Method and System for Heterogeneous Network Fusion and Intelligent Handoff Control of Tripartite Terminal Devices

Abstract

The present invention provides a method and system for heterogeneous network fusion and intelligent handover control of a three-in-one terminal device, which relates to the technical field of terminal devices, including obtaining user biometric data for identity recognition and triggering a network handover evaluation process; analyzing the user's historical network handover records to establish a prediction model; monitoring network performance indicators in real time and fusing them with the prediction results to generate decision parameters; formulating a data transmission buffering strategy and allocating resources based on a scoring value, and selecting an optimal index structure to achieve data migration and resource reconfiguration. This method effectively improves the handover efficiency of terminal devices and the data transmission stability in a heterogeneous network environment.

Description

Technical Field

[0001] The present invention relates to terminal equipment technology, and in particular to a method and system for controlling heterogeneous network fusion and intelligent switching of three-in-one terminal equipment. Background Art

[0002] With the rapid development of mobile communication technology and the popularization of intelligent terminal devices, three-in-one terminal devices that integrate multiple communication technologies such as cellular networks, WiFi and Bluetooth have appeared on the market. Such devices can work in different network environments, providing users with more flexible network access methods and more stable communication experience. Three-in-one terminal devices can switch between heterogeneous networks according to user needs and network environment conditions to ensure communication quality and user experience.

[0003] Existing three-in-one terminal devices mainly rely on simple signal strength comparison or preset rules for heterogeneous network switching, lacking in-depth consideration of user personalized needs and behavioral characteristics. Traditional network switching technology usually makes decisions based on predefined static parameters and cannot make intelligent adjustments based on actual usage scenarios and dynamic network environments, resulting in unnecessary frequent switching or switching delays during the switching process. In addition, the existing technology lacks an effective identity authentication mechanism during the network switching process, making it difficult to ensure the security and reliability of network switching.

[0004] The main defects of the existing technology include: first, the lack of identity recognition and security authentication mechanism based on user biometrics, which makes the network switching process have security risks and cannot effectively prevent unauthorized access and identity fraud risks; second, the existing technology fails to fully consider the user's historical usage habits and behavior patterns, and cannot make personalized network switching decisions based on the user's time pattern, location information and business type characteristics, resulting in poor user experience; finally, there is a lack of effective data transmission buffering strategy and resource allocation mechanism in the network switching process, which leads to data loss, service interruption and resource waste in the switching process, and cannot guarantee business continuity and service quality. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for controlling heterogeneous network integration and intelligent switching of a three-in-one terminal device, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides a three-in-one terminal device heterogeneous network integration and intelligent switching control method, including:

[0008] Obtaining biometric data of a user in a three-in-one terminal device, and inputting the biometric data into corresponding feature extraction models to generate a multi-dimensional feature vector;

[0009] Calculate the similarity between the multi-dimensional feature vector and a pre-set user biometric template library to obtain a credibility score for user identity recognition. When the credibility score is higher than a pre-set security threshold, trigger the network switching evaluation process;

[0010] In the network switching evaluation process, analyze the user's historical network switching records, and perform correlation modeling on the user's time pattern features, location information features, and service type features to obtain a user network switching probability prediction model;

[0011] Based on the output result of the user network switching probability prediction model, real-time monitor the performance metric data of the currently connected network;

[0012] Perform weighted fusion calculation on the performance metric data and the output result of the user network switching probability prediction model to generate network switching decision parameters. Determine the optimal switching time window according to the network switching decision parameters, and formulate a corresponding data transmission buffering strategy;

[0013] Calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; and select the index structure with the best performance in combination with the migration priority and the index efficiency evaluation value to achieve data migration and resource reconfiguration.

[0014] Calculating the similarity between the multi-dimensional feature vector and a pre-set user biometric template library to obtain a credibility score for user identity recognition. When the credibility score is higher than a pre-set security threshold, triggering the network switching evaluation process includes:

[0015] Calculate the similarity between the multi-dimensional feature vector and a pre-set biometric template, and perform decision-level fusion on the recognition results of different features using the D-S evidence theory to generate a fused credibility score. The decision-level fusion solves the feature contradiction problem by calculating the conflict factor between features;

[0016] Compare the fused credibility score with the current security threshold. When the fused credibility score is greater than the current security threshold, confirm that the user identity is valid and output the identity recognition result to re-trigger the network switching evaluation process.

[0017] Performing decision-level fusion on the recognition results of different features using the D-S evidence theory to generate a fused credibility score. The decision-level fusion solves the feature contradiction problem by calculating the conflict factor between features includes:

[0018] Construct a feature fusion model based on the D-S evidence theory, substitute the feature fusion model into the orthogonal sum operation formula to obtain a feature combination result, and the feature combination result reflects the support degree distribution of different features;

[0019] Calculate the feature conflict part that appears in the result of the feature combination, quantify the degree of feature conflict based on the Euclidean distance function, and the Euclidean distance function calculates the sum of the squares of the differences in the basic probability assignment values of different features to obtain the feature difference degree;

[0020] Introduce the feature difference degree into the conflict factor calculation formula of the D-S evidence theory, and use an exponential decay function to correct the original conflict factor to obtain a corrected conflict factor, and the corrected conflict factor can accurately reflect the degree of contradiction between features;

[0021] Perform normalization processing on the result of the feature combination based on the corrected conflict factor, construct a D-S fusion rule, and perform evidence synthesis on the normalized result of the feature combination and the corrected conflict factor to obtain a fusion credibility score.

[0022] Perform weighted fusion calculation on the performance index data and the output result of the user network handover probability prediction model to generate a network handover decision parameter, determine the optimal handover time window according to the network handover decision parameter, and formulate a corresponding data transmission buffering strategy including:

[0023] Perform weighted fusion calculation on the performance index data and the output result of the user network handover probability prediction model, where the weighted fusion calculation dynamically adjusts the importance of the performance index data and the output result using an adaptive weight coefficient to generate a network handover decision parameter;

[0024] Set a network state evaluation window according to the network handover decision parameter, continuously judge the handover conditions within the network state evaluation window, and determine the optimal handover time window when the network handover decision parameter continuously meets the preset handover threshold;

[0025] Formulate a data transmission buffering strategy according to the time range of the optimal handover time window and the data transmission state of the current network, and the data transmission buffering strategy realizes smooth data transmission during network handover by adjusting the buffer size of data transmission.

[0026] Formulate a data transmission buffering strategy according to the time range of the optimal handover time window and the data transmission state of the current network including:

[0027] Construct a network state feature vector, and form a historical state sequence by combining the network state feature vector with the historical state feature vectors within the historical observation window;

[0028] Input the historical state sequence into a long short-term memory network, and perform temporal feature extraction on the historical state sequence through the forget gate, input gate, and output gate of the long short-term memory network to obtain a hidden state sequence;

[0029] Construct an attention mechanism based on the hidden state sequence, calculate an energy score through the attention mechanism, normalize the energy score to obtain an attention weight, and generate a context vector based on the weighted sum of the attention weight and the historical state sequence;

[0030] Establish a comprehensive optimization objective function based on the context vector. The comprehensive optimization objective function includes average transmission delay, resource utilization rate, and throughput variance, and adaptively adjust the learning rate according to the current error;

[0031] Obtain the predicted optimal window size based on the historical state sequence, the context vector, and the comprehensive optimization objective function, and generate a data transmission buffering strategy based on the difference between the optimal window size and the window size at the previous moment.

[0032] Calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; and select the index structure with the optimal performance by combining the migration priority and the index efficiency evaluation value, and implement data migration and resource reconfiguration including:

[0033] Obtain a data score value by weighting various parameters in the data transmission buffering strategy and the data resources;

[0034] Compare the data score value with the hot data threshold and the cold data threshold, divide the data resources higher than the hot data threshold into the hot data layer, divide the data resources lower than the cold data threshold into the cold data layer, and divide the data resources between the hot data threshold and the cold data threshold into the warm data layer;

[0035] For the data resources in each data layer, calculate the migration priority based on the ratio of its access frequency to the storage space, the data value weight, and the normalized time decay exponent, and the normalized time decay exponent decreases exponentially with time growth;

[0036] Monitor the query operations in each data layer, calculate the first weighted sum of the query performance and the query weight, and calculate the second weighted sum of the resource consumption and the resource utilization rate, and divide the first weighted sum by the second weighted sum to obtain the index efficiency evaluation value;

[0037] Select the index structure with the optimal performance according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, calculate the resource allocation quota according to the index structure with the optimal performance after the migration is completed, and update the resource configuration of the storage level.

[0038] Select the index structure with the optimal performance according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, and calculate the resource allocation quota according to the index structure with the optimal performance after the migration is completed, and update the resource configuration of the storage level, including:

[0039] Obtain the load data of the index shards involved in the query operation, calculate the average value of the load data, divide the sum of the squared deviations of the load data from the average value by the number of shards and take the negative value after taking the square root to calculate the load balance degree;

[0040] Multiply the index efficiency evaluation value and the load balance degree by the first balance factor and the second balance factor respectively, and subtract the product of the index maintenance overhead and the third balance factor to obtain the fitness of the current index structure;

[0041] Calculate the fitness of the newly constructed candidate index structure, subtract the fitness of the current index structure from the fitness of the candidate index structure, and then divide by the fitness of the current index structure to obtain the relative change rate of fitness;

[0042] When the relative change rate of fitness is greater than the preset update threshold, calculate the ratio of the historical access popularity of the data object to the data size, and use the product of the ratio, the business value weight, and the time decay exponent as the data migration priority;

[0043] Calculate the proportion of the data migration priority in the total sum of the migration priorities of all data objects, and allocate the product of the proportion, the total resource pool, and the level adjustment coefficient to the corresponding storage level for updating the resource configuration of the storage level.

[0044] In the second aspect of the embodiments of the present invention,

[0045] Provide a heterogeneous network fusion and intelligent switching control system for a three-in-one terminal device, including:

[0046] The first unit is used to obtain the biometric data of the user in the three-in-one terminal device, and input the biometric data into the corresponding feature extraction model to generate a multi-dimensional feature vector;

[0047] The second unit is used to calculate the similarity between the multi-dimensional feature vector and the preset user biometric template library to obtain the credibility score of user identity recognition. When the credibility score is higher than the preset security threshold, trigger the network switching evaluation process;

[0048] The third unit is used to analyze the historical network switching records of the user in the network switching evaluation process, and perform correlation modeling on the time pattern feature, location information feature, and service type feature of the user to obtain a user network switching probability prediction model;

[0049] A fourth unit, configured to monitor in real time performance metric data of a currently accessed network based on an output result of the user network handover probability prediction model;

[0050] A fifth unit, configured to perform weighted fusion calculation on the performance metric data and the output result of the user network handover probability prediction model to generate network handover decision parameters, determine an optimal handover time window according to the network handover decision parameters, and formulate a corresponding data transmission buffering strategy;

[0051] A sixth unit, configured to calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; and select an index structure with the optimal performance in combination with the migration priority and an index efficiency evaluation value to implement data migration and resource reconfiguration.

[0052] In a third aspect of the embodiments of the present invention,

[0053] There is provided an electronic device, including:

[0054] A processor;

[0055] A memory for storing instructions executable by the processor;

[0056] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0057] In a fourth aspect of the embodiments of the present invention,

[0058] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0059] The beneficial effects of the present application are as follows:

[0060] By combining biometric recognition and network behavior analysis, the present invention realizes the intelligence and personalization of the network handover process, and improves the adaptability and user experience of a terminal device in a heterogeneous network environment.

[0061] The present invention constructs a network handover probability prediction model based on the historical behavior of a user, and performs weighted fusion on real-time performance metrics and a prediction result, making the network handover decision more accurate and efficient, and reducing the latency and data packet loss rate in the network handover process.

[0062] The present invention designs a data transmission buffering strategy and a resource allocation mechanism based on a score value. Through optimization of the migration priority and the index structure, the data migration efficiency in the network handover process is significantly improved, and the service continuity and service quality are guaranteed. Description of the Drawings

[0063] Figure 1 Schematic flowchart of the heterogeneous network fusion and intelligent handover control method for the three-in-one terminal device according to an embodiment of the present invention;

[0064] Figure 2 Schematic flowchart of the feature fusion based on the D-S evidence theory according to an embodiment of the present invention;

[0065] Figure 3 Schematic diagram for comparing the resource utilization rates in different handover scenarios according to an embodiment of the present invention;

[0066] Figure 4 Schematic flowchart of the complete process from data score calculation to final resource configuration update according to an embodiment of the present invention;

[0067] Figure 5 Schematic diagram of the load balancing performance of different methods according to an embodiment of the present invention under the condition of increasing query load. Detailed implementation manners

[0068] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0070] Figure 1 Schematic flowchart of the heterogeneous network fusion and intelligent handover control method for the three-in-one terminal device according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0071] Obtain the biometric data of the user in the three-in-one terminal device, and input the biometric data into the corresponding feature extraction model to generate a multi-dimensional feature vector;

[0072] Calculate the similarity between the multi-dimensional feature vector and the preset user biometric template library to obtain the credibility score for user identity recognition. When the credibility score is higher than the preset security threshold, trigger the network handover evaluation process;

[0073] In the network handover evaluation process, analyze the historical network handover records of the user, and perform correlation modeling on the time regularity feature, location information feature, and service type feature of the user to obtain a user network handover probability prediction model;

[0074] Based on the output result of the user network handover probability prediction model, the performance metric data of the currently accessed network is monitored in real time;

[0075] The performance metric data is weighted and fused with the output result of the user network handover probability prediction model to generate network handover decision parameters, determine the optimal handover time window according to the network handover decision parameters, and formulate corresponding data transmission buffering strategies;

[0076] Calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain the migration priority; and select the index structure with the optimal performance in combination with the migration priority and the index efficiency evaluation value to achieve data migration and resource reconfiguration.

[0077] In an alternative embodiment, calculating the similarity between the multi-dimensional feature vector and a pre-set user biometric template library to obtain a credibility score for user identity recognition. When the credibility score is higher than a preset security threshold, triggering the network handover evaluation process includes:

[0078] Calculate the similarity between the multi-dimensional feature vector and a pre-set biometric template, and perform decision-level fusion on the recognition results of different features using the D-S evidence theory to generate a fused credibility score. The decision-level fusion solves the feature contradiction problem by calculating the conflict factor between features;

[0079] Compare the fused credibility score with the current security threshold. When the fused credibility score is greater than the current security threshold, confirm that the user identity is valid and output the identity recognition result to re-trigger the network handover evaluation process.

[0080] For fingerprint features, the device extracts the minutiae features of the fingerprint, including key feature points such as endpoints and bifurcation points. Suppose the set of extracted feature points contains 32 feature points, and each feature point is represented by a position coordinate and a direction angle. Match these feature points with the corresponding feature points in the stored template, and calculate the ratio of the number of matching points to the total number of feature points as the fingerprint similarity. If 26 feature points are matched, the fingerprint similarity is 26 / 32 = 0.8125.

[0081] For face features, the device extracts the key point features of the face, such as the relative position relationship and texture features of 128 feature points such as the corners of the eyes, the tip of the nose, and the corners of the mouth. Calculate the Euclidean distance between these feature points and the difference from the stored template, and record the normalized similarity as 0.78.

[0082] For voiceprint features, the device extracts the spectral features of the sound and converts them into a 512-dimensional feature vector. By calculating the cosine similarity between this feature vector and the stored template feature vector, the voiceprint similarity is obtained as 0.75.

[0083] The device has obtained the similarity calculation results of three biometric features: fingerprint similarity 0.8125, face similarity 0.78, and voiceprint similarity 0.75. However, directly using these similarities for decision-making may lead to inaccurate results because the reliabilities of different features vary in different environments. For example, in low-light environments, the reliability of face recognition will decrease; in noisy environments, the accuracy of voiceprint recognition will be affected.

[0084] The device assigns basic probability assignment (BPA) values to each feature according to the current environmental conditions. Assume that in the current environment, the reliability of fingerprint recognition is relatively high, face recognition is the second, and voiceprint recognition is relatively low, and weights 0.5, 0.3, and 0.2 are assigned respectively.

[0085] Calculate the BPA value of fingerprint recognition. For the hypothesis of "user identity is valid", the BPA value is the product of the similarity and the weight, that is, 0.8125×0.5 = 0.4063; for the hypothesis of "user identity is invalid", the BPA value is the product of (1 - similarity) and the weight, that is, (1 - 0.8125)×0.5 = 0.0938; the remaining uncertain part is 1 - (0.4063 + 0.0938) = 0.5.

[0086] Similarly, calculate the BPA value of face recognition. For the hypothesis of "user identity is valid", the BPA value is 0.78×0.3 = 0.234; for the hypothesis of "user identity is invalid", the BPA value is (1 - 0.78)×0.3 = 0.066; the remaining uncertain part is 1 - (0.234 + 0.066) = 0.7.

[0087] Calculate the BPA value of voiceprint recognition. For the hypothesis of "user identity is valid", the BPA value is 0.75×0.2 = 0.15; for the hypothesis of "user identity is invalid", the BPA value is (1 - 0.75)×0.2 = 0.05; the remaining uncertain part is 1 - (0.15 + 0.05) = 0.8.

[0088] During the fusion process, it is necessary to handle the possible conflicts between features. The device calculates the conflict factor k12 between fingerprint and face features by multiplying the BPA value of "user identity is valid" in fingerprint recognition by the BPA value of "user identity is invalid" in face recognition, and then adding the product of the BPA value of "user identity is invalid" in fingerprint recognition and the BPA value of "user identity is valid" in face recognition, that is:

[0089] 0.4063×0.066 + 0.0938×0.234 = 0.0486。

[0090] The device calculates the BPA value after the fusion of the device calculation fingerprint and the face feature. For the hypothesis of "valid user identity", the fused BPA value is (0.4063×0.234 + 0.4063×0.7 + 0.5×0.234) / (1 - 0.0486) = 0.5421; for the hypothesis of "invalid user identity", the fused BPA value is:

[0091] (0.0938×0.066 + 0.0938×0.7 + 0.5×0.066) / (1 - 0.0486) = 0.1249;

[0092] The remaining uncertain part is 1 - (0.5421 + 0.1249) = 0.333.

[0093] The voiceprint feature is incorporated into the calculation. The conflict factor k123 between the calculation fingerprint, the face fusion result and the voiceprint feature is calculated in a similar way, and k123 = 0.0398 is obtained.

[0094] The BPA value after the complete fusion of the three features is calculated. For the hypothesis of "valid user identity", the final fused BPA value is 0.6324; for the hypothesis of "invalid user identity", the final fused BPA value is 0.1421; the remaining uncertain part is 0.2255.

[0095] Since the final fused BPA value of "valid user identity" is 0.6324, which is greater than the BPA value of "invalid user identity" of 0.1421, and the uncertain part is small, the device can conclude that the user identity is valid. This 0.6324 is the fusion credibility score.

[0096] The device compares the obtained fusion credibility score 0.6324 with the preset security threshold. Assume that the current security threshold is 0.6. Since the fusion credibility score is greater than the security threshold, the device confirms that the user identity is valid, outputs the identity recognition result, and triggers the network switching evaluation process.

[0097] In practical applications, the security threshold can be dynamically adjusted according to different security level requirements. For example, when processing highly sensitive financial transaction data, a higher security threshold such as 0.8 may be set; while when processing ordinary browsing data, a lower security threshold such as 0.5 may be used.

[0098] Compared with traditional single - feature recognition or simple multi - feature combination methods, in this embodiment, decision - level fusion is performed through the D - S evidence theory, fully considering the reliability differences of different features in various environments and the conflict situations among features, achieving more reliable and robust user identity recognition, and providing a solid foundation for the subsequent network handover evaluation process.

[0099] In an alternative embodiment, the D - S evidence theory is used to perform decision - level fusion on the recognition results of different features to generate a fusion credibility score. The decision - level fusion solves the feature contradiction problem by calculating the conflict factor among features, including:

[0100] Construct a feature fusion model based on the D - S evidence theory, substitute the feature fusion model into the orthogonal sum operation formula to obtain the feature combination result, and the feature combination result reflects the support degree distribution of different features;

[0101] Calculate the feature conflict part in the feature combination result, quantify the feature conflict degree based on the Euclidean distance function, and the Euclidean distance function calculates the sum of squares of the differences in the basic probability assignment values of different features to obtain the feature difference degree;

[0102] Introduce the feature difference degree into the conflict factor calculation formula of the D - S evidence theory, and use an exponential decay function to correct the original conflict factor to obtain the corrected conflict factor, and the corrected conflict factor can accurately reflect the contradiction degree between features;

[0103] Based on the corrected conflict factor, perform normalization processing on the feature combination result, construct a D - S fusion rule, and perform evidence synthesis on the normalized feature combination result and the corrected conflict factor to obtain a fusion credibility score.

[0104] When the device recognizes biometric features, there may be contradictions in the recognition results of different features. For example, the fingerprint recognition result highly supports user A, while the face recognition result supports user B. To solve this problem, the device needs to establish a feature fusion model and reasonably handle feature conflicts.

[0105] The device constructs a feature fusion model based on the D - S evidence theory. Taking fingerprint, face, and voiceprint biometric features as examples, the device assigns basic probability assignment values (BPA) to each feature. Suppose the support degree of fingerprint recognition for user A is 0.75, the support degree for user B is 0.05, and the uncertainty is 0.2; the support degree of face recognition for user A is 0.65, the support degree for user B is 0.1, and the uncertainty is 0.25; the support degree of voiceprint recognition for user A is 0.6, the support degree for user B is 0.15, and the uncertainty is 0.25.

[0106] The device substitutes these basic probability assignment values into the orthogonal sum operation for feature combination. During the orthogonal sum operation, the device calculates the combined result of fingerprint features and face features. The combined support degree for user A is equal to the product of the fingerprint supporting user A and the face supporting user A, plus the product of the fingerprint supporting user A and the face being uncertain, plus the product of the fingerprint being uncertain and the face supporting user A, that is:

[0107] 0.75×0.65 + 0.75×0.25 + 0.2×0.65 = 0.7175.

[0108] The device calculates that the combined support degree for user B is 0.0675, and the uncertain part is 0.05.

[0109] During the feature combination process, the device discovers that there are parts of feature conflicts, that is, the fingerprint supports user A while the face supports user B, or the fingerprint supports user B while the face supports user A. This part of the conflict value is equal to the product of the fingerprint supporting user A and the face supporting user B, plus the product of the fingerprint supporting user B and the face supporting user A, that is 0.75×0.1 + 0.05×0.65 = 0.1075. In the traditional D-S theory, this part of the conflict is simply discarded, which may lead to inaccurate fusion results.

[0110] To accurately quantify the degree of feature conflict, the device uses the Euclidean distance function to calculate the feature difference degree. The device calculates that the square of the difference in the support degree for user A between the fingerprint and the face features is (0.75 - 0.65)² = 0.01, the square of the difference in the support degree for user B is (0.05 - 0.1)² = 0.0025, and the square of the difference in the uncertain part is (0.2 - 0.25)² = 0.0025. Taking the square root of the sum of these squared differences, the feature difference degree is 0.1225.

[0111] The device introduces the feature difference degree into the calculation of the conflict factor in the D-S theory. The traditional conflict factor K = 0.1075, and the device modifies it using an exponential decay function: substituting the feature difference degree 0.1225 into the exponential decay function, assuming the device selects an exponential decay function with an adjustment parameter of 1.5, that is exp(-1.5×0.1225) = 0.8315. The device multiplies this value by the original conflict factor to obtain the modified conflict factor K' = 0.1075×0.8315 = 0.0894. This modification makes the conflict factor more accurately reflect the degree of contradiction between features. The greater the feature difference, the more significant the modification.

[0112] The device normalizes the feature combination result based on the modified conflict factor. During the normalization process, the device divides the combined support degree by (1 - K'), that is, the normalized support degree for user A is:

[0113] 0.7175 / (1 - 0.0894) = 0.7879;

[0114] The normalized support degree for User B is:

[0115] 0.0675 / (1 - 0.0894) = 0.0741; The uncertain part is 0.05 / (1 - 0.0894) = 0.0549.

[0116] The device further fuses the voiceprint feature with the combined result of the first two features. The device repeats the aforementioned calculation process, and the conflict factor of the three features is 0.1523, which is corrected to 0.1237 after correction. The final fusion result shows that the support degree for User A is 0.8642, the support degree for User B is 0.0853, and the uncertain part is 0.0505.

[0117] The device outputs the support degree of 0.8642 for User A as the fusion credibility score. Compared with the single feature or simple combination method, this fusion credibility score comprehensively considers the discrimination results of multiple features and their conflict situations, and the reliability is significantly improved.

[0118] To verify the effectiveness of this method, the device is tested under different environmental conditions. Under standard conditions (good lighting, quiet environment), the accuracy rates of the traditional D-S fusion method and this embodiment are 92.3% and 94.8% respectively, and the difference is not significant; but under complex conditions (weak lighting, noisy environment), the accuracy rate of the traditional method drops to 78.5%, while this embodiment still maintains a high accuracy rate of 89.2%, showing stronger robustness.

[0119] In the prior art, the traditional D-S evidence theory has great limitations in dealing with highly conflicting evidence. The standard D-S combination rule simply discards the conflicting evidence. When the contradiction between features is large, it will lead to the loss of valuable information and may even produce results contrary to intuition. For example, when the discrimination results of two features are completely opposite, the conflict factor is close to 1, and the denominator in the normalization process is close to 0, making the fusion result unstable or unreliable.

[0120] Figure 2 The flow chart of the feature fusion based on the D-S evidence theory for the embodiments of the present invention is as follows:

[0121] This figure shows a feature fusion and credibility evaluation process based on the D-S evidence theory. The whole process is divided into four main steps: First, a feature fusion model is constructed based on the D-S evidence theory. The input feature data is processed through the orthogonal sum operation formula to obtain the preliminary feature fusion result, which reflects the support degree distribution of different features. The second step is that the system calculates the conflicting part of the feature combination result, and uses the Euclidean distance function to quantify the conflict degree between features, and obtains the feature difference degree by calculating the difference of the basic probability assignment values. In the third step, the feature difference degree is introduced into the calculation of the D-S theory conflict factor, and the exponential decay function is used to correct the original conflict factor, so that the corrected conflict factor can accurately reflect the contradiction degree between features. Finally, based on the corrected conflict factor, the feature combination result is normalized, the D-S fusion rule is constructed and the evidence synthesis is performed, and finally the credibility score of the system is obtained. This process realizes the accurate evaluation of feature fusion in a step-by-step refinement and optimization manner, ensuring the reliability and accuracy of the fusion result.

[0122] The implementation means of the prior art usually adopt improvement methods such as the Yager rule or the Murphy average method. Although these methods alleviate the high conflict problem to a certain extent, they lack in-depth understanding and accurate quantification of the essence of the conflict. The Yager rule allocates all the conflicting parts to uncertainty, which may lead to a decrease in the discrimination ability after fusion; the Murphy average method simply averages and then fuses the evidence, ignoring the differences between evidence and the problem of different credibility.

[0123] The improvement starting point of this application is to deeply understand the essence of feature conflict, accurately quantify the feature difference by introducing the Euclidean distance function, and perform adaptive correction using the exponential decay function. This method can dynamically adjust the conflict factor according to the specific degree of difference between features, and effectively handle the high conflict situation while retaining valuable information.

[0124] The recognition accuracy of this method has increased by 15.3 percentage points in high conflict scenarios (conflict factor > 0.7), and by 9.7 percentage points in medium conflict scenarios (conflict factor 0.3 - 0.7). Especially in the multi-modal biometric fusion scenario, when the environmental factors cause a significant decline in the recognition effect of a certain feature, this method can intelligently adjust the contribution weights of each feature, maintain the stability of the overall recognition performance, and provide more reliable identity verification support for subsequent network switching decisions.

[0125] In an optional implementation manner, the performance index data is weighted and fused with the output result of the user network switching probability prediction model to generate network switching decision parameters. According to the network switching decision parameters, the optimal switching time window is determined, and the corresponding data transmission buffer strategy is formulated, including:

[0126] Perform weighted fusion calculation on the performance metric data and the output result of the user network handover probability prediction model, where the weighted fusion calculation dynamically adjusts the importance of the performance metric data and the output result using an adaptive weight coefficient to generate network handover decision parameters;

[0127] Set a network status evaluation window according to the network handover decision parameters, and continuously judge the handover conditions within the network status evaluation window. When the network handover decision parameters continuously meet the preset handover threshold, determine the optimal handover time window;

[0128] Formulate a data transmission buffering strategy according to the time range of the optimal handover time window and the current network data transmission status. The data transmission buffering strategy realizes smooth data transmission during network handover by adjusting the buffer size of data transmission.

[0129] Obtain the performance metric data of the current network, including parameters such as network latency, packet loss rate, signal strength, and bandwidth utilization. For example, in a certain embodiment, the data collected by the device is: network latency 85 milliseconds, packet loss rate 2.3%, signal strength -75dBm, and bandwidth utilization 65%. These raw data are normalized to convert to standardized performance metric values. The network latency score after normalization is 0.72 (the higher the value, the lower the latency and the better the performance), the packet loss rate normalization score is 0.85, the signal strength normalization score is 0.68, and the bandwidth utilization normalization score is 0.75.

[0130] Obtain the output result from the user network handover probability prediction model. This model predicts the probability of the user switching networks based on the user's historical network handover behavior and current context information. For example, according to the current time, location, and user activity type, the prediction model outputs that the probability of the user actively switching to the WiFi network within the next 5 minutes is 0.78, the probability of switching to the 4G network is 0.12, the probability of switching to the 5G network is 0.05, and the probability of remaining on the current network is 0.05.

[0131] Execute the weighted fusion calculation, and dynamically adjust the importance of the performance metric data and the output result of the prediction model using an adaptive weight coefficient. The adaptive weight coefficient is dynamically generated according to the current user behavior pattern and network environment. In scenarios with high requirements for service stability (such as video conferencing), the weight of the performance metric data is higher; while in the case where the user is about to move to a new environment, the weight of the output result of the prediction model is higher.

[0132] The user is detected to be watching a high-definition video and the location is relatively stable. Therefore, the weight assigned to the performance metric data is 0.65, and the weight assigned to the output result of the prediction model is 0.35. For the performance metric data internally, the weights of each metric are: network latency 0.4, packet loss rate 0.25, signal strength 0.2, and bandwidth utilization 0.15. The comprehensive score of the performance metrics is calculated as:

[0133] 0.72×0.4 + 0.85×0.25 + 0.68×0.2 + 0.75×0.15 = 0.75.

[0134] For the output result of the prediction model, the terminal device considers the benefits and costs of network switching. The benefit coefficient for switching to WiFi is 0.9 (the current network is 4G), the benefit coefficient for switching to 4G is 0.3 (already 4G currently), and the benefit coefficient for switching to 5G is 0.85. The weighted score of the prediction model result is calculated as:

[0135] 0.78×0.9 + 0.12×0.3 + 0.05×0.85 + 0.05×0 = 0.7561.

[0136] Fuse the comprehensive score of the performance metrics and the weighted score of the prediction model result:

[0137] 0.75×0.65 + 0.7561×0.35 = 0.7522.

[0138] This value is the network switching decision parameter, which reflects the necessity of network switching after comprehensively considering the current network performance and user behavior prediction.

[0139] Based on the calculated network switching decision parameter, the terminal device sets a network status evaluation window. The size of the evaluation window is dynamically adjusted according to the service type and network fluctuation. For delay-sensitive services such as online games, the evaluation window is small, such as 5 seconds; for delay-insensitive services such as file downloads, the evaluation window is large, such as 20 seconds. In this example, the user is watching a high-definition video, and the evaluation window set by the device is 10 seconds.

[0140] Within the set evaluation window, the terminal device continuously calculates the network switching decision parameter and compares it with a preset switching threshold. The preset switching threshold is set according to the service priority. For example, the switching threshold for high-priority services is 0.8, for medium-priority is 0.75, and for low-priority is 0.7. In this example, high-definition video is a medium-priority service, so the switching threshold is set to 0.75.

[0141] Within the 10-second evaluation window, the terminal device calculates the network switching decision parameter once per second, and the results are:

[0142] 0.7522, 0.7531, 0.7544, 0.7560, 0.7578, 0.7595, 0.7612, 0.7625, 0.7633, 0.7640.

[0143] Since the decision parameter has continuously exceeded the threshold of 0.75 starting from the fifth second, the terminal device determines that the optimal handover time window is between the 5th second and the 10th second, and selects the 7th second as the best handover moment because the growth trend of the decision parameter has stabilized at this time.

[0144] After determining the optimal handover time window, the terminal device formulates a data transmission buffering strategy based on the current network data transmission status. The device detects that the current data streams being transmitted are: a video stream of 5 Mbps and a background application data stream of 1.2 Mbps. To ensure smooth data transmission during network handover, the terminal device calculates the required buffer size.

[0145] For the video stream, considering the possible handover delay time (about 300 milliseconds) and the requirement of uninterrupted video playback, the device increases the video buffer to the data volume of 2 seconds, that is, 5 Mbps × 2 seconds = 10 Mb (about 1.25 MB). For the background application data stream, the device pauses the transmission of non-critical data, only retains the transmission of critical data and sets a buffer of 400 KB. In addition, the device pre-sends a handshake message to the new network to reduce the connection establishment time during actual handover.

[0146] When performing network handover, the terminal device operates in the following policy sequence: first, increase the critical service buffer, pause the transmission of non-critical services, establish a connection with the new network, redirect the data stream to the new network, and resume the transmission of non-critical services. The impact of the entire process on the user experience is almost imperceptible, and the video playback remains smooth without stuttering.

[0147] Network handover decisions are mainly based on simple signal strength thresholds or network quality metrics. For example, handover is triggered when the signal strength is below -85 dBm. This method lacks consideration of user behavior and service requirements, often resulting in unnecessary network handovers or inappropriate handover times, causing service interruptions.

[0148] The implementation means of the prior art usually adopt a fixed weight method to fuse various metrics, making it difficult to adapt to changes in the importance of metrics in different scenarios. For example, some methods fixedly allocate a weight of 40% to network latency, 30% to packet loss rate, and 30% to signal strength, and cannot adjust the weight allocation according to different service types. In addition, traditional methods lack a continuous judgment mechanism, and are prone to frequent handovers due to instantaneous network fluctuations, forming a "ping-pong effect".

[0149] Figure 3 Schematic diagram of resource utilization comparison for different handover scenarios in the embodiments of the present invention:

[0150] This figure shows the comparison of resource utilization under different network handover types. The horizontal axis represents the network handover types, including five handover scenarios: 4G→WiFi, WiFi→4G, 4G→5G, 5G→4G, and WiFi→5G. The vertical axis represents the percentage of resource utilization. The performance of four different implementation schemes is compared in the figure: the CPU utilization rate of this technical solution (triangle solid line), the memory utilization rate of this technical solution (square solid line), the CPU utilization rate of the traditional solution (circular dotted line), and the memory utilization rate of the traditional solution (asterisk dotted line). Judging from the data performance, this technical solution has obvious advantages in resource utilization in various network handover scenarios. Specifically, when switching from 4G to WiFi, the CPU and memory utilization rates of this technical solution are 12% and 18% respectively, while those of the traditional solution reach 18% and 27%; in a high-load scenario such as 4G→5G, the CPU and memory utilization rates of this technical solution are 16% and 23% respectively, significantly lower than 26% and 38% of the traditional solution. Especially in the scenario of WiFi→5G, the resource utilization rate of this technical solution remains at a low level (CPU 17%, memory 24%), while the resource consumption of the traditional solution increases significantly (CPU 28%, memory 42%), fully demonstrating the superiority of this technical solution in terms of resource efficiency.

[0151] The starting point of the improvement in this application is to introduce an adaptive weight coefficient, dynamically adjust the importance of each index according to the user behavior pattern and service type, and continuously judge by setting a network status evaluation window to avoid unnecessary handovers caused by short-term network fluctuations. In addition, this method also innovatively formulates a data transmission buffering strategy to achieve smooth data transmission during network handover by dynamically adjusting the buffer size.

[0152] The network handover success rate in typical usage scenarios has been increased from 89.3% to 97.6%, the number of unnecessary network handovers has been reduced by 75.2%, and the service interruption time has been shortened from an average of 280 milliseconds to 45 milliseconds. Especially in high-speed mobile scenarios (such as high-speed trains / subways), this method can anticipate changes in network conditions in advance, perform network handovers at the optimal time, and reduce the service interruption rate from 22.5% to 4.8%, significantly improving the user experience.

[0153] In an optional implementation manner, formulating a data transmission buffering strategy according to the time range of the optimal handover time window and the data transmission status of the current network includes:

[0154] Construct a network status feature vector, and form a historical status sequence by combining the network status feature vector with the historical status feature vectors within the historical observation window;

[0155] Input the historical state sequence into a long short-term memory network, and perform temporal feature extraction on the historical state sequence through the forget gate, input gate, and output gate of the long short-term memory network to obtain a hidden state sequence;

[0156] Construct an attention mechanism based on the hidden state sequence, calculate an energy score through the attention mechanism, normalize the energy score to obtain an attention weight, and generate a context vector according to the weighted sum of the attention weight and the historical state sequence;

[0157] Establish a comprehensive optimization objective function based on the context vector. The comprehensive optimization objective function includes average transmission delay, resource utilization rate, and throughput variance, and adaptively adjust the learning rate according to the current error;

[0158] Obtain the predicted optimal window size based on the historical state sequence, the context vector, and the comprehensive optimization objective function, and generate a data transmission buffer strategy based on the difference between the optimal window size and the window size at the previous moment.

[0159] Construct a network state feature vector including multi-dimensional metrics such as the current network bandwidth, end-to-end delay, packet loss rate, signal strength, network jitter, and service type. For example, the device detects that the current state is: downlink bandwidth 35Mbps, uplink bandwidth 12Mbps, end-to-end delay 75ms, packet loss rate 1.5%, signal strength -68dBm, network jitter 8ms, and the service type is video streaming media (encoded as 3). The terminal device normalizes these metrics to obtain the network state feature vector [0.35, 0.24, 0.65, 0.85, 0.72, 0.78, 0.3].

[0160] Combine the current network state feature vector with the historical state feature vectors within the historical observation window to form a historical state sequence. Assume that the size of the historical observation window is 10, then the historical state sequence includes 10 historical state feature vectors and 1 current state feature vector, a total of 11 time points of data. For the sake of illustration, the historical state feature vectors at the last 3 time points are listed here: at time [-2] [0.32, 0.22, 0.63, 0.82, 0.68, 0.75, 0.3], at time [-1] [0.33, 0.23, 0.64, 0.83, 0.70, 0.76, 0.3], at time [0] (current) [0.35, 0.24, 0.65, 0.85, 0.72, 0.78, 0.3].

[0161] The constructed historical state sequence is input into a Long Short-Term Memory network (LSTM) for processing. The LSTM extracts temporal features from the historical state sequence through forget gates, input gates, and output gates. The forget gate determines which information needs to be removed from the cell state, the input gate decides which information to update, and the output gate determines which information to output. In this example, the LSTM network contains 64 hidden units. Through these gating mechanisms, the LSTM can capture the long-term dependencies of network state changes and output a hidden state sequence.

[0162] For each time point t, the LSTM outputs the corresponding hidden state vector ht. Assume the dimension of the hidden state vector is 64. After processing, the terminal device obtains the hidden state vectors at 11 time points, forming a hidden state sequence H = [h0, h1, h2, ..., h10]. For simplicity of description, only the first 3 values of the hidden state vectors are listed here. For example, h8 = [0.52, 0.38, 0.41, ...], h9 = [0.54, 0.39, 0.42, ...], h10 = [0.55, 0.41, 0.43, ...].

[0163] Based on the obtained hidden state sequence, the terminal device constructs an attention mechanism. The attention mechanism assigns weights to each historical state by calculating the correlation between the current hidden state and the historical hidden states. Specifically, the terminal device calculates the energy score ei using the dot product (or other similarity calculation methods) between the current hidden state h10 and each historical hidden state hi. For example, e8 = h10·h8 = 0.963, e9 = h10·h9 = 0.978 (dot product is used here for simplified calculation).

[0164] After calculating all the energy scores, the terminal device performs normalization to obtain the attention weights. The normalization uses the softmax function to make the sum of all weights equal to 1. In this example, the normalized attention weights are:

[0165] α0 = 0.05, α1 = 0.06, ..., α8 = 0.13, α9 = 0.15, α10 = 0.17.

[0166] These weights reflect the importance of each historical state for the current prediction.

[0167] A context vector is generated based on the weighted sum of the attention weights and the historical state sequence. The context vector c is equal to the sum of the products of each historical hidden state and its corresponding attention weight. The historical states with high weights will contribute more to the context vector, enabling the model to focus on more relevant historical information.

[0168] After obtaining the context vector, the terminal device establishes a comprehensive optimization objective function based on this vector. This function includes three key metrics: average transmission delay, resource utilization, and throughput variance. In this example, the weight of the average transmission delay is 0.5, the weight of the resource utilization is 0.3, and the weight of the throughput variance is 0.2. The terminal device normalizes these three metrics and sums them with weights to obtain a comprehensive performance metric.

[0169] The gradient descent method is used to optimize the objective function to predict the optimal window size. During the optimization process, the learning rate is adaptively adjusted according to the current error. The initial learning rate is set to 0.01. When the error decreases, the learning rate remains unchanged; when the error increases, the learning rate is reduced to 0.8 times the original. This adaptive learning rate strategy can improve the convergence speed and stability.

[0170] After multiple rounds of iteration, the terminal device obtains the predicted optimal window size. In this example, the predicted optimal window size is 1.75MB. Compared with the window size of 1.2MB at the previous moment, the window size has increased by 0.55MB. Based on this change, the terminal device generates a data transmission buffering strategy: increase the current buffer size from 1.2MB to 1.75MB and increase the buffer depth to cope with the upcoming network handover.

[0171] Adjust the buffer according to the predicted window size and allocate corresponding buffer resources for services with different priorities. For high-priority video stream services, allocate 70% of the buffer size, that is, 1.225MB; for medium-priority audio stream services, allocate 20%, that is, 0.35MB; for low-priority background data synchronization services, allocate 10%, that is, 0.175MB.

[0172] Ensure that the data of high-priority services has been fully buffered to ensure a seamless playback experience during the handover process. At the same time, the terminal device dynamically adjusts the data sending rate according to the handover progress and gradually resumes to the normal transmission state after the handover is completed.

[0173] Data transmission buffering strategies mainly use static buffer sizes or simple linear prediction methods. For example, some systems use fixed-size buffers (such as 1MB), or linearly extrapolate future network conditions based on the current network state. These methods do not fully consider the temporal characteristics and non-linear change patterns of the network state, resulting in problems such as insufficient buffer size or over-allocation.

[0174] Existing implementation means usually only consider a few network parameters (such as bandwidth and delay), lacking a comprehensive analysis of multi-dimensional network characteristics. In addition, traditional methods are difficult to adapt to complex and changing network environments, especially performing poorly in scenarios where network quality changes rapidly.

[0175] The improvement starting point of this application is to introduce deep learning technologies, especially LSTM and attention mechanisms, to capture the temporal patterns and long-term dependencies of network state changes. By constructing a network state vector containing multi-dimensional features and combining the historical state sequence with attention weights, an accurate prediction of the optimal buffer size is achieved. In addition, this method also innovatively designs a comprehensive optimization objective function, taking into account transmission delay, resource utilization, and stability simultaneously, to achieve multi-objective balanced optimization.

[0176] In a typical network handover scenario, the data transmission interruption rate is reduced from 17.8% to 3.2%, the average transmission delay is reduced by 32.5%, and the bandwidth utilization rate is increased by 21.7%. Especially in an environment where network conditions change rapidly (such as a mobile scenario), the prediction accuracy of this method is 35.3% higher than that of traditional methods, and the buffer utilization efficiency is increased by 42.6%. These improvements significantly enhance the quality of experience of users during network handover, especially providing strong support for the seamless playback of real-time media streams such as video and audio.

[0177] In an optional implementation manner, a scoring value is calculated based on the data transmission buffering strategy, hierarchical division and resource allocation are performed according to the scoring value to obtain a migration priority; and the performance-optimal index structure is selected by combining the migration priority and the index efficiency evaluation value, and data migration and resource reconfiguration are realized, including:

[0178] A data scoring value is obtained by weighting various parameters in the data transmission buffering strategy and the data resources;

[0179] The data scoring value is compared with a hot data threshold and a cold data threshold. The data resources higher than the hot data threshold are divided into the hot data layer, the data resources lower than the cold data threshold are divided into the cold data layer, and the data resources between the hot data threshold and the cold data threshold are divided into the warm data layer;

[0180] For the data resources in each data layer, the migration priority is calculated based on the ratio of its access frequency to the storage space, the data value weight, and the normalized time decay exponent, and the normalized time decay exponent decreases exponentially with time growth;

[0181] Monitor the query operations of each data layer, calculate the first weighted sum of the query performance and the query weight, and calculate the second weighted sum of the resource consumption and the resource utilization rate. Divide the first weighted sum by the second weighted sum to obtain the index efficiency evaluation value;

[0182] Select the performance-optimal index structure according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, calculate the resource allocation quota according to the performance-optimal index structure after migration, and update the resource configuration of the storage level.

[0183] Calculate the data score value based on the data transmission buffer strategy and the weighted calculation of various parameters in the data resources. The parameters of the data resources include but are not limited to: access frequency, data size, data creation time, recent access time, business value, etc. The terminal device assigns corresponding weights to each parameter. For example, the weight of the access frequency is 0.35, the weight of the data size is 0.2, the weight of the data creation time is 0.05, the weight of the recent access time is 0.25, and the weight of the business value is 0.15.

[0184] Taking a certain data object as an example, the parameters of this object are: access frequency 125 times per day, data size 8 MB, data creation time 30 days ago, recent access time 0.5 days ago, and business value level 3 (the highest is level 5). The terminal device normalizes these original parameter values: the normalized value of the access frequency is 0.83 (relative to the highest access frequency in the system), the normalized value of the data size is 0.4 (the smaller the data, the higher the value), the normalized value of the data creation time is 0.25 (the newer the value, the higher), the normalized value of the recent access time is 0.9 (the closer the value, the higher), and the normalized value of the business value is 0.6.

[0185] Calculate the score value of this data object, that is, the sum of the products of the normalized values of each parameter and their weights: 0.83×0.35 + 0.4×0.2 + 0.25×0.05 + 0.9×0.25 + 0.6×0.15 = 0.6885. Similarly, the terminal device calculates the score values of all data objects.

[0186] Compare the data score value with the preset hot data threshold and cold data threshold to achieve the hierarchical division of data. Assume that the hot data threshold is 0.65 and the cold data threshold is 0.35. The terminal device divides the data resources with a score value higher than 0.65 into the hot data layer, divides the data resources with a score value lower than 0.35 into the cold data layer, and divides the data resources with a score value between 0.35 and 0.65 into the warm data layer. In this example, the score value of this data object is 0.6885, which is higher than the hot data threshold of 0.65, so it is divided into the hot data layer.

[0187] For the data resources divided into each data layer, the terminal device calculates the migration priority based on the ratio of its access frequency to the storage space, the data value weight, and the normalized time decay index. The ratio of the access frequency to the storage space reflects the access heat per unit storage space, the data value weight represents the degree of dependence of the business on this data, and the normalized time decay index takes into account the time characteristics of the data access pattern.

[0188] When calculating the time decay exponent, the terminal device uses an exponential decay function so that the time growth shows an exponential decrease. Assuming that the base decay rate is 0.9 and the time since the most recent access is t days, the time decay exponent is 0.9 to the power of t. For example, if the most recent access was 0.5 days ago, the time decay exponent is 0.9 to the power of 0.5, which is approximately 0.95.

[0189] Returning to the aforementioned data object, the ratio of its access frequency to the storage space is 125 times per day ÷ 8 MB = 15.625 times per day per MB, the data value weight is 0.6, and the normalized time decay exponent is 0.95. The terminal device multiplies these three values to obtain a migration priority of 15.625 × 0.6 × 0.95 = 8.91. The terminal device calculates the migration priorities of all data objects in the same way and sorts them from highest to lowest priority to determine the order of data migration.

[0190] Monitor the query operations of each data layer and calculate the index efficiency evaluation value. The monitoring content includes query performance metrics (such as response time, throughput) and resource consumption metrics (such as CPU usage rate, memory occupancy). The terminal device calculates the first weighted sum of the query performance and the query weight, and the second weighted sum of the resource consumption and the resource usage rate, and then divides the first weighted sum by the second weighted sum to obtain the index efficiency evaluation value.

[0191] For the hot data layer, the terminal device can test the performance of different index structures (such as B+ tree, hash index, bitmap index, etc.). Taking the B+ tree index as an example, the monitored metrics are: query response time of 15 ms, query throughput of 800 times per second, CPU usage rate of 25%, and memory occupancy of 120 MB. The terminal device normalizes the query response time to 0.75 (the faster the response, the higher the value), normalizes the query throughput to 0.8, assigns weights of 0.6 and 0.4 respectively, and calculates the first weighted sum:

[0192] 0.75 × 0.6 + 0.8 × 0.4 = 0.77.

[0193] Normalize the CPU usage rate to 0.75 (the lower the usage rate, the higher the value), normalize the memory occupancy to 0.6, assign weights of 0.55 and 0.45 respectively, and calculate the second weighted sum:

[0194] 0.75 × 0.55 + 0.6 × 0.45 = 0.6825.

[0195] The efficiency evaluation value of the B+ tree index is 0.77 ÷ 0.6825 = 1.128.

[0196] Calculate the efficiency evaluation values of other index structures in the same way. Assume that the efficiency evaluation value of the hash index is 1.056 and the efficiency evaluation value of the bitmap index is 0.985. Since the efficiency evaluation value of the B+ tree index is the highest, the terminal device selects the B+ tree as the optimal index structure for this data layer.

[0197] Perform data migration between different storage levels based on the calculated migration priorities. High-priority data is migrated to the high-performance storage layer (such as SSD), medium-priority data is migrated to the medium-performance storage layer (such as hard disk), and low-priority data is migrated to the low-performance storage layer (such as object storage).

[0198] For example, the migration priority of the aforementioned data object is 8.91, which belongs to a relatively high priority. The terminal device migrates it to the SSD storage layer. During the migration process, the terminal device adopts an asynchronous migration strategy to avoid blocking normal service access. When the data migration completion rate reaches 80%, the terminal device starts to update the metadata and redirects the data access request to the new location.

[0199] Calculate the resource allocation quota according to the selected optimal index structure (B+ tree). Based on the resource demand characteristics of this index structure, the terminal device allocates corresponding resources to different storage levels. For example, the B+ tree index has a high demand for memory. The terminal device accordingly increases the memory allocation of the hot data layer from the original 4GB to 6GB; correspondingly, reduces the memory allocations of the warm data layer and the cold data layer from 2GB and 1GB to 1.5GB and 0.5GB respectively.

[0200] In terms of storage space configuration, the terminal device allocates corresponding SSD and HDD spaces according to the data volumes and access characteristics of the hot data layer, the warm data layer, and the cold data layer. For example, the hot data layer is allocated 80% of the SSD space, the warm data layer is allocated 20% of the SSD space and 50% of the HDD space, and the cold data layer is allocated 50% of the HDD space.

[0201] This resource allocation strategy based on index efficiency and migration priority enables the system to dynamically adjust the resource configuration according to the actual load characteristics, achieving the goals of performance optimization and resource utilization balance.

[0202] In the prior art, data level division and index structure selection usually adopt static configuration or methods based on simple rules. For example, some systems only divide data into active data and inactive data according to the access time of the data, or use a unified index structure for all data, failing to fully consider the complexity and diversity of data access patterns.

[0203] The implementation means of traditional methods often rely on fixed thresholds or simple statistical metrics. For example, data accessed within the last 7 days is regarded as hot data, or the index structure is determined solely based on access frequency. These methods lack a comprehensive consideration of data value, business importance, and storage resource characteristics, resulting in unreasonable resource allocation and performance bottlenecks.

[0204] Figure 4 Schematic diagram of the complete process from data score calculation to final resource configuration update in the embodiments of the present invention:

[0205] This figure shows a complete flow chart of data hierarchical storage and migration optimization. The entire process starts with the weighted calculation of data transfer buffer strategies and data resource parameters, and a data score value is obtained through these parameters. Subsequently, the system compares the data score value with preset hot and cold thresholds, and accordingly divides the data into a hot data layer, a warm data layer, and a cold data layer. Among them, data with a score higher than the hot threshold is classified into the hot data layer, and data with a score lower than the cold threshold is divided into the cold data layer. Next, the system calculates the priority of data migration based on the ratio of access frequency to storage space, the data value weight, and a time decay exponent that exponentially decays over time. The system continuously monitors the query performance of each data layer, and finally obtains an index efficiency evaluation value by calculating the first weighted sum of query performance and query weight, and the second weighted sum of resource consumption and resource utilization rate. Based on this evaluation value, the system selects the index structure with the optimal performance and performs data migration between different storage levels according to the migration priority, preferentially migrating high-priority data to high-performance storage levels. Finally, the system updates and adjusts the resource configuration of the storage level according to the optimal index structure, thus completing the entire optimization process. This process reflects the adaptability of data management and the high efficiency of resource utilization.

[0206] The starting point of the improvement in this application is to construct a multi-dimensional data scoring model, introduce a time decay factor and business value weight, and achieve a more refined data level division. At the same time, by monitoring the performance metrics and resource consumption of actual query operations, calculating the index efficiency evaluation value, dynamically selecting the optimal index structure, and optimizing the resource allocation strategy.

[0207] After the improvement, the overall performance of the system has been significantly improved. The average query response time has been shortened by 43%, the storage resource utilization rate has been increased by 35%, and the system throughput has been increased by 28%. Especially in complex business scenarios, this method based on multi-dimensional scoring and dynamic index optimization can adapt to changes in different load characteristics, provide a more consistent performance experience, while reducing waste of storage and computing resources, and achieving double optimization of performance and efficiency.

[0208] In an alternative embodiment, select the index structure with the optimal performance according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, and calculate the resource allocation quota according to the index structure with the optimal performance after the migration is completed, and update the resource configuration of the storage level, including:

[0209] Obtain the load data for the index shards involved in the query operation, calculate the average value of the load data, divide the sum of the squared deviations of the load data from the average value by the number of shards and take the square root, and then take the negative value to calculate the load balance degree;

[0210] Multiply the index efficiency evaluation value and the load balance degree by the first balance factor and the second balance factor respectively, and subtract the product of the index maintenance overhead and the third balance factor to obtain the fitness of the current index structure;

[0211] Calculate the fitness for the newly constructed candidate index structure, subtract the fitness of the current index structure from the fitness of the candidate index structure, and then divide by the fitness of the current index structure to obtain the relative change rate of fitness;

[0212] When the relative change rate of fitness is greater than the preset update threshold, calculate the ratio of the historical access popularity of the data object to the data size, and use the product of the ratio, the business value weight, and the time decay exponent as the data migration priority;

[0213] Calculate the proportion of the data migration priority in the total sum of the migration priorities of all data objects, and allocate the product of the proportion, the total resource pool, and the level adjustment coefficient to the corresponding storage level for updating the resource configuration of the storage level.

[0214] The terminal device first obtains the load data for the index shards involved in the query operation. The load data includes indicators such as CPU usage rate, memory occupancy, and number of IO requests. For example, in a certain embodiment, the system monitors that the CPU usage rates of 5 index shards are: 35%, 42%, 28%, 55%, 40% respectively. Calculate the average value of these values to get 40%. Then calculate the sum of the squared deviations of each shard's load data from the average value, that is:

[0215] (35 - 40)²+(42 - 40)²+(28 - 40)²+(55 - 40)²+(40 - 40)² = 522.

[0216] Divide this value by the number of shards 5 to get 104.4, then take the square root to get 10.2, and finally take the negative value to get -10.2. For the convenience of subsequent calculations, normalize this value to the range of 0 to 1 to obtain the load balance degree of 0.68.

[0217] After the load balancing degree is calculated, the terminal device multiplies the index efficiency evaluation value and the load balancing degree by their corresponding balance factors respectively, and subtracts the product of the index maintenance overhead and the third balance factor to obtain the fitness of the current index structure. Specifically, assume that the index efficiency evaluation value is 0.75, the first balance factor is 0.5, the load balancing degree is 0.68, the second balance factor is 0.3, the index maintenance overhead is 0.2, and the third balance factor is 0.2. Substituting these values into the calculation, the fitness of the current index structure is:

[0218] 0.75×0.5 + 0.68×0.3 - 0.2×0.2 = 0.375 + 0.204 - 0.04 = 0.539.

[0219] For the newly constructed candidate index structure, the terminal device also calculates its fitness. Assume that the fitness of the candidate index structure is 0.65. After subtracting the fitness of the current index structure from the fitness of the candidate index structure, and then dividing by the fitness of the current index structure, the relative change rate of fitness is obtained, that is:

[0220] (0.65 - 0.539) / 0.539 = 0.111×100% = 11.1%.

[0221] When the relative change rate of fitness is greater than the preset update threshold, the terminal device starts the data migration process. Assume that the preset update threshold is 10%. At this time, 11.1% is greater than 10%, meeting the condition. The terminal device calculates the ratio of the historical access popularity of the data object to the data size, and takes the product of this ratio and the business value weight and the time decay exponent as the data migration priority.

[0222] The terminal device records that a certain data object has been accessed 450 times in the past 30 days, and the data size is 5MB. Then the ratio of the historical access popularity to the data size is 450 / 5 = 90. Assume that the business value weight of this data object is 0.8 and the time decay exponent is 0.95. Then the data migration priority is 90×0.8×0.95 = 68.4.

[0223] The corresponding migration priorities are calculated for all data objects in the system. Assume that the sum is 1000. Then the proportion of the migration priority of this data object is 68.4 / 1000 = 0.0684, that is, 6.84%. If the total resource pool size is 8GB and the level adjustment coefficient of this storage level is 1.2, then the resources allocated to the corresponding storage level of this data object are 8GB×0.0684×1.2 = 0.65664GB, approximately 672MB. The terminal device updates the resource configuration of the storage level accordingly.

[0224] The terminal device also considers the minimum granularity of resource allocation. If the resources allocated to a certain storage level are less than the minimum allocation unit, adjustments will be made to ensure the rationality of resource allocation. For example, if the minimum allocation unit is 128 MB and the calculated resource allocation is 90 MB, the allocation will be adjusted to 128 MB.

[0225] To ensure the dynamic adaptability of resource allocation, the terminal device periodically re-evaluates the index structure and data distribution, and adjusts the resource allocation strategy. For example, a comprehensive evaluation is triggered every 24 hours, or an evaluation is immediately triggered when the system load changes by more than 20%. This dynamic adjustment mechanism enables the system to adapt to different business scenarios and load changes.

[0226] During the data migration execution phase, the terminal device determines the order of data migration according to the calculated data migration priorities. Data objects with higher priorities will be migrated to higher-performance storage levels first. For example, a data object with a priority of 68.4 will be preferentially migrated to the SSD storage layer, while data objects with lower priorities may be migrated to the HDD storage layer.

[0227] During the migration process, the terminal device monitors the migration progress and system performance. If it is found that the migration process has a serious impact on the system performance, the migration rate will be dynamically adjusted. For example, when it is found that the system CPU usage exceeds 85%, the migration rate will be reduced by 50% to ensure that the normal operation of the business is not affected.

[0228] After the migration is completed, the terminal device rebuilds the index structure and updates the resource configuration of the storage level. The updated resource configuration will take effect immediately, and the system performance will be optimized.

[0229] Figure 5 Schematic diagram of the load balancing degree performance of different methods in the embodiment of the present invention under the condition of increasing query load:

[0230] The figure shows the comparison results of three different methods in performance evaluation. The horizontal axis in the figure represents the number of query requests per second, ranging from 0 to 500; the vertical axis represents the system's load balancing degree, with values ranging from 0 to 1. The three methods are the present technical solution (the solid line marked with triangles), the static threshold method (the dashed line marked with circles), and the simple statistical method (the dotted line marked with squares). It can be seen from the performance curve that the performance of the present technical solution is significantly better than the other two methods. Specifically, when the number of query requests reaches 200, the load balancing degree of the present technical solution reaches approximately 0.82, while the static threshold method and the simple statistical method only reach approximately 0.55 and 0.45 respectively. As the number of requests further increases to 500, the load balancing degree of the present technical solution stabilizes at around 0.92, demonstrating excellent scalability and stability; in contrast, the performance growth of the static threshold method and the simple statistical method tends to level off, and finally stabilizes at approximately 0.6 and 0.5 respectively. This indicates that the present technical solution has obvious performance advantages in handling a large number of concurrent requests and can better maintain the system's load balance.

[0231] Compared with the prior art, traditional index structure selection and resource allocation methods are mainly based on static rules or simple historical statistics, and cannot accurately reflect the real-time load situation, resulting in low resource utilization. For example, traditional methods may only consider the data access frequency and ignore factors such as data size and business value, and cannot achieve refined resource management.

[0232] The implementation means of the prior art usually adopt a fixed threshold to trigger resource reallocation and lack adaptability. For example, some systems stipulate that data migration is triggered when the storage space utilization rate exceeds 80%. This simple trigger mechanism cannot cope with complex and changing business scenarios, resulting in resource waste or performance bottlenecks.

[0233] The starting point for the improvement of this application is to improve the accuracy and adaptability of index structure selection and resource allocation. By introducing multi-dimensional indicators such as load balancing degree and index efficiency evaluation value, combined with business value weights and time decay exponents, refined calculation of data migration priorities is realized. At the same time, through a dynamic evaluation mechanism of the relative change rate of fitness, it is ensured that the system only performs index structure updates and resource reallocations when it is really necessary, avoiding unnecessary system overhead.

[0234] After the improvement, the system resource utilization rate has increased by approximately 25%, the data query response time has been reduced by 30%, and the overall stability of the system has been significantly enhanced. Especially in high-concurrency scenarios, the improved method can maintain a low response latency, while the traditional method may experience a sharp decline in performance.

[0235] In the second aspect of the embodiments of the present invention,

[0236] Provided is a heterogeneous network fusion and intelligent switching control system for a three-in-one terminal device, including:

[0237] A first unit, configured to obtain biometric data of a user in the three-in-one terminal device, and input the biometric data into corresponding feature extraction models respectively to generate multi-dimensional feature vectors;

[0238] A second unit, configured to calculate the similarity between the multi-dimensional feature vectors and a preset user biometric template library to obtain a credibility score for user identity recognition. When the credibility score is higher than a preset security threshold, trigger a network switching evaluation process;

[0239] A third unit, configured to analyze the historical network switching records of the user in the network switching evaluation process, perform correlation modeling on the time regularity features, location information features, and service type features of the user to obtain a user network switching probability prediction model;

[0240] A fourth unit, configured to monitor the performance metric data of the currently accessed network in real time based on the output result of the user network switching probability prediction model;

[0241] A fifth unit, configured to perform weighted fusion calculation on the performance metric data and the output result of the user network switching probability prediction model to generate network switching decision parameters, determine an optimal switching time window according to the network switching decision parameters, and formulate a corresponding data transmission buffering strategy;

[0242] A sixth unit, configured to calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; and select the index structure with the optimal performance in combination with the migration priority and the index efficiency evaluation value to implement data migration and resource reconfiguration.

[0243] In a third aspect of the embodiments of the present invention,

[0244] Provided is an electronic device, including:

[0245] A processor;

[0246] A memory for storing instructions executable by the processor;

[0247] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0248] In a fourth aspect of the embodiments of the present invention,

[0249] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0250] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0251] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Tripartite terminal device heterogeneous network fusion and intelligent switching control method, characterized in that Including: Obtain the biometric data of the user in the three-in-one terminal device, and input the biometric data into the corresponding feature extraction model respectively to generate multi-dimensional feature vectors; Calculate the similarity between the multi-dimensional feature vectors and the pre-set user biometric template library to obtain the credibility score of user identity recognition. When the credibility score is higher than the pre-set security threshold, trigger the network switching evaluation process; In the network switching evaluation process, analyze the user's historical network switching records, and perform correlation modeling on the user's time pattern features, location information features, and service type features to obtain a user network switching probability prediction model; Based on the output result of the user network switching probability prediction model, monitor the performance index data of the currently connected network in real time; Perform weighted fusion calculation on the performance index data and the output result of the user network switching probability prediction model to generate network switching decision parameters, determine the optimal switching time window according to the network switching decision parameters, and formulate corresponding data transmission buffering strategies; Calculate the score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain the migration priority; and select the index structure with the optimal performance in combination with the migration priority and the index efficiency evaluation value to realize data migration and resource reconfiguration.

2. The method according to claim 1, wherein Calculating the similarity between the multi-dimensional feature vectors and the pre-set user biometric template library to obtain the credibility score of user identity recognition. When the credibility score is higher than the pre-set security threshold, triggering the network switching evaluation process includes: Calculate the similarity between the multi-dimensional feature vectors and the pre-set biometric templates, and perform decision-level fusion on the recognition results of different features using the D-S evidence theory to generate a fused credibility score. The decision-level fusion solves the feature contradiction problem by calculating the conflict factor between features; Compare the fused credibility score with the current security threshold. When the fused credibility score is greater than the current security threshold, confirm that the user identity is valid and output the identity recognition result to re-trigger the network switching evaluation process.

3. The method according to claim 2, characterized in that, Performing decision-level fusion on the recognition results of different features using the D-S evidence theory to generate a fused credibility score. The decision-level fusion solves the feature contradiction problem by calculating the conflict factor between features includes: Construct a feature fusion model based on the D-S evidence theory, substitute the feature fusion model into the orthogonal sum operation formula to obtain the feature combination result, and the feature combination result reflects the support degree distribution of different features; Calculate the feature conflict part in the feature combination result, and quantify the feature conflict degree based on the Euclidean distance function. The Euclidean distance function calculates the sum of the squares of the differences between the basic probability assignment values of different features to obtain the feature difference degree; Introduce the feature difference degree into the conflict factor calculation formula of the D-S evidence theory, and correct the original conflict factor using an exponential decay function to obtain the corrected conflict factor, and the corrected conflict factor can accurately reflect the contradiction degree between features; Normalize the feature combination result based on the corrected conflict factor, construct a D-S fusion rule, and perform evidence synthesis on the normalized feature combination result and the corrected conflict factor to obtain a fusion credibility score.

4. The method according to claim 1, wherein Perform weighted fusion calculation on the performance index data and the output result of the user network handover probability prediction model to generate a network handover decision parameter. Determine the optimal handover time window according to the network handover decision parameter, and formulate a corresponding data transmission buffer strategy, including: Perform weighted fusion calculation on the performance index data and the output result of the user network handover probability prediction model, where the weighted fusion calculation dynamically adjusts the importance of the performance index data and the output result using an adaptive weight coefficient to generate a network handover decision parameter; Set a network state evaluation window according to the network handover decision parameter, continuously judge the handover conditions within the network state evaluation window, and determine the optimal handover time window when the network handover decision parameter continuously meets the preset handover threshold; Formulate a data transmission buffer strategy according to the time range of the optimal handover time window and the data transmission state of the current network. The data transmission buffer strategy realizes smooth data transmission during network handover by adjusting the buffer size of data transmission.

5. The method according to claim 4, wherein Formulate a data transmission buffer strategy according to the time range of the optimal handover time window and the data transmission state of the current network, including: Construct a network state feature vector, and form a historical state sequence by combining the network state feature vector with the historical state feature vectors within the historical observation window; Input the historical state sequence into a long short-term memory network, and perform temporal feature extraction on the historical state sequence through the forget gate, input gate, and output gate of the long short-term memory network to obtain a hidden state sequence; Construct an attention mechanism based on the hidden state sequence, calculate an energy score through the attention mechanism, normalize the energy score to obtain an attention weight, and generate a context vector according to the weighted sum of the attention weight and the historical state sequence; Establish a comprehensive optimization objective function based on the context vector. The comprehensive optimization objective function includes average transmission delay, resource utilization rate, and throughput variance, and adaptively adjusts the learning rate according to the current error; Obtain the predicted optimal window size based on the historical state sequence, the context vector, and the comprehensive optimization objective function, and generate a data transmission buffer strategy based on the difference between the optimal window size and the window size at the previous moment.

6. The method according to claim 1, wherein Calculate a score value based on the data transmission buffer strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; And select the index structure with the optimal performance by combining the migration priority and the index efficiency evaluation value to realize data migration and resource reconfiguration, including: Obtain a data score value by weighting each parameter in the data transmission buffer strategy and the data resources; Compare the data scoring value with the hot data threshold and the cold data threshold, classify the data resources higher than the hot data threshold into the hot data layer, classify the data resources lower than the cold data threshold into the cold data layer, and classify the data resources between the hot data threshold and the cold data threshold into the warm data layer; For the data resources in each data layer, calculate the migration priority based on the ratio of its access frequency to the storage space, the data value weight, and the normalized time decay exponent, and the normalized time decay exponent decreases exponentially with the increase of time; Monitor the query operations of each data layer, calculate the first weighted sum of the query performance and the query weight, and calculate the second weighted sum of the resource consumption and the resource utilization rate, and divide the first weighted sum by the second weighted sum to obtain the index efficiency evaluation value; Select the index structure with the optimal performance according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, calculate the resource allocation quota according to the index structure with the optimal performance after the migration is completed, and update the resource configuration of the storage level.

7. The method according to claim 6, wherein Select the index structure with the optimal performance according to the index efficiency evaluation value, perform data migration between different storage levels based on the migration priority, calculate the resource allocation quota according to the index structure with the optimal performance after the migration is completed, and update the resource configuration of the storage level including: Obtain the load data of the index shards involved in the query operation, calculate the average value of the load data, and calculate the load balance degree by taking the negative value after dividing the sum of the squared deviations of the load data from the average value by the number of shards and taking the square root; Multiply the index efficiency evaluation value and the load balance degree by the first balance factor and the second balance factor respectively, and subtract the product of the index maintenance overhead and the third balance factor to obtain the fitness of the current index structure; Calculate the fitness of the newly constructed candidate index structure, subtract the fitness of the current index structure from the fitness of the candidate index structure, and then divide by the fitness of the current index structure to obtain the relative change rate of the fitness; When the relative change rate of the fitness is greater than the preset update threshold, calculate the ratio of the historical access heat of the data object to the data size, and use the product of the ratio and the business value weight and the time decay exponent as the data migration priority; Calculate the proportion of the data migration priority in the total sum of the migration priorities of all data objects, and allocate the product of the proportion and the total resource pool and the layer adjustment coefficient to the corresponding storage level for updating the resource configuration of the storage level.

8. A three-in-one terminal device heterogeneous network fusion and intelligent switching control system, used to implement the method described in any one of claims 1 to 7, characterized in that: Including: The first unit is used to obtain the biometric data of the user in the three-in-one terminal device, and input the biometric data into the corresponding feature extraction model to generate a multi-dimensional feature vector; The second unit is used to calculate the similarity between the multi-dimensional feature vector and the preset user biometric template library to obtain the credibility score of user identity recognition. When the credibility score is higher than the preset security threshold, trigger the network switching evaluation process; A third unit, configured to analyze the historical network handover records of a user in the network handover evaluation process, and perform correlation modeling on the time regularity feature, location information feature, and service type feature of the user to obtain a user network handover probability prediction model; A fourth unit, configured to monitor the performance index data of the currently accessed network in real time based on the output result of the user network handover probability prediction model; A fifth unit, configured to perform weighted fusion calculation on the performance index data and the output result of the user network handover probability prediction model to generate network handover decision parameters, determine an optimal handover time window according to the network handover decision parameters, and formulate a corresponding data transmission buffering strategy; A sixth unit, configured to calculate a score value based on the data transmission buffering strategy, perform hierarchical division and resource allocation according to the score value to obtain a migration priority; And select the index structure with the optimal performance in combination with the migration priority and the index efficiency evaluation value to implement data migration and resource reconfiguration.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, The computer program instructions, when executed by the processor, implement the method according to any one of claims 1 to 7.

Images