Case-based reasoning method for man-machine collaborative personalized exercise training driven by causal knowledge

Abstract

The invention relates to the technical field of intelligent exercise training, and particularly discloses a causal knowledge-driven man-machine collaborative personalized exercise training case reasoning method, which comprises the following steps of: constructing a case library containing a plurality of cases, and performing causal knowledge reconstruction at different levels; searching a case most similar to the target case from a case library based on the weight of each description variable; the difference of training scheme variables between the target case and the retrieved similar cases is recognized, anti-fact intervention and inference of any training scheme variable are conducted on the target case, and one or more new cases are screened out; when an actual result generated after the new case is executed does not conform to expectation, executing an attribution process, and generating different processing suggestions according to attribution types; new cases which do not conform to expectation are stored in a case library, and causal knowledge reconstruction is periodically carried out. According to the method, causal science and human-computer interaction can be deeply fused to carry out personalized training scheme reasoning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent sports training, and more particularly to a case reasoning method for personalized sports training driven by causal knowledge and human-computer collaboration. BACKGROUND

[0002] Case reasoning technology generates causal knowledge from a case library through structural equations, and then the causal knowledge is used throughout the reasoning process to obtain a training plan. However, current case reasoning technology often cannot meet the personalized needs of sports training, especially competitive sports training, mainly because the causal knowledge supporting reasoning is insufficient, and the collaboration between coaches and athletes is not sufficient.

[0003] Therefore, how to deeply integrate causal science and human-computer interaction to reason about personalized training plans is a problem that needs to be solved by those skilled in the art. SUMMARY

[0004] In view of the above problems, the present application provides a case reasoning method for personalized sports training driven by causal knowledge and human-computer collaboration, so as to overcome the above problems or at least partially solve the above problems.

[0005] In order to achieve the above purpose, the present application adopts the following technical solutions: A case reasoning method for personalized sports training driven by causal knowledge and human-computer collaboration, comprising the following steps: S1, a case library containing a plurality of cases is constructed, and the observation data of each case is composed of description variables, training plan variables and sports performance variables; based on all observation data of the case library, causal knowledge is reconstructed at different levels; S2, the weights of each description variable are calculated on the reverse causal graph of the description variables and the sports performance variables by using the PageRank algorithm, and the most similar case to the target case is retrieved from the case library based on the weights of each description variable; S3, the differences in training plan variables between the target case and the retrieved similar case are identified, the counterfactual intervention and inference of the target case are performed on any training plan variable, and one or more new cases are selected; S4, the actual sports performance after the execution of the new case is compared with the original sports performance, when the two are inconsistent, the feedback of the coach is received, the attribution process is performed, and different processing suggestions are generated according to the attribution type; S5, when the execution result of the new case is inconsistent with the original sports performance, the new case is stored in the case library, and the causal knowledge is reconstructed periodically.

[0006] Furthermore, in S1, each case is defined as a triple (X, T, Y), where X is a set containing m descriptive variables, denoted as X = {x1, x2, ..., x...}. m These descriptive variables describe the case background, environment, or the state of the subject. T is a set containing n training scheme variables, denoted as T = {t1, t2, ..., t}. n These training program variables are the interventions or actions implemented; Y is a set containing k athletic performance variables, represented as Y = {y1, y2, ..., y...} k These scalar metrics of athletic performance are the results or outputs that need to be observed; Case Library CB For a containing p A set of cases, denoted as CB = {( X 1, T 1, Y 1), ( X 2, T 2, Y 2), ..., ( X p , T p , Y p )}.

[0007] Furthermore, in S1, the process of reconstructing causal knowledge at different levels includes: Determine the set of variables V XY = X∪Y, in the variable set V XY The causal discovery algorithm is executed to construct a description-representation causal graph; Determine the set of variables V TY = T ∪ Y In the variable set V TY The causal discovery algorithm is executed to construct the training-performance causal graph; Determine the set of variables V XTY = X ∪ T ∪ Y In the variable set V XTY The causal discovery algorithm is executed to construct a description-training-performance causal graph.

[0008] Furthermore, S2 includes: Reverse the direction of all directed edges in the causal graph to get the reverse causal graph; On the reverse causal graph, use PageRank algorithm to generate a PageRank score value for each description variable node as the weight of the description variable node; Calculate the distance between the target case and any case in the case base based on the weight of each description variable node; Sort all cases in the case base according to the distance from the target case from small to large, and select the top k cases as the retrieval result.

[0009] Further, the distance calculation formula between the target case and any case in the case base is:

[0010] Wherein, d w ( Q , C j ) represents the weighted distance between the target case Q and the jth case C j in the case base; i represents the ith description vector node, m represents that a case contains m description vector nodes, w i represents the weight of the ith description vector node, q i represents the feature value of the ith description vector in the target case, and C ji represents the feature value of the ith description vector of the jth case in the case base.

[0011] Further, the calculation process of using PageRank algorithm to calculate the weight of each description variable node is: Initialize the weights of all description variable nodes to equal values, and update the weights of each description variable node according to the following iteration formula until the weight values of all description variable nodes converge; Wherein, the weight iteration update formula is:

[0012] Wherein, is the weight of the description variable node x i The weight of the description variable node k in the first iteration; is the weight of the description variable node x i The weight of the description variable node k in the first iteration; d is the damping factor; G R is the reverse causal graph; V is the total number of nodes in the reverse causal graph; In( x i, G R ) is a set of nodes pointing to the node G R in the reversed causal graph x i ; Out( x i , G R ) is the out-degree of the node x i in the reversed causal graph G R .

[0013] Further, S3 comprises: S31, comparing the target case and any similar case on training program variables, finding out training program variables with different values, and constructing a set of difference variables Δ T j ; S32, calculating the power set of Δ T j (Δ P T j ), each element S in the power set of Δ P T j is a subset composed of one or more difference training program variables, representing a potential modification program; S33, for each element S in the power set, constructing a counterfactual intervention, which is: modifying the training program variables belonging to element S in the target case to the corresponding training program variables in the similar case, and keeping the values of other training program variables in the target case unchanged; S34, executing S31-S33 on all similar cases to generate a set of counterfactual interventions; S35, applying each counterfactual intervention in the set of counterfactual interventions to the individual causal model corresponding to the target case one by one, performing counterfactual inference, and calculating the potential values of the sports performance variables under different counterfactual interventions; S36, summarizing the potential performance under all counterfactual interventions into a list, each entry containing: a complete candidate training program, the potential value of the sports performance variable, and the amount of improvement relative to the original sports performance variable value; sorting the list in descending order according to the improvement amount, and selecting one or more new cases from it.

[0014] Further, S4 comprises: S41, comparing the actual value of the sports performance variable of the athlete after performing the new case Y actual with the original sports performance​​Y Q If Y actual is worse than Y Q , then it is determined as a failure; if no failure occurs, then the causal surprise degree is calculated; S42, when a failure is detected or the causal surprise degree exceeds the threshold, a structured interface or guided question is provided, and error attribution is performed according to the received coach feedback; the attribution type is divided into an accidental situational factor or a structural model defect; S43, different processing suggestions are generated according to the attribution type, for the accidental situational factor, the new case is marked as a situational abnormal case; for the structural model defect, the received expert knowledge of the coach is converted into one or more specific modification hypotheses of the structural causal model.

[0015] Further, the calculation process of the causal surprise degree in S41 includes: all the directed edges in the description-training-performance causal graph are reversed to obtain a corresponding full-variable reverse causal graph; the PageRank algorithm is run on the full-variable reverse causal graph to calculate a global causal weight vector X , T , Y for all variables (Y, X, D) W XTY ; the actual value of the sports performance vector is calculated using the causal weight corresponding to the sports performance variable Y Y actual and the weighted distance between the potential value Y’ Q as the causal surprise degree, and the calculation formula is:

[0016] where y i represents the i th sports performance variable, Y represents the set of sports performance variables, represents the normalized weight of the i th sports performance variable in the global causal weight vector W XTY , represents the potential value of the i th sports performance variable, represents the actual value of the i th sports performance variable.

[0017] Further, after the structural causal model is enhanced in S5, the new model is stored in the case library, and the original old model is archived. When the new model performs poorly, the old model is rolled back.

[0018] Via the technical solution, compared with the prior art, the application has the following beneficial effects: The application forms an intelligent closed-loop process capable of learning from experience, co-evolving with experts and constantly improving its own causal model from knowledge discovery of causal reconstruction, to precise matching of causal retrieval, to generation of a quantitative scheme of causal reuse, and finally through causal modification and causal reservation. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0020] Fig. 1 A flowchart of the causal knowledge driven human-machine collaborative personalized sports training case reasoning method provided in the embodiments of the present application; Fig. 2 A causal case reasoning architecture diagram provided in the embodiments of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] As shown in the drawings, Figs. 1-2 The embodiments of the present application disclose a causal knowledge driven human-machine collaborative personalized sports training case reasoning method, comprising the following steps: S1, causal reconstruction: constructing a case library containing a plurality of cases, the observation data of each case being composed of description variables, training scheme variables and sports performance variables; based on all observation data of the case library, different levels of causal knowledge reconstruction are performed; S2. Causal Retrieval: On the inverse causal graph of descriptive variables and sports performance variables, the weight of each descriptive variable is calculated using the PageRank algorithm, and the case most similar to the target case is retrieved from the case library based on the weight of each descriptive variable. S3. Causal reuse: Identify the differences in training scheme variables between the target case and the retrieved similar cases, perform counterfactual intervention and inference on the target case using arbitrary training scheme variables, and select one or more new cases; S4. Causal Modification: Compare the actual athletic performance after the new case is executed with the original athletic performance. When the two do not match, receive feedback from the coach, execute the attribution process, and generate different handling suggestions based on the attribution type. S5. Causal Preservation: When the execution result of a new case does not match the original motion performance, the new case is stored in the case library, and causal knowledge is reconstructed periodically.

[0023] The following is a further explanation of each of the above steps.

[0024] S1 Causal Reconstruction: The task of the causal reconstruction step is to generate causal knowledge from the case library.

[0025] Specifically, each case is defined as a triple (X, T, Y), where X is a set containing m descriptive variables, denoted as X = {x1, x2, ..., x...}. m These descriptive variables describe the background, environment, or state of the subject, such as the athlete's age, fatigue level, and weather conditions.

[0026] T is a set containing n training scheme variables, denoted as T = {t1, t2, ..., t}. n These training program variables are the interventions or actions implemented, such as training duration, number of strength training sets, and type of tactical drills. In causal inference, this is usually referred to as treatment.

[0027] Y is a set containing k athletic performance variables, represented as Y = {y1, y2, ..., y...} k These athletic performance scalars are the results or outputs that need to be observed, such as shooting percentage, 100-meter sprint time, and game points. In causal inference, this is usually referred to as the outcome. Case Library CB For a containing p A set of cases, denoted as CB = {( X 1, T 1, Y 1), ( X2, T 2, Y 2), ..., ( X p , T p , Y p Let V = X∪T∪Y be the set of all variables.

[0028] The overall goal of causal reconstruction is to reconstruct a three-level causal graph from the observational data of the case library CB, in order to obtain causal path knowledge at different levels. The specific process includes: Determine the set of variables V XY = X∪Y, in the variable set V XY The causal discovery algorithm is executed to construct a description-representation causal graph. G XY = ( V XY , E XY This causal diagram focuses on the relationship between intrinsic states and outcomes, revealing descriptive variables. X With athletic performance Y The causal path between them, ignoring the training scheme for now. T The impact provides knowledge about "how the system operates in the absence of specific interventions?". For example, it might reveal "fatigue levels (...)". x i This will directly affect shooting percentage. y j )decline".

[0029] Causal discovery algorithms can be PC or FCI, and the process includes: based on CB The data underwent a series of conditional independence tests. A ⊥ B | S ),in A , B ∈ V XY , S V XY \ { A , B Construct a causal framework and orient edges to produce a directed graph. G XY = ( V XY , E XY ).

[0030] Determine the set of variables VTY = T ∪ Y In the variable set V TY The causal discovery algorithm is executed to construct the training-performance causal graph. G TY = ( V TY , E TY This causal diagram focuses on the correlation between interventions and outcomes, revealing the training program. T With athletic performance Y The causal path between them, temporarily ignoring the athlete's initial state. X The differences provide a preliminary answer to the question, "Are different training programs effective?" For example, it might show that "increasing the duration of tactical drills ( t i ) and game score ( y k "Positive correlation." But this relationship may be spurious.

[0031] Determine the set of variables V XTY = X ∪ T ∪ Y In the variable set V XTY The causal discovery algorithm is executed to construct a description-training-representation causal graph. G XTY = ( V XTY , E XTY This cause-effect graph is a global, comprehensive view that includes all variables. X , T , Y A unified causal model is proposed to reveal the complex interaction pathways among these factors, including confounding and mediation effects. This diagram can verify or revise [the model / model]. G XY and G TY The findings. For example, if in G TY There are edges in t i → y j But G XTY The edge in the middle disappeared, and a path was replaced. t i ← xk → y j Then a confounding variable was identified. x k Discovering mediating effects: for example, graphs G XTY May reveal the path x i → t j → y k This means describing variables x i By influencing the choice of training program t j This, in turn, affects the final performance. y k .

[0032] S2. Causal Retrieval: The task of the causal retrieval stage is to retrieve the most similar case to the input target case from the case library based on causal knowledge. Unlike the association-based similarity measure of traditional CBR, "causal retrieval" uses the causal knowledge learned in the causal reconstruction stage to determine the weights of descriptive feature variables on the "reverse causal graph" using the PageRank algorithm. This results in root features that influence athletic performance through multiple mediating paths having higher weights in the similarity measure, while the case similarity is independent of variables that have no impact on athletic performance. For example, sprinting ability may have a greater impact on a soccer striker's goal tally, but a smaller impact on a goalkeeper's save percentage.

[0033] Specifically, this invention provides a method for determining feature weights based on a reverse causal graph and PageRank. In the original causal graph, edges point from cause to effect; in the reverse graph, edges point from effect back to cause. Therefore, a feature node that serves as a common cause for multiple performance indicators becomes an authoritative node in the reverse graph, attracting multiple incoming chains. By calculating the centrality score of this node using the PageRank algorithm, its causal importance can be quantified, thereby generating objective feature weights that contain causal knowledge. The specific implementation steps are as follows: 1) Obtain or construct a directed acyclic graph (DAG) representing the causal relationships between multiple feature variables, denoted as a causal graph. G =( V , E In this invention, this diagram is the description-representation causal graph constructed earlier. G XY ,in V Descriptive features X and performance characteristics Y The set of nodes, EA set of directed edges representing direct causal relationships.

[0034] The reverse causal graph is obtained by reversing the direction of all directed edges in the original causal graph, in which the original root cause nodes become terminal nodes of influence convergence.

[0035] 2) On the reverse causal graph, a PageRank score value is generated for each descriptive variable node as the weight of the descriptive variable node using the PageRank algorithm; specifically, the weights of all descriptive variable nodes are initialized to equal values, and the weights of the descriptive variable nodes are updated according to the following iteration formula until the weight values of all descriptive variable nodes converge: The weight iteration update formula is:

[0036] wherein, is the descriptive variable node x i The weight of the descriptive variable node k in the first iteration; x i The weight of the descriptive variable node k in the first d iteration; G is a damping factor; V R is the reverse causal graph; x is the total number of nodes in the reverse causal graph; In( G i , G R ) is a set of nodes in the reverse causal graph x R pointing to the node x i ; Out( G i , x R ) is the out-degree of the node G i in the reverse causal graph R .

[0037] X 3) The PageRank score value of each descriptive feature node x = { x 1, x 2,..., n} after convergence is taken as the maximum weight of the corresponding feature, forming an n-dimensional feature weight vector W = ( w 1, w2,..., w n ). The vector can be normalized.

[0038] 4) Calculate the distance between any case in the target case and the case base based on the weight of each description variable node, and the specific calculation formula is:

[0039] wherein, d w Q , C j ) represents the weighted distance between the target case Q and the jth case C j in the case base; i represents the ith description vector node, m represents that a case contains m description vector nodes, w i represents the weight of the ith description vector node, q i represents the feature value of the ith description vector in the target case, and C ji represents the feature value of the ith description vector of the jth case in the case base.

[0040] 5) Sort all cases in the case base according to the distance from the target case from small to large, and select the first k cases as the retrieval result, and k can be specified by the coach, and the default is 1.

[0041] S3, Causal reuse: The core task of the causal reuse link is to diagnose and optimize the training scheme for the current athlete. In this link, an innovative method based on Granular Counterfactual Inference is introduced, which aims to answer the core question: "For the current athlete, if a different training scheme is adopted, to what extent can the athletic performance be improved?"

[0042] This method uses the constructed training-performance causal graph G TY and its corresponding structural causal model (Structural Causal Model, SCM) as the calculation basis. Unlike traditional CBR, which directly reuses the entire solution, the present application regards the training scheme T j of the retrieved similar case C Cj as a set of improvement inspirations. By identifying the target case T Q and the similar case T Cj ​By analyzing the differences in training variables, counterfactual interventions can be constructed and evaluated for any subset of variables. This means the system can explore various refined "hybrid schemes," such as "changing only training intensity," "changing only training frequency," or "changing both intensity and frequency," and predict the improvement of each scheme relative to current known performance. The implementation steps of this process are as follows: S30. Obtain the target cases for current diagnosis and optimization. Q =( X Q , T Q , Y Q ),in X Q For the athlete's condition, T Q For the current training scheme, Y Q This represents the observed motion performance under this scheme. Simultaneously, it retrieves the k most similar cases output from the "causal retrieval" step. C 1, ..., C k Loading the trained-performance causal model already learned in the causal reconstruction phase. M .

[0043] For each retrieved similar case C j (j=1, ..., k) Perform the following S31-S33 operations: S31. Identify discrepancy variables: Compare the training scheme variables of the target case and any similar case, identify the training scheme variables with different values, and construct a set of discrepancy variables Δ. T j = { t i | t i ∈ T and t i,Q ≠ t i,Cj}; S32. Generate intervention subsets: Calculate the set of differential variables Δ T j power set P (Δ T j (excluding the empty set), power set P (Δ T j Each element S in ) is a subset of one or more differential training scheme variables, representing a potential modification scheme; S33. Constructing Counterfactual Interventions: Targeting Power Sets P (Δ T j For each element S in the equation, construct a counterfactual intervention: modify the training scheme variables belonging to element S in the target case to similar cases. C j The values ​​of the corresponding training scheme variables are kept constant, and the values ​​of other training scheme variables that are not part of element S in the target case remain unchanged. For example, if T Q ={ t 1=A, t 2=B, t 3=C}, T Cj ={ t 1=X, t 2=Y, t If 3 = C}, then Δ T j ={ t 1, t 2}. The generated interventions include: ① Modification t 1(do( t 1=X); ② Modify t 2(do( t 2=Y)); ③ Simultaneously modify t 1 and t 2(do( t (t1=X, t2=Y). By performing this operation on all k similar cases, the system will generate a large and diverse set of counterfactual interventions.

[0044] S34. Perform S31-S33 on all similar cases to generate a set of counterfactual interventions.

[0045] S35. Apply each counterfactual intervention from the counterfactual intervention set to the individual causal model corresponding to the target case, perform counterfactual inference, and calculate the latent values ​​of the sports performance variables under different counterfactual interventions; the counterfactual inference process is a standard three-step process: (a) Abduction: This step only needs to be performed once for the target case Q, using all the observed facts of the target case Q. X Q , T Q , Y Q ), through structural causal model M Inverse inference of exogenous noise variables representing inherent athlete traits U QThe posterior distribution or expected value. This step utilizes the most complete information to complete an accurate model of the individual athlete, explaining why this athlete, in this state, received this training, and achieved this result.

[0046] (b) Action: For the original model M Apply the counterfactual intervention currently being evaluated. For example, for the intervention do( t 1=X, t If 2=Y), then the model will determine t 1 and t The equations in equation 2 are replaced with the following: t 1 = X and t 2 = Y, resulting in a modified model. M' .

[0047] (c) Prediction: In the modified model M' In this process, the individual traits obtained in step (a) are used. U Q Descriptive variables of the target case X Q The new athletic performance is calculated in a positive direction. Y The expected value. This expected value is the potential value under this specific intervention. Y' Q .

[0048] S36. Summarize all potential performances under counterfactual interventions into a list, each entry containing: a complete candidate training scheme (i.e. T Q The modified version after a subset), the latent value of the athletic performance variable, and the improvement Δ relative to the original athletic performance variable value. Y = Y' Q - Y Q According to the lifting amount Δ Y Sort the list in descending order (e.g., shorter time, improved score rate, etc.).

[0049] Next, the ranked list of candidate solutions is presented to the coaches. This list clearly shows the various possible fine-tuning options and their quantifiable expected effects, and is presented in comparison with the current performance, making it more valuable for decision-making.

[0050] For example, assuming the current 100-meter sprint time is 10.5 seconds, the system might output: "Current performance: 10.5 seconds." Based on this, the following optimization suggestions are proposed: 1. Plan A (only increase strength training weight from 80kg to 100kg): expected time 10.4s (improve 0.1s); 2. Plan B (only increase flexibility training time by 30 minutes): expected time 10.45s (improve 0.05s); 3. Plan C (adjust both strength training weight and flexibility training): expected time 10.42s (improve 0.08s)...

[0051] Based on these information, the coach can choose one or more optimal plans as new cases to be adopted, and complete the human-computer collaborative decision-making according to his own professional knowledge and competition goals.

[0052] S4, Causal modification: the causal modification link is the core of the human-computer collaborative knowledge evolution of the present application, its task is to attribute the difference between the system prediction and the actual result after the new plan is executed and actual results are generated, and to modify the underlying causal knowledge in collaboration with the coach. This link turns "failure" or "unexpected" cases into valuable opportunities for system learning and evolution, rather than simply regarding them as errors. The core idea is: when the prediction does not match the reality, there must be a reason that has not been captured by the model. Specifically, it includes: S41, execute the new case generated by S3 to obtain new complete case data C new = ( X actual , T final , Y actual ). Wherein Y actual is the real sports performance of the athlete after implementing the new case; compare the actual value of the athlete's sports performance variable after executing the new case Y actual with the original sports performance Y Q , if Y actual is worse than Y Q , it is determined as a failure; if no failure occurs, calculate the causal unexpectedness.

[0053] This step aims to quantify the "abnormal" degree of the new case execution result, which is divided into "failure" and "unexpected" two cases: (a) Failure judgment: first, compare the actual performance of the new plan Y actual with the original performance before adopting the plan Y Q . If Yactual Worse than Y Q If the performance is worse than expected (e.g. sprint time increases, scoring rate decreases), it is determined as a failure, and the failure case is a high-value learning signal. The system will directly mark this case and enter S42 for attribution.

[0054] (b) Causal Surprise Degree calculation: If no failure occurs (i.e. Y actual ≥ Y Q , the accuracy of the prediction will be further evaluated, and the "Causal Surprise Degree" will be calculated. This measure uses the full-variable causal graph G XTY to calculate the causal weight of each variable. The specific calculation process is as follows: Reverse all directed edges in the description-training-performance causal graph to obtain the corresponding full-variable reverse causal graph; Run the PageRank algorithm on the full-variable reverse causal graph to calculate a global causal weight vector X , T , Y for all variables W XTY The reason for calculating the global causal weight vector for all variables is to globally evaluate the causal similarity between the potential value and the actual value. The specific calculation process is the same as the feature weight determination method based on reverse causal graph and PageRank proposed in the causal retrieval part of S2. Each variable in X, T, and Y has a weight value, which together constitutes the W XTY weight vector.

[0055] Use the causal weight corresponding to the sports performance variable Y to calculate the weighted distance between the actual value Y actual and the potential value Y’ Q of the sports performance vector as the causal surprise degree, and the calculation formula is:

[0056] Where y i represents the i-th sports performance variable, Y represents the set of sports performance variables, represents the normalized weight of the i-th sports performance variable in the global causal weight vector W XTY , and the normalization standard is that the total Y variable weight w yi sum is 1, represents the potential value of the i-th sports performance variable, This represents the actual value of the i-th performance variable. Y has more than one variable; for example, in a 20km race walk, performance Y might include two variables: the first half time and the second half time. The first half time influences the second half time. When there are multiple Y variables, the sum of the weights of all performance variables in Y is normalized to 1. In other words, if there is only one Y variable, no weighting is needed.

[0057] The calculated S D With preset threshold θ surprise Compare. If S D >θ surprise If the result is unexpected, it means that the existing causal structure model failed to accurately predict the outcome and attribution is required.

[0058] S42. When a failure or causal unexpectedness exceeds the threshold, a structured interface or guided question is provided, and error attribution is performed based on the received coach feedback; the attribution type is divided into incidental situational factors or structural model defects.

[0059] Among them, incidental situational factors refer to the coach's observation of a temporary, unmodeled variable. X The factors included (such as "poor sleep") suggest that this bias was caused by randomness or a specific situation, and the structural causal model itself may not be flawed.

[0060] The structural model flaw refers to the coach's decision to exclude all incidental factors and the assumption that the model may underestimate the negative impact of short-term neural fatigue following high-intensity strength training on explosive power. This directly points to the shortcomings of the current causal diagram. G XTY Or the structural equation parameters may be defective.

[0061] Below is an example of a human-computer collaborative attribution dialogue: The system outputs: "Warning: A significant unexpected event has been detected. For the plan 'increasing the strength load to 100kg,' the system predicted that the 100-meter time would increase to 10.4 seconds, but the actual observed time was 10.48 seconds. The unexpectedness exceeds the threshold. Please assist in attributing the cause." The system outputs: "During training or testing, were there any of the following unrecorded special circumstances? 1. Athlete's physical discomfort (such as a cold, muscle soreness)? 2. Abnormal venue or weather conditions? 3. Fluctuations in mental state? 4. Other special events?"

[0062] The coach typed: "The athlete reported poor sleep quality last night."

[0063] S43. Based on the attribution type, generate different treatment suggestions, and for incidental situational factors, apply them to new cases. Cnew Cases marked as contextually anomalous will be stored in the case library during subsequent causal retention, but may be assigned a lower weight or excluded during model retraining to avoid noise interference with the model. To address structural model deficiencies, the received expert knowledge from the coach is transformed into one or more specific modification hypotheses for the structural causal model, such as: "It is recommended to modify the causal graph..." G XTY "Add a negative causal edge from t_strength_load to y_start_reaction, or introduce a new mediator variable 'neural fatigue'", the system presents this hypothesis to the coach and requests confirmation.

[0064] S5. Causal Preservation: The causal preservation stage is responsible for solidifying new knowledge and experience, enabling the long-term evolution of the system. This stage ensures that each reasoning cycle provides a richer and more accurate foundation for the next. It not only preserves cases, but more importantly, it dynamically updates the causal graph of the entire system based on the conclusions of the "causal modification" stage. Specifically, it includes: S51. Case Enhancement and Conditional Storage: Case enhancement and conditional storage: This step is conditional. First, the evaluation results of the causal modification step are checked: the new solution is judged to have failed if and only if its causal unexpectedness exceeds a preset threshold θ. surprise This step is performed only when the time is right.

[0065] Execute the action: If the conditions are met, the new case will be... C new = ( X actual , T final , Y actual The case is then enhanced by attaching metadata (annotations) obtained during the collaborative attribution process. This enhanced case is then stored in the case library. CB .

[0066] Skip Operation: If the new solution does not lead to failure and the causal unexpectedness does not reach the threshold, it indicates that the existing structural causal model is sufficient to explain the outcome of this intervention. The system will not perform case storage to avoid case library redundancy and maintain model simplicity, and subsequent steps S52, S53, and S54 will also be skipped.

[0067] S52, Structural Causal Model Update: This step is triggered only after case storage was performed in S51. The operation is executed based on the conclusion of the causal modification step: If the cause is attributed to incidental situational factors, the existing cause-and-effect diagram will not be modified. GXTY and structural causal models M , maintaining stability of current knowledge.

[0068] If the cause is attributed to structural model defects, and the coach has confirmed the modification hypothesis, the model update operation will be performed. This may include: structural update: according to the coach's confirmed hypothesis, add or delete edges in the causal graph G XTY , and then re-estimate the parameters of the structural causal model based on the updated graph structure and the case base containing new cases M .

[0069] Parameter update: if the causal graph structure is unchanged, only the parameter estimation is inaccurate, new cases C new can be used to update the relevant parameters of the model M online (for example, using Bayesian update or gradient descent).

[0070] S53, model version control and archiving: in order to ensure the robustness and traceability of the system, after any structural modification of the structural causal model, the old model will not be directly overwritten. Instead, it will archive the old model and save the new model as G XTY_v1.1 . This version control mechanism has the following advantages: it can track the evolution history of system knowledge; when the updated model performs poorly, it can quickly roll back to the previous stable version. Different versions of the model can be compared and analyzed to evaluate the learning effect.

[0071] S54, periodic causal reconstruction: in addition to immediate updates triggered by "failure" or "unexpected" cases, the system should also have a periodic or case base growth-based trigger. For example, every time the case base is increased by 100 cases, the system can automatically re-execute a complete causal reconstruction process to discover more macroscopic and fundamental structural changes that may occur in the data distribution.

[0072] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the method part.

[0073] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A causal knowledge-driven human-computer collaborative personalized exercise training case reasoning method, characterized in that, Includes the following steps: S1. Construct a case library containing multiple cases, where the observation data for each case consists of descriptive variables, training scheme variables, and athletic performance variables; based on all the observation data in the case library, perform causal knowledge reconstruction at different levels. S2. On the inverse causal graph of descriptive variables and sports performance variables, the weight of each descriptive variable is calculated using the PageRank algorithm, and the case most similar to the target case is retrieved from the case library based on the weight of each descriptive variable. S3. Identify the differences in training scheme variables between the target case and the retrieved similar cases, perform counterfactual intervention and inference on the target case using arbitrary training scheme variables, and select one or more new cases; S4. Compare the actual athletic performance generated after the new case is executed with the original athletic performance. When the two do not match, receive feedback from the coach, execute the attribution process, and generate different handling suggestions based on the attribution type. S5. When the execution result of a new case does not match the original motion performance, the new case is stored in the case library and causal knowledge is reconstructed periodically.

2. The causal knowledge-driven human-machine collaborative personalized sports training case reasoning method as described in claim 1, characterized in that, In S1, each case is defined as a triple (X, T, Y), where X is a set containing m descriptive variables, denoted as X = {x1, x2, ..., x...}. m These descriptive variables describe the case background, environment, or the state of the subject. T is a set containing n training scheme variables, denoted as T = {t1, t2, ..., t}. n These training program variables are the interventions or actions implemented; Y is a set containing k athletic performance variables, represented as Y = {y1, y2, ..., y...} k These scalar metrics of athletic performance are the results or outputs that need to be observed; Case Library CB For a containing p A set of cases, denoted as CB = {( X 1, T 1, Y 1), ( X 2, T 2, Y 2), ..., ( X p , T p , Y p )}.

3. The causal knowledge-driven human-computer collaborative personalized exercise training case reasoning method as described in claim 2, characterized in that, In S1, the process of reconstructing causal knowledge at different levels includes: Determine the set of variables V XY = X∪Y, in the variable set V XY The causal discovery algorithm is executed to construct a description-representation causal graph; Determine the set of variables V TY = T ∪ Y In the variable set V TY The causal discovery algorithm is executed to construct the training-performance causal graph; Determine the set of variables V XTY = X ∪ T ∪ Y In the variable set V XTY The causal discovery algorithm is executed to construct a description-training-performance causal graph.

4. The causal knowledge-driven human-computer collaborative personalized sports training case reasoning method as described in claim 3, characterized in that, S2 include: Reverse the direction of all directed edges in the description-representation causal graph to obtain the reverse causal graph; In the reverse causal graph, the PageRank algorithm is used to generate a PageRank score for each descriptive variable node, which serves as the weight of the descriptive variable node. The distance between the target case and any case in the case library is calculated based on the weights of each descriptive variable node. Sort all cases in the case library in ascending order of their distance from the target case, and select the top k cases as the search results.

5. The causal knowledge-driven human-computer collaborative personalized sports training case reasoning method as described in claim 4, characterized in that, The formula for calculating the distance between the target case and any case in the case library is: in, d w ( Q C j ) represents the target case Q and the j-th case C in the case library. j The weighted distance between them; i represents the i-th description vector node, m represents the total number of description vector nodes in a case, w i q represents the weight of the i-th description vector node. i Let C represent the feature value of the i-th descriptive vector in the target case. ji This represents the feature value of the i-th descriptive vector of the j-th case in the case library.

6. The causal knowledge-driven human-machine collaborative personalized sports training case reasoning method as described in claim 4, characterized in that, The calculation process for the weights of each descriptive variable node using the PageRank algorithm is as follows: Initialize the weights of all descriptive variable nodes to equal values, and update the weights of each descriptive variable node according to the following iterative formula until the weight values ​​of all descriptive variable nodes converge. The formula for weight iteration update is as follows: in, To describe variable nodes x i In the k Weights for +1 iteration; To describe variable nodes x i In the k The weights for the next iteration; d It is the damping factor; G R This is a reverse cause-effect graph; | V | represents the total number of nodes in the reverse causal graph; In( x i , G R (in the reverse causal graph) G R Middle pointer node x i The set of nodes; Out( x i , G R ) is a node x i In the reverse causality diagram G R The degree of exit in.

7. The causal knowledge-driven human-computer collaborative personalized sports training case reasoning method as described in claim 1, characterized in that, S3 include: S31. Compare the training scheme variables of the target case and any similar case, identify the training scheme variables with different values, and construct a set of difference variables Δ. T j ; S32. Calculate the set of differential variables Δ T j power set P (Δ T j ), power set P (Δ T j Each element S in ) is a subset of one or more differential training scheme variables, representing a potential modification scheme; S33. For each element S in the power set, construct a counterfactual intervention, which is to modify the training scheme variable that belongs to element S in the target case to the value of the corresponding training scheme variable in the similar case, while keeping the values ​​of other training scheme variables that do not belong to element S in the target case unchanged. S34. Perform S31-S33 on all similar cases to generate a set of counterfactual interventions; S35. Apply each counterfactual intervention in the counterfactual intervention set to the individual causal model corresponding to the target case, perform counterfactual inference, and calculate the latent values ​​of the motion performance variables under different counterfactual interventions; S36. Compile all potential performances under counterfactual interventions into a list, with each entry containing: a complete candidate training program, the potential value of the performance variable, and the improvement relative to the original performance variable value; sort the list in descending order based on the improvement and select one or more new cases from it.

8. The causal knowledge-driven human-computer collaborative personalized sports training case reasoning method as described in claim 3, characterized in that, S4 include: S41. The actual values ​​of the athlete's performance variables after implementing the new case. Y actual Compared with original movement performance Y Q Compare, if Y actual Inferior Y Q If no failure occurs, it is considered a failure; if no failure occurs, the causal unexpectedness is calculated. S42. When a failure or causal unexpectedness exceeds the threshold, a structured interface or guided question is provided, and error attribution is performed based on the feedback received from the coach. The attribution type is divided into incidental situational factors or structural model defects. S43. Based on the attribution type, generate different treatment suggestions. For incidental situational factors, mark new cases as situational anomalous cases. For structural model defects, transform the received expert knowledge from the coach into one or more specific modification hypotheses for the structural causal model.

9. The causal knowledge-driven human-machine collaborative personalized sports training case reasoning method as described in claim 8, characterized in that, In S41, the calculation process for the degree of causal unexpectedness includes: Reverse all directed edges in the description-training-representation causal graph to obtain the corresponding total variable reverse causal graph; Run the PageRank algorithm on the total variable inverse causal graph for all variables ( X , T , Y Calculate a global causal weight vector. W XTY ; The actual value of the motion performance vector is calculated using the causal weights corresponding to the motion performance variable Y. Y actual With potential value Y’ Q The weighted distance between them, as the causal unexpectedness, is calculated using the following formula: Among them, y i Let Y represent the i-th athletic performance variable, and let Y represent the set of athletic performance variables. This represents the i-th performance variable in the global causal weight vector. W XTY Normalized weights in This represents the latent value of the i-th athletic performance variable. This represents the actual value of the i-th performance variable.

10. The causal knowledge-driven human-machine collaborative personalized sports training case reasoning method as described in claim 8, characterized in that, In S5, after the structural causal model is enhanced, it is stored as a new model in the case library, while the original old model is archived. When the new model performs poorly, it is rolled back to the previous old model.

Images