Supporting intent-driven cross-level video algorithm power scheduling multi-dimensional retrieval engine and method

Abstract

The application discloses a kind of multi-dimensional retrieval engine and method supporting intent-driven cross-level video algorithm power scheduling, comprising: intention analysis module, for the task intention information of user input obtained is parsed into the task feature vector consisting of key task elements;System state monitoring unit, for obtaining real-time system state information and generating system state vector;Strategy generation module, for generating element dimension weight strategy;Parallel matching module, for matching the algorithm power node in video algorithm power resource, obtain algorithm power node matching table;Result optimization module, for the optimization screening of algorithm power node according to algorithm power node matching table, obtain optimal algorithm power node and its execution parameter list.The application can realize the automatic analysis and mapping of user intention, scene adaptive dynamic strategy generation, intelligent scoring of multi-factor fusion and feedback-based continuous learning optimization, ultimately comprehensively improve the ease of use, accuracy and intelligent level of scheduling system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of computer technology, video processing and computing power scheduling, and particularly relates to a multi-dimensional retrieval engine and method supporting intent-driven cross-level video computing power scheduling. BACKGROUND

[0002] With the wide deployment of video monitoring devices (such as fixed cameras, mobile devices, aerial devices, etc.), the demand for cross-regional and cross-level video data processing has surged, and higher requirements have been put forward for the scheduling efficiency and matching accuracy of video computing power resources (such as video analysis algorithms, computing nodes, etc.). The retrieval method of the existing video computing power scheduling system is mostly based on a single dimension (such as only by region or algorithm type), which has the following defects:

[0003] 1. Single dimension: unable to simultaneously consider multi-dimensional constraints such as administrative division level, algorithm type, time characteristics, device type, and security level, resulting in large resource matching deviation;

[0004] 2. Poor dynamic adaptability: dynamic factors such as task period, resource load, and node fluctuation are not considered, making it difficult to adapt to special scenarios such as night and emergency tasks;

[0005] 3. Insufficient security control: lack of fine-grained verification of resource security level and requester authority, with data security risks;

[0006] 4. Fixed scoring mechanism: fixed weights, unable to dynamically adjust according to resource state or task attributes, affecting scheduling flexibility.

[0007] In addition, the existing retrieval method requires users to explicitly and structurally fill in all dimensional parameters, which is tedious and requires high user professional background, and in the face of sudden and cross-field emergency video analysis tasks, the response efficiency is low, and the intelligent service experience of "thinking and getting" cannot be achieved.

[0008] Therefore, how to further reduce the use threshold on the basis of realizing multi-dimensional accurate retrieval and improve the understanding ability of the system to the real intention of the user and the self-adaptive scheduling ability to the dynamic environment has become a technical bottleneck to be broken through. SUMMARY

[0009] The present application aims to solve the problems of single retrieval dimension, poor dynamic adaptability, insufficient safety control, fixed scoring mechanism, complex user operation and inability to intelligently understand business intent in the prior art, and provides a multi-dimensional retrieval engine and method supporting intent-driven cross-level video computing power scheduling, to realize automatic analysis and mapping of user intent, scene-adaptive dynamic strategy generation, intelligent scoring of multiple factors and feedback-based continuous learning optimization, and ultimately comprehensively improve the ease of use, accuracy and intelligent level of the video computing power scheduling system where the retrieval engine is located.

[0010] The technical solutions of the present application are as follows:

[0011] One, a multi-dimensional retrieval engine supporting intent-driven cross-level video computing power scheduling

[0012] An intent analysis module is configured to parse the obtained user input task intent information into a task feature vector composed of key task elements.

[0013] A system state monitoring unit is configured to obtain real-time system state information and generate a system state vector.

[0014] A strategy generation module is configured to generate an element dimension weight strategy according to the task feature vector and the system state vector.

[0015] A parallel matching module is configured to match computing power nodes in video computing power resources according to the task feature vector and the element dimension weight strategy, and obtain a computing power node matching table.

[0016] A result optimization module is configured to optimize and screen the computing power nodes according to the computing power node matching table, and obtain optimal computing power nodes and their execution parameter lists.

[0017] The strategy generation module comprises:

[0018] A strategy knowledge base is configured to store strategy base templates and strategy adjustment templates.

[0019] A first strategy generation unit is configured to select templates from the strategy knowledge base and generate an element dimension weight strategy according to the task feature vector and the system state vector.

[0020] The strategy generation module further comprises:

[0021] A strategy fine-tuning unit is configured to generate strategy fine-tuning information according to the task feature vector and the system state vector.

[0022] A second strategy generation unit is configured to adjust the element dimension weight strategy generated by the first strategy generation unit in combination with the strategy fine-tuning information, and obtain a final element dimension weight strategy.

[0023] The strategy adjustment template comprises one or more of an emergency task strategy adjustment template, a high-load strategy adjustment template, and a high-security strategy adjustment template.

[0024] In the first strategy generation unit, if it is identified that the task feature vector contains a key task element related to an emergency semantic feature or the task queue length in the system state vector exceeds a preset queue length, an element dimension weight strategy is generated according to the emergency task strategy adjustment template;

[0025] If it is identified that the system load in the system state vector is higher than a preset load threshold, an element dimension weight strategy is generated according to the high-load strategy adjustment template;

[0026] If it is identified that the task feature vector contains a key task element related to a security feature, an element dimension weight strategy is generated according to the high-security strategy adjustment template;

[0027] If two or more of the emergency task strategy adjustment template, the high-load strategy adjustment template, and the high-security strategy adjustment template are simultaneously called, the corresponding template is selected according to a preset priority order, so as to output the corresponding element dimension weight strategy;

[0028] Otherwise, an element dimension weight strategy is generated according to the strategy base template.

[0029] The pre-trained reinforcement learning model is stored in the strategy fine-tuning unit, the state of the reinforcement learning model is the system state vector and the task feature vector, the action is the strategy fine-tuning information, and the reward function R during training is constructed according to the task execution success rate and the task response delay.

[0030] In the result optimization module, the resource health degree of each computing power node in the computing power node matching table and the intention matching degree of each computing power node and the task feature vector are calculated, the task execution degree of each computing power node is calculated according to the matching degree, the resource health degree, and the intention matching degree of each computing power node, the task execution degrees of the remaining computing power nodes in the computing power node matching table are iteratively calculated, the computing power nodes in the computing power node matching table are arranged in descending order of the task execution degree, and thus the optimal computing power node and the execution parameter list thereof are output.

[0031] The cross-level video computing power scheduling multi-dimensional retrieval engine further comprises:

[0032] The feedback learning module is configured to collect execution feedback data of the whole task scheduling cycle, and optimize the intention analysis module and the strategy generation module according to the execution feedback data of the whole task scheduling cycle.

[0033] II. A video computing power scheduling system

[0034] The video computing power scheduling system comprises the multi-dimensional retrieval engine supporting the intention-driven cross-level video computing power scheduling.

[0035] Three, a multi-dimensional retrieval method supporting intention-driven cross-level video computing power scheduling

[0036] First, the task intention information input by the user is parsed into a task feature vector composed of key task elements; real-time system state information is obtained and a system state vector is generated; then, an element dimension weight strategy is generated according to the task feature vector and the system state vector; then, the computing power nodes in the video computing power resources are matched according to the task feature vector and the element dimension weight strategy, and a computing power node matching table is obtained; finally, the optimal computing power node and its execution parameter list are obtained through optimization and screening of the computing power nodes according to the computing power node matching table.

[0037] Compared with the prior art, the application has the following beneficial effects:

[0038] 1. Extremely easy to use: users do not need to learn complex dimension parameters, and can obtain accurate computing power scheduling service through natural expression, greatly reducing the use threshold.

[0039] 2. Intelligent decision-making: the system can understand user intentions and make optimal decisions in combination with real-time system state, making the scheduling process more intelligent and more suitable for actual business scenarios.

[0040] 3. Continuous evolution: the introduction of a feedback learning mechanism makes the retrieval engine no longer static, but can continuously optimize its intention understanding ability and scheduling strategy as the number of uses increases.

[0041] 4. Reliable guarantee: the resource health assessment mechanism adds a reliability verification dimension to the scheduling decision, effectively reducing the risk of task failure caused by resource node problems.

[0042] 5. More scientific decision-making: through the hybrid decision-making mechanism of the fusion strategy template and reinforcement learning, the limitations of pure threshold judgment are overcome, making the strategy generation more intelligent and practical, and significantly improving the scheduling quality in complex scenarios.

[0043] 6. More accurate evolution: a specific, multi-level parameter optimization mechanism is designed, so that the feedback learning process is no longer a black box, but a continuous self-optimization process with clear goals, algorithms and paths. The system evolution direction is clear and the effect is measurable. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The overall framework of the multi-dimensional retrieval engine supporting intention-driven cross-level video computing power scheduling is provided for the application Figure 1 .

[0045] Figure 2 An overall framework diagram of a multi-dimensional retrieval engine supporting intent-driven cross-level video computing power scheduling Figure 2 .

[0046] Figure 3 A framework diagram of a strategy generation module Figure 1 .

[0047] Figure 4 A framework diagram of a strategy generation module Figure 2 .

[0048] Figure 5 A framework diagram of a feedback learning module. DETAILED DESCRIPTION

[0049] In order for those skilled in the art to better understand the present application, so as to make a clearer limitation on the scope of the present application, the present application is described in detail below with reference to some specific embodiments of the present application and the accompanying drawings. It should be noted that the following is only some specific embodiments of the present application concept and is only a part of the embodiments of the present application, and the specific and direct description of the related structure is only for the convenience of understanding the present application, and each specific feature does not of course, directly limit the scope of the present application. The routine selection and replacement made by those skilled in the art under the guidance of the present application concept, and the reasonable arrangement and combination of several technical features under the guidance of the present application concept, should be regarded as within the scope of the present application claimed.

[0050] As shown in Figure 1 , the present application proposes a multi-dimensional retrieval engine supporting intent-driven cross-level video computing power scheduling, the retrieval engine comprising:

[0051] An intent analysis module for parsing the obtained user input task intent information into a task feature vector composed of key task elements;

[0052] In a feasible implementation, the form of the user input task intent information is natural language, structured form and / or feature picture. A pre-trained model corresponding to the form is used to parse and extract the key task elements.

[0053] Key mission elements include elements of administrative division dimension (D1), video analysis algorithm dimension (D2), time-sensitive dimension (D3), data source type dimension (D4) and security level dimension (D5). Among them, the elements of the administrative division dimension (D1) include single-level regions or nested combinations of multi-level regions, such as upper region = province, lower region = city. The elements of the video analysis algorithm dimension (D2) include single video task label or combination of multi-video task labels, such as vehicle-related research and judgment, face clustering. The elements of the time-sensitive dimension (D3) include task time period and task urgency; such as task time period is night, then subsequent priority matching of power node containing infrared algorithm; if the task intent information contains "immediately" and "involving cases", "immediately" and "involving cases" are described as task urgency. The elements of the data source type dimension (D4) include video source device types, such as fixed cameras, mobile devices, aerial devices. The elements of the security level dimension (D5) include low security level, medium security level and high security level.

[0054] The system state monitoring unit is configured to obtain real-time system state information and generate a system state vector; the real-time system state information includes system load, resource pool change rate and task queue length of the video computing power scheduling system, that is, the system state monitoring unit is configured to monitor the real-time state of the entire video computing power scheduling system.

[0055] The policy generation module is configured to generate an element dimension weight strategy according to the task feature vector and the system state vector; the element dimension weight strategy is represented as [W1, W2, W3, W4, W5], wherein W1, W2, W3, W4, and W5 are the weights of the administrative division dimension, the weights of the video analysis algorithm dimension, the weights of the time-sensitive dimension, the weights of the data source type dimension, and the weights of the security level dimension, respectively.

[0056] As shown in Figure 3 , the policy generation module includes:

[0057] The policy knowledge base is configured to store policy base templates and policy adjustment templates.

[0058] The first policy generation unit is configured to select templates from the policy knowledge base and generate an element dimension weight strategy according to the task feature vector and the system state vector.

[0059] In a feasible implementation manner, as shown in Figure 4 , the policy generation module further includes:

[0060] The policy fine-tuning unit is configured to generate policy fine-tuning information according to the task feature vector and the system state vector.

[0061] The second strategy generation unit is configured to adjust the element dimension weight strategy generated by the first strategy generation unit in combination with the policy fine-tuning information, to obtain a final element dimension weight strategy.

[0062] The policy adjustment templates include, but are not limited to, an emergency task policy adjustment template, a high-load policy adjustment template, and a high-security policy adjustment template.

[0063] In the first strategy generation unit, if it is identified that the task feature vector contains a key task element related to a preset emergency semantic feature (i.e., the task intention information contains "emergency", "immediately", or "case-related") or the length of the task queue in the system state vector exceeds a preset queue length, which is denoted as an emergency feature, the element dimension weight strategy is generated according to the emergency task policy adjustment template. The template usually significantly increases the weight of the administrative division dimension (D1) to ensure that the task is quickly dispatched to the nearest resource node, reduces communication delay, and may reduce the weight of other non-core dimensions (such as [W1, W2, W3, W4, W5] = [0.6, 0.2, 0.1, 0.0, 0.1]).

[0064] If it is identified that the system load in the system state vector is higher than a preset load threshold (such as 85%), which is denoted as a load feature, the element dimension weight strategy is generated according to the high-load policy adjustment template. Optionally, the high-load policy adjustment template calculates an increase weight of the administrative division dimension according to the system load, the increase weight is positively correlated with the load rate, and the weight of the time-sensitive dimension (D3) is reduced. The element dimension weight strategy is output after adjusting the weights of the administrative division dimension and the time-sensitive dimension based on the base strategy stored in the base policy adjustment template. The core strategy of the load policy adjustment template is to increase the weight of the administrative division dimension (D1) and reduce the weight of the time-sensitive dimension (D3). The purpose is to prioritize the "geographical location proximity" of the resource under high-load pressure, which is the most reliable scheduling principle, and temporarily relax the requirement for "optimal time period" to prioritize the stability of the system and the successful scheduling of the task.

[0065] If it is identified that the task feature vector contains a key task element related to a preset security feature, which is denoted as a security feature, the element dimension weight strategy is generated according to the high-security policy adjustment template. The core role of the template is not only to maintain the weight of the security level dimension (D5) at a relatively high level, but more importantly, to automatically trigger a series of security instructions such as encrypted channel pre-negotiation and enhanced audit logs, which work together with the weight adjustment strategy to meet the security compliance requirements of the task.

[0066] Otherwise, the element dimension weight strategy is generated according to the base policy template.

[0067] Optionally, the element dimension weight strategy [W1, W2, W3, W4, W5] = [0.4, 0.3, 0.1, 0.1, 0.1] stored in the policy base template.

[0068] In an implementable embodiment, if two or more of the emergency feature, the load feature, and the security feature occur simultaneously, the corresponding template is selected according to the preset priority order, so as to output the corresponding element dimension weight strategy, and then the second strategy generation unit is used for fine-tuning, so as to generate the final element dimension weight strategy.

[0069] Optionally, the priority order is set based on the urgency and importance of the business logic. The priority order from high to low is security> emergency> load. For example, if the security feature and the emergency feature exist simultaneously, the high security strategy adjustment template is selected to generate the strategy.

[0070] The model in the second strategy generation unit generates strategy fine-tuning information (for example: “on the basis of the high security strategy adjustment template, moderately increase the weight of the administrative division dimension (D1) to reduce the delay caused by the emergency”), and the second strategy generation unit combines the strategy generated by the second strategy generation unit and the strategy fine-tuning information to generate the final element dimension weight strategy considering all situations.

[0071] The weight dynamic adjustment unit is also stored in the policy base template, and the calculation formula of the weight change value of each dimension is as follows:

[0072] ;

[0073] ;

[0074] ;

[0075] wherein, ΔW i is the weight change value of the dimension i, α and β are two weight coefficients, α+β=1, α=0.6, and β=0.4; S i is the historical success rate of the dimension i, S base is the base success rate of the dimension i, and ΔRi is the resource pool change rate of the dimension i. The calculation window T of the resource pool change rate is 24 hours. i,t Nt is the number of available resources associated with the dimension i in the current time period t; by default, t=5 minutes. i,t-1 Nt-1 is the number of available resources associated with the dimension i in the last monitoring period t-1; optionally, the resource pool change rate monitoring period can be configured as t={1 minute: emergency task mode; 5 minutes: regular mode}. i W is the weight of the dimension i.

[0076] In one possible implementation, the specific security instructions include, but are not limited to:

[0077] 1. Encryption channel pre-negotiation: Before the dispatch instruction is issued, the engine will preferentially establish a secure communication link with the target computing power node. Specifically, if the security level Lr of the computing power node is greater than or equal to 2 (or according to the security policy configuration), a dynamic encryption channel is forcibly enabled. The encryption algorithm uses TLS 1.3 or the national SM4 algorithm, and the encryption key is dynamically generated and distributed by the platform key management system, and is valid for a single task.

[0078] 2. Audit log enhancement: The log recording level of the entire life cycle of the task scheduling and execution is upgraded to the "audit" level. Specifically:

[0079] Log content expansion: In addition to the regular log, the task requester's identity ID, task feature vector, scheduled computing power node ID, used encryption channel identifier, policy generation reason (such as "trigger high security template"), data access flow, etc. are additionally recorded.

[0080] Log storage protection: The audit log of the task is synchronized in real time to the tamper-proof log server, and a higher-privilege access control policy is set, which is only called by the security audit module.

[0081] 3. Access control enhancement: The instruction is issued to the target computing power node, and the access control list (ACL) is temporarily tightened during the task execution. For example, only requests from the retrieval engine and certain authorized users are allowed, and the default policy is restored after the task is completed.

[0082] In one possible implementation, the online node change rate in the system state vector is also considered. If it is identified that the online node change rate in the system state vector is higher than the preset change rate threshold, the high-load strategy adjustment template is called to calculate the new weight of the video analysis algorithm dimension, and the element dimension weight strategy is output after the weight adjustment based on the basic strategy stored in the strategy basic template.

[0083] Optionally, the calculation formula of the new weight W2 of the video analysis algorithm dimension is as follows:

[0084] W2 = min(0.8, W2 init ×(1+△N / 20));

[0085] Where △N is the percentage value of the node change rate, and W2 init is the initial weight value before compensation, that is, the weight of the video analysis algorithm dimension in the strategy basic template.

[0086] In one possible implementation, the element dimension weight strategy is output after the weight adjustment based on the basic strategy stored in the strategy basic template, specifically:

[0087] The new weight W2 of the video analysis algorithm dimension is greater than the initial weight value W2 before compensation init , the weight difference Δ before and after compensation is calculated, the weight change amount of other dimensions is calculated according to the weight difference Δ and the weight ratio of other dimensions except the video analysis algorithm dimension, and then the new weight of other dimensions is calculated, so that the weight sum of the new weight W2 of the video analysis algorithm dimension and the new weight of other dimensions is 1. For example, the weight ratio of other dimensions W1: W3: W4: W5 = 1: 1: 1: 1, and the weight difference Δ is 0.1, then the weight change amount of other dimensions is -0.025.

[0088] The pre-trained reinforcement learning model is stored in the policy fine-tuning unit, the state of the reinforcement learning model is the system state vector and the task feature vector, the action is the policy fine-tuning information, and the reward function R during training is constructed according to the task execution success rate U and the task response delay T, and the formula satisfies R=f(U,T). Users can configure specific reward calculation formulas according to their business needs, and R=f(U,T) covers all possible function forms for calculating rewards according to task execution success rates and task response delays, including linear weighting, nonlinear transformation (such as logarithm, exponential), threshold-based segmented functions, and various implementation methods.

[0089] The task execution success rate (U) refers to the index of a single task being successfully completed within the scheduling life cycle. Its value is a binary variable, success (i.e. U=1): the task is received by the target computing power node, and the expected format of the non-empty result set is returned within the maximum allowed time T max , and the result status code is success. Failure (i.e. U=0): the task fails to meet the success condition for any reason, including but not limited to: the computing power node is unresponsive, the execution process is timed out (T response >T max ), an error is returned, the returned result set is empty or has a format error.

[0090] The task response delay (T) refers to the time interval (unit: seconds) experienced from the start of the task instruction issued by the retrieval engine to the final selected computing power node to the reception of the first batch of valid result data returned by the computing power node. The task response delay is mainly used to measure the response agility of the scheduling strategy, rather than the total processing time of the task.

[0091] The parallel matching module is configured to match the computing power nodes in the video computing power resource according to the task feature vector and the element dimension weight strategy, and obtain a computing power node matching table.

[0092] In the matching process of the task feature vector and the computing power node, the matching process of different dimensions of key task elements is different. The matching score of each computing power node in the administrative division dimension (D1), the video analysis algorithm dimension (D2), the time sensitivity dimension (D3), the data source type dimension (D4), and the security level dimension (D5) matches the task demand, which constitutes the dimension matching value of the computing power node, used for transparency display and audit traceability.

[0093] For the administrative division dimension, the matching value of the administrative division dimension of the computing power node is determined according to the matching degree of the elements of the administrative division dimension in the task feature vector and the elements of the administrative division dimension of the computing power node. The administrative division dimension is generally the postal code of the superior region and the subordinate region. If the elements of the administrative division dimension in the task feature vector and the computing power node are completely matched, the matching value of the administrative division dimension is 1.0; if only the postal code of the superior region is matched, the matching value of the administrative division dimension is 0.7; if the two are adjacent regions (geographical distance ≤ 50km), the matching value of the administrative division dimension is 0.3; otherwise, there is no match, and the matching value of the administrative division dimension is 0.

[0094] For the video analysis algorithm dimension, the matching value of the video analysis algorithm dimension of the computing power node is determined according to the matching degree of the elements of the video analysis algorithm dimension in the task feature vector and the elements of the video analysis algorithm dimension of the computing power node. The video task label uses structured coding: <algorithm field> _ <function type> _ <version>, such as the vehicle identification label = "VEHICLE_RECOMMENDATION_V3". The calculation formula of the matching value P2 of the video analysis algorithm dimension is as follows:

[0095] P2=(BaseScore+VersionScore)*WeightFactor;

[0096] BaseScore=0.8×Ptag / Qtag;

[0097] VersionScore=0.2×PA;

[0098] Wherein, Ptag is the number of matched algorithm tags, Qtag is the number of requested algorithm tags in the task feature vector; PA is the algorithm version matching degree; WeightFactor is the label fusion weight.

[0099] In a feasible implementation, the number of matched algorithm tags is the number of algorithm tags supported by each computing power node and matched with the requested algorithm in the task feature vector. When the algorithm field and the function type are completely the same, the algorithm tags are matched.

[0100] In a feasible implementation, in the matching algorithm, if each video analysis algorithm is a core algorithm, the label weight of the video analysis algorithm is 1.2; if each video analysis algorithm is a normal algorithm, the label weight of the video analysis algorithm is 1.0; and if each video analysis algorithm is an auxiliary algorithm, the label weight of the video analysis algorithm is 0.8. The average value of the label weights of all matching algorithms is recorded as the label fusion weight.

[0101] In a feasible implementation, in the calculation of the algorithm version matching degree, if the main version of each video analysis algorithm matches the main version of the algorithm in the task feature vector, the algorithm version matching degree is added by 1.0, if the secondary version matches, the algorithm version matching degree is added by 0.5, and otherwise, the algorithm version matching degree is not added.

[0102] For the time-sensitive dimension, the matching value P3 is composed of a basic period matching part and a historical scene bonus part, and satisfies P3=BaseMatch+HistoryBonus.

[0103] The basic period matching part (BaseMatch) is used to measure the degree of coincidence between the task time period and the preset performance peak period of the computing power node. The video computing power scheduling system defines the performance peak period of each computing power node (for example, the peak period of the night-type node is 18:00-06:00, and the peak period of the daytime-type node is 06:00-18:00). If the task time period is completely located within the performance peak period of the node, BaseMatch=1.0, which is recorded as complete matching. If there is a deviation between the task time period and the performance peak period of the node, the score is given according to the deviation degree, which is recorded as partial matching. For example, if the period deviation is less than or equal to 2 hours, BaseMatch=0.4, and if the period deviation is greater than 2 hours, BaseMatch=0.1 (note that the specific deviation threshold and the corresponding score can be configured according to actual business requirements).

[0104] The historical scene bonus (HistoryBonus) is a dynamic reward score given based on the historical task execution performance of the computing power node. The video computing power scheduling system maintains a historical scene performance database to record the execution efficiency (such as processing time consumption and success rate) of different algorithm combinations in different periods and different task types (scenes). If the elements of the video analysis algorithm dimension (D2) and the time-sensitive dimension (D3) of the current task match the records marked as “optimal in the same scene” in the history of the computing power node, a reward score is given, for example, HistoryBonus=+0.2. This historical scene bonus is accumulative, that is, if the node performs optimally in multiple related scenes, a maximum bonus of X (such as X=0.4) can be accumulated.

[0105] The following illustrates the calculation process of the historical scene bonus:

[0106] Suppose a task is "vehicle recognition at night (02:00)". For a diurnal node, its base period matching score BaseMatch can be low (e.g. 0.1). But for a nocturnal node, its BaseMatch = 1.0. If this nocturnal node ranks first in efficiency in performing the "night vehicle recognition" task in historical data, it can also get a bonus of HistoryBonus = +0.2. Therefore, the final matching value of the computing power node in the time-sensitive dimension P3 = 1.0 + 0.2 = 1.2.

[0107] For the data source type dimension, the matching value of the data source type dimension of the computing power node is determined by the matching degree between the element of the data source type dimension in the task feature vector and the element of the data source type dimension of the computing power node. If the device types of the two are completely matched, the matching value of the data source type dimension is 1.0, if the device type of the computing power node is compatible with the device type of the task (such as aerial device matching mobile), the matching value of the data source type dimension is 0.6, otherwise the device types are not matched, and the matching value of the data source type dimension is 0.

[0108] For the security level dimension, if the security level of the computing power node is less than or equal to the permission level of the task requester, the matching value of the security level dimension of the computing power node is 1.0, otherwise it is 0.0.

[0109] Remove the computing power nodes with matching value of 0 in the security level dimension; based on the element dimension weight strategy, the matching values of the remaining computing power nodes in the five dimensions are weighted and summed respectively to obtain the matching degree of each computing power node; finally, the highest matching degree of the preset number of computing power nodes is retained and a computing power node matching table is generated.

[0110] In a feasible implementation, the calculation formula of the matching degree of each computing power node is as follows:

[0111] Score =∑ i (W i ×P i )+γ(1-C 当前负载 / C 最大负载 ),i=1,...,5

[0112] Where P i is the matching value of dimension i, W i is the weight of dimension i, ΣW i = 1, the load additional weight γ = 0.1, C 当前负载 is the current load of each computing power node, C 最大负载 is the maximum load of each computing power node. When C 当前负载 > 90%, the load additional weight γ is automatically zero.

[0113] In a feasible implementation, when the verification fails (i.e., the matching value of the security level dimension is 0.0), the resource ID (i.e., the computing power node ID) and the security level Lr and the requester's authority level Lu are pushed to the approval module; after receiving the pass instruction returned by the approval module, the resource is reactivated and included in the computing power node matching table; the approval process has an automatic pass threshold, and when (Lr-Lu)≤2, the automatic approval passes; for resources with a security level Lr≥2, a dynamic encryption channel is enabled during scheduling, and the encryption algorithm used is TLS 1.3 or the national standard SM4 algorithm.

[0114] The result optimization module is configured to perform optimization screening on the computing power nodes according to the computing power node matching table to obtain an optimal computing power node and a list of execution parameters thereof.

[0115] In the result optimization module, the resource health degree of each computing power node in the computing power node matching table and the intention fit degree of each computing power node with the task feature vector are calculated, the task execution degree of each computing power node is calculated according to the matching degree, the resource health degree and the intention fit degree of each computing power node, the task execution degrees of the remaining computing power nodes in the computing power node matching table are iteratively calculated, the computing power nodes in the computing power node matching table are arranged in descending order of the task execution degree, and thus the optimal computing power node and the list of execution parameters thereof are output.

[0116] Optionally, the resource health degree Healthscore is obtained based on real-time performance indicators of each computing power node, and the real-time performance indicators include CPU utilization, memory utilization and network latency. The calculation formula is as follows:

[0117] Healthscore = 1-(w1xCPU+w2xMem+w1xCPU+w3xLatency)

[0118] wherein w1, w2 and w3 are three weight coefficients, CPU is the normalized CPU utilization, Mem is the normalized memory utilization, and Latency is the normalized network latency.

[0119] Optionally, the intention fit degree of each computing power node with the task feature vector is the feature similarity between the historical task feature of each computing power node and the current task feature vector, and the specific value of the intention fit degree is determined according to the feature similarity and a preset similarity threshold.

[0120] In a feasible implementation, the task execution degree of each computing power node is calculated according to the matching degree, the resource health degree and the intention fit degree of each computing power node, and specifically includes:

[0121] The weighted sum of the matching degree, the resource health degree and the intention fit degree of each computing power node is taken as the task execution degree of the computing power node.

[0122] In an implementable embodiment, the optimal computing power node and its execution parameter list include a resource ID (i.e., a computing power node ID), an algorithm list, matching values of each dimension, a matching degree, and a task execution degree, and the optimal computing power node and its execution parameter list are output in the form of a work order.

[0123] In an implementable embodiment, the optimal computing power node and its execution parameter list are output in the form of a structured work order, which includes the following core fields:

[0124] Resource ID: a unique identifier of the computing power node finally selected for scheduling.

[0125] Algorithm list: specific video analysis algorithm identifiers and their execution order (if necessary) that the retrieval engine matches and specifies for the computing power node for this specific video task, which needs to be loaded and executed. The algorithm list is derived from the matching result of the algorithm dimension element (D2) in the task feature vector and the algorithm capability of the computing power node.

[0126] Matching degree: the total score obtained by weighting and summing the matching values of the above five dimensions according to the weight strategy.

[0127] Task execution degree: the final score after comprehensively matching the matching degree, the resource health degree, and the intention fit degree.

[0128] This work order will be issued to the target computing power node, and the executor in the node will load the specified algorithm package according to the "algorithm list" field in the work order and analyze and process the input video data.

[0129] The multi-dimensional retrieval engine for cross-level video computing power scheduling further includes:

[0130] As shown in Figure 2 , the feedback learning module is configured to collect execution feedback data of the entire task scheduling cycle and optimize the models in the intention analysis module and the strategy generation module according to the execution feedback data of the entire task scheduling cycle.

[0131] As shown in Figure 5 , the feedback learning module includes:

[0132] A data collection unit is configured to obtain execution feedback data of the entire task scheduling cycle, which includes but is not limited to task intention information, a task feature vector, a system state vector, an element dimension weight strategy, a computing power node matching table, an optimal computing power node and its execution parameter list, a computing power node finally executing a task, and task execution result data.

[0133] The data processing unit is configured to calculate a comprehensive utility value of each task according to the execution feedback data, filter the execution feedback data according to the comprehensive utility value of the task, that is, filter out task records with high comprehensive utility value (that is, successful execution, low delay, and high resource utilization), and generate a first training data set according to the task intention information and the task feature vector in the filtered execution feedback data, and generate a second training data set according to the system state vector of each task, the task feature vector, the comprehensive utility value, and the system state vector after the task execution in the filtered execution feedback data; the system state vector of each task and the task feature vector are combined as a current state, the comprehensive utility value is used as a reward, and the system state vector after the task execution is used as a next state.

[0134] In an implementable embodiment, the comprehensive utility value Utility of each task is calculated according to a task success state, an actual response delay, and resource utilization, and the specific formula is as follows:

[0135] Utility = w4 x I succes + w5 x (1-T delay / T max ) + w6 x R utilization

[0136] wherein I succes is a task success identifier (1 for success and 0 for failure), T delay is an actual response delay, T max is an allowed maximum delay, R utilization is resource utilization, and w4, w5, and w6 are three weight coefficients.

[0137] The first model optimization unit is configured to fine-tune the pre-trained model in the intention analysis module in a supervised learning manner to reduce analysis errors. The fine-tuning of the model enables the model to output a structure such as {D1:33***6, D2:VEHICLE_RECOMMENDATION_V3, D3:night, D4:fixed camera, D5:high} more stably and accurately when the model is input with “A city B district road junction camera last night” again.

[0138] The second model optimization unit is configured to fine-tune the reinforcement learning model in the strategy generation module according to the second training data set.

[0139] Optionally, the reinforcement learning model is an Actor-Critic framework, the model is fine-tuned by using the second training data set, and the model training process is a prior art, which is not described herein. The fine-tuned Actor network is deployed back to the strategy generation module to generate more optimal strategy fine-tuning information online.

[0140] A parameter optimization unit is configured to optimize the strategy generation module and the result optimization module. Optionally, the parameter optimization unit is configured to dynamically adjust the alpha and beta of the weight adjustment unit in the strategy knowledge base, adjust the adjustment rules of the high-load strategy adjustment template, and adjust the weight of the resource health degree in the result optimization module, that is, to adjust the preset parameters (such as weight coefficients, scale factors, threshold values, etc.) in the strategy generation module and the result optimization module. With the accumulation of task samples, the optimal values of these parameters are found.

[0141] In an embodiment, the optimization target is to maximize the correlation between the comprehensive utility value of the historical tasks and the task execution degree (score value) given by the engine. This target ensures that the score of the search engine can accurately predict the actual execution effect of the task.

[0142] In an embodiment, a Bayesian optimization algorithm is used to automatically search and determine a set of optimal hyperparameter combinations based on historical task data, so that the decision rules of the entire video computing power scheduling system can be continuously optimized and updated to adaptively improve the overall performance of the engine.

[0143] The application further provides a multi-dimensional search method for supporting intent-driven cross-level video computing power scheduling, which comprises the following steps:

[0144] First, the task intention information input by the user is parsed into a task feature vector composed of key task elements; real-time system state information is obtained and a system state vector is generated; then, an element dimension weight strategy is generated based on the task feature vector and the system state vector; then, the computing power nodes in the video computing power resources are matched based on the task feature vector and the element dimension weight strategy, and a computing power node matching table is obtained; finally, the computing power nodes are optimized and screened based on the computing power node matching table, and the optimal computing power nodes and their execution parameter list are obtained.

[0145] The execution feedback data of the entire task scheduling cycle is collected, and the model used for parsing and the model used for generating the strategy are optimized based on the execution feedback data of the entire task scheduling cycle.

[0146] The technical solutions of the application will be described in detail below with reference to specific examples.

[0147] A user needs to handle an urgent task, and inputs the following text in the system interface: "Immediately analyze all the videos recorded by the cameras at all intersections in B district of A city from 8:00 to 10:00 last night, and find the vehicles involved using the best license plate recognition algorithm, and pay attention to confidentiality!"

[0148] After the intention analysis module receives the above natural language instruction, an NLP model is called to perform entity and relationship recognition.

[0149] "Hangzhou Xihu District" corresponds to the administrative division dimension (D1): mapped to code 330106.

[0150] "License plate recognition algorithm" corresponds to the video analysis algorithm dimension (D2): mapped to the core algorithm label VEHICLE_PLATE_RECOGNITION, and automatically associated with its latest version.

[0151] "8pm to 10pm last night" corresponds to the time-sensitive dimension (D3): mapped to 2023-10-26 20:00:00 ~ 2023-10-26 22:00:00, and identified as a night period.

[0152] "Intersection camera" corresponds to the data source type dimension (D4): mapped to a fixed camera.

[0153] "Pay attention to confidentiality" corresponds to the security level dimension (D5): mapped to Lr≤1.

[0154] "Immediately", "case involved" correspond to the identification of emergency task signs.

[0155] The NLP model parses and outputs the task feature vector.

[0156] The system state monitoring unit obtains the current load rate of 75% (high), and the video analysis algorithm node change rate ΔN in the resource pool is 15%.

[0157] The strategy generation module receives the task feature vector, identifies the "immediately" and "case involved" emergency semantics, and forcibly calls the emergency task weight template from the strategy knowledge base: [W1, W2, W3, W4, W5] = [0.6, 0.2, 0.1, 0.0, 0.1] as the basic strategy. The reinforcement learning strategy unit outputs the fine-tuning suggestion based on the current state (load = 75%, ΔN = 15%) and task features (algorithm = license plate recognition, security level = high) on the basis of the template: further increase the D1 (administrative division) weight to 0.65 (due to high load, prioritize localization), and compensate the D2 (algorithm) weight W2 = min(0.8, 0.2(1+15 / 20)) = 0.35. Finally, the element dimension weight strategy is generated: [0.54, 0.29, 0.08, 0.0, 0.08], the value of the weight in this invention is set to two decimal places, so the sum of the weights of each dimension is about 0.99, and the theoretical value of the sum of the weights of each dimension is 1. Because the weight is used for weighted calculation, this small deviation is acceptable in engineering implementation and does not affect the relative importance of each dimension weight. At the same time of generating the strategy, the encrypted channel instruction is preset.

[0158] In the parallel matching module, the algorithm nodes in the video algorithm resource are matched according to the task feature vector and the element dimension weight strategy, and the algorithm node matching table is obtained;

[0159] Assuming that the computing power node matching table contains 3 candidate resources A, B and C. The resource health degree and intention matching degree of the three candidate resources A, B and C are calculated in the result optimization module, as follows:

[0160] Candidate resource A: 90% of the historical tasks are traffic security analysis, which is highly matched with the current "case vehicle investigation" intention, and the matching degree compensation score is +0.12; the node health degree is 0.95 (full score 1).

[0161] Candidate resource B: 70% of the historical tasks are traffic flow monitoring, which has certain relevance with the "case vehicle investigation" intention, and the matching degree compensation score is +0.08; the node health degree is 0.90.

[0162] Candidate resource C: the historical tasks are mostly general object recognition, and the matching degree compensation score is +0.03; the health degree score is 0.88.

[0163] On the basis of high matching degree, the intention matching degree and health degree of candidate resource A are also extremely high, and the task execution degree is far superior to other resources, so the task is assigned to candidate resource A, and the adaptive executor monitors that the container of resource A is successfully started within 10 seconds, and the task execution is successful.

[0164] The data acquisition unit of the feedback learning module acquires the full link data of this task. Record this task: intention text, analysis result, strategy, scoring details, execution success and low delay. The data processing unit calculates the comprehensive utility value Utility of this scheduling. Because the task is successfully executed and the delay is low, the Utility is 0.92 (full score 1.0). Start two model optimization units, this task full link data as positive sample, used to strengthen the association of NLP model to "case", "confidential" and other words and security level, urgency. The comprehensive utility value Utility is used as a reward signal, which is input into the reinforcement learning model of the strategy generation module, and the network parameters of the reinforcement learning model are updated through the policy gradient algorithm to strengthen the decision tendency of the strategy (improve the weights of D1 and D2) taken by the reinforcement learning model in the "high load + urgent task" scene. The task data is added to the data set for the next round of Bayesian optimization to further fine-tune the hyperparameters in the weight adjustment formula.

[0165] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A multi-dimensional retrieval engine supporting intent-driven cross-level video algorithmic power scheduling, characterized in that, The method comprises the following steps: An intention analysis module is configured to analyze the obtained task intention information of the user input into a task feature vector composed of key task elements; The key task elements include elements of administrative division dimension, video analysis algorithm dimension, time sensitivity dimension, data source type dimension, and security level dimension; A system state monitoring unit is configured to obtain real-time system state information and generate a system state vector; A strategy generation module is configured to generate an element dimension weight strategy according to the task feature vector and the system state vector; The strategy generation module comprises: A strategy knowledge base is configured to store a strategy base template and a strategy adjustment template; A first strategy generation unit is configured to select a template from the strategy knowledge base according to the task feature vector and the system state vector and generate an element dimension weight strategy; In the first strategy generation unit, if the task feature vector is identified to contain a key task element related to an emergency semantic feature or the task queue length in the system state vector exceeds a preset queue length, the element dimension weight strategy is generated according to an emergency task strategy adjustment template; If the system load in the system state vector is identified to be higher than a preset load threshold, the element dimension weight strategy is generated according to a high-load strategy adjustment template; If the task feature vector is identified to contain a key task element related to a security feature, the element dimension weight strategy is generated according to a high-security strategy adjustment template; If two or more of the emergency task strategy adjustment template, the high-load strategy adjustment template, and the high-security strategy adjustment template are simultaneously called, the corresponding template is selected according to a preset priority order, and the corresponding element dimension weight strategy is outputted; Otherwise, the element dimension weight strategy is generated according to the strategy base template; A parallel matching module is configured to match the computing power nodes in the video computing power resources according to the task feature vector and the element dimension weight strategy, and obtain a computing power node matching table; A result optimization module is configured to optimize and screen the computing power nodes according to the computing power node matching table, and obtain the optimal computing power nodes and their execution parameter list; In the result optimization module, the resource health degree of each computing power node in the computing power node matching table and the intention fit degree of each computing power node with the task feature vector are calculated, the task execution degree of each computing power node is calculated according to the matching degree, the resource health degree, and the intention fit degree of each computing power node, the task execution degrees of the remaining computing power nodes in the computing power node matching table are iteratively calculated, the computing power nodes in the computing power node matching table are arranged in descending order of the task execution degree, and the optimal computing power nodes and their execution parameter list are outputted. The matching degree is the total score obtained by weighting and summing the matching values of the administrative division dimension, the video analysis algorithm dimension, the time sensitivity dimension, the data source type dimension, and the security level dimension according to the weight strategy; The resource health degree is obtained based on the real-time performance indicators of each computing power node; The intention fit degree of each computing power node with the task feature vector is the feature similarity between the historical task features of each computing power node and the current task feature vector, and the specific value of the intention fit degree is determined according to the feature similarity and a preset similarity threshold.

2. The intent-driven cross-level video computing power scheduling multi-dimensional retrieval engine according to claim 1, wherein, The strategy generation module further comprises: A policy fine-tuning unit is configured to generate policy fine-tuning information according to the task feature vector and the system state vector; A second policy generation unit is configured to adjust the element dimension weight strategy generated by the first policy generation unit in combination with the policy fine-tuning information to obtain a final element dimension weight strategy.

3. The intent-driven cross-level video computing power scheduling multi-dimensional retrieval engine according to claim 2, wherein, The policy fine-tuning unit stores a pre-trained reinforcement learning model, the state of the reinforcement learning model is the system state vector and the task feature vector, the action is the policy fine-tuning information, and the reward function R during training is constructed according to the task execution success rate and the task response delay.

4. The intent-driven cross-level video computing power scheduling multi-dimensional retrieval engine according to claim 1, wherein, The cross-level video computing power scheduling multi-dimensional retrieval engine further includes: A feedback learning module is configured to collect execution feedback data of the entire task scheduling cycle and optimize the intention analysis module and the policy generation module according to the execution feedback data of the entire task scheduling cycle.

5. A video computing power scheduling system, characterized in that, The video computing power scheduling system includes the cross-level video computing power scheduling multi-dimensional retrieval engine supporting intention driving.

6. An intent-driven cross-level video computing power scheduling multi-dimensional retrieval method, characterized in that, The method is based on the cross-level video computing power scheduling multi-dimensional retrieval engine of claim 1 and includes the following steps: First, the task intention information input by the user is parsed into a task feature vector composed of key task elements; real-time system state information is obtained to generate a system state vector; then, an element dimension weight strategy is generated according to the task feature vector and the system state vector; then, the computing power nodes in the video computing power resources are matched according to the task feature vector and the element dimension weight strategy to obtain a computing power node matching table; finally, the computing power nodes are optimized and screened according to the computing power node matching table to obtain the optimal computing power nodes and their execution parameter list.

Images