Intelligent interface switching method based on multi-dimensional perception

Abstract

The invention discloses an intelligent interface switching method based on multi-dimensional perception, and belongs to the field of electric digital data processing, and the method comprises the steps: obtaining equipment data, network data and user data to obtain standard vector information; obtaining high-precision position information based on the standard vector information, and obtaining at least one candidate interface based on the high-precision position information, the current interface metadata and a preset position formula; obtaining at least one dynamic interface based on a time sequence prediction algorithm, the historical track information, the current interface metadata and the candidate interface; and obtaining interface dynamic data in real time based on the main interface and the plurality of standby interfaces, obtaining a dynamic evaluation score based on the interface dynamic data, and executing an intelligent switching action according to the dynamic evaluation score. Therefore, by implementing the method and the device, the problem that a uniform interface cannot meet the performance and compliance requirements of super-large-scale and cross-region questions and answers in the prior art can be solved, and the switching of the main interface and the standby interface or the intelligent switching of different protocols can be realized.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to an intelligent interface switching method based on multi-dimensional perception. Background Technology

[0002] AI-powered question-answering systems deployed across borders or multiple regions have become widely used globally, serving as a key infrastructure for enterprises' globalization strategies. These systems support real-time multilingual interaction and intelligent intent recognition, covering dozens of mainstream and less commonly spoken languages, significantly improving customer service efficiency and experience. They have been successfully implemented in various industries, including aviation, retail, healthcare, and finance, optimizing travel bookings and enhancing customer experience, while also contributing to the intelligentization of government services.

[0003] However, in AI question-answering systems deployed across borders or multiple regions, user network environments and regulatory requirements vary significantly across different regions. Traditional single-interface or fixed-node models struggle to balance low latency and compliance. Particularly globally, IP location often has errors ranging from hundreds to thousands of kilometers. Combined with differences in network conditions (bandwidth, packet loss rate, RTT), time zones, and language preferences across regions, a unified interface cannot meet the performance and compliance requirements of ultra-large-scale, cross-regional question-answering. Therefore, an intelligent interface switching engine is needed that comprehensively utilizes multi-dimensional perception (geographical location, network quality, regulatory compliance, etc.) to achieve primary / backup node switching or different protocol switching within milliseconds. Summary of the Invention

[0004] This invention provides an intelligent interface switching method based on multi-dimensional perception, which can solve the problem that the unified interface in the prior art cannot meet the performance and compliance requirements of ultra-large-scale and cross-regional question and answer, and realize the switching of primary and backup interfaces or different protocols in milliseconds.

[0005] This invention provides a method for intelligent interface switching based on multi-dimensional perception, comprising:

[0006] Acquire device data, network data, and user data, and obtain standard vector information based on preset structured processing methods, preset standardized processing methods, and device data, network data, and user data;

[0007] Obtain the current interface metadata, obtain high-precision location information based on standard vector information, current interface metadata and Kalman filter algorithm, and then obtain at least one candidate interface that meets the preset compliance conditions based on high-precision location information, current interface metadata and preset location formula.

[0008] Obtain historical trajectory information, and based on the time series prediction algorithm, historical trajectory information, current interface metadata and candidate interfaces, obtain at least one dynamic interface that meets the preset information conditions;

[0009] The main interface and several backup interfaces are preloaded based on dynamic interfaces;

[0010] Real-time dynamic data of the interface is obtained based on the main interface and several backup interfaces, and dynamic evaluation scores are obtained based on the dynamic data of the interface.

[0011] When the dynamic evaluation score exceeds the preset threshold range, an intelligent switching action is executed: the target interface is obtained from the main interface and several backup interfaces based on the dynamic data of the interface, and the current interface is switched to the target interface.

[0012] This invention provides an intelligent interface switching method based on multi-dimensional perception. It acquires and fuses device data, network data, and user data to obtain multi-source correlated data, combining the real-time performance of the interface network with the connection between the device context. The data is structured and standardized to improve the dimensionality and accuracy of environmental perception and the reliability of the data. Simultaneously, it achieves dual-channel parallel operation of primary and backup interfaces by filtering dynamic interfaces, improving interface availability. Furthermore, it monitors interface dynamics in real time for intelligent interface switching, realizing millisecond-level intelligent switching between primary and backup interfaces using multi-dimensional perception.

[0013] Furthermore, the process involves acquiring device data, network data, and user data, and obtaining standard vector information based on preset structured processing methods, preset standardized processing methods, and the device data, network data, and user data. This includes: acquiring device data, network data, and user data; obtaining data verification results based on device data, network data, user data, and preset verification methods; when the reliability verification results are determined to be normal, performing time-series marking and association binding based on device data, network data, and user data to obtain data packets; obtaining structured raw information based on data packets and preset structured processing methods; and obtaining standard vector information based on structured raw information, preset standardized processing methods, and structured raw information.

[0014] In the above scheme, by performing data reliability verification, time-series marking, and contextual association binding on multi-source data, the three types of data—device status, network quality, and user preferences—are multi-source associated and fused, and feature extraction is achieved through structured and standardized processing, thereby improving the dimensionality and accuracy of environmental perception.

[0015] Furthermore, the data verification results obtained based on device data, network data, user data, and preset verification methods also include: when it is determined that the data verification results are abnormal, executing a compensation strategy to obtain compensation data; performing time-series marking and association binding based on the compensation data to obtain a compensation data package; obtaining the compensation structured original information based on the compensation data package and a preset structured processing method; obtaining compensation vector information based on the compensation structured original information, a preset standardization processing method, and the structured original information, and using the compensation vector information as standard vector information.

[0016] In the above scheme, a compensation strategy is triggered when the data verification contact is determined to be abnormal, so as to compensate for the abnormal situation in the data, improve the accuracy of environmental perception data, and provide a reliable data foundation for subsequent intelligent switching of interfaces.

[0017] Furthermore, the system acquires current interface metadata, obtains high-precision location information based on standard vector information, current interface metadata, and a Kalman filter algorithm, and then acquires at least one candidate interface that meets preset compliance conditions based on the high-precision location information, current interface metadata, and a preset location formula. This includes: acquiring GPS data and IP data based on standard vector information; acquiring interface metadata, dynamically fusing GPS data and IP data based on interface metadata and a Kalman filter algorithm to obtain high-precision location information; acquiring geographic data based on high-precision location information and a preset location formula; acquiring compliance data based on geographic data and a preset multidimensional cost function; and acquiring at least one candidate interface that meets preset compliance conditions based on the compliance data.

[0018] In the above scheme, geospatial analysis based on standard vector information is used for dynamic fusion, and the Kalman filter algorithm is used for location information fusion to obtain high-precision location information. Interface nodes are calculated, sorted and selected by preset location formula and preset multidimensional cost function to obtain at least one candidate interface with preset compliance conditions, thereby realizing the topology calculation of dynamic nodes.

[0019] Furthermore, when it is determined based on the GPS data, IP data, and Kalman filter algorithm that the Kalman filter residual exceeds the preset residual range, historical location information is obtained, and single positioning data is obtained based on the historical location information, GPS data, IP data, and preset single positioning standard, and the single positioning data is used as the high-precision location information;

[0020] When it is determined that a single location data cannot be obtained based on the historical location information, GPS data, IP data, and a preset single location standard, adjacent successful location information is obtained and used as the high-precision location information.

[0021] In the above scheme, when the fusion residual is too large, that is, the error is unacceptable, the most reliable positioning data source is selected by the preset judgment criteria, and the system falls back to a single positioning data. When none of the above are available, the adjacent successful positioning information is obtained, that is, the previous successful positioning information is used as a temporary substitute. This ensures the accuracy of the positioning data, reduces the error of the positioning data, improves compliance, and enhances robustness, especially in the case of GPS drift or unavailability.

[0022] Furthermore, dynamic data from the main interface and dynamic interfaces is acquired in real time to obtain dynamic evaluation scores, including:

[0023] Real-time acquisition of dynamic data from the main interface and dynamic interfaces;

[0024] A preset normalized weighted algorithm and dynamic interface data are used to obtain dynamic evaluation scores. The dynamic interface data includes latency score, risk score, availability score, freshness score, and energy consumption score. The calculation process satisfies the following formula:

[0025] Q = w1 * S L +w2*S C +w3*S A +w4*S F +w5*S E

[0026] S L =exp(-L / 150)

[0027] S C =1-C

[0028] S A =A

[0029] S F =exp(-F / 10)

[0030] S E =exp(-E / 1)

[0031] In the formula: Q represents the dynamic evaluation score, S L Represents the delay score, where L represents the effective delay, and S represents the delay score. C C represents the risk score, and S represents the compliance index. A The interface availability score is represented by A, where A represents interface availability and S represents interface availability. F The score represents the freshness of the data, where F represents the freshness of the data and S represents the freshness of the data. E The values ​​represent energy consumption score, E represents energy efficiency, w1 represents the weight index corresponding to the latency score, w2 represents the weight index corresponding to the risk score, w3 represents the weight index corresponding to the availability score, w4 represents the weight index corresponding to the freshness score, and w5 represents the weight index corresponding to the energy consumption score.

[0032] Furthermore, the system acquires dynamic data from the main interface and several backup interfaces in real time to obtain dynamic evaluation scores. It also includes the following: when the dynamic evaluation score is determined to be within a preset threshold range, there is no need to perform intelligent switching actions.

[0033] In the above scheme, dynamic evaluation scores are calculated by real-time monitoring of the interface dynamic data and normalization and weighting methods. Based on preset judgment conditions, guidance strategies are provided for subsequent interface scheduling and switching, thereby realizing intelligent switching of interface nodes.

[0034] Furthermore, the current interface policy network is trained based on the preset policy gradient algorithm, and convergence data is obtained based on the current interface policy network. When the convergence data is determined not to meet the preset convergence conditions, the original learning parameters are obtained, and the target learning parameters are obtained based on the original learning parameters and the preset adjustment method. The current interface is then configured based on the target learning parameters.

[0035] Furthermore, the current interface policy network is trained based on a preset policy gradient algorithm, and convergence data is obtained based on the current interface policy network, including: obtaining the total number of training epochs, window data, and average reward based on the current interface policy network; the convergence data includes the average reward growth rate; the average reward growth rate is obtained based on the total number of training epochs, window data, and average reward, and its calculation process satisfies the following formula:

[0036]

[0037] In the formula: ARGR represents the average reward growth rate. This represents the average reward from round T to round Tw. This represents the average reward from round Tw to round T-2w.

[0038] When the average reward growth rate meets the preset growth rate condition and the preset duration, the convergence data is determined to not meet the preset convergence condition.

[0039] In the above scheme, the current interface policy network is continuously trained by a preset policy gradient algorithm, and the convergence data of the relevant interface is detected and obtained in real time. When the preset convergence conditions are not met, the target learning parameters are obtained so as to configure the current interface through the target learning parameters. This avoids the policy network's policy reward or reward function from stagnating or changing slowly over multiple training cycles, which would result in the current learning rate, batch size and other hyperparameter settings being insufficient to drive effective learning. This achieves the optimization of the reinforcement learning parameters of the current interface.

[0040] Furthermore, real-time monitoring information is acquired to obtain latency data based on the monitoring information. When it is determined that the latency data exceeds the preset maximum increase, a three-level rollback operation is executed.

[0041] In the above solution, a disaster recovery and auditing mechanism is introduced through an intelligent three-level rollback strategy. Real-time monitoring information is obtained to detect network latency data of the interface in real time. Then, when the latency data exceeds the preset maximum increase, the three-level rollback operation is executed in a timely manner to realize intelligent interface switching and improve the response efficiency of the question and answer system.

[0042] This invention provides an intelligent interface switching method based on multi-dimensional perception. It fuses device data, network data, and user data to obtain multi-source correlated data, combining the real-time performance of the interface network with the connection between the device context. The method introduces data fusion, reliability verification, compensation mechanisms, time-series tags, and context binding. Data processing is performed according to the data detection results, and the data is structured, standardized, and subjected to geospatial analysis, significantly improving the dimensionality and accuracy of environmental perception and data reliability. Through dynamic node topology calculation, Kalman filter positioning dynamic fusion, and compliance judgment to filter dynamic interfaces and perform pre-loading actions, combined with time-series prediction and confidence judgment, it achieves dual-channel parallel operation and intelligent switching scheduling of primary and backup interfaces, improving interface availability, reducing positioning errors, and enhancing robustness. Simultaneously, real-time quality monitoring and adaptive adjustment are performed during this process, with real-time detection of interface dynamics for intelligent interface switching, ensuring interface availability, data freshness, and regulatory compliance, reducing latency, and improving energy efficiency. Hyperparameter adjustment, disaster recovery, and auditing mechanisms are introduced to ensure stable operation during the switching process, achieving millisecond-level intelligent interface switching between primary and backup interfaces based on multi-dimensional perception. Attached Figure Description

[0043] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of an intelligent interface switching method based on multi-dimensional perception provided in this embodiment;

[0045] Figure 2 This is a schematic diagram of the intelligent interface switching process of an AI question-answering system based on multi-dimensional perception provided in this embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0048] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0049] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0050] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0051] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0052] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0053] Example 1:

[0054] See Figure 1This embodiment provides a method for intelligent interface switching based on multi-dimensional perception, including:

[0055] S1. Acquire device data, network data, and user data to obtain standard vector information based on preset structured processing methods, preset standardized processing methods, and device data, network data, and user data;

[0056] S2. Obtain the current interface metadata, obtain high-precision position information based on standard vector information, current interface metadata and Kalman filter algorithm, and then obtain at least one candidate interface that meets the preset compliance conditions based on high-precision position information, current interface metadata and preset position formula.

[0057] S3. Obtain historical trajectory information. Based on the time series prediction algorithm, historical trajectory information, current interface metadata, and candidate interfaces, obtain several dynamic interfaces that meet the preset information conditions.

[0058] S4. Preload the main interface and several backup interfaces based on several dynamic interfaces;

[0059] S5. Based on the main interface and several backup interfaces, obtain dynamic interface data in real time, and obtain dynamic evaluation scores based on the dynamic interface data;

[0060] S6. When the dynamic evaluation score exceeds the preset threshold range, execute the intelligent switching action: obtain the target interface from the main interface and several backup interfaces based on the dynamic data of the interface, and switch the current interface to the target interface.

[0061] This embodiment provides an intelligent interface switching method based on multi-dimensional perception. It acquires and fuses device data, network data, and user data to obtain multi-source correlated data, combining the real-time performance of the interface network with the connection between the device context. The data is structured and standardized to improve the dimensionality and accuracy of environmental perception and the reliability of the data. Simultaneously, by selecting at least one dynamic interface, it achieves dual-channel parallel operation of the primary and backup interfaces, improving interface availability. Furthermore, it monitors interface dynamics in real time for intelligent interface switching. Applied to an AI question-answering system, it enables millisecond-level intelligent switching between primary and backup interfaces using multi-dimensional perception.

[0062] In practical implementation, User Device Information (UserDeviceInfo) typically includes GPS coordinates, MAC address, operating system version, screen resolution, time zone, language preference, network type, and DNS resolution results. Network data typically includes data obtained from network layer resolution of client IP addresses, capture of RTT (Round-Trip Time) and packet loss rate, and extraction of TCP window expansion factor. In specific applications, GPS coordinates from the device are obtained periodically via the device API with precision fields; the MAC address is obtained from the device via system calls to obtain network card hardware information; the operating system version and resolution are obtained from the device via the device system information interface; the time zone and language preference are read from the device via system configuration; the network type and DNS resolution results are obtained by monitoring network interface status and recording DNS request responses; the client IP address from the network layer is obtained when the edge server receives requests; RTT data from the network layer is obtained by measuring during the TCP handshake and QUIC connection initialization phases; the packet loss rate from the network layer is estimated by counting TCP retransmissions; and the TCP window expansion factor from the network layer is obtained by reading the TCP header field (Window Scale Option), etc. Unlike traditional methods that typically only collect IP address or GPS location data and ignore real-time network performance and device context, this solution adopts an end-to-network collaborative collection mechanism, which for the first time integrates three types of data: "device status, network quality, and user preferences," greatly improving the dimensionality and accuracy of environmental perception.

[0063] Optionally, acquire device data, network data, and user data, and obtain standard vector information based on preset structured processing methods, preset standardized processing methods, and device data, network data, and user data. This includes: acquiring device data, network data, and user data; obtaining data verification results based on device data, network data, user data, and preset verification methods; when it is determined that there are no anomalies in the reliability verification results, performing time-series marking and association binding based on device data, network data, and user data to obtain data packets; obtaining structured raw information based on data packets and preset structured processing methods; and obtaining standard vector information based on structured raw information, preset standardized processing methods, and structured raw information.

[0064] In the specific implementation process, the reliability of multi-source data is first verified. This includes verifying the accuracy of GPS data (marking it as "low confidence" if the accuracy field is >100 meters), determining the validity of MAC addresses (excluding all zeros and virtual MAC addresses), and verifying whether DNS results exhibit hijacking behavior (such as abnormal CNAME records). Based on the reliability verification results, risk types are marked, typically categorized as follows: Low risk: repeated requests within a short period with normal behavior patterns, requiring only rapid processing; Medium risk: abnormal CNAME records returned by DNS, a large number of failed records, triggering degradation logic (such as redirecting to a backup interface / node); High risk: the presence of known attacking IPs, frequent changes in DNS policies, triggering isolation measures (such as dropping requests or adding to a blacklist), thus achieving risk classification and processing. Simultaneously, based on device fingerprints and timestamps, data repeatedly uploaded by the same device within a short period is deduplicated and filtered for security. Combined with reliability verification, suspicious IPs or abnormal DNS requests are marked to prevent probe attacks using the interface.

[0065] In the specific implementation process, data that has passed reliability verification is time-series labeled and context-associated, each data item is given a precise timestamp (UTC format can be used), and the data collection source (e.g., local / network) is recorded. DNS queries, RTT information, and IP information are combined into a data context packet to improve event reconstructability. After the data has been cleaned, verified, and bound, it is output as "structured raw information" according to a preset structured processing method (e.g., JSON structure). This structured data will be passed as input to the environmental feature standardization module to generate input vectors for subsequent models such as geographic matching, network QoS scoring, and user profiling, i.e., standard vector information.

[0066] In the specific implementation process, the process of generating standard vector information is as follows: First, the structured raw information from the previous step is processed: device fingerprint generation is performed, such as through data encryption using a secure hash algorithm, which differs from the plaintext MAC or IP method used in existing technologies, enhancing privacy compliance and facilitating blockchain recording; then, network quality is scored using a network quality scoring model, and those with lower scores are marked. Unlike the simple linear weighting of existing technologies, an exponential function fitting and packet loss ratio weighting method are used to construct the network quality scoring model, such as: score = 0.7 * exp(-rtt / 150) + 0.3 * (1 - packet_loss_ratio), where exp(-rtt / 150) indicates that as the RTT value increases, the score will increase exponentially. Packet loss is assessed by packet decay, where packet_loss_ratio represents the amount of lost packets. A score < 0.3 indicates a low rating and is labeled as poor network performance, improving sensitivity to low latency and enhancing user experience scheduling accuracy. Finally, feature vectors are synthesized using a pre-defined standardization process to obtain standardized feature vectors (vectors or JSON). First, vector synthesis is performed through multimodal heterogeneous data fusion, unlike existing technologies that rely on single dimensions such as network QoS or IP location, to improve the personalized accuracy of interface switching. Then, dynamic normalization, label mapping, and embedding encoding are used for vector standardization, unlike existing technologies that use rule-based judgment or hard coding. This method adapts to the input of machine learning models, facilitating model training and optimization for subsequent geospatial analysis and node matching, achieving the quantification and standardization of the user environment.

[0067] Optionally, obtaining data verification results based on device data, network data, user data, and preset verification methods further includes: when it is determined that the data verification result is abnormal, executing a compensation strategy to obtain compensation data; performing time-series marking and association binding based on the compensation data to obtain a compensation data package; obtaining compensation structured raw information based on the compensation data package and a preset structured processing method; obtaining compensation vector information based on the compensation structured raw information, a preset standardization processing method, and structured raw information, and using the compensation vector information as standard vector information.

[0068] In the specific implementation process, when an anomaly is detected in the data verification result, a compensation strategy is triggered. Based on the data reliability verification result, the system will activate the compensation strategy mechanism when any of the following conditions are met: GPS accuracy anomaly (GPS accuracy field > 100 meters or data latency > 10 seconds), RTT anomaly (RTT > 500ms or packet loss rate > 5%), DNS failure (DNS resolution time > 200ms or returned error code), marking suspicious IPs, etc., and will then use a backup positioning mechanism for compensation processing. The corresponding compensation methods and processes are as follows: When GPS is unavailable, WiFi fingerprint positioning (SSID + BSSID → local hotspot database) is used; when GPS accuracy offset occurs, GPS positioning data is fused with IP positioning data, and Kalman filtering is applied for location correction; when network quality anomalies occur, historical network window data is introduced, and a moving average is applied to correct abnormal fluctuations.

[0069] Optionally, the system acquires current interface metadata, obtains high-precision location information based on standard vector information, current interface metadata, and a Kalman filter algorithm, and then acquires at least one candidate interface that meets preset compliance conditions based on the high-precision location information, current interface metadata, and a preset location formula. This includes: acquiring GPS data and IP data based on standard vector information; acquiring interface metadata, dynamically fusing GPS data and IP data based on interface metadata and a Kalman filter algorithm to obtain high-precision location information; acquiring geographic data based on high-precision location information and a preset location formula; acquiring compliance data based on geographic data and a preset multidimensional cost function; and acquiring at least one candidate interface that meets preset compliance conditions based on the compliance data.

[0070] In the specific implementation process, a hybrid positioning engine performs geospatial analysis and node matching based on standard vector information, and uses a Kalman filter algorithm to fuse location information to obtain high-precision location information (target coordinate accuracy of ±50m), reducing GPS drift or IP positioning errors. The final calculation results are used for the next step of dynamic node topology calculation. The data input process of the hybrid positioning engine is as follows: GPS data and IP data are obtained from the device. GPS data includes latitude and longitude, accuracy, and timestamp. IP data is obtained by the edge access node by calling the IP geodatabase (such as MaxMind, IP2Location), with an accuracy of approximately ±20 to 100 kilometers, large error but wide coverage.

[0071] In the specific implementation process, when performing dynamic node topology calculation, high-precision user location (i.e., high-precision location information) and node metadata are input for data processing. A preset location formula (which can be the Haversine formula) is used to calculate geographical distance (i.e., geographical data). Then, based on a preset multidimensional cost function, such as Ci = α·di + β·li + γ·qi (where Ci is the multidimensional cost result, di is the geographical distance, li is the latency, and qi is the compliance level), nodes are sorted and optimized. This selects the Top N candidate nodes (corresponding to at least one candidate interface, N≥1) according to preset compliance conditions. The preset compliance condition can be that if the compliance level is too high (e.g., greater than 0.8), the node is excluded. The final result is used for the next step of interface preloading and scheduling. In application, this node metadata comes from the system-maintained node capability database (Node MetadataRegistry), obtained through integration with various systems within the AI ​​question-answering system, such as the compliance scanning system generating node compliance levels and the scheduling system collecting node load. The specific fields, their meanings, and sources are shown in Table 1 below.

[0072] Table 1 lists the fields in the node metadata, along with their meanings and sources.

[0073]

[0074]

[0075] In the specific implementation process, the node sorting and candidate node selection mechanism is as follows: First, all available nodes are sorted from smallest to largest according to the multidimensional cost function. Then, compliance is judged according to preset compliance conditions, such as: if q i If q > 0.8, then this node cannot provide service due to local regulations and should be excluded. i If q < 0.4, then the node is considered a low compliance risk node and will be retained with priority. i If the value is ∈[0.4, 0.8], it needs to be marked according to the laws and regulations of the user's location (such as cross-traffic blocking, etc.); finally, the Top N candidate nodes are selected according to the selection criteria based on the number of candidate nodes. The selection criteria for the number of candidate nodes can be as follows: based on the multidimensional cost results, the top N=5 optimal nodes are selected by default. If there are fewer than 5 available nodes, the nodes are degraded and scheduled according to priority. At the same time, the value of N can be dynamically adjusted according to the current traffic pressure (such as taking N=3 during peak hours).

[0076] Optionally, when it is determined based on the GPS data, IP data, and Kalman filter algorithm that the Kalman filter residual exceeds a preset residual range, historical location information is obtained, and single positioning data is obtained based on the historical location information, GPS data, IP data, and a preset single positioning standard, and the single positioning data is used as the high-precision location information; when it is determined based on the historical location information, GPS data, IP data, and the preset single positioning standard that single positioning data cannot be obtained, adjacent successful positioning information is obtained, and the adjacent successful positioning information is used as the high-precision location information.

[0077] In the specific implementation process, a discrete-time Kalman filter algorithm is used to dynamically fuse GPS and IP positioning to reduce positioning errors, especially to improve robustness when GPS drifts or is unavailable. If the Kalman filter residual is too large, it will fall back to a single positioning data. The priority of this single location data is set as: GPS > IP > historical location. When the fusion residual is too large (the error is unacceptable), the system will automatically switch to the most reliable data source based on accuracy, timeliness, and scenario. If both are unavailable, it may fall back to the last successful positioning record (i.e., adjacent successful positioning information) as a temporary substitute, ensuring the accuracy of positioning data, reducing positioning data errors, improving compliance, and especially improving robustness when GPS drifts or is unavailable.

[0078] In the specific implementation process, after selecting the candidate interfaces of ToP N, intelligent interface preloading and dynamic scheduling are carried out. A predictive cache preheating processing method is adopted. Based on the candidate interface and the user's historical trajectory information, combined with the time series prediction algorithm (LSTM time series prediction) and preset information conditions, preloading is performed to load interface resources in advance, so as to significantly reduce the switching latency. The time series prediction process using the time series prediction algorithm is as follows: An LSTM time series prediction model is constructed based on LSTM (Long Short-Term Memory Neural Network). The input prediction data includes: user historical trajectory data, such as recording the node paths and time information of the user's past N successful visits, forming a time series input; auxiliary context features, such as time features (hours, dates, etc.), location information (latitude and longitude, etc.) embedding; and the current device network status (such as RTT, packet loss rate). The constructed hierarchical structure includes an input layer, a Dropout layer (regularization technique), two LSTM layers (with 128 hidden units), and a fully connected layer. The input layer is used to input the time series (historical nodes and features), the Dropout layer is used to prevent overfitting, and the fully connected layer, combined with softmax (normalized exponential function), outputs the probability distribution of candidate nodes.

[0079] In the specific implementation process, when predicting confidence, the confidence level is defined as the softmax probability value of the Top-1 node. The pre-set confidence conditions are: if the confidence level ≥ 0.75, the prediction is considered reliable, and interface preheating is performed; if the confidence level ≤ 0.5, the prediction is considered unstable, and cache loading is not triggered. The interface preloading mechanism is explained as follows: preloading involves establishing short connections / handshake channels with the corresponding nodes and loading necessary protocol resources (such as QUIC connection cache, TLS certificate, and interface mapping table). The preloading time window includes initiating the preloading action 500ms before prediction, utilizing the main thread / IO thread to preheat and cache the connection state. Successful prediction will significantly reduce switching latency and provide data input from multiple preheated nodes / protocol channels for subsequent dual-channel load balancing. Dual-channel load balancing is based on multiple pre-warmed node / protocol channels, and uses a data processing method that combines a complete response and improves availability by using parallel primary and backup interfaces (HTTP / 2, WebSocket, QUIC pre-connection), prioritizing the fastest response channel, and switching to the backup channel if the primary channel is unreliable (if the primary channel does not respond within 50ms, the backup channel is immediately activated).

[0080] Optionally, dynamic evaluation scores can be obtained in real time based on the main interface and dynamic interfaces, including:

[0081] Real-time acquisition of dynamic data from the main interface and dynamic interfaces;

[0082] A preset normalized weighted algorithm and dynamic interface data are used to obtain dynamic evaluation scores. The dynamic interface data includes latency score, risk score, availability score, freshness score, and energy consumption score. The calculation process satisfies the following formula:

[0083] Q = w1 * S L +w2*S C +w3*S A +w4*S F +w5*S E

[0084] S L =exp(-L / 150)

[0085] S C =1-C

[0086] S A =A

[0087] S F =exp(-F / 10)

[0088] S E =exp(-E / 1)

[0089] In the formula: Q represents the dynamic evaluation score, S L Represents the delay score, where L represents the effective delay, and S represents the delay score. C C represents the risk score, and S represents the compliance index. A The interface availability score is represented by A, where A represents interface availability and S represents interface availability. F The score represents the freshness of the data, where F represents the freshness of the data and S represents the freshness of the data. E The values ​​represent energy consumption score, E represents energy efficiency, w1 represents the weight index corresponding to the latency score, w2 represents the weight index corresponding to the risk score, w3 represents the weight index corresponding to the availability score, w4 represents the weight index corresponding to the freshness score, and w5 represents the weight index corresponding to the energy consumption score.

[0090] Optionally, dynamic evaluation scores can be obtained in real time based on the main interface and several backup interfaces. This also includes: when it is determined that the dynamic evaluation score does not exceed the preset threshold range, there is no need to perform intelligent switching action.

[0091] In the specific implementation process, a real-time quality monitoring and adaptive adjustment mechanism is introduced, including dynamic QoS assessment and reinforcement learning parameter optimization. Dynamic QoS assessment is achieved by monitoring the dynamic data of the interface in real time and calculating the dynamic assessment score using normalization and weighting methods. The interface dynamic data typically refers to relevant dynamic assessment indicators, including effective latency, regulatory compliance index, interface availability, data freshness, and energy efficiency. These data are normalized and weighted to obtain the QoS score (i.e., the dynamic assessment score), which guides the interface scheduling strategy according to a preset threshold range. Among the various indicator data, S... L Represents the latency score (exponential decay), where L represents the effective latency, i.e., the actual RTT of the user request (including network and processing latency), usually measured in milliseconds (ms), and S... C C represents the risk score; C represents the compliance index, which is the current compliance level of the interface (the degree of compliance with local regulations), with a value range of 0 to 1. The lower the value, the safer, the lower the risk, the higher the compliance, and the higher the corresponding score; S A This represents the interface availability score (the raw value of A), where A represents interface availability, i.e., the connection success rate of the interface, and the value ranges from 0 to 1; S F The value represents the freshness score, where F represents data freshness, i.e., the degree of synchronization between the interface data and the main source data. The unit of the value is seconds (s), and the lower the value, the better, meaning the newer the data, the better; S EThe values ​​represent energy consumption scores, where E represents energy efficiency, i.e., the average power consumption per response, measured in mW / response. A lower value is better, meaning lower energy consumption results in a higher score. w1 represents the weight index corresponding to the latency score, w2 represents the weight index corresponding to the risk score, w3 represents the weight index corresponding to the availability score, w4 represents the weight index corresponding to the freshness score, and w5 represents the weight index corresponding to the energy consumption score. The weights can be adjusted according to the actual situation. An example of a weight is: [w1, w2, w3, w4, w5] = [0.3, 0.2, 0.2, 0.15, 0.15].

[0092] Optionally, the current interface policy network is trained based on a preset policy gradient algorithm, and convergence data is obtained based on the current interface policy network. When the convergence data is determined not to meet the preset convergence conditions, the original learning parameters are obtained, and the target learning parameters are obtained based on the original learning parameters and the preset adjustment method. The current interface is then configured based on the target learning parameters.

[0093] Optionally, the current interface policy network is trained based on a preset policy gradient algorithm, and convergence data is obtained based on the current interface policy network, including: obtaining the total number of training epochs, window data, and average reward based on the current interface policy network; the convergence data includes the average reward growth rate; the average reward growth rate is obtained based on the total number of training epochs, window data, and average reward, and its calculation process satisfies the following formula:

[0094]

[0095] In the formula: ARGR represents the average reward growth rate. This represents the average reward from round T to round Tw. This represents the average reward from round Tw to round T-2w.

[0096] When the average reward growth rate meets the preset growth rate condition and the preset duration, the convergence data is determined to not meet the preset convergence condition.

[0097] In the specific implementation process, the current interface policy network is continuously trained through a preset policy gradient algorithm, such as PPO, and the convergence data of the relevant interface is detected and acquired in real time. This includes QoS indicator data, node status data, and user behavior data. When the preset convergence condition is not met, i.e., convergence is slow, the hyperparameters are adjusted through adaptive decision-making: target learning parameters are obtained to configure the current interface. Specifically, "slow convergence" refers to the policy network's policy reward (or reward function) stagnating or changing slowly over multiple training periods, meaning the current learning rate, batch size, and other hyperparameter settings are insufficient to drive effective learning. Pre-set rules for judging slow convergence are provided, for example: an average reward growth rate (ARGR) < 0.02 (i.e., a growth rate less than 2%) for three consecutive window periods is considered slow convergence. An example calculation is as follows: if the average reward for rounds 40-50 is 0.582, and the average reward for rounds 50-60 is 0.572, then at this time... If the observation window is less than 2% for three consecutive observations, then the convergence is considered slow.

[0098] In the specific implementation process, the target learning parameters include the learning rate and the exploration epsilon. The initial value of the learning rate (η) is assumed to be 0.0003 (default PPO setting). If convergence is slow, the learning rate can be automatically adjusted according to a preset adaptive adjustment method, such as: η new =η old *1.5 Note that the value of η needs to be controlled to an upper limit not exceeding 0.002 to prevent training oscillations. The exploration rate is used to control the trade-off between "exploration" and "exploitation" in the policy network. Assuming the initial exploration rate is ε = 0.1, if convergence is slow, it indicates that the policy over-relys on historical experience. The exploration rate can be automatically adjusted according to a preset adaptive adjustment method, such as: ε new =min(ε old +0.05, 0.3).

[0099] Optionally, real-time monitoring information can be acquired to obtain latency data. When the latency data is determined to exceed the preset maximum increase, a three-level rollback operation is executed.

[0100] In the specific implementation process, a disaster recovery and auditing mechanism is introduced through an intelligent three-level rollback strategy. Real-time monitoring information is obtained to detect network latency data of the interface in real time. Then, based on preset rollback trigger conditions, a three-level rollback operation is executed promptly when the latency data exceeds a preset maximum increase, thus constructing an intelligent rollback strategy. For example, the preset rollback trigger condition is a latency increase > 200% after switching interfaces, and the three-level rollback is: backup node in the same region → local cache → global default interface.

[0101] In the specific implementation process, the disaster recovery and auditing mechanism introduced also includes full-link tracing (blockchain logs). Detailed data is entered and recorded for each request or interface switching operation, and the data is written into the blockchain smart contract. It is tamper-proof and judged and processed based on preset audit conditions. If writing to the chain fails, it is retried or a buffer is recorded to achieve financial-grade auditing.

[0102] This embodiment provides an intelligent interface switching method based on multi-dimensional perception. It fuses device data, network data, and user data to obtain multi-source correlated data, combining the real-time performance of the interface network with the connection between the device context. The method introduces data fusion, reliability verification, compensation mechanisms, time-series tags, and context binding. Data processing is performed according to the data detection results, and the data is structured, standardized, and subjected to geospatial analysis, significantly improving the dimensionality and accuracy of environmental perception and data reliability. Dynamic interface selection and pre-loading are achieved through dynamic node topology calculation, Kalman filter positioning, and compliance judgment. Combined with time-series prediction and confidence judgment, dual-channel parallel operation and intelligent switching scheduling of primary and backup interfaces are realized, improving interface availability, reducing positioning errors, and enhancing robustness. Simultaneously, real-time quality monitoring and adaptive adjustment are performed, with real-time detection of interface dynamics for intelligent interface switching, ensuring interface availability, data freshness, and regulatory compliance, reducing latency, and improving energy efficiency. Hyperparameter adjustment, disaster recovery, and auditing mechanisms are introduced to ensure stable operation during the switching process, achieving millisecond-level intelligent interface switching between primary and backup interfaces based on multi-dimensional perception.

[0103] Example 2:

[0104] See Figure 2 This embodiment provides an intelligent interface switching process for an AI question-answering system based on multi-dimensional perception, implemented using the aforementioned intelligent interface switching method based on multi-dimensional perception, including:

[0105] start;

[0106] S21. Data acquisition verification;

[0107] S22. Determine if the GPS / IP data is valid: if yes, proceed to step S231; if no, proceed to step S232.

[0108] S231. Perform geographic data augmentation;

[0109] S232. If the fallback process logs or marks the data as invalid (NULL), the compensation mechanism will be triggered.

[0110] S24. Construct the geospatial analysis model GeoContext and calculate the confidence level;

[0111] S25, Decision Engine: Determines data center region policies (region policies) based on specific coding rules (regionCode) / confidence levels;

[0112] S26, Model routing distribution weighted round-robin;

[0113] S27. Call the AI ​​model to obtain the results;

[0114] S28, Model Response Tracking;

[0115] S29. Determine if there is an anomaly: if yes, proceed to step S210; otherwise, proceed directly to step S211.

[0116] S210, Triggering downgraded circuit breaker selection of backup model;

[0117] S211, Return the result;

[0118] Finish.

[0119] In the specific implementation process, the geospatial analysis in step S231 is to enhance the geographic features, which requires querying the IP address and calculating based on Haversine. In step S232, the fallback process records logs or marks NULL, and also includes triggering a compensation mechanism to perform data compensation.

[0120] Example 3:

[0121] Based on the above-described method embodiments, this invention provides an intelligent interface switching system based on multi-dimensional perception, comprising:

[0122] The data processing module is used to clean and compensate the raw collected data to output JSON structured data.

[0123] The feature extraction module is used to quantify the user environment based on JSON structured data to output a standardized feature vector;

[0124] The node calculation module is used to calculate and select nodes based on standardized feature vectors and node metadata to output node matching results.

[0125] The policy training module is used to reinforce learning the input state space based on standardized feature vectors and QoS data to output policy training.

[0126] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the intelligent interface switching method based on multi-dimensional perception provided by any of the above method item embodiments of the present invention.

[0127] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent interface switching based on multi-dimensional perception, characterized in that, include: Acquire device data, network data, and user data, and obtain standard vector information based on a preset structured processing method, a preset standardized processing method, and the device data, network data, and user data; Obtain the current interface metadata, obtain high-precision location information based on the standard vector information, the current interface metadata and the Kalman filter algorithm, and then obtain at least one candidate interface that meets the preset compliance conditions based on the high-precision location information, the current interface metadata and the preset location formula. Obtain historical trajectory information, and based on the time-series prediction algorithm, the historical trajectory information, the current interface metadata, and the candidate interfaces, obtain at least one dynamic interface that meets the preset information conditions; Based on the aforementioned dynamic interface, the main interface and several backup interfaces are preloaded; Based on the main interface and several backup interfaces, dynamic data of the interfaces are obtained in real time, and dynamic evaluation scores are obtained based on the dynamic data of the interfaces. When the dynamic evaluation score is determined to exceed the preset threshold range, an intelligent switching action is performed: based on the dynamic data of the interface, the target interface among the main interface and several backup interfaces is obtained, and the current interface is switched to the target interface.

2. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, The acquisition of device data, network data, and user data, based on a preset structured processing method, a preset standardized processing method, and the acquisition of standard vector information from the device data, network data, and user data, includes: Acquire device data, network data, and user data, and obtain data verification results based on the device data, network data, user data, and a preset verification method; When it is determined that the reliability verification result is not abnormal, time-series marking and association binding are performed based on the device data, network data, and user data to obtain data packets; Structured raw information is obtained based on the data package and the preset structured processing method; Standard vector information is obtained based on the structured raw information and the preset standardization processing method.

3. The intelligent interface switching method based on multi-dimensional perception as described in claim 2, characterized in that, The process of obtaining data verification results based on the device data, network data, user data, and a preset verification method further includes: When the data verification result is determined to be abnormal, a compensation strategy is executed to obtain compensation data; Based on the compensation data, time-series marking and association binding are performed to obtain the compensation data package; The original structured information of the compensation is obtained based on the compensation data package and the preset structured processing method; Based on the aforementioned structured original information and the preset standardized processing method, compensation vector information is obtained, and the compensation vector information is used as the standard vector information.

4. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, The process of obtaining current interface metadata, obtaining high-precision location information based on the standard vector information, current interface metadata, and Kalman filtering algorithm, and then obtaining at least one candidate interface that meets preset compliance conditions based on the high-precision location information, current interface metadata, and preset location formula includes: GPS data and IP data are obtained based on the aforementioned standard vector information; Obtain interface metadata, and dynamically fuse the GPS data and IP data based on the interface metadata and Kalman filter algorithm to obtain high-precision location information; Geographic data is obtained based on the high-precision location information and the preset location formula; Compliance data is obtained based on the aforementioned geographic data and a preset multidimensional cost function; Based on the compliance data, at least one candidate interface that meets the preset compliance conditions is obtained.

5. The intelligent interface switching method based on multi-dimensional perception as described in claim 4, characterized in that, Also includes: When the Kalman filter residual is determined to exceed the preset residual range based on the GPS data, IP data, and Kalman filter algorithm, historical location information is obtained. Based on the historical location information, GPS data, IP data, and preset single positioning standard, single positioning data is obtained and used as the high-precision location information. When it is determined that a single location data cannot be obtained based on the historical location information, GPS data, IP data, and a preset single location standard, adjacent successful location information is obtained and used as the high-precision location information.

6. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, The step of acquiring dynamic interface data in real time based on the main interface and the dynamic interface, and then acquiring a dynamic evaluation score based on the dynamic interface data, includes: Real-time acquisition of interface dynamic data based on the main interface and dynamic interface; A preset normalized weighted algorithm and the interface dynamic data are used to obtain a dynamic evaluation score, wherein the interface dynamic data includes latency score, risk score, availability score, freshness score, and energy consumption score, and the calculation process satisfies the following formula: Q=w1*S L +w2*S C +w3*S A +w4*S F +w5*S E S L =exp(-L / 150) S C =1-C S A =A S F =exp(-F / 10) S E =exp(-E / 1) In the formula: Q represents the dynamic evaluation score, S L Represents the delay score, where L represents the effective delay, and S represents the delay score. C Represents risk score, C represents compliance index, S A The interface availability score is represented by A, where A represents interface availability and S represents interface availability. F The score represents the freshness of the data, where F represents the freshness of the data and S represents the freshness of the data. E The values ​​represent energy consumption score, E represents energy efficiency, w1 represents the weight index corresponding to the latency score, w2 represents the weight index corresponding to the risk score, w3 represents the weight index corresponding to the availability score, w4 represents the weight index corresponding to the freshness score, and w5 represents the weight index corresponding to the energy consumption score.

7. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, The method of obtaining dynamic data of the interface in real time based on the main interface and several backup interfaces, and obtaining a dynamic evaluation score based on the dynamic data of the interface, further includes: when it is determined that the dynamic evaluation score does not exceed the preset threshold range, there is no need to perform an intelligent switching action.

8. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, Also includes: The current interface policy network is trained based on a preset policy gradient algorithm, and convergence data is obtained based on the current interface policy network. When the convergence data does not meet the preset convergence conditions, the original learning parameters are obtained, and the target learning parameters are obtained based on the original learning parameters and the preset adjustment method. The current interface is then configured based on the target learning parameters.

9. The intelligent interface switching method based on multi-dimensional perception as described in claim 8, characterized in that, The process of training the current interface policy network based on a preset policy gradient algorithm and obtaining convergence data based on the current interface policy network includes: Based on the current interface policy network, obtain the total number of training rounds, window data, and average reward; The convergence data includes the average reward growth rate; The average reward growth rate is obtained based on the total number of training rounds, window data, and average reward, and its calculation process satisfies the following formula: In the formula: ARGR represents the average reward growth rate. This represents the average reward from round T to round Tw. This represents the average reward from round Tw to round T-2w. When the average reward growth rate meets the preset growth rate condition and the preset duration, it is determined that the convergence data does not meet the preset convergence condition.

10. The intelligent interface switching method based on multi-dimensional perception as described in claim 1, characterized in that, include: Real-time monitoring information is acquired to obtain delay data based on the monitoring information. When the delay data is determined to exceed the preset maximum increase, a three-level rollback operation is executed.

Images