A method for comparing and extracting device feature data based on decision tree

Abstract

The present invention discloses a device feature data comparison and extraction method based on a decision tree, which relates to the field of device feature data comparison and extraction. The method comprises: performing hierarchical processing on device features, screening out main factors affecting device performance, and constructing a hierarchical decision tree structure; performing feature clustering analysis on the decision tree structure, merging similar features into feature groups, generating new synthetic features based on the feature groups, and adaptively adjusting the decision tree structure; setting initial weights of the importance of features in nodes, adopting an online learning algorithm, and dynamically adjusting the importance weights of features according to real-time data feedback; obtaining real-time operation data of devices to be compared, and selecting an optimal feature data comparison scheme based on the feature weights of decision path nodes of each decision tree; and dynamically deleting redundant nodes of the decision tree based on the selection of the historical optimal feature data comparison scheme.

Description

Technical Field

[0001] The present invention relates to the field of device feature data comparison and extraction, and in particular to a device feature data comparison and extraction method based on a decision tree. Background Art

[0002] In modern manufacturing and industrial environments, equipment performance monitoring and fault detection have become increasingly important. With the development of the Internet of Things and big data technologies, it is now possible for equipment to collect and analyze large amounts of real-time data. Traditional equipment monitoring methods typically rely on static feature selection and offline analysis, making them difficult to adapt to the ever-changing conditions and characteristics of equipment operation.

[0003] How to establish a decision tree-based device feature data comparison and extraction method that can efficiently extract and compare device data through adaptive feature selection and real-time feedback mechanisms, thereby quickly identifying performance differences and potential faults, is an urgent problem that needs to be solved. Summary of the Invention

[0004] In order to solve the above technical problems, a device feature data comparison and extraction method based on a decision tree is provided. This technical solution solves the problems raised in the above background technology.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0006] A device feature data comparison and extraction method based on a decision tree, comprising:

[0007] Perform hierarchical processing on device features, screen out the main factors affecting device performance, and build a hierarchical decision tree structure;

[0008] Perform feature clustering analysis on the decision tree structure, merge similar features into feature groups, generate new synthetic features based on the feature groups, and perform adaptive structural adjustments on the decision tree;

[0009] Set the initial importance weight of the features in the node, and use the online learning algorithm to dynamically adjust the importance weight of the features based on real-time data feedback;

[0010] Obtain the real-time operating data of the equipment to be compared, and select the optimal feature data comparison solution based on the feature weights of the decision path nodes of each decision tree;

[0011] Based on the selection of the historical optimal feature data comparison scheme, redundant nodes of the decision tree are dynamically deleted.

[0012] Preferably, the hierarchical processing of device features, screening out the main factors affecting device performance, and constructing a hierarchical decision tree structure specifically includes:

[0013] Get the comparison device set M;

[0014] Perform statistical analysis on the historical operating data of the devices in the device set M, screen out the main factors affecting the device performance, and establish the corresponding relationship between the device and its main feature set;

[0015] Among them, the main features are evaluated by the information gain algorithm on all the main factors affecting the performance of the equipment, and the feature with the largest information gain is selected as the main feature;

[0016] The identified features are systematically divided into layers, with the first layer recording the main features of the device;

[0017] Based on the specific operating data of the equipment, the subsequent layer structure is subjected to in-depth information gain evaluation and feature classification. Based on the main features, it is further refined to identify specific performance indicators, and the specific performance indicators are filled into the subsequent layers as refined features.

[0018] Determine and construct the decision tree structure based on the feature hierarchy division, specifically:

[0019] Use the first-level features as the root node of the decision tree;

[0020] Based on each main feature root node, select secondary features for splitting in turn, and recursively construct the child nodes of each node. Each node represents a different performance indicator of the node on the previous layer. The construction ends when all feature levels are expressed.

[0021] Mark nodes without child nodes as leaf nodes, which store the performance status of the device.

[0022] A decision tree is established by comparing all devices in the device set M, which is marked as a decision tree set.

[0023] Preferably, the performing feature cluster analysis on the decision tree structure, merging similar features into feature groups, generating new synthetic features based on the feature groups, and performing adaptive structural adjustment on the decision tree specifically includes:

[0024] Obtain a set of decision trees, and perform complexity evaluation on the decision trees in the set based on the scale complexity measurement algorithm and leaf node distribution;

[0025] The above decision tree complexity evaluation calculation expression is:

[0026]

[0027] Where, C(T i ) is the complexity evaluation value of the i-th decision tree, n is the total number of decision tree sets, m is the number of subtrees of the i-th decision tree, L(Tc j ) is the number of leaf nodes in the jth subtree, Ltotal is the total number of child nodes, σ j 2 is the variance of the number of nodes at each level in the j-th subtree, c is the number of feature categories, p k is the proportion of the k-th feature category in the leaf node;

[0028] Set a complexity adjustment threshold, filter out decision trees whose complexity calculation results exceed the complexity adjustment threshold, and establish a complex decision tree set V based on the screening results;

[0029] Perform feature similarity analysis on the decision trees in the complex decision tree set V, and identify the similarity between features by calculating the Pearson correlation coefficient between the features;

[0030] The K-means clustering algorithm is used for iterative calculation to reduce the feature dimension of each complex decision tree and divide similar features into multiple cluster groups;

[0031] Taking feature similarity as weight, perform weighted aggregation on the clustering results in each feature clustering group to generate new synthetic features;

[0032] The information gain of the new synthetic features is re-evaluated, and the decision tree structure is adaptively adjusted according to the order of information gain.

[0033] Preferably, the setting of the initial weight of the importance of the feature in the node, using an online learning algorithm, and dynamically adjusting the importance weight of the feature according to real-time data feedback specifically includes:

[0034] The information gain calculation result corresponding to each node feature in the decision tree is used as the initial weight of the feature's importance;

[0035] Real-time monitoring and comparison of running devices in the device set M, obtaining real-time device operation data of the running devices, inputting the real-time device operation data into the decision tree, and obtaining the device performance status prediction results;

[0036] By comparing the deviation between the current device performance state and the expected performance state, the contribution feedback of different features to the state deviation is quantified, and the contribution deviation feedback value of each feature is calculated. The specific expression is:

[0037]

[0038] Where Δω z is the contribution bias feedback value of the z-th feature, η is the learning rate, is the difference between the actual equipment performance status y and the expected performance status Deviation quantization value, is the model prediction output For the zth feature Xz sensitivity;

[0039] Among them, the feature sensitivity value is set based on the proportion of the initial weight of the feature's importance to the cumulative value of the initial weight of all feature importances;

[0040] Based on the real-time feedback of feature contribution deviation results, the importance weight of the corresponding feature is dynamically adjusted according to the contribution deviation ratio and scale.

[0041] Preferably, the step of obtaining the real-time operation data of the device to be compared and selecting the optimal feature data comparison scheme based on the feature weights of the decision path nodes of each decision tree specifically includes:

[0042] Obtain real-time operating data of the device to be compared, input the real-time operating data into all decision trees in the decision tree set, and compare and analyze the real-time operating data with the nodes in the decision tree based on the node characteristics and splitting rules in the decision tree;

[0043] For a given device D to be compared, after inputting the real-time operation data of the device, its decision path set is PD = {PD1…PD r …PD m}, where PD r The decision path determined for the rth decision tree;

[0044] Extract the importance weights of the features of the nodes passed by each decision path in the decision path set, calculate the cumulative value of the feature importance weights of each decision path, select and record the decision path with the largest cumulative weight value as the optimal decision path, and mark the feature expression set corresponding to the nodes passed by the optimal path as the optimal feature data comparison scheme of device D.

[0045] Preferably, the dynamically deleting redundant nodes of the decision tree based on the selection of the historical optimal feature data comparison scheme specifically includes:

[0046] Set the decision tree node adjustment time window size, and obtain the historical optimal decision path set of each decision tree record in the decision tree set in the previous time window of each adjustment time window;

[0047] The optimal decision path nodes of the decision tree are overlapped and reproduced, the number of times the historical optimal decision path passes through each node is marked as the number of active times of the node, and the minimum active number threshold of the node is set;

[0048] Nodes whose active times are less than the minimum active times threshold are pruned, and the feature information corresponding to the nodes is deleted.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] Perform feature clustering analysis on the decision tree structure, merge similar features into feature groups, generate new synthetic features based on feature groups, reduce redundant features, and thus reduce the complexity of the decision tree, while improving the model's generalization ability when faced with data from different devices. Set the initial importance weights of features in the node, and use an online learning algorithm to dynamically adjust the importance weights of features based on real-time data feedback to avoid the model relying on outdated features, thereby reducing the risk of overfitting. Based on the feature weights of the decision path nodes of each decision tree, select the optimal feature data comparison scheme to ensure that the features compared each time are the most relevant features under the current device status, thereby improving the accuracy of the comparison. Based on the selection of the historical optimal feature data comparison scheme, dynamically delete redundant nodes in the decision tree to ensure that the model maintains the simplest structure and improve the computational efficiency of the model.

[0051] A decision tree-based device feature data comparison and extraction method is implemented to efficiently extract and compare device data through adaptive feature selection and real-time feedback mechanism, thereby quickly identifying performance differences and potential faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a device feature data comparison and extraction method based on a decision tree according to the present invention;

[0053] Figure 2 The present invention performs hierarchical processing on device features, screens out the main factors affecting device performance, and constructs a hierarchical decision tree structure flow chart;

[0054] Figure 3 The present invention performs feature cluster analysis on the decision tree structure, merges similar features into feature groups, generates new synthetic features based on the feature groups, and performs adaptive structural adjustment on the decision tree.

[0055] Figure 4 A flowchart for setting the initial weight of the importance of features in a node of the present invention, using an online learning algorithm to dynamically adjust the importance weight of features based on real-time data feedback;

[0056] Figure 5 A flowchart of the present invention for obtaining real-time operation data of the device to be compared and selecting the optimal feature data comparison solution based on the feature weights of the decision path nodes of each decision tree;

[0057] Figure 6 This is a flow chart for dynamically deleting redundant nodes in a decision tree based on the selection of the historical optimal feature data comparison solution of the present invention. DETAILED DESCRIPTION

[0058] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0059] Reference Figure 1 As shown, a device feature data comparison and extraction method based on a decision tree includes:

[0060] Perform hierarchical processing on device features, screen out the main factors affecting device performance, and build a hierarchical decision tree structure;

[0061] Perform feature clustering analysis on the decision tree structure, merge similar features into feature groups, generate new synthetic features based on the feature groups, and perform adaptive structural adjustments on the decision tree;

[0062] Set the initial importance weight of the features in the node, and use the online learning algorithm to dynamically adjust the importance weight of the features based on real-time data feedback;

[0063] Obtain the real-time operating data of the equipment to be compared, and select the optimal feature data comparison scheme based on the feature weights of the decision path nodes of each decision tree;

[0064] Based on the selection of the historical optimal feature data comparison scheme, redundant nodes of the decision tree are dynamically deleted.

[0065] Reference Figure 2 As shown in the figure, the device features are hierarchically processed to screen out the main factors affecting device performance. The hierarchical decision tree structure is constructed, specifically including:

[0066] Get the comparison device set M;

[0067] Perform statistical analysis on the historical operating data of the devices in the device set M, screen out the main factors affecting the device performance, and establish the corresponding relationship between the device and its main feature set;

[0068] Among them, the main features are evaluated by the information gain algorithm on all the main factors affecting the performance of the equipment, and the feature with the largest information gain is selected as the main feature;

[0069] The information gain evaluation expression is:

[0070]

[0071] Where IG(T,A) is the information gain evaluation value of feature A, H(T) is the entropy of data set T, and T q is the subset divided according to the value q of feature A;

[0072] The identified features are systematically divided into layers, with the first layer recording the main features of the device;

[0073] Based on the specific operating data of the equipment, the subsequent layer structure is subjected to in-depth information gain evaluation and feature classification. Based on the main features, it is further refined to identify specific performance indicators, and the specific performance indicators are filled into the subsequent layers as refined features.

[0074] This approach ensures that each step of the selection is made on the basis of maximizing information, thereby improving the effectiveness of the decision tree;

[0075] Determine and construct the decision tree structure based on the feature hierarchy division, specifically:

[0076] Use the first-level features as the root node of the decision tree;

[0077] Based on each primary feature root node, secondary features are selected for splitting, and child nodes are recursively constructed for each node. Each node represents a different performance indicator of the node in the previous layer. Construction ends when all feature levels are expressed. For example, if the root node is temperature (T), child nodes can be further refined based on different temperature ranges (such as low temperature, medium temperature, and high temperature). These child nodes may include the performance of the device under different temperature conditions, such as load (L), energy consumption (P), and efficiency (E).

[0078] Nodes without child nodes are marked as leaf nodes. Leaf nodes store the performance status of the device. When recursively constructing child nodes, the dataset is continuously split into multiple subsets until the preset stopping conditions are reached, such as reaching the maximum depth, insufficient number of samples, or information gain tending to zero.

[0079] Each node split is based on the value range of the current feature, and the priority of the split is determined according to the influence of each feature.

[0080] A decision tree is established by comparing all devices in the device set M, which is marked as a decision tree set.

[0081] Reference Figure 3 As shown in the figure, feature clustering analysis is performed on the decision tree structure, similar features are merged into feature groups, new synthetic features are generated based on the feature groups, and adaptive structural adjustment of the decision tree is performed. Specifically, the following steps are involved:

[0082] Obtain a set of decision trees, and perform complexity evaluation on the decision trees in the set based on the scale complexity measurement algorithm and leaf node distribution;

[0083] The above decision tree complexity evaluation calculation expression is:

[0084]

[0085] Where, C(T i) is the complexity evaluation value of the i-th decision tree, n is the total number of decision tree sets, m is the number of subtrees of the i-th decision tree, L(Tc j ) is the number of leaf nodes in the jth subtree, L total is the total number of child nodes, σ j 2 is the variance of the number of nodes at each level in the j-th subtree, c is the number of feature categories, p k is the proportion of the k-th feature category in the leaf node;

[0086] Set a complexity adjustment threshold, filter out decision trees whose complexity calculation results exceed the complexity adjustment threshold, and establish a complex decision tree set V based on the screening results;

[0087] Perform feature similarity analysis on the decision trees in the complex decision tree set V. By calculating the Pearson correlation coefficient between the features, the similarity between the features can be identified. By calculating the Pearson correlation coefficient, the similarity between the features can be quantified. A value close to 1 indicates a positive correlation, and a value close to -1 indicates a negative correlation.

[0088] The K-means clustering algorithm is used for iterative calculation to reduce the feature dimension of each complex decision tree and divide similar features into multiple cluster groups. The K-means clustering algorithm expression is:

[0089]

[0090] Where, e is the number of clusters, X l Indicates belonging to cluster C l The characteristics of μ i As the cluster center, the optimal clustering solution of the features is finally determined through continuous iterative calculation;

[0091] Using feature similarity as weight, the clustering results in each feature cluster group are weighted and summarized to generate new synthetic features that can comprehensively reflect the performance of the device;

[0092] The information gain of the new synthetic features is re-evaluated, and the decision tree structure is adaptively adjusted according to the order of information gain.

[0093] Through feature clustering analysis, the complexity of device features is effectively reduced, reducing feature redundancy. By analyzing similar features, clustering, and synthesizing new features, the generated synthetic features demonstrate good discriminative power in the decision tree model, enhancing the model's interpretability and computational efficiency.

[0094] Reference Figure 4As shown in the figure, the initial weight of the importance of the features in the node is set, and the online learning algorithm is used to dynamically adjust the importance weight of the features based on real-time data feedback. Specifically,

[0095] The information gain calculation result corresponding to each node feature in the decision tree is used as the initial weight of the feature's importance;

[0096] Real-time monitoring and comparison of running devices in the device set M, obtaining real-time device operation data of the running devices, inputting the real-time device operation data into the decision tree, and obtaining the device performance status prediction results;

[0097] By comparing the deviation between the current device performance state and the expected performance state, the contribution feedback of different features to the state deviation is quantified, and the contribution deviation feedback value of each feature is calculated. The specific expression is:

[0098]

[0099] Where Δω z is the contribution bias feedback value of the z-th feature, η is the learning rate, is the difference between the actual equipment performance status y and the expected performance status Deviation quantization value, is the model prediction output For the zth feature X z sensitivity;

[0100] Among them, the feature sensitivity value is set based on the proportion of the initial weight of the feature's importance to the cumulative value of the initial weight of all feature importances;

[0101] Based on the real-time feedback of feature contribution deviation results, the importance weight of the corresponding feature is dynamically adjusted according to the contribution deviation ratio and scale. The weight adjustment can be regarded as the model correcting its understanding of the device through learning, so as to better adapt to the new device status.

[0102] Real-time feedback processing isn't a one-time process. As the device continues to operate and data continues to flow in, the model will continually fine-tune the weights of each feature. This dynamic process ensures that the model can maintain accurate predictions of device performance even as device data continues to change.

[0103] Reference Figure 5 As shown, the real-time operation data of the device to be compared is obtained, and based on the feature weights of the decision path nodes of each decision tree, the optimal feature data comparison scheme is selected, specifically including:

[0104] Obtain the real-time operation data of the device to be compared, input the real-time operation data into all decision trees in the decision tree set, and compare and analyze the real-time operation data with the nodes in the decision tree based on the node characteristics and splitting rules in the decision tree. During the analysis process, each node on the path will be split according to the characteristic value of the input data. By comparing the characteristic value and the splitting point, the decision path of the data flow is determined.

[0105] For a given device D to be compared, after inputting the real-time operation data of the device, its decision path set is PD = {PD1…PD r …PD m}, where PD r The decision path determined for the rth decision tree;

[0106] Extract the importance weights of the features of the nodes passed by each decision path in the decision path set, calculate the cumulative value of the feature importance weights of each decision path, select and record the decision path with the largest cumulative weight value as the optimal decision path, and mark the feature expression set corresponding to the nodes passed by the optimal path as the optimal feature data comparison scheme of device D.

[0107] This process enables efficient analysis of device operational data and selects the most representative feature sets from each device's feature data comparison, improving the accuracy and real-time nature of the comparison analysis. Through dynamic feedback and optimization, the decision tree structure is continuously adjusted and improved to ensure the accuracy and efficiency of the comparison process.

[0108] Reference Figure 6 As shown in the figure, based on the selection of the historical optimal feature data comparison scheme, the dynamic deletion of redundant nodes in the decision tree specifically includes:

[0109] Set the time window size for decision tree nodes. This time window can be set based on actual business needs, for example, hourly, daily, or weekly. The size of the time window affects the frequency of calculating the historical optimal decision path and optimizing the decision tree. The time window can be fixed or dynamically adjusted based on model performance. When model performance (such as prediction accuracy or operational efficiency) fluctuates, the time window can be shortened or extended accordingly to adapt to changing business needs.

[0110] In each new time window, the historical optimal decision path from the previous time window is reproduced. By comparing the nodes of each historical optimal path, it is possible to determine which nodes appear repeatedly in multiple time windows or multiple decision trees.

[0111] The decision tree is recreated by reproducing the nodes along the optimal decision path. The number of times the historical optimal decision path passes through each node is marked as the node's active count. A node with more frequent occurrences contributes more to the model's overall prediction. A minimum threshold for node active count is set. Typically, the threshold can be adjusted dynamically based on the actual application scenario and model complexity. For example, if model efficiency is prioritized, a higher threshold can be set to increase pruning efforts; if model prediction accuracy is prioritized, a lower threshold can be set to retain more nodes.

[0112] Nodes whose activity counts are less than the minimum active count threshold are pruned, and the corresponding feature information is deleted. The purpose of node pruning is to reduce model complexity and improve computational efficiency by removing redundant or inefficient split nodes. Pruning directly removes nodes from the decision tree and adjusts the tree structure. Deleting features not only reduces the model's dimensionality but also prevents redundant features from interfering with the model's prediction process. During the pruning process, important feature nodes are ensured to remain intact, ensuring that the model's predictive performance is not weakened.

[0113] Furthermore, the present solution also proposes a device feature data comparison and extraction method storage medium based on a decision tree, on which a computer-readable program is stored. When the computer-readable program is called, the above-mentioned device feature data comparison and extraction method based on a decision tree is executed.

[0114] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state disk (SSD).

[0115] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A device feature data comparison and extraction method based on a decision tree, characterized in that: include: Perform hierarchical processing on device features, screen out the main factors affecting device performance, and build a hierarchical decision tree structure; Obtain a set of decision trees, and perform complexity evaluation on the decision trees in the set based on the scale complexity measurement algorithm and leaf node distribution; The above decision tree complexity evaluation calculation expression is: Where, is the complexity evaluation value of the i-th decision tree, n is the total number of decision tree sets, is the number of subtrees of the i-th decision tree, is the number of leaf nodes in the j-th subtree, is the total number of child nodes, is the variance of the number of nodes at each level in the j-th subtree, is the number of feature categories, is the proportion of the k-th feature category in the leaf node; Set a complexity adjustment threshold, filter out decision trees whose complexity calculation results exceed the complexity adjustment threshold, and establish a complex decision tree set V based on the screening results; Perform feature similarity analysis on the decision trees in the complex decision tree set V, and identify the similarity between features by calculating the Pearson correlation coefficient between the features; The K-means clustering algorithm is used for iterative calculation to reduce the feature dimension of each complex decision tree and divide similar features into multiple cluster groups; Taking feature similarity as weight, perform weighted aggregation on the clustering results in each feature clustering group to generate new synthetic features; Re-evaluate the information gain of the new synthetic features and adjust the decision tree structure adaptively according to the order of information gain; Set the initial importance weight of the features in the node, and use the online learning algorithm to dynamically adjust the importance weight of the features based on real-time data feedback; Obtain the real-time operating data of the equipment to be compared, and select the optimal feature data comparison scheme based on the feature weights of the decision path nodes of each decision tree; Based on the selection of the historical optimal feature data comparison scheme, redundant nodes of the decision tree are dynamically deleted.

2. The device feature data comparison and extraction method based on a decision tree according to claim 1 is characterized in that: The hierarchical processing of device features, screening out the main factors affecting device performance, and constructing a hierarchical decision tree structure specifically includes: Get the comparison device set M; Perform statistical analysis on the historical operating data of the devices in the device set M, screen out the main factors affecting the performance of the devices, and establish the corresponding relationship between the devices and their main feature sets; Among them, the main features are evaluated by the information gain algorithm on all the main factors affecting the performance of the equipment, and the feature with the largest information gain is selected as the main feature; The identified features are systematically divided into layers, with the first layer recording the main features of the device; Based on the specific operating data of the equipment, the subsequent layer structure is subjected to in-depth information gain evaluation and feature classification. Based on the main features, it is further refined to identify specific performance indicators, and the specific performance indicators are filled into the subsequent layers as refined features. Determine and construct the decision tree structure based on the feature hierarchy division, specifically: Use the first-level features as the root node of the decision tree; Based on each main feature root node, select secondary features for splitting in turn, and recursively construct the child nodes of each node. Each node represents a different performance indicator of the node on the previous layer. The construction ends when all feature levels are expressed. Mark nodes without child nodes as leaf nodes, which store the performance status of the device. A decision tree is established by comparing all devices in the device set M, which is marked as a decision tree set.

3. The device feature data comparison and extraction method based on a decision tree according to claim 2 is characterized in that: The initial weight of the importance of the feature in the node is set, and the online learning algorithm is used to dynamically adjust the importance weight of the feature according to real-time data feedback. Specifically, the following steps are performed: The information gain calculation result corresponding to each node feature in the decision tree is used as the initial weight of the feature's importance; Real-time monitoring and comparison of running devices in the device set M, obtaining real-time device operation data of the running devices, inputting the real-time device operation data into the decision tree, and obtaining the device performance status prediction results; By comparing the deviation between the current device performance state and the expected performance state, the contribution feedback of different features to the state deviation is quantified, and the contribution deviation feedback value of each feature is calculated. The specific expression is: Where, is the contribution deviation feedback value of the z-th feature, is the learning rate, Is the actual device performance status Expected performance status Deviation quantization value, is the model prediction output For the zth feature sensitivity; Among them, the feature sensitivity value is set based on the proportion of the initial weight of the feature's importance to the cumulative value of the initial weight of all feature importances; Based on the real-time feedback of feature contribution deviation results, the importance weight of the corresponding feature is dynamically adjusted according to the contribution deviation ratio and scale.

4. The device feature data comparison and extraction method based on a decision tree according to claim 3 is characterized in that: The step of obtaining the real-time operation data of the device to be compared and selecting the optimal feature data comparison scheme based on the feature weights of the decision path nodes of each decision tree specifically includes: Obtain real-time operating data of the device to be compared, input the real-time operating data into all decision trees in the decision tree set, and compare and analyze the real-time operating data with the nodes in the decision tree based on the node characteristics and splitting rules in the decision tree; For a given device D to be compared, after inputting the real-time operation data of the device, its decision path set is ,in The decision path determined for the rth decision tree; Extract the importance weights of the features of the nodes passed by each decision path in the decision path set, calculate the cumulative value of the feature importance weights of each decision path, select and record the decision path with the largest cumulative weight value as the optimal decision path, and mark the feature expression set corresponding to the nodes passed by the optimal path as the optimal feature data comparison scheme of device D.

5. The device feature data comparison and extraction method based on a decision tree according to claim 4 is characterized in that: The dynamic deletion of redundant nodes in the decision tree based on the selection of the historical optimal feature data comparison solution specifically includes: Set the decision tree node adjustment time window size, and obtain the historical optimal decision path set of each decision tree record in the decision tree set in the previous time window of each adjustment time window; The optimal decision path nodes of the decision tree are overlapped and reproduced, the number of times the historical optimal decision path passes through each node is marked as the number of active times of the node, and the minimum active number threshold of the node is set; Nodes whose active times are less than the minimum active times threshold are pruned, and the corresponding feature information of the nodes is deleted.

Images