensors 1-6, 2-5, and 3-4, respectively, including the period of the 1st channel vibration sensor waveform data and the period of the 6th channel vibration sensor waveform data, the period of the 2nd channel vibration sensor waveform data and the period of the 5th channel vibration sensor waveform data, the period of the 3rd channel vibration sensor waveform data and the period of the 4th channel vibration sensor waveform data, ..., the number of inflection points of the 1st channel vibration sensor waveform data and the number of inflection points of the 6th channel vibration sensor waveform data, the number of inflection points of the 2nd channel vibration sensor waveform data and the number of inflection points of the 5th channel vibration sensor waveform data, the number of inflection points of the 3rd channel vibration sensor waveform data and the number of inflection points of the 4th channel vibration sensor waveform data; because each group of related flexible piezoelectric sensors obtains the information gain ratios of 5 signal features, there are a total of 15 information gain ratios of signal features; and then the coefficient of variation C of the same signal feature is calculated. v , and the coefficients of variation of the five signal features are obtained respectively, as shown in Table 2. Among them, the coefficient of variation of the period is the smallest, indicating that the period can represent the correlation between these flexible piezoelectric sensors. The period is selected as the signal indicator as the causal relationship between the known flexible piezoelectric sensors 1-6, 2-5, and 3-4, and the causal relationship of channel 1 period-channel 6 period, channel 2 period-channel 5 period, and channel 3 period-channel 4 period is added to the network to realize the third correction of the Bayesian network proposed in this invention.

[0078] Table 2

[0079]

[0080] S8, performing parameter learning on the Bayesian network obtained in step S7 through Laplace smoothing and maximum likelihood estimation to obtain quantitative relationships between nodes.

[0081] S9, perform local search optimization on the signal feature node sequence given in step S4 through the validation set, further optimize the model structure, and obtain an improved Bayesian network classification model, such as Figure 3 shown.

[0082] Local search optimization can use algorithms such as simulated annealing and gradient descent.

[0083] S10, using the test set to test the accuracy of the improved Bayesian network classification model obtained in step S9.

[0084] The comparison of the accuracy of the improved Bayesian network classification model obtained in this embodiment with other machine learning methods is shown in Table 3, where other machine learning methods include: SVM, KNN, decision tree, and Bayesian network.

[0085] Table 3

[0086]

[0087] S11, classification is performed using the improved Bayesian network classification model.

[0088] New knee joint data is input, and the new knee joint data only includes the vibration sensor waveform data and the height and weight in the knee joint function evaluation table data; signal features are obtained from the vibration sensor waveform data; this embodiment uses 30 signal features and the height and weight as inputs of the improved Bayesian network classification model, first derives the information in the knee joint function evaluation table, and then uses the information in the joint function evaluation table and the height and weight again as inputs of the improved Bayesian network classification model, and finally outputs the value of the label node, i.e., the first state, the second state or the third state, to classify the new knee joint data.

[0089] Finally, it should be noted that the embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying knee joint data based on a Bayesian network improved by a knowledge graph, characterized by: It includes the following steps: S1, obtain the labeled knee joint dataset; Select knee joint data from the knee joint database and label the selected knee joint data to form a knee joint dataset; S2, feature extraction of knee joint dataset; The features include signal features and functional features, wherein the signal features are obtained based on the vibration sensor waveform data, and the functional features are obtained from the knee joint function evaluation table data; S3, divide the knee joint dataset into training set, validation set and test set; S4, building a Bayesian network; The features, fields, and labels of the indicator data in step S2 are selected as nodes of the Bayesian network, the node order for Bayesian network training is given, and the initial Bayesian network structure is obtained through learning the training set, wherein the search strategy is the hill climbing algorithm, and the scoring function is the K2 scoring function; S5, modifying the Bayesian network of step S4 based on expert knowledge; S6, judging the common causal strength based on the signal features and modifying the Bayesian network in step S5; The specific method for judging the strength of common causality is as follows: I(A,B)=H(B)-H(B|A) (4) Among them, H(A) is the information entropy, i is the value of feature A, n is the value range of A, p i is the probability of the i-th value, H(B|A) is the conditional entropy of B with respect to A, p(B|A) represents the probability of B taking the value under condition A, I(A,B) is the information gain, IGR is the information gain ratio, C v is the coefficient of variation, where σ is the standard deviation of the information gain ratio and μ is the mean; After obtaining the information gain ratio of the two signal features of each channel of the vibration sensor waveform data according to formulas (2)-(5), the standard deviation and mean of the information gain ratio of the same two signals in all channels of the vibration sensor waveform data are calculated. Finally, the coefficient of variation between the two signals is obtained according to formula (6), and the coefficient of variation is used as an indicator of the common causal strength. The coefficients of variation are sorted and selected from small to large according to a preset number, and the selected causal relationship is added to the Bayesian network obtained in step S5. S7, selecting the signal indicator with the strongest causal relationship based on the common causal strength for the relevant vibration sensors, and modifying the Bayesian network obtained in step S6; The signal index with the strongest causal relationship is selected by the common causal strength, and the signal index that characterizes the obvious common causal relationship between known related flexible piezoelectric sensors is determined. The specific method is to select two known related flexible piezoelectric sensors, calculate the information gain ratio of the same signal features of the vibration sensor waveform data of the corresponding channels of the two related flexible piezoelectric sensors, and obtain the coefficient of variation C of the corresponding signal features based on the information gain ratio of the same signal features of all related flexible piezoelectric sensors. v ,The signal feature with the smallest coefficient of variation is selected as the causal relationship and added to the corresponding signal features of the known related flexible piezoelectric sensors; S8, performing parameter learning on the Bayesian network obtained in step S7 through Laplace smoothing and maximum likelihood estimation to obtain quantitative relationships between nodes; S9, performing local search optimization on the signal feature node sequence given in step S4 using the validation set, further optimizing the Bayesian network classification model structure, and obtaining an improved Bayesian network classification model; S10, using the test set to test the accuracy of the improved Bayesian network classification model; S11, classification is performed using the improved Bayesian network classification model.

2. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 1, characterized in that: Each knee joint data in S1 is composed of three types of data: vibration sensor waveform data, index data, and knee joint function evaluation table data; the labels are divided into three types: first state, second state and third state.

3. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 2, characterized in that: The vibration sensor waveform data includes six channels of data, each of which comes from six flexible piezoelectric sensors on the same wearable kneepad. The six flexible piezoelectric sensors are evenly distributed on the wearable kneepad to collect vibration data and save it in a knee joint database. The knee joint function evaluation form data were obtained based on the processed Lysholm table, which at least included height, weight, and graded data on lameness, support, locking, joint instability, pain, swelling, difficulty climbing stairs, and difficulty squatting; The index data is calculated based on the vibration sensor waveform data or the knee joint function evaluation table data.

4. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 1, characterized in that: The specific content of S2 is: After preprocessing the vibration sensor waveform data, five signal features are extracted from each channel of the vibration sensor waveform data: period, energy, kurtosis, correlation and number of inflection points; Ten functional characteristics were extracted from the knee function evaluation form data: height, weight, limp, support, locking, joint instability, pain, swelling, difficulty climbing stairs, and difficulty squatting.

5. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 1, characterized in that: The causal relationships determined by the expert knowledge in S5 include: BMI and label, lysholm table score and label, each functional feature and lysholm table score, height and BMI, weight and BMI, swelling and joint instability, pain and difficulty squatting, pain and difficulty climbing stairs, pain and swelling, limp and support, height and weight. The above causal relationships are all caused by the former and the latter is the effect. If the Bayesian network fails to train these causal relationships through data-driven training in step S4, the missing causal relationships need to be supplemented based on expert knowledge.

6. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 1, characterized in that: The S11 is classified using an improved Bayesian network classification model, specifically: New knee joint data is input, and the new knee joint data only includes the vibration sensor waveform data and the height and weight in the knee joint function evaluation table data; signal features are obtained from the vibration sensor waveform data; the signal features and height and weight are used as the input of the final Bayesian network structure classification model, and the functional feature values ​​in the knee joint function evaluation table are first derived. Then, the functional feature values ​​in the joint function evaluation table and the height and weight are again used as the input of the improved Bayesian network classification model, and finally the value of the label node is output to classify the new knee joint data.

7. The method for classifying knee joint data based on a Bayesian network improved by a knowledge graph according to claim 1, characterized in that: The local search optimization in S9 adopts simulated annealing or gradient descent algorithm.

Images