Practical operation data scoring method and system

Abstract

The invention relates to the technical field of automatic scoring, in particular to a practical operation data scoring method and system, and the method comprises the steps: obtaining to-be-scored practical operation data and positive real operation data, respectively extracting original practical operation features in the to-be-scored practical operation data and original positive real operation features in the positive real operation data, according to the original practical operation features and the original correct practical operation features, domain constraint enhancement features are determined, the middle domain constraint enhancement features and the domain constraint enhancement features are used for enhancing the original practical operation features conforming to domain rules, the original practical operation features and the domain constraint enhancement features are fused to obtain first domain adaptive fusion features, and the first domain adaptive fusion features are fused to obtain second domain adaptive fusion features; and performing fusion processing on the original correct operation feature and the domain constraint enhancement feature to obtain a second domain adaptive fusion feature, and generating a score corresponding to the to-be-scored practical operation data according to the first domain adaptive fusion feature and the second domain adaptive fusion feature. Exclusive specifications of the practical operation field are mined through field constraint enhanced features, and the scoring accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of automatic scoring technology, and in particular to a scoring method and system for practical data. Background Technology

[0002] In the field of feature encoding for practical training, existing technologies often use unmodified basic convolutional layers or long short-term memory networks to extract features. These features are then fused through simple concatenation or weighting. The encoding process employs standard attention mechanisms, which fail to consider the specific specifications of the practical field. This results in poor feature targeting and a disconnect from the practical business scenario. Consequently, subsequent bias judgments based on these features are prone to misjudgment and omission, and the scoring results lack accuracy. Summary of the Invention

[0003] To address the aforementioned problems in the prior art, this invention provides a scoring method and system for practical data, which solves the problem that the prior art does not consider the specific specifications of the practical field, resulting in poor feature targeting and thus inaccurate scoring.

[0004] According to one aspect of the present invention, a method for scoring practical data is provided, comprising: Obtain the practical data to be scored and the correct practical data; Extract the original practical features from the practical data to be scored and the original correct practical features from the correct practical data respectively; Based on the original practical features and the original correct practical features, domain constraint enhancement features are determined; wherein, the domain constraint enhancement features are used to enhance the original practical features that conform to the domain rules; The original practical features are fused with the domain constraint enhancement features to obtain the first domain adaptation fusion feature; The original correct operation features are fused with the domain constraint enhancement features to obtain the second domain adaptation fusion features; Based on the first domain adaptation and fusion features and the second domain adaptation and fusion features, a score is generated corresponding to the practical data to be scored.

[0005] According to another aspect of the present invention, a scoring system for practical data is provided, comprising: The acquisition module is used to acquire the practical data to be scored and the correct practical data. The extraction module is used to extract the original practical features from the practical data to be scored and the original correct practical features from the correct practical data, respectively. The determination module is used to determine the domain constraint enhancement features based on the original practical features and the original correct practical features; wherein the domain constraint enhancement features are used to enhance the original practical features that conform to the domain rules; The fusion module is used to fuse the original practical features with the domain constraint enhancement features to obtain a first domain adaptation fusion feature; and to fuse the original correct practical features with the domain constraint enhancement features to obtain a second domain adaptation fusion feature. The generation module is used to generate a score corresponding to the practical data to be scored based on the first domain adaptation and fusion features and the second domain adaptation and fusion features.

[0006] The beneficial effects of this invention are reflected in the fact that by introducing domain constraints to enhance features, it can uncover the specific specifications of the practical domain (such as the parameter range of core operations and the compliance requirements of step execution), so that the fused features can get rid of the limitations of pure data-driven, closely fit the practical business scenario, and specifically solve the problem that traditional feature fusion only relies on data distribution and is detached from domain specifications, thereby improving the accuracy of scoring. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the scoring method for practical data provided by the present invention.

[0008] Figure 2 This is a flowchart illustrating the domain constraint enhancement feature determination method provided by the present invention.

[0009] Figure 3 This is a flowchart illustrating the scoring generation method provided by the present invention.

[0010] Figure 4 This is a schematic diagram of the scoring system for practical data provided by the present invention.

[0011] Figure 5 This is a schematic diagram of the structure of a scoring system for practical data provided in another embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Figure 1 This is a flowchart illustrating the scoring method for practical data provided by the present invention. Figure 2 This is a flowchart illustrating the domain constraint enhancement feature determination method provided by the present invention. Figures 1-2 As shown, the scoring methods for practical data include: S110: Obtain the actual operation data to be scored and correct the actual operation data.

[0014] In this embodiment of the invention, after obtaining the practical data to be scored and the correct practical data, the practical data to be scored is subjected to anomaly filtering, time-series alignment, and standardization preprocessing to obtain preprocessed practical data to be scored and preprocessed correct practical data, respectively. The purpose is to remove interfering data from the practical data to be scored and the correct practical data, thereby preventing inaccurate scores from the practical data to be scored.

[0015] The practical data to be scored can be the data generated by the trainees during actual operation. Correct practical data can serve as reference data for evaluating the trainees' actual operation to determine whether their operation was correct.

[0016] S120: Extract the original practical features from the practical data to be scored and the original correct practical features from the correct practical data.

[0017] In this embodiment of the invention, since the representation of local operational details and overall temporal logic features is equally important in practical scenarios, and a single network is difficult to capture them all, this invention uses an improved convolutional layer-long short-term memory network combined with global average pooling to extract original features. The original features extracted by this combination can simultaneously cover local operational details and overall temporal logic features, effectively solving the problems of pure convolutional layers ignoring temporal sequence and pure long short-term memory networks failing to capture enough details.

[0018] Specifically, based on convolutional layers This method captures local operational features in the practical data to be scored, such as button clicks and parameter adjustments. Convolution operations extract local features from the practical data through a sliding window, thus capturing local operational patterns during the practical process.

[0019] Furthermore, based on Long Short-Term Memory (LSTM) networks The LSTM network extracts temporal dependencies between operational steps corresponding to each detailed feature to obtain temporally related features. Through its gating mechanism (forget gate, input gate, output gate), LSTM can remember important information in long sequences and forget unimportant information, thereby capturing the temporal dependencies between operational steps. Therefore, the LSTM network can model the operational sequence, operational intervals, and logical relationships between steps during a student's practical training, thus generating features containing temporal information.

[0020] Furthermore, through global average pooling Simplified convolutional layer-based The extracted detailed features are used to obtain the pooled features. By employing global average pooling, the dimensionality of the detailed features can be simplified while retaining the overall feature information of the trainee's practical process.

[0021] Furthermore, the temporal correlation features output by the Long Short-Term Memory (LSTM) network are fused with the pooled features to obtain the original practical features. Specifically, the temporal correlation features output by the LSTM network are weighted with the pooled features to obtain the original practical features. The method for extracting the original correct practical features is the same as the method for extracting features from the practical data to be scored, and will not be elaborated here. Therefore, by combining the improved convolutional layer-LSTM network with global average pooling, both the local details and temporal logic of the practical operations are taken into account, resulting in a more comprehensive feature representation.

[0022] S130: Determine the domain constraint enhancement features based on the original practical features and the original correct practical features.

[0023] In this embodiment of the invention, to address the problem that traditional feature fusion relies solely on data distribution and is detached from domain specifications, this invention calculates domain-constrained enhanced features using original practical features and original correct practical features. These domain-constrained enhanced features can uncover specific specifications within the practical domain (such as the parameter range of core operations and compliance requirements for step execution), enabling the fused features to overcome the limitations of purely data-driven approaches and closely align with practical business scenarios.

[0024] Furthermore, domain rules are rules that guide, constrain, and evaluate operational behavior within a specific domain. Examples include parameter thresholds in safety procedures, strict temporal dependencies in operational processes, and precision ranges in quality standards. Therefore, domain constraint enhancement features are essentially the encoded features of various domain rule constraints in the high-dimensional feature space of a neural network.

[0025] Specifically, S130 includes the following steps: S131: Based on the second multilayer perceptron, obtain the first transformation feature according to the original practical features.

[0026] In this embodiment of the invention, a multilayer perceptron... By using a hidden layer structure, the original practical features are nonlinearly transformed to obtain the first transformed features, which are semantic feature vectors that have been filtered and enhanced by domain knowledge.

[0027] Specifically, during the training of the original practical features, the multilayer perceptron encodes and stores domain knowledge (such as standard operating procedures, safety regulations, and logical relationships between parameters) in the specific practical scenarios into its network weights. In processing the original practical features, the knowledge weights corresponding to the domain knowledge stored in the multilayer perceptron are used to interpret and annotate the input original practical features, and finally output the first transformed feature.

[0028] Furthermore, the degree of conformity to domain rules during the original practical operation can be determined through the first transformation feature. It's understandable that the original correct operation feature only contains the actual action, such as a student clicking button A in the 3rd second, with parameter X set, but it's unknown whether the actual operation was correct or conformed to the rules, such as whether the click was timely or whether parameter X was dangerous. However, the multilayer perceptron can label the original correct operation feature with domain rules, outputting the first transformation feature. For example, if the domain rules include temporal compliance, the value of the temporal compliance dimension of the original correct operation feature might be -0.8 (indicating severe delay), and the value of the parameter safety dimension might be 0.9 (indicating safety).

[0029] S132: Based on the third multilayer perceptron, the second transformation feature is obtained according to the original correct operation features.

[0030] In this embodiment of the invention, the multilayer perceptron performs a nonlinear transformation on the original correct operation features through a hidden layer structure to obtain a second transformed feature, wherein the second transformed feature is a semantic feature vector that has been filtered and enhanced by domain knowledge.

[0031] Specifically, during the training of the original correct operation features, the multilayer perceptron encodes and stores domain knowledge (such as standard operating procedures, safety regulations, and logical relationships between parameters) in the specific practical scenarios into its network weights. In processing the original correct operation features, the knowledge weights corresponding to the domain knowledge stored in the multilayer perceptron are used to interpret and annotate the input original correct operation features, and finally output the second transformation feature.

[0032] S133: Based on the first transformation feature, the original practical feature is calibrated to obtain the calibrated original practical feature.

[0033] In this embodiment of the invention, the original practical features are calibrated by calculating the Hadamard product of the first transformed feature and the original practical features. Specifically, the first transformed feature serves as a weight to scale the original practical features. For example, if a rule in a certain dimension is severely violated (MLP output ≈ -1), then the original practical features of that dimension are preserved but their signs are reversed, indicating that the operation occurred but was performed incorrectly, and the error is clearly marked.

[0034] S134: Based on the second change feature, calibrate the original correct operation feature to obtain the calibrated original correct operation feature.

[0035] In this embodiment of the invention, the original correct operational features are calibrated by calculating the Hadamard product of the second transformed feature and the original correct operational features. Specifically, the first transformed feature is used as a weight to scale the original operational features. For example, if the rule compliance of a certain dimension is high (MLP output ≈ 1), then the original correct operational features of that dimension are preserved or enhanced, indicating that the operation has occurred and was performed correctly.

[0036] S135: Compute the domain constraint enhancement features based on the calibrated original practical features and the calibrated original correct practical features.

[0037] In this embodiment of the invention, the domain constraint enhancement feature is calculated by summing the calibrated original practical features and the calibrated original correct practical features, thereby fusing the calibrated original practical features and the calibrated original correct practical features.

[0038] S140: The original practical features and the domain constraint enhancement features are fused together to obtain the first domain adaptation fusion features.

[0039] S150: The original correct operation features are fused with the domain constraint enhancement features to obtain the second domain adaptation fusion features.

[0040] S160: Generate scores corresponding to the practical data to be scored based on the adaptation and fusion features of the first domain and the adaptation and fusion features of the second domain.

[0041] Figure 3 This is a flowchart illustrating the scoring generation method provided by the present invention. Figure 3 As shown, S160 may include: S161: Transpose the first domain adaptation fusion feature to obtain the transposed fusion feature.

[0042] S162: Based on the first multilayer perceptron, obtain the operation rhythm adaptive attention coefficient according to the transposed fusion feature and the second domain adaptation fusion feature. The operation rhythm adaptive attention coefficient is used to dynamically allocate the weights corresponding to different operation rhythms.

[0043] In this embodiment of the invention, an interaction matrix is ​​obtained by calculating the Hadamard product of the transposed fusion features and the second domain-adapted fusion features. The interaction matrix captures the temporal correlation strength and pattern differences between student operations and standard operations in each feature dimension. The first multilayer perceptron (MLP) extracts student practical features with operation rhythm offsets from the interaction matrix, such as device start-up and shutdown.

[0044] Furthermore, the Softmax function normalizes the rhythm-shifted student operation features output by the MLP into a probability distribution vector, namely the operation rhythm adaptive attention coefficient. The operation rhythm adaptive attention coefficient can dynamically adjust the weight allocation of the attention window according to the different operation rhythms of different students, strengthening the feature representation of core operations (such as key parameter calibration, device start / stop) and weakening the interference of redundant operations (such as repeated confirmations, irrelevant clicks). Compared with the basic attention mechanism, it is more in line with the personalized characteristics of student operation, solving the problem that the fixed window of the traditional attention mechanism cannot adapt to the individual rhythm differences of students. The close cooperation of each link ultimately enables the output encoded features to accurately represent the core differences between the student and the correct operation.

[0045] S163: Based on the adaptive attention coefficient of operation rhythm, the first domain adaptation fusion feature is encoded to obtain the encoded student operation feature.

[0046] S164: Based on the adaptive attention coefficient of the operation rhythm, the second domain adaptation fusion feature is encoded to obtain the encoded correct operation feature.

[0047] S165: Generate scores corresponding to the practical data to be scored based on the coded student operation characteristics and the coded correct operation characteristics.

[0048] In this embodiment of the invention, a student operation score is generated using a fully connected layer and a sigmoid function, based on the encoded student operation features and the encoded correct operation features. Specifically, the encoded student operation features and the encoded correct operation features are concatenated to obtain the concatenated overall operation features, i.e., through... A concatenation operation is performed. Then, based on a fully connected layer (FC), the overall operational features of the concatenated model are extracted to obtain the overall operational deviation features. Based on the sigmoid function, the overall operational deviation features are predicted to obtain the student's operational score.

[0049] In one embodiment, S162 can be specifically implemented as follows: calculating the first Hadamard product based on the transposed fusion feature and the second domain-adaptive fusion feature; extracting the difference features in the first Hadamard product based on the first multilayer perceptron; and converting the difference features into an operation rhythm adaptive attention coefficient.

[0050] In this embodiment of the invention, an improved operation rhythm adaptive attention mechanism is used, and the operation rhythm adaptive attention coefficients are obtained by using the Hadamard product between the transpose of the first domain adaptation fusion feature and the second domain adaptation fusion feature, through a multilayer perceptron and a Softmax function.

[0051] In one embodiment, S163 can be specifically implemented as follows: based on the attention mechanism, determine the dependency relationship between different features in the first domain adaptation fusion feature; based on the operation rhythm adaptive attention coefficient, adjust the dependency relationship to obtain the adjusted first domain adaptation fusion feature; retain the original student features in the adjusted first domain adaptation fusion feature; wherein, the encoded student operation features include the original student features.

[0052] In this embodiment of the invention, an attention mechanism is used. Dependencies between different features in the first domain adaptation fusion feature are extracted to obtain dependency features. The second Hadamard product of the dependency features and the adaptive attention coefficient of the operational rhythm is calculated to obtain the adjusted first domain adaptation fusion feature. By dynamically adjusting the weights of different time steps, the scoring can adaptively assign higher attention weights to time steps with significant rhythm differences based on the actual operational rhythm of the learner, thereby more accurately focusing on core operational steps and improving scoring accuracy.

[0053] Furthermore, the sum between the second Hadamard product and the first domain adaptation fusion feature is calculated to retain the original learner features in the adjusted first domain adaptation fusion feature, thus preventing the attention mechanism from filtering out important but irrelevant operational features. The encoded correct operational features follow the same steps as above and will not be elaborated here.

[0054] In one embodiment, before S140, the scoring method for the practical data can be specifically implemented as follows: concatenating the original practical features with the original correct practical features to obtain the concatenated practical features; based on the fourth multilayer perceptron, obtaining the third change feature according to the concatenated practical features; calculating the practical scenario constraint enhancement gating weight based on the third change feature; wherein, the practical scenario constraint enhancement gating weight is used to dynamically adjust the proportion of the original practical features and the original correct practical features during fusion; wherein, S140 can be specifically implemented as follows: fusing the original practical features with the practical scenario constraint enhancement gating weight and the domain constraint enhancement feature together to obtain the first domain adaptation fusion feature.

[0055] In this embodiment of the invention, the original practical features and the original correct practical features are spliced ​​together. The process involves concatenating practical features to obtain the final operational features. A multilayer perceptron is used to extract the relationship patterns between the original operational features and the original correct operational features, thus obtaining relationship weights. Then, a sigmoid function is used to transform these relationship weights to the 0-1 range, resulting in operational scenario constraint-enhanced gating weights. By introducing this operational scenario constraint-enhanced gating mechanism, domain norms are integrated into the feature fusion process, making the features more aligned with the business essence.

[0056] Furthermore, the original practical features are fused together with the practical scenario constraint-enhanced gating weights and the domain constraint-enhanced features to obtain the first domain-adaptive fusion features. S150 can be specifically implemented as follows: the original correct practical features are fused together with the practical scenario constraint-enhanced gating weights and the domain constraint-enhanced features to obtain the second domain-adaptive fusion features.

[0057] Furthermore, the third Hadamard product of the original practical features and the practical scenario constraint enhancement gating weights is calculated. Then, the sum of the third Hadamard product and the domain constraint enhancement features is calculated to obtain the first domain adaptation fusion feature. The calculation method for the second domain adaptation fusion feature is the same as described above, and will not be repeated here. That is, the original operational features in the above method are replaced with the original correct practical features.

[0058] Among them, the gating weight is used to dynamically adjust the proportion of student features and correct features in the fusion process, so as to avoid the fusion feature bias caused by the excessive proportion of a certain type of feature.

[0059] In one embodiment, the scoring method for practical data can be specifically implemented as follows: based on the coded student operation characteristics and the coded correct operation characteristics, key operation deviation characteristics are obtained; based on the operation deviation characteristics, student operation error identification results are generated.

[0060] In this embodiment of the invention, since timing compliance is the core dimension for judging operational deviations in practical assessments, the operational timing constraint function can filter out invalid deviations caused by non-timing factors, while the deviation weight calibration item can make core operational deviations (such as incorrect adjustment of key parameters or incorrect execution of core steps) dominate the features, avoiding non-core deviations (such as click errors on irrelevant interfaces) from masking core problems. This effectively solves the problem that traditional fully connected networks lack business orientation in weight allocation and weaken core deviations when constructing deviation features.

[0061] Specifically, based on key operational deviation features, a classification network (Softmax) is used to distinguish deviation types. First, an operational deviation penalty coefficient is calculated, which reflects the degree of deviation based on the quantification ratio of the deviation features. The formula for calculating the operational deviation penalty coefficient is as follows: ,in, This refers to the original deviation features between the encoded learner's operational characteristics and the encoded correct operational characteristics. It is a fully connected layer. To obtain the L2 norm, These are the correct operational features after encoding.

[0062] Furthermore, based on the fully connected layer, the key operational deviation features are processed to obtain the processed deviation features. The processed deviation features are weighted with the operational deviation penalty coefficient, and the weighted features are then multiplied by the key operational deviation features to obtain the key deviation fusion features.

[0063] Furthermore, the key deviation fusion features are processed using the Softmax function to obtain the probability of each deviation type. Finally, the deviation type corresponding to the index of the maximum probability is used as the error identification result, where the index corresponds to the deviation type.

[0064] The operational deviation penalty coefficient assigns higher weight to operations with more severe deviations (such as completely deviating from correct parameter settings or timing errors that completely reverse core steps) in type determination. This ensures that the determination results align with the business logic of "matching the degree of deviation with the error level" in practical assessments. This solves the problem of traditional classification networks relying solely on feature similarity to determine types, ignoring the degree of deviation and business weights. The close coordination of each stage ensures that error identification results accurately correspond to specific deviation types in practical operations. Therefore, by combining the penalty coefficient-weighted features in the classification network for determination, the accuracy of deviation type differentiation is significantly improved, enabling error identification results to accurately correspond to specific problems in practical operations. This provides a precise basis for subsequent error reporting and manual review.

[0065] In one embodiment, the scoring method for practical data can be specifically implemented as follows: determining the absolute difference features between the encoded student operation features and the encoded correct operation features; performing weight calibration on the absolute difference features based on the operation temporal constraint function to obtain the original deviation features; extracting the temporal weighted deviation features from the original deviation features based on a fully connected layer; semantically aligning the encoded student operation features with the encoded correct operation features to obtain aligned operation features; and generating key operation deviation features based on the original deviation features and the aligned operation features.

[0066] In this embodiment of the invention, an improved fully connected network is used to construct key operation deviation features. The core function of the newly added deviation weight calibration term is to enhance the representation capability of core operation deviations. Specifically, it is the absolute difference feature between the pre-encoded student operation features and the pre-encoded correct operation features. The calculation method of the absolute difference feature includes: calculating the difference feature between the pre-encoded student operation features and the pre-encoded correct operation features, and calculating the absolute value of the difference feature as the absolute difference feature.

[0067] Furthermore, by applying a temporal constraint function to weight the absolute difference features, the core deviation features in the temporal dimension of the practical scenario are highlighted. Then, a fully connected layer processes the original deviation features and the encoded student operation features separately to obtain temporal deviation features and key student operation features. Therefore, by improving the fully connected network and introducing a temporal constraint function, the constructed key operation deviation features focus on core deviations at the business level, filtering out invalid interference deviations, and the feature representation is more aligned with the evaluation logic of practical assessments.

[0068] Furthermore, the encoded correct operation features are inverted to obtain inverted correct operation features. Then, the inverted correct operation features are processed by the Softmax function to obtain the probability of each operation type.

[0069] Furthermore, the fourth Hadamard product of the trainee's key operational features and the probability of each operational type is calculated, and then the weighted sum of the fourth Hadamard product and the temporal deviation features is calculated to obtain the key operational deviation features.

[0070] In one embodiment, the method for constructing the operation timing constraint function includes: mapping the first domain adaptation fusion features to obtain the mapped fusion features; determining the anomaly degree value corresponding to the mapped fusion features based on a multilayer perceptron; converting the anomaly degree value into a timing weight; evaluating the dependency relationship between the mapped fusion features at different time steps; and constructing the operation timing constraint function according to the timing weight and the dependency relationship between the mapped fusion features at different time steps.

[0071] In this embodiment of the invention, since timing compliance is the core dimension for judging operational deviations in practical assessments, the operational timing constraint function can filter out invalid deviations caused by non-timing factors, while the deviation weight calibration item can make core operational deviations (such as incorrect adjustment of key parameters or incorrect execution of core steps) dominate the features, avoiding non-core deviations (such as click errors on irrelevant interfaces) from masking core problems. This effectively solves the problem that traditional fully connected networks lack business orientation in weight allocation and weaken core deviations when constructing deviation features.

[0072] Furthermore, based on the embedding layer The first domain adaptation fusion feature is mapped from high-dimensional feature to low-dimensional space to obtain the mapped fusion feature. The absolute value of the mapped fusion feature is processed to eliminate the influence of the sign of the feature value, so that the subsequent network can focus more on the size (i.e. intensity) of the feature and avoid the weakening of the temporal constraint signal due to positive and negative cancellation.

[0073] Furthermore, a nonlinear transformation is applied to the fused features after absolute value processing using a multilayer perceptron (MLP) to learn the complex nonlinear relationships between features and capture the implicit patterns of temporal violations in the feature space. The transformed features are then negativeed, and based on the Sigmoid function, the negativeed features are mapped to a preset interval (0,1). Negation increases the original MLP output (indicating a more pronounced potential temporal anomaly) and decreases the Sigmoid output. The Sigmoid output value serves as a temporal weight; a smaller weight indicates a higher likelihood of temporal constraint violations, thus suppressing them in subsequent weighting.

[0074] Furthermore, the first-domain adaptation fusion features are transposed, and then mapped from high-dimensional features to low-dimensional space based on the embedding layer to obtain transposed fusion features. Then, the transposed fusion features are predicted based on the Softmax function to obtain the corresponding probabilistic features. The Softmax output is used to emphasize the influence of important temporal nodes.

[0075] Furthermore, the probabilistic features corresponding to the transposed fusion features are multiplied by a preset coefficient to obtain product features. These product features are then weighted with temporal weights to calibrate the weights of the absolute difference features. The above calculation process is the calculation process for the operational temporal constraint function. The preset coefficient is 0.3.

[0076] Figure 4 This is a schematic diagram of the scoring system for practical data provided by the present invention. Figure 4 As shown, the scoring system for practical data includes: an acquisition module for acquiring practical data to be scored and correct practical data; an extraction module for extracting the original practical features from the practical data to be scored and the original correct practical features from the correct practical data; a determination module for determining domain constraint enhancement features based on the original practical features and the original correct practical features; wherein, the domain constraint enhancement features are used to enhance the original practical features that conform to the domain rules; a fusion module for fusing the original practical features and the domain constraint enhancement features to obtain a first domain-adaptive fusion feature; fusing the original correct practical features and the domain constraint enhancement features to obtain a second domain-adaptive fusion feature; and a generation module for generating the score corresponding to the practical data to be scored based on the first domain-adaptive fusion feature and the second domain-adaptive fusion feature.

[0077] Figure 5 This is a schematic diagram of the structure of a scoring system for practical data provided in another embodiment of the present invention. (See diagram below.) Figure 5As shown, the generation module may include: a transpose unit, used to transpose the first domain adaptation fusion features to obtain transposed fusion features; an acquisition unit, used to acquire an operation rhythm adaptive attention coefficient based on the first multilayer perceptron, according to the transposed fusion features and the second domain adaptation fusion features; wherein the operation rhythm adaptive attention coefficient is used to dynamically allocate weights corresponding to different operation rhythms; a first encoding unit, used to encode the first domain adaptation fusion features based on the operation rhythm adaptive attention coefficient to obtain encoded student operation features; a second encoding unit, used to encode the second domain adaptation fusion features based on the operation rhythm adaptive attention coefficient to obtain encoded correct operation features; and a scoring unit, used to generate a score corresponding to the practical data to be scored based on the encoded student operation features and the encoded correct operation features.

[0078] In one embodiment, the generation module may be specifically configured to: calculate a first Hadamard product based on the transposed fusion features and the second domain-adaptive fusion features; extract the differential features in the first Hadamard product based on a first multilayer perceptron; and convert the differential features into an operation rhythm adaptive attention coefficient.

[0079] In one embodiment, the generation module may be specifically configured to: determine the dependency relationship between different features in the first domain adaptation fusion feature based on the attention mechanism; adjust the dependency relationship based on the operation rhythm adaptive attention coefficient to obtain the adjusted first domain adaptation fusion feature; retain the original student features in the adjusted first domain adaptation fusion feature; wherein, the encoded student operation features include the original student features.

[0080] In one embodiment, the determining module may be specifically configured as follows: based on a second multilayer perceptron, obtain a first transformation feature according to the original operational features; based on a third multilayer perceptron, obtain a second transformation feature according to the original correct operational features; calibrate the original operational features based on the first transformation feature to obtain calibrated original operational features; calibrate the original correct operational features according to the second transformation feature to obtain calibrated original correct operational features; and calculate the domain constraint enhancement feature based on the calibrated original operational features and the calibrated original correct operational features.

[0081] In one embodiment, before the fusion module, the scoring system for practical data can be specifically configured as follows: concatenating the original practical features with the original correct practical features to obtain concatenated practical features; based on the fourth multilayer perceptron, obtaining the third change feature according to the concatenated practical features; and calculating the practical scenario constraint enhancement gating weight based on the third change feature; wherein the practical scenario constraint enhancement gating weight is used to dynamically adjust the proportion of the original practical features and the original correct practical features during fusion; wherein the fusion module can be specifically configured as follows: performing fusion processing on the original practical features, the practical scenario constraint enhancement gating weight, and the domain constraint enhancement features to obtain the first domain-adaptive fusion feature.

[0082] In one embodiment, the scoring system for practical data can be specifically configured as follows: based on the coded student operation characteristics and the coded correct operation characteristics, key operation deviation characteristics are obtained; based on the operation deviation characteristics, student operation error identification results are generated.

[0083] In one embodiment, the scoring system for practical data can be specifically configured as follows: determining the absolute difference features between the encoded student operation features and the encoded correct operation features; performing weight calibration on the absolute difference features based on the operation time-series constraint function to obtain the original deviation features; extracting the time-series deviation features from the original deviation features and the student key operation features from the encoded student operation features based on a fully connected layer; and generating key operation deviation features based on the time-series deviation features and the student key operation features.

[0084] In one embodiment, the system for constructing the operation timing constraint function can be specifically configured as follows: mapping the first domain adaptation fusion feature to obtain the mapped fusion feature; determining the anomaly degree value corresponding to the mapped fusion feature based on a multilayer perceptron; converting the anomaly degree value into a timing weight; evaluating the dependency relationship between the mapped fusion feature and different time steps; and constructing the operation timing constraint function according to the timing weight and the dependency relationship between the mapped fusion feature and different time steps.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A scoring method for practical data, characterized in that, include: Obtain the practical data to be scored and the correct practical data; Extract the original practical features from the practical data to be scored and the original correct practical features from the correct practical data respectively; Based on the original practical features and the original correct practical features, domain constraint enhancement features are determined; wherein, the domain constraint enhancement features are used to enhance the original practical features that conform to the domain rules; The original practical features are fused with the domain constraint enhancement features to obtain the first domain adaptation fusion feature; The original correct operation features are fused with the domain constraint enhancement features to obtain the second domain adaptation fusion features; Based on the first domain adaptation and fusion features and the second domain adaptation and fusion features, a score is generated corresponding to the practical data to be scored.

2. The scoring method for practical data according to claim 1, characterized in that, The step of generating a score corresponding to the practical data to be scored based on the first domain adaptation fusion feature and the second domain adaptation fusion feature includes: The first domain adaptation fusion feature is transposed to obtain the transposed fusion feature; Based on the first multilayer perceptron, an operation rhythm adaptive attention coefficient is obtained according to the transposed fusion feature and the second domain-adaptive fusion feature; wherein, the operation rhythm adaptive attention coefficient is used to dynamically allocate the weights corresponding to different operation rhythms. Based on the adaptive attention coefficient of the operation rhythm, the first domain adaptation fusion feature is encoded to obtain the encoded student operation feature. Based on the adaptive attention coefficient of the operation rhythm, the second domain adaptation fusion feature is encoded to obtain the encoded correct operation feature; Based on the coded student operation characteristics and the coded correct operation characteristics, a score is generated corresponding to the practical data to be scored.

3. The scoring method for practical data according to claim 2, characterized in that, The step of obtaining the adaptive attention coefficient for operation rhythm based on the first multilayer perceptron and the transposed fusion feature and the second domain-adaptive fusion feature includes: Calculate the first Hadamard product based on the transposed fusion feature and the second domain-adapted fusion feature; Based on the first multilayer perceptron, extract the differential features in the first Hadamard product; The aforementioned differences are transformed into an adaptive attention coefficient for operational rhythm.

4. The scoring method for practical data according to claim 2, characterized in that, The process of encoding the first domain-adaptive fusion features based on the adaptive attention coefficient of the operation rhythm to obtain the encoded student operation features includes: Based on the attention mechanism, the dependencies between different features in the first domain adaptation fusion feature are determined; Based on the adaptive attention coefficient of the operation rhythm, the dependency relationship is adjusted to obtain the adjusted first domain adaptation fusion feature; The original student features in the adjusted first domain adaptation fusion features are retained; wherein, the encoded student operation features include the original student features.

5. The scoring method for practical data according to claim 1, characterized in that, The step of determining the domain constraint enhancement features based on the original practical features and the original correct practical features includes: Based on the second multilayer perceptron, the first transformation feature is obtained according to the original practical features; Based on the third multilayer perceptron, the second transformation feature is obtained according to the original correct operation feature; Based on the first transformation feature, the original practical feature is calibrated to obtain the calibrated original practical feature; Based on the second change feature, the original correct operation feature is calibrated to obtain the calibrated original correct operation feature; Based on the calibrated original operational features and the calibrated original correct operational features, the domain constraint enhancement features are calculated.

6. The scoring method for practical data according to claim 1, characterized in that, Before fusing the original practical features with the domain constraint enhancement features to obtain the first domain-adaptive fusion features, the method further includes: The original practical features are concatenated with the original correct practical features to obtain the concatenated practical features. Based on the fourth multilayer perceptron, the third change feature is obtained according to the spliced ​​operational features; Based on the third change feature, the practical scenario constraint enhancement gating weight is calculated; wherein, the practical scenario constraint enhancement gating weight is used to dynamically adjust the proportion of the original practical feature and the original correct practical feature during fusion. The step of fusing the original practical features with the domain constraint enhancement features to obtain the first domain adaptation fusion feature includes: The original practical features are fused together with the practical scenario constraint enhancement gating weights and the domain constraint enhancement features to obtain the first domain adaptation fusion features.

7. The scoring method for practical data according to claim 2, characterized in that, Also includes: Based on the encoded student operation characteristics and the encoded correct operation characteristics, key operation deviation characteristics are obtained; Based on the operational deviation characteristics, the student's operational error identification results are generated.

8. The scoring method for practical data according to claim 7, characterized in that, The step of obtaining key operational deviation features based on the encoded student operational features and the encoded correct operational features includes: Determine the absolute difference between the coded student operational characteristics and the coded correct operational characteristics; Based on the operational timing constraint function, the absolute difference features are weighted and calibrated to obtain the original deviation features; Based on the fully connected layer, the temporal deviation features in the original deviation features and the key student operation features in the encoded student operation features are extracted respectively. Based on the timing deviation characteristics and the student's key operation characteristics, key operation deviation characteristics are generated.

9. The scoring method for practical data according to claim 8, characterized in that, The method for constructing the operation timing constraint function includes: The first domain adaptation and fusion features are mapped to obtain the mapped fusion features; Based on a multilayer perceptron, the anomaly level value corresponding to the mapped fused feature is determined; The anomaly severity value is converted into a time-series weight; Evaluate the dependencies of the mapped fused features across different time steps; The operation timing constraint function is constructed based on the time-series weights and the dependencies between the mapped fusion features at different time steps.

10. A scoring system for practical data, characterized in that, include: The acquisition module is used to acquire the practical data to be scored and the correct practical data. The extraction module is used to extract the original practical features from the practical data to be scored and the original correct practical features from the correct practical data, respectively. The determination module is used to determine the domain constraint enhancement features based on the original practical features and the original correct practical features; wherein the domain constraint enhancement features are used to enhance the original practical features that conform to the domain rules; The fusion module is used to fuse the original practical features with the domain constraint enhancement features to obtain a first domain adaptation fusion feature; and to fuse the original correct practical features with the domain constraint enhancement features to obtain a second domain adaptation fusion feature. The generation module is used to generate a score corresponding to the practical data to be scored based on the first domain adaptation and fusion features and the second domain adaptation and fusion features.

Images