Intelligent monitoring system and method for physical training

Abstract

The invention provides an intelligent monitoring system and method for physical training, and the system comprises a data collection module, a data processing module, an intelligent monitoring module and a safety alarm module, and the method comprises the steps: collecting related data of the physical training based on the data collection module, preprocessing the collected related data based on the data processing module, and sending the preprocessed data to the intelligent monitoring module; based on the preprocessed related data, intelligent monitoring is conducted through the intelligent monitoring module, an intelligent monitoring result is obtained, the safety alarm module judges whether physical training is abnormal or not according to the intelligent monitoring result, and if yes, an alarm is given. Intelligent monitoring of physical training can be achieved, the safety of physical training can be guaranteed, and use is convenient.

Description

Technical Field

[0001] The present invention relates to the technical field of sports training, and in particular to a sports training intelligent monitoring system and method. Background Art

[0002] With the rapid development of science and technology, the methods and means of sports training are also constantly innovating. Traditional sports training methods mainly rely on the experience of coaches, subjective feedback from athletes, and some basic physiological monitoring equipment. However, these methods often have certain limitations, such as incomplete data collection, inaccurate data analysis, and lagging feedback mechanisms, which cannot meet the growing training needs of high-level athletes and the high-precision requirements for training effects.

[0003] At present, although there are many sports monitoring devices on the market, most of them only focus on a single physiological indicator (such as heart rate monitoring) or sports parameters (such as step counting), lacking intelligent monitoring and comprehensive evaluation of the entire training process of athletes, and unable to predict the risks that occur during sports training and reduce the occurrence of accidents. Therefore, it is very necessary to design an intelligent monitoring system and method for sports training. Summary of the invention

[0004] In order to overcome the deficiencies of the prior art, an object of the present invention is to provide a sports training intelligent monitoring system and method.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] The present invention provides a sports training intelligent monitoring system, comprising:

[0007] Data collection module, used to collect sports training related data;

[0008] A data processing module, used for preprocessing relevant data;

[0009] Intelligent monitoring module, used for intelligent monitoring based on relevant data;

[0010] The safety alarm module is used to determine whether the physical training is abnormal based on the intelligent monitoring results, and to issue an alarm if it is abnormal.

[0011] The present invention also provides a sports training intelligent monitoring method, comprising:

[0012] Step 100: Collecting sports training related data based on a data collection module;

[0013] Step 200: pre-processing the collected relevant data based on the data processing module;

[0014] Step 300: Based on the pre-processed relevant data, intelligent monitoring is performed through an intelligent monitoring module to obtain intelligent monitoring results;

[0015] Step 400: The safety alarm module determines whether the physical training is abnormal based on the intelligent monitoring results, and if so, issues an alarm.

[0016] Preferably, in step 100, the physical training related data includes external temperature, humidity, physical training type, trainee's gender, height, age, weight, vital capacity, heart rate and exercise time.

[0017] Preferably, in step 200, the collected relevant data is preprocessed based on the data processing module, specifically:

[0018] Pre-process the collected external temperature, humidity, vital capacity and heart rate of the trainee, including:

[0019] Detect data errors based on a multi-rule multi-type error data monitoring algorithm;

[0020] The detected errors are repaired by a multi-rule based multi-type error data repair algorithm;

[0021] The dataset is constructed based on the repaired data and other data from sports training related data.

[0022] Preferably, in step 300, based on the pre-processed relevant data, intelligent monitoring is performed by an intelligent monitoring module to obtain intelligent monitoring results, specifically:

[0023] Build DeepFM model and LightGBM model;

[0024] Divide the data set into training set and test set;

[0025] The training set is divided into 5 parts, any 4 parts are used to train the DeepFM model and the LightGBM model, and the remaining 1 part is used to verify the model. Finally, the prediction results of each data in the training set are obtained, and the 5 prediction results are merged. The prediction results of the DeepFM model and the LightGBM model are recorded as new features 1 and new features 2 respectively;

[0026] The test set is divided into 5 parts, and the DeepFM model and the LightGBM model are input respectively to obtain the prediction results of the 5 test sets. The prediction results of the test set are obtained by adding and averaging. The prediction results of the DeepFM model and the LightGBM model are recorded as prediction 1 and prediction 2 respectively.

[0027] New feature 1 and new feature 2 are concatenated as new features, and corresponding labels are added as the training data set of the naive Bayes classifier. Prediction 1 and prediction 2 are concatenated as the test data set of the naive Bayes classifier to obtain the final prediction result.

[0028] Preferably, in step 400, the safety alarm module determines whether the physical training is abnormal according to the intelligent monitoring result, and if abnormal, an alarm is issued, specifically:

[0029] The final prediction result is obtained. If it is greater than the preset threshold, the physical training is judged to be abnormal and the safety alarm module issues an alarm.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] The present invention provides a sports training intelligent monitoring system and method, the system includes a data acquisition module, a data processing module, an intelligent monitoring module and a safety alarm module, the method includes collecting sports training related data based on the data acquisition module, preprocessing the collected related data based on the data processing module, performing intelligent monitoring through the intelligent monitoring module based on the preprocessed related data, obtaining intelligent monitoring results, and the safety alarm module determines whether the sports training is abnormal according to the intelligent monitoring results, and if abnormal, alarms. The present invention can realize intelligent monitoring of sports training, can ensure the safety of sports training, and is easy to use. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0034] Figure 2 It is a schematic diagram of FM structure;

[0035] Figure 3 It is a schematic diagram of the Deep structure;

[0036] Figure 4 Schematic diagram for converting an input vector into a low-dimensional dense vector;

[0037] Figure 5 This is a schematic diagram of the DeepFM structure;

[0038] Figure 6 This is a schematic diagram of the LightGBM model structure;

[0039] Figure 7 Schematic diagram of the Stacking fusion process. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The purpose of the present invention is to provide a sports training intelligent monitoring system and method, which can realize sports training intelligent monitoring, ensure the safety of sports training, and be easy to use.

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] The present invention provides a sports training intelligent monitoring system, comprising:

[0044] Data collection module, used to collect sports training related data;

[0045] A data processing module, used for preprocessing relevant data;

[0046] Intelligent monitoring module, used for intelligent monitoring based on relevant data;

[0047] The safety alarm module is used to determine whether the physical training is abnormal based on the intelligent monitoring results, and to issue an alarm if it is abnormal.

[0048] like Figure 1 As shown, the present invention also provides a sports training intelligent monitoring method, comprising:

[0049] Step 100: Collecting sports training related data based on a data collection module;

[0050] Step 200: pre-processing the collected relevant data based on the data processing module;

[0051] Step 300: Based on the pre-processed relevant data, intelligent monitoring is performed through an intelligent monitoring module to obtain intelligent monitoring results;

[0052] Step 400: The safety alarm module determines whether the physical training is abnormal based on the intelligent monitoring results, and if so, issues an alarm.

[0053] In step 100, the physical training related data includes external temperature, humidity, physical training type, gender, height, age, weight, vital capacity, heart rate and exercise time of the trainee.

[0054] In step 200, the collected relevant data is preprocessed based on the data processing module, specifically:

[0055] Pre-process the collected external temperature, humidity, vital capacity and heart rate of the trainee, including:

[0056] Detect data errors based on a multi-rule multi-type error data monitoring algorithm;

[0057] The detected errors are repaired by a multi-rule based multi-type error data repair algorithm;

[0058] constructing a dataset based on the repaired data and other data from sports training related data;

[0059] Next, we will first introduce the multi-type error data monitoring algorithm based on multiple rules:

[0060] Error detection is the first step in data cleaning. Due to the complex relationship between rules, we first consider how to deal with a rule set with potential rule conflicts and how to apply multiple rules to find erroneous data after resolving the conflicts.

[0061] First, let’s introduce conflict resolution:

[0062] 1. Rule Set Harmony Index

[0063] There are five types of relationship analysis between rules, namely logical order determination, empty attribute intersection, rule conflict, rule redundancy, and rule mutual exclusion. The first two relationships will not destroy the harmony of the rule set, while the last three will affect the harmony of the rule set. Below we give the harmony index of the rule set;

[0064] The rule set harmony index is given by the multidimensional time series data x∈T M×N and a set of rules R of size n defined on it set = {R1, R2, ..., R n}, and give the formula for calculating the harmony index of its rule set:

[0065]

[0066] The explanation of this formula is as follows:

[0067] (1) The numerator uses a summation sign to represent all values ​​of the parameters i and l that satisfy the definition;

[0068] (2) For the numerator, it means that when studying the determined rule subset, it can be judged that this rule subset either has no conflict or one of the three events A, B, and C occurs. Then, a simple addition can be used to measure the number of problems, where P(A) is a continuous value, and P(B) and P(C) are discrete values.

[0069] (3) For the denominator, we need to select l+1 rules from n rules to determine the relationship, and l+1 ≥ 2. The upper bound of this sampling is the number of combinations of n minus the case of l+1 = 0 or 1, so it is expressed as 2 n -n-1.

[0070] (4) The value range of this number is between 0 and 1. The closer the value is to 0, the more harmonious the rule set is. When it is equal to 0, it means that there is no conflict in the rule set. On the contrary, when the value is close to 1, it means that the rule set has serious conflicts. In the extreme case of 1, it means that any rule subset is problematic.

[0071] For a rule set, rule mutual exclusion is much more serious than rule conflict, and rule conflict is much more serious than rule redundancy. Mutual exclusion can be considered a strong conflict, and redundancy can be considered a weak conflict. In the formula, these three situations are treated equally, which is inappropriate in data cleaning. In addition, the value obtained by this formula may be too small, and the observation effect is not good.

[0072] X∈T M×N and a set of rules R of size n defined on it set = {R1, R2, ..., R n}

[0073] Therefore, the updated rule set harmony index is M×N and a set of rules R of size n defined on it ser = {R1, R2, ..., R n}, and the added weight set W rset = {W A , W B , W C , W K}, we use the triple XRW = {X, R set , W rset} to indicate.

[0074] Weighted rule set harmony index for XRW = {X, R set , W rset The formula for calculating the harmony index is given as follows:

[0075]

[0076] This formula is explained as follows:

[0077] (1)W A , W B , W C Respectively represent the weights of events A, B, and C, which are used to adjust the influence coefficients of the three events on the harmony index;

[0078] (2) When n is large enough in the denominator, the exponential level is much larger than the linear level. Consider keeping only 2 n .

[0079] (3) When the rule set is large and conflicts are few, this value will be very small and difficult to observe. k is to multiply it by a coefficient, according to the specific situation W k It can be a fixed value or a function of n.

[0080] (4) After adjustment, the range of calculated values ​​may theoretically exceed 1, but the probability of this happening is very low, and the observed effect is better after adjustment.

[0081] 2. Digestion operation

[0082] Given XRW = {X, R set ,W rset}, for R set If there is a mapping r:R set →R' set , so that R s et' no longer has quality problems, then r is called R s The digestion operation of et, digestion operation r = {r1, r2, ...}, where r i Represents the rule R i ∈R s et an operation;

[0083] R i To explain, we have the highest tolerance for rule redundancy. Rule redundancy only reflects that some rules may be redundant, but it does not mean that redundant rules are worthless, because redundant rules can also lead to wrong data.

[0084] Mutual exclusion of rules is intolerable. Mutual exclusion of rules means that no correct data can pass. The tolerance of rule conflicts depends on the size of P(A). Mutual exclusion and rule conflicts are essentially too high a constraint on data. Constraints need to be relaxed in order to eliminate mutual exclusion and reduce the degree of rule conflicts, so that D i It is rule R i The correct data set defined, r i That is, we need to make D iExpanded, formalized as follows:

[0085]

[0086] After defining the resolution operation, we need to consider the resolution cost. Here, we need to consider three dimensions for the resolution cost: the number of modified rules, the extent of rule modification, and the different levels of trust in different rules. The resolution cost calculation formula is given below:

[0087]

[0088] The explanation of this formula is as follows:

[0089] (1)W i Represents the rule R i The trust level is a fixed value ranging from 0 to 1.

[0090] (2) For the molecular part, only D' i >D i When W i It will enter the cumulative sum.

[0091] (3) Indicates that after operation r i The original correct data range accounts for the proportion of the modified data, and the value range is 0-1, D i is a fixed value, D' i Is the change value. Special note D' i →∞ means that rule R is deleted. i .

[0092] (4) The numerator indicates that selecting a small number of rules with low confidence will make the cost smaller, and the denominator indicates the degree of change to the initial rule set, D' i The bigger it is, the greater the cost.

[0093] (5) When no rule is adjusted, the denominator is n, the numerator is 0, and the cost of elimination is 0. When the rule set is adjusted greatly, the denominator tends to 0, the numerator tends to the sum of the trust of all rules, and the cost of elimination is →∞. This performance is in line with expectations, that is, the number of modified rules is small, the modification range is small, and the cost of elimination is reduced if the rules with high trust are not adjusted.

[0094] Based on the above discussion on the rule harmony index and the resolution operation, it can be concluded that the goal of the optimization problem is to make both ha and cost as small as possible, so the formal definition is:

[0095] r:R set →R' set

[0096] ha <threshold and min{S=ha×cost}

[0097] Next, we will introduce error data detection:

[0098] 1. Wrong data type

[0099] Error data detection refers to the discovery of data points in the data that do not meet the constraints. These data points that do not meet the expectations can be called anomalies, outliers, out-of-limit values, etc. There are many reasons for the errors. The simple reason may be that the sensor ages and errors occur when collecting data.

[0100] Based on this, error data can be divided into three types: single point error, continuous error and context error;

[0101] (1) Single point error: This is the most common type of error, manifested as an error in a single data point that deviates from other correct data. Typical examples include null values, zero values, and out-of-limit values. It can also be divided into single point large errors and single point small errors. A single point large error can be considered as a large deviation between the observed value of a data point at a certain moment and the expected value; a single point small error can be considered as a small deviation from the expected value of a data point but showing a significant difference, which can be a non-out-of-limit mutation value.

[0102] (2) Continuous errors: In the time series data center, data points at multiple consecutive moments appear abnormal. This type of error is also relatively common. For example, a sensor crashes for a short period of time, resulting in missing data for a period of time.

[0103] (3) Contextual error: This is a special type of error, in which the data is considered to be erroneous in a specific context. It can also be considered a conditional error. The following two attributes are needed to describe contextual errors:

[0104] Context attributes: Context attributes are used to determine the specific environment of the data. For example, for the speed constraint of time series data, the previous and next data of this data point are the constraint environment; the data in a certain range where each data point in the box plot is located is the context environment constraining this data point;

[0105] Behavioral attributes: Behavioral attributes are defined on the specific meaning of data points.

[0106] 2. Error data detection under multiple rules

[0107] For a given set of rules and a multidimensional time series data X∈T M×NA simple and direct idea is to apply each rule to the data set in turn, and then all the erroneous data fragments can be obtained. However, this has problems. On the one hand, the data set is often very large, and a very large data set will get a large set of erroneous fragments at one time, which is not easy to observe and will make users anxious. On the other hand, if only multiple erroneous data fragments are displayed, users will be confused and unable to analyze the cause of the error.

[0108] Consider that each rule only acts on a very small fragment of the data set at a time. After considering all the rules, we can detect the data of a time window at a time. The size of the time window depends on the scope of all the rules. This effectively simplifies the problem. In addition, for each error fragment, in order to observe, consider that an error consists of an error fragment and a violated rule, which can be defined as e k ={{d},R i}, {d} represents the point set of erroneous data points;

[0109] In each error obtained, the error fragment must contain erroneous data points, but not every data point is erroneous. Considering that an erroneous data point often violates multiple rules, it is proposed that data points that appear repeatedly in the error fragment in a time window are more likely to be error points, so erroneous data detection can be achieved.

[0110] Next, we will introduce the multi-type error data repair algorithm based on multiple rules:

[0111] The present invention provides two methods, one of which is to perform brute force solution, select a given set of rules, and sequentially execute the constraint judgment function B and repair function F of each rule on the entire data set;

[0112] Another method is to introduce hypergraph theory to repair erroneous data. The specific steps are as follows:

[0113] stepl: Design the time window size to be size, and the hypergraph G =<V,E,W> initialization;

[0114] Step 2: Use multiple rules to detect and put each hyperedge into the hypergraph;

[0115] Step 3: Determine the sparsity of the multi-rule data repair hypergraph. If it is a dense hypergraph, proceed to the next step. If it is a sparse hypergraph, proceed to step 5.

[0116] Step 4: Call the dense hypergraph processing algorithm to convert the dense hypergraph into a sparse hypergraph;

[0117] Step 5: Call the algorithm for processing sparse hypergraphs to repair each data point in the sparse graph. Each data point must be constrained by all relevant rules at the same time.

[0118] In step 300, based on the pre-processed relevant data, intelligent monitoring is performed through the intelligent monitoring module to obtain intelligent monitoring results, specifically:

[0119] Build DeepFM model and LightGBM model;

[0120] Divide the data set into training set and test set;

[0121] The training set is divided into 5 parts, any 4 parts are used to train the DeepFM model and the LightGBM model, and the remaining 1 part is used to verify the model. Finally, the prediction results of each data in the training set are obtained, and the 5 prediction results are merged. The prediction results of the DeepFM model and the LightGBM model are recorded as new features 1 and new features 2 respectively;

[0122] The test set is divided into 5 parts, and the DeepFM model and the LightGBM model are input respectively to obtain the prediction results of the 5 test sets. The prediction results of the test set are obtained by adding and averaging. The prediction results of the DeepFM model and the LightGBM model are recorded as prediction 1 and prediction 2 respectively.

[0123] New feature 1 and new feature 2 are concatenated as new features, and corresponding labels are added as the training data set of the naive Bayes classifier. Prediction 1 and prediction 2 are concatenated as the test data set of the naive Bayes classifier to obtain the final prediction result.

[0124] The above process is introduced in detail:

[0125] 1. DeepFM model construction

[0126] DeepFM is a model that integrates deep neural networks and factor decomposition machines. It can learn the correlation between features and high-order feature combinations, which helps to improve prediction accuracy. The factor decomposition machine can provide embedding vectors of original features, so that the weights of features can be better understood and adjusted. At the same time, DeepFM adopts a shared weight approach to reduce the parameters and computational costs of the neural network model, while having faster training and prediction speeds. The DeepFM model has a good application prospect and is widely used in recommendation systems, search ranking, advertising recommendation, natural language processing and other fields, and has achieved good results. However, FM and its derivative algorithms are rarely studied in the field of sports. Therefore, based on the above advantages, the present invention constructs a DeepFM model to predict the probability of occurrence of sports training risk events. The construction process can be divided into three stages: feature combination, efficient feature representation and classification prediction. DeepFM first combines the input features to represent the interaction between them. This is achieved through the factorization machine (FM) model. The FM model factors the interactions between features to obtain a series of cross factors to represent the interactions between features. In simple terms, FM represents the interaction between any two features in a vector as the product of their feature factors. By combining features in this way, DeepFM can better capture the high-order relationships between features, which are usually not captured by linear models. Then, in order to represent these interactive features, DeepFM uses a neural network model. Specifically, it uses the combined features as the input of the neural network and uses a multi-layer perceptron (MLP) to learn efficient feature representations. This can automatically learn the nonlinear representation of features through the neural network architecture, enhancing the factorization machine's learning and representation capabilities for feature representation and feature interaction. Finally, DeepFM uses the feature vectors obtained from the previous two processes as the input of the classification model, uses a fully connected layer for classification, and outputs the probability results through the final sigmoid function;

[0127] In summary, the construction process of the DeepFM model is to combine the factorization machine and the neural network to perform multiple stages such as feature decomposition, cross-feature learning, and feature mapping, so as to improve the model's learning and abstraction capabilities of data features, thereby improving the precision and accuracy of its classification prediction. The DeepFM model consists of two parts: the FM component and the Deep component, which respectively realize low-order and high-order feature interactions;

[0128] (1) FM component: The FM component is mainly used to perform special processing and feature crossover on low-dimensional feature vectors. By decomposing the latent factors, it learns the connection and weight between the implicit attributes of the features. It can also adaptively adjust parameters during model training to improve the performance of the model. The core idea of ​​the FM component is to perform second-order feature crossover on the feature vector to reduce the number of model parameters and improve the model's understanding and expression of feature crossover. Specifically, in FM, each dimension in the feature vector is represented as an implicit factor, and the prediction result of the feature vector is obtained by performing outer product operations and cumulative summation between the implicit factors. Mathematically, the representation of the FM model is shown in the following formula.

[0129]

[0130] The first two items are the linear model part, and the last item is the second-order feature interaction part, w o is the bias term, x i represents the i-th component of the eigenvector x, w i represents the parameter of the i-th feature, n is the number of variables, w ij represents the coefficient multiplied by the i-th and j-th features;

[0131] In the dataset of the present invention, many features are discrete. Therefore, these discrete features need to be one-hot encoded, which will make the data particularly sparse. When the value of a series of discrete data is very large, for example, 1 million, after one-hot encoding, the feature space increases by 1 million dimensions, and the data will be too sparse. In the second-order feature combination part, the data is too sparse and it is impossible to learn w ij ; To solve the above problem, an implicit vector v is introduced for each feature i =(v i1 ,v i2' …,v ik )(k is a hyperparameter), and use <v i ,v j >Replace w i j, at this time, w i The solution of j is converted to v i and v j The converted formula is as follows:

[0132]

[0133] For each variable, the FM component learns a one-dimensional vector of fixed size k through the embedding layer. Therefore, the variable x i and x j The weights between them can be expressed by the feature correspondence vector v i and v j The inner product of the original FM is O(kn2 ), through conversion, the time complexity is changed from O(kn 2 ) is reduced to O(kn), and the formula conversion process is shown as follows:

[0134]

[0135] In the formula, <v i ,v j > represents a vector v i and v j The inner product of i f represents the vector v i The fth component of , after such transformation and decomposition, can be solved by stochastic gradient descent (SGD), as shown below:

[0136]

[0137] The detailed structure of FM is as follows Figure 2 As shown;

[0138] Before one-hot encoding, each feature belongs to a feature domain. In the sparse feature part, the discrete data feature domain after one-hot encoding contains multiple columns, while the continuous feature remains unchanged, or one column corresponds to one feature domain. The feature domain can be used to convert the sparse large matrix into a dictionary and two small matrices during storage, thereby achieving the purpose of anti-sparseness and saving storage space: the small dictionary adds an index mark to each eigenvalue of the discrete feature, and each continuous feature is only assigned a feature index identifier. The first small matrix is ​​the feature index matrix, whose length is the sample length, and each column represents the index identifier of the sample eigenvalue in the dictionary. The sparsity problem caused by one-hot encoding, one column replaces all columns in a feature domain, and the second small matrix is ​​the eigenvalue matrix, which stores the eigenvalues ​​of the corresponding index position. The recorded discrete eigenvalues ​​are all 1, not 0, and the continuous type stores the original value. Through the above conversion, the feature domain of the corresponding feature can be easily found to participate in the calculation. Then introduce the dense embedding layer to convert the discrete vector of the sparse feature part into a low-dimensional dense vector;

[0139] In the FM layer, the plus sign corresponds to the first-order feature calculation. In the figure, it is connected to the sparse feature part through a black line with a weight of w. i , the cross circle corresponds to the interaction calculation between each feature, and the red line represents its connection with the dense embedding, which will use the above dictionary, feature index matrix and eigenvalue matrix to calculate the interaction features;

[0140] (2) Deep component: The Deep component refers to the neural network part, which accepts normalized features as input. It is mainly used to perform high-dimensional representation of cross features and extract deep semantic information. It can also perform classification prediction tasks. The Deep component contains multiple layers of fully connected layers and activation functions. The ReLU activation function is usually used. The input part is the input feature converted into a fixed-dimensional vector representation through the embedding layer. The dimension of the vector is determined by the size of the Embedding. The input features of the Deep component can be new features obtained after feature crossover or original features. These encoded feature vectors are input into a multi-layer fully connected neural network (Multi-Layer Perceptron, MLP). The input feature vectors are nonlinearly transformed through the fully connected layer to improve the feature abstraction ability layer by layer. The output of each layer is used as the input of the next layer. The output layer is a separate sigmoid function layer that outputs the probability prediction result. The detailed structure of the Deep component is as follows: Figure 3 As shown;

[0141] The Deep component and the FM component will share the low-dimensional dense vector of the dense embedding part, which is converted from the input vector of the sparse feature part, such as Figure 4 Each feature field will be transformed through such a fully connected network, regardless of the length of each feature field before (the length after one-hot encoding), it will be converted into a vector of length k, where v i j is not part of the FM hidden vector, but k here has the same length as the hidden vector, which is convenient for subsequent calculations. All transformed low-dimensional vectors are merged, that is, a (0) =(e1,e2,...,e m ), as the output of the dense embedding layer and the input of the hidden layer;

[0142] In terms of risk prediction, the DNN model can utilize a large amount of historical data and structured data to learn the complex nonlinear relationship between features, thereby improving the prediction precision and accuracy. Therefore, DNN is used as the deep part of the model, and the formula of the output layer is as follows;

[0143] a (H+1) =σ(w (H) a (H) +b (H) )

[0144] Where H is the number of hidden layers, a (l) is the output of the ιth hidden layer, w (l) and b (l) are weights and biases, respectively. Figure 3 The circle in the hidden layer represents the activation function ReLU, and the formula is as follows:

[0145]

[0146] It can avoid the gradient vanishing problem caused by the calculation of the sigmoid function, but in the case of binary classification, sigmoid usually appears in the last layer to obtain the probability value. The formula is as follows:

[0147]

[0148] Finally, the outputs of the FM part and the Deep part will be integrated in the output unit part and the result will be calculated by sigmoid. The formula is as follows. The whole process of the DeepFM model is as follows Figure 5 shown.

[0149]

[0150] 2. LightGBM model construction

[0151] LightGBM is a gradient boosting framework that utilizes a research rule set that is primarily based on trees. It is a boosting algorithm that uses leaf-grown trees rather than horizontally grown trees. Compared with alternative algorithms such as XGBoost and CatBoost, LightGBM is more efficient and accurate. The amount of data in the sports risk dataset is relatively large. The LightGBM model no longer divides the data when processing large amounts of data, thereby reducing the amount of computation. Therefore, with the use of this classifier, sports training can be predicted quickly and efficiently.

[0152] The LightGBM algorithm is an optimized version model of the gradient boosting decision tree. It is essentially a Boosting type algorithm in ensemble learning. It uses serial sequential connections for base learners and continuously adds new base learners to update the weights of misclassified samples. It uses misjudgment experience to improve the model and iteratively learns to obtain the minimum loss function result. Its main hyperparameters are also related to the parameters in the decision tree. Usually, the hyperparameters of the tree model are divided into the following four categories: training speed parameters, tree structure control, accuracy improvement, and overfitting control parameters for adjustment. The main hyperparameters of LightGBM are shown in Table 1.

[0153] Table 1 LightGBM hyperparameters

[0154]

[0155] In the process of building the LightGBM model, data preprocessing and feature selection have a crucial impact on the accuracy and performance of the model. In addition, the parameter setting and tuning of the model will also affect the accuracy and stability of the prediction results. The LightGBM model mainly includes a histogram-based decision tree algorithm, a Leaf-Wise leaf growth strategy with depth restrictions, unilateral gradient sampling, and direct support for category features. The detailed structure of the LightGBM model is as follows: Figure 6 shown.

[0156] 3. Stacking model integration of DeepFM and LightGBM

[0157] This model is a fusion model based on DeepFM and LightGBM, combining the advantages of deep learning and gradient boosting tree algorithms. The DeepFM algorithm has the characteristics of easy model training and high model accuracy, and can learn high-order interaction information between input features, while the LightGBM algorithm has high efficiency and accuracy, and can maintain high-order accuracy when processing large-scale data. Therefore, combining the two algorithms can give full play to their respective advantages and improve the prediction ability of the model. The fusion process is as follows: Figure 7 shown.

[0158] In step 400, the safety alarm module determines whether the physical training is abnormal according to the intelligent monitoring results. If abnormal, an alarm is issued, specifically:

[0159] Get the final prediction result. If it is greater than the preset threshold, the sports training is judged to be abnormal, and the safety alarm module will sound an alarm;

[0160] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0161] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A sports training intelligent monitoring system, characterized in that: include: Data collection module, used to collect sports training related data; A data processing module, used for preprocessing relevant data; Intelligent monitoring module, used for intelligent monitoring based on relevant data; The safety alarm module is used to determine whether the physical training is abnormal based on the intelligent monitoring results, and to issue an alarm if it is abnormal.

2. A sports training intelligent monitoring method, characterized in that: include: Step 100: Collecting sports training related data based on a data collection module; Step 200: pre-processing the collected relevant data based on the data processing module; Step 300: Based on the pre-processed relevant data, intelligent monitoring is performed through an intelligent monitoring module to obtain intelligent monitoring results; Step 400: The safety alarm module determines whether the physical training is abnormal based on the intelligent monitoring results, and if so, issues an alarm.

3. The method according to claim 2, characterized in that In step 100, the physical training related data includes external temperature, humidity, physical training type, gender, height, age, weight, vital capacity, heart rate and exercise time of the trainee.

4. The method according to claim 3, characterized in that In step 200, the collected relevant data is preprocessed based on the data processing module, specifically: Pre-process the collected external temperature, humidity, vital capacity and heart rate of the trainee, including: Detect data errors based on a multi-rule multi-type error data monitoring algorithm; The detected errors are repaired by a multi-rule based multi-type error data repair algorithm; The dataset is constructed based on the repaired data and other data from sports training related data.

5. The method according to claim 4, characterized in that In step 300, based on the pre-processed relevant data, intelligent monitoring is performed through the intelligent monitoring module to obtain intelligent monitoring results, specifically: Build DeepFM model and LightGBM model; Divide the data set into training set and test set; The training set is divided into 5 parts, any 4 parts are used to train the DeepFM model and the LightGBM model, and the remaining 1 part is used to verify the model. Finally, the prediction results of each data in the training set are obtained, and the 5 prediction results are merged. The prediction results of the DeepFM model and the LightGBM model are recorded as new features 1 and new features 2 respectively; The test set is divided into 5 parts, and the DeepFM model and the LightGBM model are input respectively to obtain the prediction results of the 5 test sets. The prediction results of the test set are obtained by adding and averaging. The prediction results of the DeepFM model and the LightGBM model are recorded as prediction 1 and prediction 2 respectively. New feature 1 and new feature 2 are concatenated as new features, and corresponding labels are added as the training data set of the naive Bayes classifier. Prediction 1 and prediction 2 are concatenated as the test data set of the naive Bayes classifier to obtain the final prediction result.

6. The method according to claim 5, characterized in that In step 400, the safety alarm module determines whether the physical training is abnormal according to the intelligent monitoring results. If abnormal, an alarm is issued, specifically: The final prediction result is obtained. If it is greater than the preset threshold, the physical training is judged to be abnormal and the safety alarm module issues an alarm.

Images