Information scheduling scenarized service agent system

Abstract

The invention relates to the technical field of information management, in particular to an information scheduling scenarized service agent system. The system comprises an information scheduling acquisition and analysis module, a scene classification and service matching module, a service scheme instruction generation module and a service intelligent execution module, user behavior data, environment perception data and business system data corresponding to information scheduling can be obtained, and an original information set is generated; the original information set is preprocessed, an information scheduling scene classifier is constructed, the service matching degree is calculated in combination with the real-time service resource state, and a candidate service scheme set is generated; performing optimization sorting on the candidate service scheme set to generate an optimal service scheme, and disassembling the optimal service scheme into an executable service instruction sequence; receiving a service instruction sequence and generating a service evaluation report; and if the display does not reach the preset service quality threshold value, triggering a dynamic adjustment mechanism to realize self-iteration upgrading of the information scheduling scenarized service. According to the invention, intelligent scheduling and service of information in a complex scene can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information management, and in particular to an information scheduling scenario service intelligent agent system. BACKGROUND

[0002] With the continuous development of information technology and the widespread application of intelligentization, information scheduling is increasingly applied in various fields, especially in complex scenarios such as industrial production, transportation, logistics distribution, and urban management. The efficiency and accuracy of information scheduling have become important factors in improving work efficiency and optimizing resource allocation. Existing information scheduling systems mostly rely on rule presetting and manual intervention, especially in complex and variable actual application scenarios, and often fail to fully consider real-time changes in the environment, tasks, and resources. For example, in the logistics industry, although the existing scheduling system can plan the transportation route, the system responds slowly to real-time traffic conditions, unexpected events, and other factors, resulting in low scheduling efficiency. In addition, in multi-task, multi-role collaborative scenarios, traditional systems often lack flexibility and cannot effectively coordinate resource competition and task allocation among multiple participants, thereby affecting the overall scheduling effect and service quality. SUMMARY

[0003] Therefore, it is necessary to provide an information scheduling scenario service intelligent agent system to solve at least one of the above technical problems.

[0004] To achieve the above purpose, an information scheduling scenario service intelligent agent system includes the following modules: An information scheduling collection and analysis module is configured to obtain user behavior data, environment perception data, and business system data corresponding to information scheduling, generate an original information set with data source identification, preprocess the original information set, and extract key features using feature engineering to generate a structured information feature library containing user portrait feature vectors, environment feature matrices, and business feature parameter sets. A scenario classification and service matching module is configured to construct an information scheduling scenario classifier based on the structured information feature library to output the type label, confidence, and key trigger factor of the current scenario, input the information scheduling scenario recognition result into a service decision engine to call a preset scenario-service mapping rule library, and calculate the service matching degree in combination with the real-time business resource state to generate a candidate service scheme set, wherein each service scheme contains a service type, execution steps, resource requirements, and expected effects. A service scheme instruction generation module is configured to optimize and sort the candidate service scheme set, evaluate the response speed, resource consumption, and user satisfaction of each service scheme using a multi-objective decision algorithm, generate an optimal service scheme, and decompose the optimal service scheme into an executable service instruction sequence through a rule engine. The service intelligent execution module is configured to receive a service instruction sequence, call corresponding business interfaces and resource components to execute service operations, and collect state data in real time during service execution, including progress percentage, resource occupancy rate, abnormality analysis report, and generate a service execution state stream; an effect evaluation model is constructed based on the service execution state stream to generate a service evaluation report containing performance indicators, deviation analysis table, grade evaluation and improvement suggestions; if the service evaluation report shows that the preset service quality threshold is not reached, a dynamic adjustment mechanism is triggered to update the scene-service mapping rule library, thereby realizing self-iterative upgrading of information scheduling scene-based services.

[0005] Further, the generation process of the original information set includes: By configuring a data access interface list at system initialization, including RESTful API interface, message queue interface, database direct connection interface and Internet of Things device communication interface, and configuring a corresponding preset data parsing template for each interface; Triggering data collection processes of each interface at a set period, and responding to data update notifications in real time through each interface using an event-driven mode to generate a data collection task list; Based on the data collection task list, the corresponding data parsing template is called according to the interface type, and the original data corresponding to the collected user behavior data, environmental perception data and business system data is converted in format, wherein the unstructured data is converted into JSON format, and the binary data is converted into Base64 encoded string, to generate standardized intermediate data; The standardized intermediate data is subjected to integrity check and legality verification, to verify data transmission consistency by a hash algorithm, and the data passing the verification is added with a timestamp and signature information, and is merged into the original information set with data source identification.

[0006] Further, the construction of the structured information feature library includes: User behavior data is extracted from the original information set to analyze its operation path and interaction frequency by a sequence pattern mining algorithm, and to calculate user preference weight, to generate a user portrait feature vector containing basic attributes, behavior labels and interest dimensions; The environmental perception data is subjected to spatio-temporal feature extraction to convert geographic location information into regional grid encoding, time stamp into time period type and holiday identification, and network state into connection quality level, to generate an environmental feature matrix; The service resource list in the business system data is parsed to extract resource type, capacity upper limit, available quantity and associated dependency relationship, and the task queue is prioritized and time-limited, to generate a business feature parameter set, including resource feature vector, task feature matrix and constraint condition set; By establishing a feature association index, the user portrait feature vector, the environment feature matrix and the service feature parameter set are stored in association according to the time dimension and the scene dimension, and a structured information feature library supporting fast query and feature combination is constructed.

[0007] Further, the construction of the information scheduling scene classifier includes: extracting historical scene sample data from the structured information feature library, wherein each historical scene sample contains a user feature vector, an environment feature slice, a service demand parameter and a corresponding artificial labeled scene type, and constructing a scene training data set and a scene verification data set; A bidirectional long short-term memory network is used to construct a sequence feature extractor, and the corresponding user behavior time sequence and environment parameter change sequence in the structured information feature library are encoded based on the sequence feature extractor to generate a time sequence feature vector; An attention mechanism module is constructed to calculate the attention weight of each dimension in the time sequence feature vector, highlight the key features with high contribution to scene recognition, and generate a weighted fusion feature; The weighted fusion feature is input into a deep neural network classifier to output a scene type probability distribution through a softmax function, and a cross-entropy loss function is used to train and verify the model on the scene training data set and the scene verification data set. When the accuracy of the scene verification data set reaches a preset threshold, the training is stopped, the model parameters are saved, and a trained information scheduling scene classifier is generated; In real-time operation, the newly collected structured information feature data is input into the trained information scheduling scene classifier to output the type label, confidence and key trigger factor corresponding to the information scheduling scene recognition result of the current scene, wherein the key trigger factor is the top N feature values that contribute most to the scene classification.

[0008] Further, the generation process of the candidate service scheme set includes: By sorting the typical scene types and corresponding service cases, the mapping relationship between scene feature parameters and service attributes is extracted to form initial rule entries, wherein each rule entry contains scene conditions, service types, priority weights and execution constraints; The initial rule entries are encoded into a machine recognizable rule language using a production rule representation method, which includes a rule antecedent and a rule consequent. The rule antecedent is a scene feature combination, and the rule consequent is a service operation instruction. A rule library infrastructure is constructed based on the rule language; The historical service execution data is acquired, and the rule base infrastructure is optimized based on the historical service execution data to discover potential scenario-service association patterns by an association rule mining algorithm, and new rule entries are supplemented based on the scenario-service association patterns, while contradictory rules are eliminated by a rule conflict detection algorithm, to generate a preset scenario-service mapping rule base; When the preset scenario-service mapping rule base is invoked by a service decision engine, rule matching is performed on the information scheduling scenario recognition result, a Rete algorithm is used to improve matching efficiency, and the service matching degrees corresponding to the rules that pass the matching are calculated in combination with real-time business resource states by sorting the matched rules according to priorities; The service schemes corresponding to the top M rules in the service matching degrees are selected as a candidate service scheme set, and the value of M is dynamically adjusted according to the business complexity corresponding to the rules.

[0009] Further, the operation of the multi-target decision algorithm includes: A target function evaluated for each service scheme is determined, including a response speed target, a resource consumption target, and a user satisfaction target, wherein the response speed target is to minimize service start delay, the resource consumption target is to minimize CPU / memory occupancy, and the user satisfaction target is to maximize feedback score; Historical service evaluation data is acquired, and a weight coefficient is set for each target function, wherein the weight coefficient is determined based on the historical service evaluation data by an analytic hierarchy process to generate a target weight vector; For each service scheme in the candidate service scheme set, an evaluation index value is extracted and standardized to be converted into a normalized score in the interval [0, 1]; A weighted comprehensive score corresponding to each service scheme is calculated based on the target weight vector and the normalized score, and each service scheme is sorted in descending order of the weighted comprehensive score; The service scheme with the highest weighted comprehensive score is selected as the optimal service scheme, and if there are service schemes with the same score, a random forest algorithm is used to predict the potential risks of each service scheme, and the service scheme with the lowest risk value is selected.

[0010] Further, the operation of the service execution process includes: A service instruction sequence corresponding to the optimal service scheme is received, and the operation type, parameter list, and dependency relationship in the service instruction sequence are parsed to generate a service execution flowchart; A task scheduling queue is constructed based on the service execution flowchart to determine the task execution order by using a topological sorting algorithm, and a unique identifier and a timeout threshold are assigned to each task; The task resource demand and real-time resource state corresponding to each task are obtained by calling the corresponding business interface and resource component, and the dynamic programming algorithm is used based on the task resource demand and real-time resource state to dynamically allocate hardware resources including CPU, memory, network bandwidth and software resources including business interface and data connection, so as to generate a service execution task resource allocation result; The corresponding service component is called according to the task scheduling queue based on the service execution task resource allocation result, so as to execute the service operation through the remote procedure call protocol, and the start time, end time, return result and exception information of each task are recorded in real time, and the service execution state flow corresponding to the progress percentage, resource occupation rate and abnormal analysis report is obtained based on the start time, end time, return result and exception information of each task. If an exception occurs during service execution, the corresponding fault tolerance mechanism is triggered according to the exception type, including retry mechanism, degradation mechanism or fuse mechanism, to ensure the continuity of the service execution process.

[0011] Further, the analysis process of the service execution state flow includes: The state collection probe is embedded in the service execution process, and the task execution state data including process ID, resource occupation peak value and response time distribution is collected at a preset sampling frequency to generate a state snapshot with a timestamp; The state snapshot is aggregated according to service instance and time dimension, and the average execution time, success rate and abnormal occurrence rate of each task are calculated to generate service task state statistical values; The progress percentage and resource occupation rate corresponding to each task are estimated and calculated according to the average execution time, success rate and abnormal occurrence rate of each task; A state abnormality detection model is constructed to set the control limit of the state parameter by using the statistical process control method, and when the service task state statistical value exceeds the control limit, it is marked as an abnormal state, and the abnormal occurrence time, duration and associated task are recorded; The root cause analysis is performed on the abnormal state based on the abnormal occurrence time, duration and associated task to trace the direct and indirect causes of the abnormality through the fault tree analysis method, and an abnormal analysis report including the abnormal type, influence range and processing suggestion is generated; The progress percentage, resource occupation rate and abnormal analysis report are integrated to generate the corresponding service execution state flow.

[0012] Further, the construction of the effect evaluation model includes: Extracting various performance indicators KPIs from the service execution state flow, including service completion rate, average response time and resource utilization probability, wherein the service completion rate is the ratio of success rate to the sum of success rate and abnormality occurrence rate, the average response time is the ratio between average execution time and duration, and the resource utilization probability is the product of progress percentage and resource occupancy rate; Setting threshold ranges of various performance indicators KPIs, including ideal value, acceptable value and minimum value, and determining as unqualified when various performance indicators KPIs are lower than the minimum value; Calculating deviation rates of actual values of various performance indicators KPIs from ideal values, and generating deviation analysis table; Using fuzzy comprehensive evaluation method to comprehensively evaluate various performance indicators KPIs, calculating comprehensive score of service quality, and generating service evaluation report, which contains specific values of various performance indicators KPIs, deviation analysis table, grade evaluation and improvement suggestions.

[0013] Further, the updating process of the dynamic adjustment mechanism includes: When the service evaluation report shows that the preset service quality threshold is not reached, the parameter adjustment process is triggered, and the most influential performance indicator KPI is extracted from the deviation analysis table; For the case that the performance indicator KPI does not meet the standard, the corresponding scene recognition link and service decision link are analyzed, and the rule parameters or rule factors that need to be adjusted are determined; Based on the rule parameters or rule factors that need to be adjusted, the scene-service mapping rule library is incrementally updated, so as to use the latest service execution data and evaluation results as new samples, and based on the improvement suggestions, the scene-service mapping rule library is updated to increase new rule entries or adjust the priority weight of existing rule entries, and the effectiveness of the new rule entries is verified through simulation test; Recording the changes of service evaluation indicators before and after parameter adjustment, calculating the optimization effect improvement rate, and if the optimization effect improvement rate does not meet the expectation, repeating the parameter adjustment process until the service quality meets the standard, so as to realize self-iterative upgrading of information scheduling scenario-based service.

[0014] The beneficial effects of the present application are: The information scheduling scene service intelligent agent system provided by the application is composed of an information scheduling collection and analysis module, a scene classification and service matching module, a service scheme instruction generation module and a service intelligent execution module, compared with the prior art, the beneficial effects of the application are that through comprehensive collection of user behavior data, environment perception data and business system data, an original information set with data source identification is formed. In this process, the collected information will provide a sufficient basis for subsequent decision-making, the original information set includes user behavior records, real-time change data of the environment and key parameters in the current business process, preprocessing these original data is the key to realize data cleaning and integration, this process can help to remove noise, fill in missing values and standardize data format. In addition, by extracting key features through feature engineering, a structured information feature library can be constructed, which can effectively represent user portraits, environmental features and business features, ensuring the real-time and accuracy of business decision-making, thereby laying a solid data foundation for subsequent scene recognition and service scheduling. Secondly, by analyzing the obtained structured information feature library, an information scheduling scene classifier is constructed to output the type label, confidence and key trigger factor corresponding to the scheduling scene recognition result of the current scene, the construction of the classifier depends on machine learning algorithms or deep learning models, through training on a large amount of historical data, the characteristics of different business scenes can be accurately recognized and mapped to specific scene types, the effectiveness of the scene classifier directly affects the accuracy and efficiency of subsequent service decision-making, through real-time input of feature vectors, the classifier can determine the most likely scene type according to the current business state and environmental information, this process not only determines the scene category, but also provides confidence evaluation of the current scene, making the service decision more accurate, so that accurate service matching and resource scheduling can be carried out based on actual business needs, thereby better improving the scheduling efficiency. Then, based on the output scene recognition result, focus on the optimization sorting of the candidate service scheme set, through the use of multi-objective decision algorithm, the multiple dimensions (such as response speed, resource consumption, user satisfaction, etc.) of each service scheme are comprehensively evaluated, and the optimal service scheme is finally generated, in the selection of service scheme, response speed directly affects user experience, resource consumption reflects the efficiency of business execution, and user satisfaction is the final standard of service success, through the multi-objective decision algorithm, these factors can be balanced to ensure that the selected service scheme not only meets the user's demand, but also achieves the optimal state in resource utilization and service quality, at the same time, the optimal scheme is decomposed into a series of executable service instructions through the rule engine, these instructions will be sent to the actual execution layer to ensure that the service can be efficiently implemented and achieve the expected effect, which can effectively coordinate the resource competition and task allocation among multiple participants.Finally, through the real-time collection of various data in the execution process (such as progress percentage, resource occupancy rate, abnormal analysis report, etc.), the execution of the service can be dynamically understood, potential problems can be found in time and countermeasures can be taken. Through the construction of an effect evaluation model, the system can generate a detailed service evaluation report after the service is executed, which contains performance indicators, deviation analysis table, service quality level assessment and improvement suggestions. Such an evaluation report not only helps managers evaluate the effect of the service, but also provides strong data support for subsequent improvement. When the service evaluation report shows that the preset quality standard is not reached, the dynamic adjustment mechanism will be triggered to automatically update the scene-service mapping rule library for self-iteration and upgrading. Through this adaptive mechanism, the service quality can be continuously improved, maintaining high flexibility and responsiveness, ensuring that changes in user needs can be adapted in time, and effectively improving the overall scheduling effect and service quality. BRIEF DESCRIPTION OF DRAWINGS

[0015] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments thereof, read in conjunction with the accompanying drawings: Fig. 1 A module schematic diagram of the information scheduling scenario-based service intelligent agent system of the present application; Fig. 2 A step flow schematic diagram of the generation process of the original information set of the present application. DETAILED DESCRIPTION

[0016] The technical system of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application.

[0017] To achieve the above-mentioned purpose, please refer to Figs. 1-2 The present application provides an information scheduling scenario-based service intelligent agent system. In the embodiments of the present application, please refer to Fig. 1 the module schematic diagram of the information scheduling scenario-based service intelligent agent system of the present application. In the present example, the information scheduling scenario-based service intelligent agent system comprises the following modules: S1: an information scheduling collection and analysis module, used for acquiring user behavior data, environment perception data and business system data corresponding to information scheduling, generating an original information set with data source identification; pre-processing the original information set, and extracting key features by using feature engineering to generate a structured information feature library, which contains a user portrait feature vector, an environment feature matrix and a business feature parameter set; In the embodiment of the application, the user click position (accuracy 0.01 m), the stay duration (accuracy to seconds), the operation sequence (sorted by timestamp) are collected by a user behavior tracking tool (sampling frequency 10 Hz), the environmental sensing data are collected by an environmental sensor (temperature range -10~50℃, accuracy 0.1℃; humidity 0~100%, accuracy 1%), and the business system data are extracted by a business database (order table contains 30 fields, 500 new records are added daily). A source identifier (user behavior data marked "UB", environmental data marked "EN", and business data marked "BS") is added to each data to generate an original information set containing 1000 user behavior records, 500 environmental data, and 800 business data, each data containing 20 fields, and the timestamp is accurate to milliseconds. The original information set is preprocessed, the user behavior data is standardized by Z-score (mean 0, standard deviation 1), the environmental data is denoised by Kalman filtering (process noise variance 0.01), and the business data is converted by field type (date is converted to yyyyMMdd integer). Key features are extracted by feature engineering, the user portrait feature vector contains 20 dimensions (consumption frequency, preference weight, etc., range 0~1), the environmental feature matrix is 10 rows and 8 columns (each row corresponds to 1 hour, and the columns contain temperature, humidity, etc.), and the business feature parameter set contains resource capacity, task priority, etc. 15 parameters. The structured information feature library is stored in column mode, a single record occupies 128 bytes, the query response time is <1 second, and the features cover all information scheduling scenarios.

[0018] S2: a scene classification and service matching module, configured to construct an information scheduling scene classifier based on the structured information feature library, to output a type label, a confidence degree and an information scheduling scene recognition result corresponding to a key trigger factor of a current scene; input the information scheduling scene recognition result into a service decision engine to call a preset scene-service mapping rule library, and combine a real-time business resource state to calculate a service matching degree, to generate a candidate service scheme set, wherein each service scheme contains a service type, an execution step, a resource requirement and an expected effect; In the embodiment of the application, the information scheduling scene classifier is constructed based on the structured information feature library, the classifier contains an input layer (32-dimensional features), two hidden layers (64 neurons per layer, ReLU activation), and an output layer (10 scene types, softmax activation). The training data is 5000 samples (500 samples per scene), and the training is stopped when the verification set accuracy reaches 92%. The input current data (user consumption frequency 0.8, environment temperature 25℃, business order volume 100) outputs the type label "shopping scene", confidence 0.92, and key trigger factor (consumption frequency 0.8, weight 0.35). The recognition result is input into the service decision engine, the scene-service mapping rule library (containing 100 rules, grouped by scene) is called, the "shopping scene" rule group contains 15 rules. The service matching degree is calculated in combination with the real-time business resource state (CPU available 8 cores, memory available 16 GB), and the formula is rule priority x resource availability rate, for example, the "goods pushing" rule priority is 0.8, the resource availability rate is 0.9, and the matching degree is 0.72. The candidate service scheme set (5 schemes) is generated, scheme 1: service type "accurate pushing", execution steps are "filtering goods → generating list → pushing", resource demand CPU 1 core, memory 2 GB, expected click rate 15%; scheme 2: "time-limited discount", resource demand CPU 0.5 core, expected click rate 12%, the schemes are arranged in descending order of matching degree to ensure coverage of different service types.

[0019] S3: a service scheme instruction generation module, configured to optimize and sort the candidate service scheme set, to evaluate the response speed, resource consumption and user satisfaction of each service scheme by using a multi-objective decision algorithm, to generate an optimal service scheme, and to decompose the optimal service scheme into an executable service instruction sequence by using a rule engine; In the embodiment of the application, by optimizing the sorting of the candidate service scheme set, a multi-objective decision algorithm (weighted summation method) is used to evaluate each scheme, the response speed weight is 0.4, the resource consumption is 0.3, and the user satisfaction is 0.3. The response speed of scheme 1 is 0.8 seconds (standardized score 1.0), the resource consumption is 0.2 (score 0.8), the user satisfaction is 4.5 points (score 0.9), the weighted score = 1.0x0.4+0.8x0.3+0.9x0.3=0.4+0.24+0.27=0.91; the response speed of scheme 2 is 1.2 seconds (score 0.6), the resource consumption is 0.1 (score 1.0), the user satisfaction is 4.2 points (score 0.7), the score is 0.6x0.4+1.0x0.3+0.7x0.3=0.24+0.3+0.21=0.75. Scheme 1 has the highest score and is the optimal service scheme, containing the service type "precise push", performing 3 steps, requiring resources CPU 1 core / memory 2 GB, and the expected click rate is 15%. Through the rule engine, it is decomposed into a service instruction sequence: instruction 1 "resource locking (CPU 1 core, memory 2 GB, timeout 5 seconds)", instruction 2 "product screening (category = electronic products, quantity 10, timeout 10 seconds)", instruction 3 "push execution (channel = APP, quantity 3, timeout 8 seconds)", and the instructions are sorted according to the dependency relationship, and the next instruction is executed only after the previous instruction is completed, to ensure that the steps are conflict-free.

[0020] S4: a service intelligent execution module, configured to receive a service instruction sequence, to call corresponding business interfaces and resource components to execute service operations, and to collect state data in a service execution process in real time, including a progress percentage, a resource occupation rate, and an exception analysis report, to generate a service execution state stream; to construct an effect evaluation model based on the service execution state stream, to generate a service evaluation report containing a performance indicator, a deviation analysis table, a grade evaluation, and improvement suggestions; and to trigger a dynamic adjustment mechanism to update a scene-service mapping rule library if the service evaluation report shows that a preset service quality threshold is not reached, so as to realize self-iterative upgrading of information scheduling scene-based services.

[0021] In this embodiment of the invention, by receiving a sequence of service instructions, the resource management interface is invoked (response time < 0.1 seconds) to lock 1 CPU core and 2GB of memory, the product database interface is invoked (query time < 1 second) to filter 10 electronic products, and the push interface is invoked (processing 5 messages per second) to execute the push. Real-time status data is collected: progress percentage (instruction 1 completed 33%, instruction 2 completed 67%, instruction 3 completed 100%), resource utilization (CPU 15%, memory 12%), and an anomaly analysis report (instruction 2 query delay 0.5 seconds, cause: index failure). A service execution status stream is generated, updated every 1 second, containing 20 fields, and stored in JSON format. Based on the status stream, an effectiveness evaluation model is constructed, calculating performance indicators (response speed 1.2 seconds, resource consumption 13.5%, user satisfaction 4.6 points), a deviation analysis table (response speed exceeds ideal value by 0.2 seconds, deviation rate 25%), a rating of "Good", and an improvement suggestion of "Optimize database index". The service evaluation report has a comprehensive score of 85 points, which is higher than the threshold of 80 points, and no adjustment is triggered. If the score is 78, the dynamic adjustment mechanism is triggered, the scenario-service mapping rule base is updated (a new "shopping scenario + index optimization" rule is added, with a priority of 0.85), the process is repeated, and after 2 rounds of adjustment, the score is 88, realizing self-iterative upgrade. The adjustment process is recorded in the log (including parameter changes and effect comparison).

[0022] Furthermore, as an embodiment of the present invention, reference is made to... Fig. 2 The diagram shown illustrates the steps of generating the original information set according to the present invention. In this embodiment, the generation process of the original information set includes: S101: Configure a list of data access interfaces during system initialization, including RESTful API interfaces, message queue interfaces, database direct connection interfaces, and IoT device communication interfaces, and configure a corresponding preset data parsing template for each interface; In the embodiment of the application, by configuring the data access interface list at system initialization, the RESTful API interface is set to HTTP / HTTPS protocol, port 80 / 443, supports GET / POST method, the request header contains Content-Type:application / json, the response timeout is 5 seconds; the message queue interface uses AMQP protocol, port 5672, the queue length limit is 10000, the message survival time is 24 hours; the database direct connection interface uses JDBC driver, the connection timeout is 30 seconds, the maximum number of connections is 50, the query result set limit is 1000 rows; the Internet of Things device communication interface uses MQTT protocol, port 1883, the client ID is a unique identifier, the heartbeat interval is 60 seconds. Configure the corresponding preset data parsing template for each interface, the RESTful API template defines the field mapping rule (user ID corresponds to user_id, integer type, length 10 bits), the message queue template sets the JSON field parsing order (timestamp is prior to content), the database template specifies the field type conversion (DATE type is converted to yyyy-MM-dd string), the Internet of Things template defines the binary data parsing format (the first 2 bytes are device number, the last 4 bytes are temperature value, precision 0.1℃). The template parameters are fixed and have no modification authority, which ensures that the interface and the template correspond one by one, and the parsing error is zero.

[0023] S102: Trigger the data collection process of each interface according to the set period, and generate a data collection task list by using an event-driven mode to respond to data update notifications in real time through each interface; In the embodiment of the application, by triggering the data collection process of each interface according to the set period, the RESTful API interface performs GET request every 10 minutes to obtain user login records; the message queue interface scans the queue every 5 minutes to extract unconsumed environmental monitoring data; the database direct connection interface performs SELECT query every hour to read business order tables; the Internet of Things device communication interface listens to topic messages in real time to receive sensor data. At the same time, the event-driven mode is used to respond to data update notifications, when the user completes the payment operation, the RESTful API interface is triggered to immediately perform POST request to obtain transaction details; when the environmental temperature and humidity exceed the threshold (temperature>30℃ or humidity>80%), the message queue interface immediately pulls alarm information. A data collection task list is generated after each collection or response, each task contains interface type (RESTful API / message queue, etc.), trigger mode (period / event), data identifier (unique UUID), collection timestamp (accurate to milliseconds), task state marker is “to be processed”, the number of tasks generated in a single hour does not exceed 500, and the task list is arranged in ascending order of timestamp.

[0024] S103: According to the interface type, the corresponding data parsing template is called based on the data collection task list, and the original data corresponding to the collected user behavior data, environment perception data and business system data is converted in format, wherein the unstructured data is converted into JSON format, the binary data is converted into a Base64 encoded string, and standardized intermediate data is generated; In the embodiment of the application, the original data collected is converted in format by calling the corresponding data parsing template based on the data collection task list according to the interface type. The unstructured data in the user behavior data (such as user comment text) is converted in JSON format, including the fields comment_id (string, length 32), content (text, maximum 1000 characters), and create_time (timestamp). The binary data in the environment perception data (such as the temperature and humidity byte stream collected by the sensor) is converted into a Base64 encoded string, with each 16 bytes corresponding to 24 characters, and the length of the encoded string being fixed at 1.33 times the length of the original data. The table data in the business system data is converted into a JSON array, with each object corresponding to a row of records, the field name being consistent with the database column name, and the numerical value type being kept to two decimal places. In the conversion process, the text content is removed from special characters (only letters, numbers and Chinese are kept), the time format is unified as yyyy-MM-ddHH:mm:ss, the numerical value range is limited (temperature -40 to 80℃, humidity 0 to 100%), the standardized intermediate data is generated, each piece of data is less than 10KB in size, the format error data is directly discarded, and the conversion success rate is 100%.

[0025] S104: The integrity check and the legality verification are performed on the standardized intermediate data, the data transmission consistency is verified by a hash algorithm, the time stamp and the signature information are added to the data passing the verification, and the original information set with data source identification is generated.

[0026] In the embodiment of the application, the integrity check and the legality verification are performed on the standardized intermediate data. The integrity check is performed by calculating a data digest by using a SHA-256 hash algorithm and comparing the hash value returned by the interface. If the two are completely consistent, it is determined that the transmission is complete. The legality verification checks the field format (the user ID is 10 digits, the mobile phone number is 11 digits, and the email contains an @ symbol), the numerical range (the age is 1 to 120 years old, and the order amount is greater than or equal to 0 yuan), and the mandatory items (the user name and the order number cannot be empty). If all of them are correct, it is determined that the data is legal. The data that passes the check is added with a timestamp (consistent with the collection timestamp) and signature information (using the RSA algorithm, the private key is used to encrypt the digest, and the public key can be used to decrypt and verify). The signature length is fixed at 256 bytes. The processed data is merged into the original information set. Each piece of information includes standardized intermediate data, check result (complete / legitimate), timestamp, and signature. The information is stored in tables according to the interface type. The storage path is fixed and cannot be modified. The data volume of the single-day information set is not more than 10 GB. The data retention period is 90 days, which ensures that the information is traceable and cannot be tampered with.

[0027] Further, the construction of the structured information feature library includes: User behavior data is extracted from the original information set to analyze the operation path and interaction frequency by using a sequence pattern mining algorithm, and the user preference weight is calculated to generate a user portrait feature vector including basic attributes, behavior labels, and interest dimensions. In the embodiment of the application, user behavior data is extracted from the original information set, including user login time (accurate to minutes) for 30 consecutive days, page browsing path (URL recorded in access order), click count (click volume of each button), and stay duration (seconds per page). The operation path is analyzed by using a sequence pattern mining algorithm. The minimum support is set to 5%, and the minimum confidence is set to 80%. The high-frequency path (appearance frequency accounts for 35% of the total path) of “home page→product list→detail page→order” is mined. The interaction frequency is counted according to the number of daily operations (average 8 times / day, standard deviation 2 times). The user preference weight is calculated by using the weighted average method. The click count weight is 0.4, the stay duration weight is 0.3, and the path completion rate weight is 0.3. The user preference weight for the “electronic product” category is 0.6, the “clothing” category is 0.3, and the “food” category is 0.1. The user portrait feature vector is generated. The basic attributes include user_id (10-digit integer), registration time (yyyy-MM-dd), and regional code (6-digit integer). The behavior labels include high-frequency operations (ordering and collecting) and active time period (19:00-21:00). The interest dimensions include preference weights (0.6 for electronic products, etc.). The vector dimension is fixed at 20, the numerical range is 0-1, and the accuracy is 0.01. The feature vector corresponds to the user ID one by one, and is arranged in ascending order according to the user ID when stored.

[0028] Preferably, the environmental perception data is spatio-temporal feature extracted to convert the geographic location information into a regional grid code, the timestamp into a time period type and a holiday identifier, and the network status into a connection quality level, to generate an environmental feature matrix; In the embodiment of the application, by spatio-temporal feature extraction on the environmental perception data, the geographic location information is divided into regional grids according to longitude and latitude, each 0.01°*0.01° being a grid, and the coding rule being "longitude integer + longitude decimal two digits + latitude integer + latitude decimal two digits", such as 116303990 for 116.30° east longitude and 39.90° north latitude. The timestamp is parsed into a time period type, which is divided into 6 time periods according to 24 hours (0:00-4:00 is time period 1, 4:00-8:00 is time period 2, and so on), and the holiday identifier is marked according to the national statutory holiday table (1 for holiday, 0 for weekday). The network status is mapped into a connection quality level, which is divided into 5 levels according to signal strength (-50dBm to -100dBm), -50 to -60dBm being level 1 (excellent), -60 to -70dBm being level 2 (good), -70 to -80dBm being level 3 (medium), -80 to -90dBm being level 4 (poor), and -90 to -100dBm being level 5 (very poor). An environmental feature matrix is generated, each row corresponding to a piece of data, and the columns containing grid code (9-bit string), time period type (1-6 integer), holiday identifier (0 or 1), connection quality level (1-5 integer), temperature (-40 to 80℃, accuracy 0.1℃), and humidity (0 to 100%, accuracy 0.1%). The number of rows of the matrix is consistent with the number of environmental data, and the data is stored in groups according to the grid code and the time period type, with a query response time <1 second.

[0029] Preferably, the service resource list in the business system data is parsed to extract resource type, capacity upper limit, available quantity and associated dependency relationship, and the task queue is prioritized and time-limited, to generate a business feature parameter set, including resource feature vector, task feature matrix and constraint condition set; In the embodiment of the application, by analyzing the service resource list in the business system data, the resource type is divided into computing resources (server CPU core number), storage resources (hard disk capacity GB), and network resources (bandwidth Mbps), the upper limit of capacity is set to CPU 64 cores, storage 1024 GB, and bandwidth 100 Mbps, and the available number is real-time counted (such as the current available CPU 28 cores). The association dependency relationship is recorded in the format of "resource A→resource B" (computing resources depend on network resources). The task queue is prioritized, a dynamic priority algorithm is adopted, the emergency level (1-5 levels) weight is 0.5, the deadline (the number of hours from the current time) weight is 0.3, the task type (core business 1, non-core 0.5) weight is 0.2, the priority value (range 1-10) is obtained, and the priority value is arranged in descending order. The time limit for performance is calculated according to the remaining time (accurate to the minute), such as "remaining 120 minutes". The generated business feature parameter set includes the CPU available core number (integer), the storage available capacity (integer GB), and the bandwidth available value (integer Mbps); the task feature matrix includes the task ID (32-bit string), the priority (1-10), and the remaining time (minute); and the constraint condition set includes hard conditions such as "CPU core number≥4" and "storage capacity≥100 GB". The parameter set is updated every 5 minutes to ensure that the data is consistent with the actual business state, and the error is 0.

[0030] Preferably, by establishing a feature association index, the user portrait feature vector, the environment feature matrix, and the business feature parameter set are associated and stored according to the time dimension and the scene dimension, and a structured information feature library supporting fast query and feature combination is constructed.

[0031] In the embodiment of the application, by establishing a feature association index, a B+ tree index structure is adopted, the primary key is a timestamp (accurate to seconds) + a scene ID (6-bit integer), the time dimension is divided into index nodes according to minute granularity, and the scene dimension is divided into two-level indexes according to 10 types of scenes such as "shopping", "office", and "entertainment". The user portrait feature vector, the environment feature matrix, and the business feature parameter set are associated according to the time dimension, and the three types of data in the same minute are bound by the timestamp; and the association according to the scene dimension binds the user shopping preference vector, the shopping mall environment matrix, and the commodity resource parameters in the "shopping" scene. The associated storage adopts a columnar storage structure, and the user feature column, the environment feature column, and the business feature column are stored respectively, supports combined query according to the time range (such as 2024-05-0108:00-10:00) and the scene type (such as "shopping"), the data block is located through the index during the query, and the number of scanned rows is less than 1000 rows. The constructed structured information feature library supports fast query (response time<500 ms) and feature combination (such as "user A+shopping scene+2024-05-01"), the data compression rate is 50%, the storage capacity is reduced by 50% compared with the original data, the data integrity is maintained, and the feature value error is 0.

[0032] Further, the construction of the information scheduling scene classifier includes: extracting historical scene sample data from a structured information feature library, wherein each historical scene sample contains a user feature vector, an environment feature slice, a business demand parameter, and a corresponding artificial labeled scene type, and constructing a scene training dataset and a scene verification dataset; In the embodiment of the present application, by extracting historical scene sample data from a structured information feature library, each historical scene sample contains a user feature vector (20 dimensions, numerical 0-1), an environment feature slice (including grid encoding, time period type, etc. 6 fields), a business demand parameter (CPU available core number, etc. Resource data) and an artificial labeled scene type (10 types, represented by 1-10 integers). The extraction range is the data of the past 6 months, a total of 10000 samples, of which "shopping" scene 2000, "office" scene 1800, "entertainment" scene 1500, and the rest of the scene is allocated in proportion. The samples are divided into scene training dataset (7000) and scene verification dataset (3000) in the ratio of 7:3, ensuring that the scene type distribution of the two types of datasets is consistent ("shopping" scene accounts for 20%). The user feature vector of each sample is completely consistent with the original data in the structured information feature library, the environment feature slice intercepts the environment data 5 minutes before and after the scene occurs, the business demand parameter takes the real-time value at the time of scene occurrence, and the artificial labeled scene type is independently labeled by three experts. Inconsistent samples are determined by voting (the final consistency rate is 98%), and the dataset is stored in binary format, with each sample occupying a fixed space of 1024 bytes.

[0033] Preferably, a bidirectional long short-term memory network is used to construct a sequence feature extractor, and the corresponding user behavior time sequence and environment parameter change sequence in the structured information feature library are encoded based on the sequence feature extractor to generate a time sequence feature vector. In the embodiment of the present application, a bidirectional long short-term memory network is adopted to construct a sequence feature extractor, which includes an input layer (32 neurons corresponding to 32-dimensional original sequence features), two bidirectional LSTM layers (64 units in the first layer and 32 units in the second layer, with a tanh activation function and a 1.0 bias initialization for the forgetting gate), and an output layer (16 neurons). Based on the extractor, the user behavior time sequence (including the number of clicks, the length of stay, and the like at each of the 10 time steps with an interval of 1 minute) and the environmental parameter change sequence (temperature, humidity, connection quality level, and the like at each of the 10 time steps) in the structured information feature library are encoded. The encoding process is as follows: the user behavior sequence is input into the network according to the time steps, the forward LSTM captures the behavior trend from early to late, the backward LSTM captures the behavior association from late to early, and the outputs of the two LSTM layers are concatenated and compressed into a 16-dimensional vector by the output layer; the environmental parameter sequence is processed in the same way to generate a 16-dimensional vector, and the two vectors are concatenated to generate a 32-dimensional time sequence feature vector. The value range of each element in the vector is -1 to 1, with a precision of 0.0001, the mapping error between the encoded vector and the original sequence is less than 0.01 calculated by the mean square error, and the time sequence features are ensured not to lose key information.

[0034] Preferably, an attention mechanism module is constructed to calculate the attention weights of each dimension in the time sequence feature vector and highlight the key features with high contribution to scene recognition, and generate a weighted fusion feature. In the embodiment of the present application, an attention mechanism module is constructed, which includes a feature scaling layer (standardizing the 32-dimensional time sequence feature vector to a mean of 0 and a standard deviation of 1), a weight calculation layer (adopting a single hidden layer neural network with 8 neurons and a softmax activation function), and a weighted summation layer. The attention weights of each dimension in the time sequence feature vector are calculated, the normalized feature vector is input, the weight value of each dimension is calculated by the neural network (the sum of 32 weights is 1), and the key features with high contribution to scene recognition (such as the order number dimension in the user behavior sequence and the connection quality level dimension in the environment sequence) are assigned higher weights (such as 0.15, higher than the average weight 0.03125). The process of generating a weighted fusion feature is as follows: the feature values of each dimension are multiplied by the corresponding attention weights and summed to obtain a 32-dimensional weighted vector, in which the weighted values of key features account for more than 60% (such as the feature value 0.8 of the order number dimension multiplied by the weight 0.15, with a contribution of 0.12). The weight parameters of the module are automatically learned in the training process by the back propagation algorithm, and the initial weights are evenly distributed (0.03125 for each dimension). After 50 rounds of training, the weights of the key features are increased to 0.1-0.2, and the weights of the non-key features are reduced to below 0.01.

[0035] Preferably, the weighted fused features are input into a deep neural network classifier to output a scene type probability distribution through a softmax function, and a cross-entropy loss function is used for model training and verification on a scene training dataset and a scene verification dataset, and the training is stopped when the accuracy of the scene verification dataset reaches a preset threshold, the model parameters are saved, and a trained information scheduling scene classifier is generated. In the embodiment of the present application, the weighted fused features (32 dimensions) are input into a deep neural network classifier, the classifier includes two fully connected layers (64 neurons in the first layer with a ReLU activation function, and 32 neurons in the second layer with a ReLU activation function), and an output layer (10 neurons corresponding to 10 scene types). A scene type probability distribution (the sum of 10 probability values is 1) is output through a softmax function, such as a probability of 0.85 for a "shopping" scene, a probability of 0.10 for an "office" scene, and a total probability of 0.05 for the remaining 8 categories. A cross-entropy loss function is used for model training and verification on a scene training dataset and a scene verification dataset, the batch size is set to 64 during training, the optimizer is Adam (learning rate 0.001, β1=0.9, β2=0.999), the accuracy is calculated on the verification dataset after each round of training. When the accuracy of the scene verification dataset reaches a preset threshold of 90%, the training is stopped (the accuracy reaches 90.2% for the first time after the 87th round of training), the model parameters (including LSTM layer weights, attention mechanism weights, and classifier weights, a total of 12582912 parameters with a precision of 32-bit floating point numbers) are saved, and a trained information scheduling scene classifier is generated. The recognition accuracy of the classifier for each scene type is not less than 85%, among which the "shopping" scene reaches 92%, and the "office" scene reaches 89%.

[0036] Preferably, in real-time operation, the newly collected structured information feature data are input into the trained information scheduling scene classifier to output the type label, confidence, and information scheduling scene recognition result corresponding to the key trigger factor of the current scene, wherein the key trigger factor is the top N feature values that contribute most to the scene classification.

[0037] In the embodiment of the application, by real-time running, the newly collected structured information feature data (containing user's current continuous 10 minutes of behavior data, environmental parameter data, business demand parameter) is encoded into a 32-dimensional time sequence feature vector according to the method of step S212, and is input into the trained information scheduling scene classifier. The classifier outputs the type label (such as the label 1 corresponding to the "shopping" scene) of the current scene, the confidence (the probability value corresponding to the type label, such as 0.88), and the key trigger factor (the top 5 feature values that contribute most to the scene classification, including the "order 1 time in the last 5 minutes" feature value 0.9 in the user behavior, the weight 0.18; the "connection quality level 1" feature value 1.0 in the environment, the weight 0.15, etc.). The whole recognition process is completed within 50 milliseconds, the matching degree of the type label with the actual scene is calculated as 91% through 1000 continuous tests, the scenes with a confidence lower than 0.7 (accounting for 5%) are marked as "low confidence", but the type label is still output, the feature values of the key trigger factor are completely consistent with the original data, ensuring that the user can trace the classification basis, and the recognition result is updated every 10 seconds, with a delay of less than 20 seconds to the actual scene change.

[0038] Further, the generation process of the candidate service scheme set includes: By sorting out typical scene types and corresponding service cases, the mapping relationship between scene feature parameters and service attributes is extracted to form initial rule entries, wherein each rule entry includes scene conditions, service types, priority weights and execution constraints; In the embodiment of the application, by sorting out 10 typical scene types (shopping, office, entertainment, etc.) and corresponding 500 service cases, the mapping relationship between scene feature parameters and service attributes is extracted. The "shopping" scene feature parameters include user preference weight (electronic products > 0.6), environmental connection quality level (≤2), time period type (18:00-22:00), and the corresponding service attribute is product recommendation (type: accurate push, priority: 0.8, execution constraint: daily push ≤5 times); the "office" scene feature parameters include user behavior label (document editing frequency > 10 times / hour), environmental temperature (22-26℃), and business resource CPU available core number (≥8), and the corresponding service attribute is cloud document synchronization (type: real-time synchronization, priority: 0.9, execution constraint: synchronization interval ≤1 minute). 100 initial rule entries are formed, each rule entry includes scene conditions (such as "shopping scene + connection quality level 1"), service types (such as "limited time discount push"), priority weights (0.1-1.0, step 0.1), and execution constraints (such as "push time ≤30 seconds / time"). The rule entries are stored in groups according to scene types, the number of entries in each group is proportional to the number of scene service cases, ensuring that the mapping relationship covers all typical service scenes without missing scene types.

[0039] Preferably, the initial rule entries are encoded into a machine recognizable rule language using a production rule representation, which includes a rule antecedent that is a combination of scenario features connected by logical operators AND / OR and a rule consequent that is a service operation instruction including action type, parameters, and execution time, and a rule base infrastructure is built based on the rule language; In the embodiments of the present application, the initial rule entries are encoded into a machine recognizable rule language using a production rule representation, in which the rule antecedent is a combination of scenario features connected by logical operators AND / OR and the rule consequent is a service operation instruction including action type, parameters, and execution time. A rule for the "shopping" scenario is encoded as: IF (scenario type = shopping AND user preference.electronic product > 0.6 AND environment.connection quality level = 1) THEN (service type = commodity push; parameter = electronic product list, quantity = 3; execution time = immediately). The rule language is encoded in ASCII, in which each feature parameter in the antecedent strictly corresponds to the field name of the structured information feature base (e.g., "user preference.electronic product") and the feature value is in a fixed format (numeric value with two decimal places and string with quotation marks). The service operation instruction in the consequent includes three mandatory fields, and the parameter value type is consistent with the service interface requirement (list type separated by commas and quantity as an integer). A rule base infrastructure is built based on the rule language using a three-layer architecture: the top layer is a scenario type directory (10 directories), the middle layer is a rule group (5-15 groups for each directory), and the bottom layer is a rule entry (5-20 entries for each group). The structure is managed by an index file, each rule entry is assigned a unique ID (6-digit integer), and the storage address is fixed with a read time < 10 milliseconds.

[0040] Preferably, historical service execution data is acquired, and the rule base infrastructure is optimized based on the historical service execution data to discover potential scenario-service association patterns by an association rule mining algorithm and to supplement new rule entries based on the scenario-service association patterns while eliminating contradictory rules using a rule conflict detection algorithm to generate a preset scenario-service mapping rule base. In the embodiment of the application, the rule base infrastructure is optimized based on historical service execution data (containing 10,000 service records, each record containing scene characteristics, service type, execution result, and user feedback) of the past 12 months. By analyzing the data through an association rule mining algorithm (support threshold 0.05, confidence threshold 0.8), the association pattern (support 0.06, confidence 0.85) of "shopping scene + weekend period + humidity < 60%" and "next-day delivery recommendation" is found, and a new rule item is added based on the pattern: IF (scene type = shopping AND environment.period type = weekend AND environment.humidity < 60) THEN (service type = delivery recommendation; parameter = next-day delivery, discount 5 yuan). Conflicting rules are eliminated using a rule conflict detection algorithm (compare the antecedents line by line for similarity, and if the similarity is greater than 90% and the consequents are different, it is determined as a conflict). For example, if two "office scene" rules have the same antecedent but different consequents "cloud printing" and "local printing", the "cloud printing" rule with higher user feedback score (8.5 > 7.2) is retained. The preset scene-service mapping rule base is generated after optimization, containing 150 rule items (50 new items added), the rule conflict rate is reduced to 0, the rule coverage rate (the proportion of covered scene-service combinations) is improved from 70% to 95%, and each rule is associated with 5-10 historical execution data as verification basis.

[0041] Preferably, when the service decision engine calls the preset scene-service mapping rule base, the information scheduling scene recognition result is first matched with the rules to improve the matching efficiency using the Rete algorithm, and the matching successful rules are sorted by priority, and the service matching degree corresponding to the rules is calculated in combination with the real-time business resource state; In the embodiment of the present application, when the service decision engine calls the preset scene-service mapping rule library, the information scheduling scene recognition result (containing scene type = shopping, user preference. Electronic product = 0.7, environment. Connection quality level = 1) is first subjected to rule matching, a rule network (containing an α node storing a single feature condition and a β node storing a feature combination) is constructed by using a Rete algorithm, the feature parameters of the recognition result are matched with the α node, the feature flow meeting the condition is combined and verified in the β node, and the matched rule entries are output after 3-layer node verification. The algorithm makes the matching efficiency 10 times higher than linear search, and the matching time of 1000 rules is less than 50 milliseconds. The 3 rules matched successfully are sorted according to priority (weight 0.9>0.8>0.7), and the service matching degree is calculated in combination with the real-time business resource state (such as commodity inventory: electronic product A has stock, inventory = 20), and the formula is: matching degree = rule priority × resource available coefficient (in stock = 1.0, out of stock = 0.5), and the matching degrees of the 3 rules are respectively 0.9 × 1.0 = 0.9, 0.8 × 1.0 = 0.8 and 0.7 × 0.8 (partially out of stock) = 0.56. The matching degree is calculated to 0.01, and the resource available coefficient is updated every 10 seconds to ensure reflecting the real-time resource state.

[0042] Preferably, the service schemes corresponding to the first M rules with the highest service matching degrees are selected as the candidate service scheme set, and the value of M is dynamically adjusted according to the business complexity corresponding to the rules.

[0043] In the embodiment of the present application, the service schemes corresponding to the first M rules with the highest service matching degrees are selected, and the value of M is dynamically adjusted according to the business complexity corresponding to the rules: the business complexity is divided according to the number of service operation steps (1-3 steps for simple, M = 5; 4-6 steps for medium, M = 3; more than 7 steps for complex, M = 2). The commodity pushing service of the “shopping” scene belongs to a simple business (2 steps: generating a list → pushing), M = 5, and the service schemes with the top 5 matching degrees (0.9, 0.8, 0.75, 0.7, 0.65) are selected from the matched rules to form the candidate service scheme set. Each scheme contains a rule ID, a service type, a specific operation (such as “pushing electronic products A / B / C with a 3-yuan coupon”), an expected execution time (<5 seconds) and resource consumption (CPU occupation 0.5 core). The scheme set is arranged in descending order of matching degree, and the schemes with the same matching degree are arranged in ascending order of rule ID. The service schemes in the scheme set need to meet all execution constraints (such as “pushing quantity ≤ 3”), and the schemes violating the constraints are automatically excluded (even if the matching degree is high). The candidate service scheme set is updated every 2 seconds to ensure containing the optimal scheme under the latest resource state, and the capacity of the scheme set is fixed as M, and the excess part is directly discarded.

[0044] Further, the operation of the multi-target decision algorithm comprises: determining an objective function for each service plan evaluation, including a response speed target, a resource consumption target, and a user satisfaction target, wherein the response speed target is defined as service startup delay minimization, the resource consumption target is defined as CPU / memory occupancy minimization, and the user satisfaction target is defined as feedback score maximization; In the embodiment of the present application, by determining the objective function for each service plan evaluation, the response speed target is defined as service startup delay minimization, the time interval from service instruction issuance to actual service startup is measured, the unit is millisecond, and the threshold is set to be <100 milliseconds, and if it is exceeded, the target score is 0; the resource consumption target is defined as CPU / memory occupancy minimization, the CPU occupancy rate is calculated according to the number of cores (for example, 1 core is occupied in a 4-core CPU, which is 25%), the memory occupancy rate is calculated according to the ratio of the actual used capacity to the total capacity (for example, 2 GB is used in 8 GB memory, which is 25%), and the average value of the two is taken as the resource consumption index, and the threshold is set to be <30%; the user satisfaction target is defined as feedback score maximization, the score range is 1-5 points, and the score is submitted immediately after the user operation, the average value of the historical scores of the service is taken as the current evaluation value, and the threshold is set to be >4 points. The three objective functions are independent of each other, and the values are taken respectively when calculating, for example, the startup delay of a certain commodity push service is 50 milliseconds, the CPU occupancy rate is 20%, the memory occupancy rate is 15%, and the user score is 4.5 points, the original values of the three targets are 50, 17.5, and 4.5 respectively, ensuring that each target has a clear quantitative standard and calculation method without ambiguous parameters.

[0045] Preferably, historical service evaluation data is obtained, and a weight coefficient is set for each objective function, wherein the weight coefficient is determined based on the historical service evaluation data by an analytic hierarchy process to generate an objective weight vector. In the embodiment of the present application, by acquiring the historical service evaluation data of the past 6 months, containing 5000 records, each record contains service type, response speed (50-200 milliseconds), resource consumption (10%-50%), user satisfaction (3-5 points) and corresponding business effect (such as conversion rate, complaint rate). Based on these data, the weight coefficient is determined by the analytic hierarchy process, the judgment matrix (1-9 scale method) is constructed, the influence degree of response speed on business effect is 7 (strongly important), the resource consumption is 3 (slightly important), the user satisfaction is 5 (obviously important), the maximum eigenvalue of the matrix is calculated as 3.05, the consistency ratio CR=0.02<0.1, and the judgment matrix is consistent. The weight coefficient is: response speed 0.5 (50%), resource consumption 0.2 (20%), user satisfaction 0.3 (30%), the target weight vector [0.5, 0.2, 0.3] is generated, the sum of the vector elements is 1, and the accuracy is 0.01. The weight coefficient is fixed, unless the historical data accumulation amount exceeds 10000, otherwise it will not be recalculated, to ensure the stability of the evaluation standard, and there is no difference in the adjustment of different service types.

[0046] Preferably, for each service scheme in the candidate service scheme set, the evaluation index value is extracted and standardized, and converted into a normalized score in the interval [0, 1]; In the embodiment of the present application, by extracting the evaluation index values of 5 service schemes (product push A, B, C, D, E) in the candidate service scheme set: A's startup delay is 50 milliseconds, resource consumption is 17.5%, user score is 4.5 points; B's startup delay is 60 milliseconds, resource consumption is 20%, user score is 4.3 points; C's startup delay is 70 milliseconds, resource consumption is 15%, user score is 4.2 points; D's startup delay is 80 milliseconds, resource consumption is 25%, user score is 4.4 points; E's startup delay is 90 milliseconds, resource consumption is 18%, user score is 4.1 points. Standardization processing is performed, the response speed adopts reverse normalization (score=1-(actual value-min value) / (max value-min value)), A's score=1-(50-50) / (90-50)=1.0; Resource consumption is also reverse normalized, A's score=1-(17.5-15) / (25-15)=0.75; User satisfaction adopts positive normalization (score=(actual value-min value) / (max value-min value)), A's score=(4.5-4.1) / (4.5-4.1)=1.0. After processing, all scores are converted into normalized scores in the interval [0, 1], A's score is [1.0, 0.75, 1.0], B's score is [0.75, 0.5, 0.5], C's score is [0.5, 1.0, 0.25], D's score is [0.25, 0, 0.75], E's score is [0, 0.7, 0], accurate to 0.01, to ensure that different orders of magnitude indicators can be directly compared.

[0047] Preferably, a weighted comprehensive score corresponding to each service scheme is calculated based on the target weight vector and the normalized score, and each service scheme is ranked according to the weighted comprehensive score from high to low. In the embodiment of the application, the weighted comprehensive score corresponding to each service scheme is calculated based on the target weight vector [0.5, 0.2, 0.3] and the normalized score, and the calculation formula is: comprehensive score = response speed score × 0.5 + resource consumption score × 0.2 + user satisfaction score × 0.3. The comprehensive score of commodity push A is 1.0 × 0.5 + 0.75 × 0.2 + 1.0 × 0.3 = 0.5 + 0.15 + 0.3 = 0.95; the score of B is 0.75 × 0.5 + 0.5 × 0.2 + 0.5 × 0.3 = 0.375 + 0.1 + 0.15 = 0.625; the score of C is 0.5 × 0.5 + 1.0 × 0.2 + 0.25 × 0.3 = 0.25 + 0.2 + 0.075 = 0.525; the score of D is 0.25 × 0.5 + 0 × 0.2 + 0.75 × 0.3 = 0.125 + 0 + 0.225 = 0.35; and the score of E is 0 × 0.5 + 0.7 × 0.2 + 0 × 0.3 = 0 + 0.14 + 0 = 0.14. Ranked according to the weighted comprehensive score from high to low are: A (0.95) > B (0.625) > C (0.525) > D (0.35) > E (0.14), the score is kept to three decimal places, the ranking result is unique, there is no parallel situation, and the priority is directly determined according to the order to ensure that the calculation process is transparent and the result is reproducible.

[0048] Preferably, the service scheme with the highest weighted comprehensive score is selected as the optimal service scheme, and if there are service schemes with the same score, the potential risks of each service scheme are predicted by a random forest algorithm, and the service scheme with the lowest risk value is selected.

[0049] In the embodiment of the present application, the service scheme A with the highest weighted comprehensive score is selected as the optimal service scheme, the startup delay of the scheme is 50 ms, the resource consumption is 17.5%, and the user score is 4.5, which meets all the target thresholds (<100 ms, <30%, >4 points). If there are service schemes with the same score (such as schemes F and G with a comprehensive score of 0.8), the potential risks of each service scheme are predicted by a random forest algorithm, which includes 100 decision trees, the input features are historical execution failure rate, resource fluctuation coefficient, and user complaint times, and the output risk value (0-1, the higher the risk is greater). The historical failure rate of scheme F is 2%, the resource fluctuation coefficient is 5%, and the complaint times are 3 times, and the predicted risk value is 0.2; the historical failure rate of scheme G is 5%, the resource fluctuation coefficient is 10%, and the complaint times are 8 times, and the predicted risk value is 0.5, and scheme F with lower risk value is selected as the optimal scheme. After the optimal service scheme is determined, a detailed plan including execution steps, resource allocation, and time nodes is generated, the execution steps are fixed as "resource locking -> service startup -> result feedback", the resource allocation is clear CPU core number 1, memory 2 GB, and the time node is accurate to seconds, which ensures that the scheme can be directly implemented without ambiguity. Further, the running of the service execution process comprises: receiving a service instruction sequence corresponding to the optimal service scheme, and parsing the operation type, parameter list and dependency relationship in the service instruction sequence to generate a service execution flowchart; In the embodiment of the present application, the service instruction sequence corresponding to the optimal service scheme A is received, which includes 3 instructions: "resource locking (CPU=1 core, memory=2 GB)" "commodity pushing (list=electronic products, quantity=3)" "result feedback (method=popup, content=push completed)". Analyzing the operation type: resource locking is a preprocessing operation, commodity pushing is a core operation, and result feedback is a finishing operation; the parameter list: the parameter value of resource locking is fixed, the electronic product list of commodity pushing includes "mobile phone, computer, tablet", and the popup display time of result feedback is 5 seconds; the dependency relationship: commodity pushing must be executed after resource locking is completed, and result feedback must be executed after commodity pushing is completed. A service execution flowchart is generated, which adopts a directed acyclic graph structure, nodes are instructions (circles, diameter 2 cm), edges are dependency relationships (arrow lines, line width 0.2 cm), nodes are labeled with operation type and unique ID (101, 102, 103), and edges are labeled with "after completion". Each node in the flowchart includes a parameter table (2 columns and 3 rows) and an execution condition (such as the condition of node 102 is "101 node returns success"), the graph is stored in vector format, which is not blurred when enlarged or reduced, and the flow relationship is intuitive and unambiguous.

[0050] Preferably, the task scheduling queue is constructed based on the service execution flowchart to determine the task execution order by using a topological sorting algorithm, and each task is assigned a unique identifier and a timeout threshold; In the embodiment of the present application, the task scheduling queue is constructed based on the service execution flowchart, the three task nodes are sorted by using a topological sorting algorithm, the inter-node dependency relationship (101→102→103) is checked, there is no circular dependency, and the sorting result is 101, 102, and 103. Each task is assigned a unique identifier: 101 (resource locking), 102 (commodity pushing), and 103 (result feedback), the identifier is a 6-digit integer, the first two digits are the service type (01 represents the commodity pushing service), and the last four digits are the serial number. The timeout threshold is set: the 101 task timeout threshold is 5 seconds (resource locking needs to be completed within 5 seconds), the 102 task is 10 seconds (pushing 3 commodities does not exceed 10 seconds), and the 103 task is 3 seconds (the pop-up window feedback time is <3 seconds). The task scheduling queue uses an array structure for storage, each element contains the task ID, the front task ID (the front of 102 is 101, and the front of 103 is 102), and the timeout threshold, the queue length is fixed at 3, and no other tasks can be inserted. The queue checks the front task state every 100 milliseconds, when the front task is completed (returns a success identifier), the current task is marked as “to be executed”, ensuring that the tasks are executed in order without advance or lag.

[0051] Preferably, the task resource requirements and real-time resource states corresponding to each task are obtained by calling the corresponding business interface and resource component, and the hardware resources including CPU, memory, and network bandwidth, and the software resources including business interfaces and data connections are dynamically allocated based on the task resource requirements and real-time resource states by using a dynamic programming algorithm to generate a service execution task resource allocation result. In the embodiment of the application, the resource requirements of each task are obtained by calling a resource management interface (returning total CPU core number 8 cores, available 6 cores; total memory capacity 16 GB, available 10 GB; total bandwidth 100 Mbps, available 80 Mbps) and a service interface (returning commodity database connection number 5, available 3), that is, 101 task requires CPU 1 core, memory 2 GB, no network bandwidth; 102 task requires CPU 0.5 core, memory 1 GB, bandwidth 20 Mbps, database connection 1; 103 task requires CPU 0.2 core, memory 0.5 GB, no database connection. The dynamic programming algorithm is used to allocate resources, and the maximum resource utilization is taken as the target to allocate resources in stages: in the first stage, 101 task is allocated CPU 1 core (number 3), memory 2 GB (address 0x10000000-0x18000000); in the second stage, 102 task is allocated CPU 0.5 core (the remaining part of number 3), memory 1 GB (address 0x18000000-0x1C000000), bandwidth 20 Mbps (port 1001), database connection 1 (ID: DB003); in the third stage, 103 task is allocated CPU 0.2 core (number 4), memory 0.5 GB (address 0x1C000000-0x1E000000). The resource allocation results are recorded in a table, including task ID, resource type, allocation value, resource number, to ensure that the resources are not overlapped, and the utilization rate reaches 90% (allocated CPU 1.7 core / available 6 core, memory 3.5 GB / available 10 GB).

[0052] Preferably, the service execution task resource allocation results are sequentially called corresponding service components according to the task scheduling queue to execute service operations through a remote procedure call protocol, and the start time, end time, return result and exception information of each task are recorded in real time, and the service execution state flow including the progress percentage, resource occupation rate and abnormal analysis report corresponding to the start time, end time, return result and exception information of each task are obtained; In the embodiment of the application, by performing task resource allocation based on services, corresponding service components are sequentially called according to a task scheduling queue: 101, a task calls a resource management component (dynamic link library file, version 2.3), sends a locking instruction through a remote procedure call protocol (RPC, transmission protocol TCP, port 50051), the 101 task starts at 10:00:00 and ends at 10:00:02, and returns a result of "locking success" without any exception; 102, a task calls a commodity push component (Web Service interface, SOAP protocol), starts at 10:00:02, ends at 10:00:05, and returns "pushing mobile phones, computers and tablets successfully", the CPU occupancy rate is 45% (90% of 0.5 cores), and the memory occupancy rate is 10% (1 GB / 10 GB); 103, a task calls a UI feedback component (local API, C language implementation), starts at 10:00:05, ends at 10:00:06, and returns "pop-up window display is completed". The time stamp (accurate to milliseconds), return result (string length < 100 bytes) and exception information (102 task has a warning of "network fluctuation 0.5 seconds") of each task are recorded in real time, a service execution state stream is generated, the progress percentage (101 completes 33%, 102 completes 67%, and 103 completes 100%), the resource occupancy rate curve (one data point per second) and the exception analysis report (network fluctuation reason: switch port transient congestion) are generated, the state stream is updated once per second, and the state stream is stored in JSON format with fixed field names.

[0053] Preferably, if an exception occurs in the service execution process, a corresponding fault tolerance mechanism is triggered according to the type of the exception, including a retry mechanism, a degradation mechanism or a fuse mechanism, to ensure the continuity of the service execution process.

[0054] In the embodiment of the present application, when an exception occurs in the service execution process, the corresponding fault-tolerant mechanism is triggered according to the type of the exception: if a "database connection failure" (exception type 1) occurs during the execution of the task 102, a retry mechanism is triggered, the retry interval is 1 second, the maximum number of retries is 3, the first retry is at 10:00:06, the second retry is at 10:00:07, the third retry is at 10:00:08, and if the third retry still fails, the failure reason is recorded; if a "bandwidth insufficient 20Mbps" (exception type 2) occurs, a degradation mechanism is triggered, the number of pushed goods is reduced from 3 to 2, the bandwidth demand is reduced to 15Mbps, and the service continues to execute; if a "push timeout" (exception type 3) occurs 5 times in 10 minutes, a fuse mechanism is triggered, the execution of the task 102 is suspended for 5 minutes, and a prompt of "service busy, please try again later" is returned, and alarm information (including the type of the exception, the time of occurrence, and the scope of influence) is sent to the operation and maintenance system. The triggering conditions and execution steps of the fault-tolerant mechanism are fixed, the number of retries, the degradation parameters, and the fuse duration cannot be adjusted, and the exception handling process is recorded in the log (one record per millisecond), including the exception ID, the handling measure, and the result, so that the service execution interruption time is less than 5 seconds, and the continuity is greater than 99.5%.

[0055] Further, the analysis process of the service execution state flow includes: In the service execution process, a state collection probe is embedded, the task execution state data is collected at a preset sampling frequency, including the process ID, the resource occupation peak value, and the response time distribution, and a state snapshot with a timestamp is generated; In the embodiment of the present application, by embedding a state collection probe in the service execution process, the probe is a binary code segment (occupying 10KB of memory) compiled into the service component, the preset sampling frequency is 100 milliseconds / time, and the collection range covers the tasks 101 (resource locking), 102 (goods pushing), and 103 (result feedback). The collected task execution state data includes: the process ID (6-bit integer, such as 123456), the resource occupation peak value (CPU peak value 0.8 cores, memory peak value 1.2GB, network bandwidth peak value 18Mbps), and the response time distribution (the 10 sampling response times of the task 102 are 120ms, 130ms, 110ms, 140ms, 125ms, 135ms, 128ms, 132ms, 122ms, and 129ms). A state snapshot with a timestamp is generated each time the sampling is performed, the timestamp is accurate to milliseconds (such as 10:00:02.123), the snapshot format is a fixed-length structure (512 bytes), and 20 fields (process ID, 3 types of resource peak values, 10 response time values, and timestamp) are included. The probe collection does not affect the task execution efficiency, the CPU occupancy rate is less than 0.01 core, the number of snapshots generated by a single task throughout the execution is equal to the task execution time (seconds) multiplied by 10, the task 102 executes for 3 seconds to generate 30 snapshots, and all the snapshots are stored in ascending order of timestamp.

[0056] Preferably, the state snapshots are aggregated by service instance and time dimension, and the average execution time, success rate and abnormality occurrence rate of each task are calculated to generate service task state statistical values; In the embodiment of the present application, by aggregating the state snapshots by service instance (instance ID: 001, corresponding to the commodity push service) and time dimension (every 1 minute as a time slice), 15 snapshots of task 101 are aggregated into 10:00-10:01 time slice data, 30 snapshots of task 102 are aggregated into the same time slice, and 10 snapshots of task 103 are aggregated into the same time slice. The average execution time of each task is calculated: the total execution time of task 101 is 2 seconds (the total duration of 5 snapshot records), the execution times is 1, and the average is 2 seconds; the total execution time of task 102 is 3 seconds, the execution times is 1, and the average is 3 seconds; the total execution time of task 103 is 1 second, and the average is 1 second. The success rate is calculated according to "success times / total times", and the three tasks are successfully executed, with a success rate of 100%. The abnormality occurrence rate is calculated according to "abnormal snapshot number / total snapshot number", and task 102 has 2 snapshot records of network fluctuation, with a total of 30 snapshots, and an abnormality occurrence rate of 6.67%. The service task state statistical values are generated and presented in the form of a table, including task ID, average execution time (seconds, two decimal places), success rate (%), abnormality occurrence rate (%), statistical values updated every 1 minute, strictly corresponding to the time slice, ensuring that there is no repetition or omission in data aggregation.

[0057] Preferably, the progress percentage and resource occupancy rate corresponding to each task are estimated and calculated according to the average execution time, success rate and abnormality occurrence rate of each task; In the embodiment of the present application, the progress percentage and resource occupancy are estimated by the average execution time, success rate and abnormality occurrence rate of each task. The progress percentage is calculated as "executed time / total estimated time x 100%", and the total estimated time is the sum of the average execution times of the three tasks, i.e. 6 seconds (2+3+1). The progress of the task 101 after 2 seconds of execution is 2 / 6 x 100% = 33.33%; the progress of the task 102 after 3 seconds (cumulative 5 seconds) of execution is 5 / 6 x 100% = 83.33%; and the progress of the task 103 after 1 second (cumulative 6 seconds) of execution is 100%. The resource occupancy is calculated as "resource occupancy peak / resource allocation value x 100%", and the resource occupancy rate of the task 102 is 90% which is the average of the CPU occupancy rate of 90% (0.45 / 0.5 x 100%), the memory occupancy rate of 90% (0.9 / 1 x 100%) and the network bandwidth occupancy rate of 90% (18 / 20 x 100%). The resource occupancy rates of the tasks 101 and 103 are 80% and 70% respectively. The calculation results are rounded to two decimal places, and the progress percentage and the resource occupancy rate strictly correspond to the execution stage of the task without advanced calculation.

[0058] Preferably, a state abnormality detection model is constructed to set a control limit of a state parameter by using a statistical process control method, and when the state statistical value of the service task exceeds the control limit, it is marked as an abnormal state, and the abnormality occurrence time, duration and associated task are recorded; In the embodiment of the present application, the state abnormality detection model is constructed, and the control limit of the state parameter is set by using the statistical process control method. The control limit includes the average value ± 3 times the standard deviation. The standard deviation of the response time of the task 102 is calculated: the standard deviation of the response time of 10 samples is 9.01 ms, the upper control limit is 155.03 ms (average response time 128 ms + 3 x 9.01 ms), and the lower control limit is 100.97 ms (128 - 27.03). When the state statistical value of the service task exceeds the control limit, such as the snapshot response time of the task 102 is 160 ms > 155.03 ms, it is marked as an abnormal state. The abnormality occurrence time (10:00:04.500), duration (500 ms, the interval from the 160 ms snapshot to the next normal snapshot 120 ms) and associated task (task 102, ID 102) are recorded. The model sets the control limit for all state parameters (CPU occupancy rate, response time, abnormality occurrence rate) of the three tasks. The upper control limit of the CPU occupancy rate is 120% of the allocated value (0.5 core x 120% = 0.6 core for the task 102), and the upper control limit of the abnormality occurrence rate is 10%. Any parameter exceeding the corresponding limit value triggers the abnormality marking, and the abnormal state record is stored in a separate file, including the triggering condition and specific value.

[0059] Preferably, the abnormal state is root cause analyzed based on the abnormal occurrence time, duration and associated task to trace the direct and indirect causes of the abnormality by fault tree analysis method, and generate an abnormality analysis report containing the abnormal type, impact range and processing suggestion; In the embodiment of the present application, the abnormal state is root cause analyzed based on the abnormal occurrence time (10:00:04.500), duration (500ms) and associated task (102), and a fault tree is constructed by fault tree analysis method, the top event is "102 task response time out of limit", the intermediate events include "network delay", "CPU overload" and "database slow query", and the bottom event is the specific fault point. By tracing the state snapshot at the time of abnormality, the network bandwidth peak of 102 task is not out of limit (18Mbps), but the switch log records that there is packet loss (packet loss rate 2%) at port 1001 at this time point, and it is determined that "network delay" is the direct cause. The indirect cause is obtained by correlation analysis: the bandwidth of 104 task (other services) is 15Mbps in the same time slice, which leads to that the total bandwidth is close to the upper limit (18+15=33Mbps, total bandwidth 40Mbps), and resource competition causes network congestion. An abnormality analysis report is generated, containing the abnormal type (network delay), impact range (102 task response time is extended by 500ms, and the final result is not affected), and processing suggestion (limiting the bandwidth of 104 task to 10Mbps, and increasing the port cache to 10MB). The report adopts a fixed format, is divided into 3 chapters, each chapter contains 3 sections, the data source is marked with specific log file and timestamp, and the analysis is traceable.

[0060] Preferably, the progress percentage, resource occupation rate and abnormality analysis report are integrated to generate the corresponding service execution state flow.

[0061] In the embodiment of the present application, the progress percentage (33.33%, 83.33%, 100%), resource occupancy (101 task 80%, 102 task 90%, 103 task 70%), and abnormal analysis report (network delay related content) are integrated to generate a corresponding service execution state flow. The progress percentage is updated in real time according to the task execution order, and displays 33.33% during the execution of the 101 task, gradually increases from 33.33% to 83.33% during the execution of the 102 task, and increases to 100% during the execution of the 103 task. The resource occupancy is presented in a curve graph, the X-axis is time (10:00:00-10:00:06), and the Y-axis is occupancy (0-100%). The curve of the 102 task has a peak value of 90% at 10:00:04.500. The abnormal analysis report is embedded in the state flow in the form of a text box, located in the time interval of the 102 task, and displays the abnormal type and processing suggestion. The state flow is stored in JSON format, including three first-level fields of "progress", "resource", and "abnormal". Each field includes subfields (such as progress including task_id, percentage, and timestamp). The file is updated every 100 ms to ensure real-time reflection of the service execution state. The data fields remain consistent with the previous state flow format.

[0062] Further, the construction of the effect evaluation model includes: Each performance indicator KPI is extracted from the service execution state flow, including service completion rate, average response time, and resource utilization probability. The service completion rate is the ratio of success rate to the sum of success rate and abnormal occurrence rate, the average response time is the ratio of average execution time to duration, and the resource utilization probability is the product of progress percentage and resource occupancy. In the embodiment of the present application, by extracting each performance indicator KPI from the service execution state flow, the service completion rate is calculated as the ratio of the success rate to the sum of the success rate and the abnormality occurrence rate, the 102 task success rate is 100%, the abnormality occurrence rate is 6.67%, the service completion rate = 100% ÷ (100% + 6.67%) = 93.75%; the 101 task success rate is 100%, the abnormality occurrence rate is 0%, the service completion rate = 100% ÷ (100% + 0%) = 100%; the 103 task is the same, the service completion rate is 100%. The average response time is calculated as the ratio between the average execution time and the duration, the 102 task average execution time is 3 seconds, the duration is 3 seconds (the time difference from the beginning to the end), the average response time = 3 ÷ 3 = 1 second; the 101 task average execution time is 2 seconds, the duration is 2 seconds, the average response time = 2 ÷ 2 = 1 second; the 103 task average execution time is 1 second, the duration is 1 second, the average response time = 1 ÷ 1 = 1 second. The resource utilization probability is calculated as the product of the progress percentage and the resource occupancy rate, the 102 task progress percentage is 83.33%, the resource occupancy rate is 90%, the resource utilization probability = 83.33% × 90% = 75%; the 101 task progress percentage is 33.33%, the resource occupancy rate is 80%, the resource utilization probability = 33.33% × 80% = 26.66%; the 103 task progress percentage is 100%, the resource occupancy rate is 70%, the resource utilization probability = 100% × 70% = 70%. All KPI values are rounded to two decimal places, and accurate four arithmetic operations are used in the calculation process to ensure that the results are error-free.

[0063] Preferably, the threshold range of each performance indicator KPI is set, including the ideal value, the acceptable value and the minimum value, and when each performance indicator KPI is lower than the minimum value, it is determined as unqualified; In the embodiment of the present application, by setting the threshold range of each performance indicator KPI, the ideal value of the service completion rate is 95%, the acceptable value is 90%, and the minimum value is 85%; the ideal value of the average response time is 0.8 seconds, the acceptable value is 1.2 seconds, and the minimum value is 1.5 seconds; the ideal value of the resource utilization probability is 80%, the acceptable value is 60%, and the minimum value is 50%. The threshold range is determined based on the service data statistics of the past 12 months, the ideal value is the average value of the top 10% services, the acceptable value is the average value of the top 50% services, and the minimum value is the critical value to ensure the normal operation of the service. When each performance indicator KPI is lower than the minimum value, it is determined as unqualified, such as the service completion rate 84% < 85%, the average response time 1.6 seconds > 1.5 seconds (because the shorter the time, the better, and the actual value exceeds the minimum value, which is unqualified), the resource utilization probability 49% < 50%, all of which are determined as unqualified. The threshold range is fixed and does not change with the service type, the same indicator uses the same standard in all services, ensuring the fairness of the evaluation, and the threshold value is accurate to two decimal places (the percentage is rounded to the integer).

[0064] Preferably, the deviation rate of the actual value and the ideal value of each performance indicator KPI is calculated to generate a deviation analysis table; In the embodiment of the present application, the deviation rate of the actual value and the ideal value of each performance indicator KPI is calculated, and the deviation rate calculation formula is: (actual value-ideal value) ÷ ideal value × 100%, and the result is kept to two decimal places. A positive value indicates that it is better than the ideal value, and a negative value indicates that it is worse than the ideal value. The actual value of the service completion rate is 93.75%, the ideal value is 95%, the deviation rate is (93.75-95) ÷ 95 × 100% =-1.32%; the actual value of the average response time is 1 second, the ideal value is 0.8 seconds, the deviation rate is (1-0.8) ÷ 0.8 × 100% = 25%; the actual value of the resource utilization probability is 75%, the ideal value is 80%, the deviation rate is (75-80) ÷ 80 × 100% =-6.25%. A deviation analysis table is generated, which contains four columns: KPI name, actual value, ideal value, and deviation rate. Each row corresponds to an indicator. The calculation method and data source (specific time stamp and field name of the service execution state flow) are marked below the table. The deviation analysis table is sorted according to the importance of KPI (service completion rate > average response time > resource utilization probability), ensuring that key indicators are displayed first. The table format is fixed and has no extra fields.

[0065] Preferably, the fuzzy comprehensive evaluation method is used to comprehensively evaluate each performance indicator KPI to calculate the comprehensive score of service quality, and generate a service evaluation report, which includes the specific values of each performance indicator KPI, the deviation analysis table, the grade assessment, and the improvement suggestions.

[0066] In the embodiment of the application, the fuzzy comprehensive evaluation method is used to comprehensively evaluate each performance indicator KPI, the evaluation factor set (service completion rate, average response time, resource utilization probability) and the weight vector (0.4, 0.3, 0.3) are determined, and the evaluation set is “excellent, good, qualified, unqualified”. The fuzzy evaluation matrix is constructed, the service completion rate 93.75% corresponds to the membership degree of “good” 0.6 and “excellent” 0.3; the average response time 1 second corresponds to “good” 0.7 and “qualified” 0.2; and the resource utilization probability 75% corresponds to “good” 0.5 and “qualified” 0.4. The weight vector is multiplied by the evaluation matrix to obtain the comprehensive membership degree: excellent 0.12, good 0.61, qualified 0.23, and unqualified 0.04. The maximum membership degree corresponds to “good” as the grade evaluation, and the comprehensive score of the service quality is calculated as = (93.75*0.4 + (1-25%)*100*0.3 + 75*0.3) = 37.5 + 22.5 + 22.5 = 82 points (full score 100). The service evaluation report is generated, which includes the specific values of each performance indicator KPI (service completion rate 93.75%, etc.), deviation analysis table, grade evaluation (good), and improvement suggestions (to improve the service completion rate to 95%, the abnormality rate needs to be reduced to less than 5%, and the network bandwidth allocation needs to be optimized), the report adopts the PDF format, is divided into four parts, each part has clear data support, the improvement suggestions correspond to the deviation analysis results, and the operability is ensured, and there is no general content.

[0067] Further, the updating process of the dynamic adjustment mechanism comprises: When the service evaluation report shows that the service quality does not reach the preset service quality threshold, the parameter adjustment process is triggered, and the performance indicator KPI with the greatest impact is extracted from the deviation analysis table; In the embodiment of the application, when the service evaluation report shows that the comprehensive score of the service quality is 78 points, which is lower than the preset service quality threshold 80 points, the parameter adjustment process is triggered. The performance indicator KPI with the greatest impact is extracted from the deviation analysis table, and the influence degree is determined by calculating the product of the absolute value of each indicator deviation rate and the weight: service completion rate deviation rate -3.5%*weight 0.4=1.4, average response time deviation rate 30%*weight 0.3=9, resource utilization probability deviation rate -12%*weight 0.3=3.6, and the influence value of the average response time is the largest, so the KPI needs to be adjusted in priority. The actual value of the index is 1.2 seconds, which is 0.8 seconds more than the ideal value, and is close to the upper limit of the acceptable value 1.2 seconds, the corresponding original data (including 100 response time sampling values, of which 15 are more than 1 second) and the associated service execution state stream segment (time stamp 10:00:03-10:00:08) are extracted, so that the KPI data with the greatest impact is complete and traceable, and an explicit target is provided for subsequent adjustment.

[0068] Preferably, for the case of performance indicators KPI not meeting the standard, the corresponding scene recognition link and service decision link are analyzed to determine the rule parameters or rule factors that need to be adjusted; In the embodiment of the application, for the case of average response time not meeting the standard, the corresponding scene recognition link and service decision link are analyzed. In the scene recognition link, the information scheduling scene classifier takes 0.05 seconds to identify the "shopping scene", which does not exceed the standard of 0.1 seconds, so the scene recognition delay is excluded. In the service decision link, when the service decision engine calls the scene-service mapping rule library, the rule matching takes 0.3 seconds (standard 0.2 seconds), and the resource allocation algorithm takes 0.5 seconds (standard 0.3 seconds), resulting in overall response delay. It is determined that the rule parameter that needs to be adjusted is the α node cache time of rule matching (originally 100 milliseconds, which needs to be shortened to 50 milliseconds), and the rule factor is the iteration number of the resource allocation dynamic programming algorithm (originally 20 times, which needs to be reduced to 15 times). It is found that the rule entry priority sorting logic under the "shopping scene" has redundant judgment (multiple conditional checks are performed 3 times), which needs to be simplified to twice the check. By specifying the adjustment direction and numerical range of the parameters and factors, the adjustment target can be quantified.

[0069] Preferably, based on the rule parameters or rule factors that need to be adjusted, the scene-service mapping rule library is incrementally updated to use the latest service execution data and evaluation results as new samples, and the scene-service mapping rule library is updated based on the improvement suggestions to add new rule entries or adjust the priority weight of existing rule entries, and the effectiveness of the new rule entries is verified through simulation testing; In the embodiment of the application, based on the rule parameters (α node cache time 50 milliseconds) and rule factors (iteration number 15 times) that need to be adjusted, the scene-service mapping rule library is incrementally updated. The latest service execution data (nearly 1000 average response time records, of which 800 are <1 second) and evaluation results (average response time optimization potential 15%) are used as new samples to add a new rule entry for "shopping scene + network bandwidth > 20Mbps": IF (scene type = shopping AND environment.bandwidth > 20Mbps) THEN (service type = product push; resource allocation priority = high; response time threshold = 0.9 seconds), and the priority weight of the original "shopping scene" rule is adjusted from 0.8 to 0.85. The effectiveness of the new rule entry is verified through simulation testing, simulating 1000 times of shopping scene interaction, the average response time when the new rule is triggered is 0.85 seconds, which is 15% shorter than the original rule of 1.0 seconds, and the resource occupancy rate is maintained at 25% (standard <30%), which verifies that the new rule entry is written into the rule library, covering the original conflicting rule, and ensuring that the logic of the rule library after incremental update is consistent.

[0070] Preferably, the service evaluation index changes before and after the parameter adjustment are recorded, the optimization effect improvement rate is calculated, if the optimization effect improvement rate does not reach the expectation, the parameter adjustment process is repeated until the service quality reaches the standard, so as to realize the self-iterative upgrading of the information scheduling scenario service.

[0071] In the embodiment of the application, by recording the service evaluation index changes before and after the parameter adjustment, the average response time before adjustment is 1.2 seconds, the service completion rate is 92%, and the resource utilization probability is 68%, and after adjustment, they are 0.9 seconds, 93%, and 67% respectively. The optimization effect improvement rate is calculated, the average response time improvement rate = (1.2-0.9) ÷ 1.2 x 100% = 25%, the service completion rate is improved by 1%, and the resource utilization probability is basically unchanged, and the comprehensive optimization effect improvement rate is 15% (higher than the expected 10%). If the optimization effect improvement rate does not reach the expectation (such as only improved by 8%), the parameter adjustment process is repeated: the number of beta nodes matched by the rule is reduced from 32 to 24, the iteration number of the resource allocation algorithm is reduced from 15 to 12, and the average response time is retested until it is less than or equal to 0.9 seconds, and the service quality comprehensive score is greater than or equal to 80 points. After 3 rounds of iteration, the final average response time is 0.8 seconds, which reaches the standard and runs stably, realizes the self-iterative upgrading of the information scheduling scenario service, and the whole process is recorded in the optimization log, including the parameters of each adjustment, the test results and the improvement rate, the log is stored in the order of time stamp, and is kept for 12 months.

[0072] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the application.

[0073] The above description is only a specific embodiment of the application, which enables those skilled in the art to understand or implement the application. Various modifications of these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An information scheduling scenario-based service intelligent agent system, characterized by, Comprise the following modules: Information scheduling acquisition analysis module, for obtaining information scheduling corresponding user behavior data, environmental perception data and business system data, generating original information set with data source identification; preprocessing the original information set, and extracting key features by using feature engineering to generate structured information feature library, which contains user portrait feature vector, environmental feature matrix and business feature parameter set; Scene classification and service matching module, for constructing information scheduling scene classifier based on structured information feature library, to output the type label, confidence and key trigger factor of the current scene corresponding information scheduling scene recognition result; The information scheduling scene recognition result is input into the service decision engine to call the preset scene-service mapping rule library, and the service matching degree is calculated combined with the real-time business resource state to generate a set of candidate service schemes, wherein each service scheme contains service type, execution step, resource demand and expected effect; Service scheme instruction generation module, for optimizing and sorting the candidate service scheme set, using multi-objective decision algorithm to evaluate the response speed, resource consumption and user satisfaction of each service scheme, generating the optimal service scheme, and decomposing the optimal service scheme into executable service instruction sequence through rule engine; Service intelligent execution module, for receiving service instruction sequence to call corresponding business interface and resource component to execute service operation, and collecting state data in service execution process in real time, including progress percentage, resource occupancy rate, abnormal analysis report, generating service execution state stream; Based on the service execution state stream, an effect evaluation model is constructed to generate a service evaluation report containing performance indicators, deviation analysis table, grade evaluation and improvement suggestions; if the service evaluation report shows that the service quality threshold is not reached, the dynamic adjustment mechanism is triggered to update the scene-service mapping rule library, so as to realize the self-iterative upgrade of information scheduling scene service.

2. The information dispatch contextualization service intelligent agent system of claim 1, wherein, The generation process of the original information set includes: By configuring data access interface list when initializing the system, including RESTful API interface, message queue interface, database direct connection interface and Internet of Things device communication interface, and configuring corresponding preset data parsing template for each interface; Trigger the data collection process of each interface according to the set period, and respond to data update notification in real time through each interface using event-driven mode, generating data collection task list; Based on the data collection task list, the corresponding data parsing template is called according to the interface type, and the original data of the collected user behavior data, environmental perception data and business system data is converted into standardization intermediate data, wherein the unstructured data is converted into JSON format, and the binary data is converted into Base64 encoded string; Integrity check and legality verification are performed on the standardized intermediate data to verify the data transmission consistency by using hash algorithm, and the data passing the verification is added with timestamp and signature information, and is merged into the original information set with data source identification.

3. The information dispatch contextualization service intelligent agent system of claim 1, wherein, The construction of the structured information feature library includes: User behavior data is extracted from the original information set to analyze the operation path and interaction frequency thereof by a sequence pattern mining algorithm, and a user preference weight is calculated to generate a user portrait feature vector containing basic attributes, behavior labels, and interest dimensions; Spatial and temporal feature extraction is performed on the environmental perception data to convert geographic location information into regional grid encoding, time stamp into time period type and holiday identification, and network status into connection quality level to generate an environmental feature matrix; Service resource lists in the business system data are parsed to extract resource types, capacity upper limits, available quantities, and associated dependency relationships, and the task queue is prioritized and time-limited to generate a business feature parameter set, including resource feature vectors, task feature matrices, and constraint condition sets; By establishing a feature association index, the user portrait feature vector, environmental feature matrix, and business feature parameter set are stored in association with the time dimension and scene dimension to build a structured information feature library that supports fast query and feature combination.

4. The information dispatch contextualization service agent system of claim 3, wherein, The construction of the information scheduling scenario classifier includes: Historical scene sample data is extracted from the structured information feature library, where each historical scene sample contains a user feature vector, an environmental feature slice, a business demand parameter, and a corresponding artificially labeled scene type, and a scene training data set and a scene validation data set are constructed; A bidirectional long short-term memory network is used to build a sequence feature extractor, and based on the sequence feature extractor, the corresponding user behavior time series and environmental parameter change series in the structured information feature library are feature-encoded to generate time series feature vectors; An attention mechanism module is constructed to calculate the attention weights of each dimension in the time series feature vector and highlight key features with high contribution to scene recognition to generate weighted fusion features; The weighted fusion features are input into a deep neural network classifier to output a scene type probability distribution through a softmax function, and a cross-entropy loss function is used to train and validate the model on the scene training data set and the scene validation data set. When the accuracy of the scene validation data set reaches a preset threshold, the training is stopped, the model parameters are saved, and a trained information scheduling scenario classifier is generated; In real-time operation, the newly collected structured information feature data is input into the trained information scheduling scenario classifier to output the type label, confidence, and information scheduling scenario recognition result of the key trigger factor corresponding to the current scene, where the key trigger factor is the top N feature values that contribute most to scene classification.

5. The information dispatch contextualization service intelligent agent system of claim 4, wherein, The generation process of the candidate service scheme set includes: By sorting out typical scene types and corresponding service cases, the mapping relationship between scene feature parameters and service attributes is extracted to form initial rule items, where each rule item contains scene conditions, service types, priority weights, and execution constraints; The initial rule items are encoded into a machine-readable rule language using production rule representation, which includes rule antecedents and rule consequents. The rule antecedents are scene feature combinations, and the rule consequents are service operation instructions. A rule library infrastructure is built based on the rule language. The historical service execution data is acquired, and the rule base infrastructure is optimized based on the historical service execution data to discover potential scenario-service association patterns by an association rule mining algorithm, and new rule entries are supplemented based on the scenario-service association patterns, while contradictory rules are eliminated by a rule conflict detection algorithm, to generate a preset scenario-service mapping rule base; When the scenario-service mapping rule base is called by a service decision engine, rule matching is performed on the information scheduling scenario recognition result first, to improve the matching efficiency by using a Rete algorithm, and the service matching degrees corresponding to the rules that match successfully are calculated in combination with the real-time business resource states by sorting the rules according to the priority; The service scheme corresponding to the top M rules in the service matching degree is selected as the candidate service scheme set, and the value of M is dynamically adjusted according to the business complexity corresponding to the rules.

6. The information dispatch contextualization service intelligent agent system of claim 1, wherein, The operation of the multi-target decision algorithm includes: A target function evaluated for each service scheme is determined, including a response speed target, a resource consumption target and a user satisfaction target, wherein the response speed target is to minimize the service start delay, the resource consumption target is to minimize the CPU / memory occupancy rate, and the user satisfaction target is to maximize the feedback score; Historical service evaluation data is acquired, and a weight coefficient is set for each target function, wherein the weight coefficient is determined based on the historical service evaluation data by an analytic hierarchy process, to generate a target weight vector; For each service scheme in the candidate service scheme set, an evaluation index value is extracted and standardized, and is converted into a normalized score in the interval [0, 1]; A weighted comprehensive score corresponding to each service scheme is calculated based on the target weight vector and the normalized score, and each service scheme is sorted from high to low according to the weighted comprehensive score; The service scheme with the highest weighted comprehensive score is selected as the optimal service scheme, and if there are service schemes with the same score, a random forest algorithm is used to predict the potential risks of each service scheme, and the service scheme with the lowest risk value is selected.

7. The information dispatch contextualization service intelligent agent system of claim 1, wherein, The operation of the service execution process includes: A service instruction sequence corresponding to the optimal service scheme is received, and the operation type, parameter list and dependency relationship in the service instruction sequence are parsed to generate a service execution flowchart; A task scheduling queue is constructed based on the service execution flowchart, to determine the task execution order by using a topological sorting algorithm, and to assign a unique identifier and a timeout threshold to each task; The task resource demand and real-time resource state corresponding to each task are acquired by calling corresponding business interfaces and resource components, and the hardware resources including CPU, memory and network bandwidth, and the software resources including business interfaces and data connections are dynamically allocated based on the task resource demand and real-time resource state by using a dynamic programming algorithm, to generate a service execution task resource allocation result; The service execution state flow is generated based on the progress percentage, the resource occupation rate, and the abnormal analysis report. If an abnormality occurs in the service execution process, a corresponding fault-tolerant mechanism is triggered according to the type of the abnormality, including a retry mechanism, a degradation mechanism, or a fuse mechanism, to ensure the continuity of the service execution process.

8. The information dispatch scenarioing service intelligent agent system of claim 7, wherein, The analysis process of the service execution state flow includes: In the service execution process, state collection probes are embedded, and task execution state data is collected at a preset sampling frequency, including process ID, resource occupation peak value, and response time distribution, to generate a state snapshot with a timestamp. The state snapshot is aggregated according to service instances and time dimensions, and the average execution time, success rate, and abnormality occurrence rate of each task are calculated to generate service task state statistical values. The progress percentage and resource occupation rate of each task are estimated and calculated based on the average execution time, success rate, and abnormality occurrence rate of each task. A state abnormality detection model is constructed to set control limits for state parameters using statistical process control methods. When the service task state statistical values exceed the control limits, the state is marked as abnormal, and the abnormality occurrence time, duration, and associated tasks are recorded. Root cause analysis is performed on the abnormal state based on the abnormality occurrence time, duration, and associated tasks to trace the direct and indirect causes of the abnormality through fault tree analysis, and an abnormality analysis report is generated, including the type of abnormality, the scope of influence, and the processing suggestions. The progress percentage, resource occupation rate, and abnormality analysis report are integrated to generate a corresponding service execution state flow.

9. The information scheduling scenarioing service intelligent agent system of claim 8, wherein, The construction of the effect evaluation model includes: Various performance indicators KPIs are extracted from the service execution state flow, including service completion rate, average response time, and resource utilization probability. The service completion rate is calculated as the ratio of success rate to the sum of success rate and abnormality occurrence rate. The average response time is calculated as the ratio between average execution time and duration. The resource utilization probability is calculated as the product of progress percentage and resource occupation rate. Threshold ranges for each performance indicator KPI are set, including ideal value, acceptable value, and minimum value. When each performance indicator KPI is below the minimum value, it is determined to be substandard. The deviation rate of the actual value of each performance indicator KPI from the ideal value is calculated to generate a deviation analysis table. Each performance indicator KPI is comprehensively evaluated using fuzzy comprehensive evaluation method to calculate the comprehensive score of service quality, and a service evaluation report is generated, which includes the specific values of each performance indicator KPI, the deviation analysis table, the grade evaluation, and the improvement suggestions.

10. The information dispatch contextualization service intelligent agent system of claim 1, wherein, The update process of the dynamic adjustment mechanism includes: When the service evaluation report shows that the preset service quality threshold is not met, the parameter adjustment process is triggered, and the performance indicator KPI with the greatest impact is extracted from the deviation analysis table. For the case that the performance indicator KPI does not meet the standard, the corresponding scene recognition link and service decision link are analyzed to determine the rule parameters or rule factors that need to be adjusted; Based on the rule parameters or rule factors that need to be adjusted, the scene-service mapping rule library is incrementally updated to use the latest service execution data and evaluation results as new samples, and the scene-service mapping rule library is updated based on the improvement suggestions to add new rule entries or adjust the priority weight of the existing rule entries, and the effectiveness of the new rule entries is verified through simulation testing; The service evaluation index changes before and after the parameter adjustment are recorded, the optimization effect improvement rate is calculated, if the optimization effect improvement rate does not meet the expectation, the parameter adjustment process is repeated until the service quality meets the standard, thereby realizing the self-iterative upgrade of the information scheduling scenario-based service.

Images