An intelligent system evaluation method based on feedback feature extraction and causal inversion identification
Through closed-loop feedback feature extraction and causal inversion data identification, the problem of information overload and causal influencing in the unmanned intelligent system is solved, which significantly improves the efficiency and accuracy of decision-making evaluation and reduces prediction errors.
Patent Information
- Application Number
- CN202411314096.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-20
AI Technical Summary
Unmanned autonomous intelligent systems are prone to falling into the problems of information overload and inconsistent causal relationships when processing large amounts of data, resulting in low efficiency and poor accuracy in decision-making evaluation.
Using a method based on closed-loop feedback feature extraction and causal inversion data identification, key features are identified inversely through the load matrix, the data volume requirement is reduced, and inconsistent data is identified and eliminated through multiple BP neural network models to improve decision evaluation efficiency and accuracy.
The number of input features is effectively reduced, the efficiency of decision-making evaluation is improved, and the decision-making accuracy is improved. The model prediction error is reduced by 13.24%.
Smart Images

Figure CN118821874B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of decision-making evaluation, and in particular to an intelligent system evaluation method based on feedback feature extraction and causal inversion identification. Background Art
[0002] As a typical autonomous intelligent system, unmanned systems are gradually becoming an important tool in all walks of life driven by information network systems and artificial intelligence technologies. With the continuous advancement of science and technology, the trend of unmanned and intelligent is changing the way of operation in many fields, especially in logistics, agriculture, monitoring, rescue and other aspects. Building efficient unmanned systems has become a key factor in promoting industry innovation. As a new working mode, the unmanned cluster system consists of multiple simple unmanned platforms and auxiliary units. It can demonstrate complex operating capabilities and a high degree of autonomy through information exchange and collaborative operations. This clustering method not only makes the operation more three-dimensional and diversified, but also significantly improves work efficiency and adaptability, allowing unmanned systems to perform tasks more efficiently and accurately.
[0003] In the modern complex and fast-paced working environment, unmanned systems need to effectively filter and process key information to support timely decision-making and action. However, with the advancement of data acquisition technology and the development of sensor networks, the amount of data that unmanned systems can obtain has increased rapidly. These data may cover a variety of features, such as time series, spatial data, text data, image data, etc. Each data type may have multiple dimensions, which increases the complexity of information management and analysis. As a result, unmanned systems may fall into a state of information overload when processing large amounts of data, that is, the dimensional overload problem.
[0004] Secondly, inconsistent causal relationships in data are also a major challenge, especially in complex and dynamic environments. The data collected by unmanned systems often come from multiple sources and multiple sensors. These data may be affected by multiple factors, such as environmental changes, sensor errors, data loss, etc., which makes the causal relationships shown in the data deviate or inconsistent from the actual situation or system requirements, making it difficult to accurately capture and analyze the causal relationships between data. Therefore, even with abundant data, unmanned systems face the challenge of how to extract effective information from them and make accurate judgments. Summary of the invention
[0005] In response to the challenges of dimensional overload and inconsistent data causal relationships in the decision-making evaluation of unmanned autonomous intelligent systems. The present invention proposes a method based on closed-loop feedback feature extraction and causal inversion data identification. On the one hand, through closed-loop feedback based on the load matrix, the new simplified features are mapped back to the original feature space, and only a small number of key features are retained to reduce the data volume required for the decision-making evaluation of the unmanned autonomous intelligent system, thereby greatly improving the decision-making evaluation efficiency. On the other hand, through the causal inversion data identification method based on data label consistency, a small number of causal inversion data can be identified and excluded, thereby improving the decision-making evaluation accuracy of the unmanned autonomous intelligent system.
[0006] The present invention first transforms the original N-dimensional data into new N-dimensional data using dimensionality reduction technology, selects M key dimensions from the new N-dimensional data, and uses the load matrix or correlation coefficient matrix to normalize and reversely identify F original key dimensions. Then, multiple samples are extracted from the complete data set; the BP neural network is used as the baseline modeling method, and a sub-model is constructed for each sample; the error is calculated in each model to form an error column; the error column is clustered; the cluster core is compared with the threshold for selection, and the class above the threshold is eliminated, otherwise it is retained. Finally, (1) the complete data set; (2) the data set after closed-loop feedback feature extraction; (3) the data set after causal inversion data identification; (4) the data set after closed-loop feedback feature extraction and causal inversion data identification are used as training sets to establish a BP neural network model independent of the above-mentioned BP sub-model, and the absolute error on the test set is calculated to verify the effectiveness of the method.
[0007] The present invention proposes an autonomous intelligent system decision-making evaluation method based on closed-loop feedback feature extraction and causal inversion data identification, which includes the following steps:
[0008] (1) Through UAV field and simulation tests, we collected an original training set D(P×N) with N input features and P groups of data, including features such as recognition and detection capabilities, and a test set T(Q×N) with N input features and Q groups of data.
[0009] (2) Use dimensionality reduction techniques to identify dimensionality reduction features in the original training set D(P×N), and then reversely identify the key features of the original training data set through the load matrix Ω. After closed-loop feedback feature extraction, the original training set D(P×N) is simplified to include only F key features, where F< <N。
[0010] (3) Randomly sample from the original training data set D(P×N) to generate multiple sub-training sets, build multiple training BP sub-models and use these sub-training sets for training. Each sub-model obtains its prediction evaluation results and generates errors for the data in each sub-training set, and then compares these errors to identify and eliminate abnormal data. After causal inversion data identification, the original training set D(P×N) is simplified to include only S groups of data, where S< <P。
[0011] (4) The original training set D(P×N) is simplified to D(S×F) with F key features and S groups of data. A test BP model is constructed and trained using the simplified training set to obtain the evaluation results of the intelligent system.
[0012] (5) The trained test BP model predicts and verifies the evaluation results of each set of test data in the test set.
[0013] Preferably, step (2) is specifically:
[0014] (2.1) Given the original training set D(P×N) denoted as D, define an N×P matrix B(N×P) such that:
[0015] (1)
[0016] Transform so that the mean of each column in B is 0, then: (2)
[0017] in, is the covariance matrix of D, that is Equal to the covariance matrix of D , the principal component of D is The eigenvector of B is decomposed into singular values to obtain:
[0018] (3)
[0019] Among them, U is an orthogonal matrix, Λ is a diagonal matrix, and V is also an orthogonal matrix. Transform again and get:
[0020] (4)
[0021] That is to say, The rows in (columns in V) are a set of orthogonal bases that transform the matrix B into Z. The columns of the matrix V are The eigenvectors of , which are also the principal components of D.
[0022] (2.2) For the principal components of D calculated in step (2.1), that is, the covariance matrix , calculate the eigenvalues corresponding to each eigenvector and sort them, select F new features with eigenvalues greater than 1, and identify them as key new features, where F< <N。
[0023] (2.3) Load matrix Ω = ( ) reflects the The original features and Correlation between new features: (5)
[0024] in, and Respectively represent The original data Data and The new data data. and Respectively represent The average value of the P group of original data and the The average value of the P groups of new data, for any and , both have .
[0025] (2.4) Normalize the value of Ω to get |Ω| norm , in order to further identify the original key features.
[0026] (2.5) Use |Ω| norm To reversely identify the key original features. So far, F key features have been identified from N original features.
[0027] Preferably, step (3) is:
[0028] (3.1) A total of K samples are randomly extracted from the original training data set in a fixed ratio. Note that each sample should cover more than 80% of the entire P group data and K>30 to meet the minimum statistical requirements.
[0029] (3.2) K BP neural network sub-models constructed using K sample training.
[0030] (3.3) When the data of group p in the training set P is not included in the data of group Among the samples, that is, In the model, the error When the pth group of data is included in the Among the samples, that is, When the model is used, the error is calculated as follows:
[0031] (6)
[0032] in, represents the actual output of the pth group of data, Indicates The prediction output of the model for the pth group of data is: is the absolute value function.
[0033] Calculate the average error of the p-th group of data in all K samples:
[0034] (7)
[0035] in, express The number of Calculated in formula (6). So far, the errors of all p data form an error column .
[0036] (3.4) Form clusters from the error column E. Use a clustering algorithm to form clusters from the error column Clustering .
[0037] (3.5) Set the preset threshold With clustering The cores of each category in are compared. Cores above the threshold indicate that the relevant clusters should be removed, otherwise they are retained. So far, the S group data is retained in the P group original data.
[0038] Beneficial effects of the present invention:
[0039] 1. The feature extraction based on closed-loop feedback of the present invention is built on the traditional dimensionality reduction technology, breaking through the limitation that the features extracted after the traditional dimensionality reduction no longer have the physical meaning of the original feature space. First, the original features with physical meaning but related to each other are converted into new features without physical meaning but unrelated, and then through rigorous and unique analytical mathematical derivation, the new features are reversely mapped back to the original space, and quantitatively traced back to the original features with physical meaning, realizing a complete feedback cycle from "original features with physical meaning but related to each other" to "new features without physical meaning but unrelated" and then to "key original features with physical meaning".
[0040] Second, the present invention breaks through the inertial assumption that the default data conforms to a complete causal logical relationship in the decision-making evaluation of traditional unmanned autonomous intelligent systems, and proposes a causal inversion data identification method based on data label consistency. It generates multiple models to generate multiple errors for each set of data to achieve causal inversion data identification for inconsistency detection. These errors are then used to create multiple clusters, and their cores are juxtaposed relative to a predetermined threshold to determine whether the corresponding data is consistent, and thus determine whether it should be retained or excluded. This innovative causal inversion data identification mechanism is independent and does not require supplementary information. Compared with other subjective methods, it has theoretical advantages.
[0041] 3. This paper studies a case study of cooperative combat decision-making evaluation of drone swarms, and uses the method proposed in this paper and BPNN to build a model for comparative verification. The conclusions are as follows: (1) Closed-loop feedback feature extraction can effectively help reduce 43.75% of input features; (2) Causal inversion data identification can efficiently identify 17 groups of inconsistent data; (3) By screening out a compact structure that is only 53.86% of the original data, the model prediction error is reduced from the original MAE=0.1360 to MAE=0.1180, a reduction of 13.24%. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a flowchart of closed-loop feedback feature extraction in the method of the present invention;
[0043] Figure 2 It is a flowchart of the causal inversion data identification based on the method of the present invention;
[0044] Figure 3 It is a process diagram of inconsistent data identification in the case of UAV swarm collaborative combat decision-making evaluation;
[0045] Figure 4 This is a graph showing the results of verifying four data sets on the test set in the UAV swarm collaborative combat decision-making evaluation case. DETAILED DESCRIPTION
[0046] The following is a detailed description of the embodiments and test results of the inventive method in conjunction with the accompanying drawings:
[0047] A decision-making evaluation method for an autonomous intelligent system based on closed-loop feedback feature extraction and causal inversion data identification includes the following steps:
[0048] (1) Through a series of UAV field and simulation tests, we collected an original training set D(P×N) with N input features and P groups of data, including features such as recognition and detection capabilities, and a test set T(Q×N) with N input features and Q groups of data.
[0049] (2) Feature extraction based on closed-loop feedback. Dimensionality reduction technology is used to identify the reduced-dimensional features in the original training set D(P×N), and then the key features of the original training data set are reversely identified through the load matrix Ω. After closed-loop feedback feature extraction, the original training set D(P×N) is simplified to include only F key features, where F< <N。
[0050] (3) Data identification based on causal inversion. Random sampling is performed from the original training data set D(P×N) to generate multiple sub-training sets, and multiple training BP sub-models are constructed and trained using these sub-training sets. Each sub-model obtains its prediction evaluation results and generates errors for the data in each sub-training set, and then compares these errors to identify and eliminate abnormal data. After causal inversion data identification, the original training set D(P×N) is simplified to include only S groups of data, where S< <P。
[0051] (4) Construct a test BP model and train the model using the simplified training data set. After steps (2) and (3), the original training set D(P×N) is simplified to a compact structure, that is, D(S×F) with only F key features and S sets of data. D(S×F) is used to train the test BP model and obtain the evaluation results of the unmanned intelligent system.
[0052] (5) The trained test BP model predicts and verifies the evaluation results of each set of test data in the test set.
[0053] Preferably, step (2) is specifically:
[0054] (2.1) Transform the original features into new features. Given the original training set D(P×N) represented as D, define an N×P matrix B(N×P) to satisfy:
[0055] (1)
[0056] Transform so that the mean of each column in B is 0, then: (2)
[0057] in, is the covariance matrix of D, that is Equal to the covariance matrix of D , the principal component of D is The eigenvector of B is decomposed into singular values to obtain:
[0058] (3)
[0059] Among them, U is an orthogonal matrix, is a diagonal matrix, is also an orthogonal matrix. Transform again and we get:
[0060] (4)
[0061] That is to say, The rows in (columns in V) are a set of orthogonal bases that transform the matrix B into Z. The columns of the matrix V are The eigenvectors of , which are also the principal components of D.
[0062] (2.2) Identify key new features. For the principal components of D calculated in step (2.1), that is, the covariance matrix , calculate the eigenvalues corresponding to each eigenvector and sort them, select F new features with eigenvalues greater than 1, and identify them as key new features, where F< <N。
[0063] The specific process of calculating the eigenvalue corresponding to each eigenvector is: covariance matrix The eigenvector of , calculate the eigenvalue corresponding to each eigenvector , No. feature vector , according to the eigenvalue And the definition of eigenvectors, construct characteristic polynomials:
[0064] (5)
[0065] in, is the unit matrix, solve this polynomial and get the eigenvector The corresponding eigenvalue .
[0066] (2.3) Calculate the load matrix Ω. Load matrix Ω = ( ) reflects the The original features and Correlation between new features: (6)
[0067] in, and Respectively represent The original data Data and The new data data. and Respectively represent The average value of the P group of original data and the The average value of the P groups of new data, for any and , both have .
[0068] (2.4) Normalization. Since the load matrix is in the range of [-1, 1], it is necessary to normalize the value of Ω to obtain |Ω| norm , in order to further identify the original key features.
[0069] (2.5) Identify the original key features. For principal component analysis PCA, since each new feature is a linear combination of the original features, we can directly use |Ω| norm To reversely identify the key original features. So far, F key features have been identified from N original features.
[0070] The use of |Ω| norm The specific process of reverse identification of key original features is as follows: select |Ω| norm The original features whose absolute values are greater than the set threshold t are regarded as the original features that contribute most to the new features, that is, the key original features.
[0071] Preferably, step (3) is:
[0072] (3.1) Random sampling from the original training set D(P×N). A total of K samples are randomly extracted from the original training data set according to a fixed ratio. Note that each sample should cover more than 80% of the entire P group data and K>30 to meet the minimum statistical requirements.
[0073] (3.2) Use multiple sub-datasets to build multiple models. Use K samples to train the constructed K BP neural network sub-models. In this step, the specific method of machine learning model construction is not mandatory, as long as the constructed machine learning model has a relatively high accuracy.
[0074] (3.3) Calculate the error of each sub-model. Among the samples, that is, In the model, the error When the pth group of data is included in the Among the samples, that is, When the model is used, the error is calculated as follows:
[0075] (7)
[0076] in, represents the actual output of the pth group of data, Indicates The prediction output of the model for the pth group of data is: is the absolute value function.
[0077] Calculate the average error of the p-th group of data in all K samples:
[0078] (8)
[0079] in, express The number of Calculated in formula (7). So far, the errors of all p data form an error column .
[0080] (3.4) Form clusters from the error column E. Use a clustering algorithm to form clusters from the error column Clustering .
[0081] (3.5) Inconsistency test by comparing the threshold value with the kernel. With clustering The cores of each category in are compared. Cores above the threshold indicate that the relevant clusters should be removed, otherwise they are retained. So far, the S group data is retained in the P group original data.
[0082] Preferably, step (5) is specifically:
[0083] For the test set containing Q group data, calculate the mean absolute error for verification:
[0084] (9)
[0085] in, represents the actual output of the qth test set, Represents the predicted output of the qth group of data.
[0086] The flowchart of the closed-loop feedback feature extraction method of the present invention is as follows: Figure 1 As shown in the figure, the core part is: using dimensionality reduction technology to transform the original N-dimensional data into new N-dimensional data; selecting M key dimensions in the new N-dimensional data; using the load matrix or the correlation coefficient matrix, after normalization, reversely identifying F original key dimensions.
[0087] The flowchart of the method of the present invention based on causal inversion data identification is as follows: Figure 1 and Figure 2 As shown in the figure, the core part is: multiple sampling in the complete data set to extract multiple samples; construct a sub-model for each sample; calculate the error in each model to form an error column; cluster the error column; compare the clustering core with the threshold to select, and eliminate the class above the threshold, otherwise retain it.
[0088] The following is a detailed introduction on how to perform feature extraction based on closed-loop feedback in the present invention, and the specific steps are as follows.
[0089] 1. Taking a UAV swarm collaborative combat decision-making evaluation case as an example, we first establish a data set, in which the input is divided into four categories, namely stability capability, environmental adaptability, identification and detection capability, and data transmission capability. The stability capability includes: x1: battery life, x2: wind resistance coefficient; environmental adaptability includes: x3: high temperature resistance coefficient, x4: low temperature resistance coefficient, x5: waterproof coefficient; identification and detection capabilities include: x6: detection distance, x7: single-machine detection probability, x8: identification distance, x9: single-machine identification probability, x10: single-machine identification probability, x11: single-machine identification probability, x12: single-machine identification probability, x13: single-machine identification probability, x14: single-machine identification probability, x15: single-machine identification probability, x16: single-machine identification probability, x17: single-machine identification probability, x18: single-machine identification probability, x19: single-machine identification probability, x21: single-machine identification probability, x22: single-machine identification probability, x23: single-machine identification probability, x24: single-machine identification probability, x25: single-machine identification probability, x26 ...4: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: single-machine identification probability, x24: 10 : Maximum number of targets monitored simultaneously, x 11 : Response time; Data transmission capabilities include: x 12 : Transmission distance, x 13 : Transmission bandwidth, x 14 : Communication quality, x 15 : Communication rate, x 16 : Data delay; the output is the unmanned cluster collaborative reconnaissance capability. The data set contains 500 sets of data, 400 sets are randomly selected as training sets and 100 sets as test sets.
[0090] 2. After PCA transformation, the original 16-dimensional feature space is transformed into a new 16-dimensional feature space {V1-V 16}, and select 8 principal components in the new space according to the principle that the eigenvalue is greater than 1. Note that these 8 principal components are independent of each other, but have no clear physical meaning.
[0091] 3. Calculate the load matrix between the 8 principal components and the 16 original features. According to the values of the elements in the load matrix, select 9 key features from the original features, namely x2, x3, x5, x6, x7, x8, x 10 , x 11 , x 15 .
[0092] The following combination Figure 3 The present invention introduces in detail how to perform data identification based on causal inversion, and the specific steps are as follows.
[0093] 1. The present invention selects a 90% extraction ratio, extracts 90% of the data from the divided training set to form a training sub-data set, and repeats 50 times to generate 50 sub-data sets.
[0094] 2. Select BP neural network as the modeling method to build the model. Each sub-dataset has its own sub-model, thus generating 50 sub-models. Calculate the absolute error between the predicted output and the actual output of the 50 sub-models for their respective training sets. .in represents the true output of the pth group of data, Represents the predicted output of the k-th model for the p-th group of data.
[0095] 3. Absolute error calculation of 50 models The average absolute error is shown in formula (11):
[0096] (9)
[0097] The average error column of 400 groups of data was obtained. The error column was clustered using the knn algorithm, and the threshold v=0.27 was taken. The threshold was compared with the kernel of each cluster, and 17 groups of inconsistent data were finally determined.
[0098] Figure 4 The results of the validation on the test set are shown. Using the BP neural network as the baseline modeling method, the absolute errors of the models on the test set are calculated based on (1) the complete data set; (2) the data set after closed-loop feedback feature extraction; (3) the data set after causal inversion data identification; and (4) the data set after closed-loop feedback feature extraction and causal inversion data identification. Table 1 shows the comprehensive results of the data set before and after closed-loop feedback feature extraction and causal inversion data identification.
[0099] Table 1 Comparison results of reconnaissance capability prediction using BPNN
[0100]
[0101] according to Figure 4 As can be seen from Table 1, after the method proposed in the present invention is used in the UAV swarm collaborative combat decision-making evaluation case, 43.75% of the input information can be significantly reduced by closed-loop feedback feature extraction alone, and the MAE on the test set only increases from 0.1360 to 0.1388, an increase of 2.59%. Only through causal inversion data identification, 4.25% of inconsistent data can be identified and excluded, and the MAE on the test set can be reduced from 0.1360 to 0.1104, a decrease of 18.82%. Through closed-loop feedback feature extraction plus causal inversion data identification, the input information can be reduced by 46.14%, and the MAE on the test set decreases from 0.1360 to 0.1189, a decrease of 13.24%.
Claims
1. An intelligent system evaluation method based on feedback feature extraction and causal inversion identification, characterized in that: The following steps are involved: Step 1: Through UAV and simulation experiments, collect N input features including recognition and detection capability features, the original training set D(P×N) of P group data, and the test set T(Q×N) of Q group data; Step 2: Use dimensionality reduction technology to identify the dimensionality reduction features of the data in the original training set D (P × N), and then reversely identify the key features of the original training data set through the load matrix Ω. The original training set D (P × N) is simplified to include F key features; the specific implementation process is: Step 2.1, the original training set D (P × N) is represented as D, and an N × P matrix B is defined to satisfy: Transform so that the mean of each column in B is 0, then: Among them, C D is the covariance matrix of D, that is, B T B is equal to the covariance matrix C of D D , the principal component of D is C D The eigenvector of B is decomposed into singular values to obtain: B=UΛV T (3) Among them, U is an orthogonal matrix, Λ is a diagonal matrix, and V is also an orthogonal matrix; transform again and get: V T B T =U T Λ=Z (4) V T The rows in are a set of orthogonal bases that transform the matrix B into Z; the columns of the matrix V are C D The eigenvector of , which is the principal component of D; Step 2.2: For the principal components of D calculated in step 2.1, that is, the covariance matrix C D , calculate the eigenvalues corresponding to each eigenvector and sort them, select F new features with eigenvalues greater than 1, and identify them as key new features, where F< <N; Step 2.3: Loading matrix Reflects the correlation between the nth original feature and the nth' new feature: Among them, x k,n and They represent the kth data of the nth original data and the kth data of the n'th new data respectively; and They represent the average value of the P group of n-th original data and the average value of the P group of n'-th new data, respectively. For any n and n', there is δ∈(0,1]; Step 2.4, normalize the value of Ω to get |Ω| norm ; Step 2.
5. Use |Ω| norm Reversely identify the key original features and identify F key features from N original features; The use of |Ω| norm The specific process of reverse identification of key original features is as follows: select |Ω| norm The original features whose absolute values are greater than the set threshold t are regarded as the original features that contribute most to the new features, that is, the key original features; Step 3: Randomly sample from the original training data set D(P×N) to generate multiple sub-training sets, build multiple training BP sub-models and use the sub-training sets for training. Each sub-model obtains the prediction evaluation results and generates errors for the data in each sub-training set. The errors are compared to identify and eliminate abnormal data. The original training set D(P×N) is simplified to include S groups of data. The specific implementation process is as follows: Step 3.1, randomly extract K samples from the original training set D(P×N) in a fixed ratio; each sample should cover more than 80% of the entire P group data and K>30; Step 3.2, use K samples to train the constructed K BP neural network sub-models; Step 3.3, calculate the error of each sub-model. When the p-th group of data in the training set P is not included in the k-th sample, that is, in the k-th model, the error e p,k =0; when the pth group of data is included in the kth sample, that is, the kth model, the error calculation method is as follows: in, represents the actual output of the pth group of data, represents the predicted output of the k-th model for the p-th group of data, and abs(·) is the absolute value function; Calculate the average error of the p-th group of data in all K samples: Among them, N * Indicates e p,k ≠0, e p,k In formula (6), the errors of all p data form an error column E = {e p }; Step 3.4: Use clustering algorithm to extract the error column E = {e p } to form clusters C = {c1, c2, c3, ...}; Step 3.5: Compare the preset threshold v with the core of each category in cluster C; cores above the threshold indicate that the relevant clusters should be removed, otherwise they are retained, and S groups of data are retained in P group of original data, where S< <P; Step 4: The original training set D(P×N) is simplified to D(S×F) with F key features and S groups of data. A test BP model is constructed and trained using the simplified training set to obtain the evaluation results of the intelligent system. Step 5: The trained test BP model is predicted and verified using the test set.
2. The intelligent system evaluation method based on feedback feature extraction and causal inversion identification according to claim 1 is characterized in that: In step 2.2, the specific process of calculating the eigenvalue corresponding to each eigenvector is as follows: D The eigenvectors of are {α1...,α p ,...,α P }, calculate the eigenvalue λ corresponding to each eigenvector, the pth eigenvector α p , according to the eigenvalue λ p And the definition of eigenvectors, construct characteristic polynomials: (C D -λ p E) α p =0 (8) Where E is the unit matrix. Solving this polynomial, we get the eigenvector α p The corresponding eigenvalue λ p .
