Performance risk intelligent monitoring and collaborative response method and system

Abstract

The invention relates to the technical field of data processing and risk monitoring, and discloses a performance risk intelligent monitoring and collaborative response method and system, and the method comprises the steps: an edge sensing node calculates the compression characteristics of performance flow data through employing a mean aggregation algorithm according to a current sliding window parameter, and transmits the compression characteristics to a cloud end; the cloud decision-making platform executes data standardization management, calculates performance risk probability by using a model reasoning algorithm, and calculates feature uncertainty indexes by using a game theory interpretation method and a variance statistical method; a feedback control strategy module generates a self-adaptive sampling control instruction according to the index, drives an edge node to dynamically update a sliding window parameter, and constructs an acquisition feedback closed loop; and the collaborative response module calculates a response priority by utilizing multi-dimensional weighted logic, calculates desensitized data by utilizing a differential privacy algorithm and then executes response. According to the invention, adaptive adjustment of data acquisition granularity and safe cooperative processing of risk events are realized.

Description

Technical Field

[0001] This invention relates to the field of data processing and risk monitoring technology, specifically to a method and system for intelligent monitoring and collaborative response to performance risks. Background Technology

[0002] In the current e-commerce and supply chain management system, risk monitoring during the fulfillment process has initially shifted from passive traceability to proactive early warning by integrating sensor technology, edge computing, and artificial intelligence algorithms. However, with the increasing complexity of business operations, existing technological solutions have revealed systemic limitations in data fusion, model application efficiency, and response mechanism coordination.

[0003] At the data collection and governance level, data related to order management, logistics tracking, and customs clearance in the fulfillment chain are often scattered across heterogeneous business systems. Inconsistent device communication protocols lead to breakpoints in the data collection chain, making it difficult to locate abnormal events in a timely manner. Simultaneously, due to the lack of unified data governance standards and cleaning mechanisms, low-quality data from the physical sensing end, generated by sensor calibration deviations or environmental noise, directly flows into the processing system, resulting in increased false alarm rates and reduced accuracy in subsequent model training.

[0004] At the application level of risk analysis models, in pursuit of predictive accuracy, existing systems often employ complex deep learning or ensemble algorithms. These models typically exhibit black-box characteristics, lacking interpretable feedback on the decision-making logic. This makes it difficult for business personnel to trace the causes of risks based on the prediction results, and fails to meet the compliance requirements for decision-making transparency in some industries. Furthermore, such models have weak scenario adaptability. When the physical environment or equipment model changes, it often requires a long period of retraining, making it unable to quickly adapt to changing performance scenarios.

[0005] At the system architecture and resource configuration level, edge computing has been widely introduced to reduce transmission latency. However, the fixed-frequency data acquisition and processing strategy has led to high costs for edge node hardware deployment and maintenance. Moreover, in environments with fluctuating network signals or limited bandwidth, rigid data transmission mechanisms can easily cause delays or loss of critical monitoring information, making it difficult to achieve a balance between real-time performance and resource costs.

[0006] At the risk response level, because the business systems of departments such as procurement, logistics, and finance are independent of each other, the information flow after an alert is triggered often relies on manual intervention or non-standardized collaborative processes, resulting in a delayed response. Furthermore, in cross-departmental or cross-enterprise collaborative handling processes, the lack of robust data anonymization and privacy protection mechanisms makes it difficult for all parties to achieve efficient information sharing and coordination while ensuring data security, ultimately affecting the optimal timing for risk management. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent monitoring and collaborative response to performance risks, which solves the problems of low resource efficiency due to lack of adaptability in data collection, lack of interpretability in model prediction, low efficiency in cross-system collaborative response, and poor privacy and security in existing performance risk monitoring.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method and system for intelligent monitoring and collaborative response to performance risks.

[0009] The first aspect of this invention provides a method for intelligent monitoring and collaborative response to performance risks, which is applied to a system including edge sensing nodes, communication networks and cloud decision-making platforms.

[0010] This method collects fulfillment flow data through edge sensing nodes, calculates compressed feature data of the data within the current time window using a mean aggregation algorithm based on the currently effective sliding window parameters, and sends it to the cloud decision-making platform via a communication network. This step utilizes a dynamic window mechanism to aggregate the raw data, reducing the amount of data transmission.

[0011] The data access and governance module of the cloud decision-making platform receives compressed feature data and performs standardized governance. It uses data cleaning and fusion methods to align the fulfillment flow data with the structured order data of the business system, calculates and outputs standardized feature vectors.

[0012] During the risk calculation phase, the performance risk prediction module inputs standardized feature vectors into the trained risk prediction model and uses the model inference algorithm to calculate the performance risk probability. At the same time, in order to quantify the fluctuation of the influence of features on the prediction results, the marginal contribution value of each feature to the model output is calculated using game theory interpretation methods, and based on the marginal contribution value in the historical backtracking period, the variance statistics method is used to calculate the feature uncertainty index.

[0013] Based on the feature uncertainty index, the feedback control strategy module generates adaptive sampling control instructions. The edge sensing node receives these instructions and updates the sliding window parameters, thereby constructing a closed-loop feedback mechanism that adjusts the data acquisition granularity based on the feature uncertainty index, enabling dynamic adjustment of the edge acquisition strategy according to the model calculation results.

[0014] The collaborative response module calculates the response priority of risk events based on the probability of performance risk and the stored business rule matrix, and then performs collaborative response after calculating de-identified data for fields containing sensitive information using a differential privacy algorithm.

[0015] In one optional implementation, the process of edge sensing nodes collecting fulfillment flow data includes: establishing a connection with sensor components through a hardware interface and polling to receive raw binary data frames; calling the parsing driver protocol to decode based on the device identifier and calculating decimal values ​​using a base conversion algorithm; setting up a time synchronization buffer pool, assigning the same system acquisition timestamp to values ​​within the same polling cycle, and encapsulating them into tuple format as the basic unit for compression processing.

[0016] In one optional implementation, the data access and governance process includes: establishing a sliding observation window and using streaming statistical algorithms to calculate the arithmetic mean and sample standard deviation; using the 3-Sigma anomaly detection formula to determine whether the deviation between the current data value and the arithmetic mean is greater than three times the sample standard deviation, and if so, determining it as an outlier and repairing it using the time-series neighborhood mean filling method; fusing fulfillment flow data and structured order data through a dual-stream correlation operator; and using a weighted summation formula to calculate an integrity score, that is, accumulating the product of the business weight of the feature field and the existence indicator variable, and marking records whose scores meet the integrity threshold as valid samples.

[0017] In one optional implementation, the calculation of the performance risk probability includes: loading an ensemble learning model containing multiple regression decision trees; mapping standardized feature vectors to decision trees, and using the tree model to calculate the sum of the weights of the leaf nodes to obtain the original prediction score; and using the Logistic function formula to calculate the performance risk probability, specifically by calculating the natural exponential function value after taking the negative of the original prediction score, and then calculating the reciprocal of the sum of the original prediction score and the original prediction score.

[0018] In one optional implementation, the calculation of the characteristic uncertainty index includes: extracting the marginal contribution value sequence within the historical backtracking period; calculating the mean contribution using the arithmetic mean formula; calculating the dispersion of the sequence using the sum of squared deviations formula, i.e., accumulating the square of the difference between the marginal contribution value and the mean contribution at each time step; and dividing the sum by the number of time steps using the variance definition formula to obtain the index.

[0019] In one optional implementation, the process of generating adaptive sampling control instructions includes: using a ranking comparison algorithm to identify the feature with the largest feature uncertainty index as the target feature; using threshold comparison logic to compare the index with a stored uncertainty threshold; and when the index is higher than the threshold, generating an instruction to instruct the edge sensing node to reduce the sliding window parameter. The payload of the instruction includes a target device identifier, an identifier of the sensor channel to be adjusted, and an identifier of the control parameter type.

[0020] In one optional implementation, the process of updating the sliding window parameters includes: parsing instructions from the edge-aware node, reading the baseline window size and normalization factor; calculating the updated parameters using the exponential decay formula; calculating the ratio of the feature uncertainty index to the normalization factor using division; calculating the product of this ratio and a negative adjustment sensitivity coefficient using multiplication; calculating the exponential decay coefficient of this product using the natural exponential function; and calculating the integer part of the product of the system baseline window size and the coefficient using multiplication and rounding operations.

[0021] In one optional implementation, the calculation of response priority includes: creating a work order when the risk probability exceeds the alarm trigger threshold; counting the number of affected orders and the value of goods, and calculating the affected business scope score using a segmented mapping function; calculating the remaining time using a time difference formula and calculating the task urgency score using a time decay function; and summing the two scores by multiplying them by weight coefficients using a linear weighted summation formula.

[0022] In one alternative implementation, the data anonymization process includes: identifying sensitive fields and determining privacy budget parameters; calculating perturbation noise using a Laplace mechanism based on the privacy budget parameters and the global sensitivity of the query function; specifically, generating random noise values ​​that conform to a Laplace distribution, wherein the scale parameter is determined by dividing the global sensitivity by the privacy budget parameters; and superimposing the noise values ​​onto the original data using an additive perturbation formula.

[0023] A second aspect of the present invention provides an intelligent monitoring and collaborative response system for performance risks, the system comprising edge sensing nodes, a communication network and a cloud decision-making platform.

[0024] Edge sensing nodes are used to perform data acquisition, calculate compressed feature data using a mean aggregation algorithm based on the currently effective sliding window parameters, and dynamically adjust the sliding window parameters in response to remote commands. The communication network is used to provide data uplink transmission channels and command downlink transmission channels.

[0025] The cloud-based decision-making platform includes a data access and governance module, a performance risk prediction module, a feedback control strategy module, and a collaborative response module.

[0026] The data access and governance module is used to calculate standardized feature vectors using data cleaning and fusion methods; The performance risk prediction module is used to calculate the performance risk probability using model inference algorithms and to calculate the characteristic uncertainty index using variance statistics methods. The feedback control strategy module generates adaptive sampling control commands based on the characteristic uncertainty index; The collaborative response module is used to calculate response priorities using multidimensional weighted logic and to calculate de-identified data using differential privacy algorithms.

[0027] This invention provides a method and system for intelligent monitoring and collaborative response to performance risks. It has the following beneficial effects: 1. This invention constructs a closed-loop feedback mechanism that drives edge sampling parameter adjustment based on cloud-based feature uncertainty indicators. Uncertainty is quantified by calculating the variance of the model's marginal contribution to features, and instructions are generated accordingly to enable edge sensing nodes to dynamically update sliding window parameters using an exponential decay algorithm. This mechanism automatically increases the temporal resolution of data acquisition when the model's judgment is unstable or when external environmental fluctuations are severe, while automatically performing mean aggregation to reduce data volume when the state is stable. This effectively reduces the transmission load and storage pressure on the communication network while ensuring the ability to capture sudden performance risk events, solving the problem of balancing resource consumption and monitoring accuracy in fixed-frequency acquisition.

[0028] 2. This invention utilizes a strategy combining game theory explanations with statistical variance to enhance the transparency of risk prediction models and the reliability of control bases. By decomposing the complex inference process of the ensemble learning model into the marginal contribution values ​​of features, and further using variance statistics to calculate feature uncertainty indicators, the system not only outputs the risk probability but also quantifies the sources of fluctuation in the current prediction results. This technical feature transforms the internal state of the black-box model into a numerical indicator that can be used for process control, ensuring that subsequent adaptive sampling control commands are generated based on the model's actual cognitive state, rather than relying solely on simple threshold triggers, thus improving the system's adaptability to complex business scenarios.

[0029] 3. This invention optimizes the collaborative response efficiency and data security for risk events by combining multidimensional weighted logic with differential privacy algorithms. The system can automatically calculate response priorities based on the scope of business impact and the urgency of tasks using a linear weighted model, ensuring that high-value or high-urgency risks are addressed first. Simultaneously, during cross-departmental or cross-system collaboration, the Laplace mechanism is used to add perturbation noise to sensitive data fields, achieving privacy protection while preserving the statistical characteristics of the data. This allows the system to share risk data without disclosing specific sensitive information, supports multi-party collaborative business processes, and resolves the conflict between data sharing and privacy protection in risk management. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a schematic diagram of the overall system architecture of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see the appendix Figure 2 This invention provides a method and system for intelligent monitoring and collaborative response to performance risks. The system includes edge sensing nodes, a communication network, and a cloud decision-making platform.

[0033] Edge sensing nodes are deployed at the fulfillment operation site, including but not limited to transport vehicle terminals, warehouse environment monitors, or logistics transit station gateways. Edge sensing nodes are equipped with sensor arrays and local computing units to perform continuous acquisition of physical environment data and primary data compression processing. Edge sensing nodes have the ability to receive and respond to remote control commands and can dynamically adjust local data acquisition frequency and compression parameters.

[0034] The communication network connects edge sensing nodes with the cloud decision-making platform, providing data uplink transmission channels and command downlink transmission channels. The communication network adopts 5G network, satellite link or industrial private network.

[0035] The cloud-based decision-making platform is deployed in a server cluster or cloud computing environment. It aggregates heterogeneous data from multiple sources, executes deep learning model inference, calculates risk indicators and model uncertainty indicators, and generates and distributes adaptive sampling control commands via communication networks based on the calculation results.

[0036] Furthermore, the cloud-based decision-making platform includes a data access and governance module, a performance risk prediction module, a feedback control strategy module, and a collaborative response module. The data access and governance module is used for data cleaning and fusion. The performance risk prediction module runs machine learning models and calculates the variance of the SHAP value. The feedback control strategy module generates window adjustment parameters for edge-aware nodes. The collaborative response module is responsible for alarm distribution and privacy anonymization at the business level.

[0037] Please see the appendix Figure 1 The method, when applied to the aforementioned system, includes the following steps: S1. Edge sensing nodes continuously collect performance flow data of the monitored targets. Based on the currently active sliding window parameters, the edge sensing nodes perform mean aggregation on the performance flow data within a specified time window to generate compressed feature data. The edge sensing nodes then transmit this compressed feature data to the cloud-based decision platform via the communication network. This process achieves edge-side dimensionality reduction of the original high-frequency data.

[0038] S2, the data access and governance module of the cloud-based decision-making platform receives compressed feature data and simultaneously acquires structured order data from the business system. The data access and governance module performs standardized governance operations on the above data. These operations include outlier identification and imputation based on statistical distribution patterns, and quality score calculation based on field completeness. This step outputs a fused standardized feature vector.

[0039] S3, the performance risk prediction module of the cloud decision-making platform will input the standardized feature vector into the preset risk prediction model to calculate the performance risk probability at the current time step.

[0040] Meanwhile, the performance risk prediction module uses game theory to calculate the marginal contribution of each feature to the model output. The module extracts a sequence of marginal contribution values ​​within a preset backtracking period and calculates the variance of this sequence. This variance is defined as a feature uncertainty index, used to characterize the stability of the model's dependence on the current feature.

[0041] S4, the feedback control strategy module of the cloud decision-making platform is based on the characteristic uncertainty index to generate adaptive sampling control instructions.

[0042] The feedback control strategy module compares the feature uncertainty index with a preset threshold. When the feature uncertainty index is higher than the preset threshold, it indicates that the model's current judgment basis is unstable, and the adaptive sampling control command instructs the edge sensing nodes to reduce the sliding window parameter to improve the data temporal resolution. When the feature uncertainty index is lower than the preset threshold, it indicates that the model's judgment basis is stable, and the adaptive sampling control command instructs the edge sensing nodes to increase the sliding window parameter to reduce the transmission load.

[0043] The edge sensing node receives the adaptive sampling control command and immediately applies the updated sliding window parameters to execute the subsequent step S1. This step constructs a closed-loop feedback chain where the cognitive state of the cloud model drives the physical perception at the edge.

[0044] The collaborative response module of the S5 cloud decision-making platform calculates the response priority of risk events based on the probability of performance risk and the preset business rule matrix.

[0045] The collaborative response module generates processing orders based on response priorities. During the distribution of processing orders or external data sharing, the collaborative response module performs differential privacy processing on data fields containing sensitive information. Differential privacy processing achieves compliant data interaction by introducing Laplace noise to mask individual data characteristics.

[0046] In this embodiment of the invention, the edge sensing node in step S1 continuously collects the fulfillment flow data of the target monitored object, specifically through a pre-built protocol adaptation module and data buffer queue within the edge sensing node. To adapt to the differences in data formats of different brands and models of sensors in logistics fulfillment scenarios, step S1 further includes the following sub-steps.

[0047] S101. The edge sensing node establishes a physical connection with the sensor components through a hardware interface and performs data polling. The hardware interface includes an RS-485 bus interface, a CAN bus interface, and an analog input interface. The sensor components include a temperature and humidity probe for monitoring the status of cold chain goods, an accelerometer for monitoring vehicle transport status, and a GPS receiver for monitoring geographical location. The edge sensing node sends read commands to each sensor component connected to the bus according to a preset sampling frequency. The electrical connection method and physical signal conditioning circuit design of the sensor components are well known to those skilled in the art and will not be described in detail here.

[0048] S102. The edge sensing node receives the raw binary data frames returned by the sensor components and decodes the messages by calling the corresponding parsing driver protocol according to the device ID. For vehicle controller area network bus data, the edge sensing node adopts the SAE J1939 protocol or OBD-II protocol specification to parse the extended frame ID and data field in the data frame and extract engine speed, instantaneous fuel consumption, and vehicle speed data. For industrial fieldbus devices, the edge sensing node adopts the Modbus RTU or OPCUA protocol and verifies the legality of the data frame through CRC cyclic redundancy check. After the verification is passed, the hexadecimal value is read from the holding register or input register according to the preset point mapping table. The parsing driver protocol converts the raw hexadecimal value into a decimal value with physical dimensions.

[0049] S103. The edge sensing node performs spatiotemporal alignment and formatted encapsulation on the parsed heterogeneous data. Due to the different response delays of different sensor components, the timestamps of the data arriving at the edge sensing node may vary. The edge sensing node sets up a millisecond-level time synchronization buffer to uniformly assign the same system acquisition timestamp to decimal values ​​acquired within the same polling cycle. The edge sensing node then encapsulates the data into a standardized tuple format. ,in For the time of collection, As a unique identifier for the feature, This represents the original physical measurement value. This tuple format serves as the basic input unit for subsequent dynamic window compression processing, ensuring that the data stored in the local buffer queue has a unified data structure.

[0050] In this embodiment of the invention, the edge sensing node performs mean aggregation on the fulfillment flow data within a specified time window based on the currently effective sliding window parameters. Specifically, this involves the edge computing unit sharding and statistically compressing the time-series data in the buffer queue. This process is executed by the sliding window algorithm logic running inside the edge sensing node, aiming to reduce data transmission frequency while preserving key trend characteristics. Step S1 also includes the following sub-steps.

[0051] S104. The edge-aware node reads the window size parameter from the local configuration storage area. The window size parameter The number of raw sampling points contained in a single compression unit is defined. During the initial operation phase of the system, the edge-aware nodes use preset default values ​​(e.g., ...). The edge-aware node continuously monitors control commands from the cloud-based decision-making platform, and updates the window size parameter in its local storage area in real time when it receives a new control command. This changes the granularity of subsequent data processing.

[0052] S105, Edge-aware nodes based on window size parameter The original data in the buffer queue is sliced. When the amount of data in the buffer queue reaches... At the specified time, the edge sensing node locks onto the data segment and calculates the arithmetic mean within that segment to generate the first data segment. Mean data of each transmission window .

[0053] The calculation for this dynamic compression process is based on the following formula: ; in, This indicates the first data uploaded by the edge sensing node to the cloud. The statistical mean of each transmission window; This parameter represents the window size in effect at the current time step. It is a positive integer and its value is dynamically determined by the cloud feedback mechanism. Indicates an index in the original data sequence; Indicates the number of data collected by the sensor Index data values ​​in a single original data sequence; and They represent the first The start and end indices of each transmission window in the original data sequence.

[0054] Based on the above calculations, the edge sensing node will Each original data point is mapped to a single mean data point. This achieves dimensionality reduction processing of the data stream. For other statistical characteristics besides the mean (such as maximum and minimum values), those skilled in the art can extend the above formula for synchronous calculation according to actual business needs, and the principle is consistent with the above mean calculation logic.

[0055] S106, The edge sensing node will calculate the mean data. The data is encapsulated into an uplink communication message. The message payload includes the window sequence number, window end timestamp, and average data. The edge sensing node sends this message to the cloud decision-making platform via MQTT or HTTP protocol. After the data transmission is complete, the edge sensing node releases the processed data from the buffer queue. The memory space occupied by each original data point is allocated, and the sliding window is pushed forward to wait for the accumulation of the next batch of data.

[0056] In this embodiment of the invention, the cloud-based decision-making platform described in step S2 receives compressed feature data and simultaneously acquires structured order data from the business system. This is specifically achieved by constructing a real-time data pipeline that includes message middleware and change data capture components. Considering the differences in data generation mechanisms in the performance monitoring scenario—namely, the continuous streaming characteristics of sensor data and the discrete transactional characteristics of business order data—step S2 further includes the following sub-steps.

[0057] S201. The data access and governance module of the cloud decision-making platform starts the message middleware to establish an uplink data listening channel for edge sensing nodes. Compressed data packets sent by edge sensing nodes are routed to specific topic partitions. This message queue acts as a buffer layer for traffic smoothing and valley filling, capable of handling high-concurrency write requests from a massive number of edge devices. The data access and governance module adopts a publish / subscribe model, decoupling data production on the edge side from data consumption in the cloud. For the specific selection and cluster deployment of the message middleware, those skilled in the art can use mature components such as Kafka or Pulsar. Their cluster maintenance and replication mechanisms are well-known technologies in the field and will not be elaborated upon here.

[0058] S202. The cloud-based decision-making platform accesses business-related databases via change data capture technology. These databases include an ERP system database storing order information, a WMS system database storing cargo information, and a customs database storing customs clearance status. The data access and governance module deploys a data capture connector to monitor the underlying binary logs of the business-related databases in real time. When write operations such as order status updates, inventory deductions, or logistics node changes occur in the business system, the data capture connector captures the change event and converts it into a standard JSON format data stream containing the operation type and the values ​​before and after the change. This mechanism avoids the latency and database performance degradation caused by traditional polling methods, ensuring that structured business-level data can enter the processing system with millisecond-level latency.

[0059] The S203 cloud-based decision-making platform utilizes a streaming computing engine to unify sensor stream data from message queues and business change stream data from data capture connectors into the computing topology. The streaming computing engine is configured with multi-source data source operators to deserialize and parse heterogeneous input streams. The engine maps sensor data streams to time-series objects and business change streams to state table objects. In this way, the data access and governance module logically unifies the perceived data from the physical world and the business data from the digital world into a real-time event stream to be processed, providing a homogeneous data foundation for subsequent anomaly cleansing and fusion computing.

[0060] The data access and governance module performs statistical anomaly identification and repair on the continuous indicator data obtained from the streaming computing engine. This process aims to remove outliers caused by sensor electrical noise or network transmission jitter, ensuring the stability of subsequent model inputs. Step S2 further includes the following sub-steps.

[0061] S204. The data access and governance module establishes a sliding observation window for statistical calculations. For any specific continuous monitoring indicator (e.g., cold chain temperature or vehicle speed), the module maintains a sliding observation window of length in memory. The total number of historical data points. The module calculates the arithmetic mean of the data sequence within the observation window in real time. with sample standard deviation .

[0062] The above statistical characteristics are calculated according to the following formula: ; ; in, This represents the arithmetic mean of the data sequence within the observation window; The standard deviation of the data sequence within the observation window represents the sample standard deviation and is used to characterize the dispersion of the data. The observation window displays the total number of historical data points, a value that is preset based on the sampling frequency. Indicates the first in the observation window A historical data value.

[0063] S205, the data access and governance module uses calculated statistical characteristics to perform anomaly detection on the currently accessed real-time data points. The module substitutes the current data value to be detected into... The verification is performed in the discrimination logic.

[0064] The calculation basis for anomaly detection is the following inequality: ; If the above inequality holds, meaning the absolute value of the deviation between the current data value and the mean exceeds three times the standard deviation, the data access and governance module determines that the data point is a statistically significant outlier. This value fails to reflect the true physical state and is caused by noise interference.

[0065] S206. The data access and governance module performs time-series neighborhood mean filling repair on data points identified as abnormal. The module discards the original abnormal values ​​and retrieves the preceding and following valid data or the mean data within adjacent time windows that are adjacent to the abnormal point on the time axis to replace the value at the abnormal position. The repaired data sequence maintains temporal continuity and is input as cleaned fused data into the subsequent integrity scoring model. For non-numerical discrete state anomalies, the module performs logical verification according to the preset transition rules of the business state machine.

[0066] In this embodiment of the invention, after cleaning and repairing single-point data, the data access and governance module performs cross-domain data association and fusion, and conducts a quality quantification assessment of the fused data records. Step S2 also includes the following sub-steps.

[0067] S207, the data access and governance module constructs a multi-dimensional correlation model of orders, logistics, and inventory, aligning discrete business data with time-series sensing data in both temporal and spatial dimensions. The module uses the order's unique identifier and logistics tracking number as the primary key, and the timestamp as the secondary key, to perform a wide table construction operation. Specifically, the module maps order fulfillment requirements from the ERP system, physical attributes of goods from the WMS system, and real-time environmental parameters from edge sensing nodes to the same feature row. This process is implemented through the dual-stream Join operator or the dimension table Join operator of the streaming computing engine, thereby transforming multi-source heterogeneous data into a unified feature vector suitable for model input.

[0068] The S208 data access and governance module uses a built-in data integrity scoring model to assess the quality of the generated feature vectors. Due to network latency or missing data entry from business systems, the feature vectors may contain missing fields. To prevent low-quality data from interfering with the accuracy of subsequent risk prediction models, the module calculates an integrity score for each fused data record. .

[0069] The integrity score is calculated according to the following formula: ; in, This represents the completeness score of a single data record. The score typically ranges from 0 to 1 and is used to quantify the completeness of the information in the record. It represents the total number of data fields contained in the feature vector, covering all feature dimensions from business attributes to sensor readings; Indicates the first The business weights in each original data sequence are pre-configured by the system, reflecting the importance of that field to the risk prediction task, and the sum of the weights of all fields satisfies the normalization condition. ; An indicator variable exists for the field, used to identify the first... Does the first original data sequence have a valid value? If the original data sequence is not empty and has a valid format, then If the field is missing or empty, then .

[0070] S209, The data access and governance module calculates the integrity score. Implement a data routing strategy. The module sets a minimum integrity threshold based on the record's integrity score. When the record's integrity score is below this threshold, it is marked as unusable data and routed to a queue awaiting manual verification; A record is marked as a valid sample if its value is above or equal to this threshold. Only data records marked as valid samples are allowed to be input into the cloud-based early warning model for risk probability calculation. Through this mechanism, the system achieves pre-emptive quality control based on information entropy contribution during the data fusion phase.

[0071] In this embodiment of the invention, after receiving the cleaned and fused effective feature vectors, the cloud-based decision platform initiates a risk prediction and uncertainty quantification process. As the core computational unit of this process, the performance risk prediction module uses a gradient boosting decision tree algorithm to construct a classifier and estimate the probability of performance anomalies (such as goods spoilage, transportation delays, and compliance inspection failures) that may occur at the current supply chain node. Step S3 further includes the following sub-steps.

[0072] S301, The performance risk prediction module loads the pre-trained ensemble learning model and receives the feature vector to be predicted. This feature vector The data is derived from the wide table constructed in the previous embodiment and includes multi-dimensional input variables, specifically covering environmental perception features (average temperature, vibration amplitude), business constraint features (remaining fulfillment time, customs clearance priority), and static attribute features (goods type, packaging grade). The model will convert the feature vectors... In the combined structure of multiple regression decision trees, each decision tree corresponds to a specific feature partitioning logic and leaf node weights.

[0073] S302, the performance risk prediction module performs inference calculations using an additive model. The model obtains the original predicted value of the sample by accumulating the scores of all tree nodes in the decision tree. Unlike the binary judgment of a single decision tree, this ensemble model iteratively optimizes the objective function so that each newly generated decision tree is used to fit the residual of the previous round of predictions.

[0074] The model's raw output The calculation process follows the formula below: ; in, The model represents the input feature vector The original output; This represents the total number of decision trees in the ensemble model, and this value determines the model's complexity and fitting ability. Indicates the first The weight value of the corresponding leaf node in the tree reflects the contribution of the leaf node to the final prediction result; Indicates the first The leaf node region structure of the tree defines the rules for partitioning the feature space; This is an indicator function used to determine the eigenvectors. Whether it falls into the first Specific leaf node regions of a tree If the rules of the tree determine that a sample falls into the region, the function value is 1; otherwise, it is 0.

[0075] S303, the performance risk prediction module maps the original prediction score to a standardized probability of risk occurrence. Because the original output... The value range of is in the real number field, which is not convenient for direct use in business decision-making. The module uses the Logistic function to perform a nonlinear transformation on it.

[0076] Probability of risk occurrence The calculation follows the formula below: ; in, This represents the probability of a risk event (such as default or damage) occurring during the current performance of the contract, and its value is strictly limited to the range of (0,1). is the base of the natural logarithm.

[0077] Through the above steps, the system maps complex heterogeneous features into a single-dimensional quantified probability index. Subsequent decision-making logic will directly determine whether to trigger the early warning process based on this probability value. For the specific training process and hyperparameter tuning (such as learning rate, maximum tree depth, etc.) of the XGBoost algorithm, those skilled in the art can refer to the relevant open-source framework documentation. Its model training method is a well-known technology in this field and will not be elaborated here.

[0078] To address the black-box nature of ensemble learning models in the decision-making process and to provide foundational data for subsequent uncertainty quantification, the performance risk prediction module further incorporates the SHAP attribution analysis mechanism. The system not only outputs a single risk probability but also calculates the marginal contribution of each feature to the prediction result for each predicted sample. Step S3 also includes the following sub-steps.

[0079] S304. The performance risk prediction module initializes the tree-based interpreter. For the XGBoost model structure constructed in the preceding steps, the module calculates the average marginal contribution of each feature by traversing the decision tree path. The system loads the trained tree structure file and constructs an interpreter object in memory, which can accept feature vectors of the same dimension as the model input.

[0080] S305, the performance risk prediction module performs attribution calculations on the input feature vector at the current moment. For each input feature... (For example, current temperature, vehicle speed, and remaining order time), the system calculates its Shapley value in the current prediction instance.

[0081] In this embodiment, the symbol is defined. For at any time feature The contribution value to the model output. This value characterizes the feature. The degree to which the actual observed value, relative to the baseline expected value of the feature, raises or lowers the model's predicted output (Log-odds).

[0082] like This indicates that the feature increases the probability of identifying performance risk at the current moment (e.g., excessively high temperature increases risk). like This indicates that the feature reduces the probability of determining performance risk at the current moment (e.g., sufficient remaining time reduces the risk of default).

[0083] For specific Shapley value calculation algorithms, i.e. how to allocate the predicted total payoff through the marginal contribution principle in combinatorial game theory, those skilled in the art can refer to the standard open-source implementation library of the SHAP algorithm or relevant classic literature. Its calculation principle is a well-known technology in this field and will not be elaborated here.

[0084] S306, The performance risk prediction module outputs a feature contribution matrix. After calculation, the system transforms the original input feature vector into the corresponding contribution vector. This vector not only demonstrates the dominant factors of current risk to business personnel, but more importantly, it serves as a core intermediate variable for subsequent feature uncertainty calculations. It is cached in the time-series database to support subsequent quantitative analysis of the stability of the model's perceived state. Through this step, the system decouples the complex nonlinear model decision logic into the linear superposition contribution of each feature, achieving a dimensional progression from prediction to interpretation.

[0085] After obtaining the single-point SHAP contribution values ​​of each feature in the aforementioned steps, the performance risk prediction module further performs an uncertainty measurement based on time-series volatility. This process aims to identify features that have an unstable impact on the model output, i.e., scenarios where the model's dependence on a certain feature changes drastically within a short period of time. Such drastic changes usually indicate that the environment is in a non-steady state or that there are potential abnormal disturbances; therefore, the system defines this as feature uncertainty. Step S3 further includes the following sub-steps.

[0086] S307, Performance Risk Prediction Module: Based on the characteristics of the target Extracting the most recent data from the time series database Observe the SHAP contribution value sequence during the backtracking period. The system sets the observation backtracking period. As a sliding window, the length of this window determines the sensitivity to uncertainty assessment. The data sequence within the window includes data from time [time value missing]. Multiple consecutive contribution sampling points moving forward.

[0087] S308, the performance risk prediction module uses a statistical variance algorithm to calculate the characteristic uncertainty index. The system quantifies the fluctuation range of the characteristic contribution value by calculating the dispersion of the sequence.

[0088] Characteristic uncertainty index The calculation follows the formula below: ; in, Representation of features The feature uncertainty index, the larger the value, the more unstable the contribution of the feature to the risk prediction result is in the time dimension; This represents the variance calculation function; This indicates the observation backtracking period, which is the number of past time steps involved in the calculation; Index variable representing time backtracking; Representation of features At any moment The SHAP value (i.e., the marginal contribution value calculated in the previous embodiment); Representation of features The average SHAP contribution during this backtracking period.

[0089] S309, The performance risk prediction module will calculate the results. The value is output to the feedback control strategy module. In this step, the system establishes a mapping relationship from numerical fluctuations to uncertainty semantics. When When the value increases, it indicates that the cloud model has detected that the weight of this feature (such as vehicle location or warehouse temperature) in risk assessment fluctuates wildly within a short period of time. This oscillation is identified by the system as a precursor to monitoring blind spots or sudden environmental changes. The system identifies among all current features... The feature with the largest variance value is selected as the target feature, and this maximum variance value is labeled as... This serves as the core basis for subsequently triggering adjustments to the sampling frequency at the edge. Through this mechanism, the system not only perceives the level of risk but also the degree of uncertainty regarding the basis for risk judgment, thereby achieving meta-cognition of monitoring quality.

[0090] In this embodiment of the invention, after completing the quantification calculation of feature uncertainty in the cloud, the system constructs a reverse feedback control loop to transform the high-level model perception state into low-level device control commands. This process realizes the transformation from pure data monitoring to proactive perception, ensuring that the computing resource configuration at the edge always matches the current environmental risk perception needs. Step S4 specifically includes the following sub-steps.

[0091] S401, the feedback control module generates dynamic adjustment commands. The cloud server monitors in real time the characteristic uncertainty indicators calculated from the preceding steps. When this indicator suggests a cognitive bias in the model regarding specific characteristics of the currently monitored object (such as the temperature distribution or vibration frequency of a cold chain vehicle), the system automatically constructs a downlink control message. This control message is encapsulated in a lightweight JSON data format or binary protocol, and its payload includes a unique identifier for the target edge device, the sensor channel ID to be adjusted, and the corresponding control parameter type identifier. During this process, the feedback control module maps abstract uncertainty values ​​to specific sampling granularity requirements, establishing an inverse correlation between risk uncertainty and data acquisition frequency. That is, for feature dimensions with high uncertainty, the system determines that denser data points are needed to eliminate cognitive ambiguity.

[0092] S402. Commands are issued via message middleware. The cloud establishes a persistent connection with the edge gateway using the MQTT or CoAP protocol. The system publishes the generated adjustment commands to specific topics subscribed to by the devices. To ensure the reachability of commands in weak network environments, a QoS level mechanism is configured in the issuance process. For sampling adjustment commands involving core risk characteristics, the system marks them as QoS2 level to ensure accurate one-time delivery of messages and prevent command loss or duplicate execution due to network jitter.

[0093] S403. Edge sensing nodes parse commands and lock configuration registers. After receiving a control message pushed from the cloud, the edge gateway deployed on the transport vehicle or warehouse controller starts the command parser to unpack and verify the message. The gateway extracts the specified sensor address and parameter update request from the message and suspends the current data acquisition thread. At this time, the edge system enters the configuration update state, locating the current sampling window size parameter stored in memory. The configuration register or variable handle is then used. This step prepares the physical device for responding to the cloud logic, laying the execution foundation for subsequently overwriting the original sampling parameters with specific algorithmic values. Through these steps, the system establishes a closed-loop link between cloud computing, network transmission, and edge execution, enabling the perception strategy of the physical world to be dynamically orchestrated in real time according to the model state of the digital world.

[0094] After parsing and acquiring the characteristic uncertainty indicators sent from the cloud, the edge device does not simply switch preset levels. Instead, it dynamically generates sampling window parameters adapted to the current environmental fluctuations through the calculation logic of continuous functions. This mechanism establishes a nonlinear mapping relationship between model cognitive anisotropy and data acquisition granularity, ensuring that information capture density is automatically increased when risk characteristics are ambiguous. Step S4 further includes the following sub-steps.

[0095] S404: The edge gateway reads the reference parameters and performs normalization. The device retrieves the preset system reference window size from its local non-volatile memory. and normalization factor Among them, the system reference window size Represents the standard compression granularity under steady-state environmental conditions (i.e., when the model has high determinism in feature determination), aiming to minimize data transmission bandwidth; normalization factor. The value is usually taken as the maximum variance value in the historical observation record of this feature, which is used to map the current real-time fluctuation value to the standard dimension interval of [0,1] to eliminate the influence of different physical quantities (such as temperature in degrees Celsius and vibration in Hertz number) on the calculation process.

[0096] S405: The edge gateway calculates the new sampling window size based on the exponential decay model. The system processor performs floating-point operations based on the current target feature uncertainty variance value. Calculate the updated window parameters .

[0097] The calculation formula is as follows: ; in, This represents the updated sampling compression window parameter, which determines how many original acquisition points will be aggregated into a transmission unit at the edge. This represents the system's baseline window size, serving as an initial reference value for calculations. This represents the variance of the uncertainty of the target characteristics at the current moment. Indicates the normalization factor; This represents the adjustment sensitivity coefficient, and This coefficient is used to control the decay rate of the window size as uncertainty changes. The larger the value, the more violently the system reacts to uncertainty, and the faster the window shrinks; Represented by natural constant An exponential function with base 0.

[0098] S406, The edge gateway updates the configuration parameters of the data aggregation module. The calculated... The value is usually a floating-point number. The system rounds it down to the nearest positive integer and writes it into the control variable of the data acquisition driver. According to the formula logic, when the target feature uncertainty variance value... As the exponent increases, the absolute value of the negative value of the exponent part increases, leading to... The function value decreases, thus making the final... Value less than the baseline window size This means that the number of raw data points contained in each transmission unit is reduced, and the frequency at which the edge reports data to the cloud (i.e., the frequency of window sliding) increases accordingly, provided that the original sampling frequency remains unchanged. This mechanism achieves focus, that is, when the model considers the environmental state to be difficult to determine, it automatically provides richer and finer-grained temporal feature support through high-frequency sampling, thereby helping the model eliminate uncertainty in subsequent time steps, while automatically reverting to low-frequency sampling to save resources when the environmental state is clear.

[0099] After the cloud-based model determines that the probability of performance risk exceeds a preset security threshold, the collaborative response module does not immediately perform equal processing on all alarms. Instead, it introduces a multi-dimensional weighted scoring mechanism to quantify and rank concurrent risk events. This process aims to address the issue of limited operational resources in high-concurrency scenarios, ensuring that system resources are prioritized for tasks with the greatest impact on business and the most urgent processing time requirements. Step S5 specifically includes the following sub-steps.

[0100] S501, the collaborative response module initiates the risk assessment trigger logic. The system monitors the probability of risk occurrence output from previous steps in real time. When a logistics node or vehicle is detected... When the value exceeds the system's configured alarm threshold (e.g., 0.75), a risk-pending work order is automatically created, and the context data of that work order is initialized. The system uses a unique index key to query the business information associated with the risk entity, including but not limited to the associated order number, goods type, consignee level, and contractually agreed delivery deadline.

[0101] S502, the collaborative response module quantifies the scope and urgency of business impact. To transform unstructured business urgency into calculable values, the system calculates a score for the affected business scope. Task urgency rating .

[0102] for The system iterates through the associated order database to count the total number of downstream orders and the total value of goods affected by the risk node. The system uses a segmented mapping function to map order amounts or quantities to standard scores from 0 to 100. The higher the score, the greater the economic compensation or customer complaints that would result if a substantial breach of contract were to occur.

[0103] for The system calculates the remaining time until the contract delivery deadline. The system uses an inverse proportional function or a time decay function to... The score is calculated based on the remaining time; the shorter the time, the higher the score. The higher the score, the more immediate intervention is required for the task; otherwise, irreversible overdue defaults will occur.

[0104] S503, the collaborative response module calculates the overall early warning processing priority. The system synthesizes the final priority score through a linear weighted model to determine the order of the work order in the operation and maintenance processing queue.

[0105] Early warning handling priority scoring The calculation follows the formula below: ; in, This indicates the priority score for handling early warnings. The higher the score, the more likely the system will place the risk event at the front of the processing queue. The score indicates the scope of affected business operations and is used to characterize the severity of the risk after it occurs. The score indicates the urgency of the task and is used to characterize the timeliness requirements for processing. Normalized weighting coefficients representing the scope of influence; Normalized weighting coefficients representing the degree of urgency; Under normal circumstances System administrators can flexibly adjust the weight allocation according to business strategies.

[0106] Through this calculation process, the system compares high-value but time-available risk events with low-value but soon-to-be-overdue risk events under a unified dimension, generating a dynamic priority list that guides the automated scheduling system or human customer service to execute intervention measures in the optimal order.

[0107] To ensure the security of data flow across departments or networks, the system does not directly transmit raw data containing precise geographical locations or specific transaction amounts when executing collaborative responses. Instead, it introduces differential privacy technology to perform real-time anonymization of sensitive fields. This mechanism preserves the statistical characteristics of the data (such as overall distribution trends) to meet macro-analysis needs, while using mathematical methods to mask precise individual information and prevent reverse engineering attacks. Step S5 also includes the following sub-steps.

[0108] S504, the collaborative response module identifies sensitive data fields and determines the privacy budget. The system first scans the upcoming warning work order data packet, identifying key privacy fields based on a pre-defined sensitivity level classification table. Typical sensitive fields include, but are not limited to: real-time GPS coordinates (latitude and longitude data) of vehicles or goods, specific order amounts for high-value goods, and customer contact information. For different types of sensitive fields, the system allocates corresponding privacy budget parameters. This parameter determines the balance between the strength of privacy protection and data availability: The smaller the value, the greater the added noise and the higher the level of privacy protection, but the accuracy of the data decreases accordingly.

[0109] The S505 collaborative response module generates perturbation noise based on the Laplace mechanism. After determining the target value to be desensitized, the system calculates the noise component to be added based on the global sensitivity and privacy budget of the field. The system uses the Laplace mechanism for implementation. Differential privacy protection ensures that the presence or absence of any single record in the output will not change the probability distribution of the output.

[0110] Desensitized data values The calculation formula is as follows: ; in, This indicates the data value after anonymization, which will be used for subsequent cross-network transmission or storage; This refers to raw, sensitive data containing real information (such as actual GPS longitude or order amount). The position parameter is Scale parameters are The Laplace probability distribution function; here, the position parameter Setting it to 0 ensures that the expected value of the noise is 0, thus maintaining the unbiasedness of the data; This indicates the global sensitivity of the query function, which is the maximum change in the function's output value when any record is deleted or added to the dataset; for the direct transfer of a single numeric record, this value is usually the maximum range of data that can be changed. This represents the privacy budget parameter, which controls the upper limit of the risk of privacy leaks.

[0111] S506, the collaborative response module performs noise superposition and data reconstruction. The processor calls the random number generator to generate specific noise values ​​based on the Laplace distribution described above, and then adds these noise values ​​to the original data in an additive manner. Up. The system then used the generated The corresponding fields in the original data packet are replaced to generate a de-identified secure data copy. This copy is then encrypted and sent to an external monitoring platform or a third-party logistics service provider. Due to the introduction of noise, attackers cannot accurately deduce the original vehicle location or cargo value by analyzing the intercepted data. However, when performing regional heat map analysis or macro-level freight statistics, since the mathematical expectation of noise is zero, the aggregation results still remain within the statistically valid range, thus achieving a dual guarantee of data privacy protection and business utility.

[0112] In this embodiment of the invention, the computer device can be embodied as an edge computing gateway or vehicle-mounted terminal deployed at a logistics site, or a cloud server or server cluster deployed in a control center. The computer device mainly includes a memory, a processor, and a network interface connected via a system bus.

[0113] The memory, as a computer-readable storage medium, is used to store the operating system, network communication protocol stack, and computer program instructions for implementing the edge-cloud collaborative monitoring method based on model uncertainty feedback in the above embodiments of the present invention. The memory consists of volatile memory (such as high-speed random access memory, RAM) and non-volatile memory (such as read-only memory, flash memory, or disk). The non-volatile memory is used to permanently store the aforementioned logical instructions and historical configuration parameters; the volatile memory provides high-speed cache space for data processing during processor operation, specifically for temporarily storing real-time streaming data queues within the sliding window, dynamically constructed SHAP value matrices, and intermediate calculation results. The memory is functionally divided into multiple logical areas, including a data acquisition and preprocessing area, a model inference area, an uncertainty calculation area, and a feedback control area.

[0114] The processor is the core of a computer device, consisting of one or more central processing units, graphics processing units, or dedicated AI acceleration chips. The processor reads the computer program temporarily stored in memory via the system bus and executes a series of specific arithmetic and logical operations to implement the technical solution of this invention. The specific execution logic is as follows: Adaptive compression of edge data is performed. The processor acquires continuous raw time-series data from the input interface and reads the currently active sampling window parameters from the register. Based on the data aggregation logic defined in the preceding embodiments, the processor sums and averages multiple raw sampling points within a specified window length, thereby converting the high-frequency physical sensing signal into a low-frequency mean feature vector. This process effectively reduces the data transmission dimensionality while preserving the core statistical characteristics within the time window.

[0115] The processor performs risk prediction and uncertainty quantification. It inputs the processed feature vectors into a pre-loaded ensemble learning model for inference, calculating the probability of performance risk in the current state. Simultaneously, the processor extracts the marginal contribution value of each feature at the current time step by parsing the model structure. The processor further invokes statistical operation instructions to calculate the variance of the SHAP value sequence within a specific backtracking period. This variance calculation process quantifies the degree of fluctuation in the model's perception of current environmental features, i.e., the feature uncertainty index defined in the aforementioned embodiment.

[0116] The processor performs closed-loop feedback control and security response. It compares the calculated uncertainty index with preset benchmark parameters. Based on the exponential decay control model constructed in the aforementioned embodiments, the processor calculates a new sampling window value. This calculation logic is as follows: when the uncertainty index increases, the processor outputs a smaller window parameter value to increase the sampling frequency; conversely, it increases the window parameter value. Furthermore, when cross-node data transmission is involved, the processor, based on differential privacy protection logic, generates noisy data conforming to a Laplace distribution through a random number generation algorithm and superimposes it onto the original sensitive fields, completing the desensitization process while ensuring the usability of the data's statistical characteristics.

[0117] The network interface provides a data transmission channel between computer devices and external communication networks (such as 5G cellular networks, industrial Ethernet, and satellite communication links). When the computer device is operating in edge mode, the network interface is responsible for receiving signals from the sensor array and sending compressed data packets to the cloud; when the computer device is operating in cloud mode, the network interface is responsible for broadcasting sampling control commands to each edge node and sending early warning information containing de-identified data to relevant business systems.

[0118] For the common components included in this computer device, such as the power management module, cooling system, and human-machine interface, those skilled in the art can make conventional selections and configurations according to actual application scenarios (such as vehicle environments or computer room environments). The specific implementation does not affect the execution of the core algorithm logic of this invention, and will not be elaborated here. Through the collaborative work of the above hardware architecture, the device can transform the physical world's logistics data flow into a computable model in digital space, and adjust the sampling behavior of the physical device in real time based on the calculation results, realizing adaptive closed-loop control between the edge and the cloud.

[0119] In this embodiment of the invention, the computer-readable storage medium can be any non-volatile or volatile medium capable of storing program code, including but not limited to a Universal Serial Bus flash drive, a portable hard drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk. Computer program instructions are stored in this computer-readable storage medium. When these computer program instructions are read and executed by a processor, the computer system implements the edge-cloud collaborative monitoring method based on model uncertainty feedback as described in the above embodiments of the invention. The specific execution steps are as follows: The process involves windowed aggregation and preprocessing of the data stream. Instructions in the storage medium first direct the processor to initialize sampling window parameters and establish a loop-based reading mechanism to acquire time-series data sent by the sensor. The instructions contain the logic for executing the data aggregation algorithm, which performs arithmetic averaging on discrete high-frequency sampling points within the current window length to generate feature vectors. The instructions also include logical judgment code to check the timestamp continuity of the data stream, ensuring that the data input to the model conforms to the temporal sequence logic.

[0120] This system performs a dual assessment of risk and uncertainty based on attribution analysis. The instructions call upon a pre-stored ensemble learning model file to perform inference operations on feature vectors, deriving a predicted performance risk value. Crucially, the program instructions also include an interpretive computational logic for real-time calculation of the marginal contribution of input features to the model's output. The program code further instructs the processor to perform variance calculations on the SHAP value sequence within a preset historical period. Through this calculation, the computer can quantify the volatility of the current feature's impact on the model's prediction results, thus obtaining a feature uncertainty index, rather than just a single risk probability output.

[0121] This section implements adaptive closed-loop adjustment of the sampling strategy. The instructions in this section implement the closed-loop feedback control logic. The program code reads the uncertainty index and substitutes it into the preset exponential decay control algorithm. The instructions instruct the processor to compare the difference between the current uncertainty and the baseline value: when an increase in the variance of feature attribution is detected (i.e., the model's judgment of environmental changes tends to be unstable), the program automatically calculates a smaller window parameter value, thereby instructing the hardware interface to increase the data acquisition frequency; conversely, when feature attribution is stable, the program calculates a larger window parameter value to reduce resource consumption. This instruction implements inverse control from data calculation to physical acquisition.

[0122] Perform differential privacy masking and collaborative response. When the program determines that it needs to send warning data to an external system or perform cross-node transmission, the instruction triggers a privacy protection subroutine. This subroutine contains random number generation logic to generate noise values ​​that conform to a specific probability distribution (such as a Laplace distribution) based on preset privacy budget parameters. The instruction instructs the processor to perform an addition operation, superimposing the generated noise value onto sensitive fields such as geographical location and amount to generate a masked data packet. Subsequently, the program instruction controls the network interface to send the masked data packet to the specified target address.

[0123] For the specific file system format and data reading protocol of the computer-readable storage medium, those skilled in the art can refer to general IEEE standards or industry specifications, which are well-known technologies in the field and will not be elaborated further here. This storage medium, by storing the aforementioned specific computer program instructions, enables general-purpose computing hardware to be reconfigured into a dedicated device with adaptive monitoring and privacy protection functions.

[0124] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0125] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0126] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0127] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0128] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0129] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0130] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0131] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent monitoring and collaborative response to performance risks, characterized in that, Includes the following steps: The system continuously collects performance flow data of the target monitored object through edge sensing nodes, and calculates the compressed feature data of the performance flow data within the current time window using the mean aggregation algorithm based on the currently effective sliding window parameters, and sends it to the cloud decision platform through the communication network. The data access and governance module of the cloud decision-making platform receives the compressed feature data and performs standardized governance, and calculates the standardized feature vector using data cleaning and fusion methods. The performance risk prediction module of the cloud decision-making platform inputs the standardized feature vector into the trained risk prediction model, calculates the performance risk probability using the model inference algorithm, calculates the marginal contribution value of each feature to the model output using the game theory interpretation method, and calculates the feature uncertainty index based on the marginal contribution value using the variance statistics method. The feedback control strategy module of the cloud decision-making platform generates an adaptive sampling control instruction based on the feature uncertainty index. The edge sensing node receives the adaptive sampling control instruction and updates the sliding window parameters, thus constructing a closed-loop feedback link driven by the cognitive state of the cloud model to the physical perception of the edge. The collaborative response module of the cloud-based decision-making platform calculates the response priority of risk events based on the performance risk probability and the stored business rule matrix using multidimensional weighted logic, and performs collaborative response after calculating de-identified data for fields containing sensitive information using a differential privacy algorithm.

2. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The edge sensing node continuously collects the performance flow data of the target monitored object, and the performance flow data includes: The edge sensing node establishes a physical connection with the sensor component through a hardware interface and performs data polling, receiving the raw binary data frames returned by the sensor component; The original binary data frame is decoded by calling the corresponding parsing driver protocol based on the device identifier, and a decimal value with physical dimensions is calculated using a number system conversion algorithm. A time synchronization buffer pool is set up, and the decimal values ​​obtained within the same polling cycle are assigned the same system acquisition timestamp and encapsulated into a standardized tuple format as the basic input unit for dynamic window compression processing.

3. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The data access and governance module receives the compressed feature data and performs standardization governance, including: A sliding observation window is established for statistical calculation, and the arithmetic mean and sample standard deviation of the data sequence within the sliding observation window are calculated using a streaming statistical algorithm. The 3-Sigma anomaly detection formula is used to calculate whether the current data value is an anomaly. If the absolute value of the deviation between the current data value and the arithmetic mean is greater than three times the sample standard deviation, the current data point is determined to be an anomaly, and the time-series neighborhood mean filling method is used to calculate the repaired value. The fulfillment flow data is aligned and integrated with the structured order data of the business system through a dual-flow association operator; The integrity score of the fused feature vector is calculated using a weighted summation formula. The calculation process is as follows: the business weight of each feature field is calculated by multiplication and the product of the existence indicator variable is calculated by multiplication, and the sum of all products is calculated by accumulation as the integrity score. Records with integrity scores higher than or equal to the integrity threshold are marked as valid samples and used as the standardized feature vector.

4. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The calculation of the performance risk probability using the model inference algorithm includes: Load the trained ensemble learning model, which contains a combined structure of multiple regression decision trees; The standardized feature vectors are mapped to the multiple regression decision trees, and the original prediction score is calculated using the tree model accumulation logic. The calculation process involves accumulating the weight values ​​of all leaf nodes of the decision trees. The performance risk probability is calculated using the Logistic function formula. The calculation process is as follows: the natural exponential function value after the original predicted score is negative is calculated using exponential operation, and the reciprocal operation is used to calculate the reciprocal of the sum of the original and natural exponential function values ​​to obtain the performance risk probability with a value in the range of 0 to 1.

5. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The method of calculating the characteristic uncertainty index using variance statistics includes: Extract the marginal contribution values ​​within the historical backtracking period to form a sequence; The average contribution of the sequence within the historical backtracking period is calculated using the arithmetic mean formula. The dispersion of the sequence is calculated using the sum of squared deviations formula. The calculation process is as follows: the difference between the marginal contribution value and the mean contribution value at each time step is calculated using subtraction, the square of the difference is calculated using squaring, and the sum of the squares of the difference is calculated using accumulation. The characteristic uncertainty index is calculated using the variance definition formula. The calculation process involves dividing the sum by the number of time steps contained in the historical backtracking period.

6. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The feedback control strategy module of the cloud-based decision-making platform generates adaptive sampling control instructions based on the characteristic uncertainty index. The generation of these adaptive sampling control instructions includes: The feature with the highest uncertainty index among the current features is calculated and identified using a ranking comparison algorithm as the target feature. The judgment result is calculated using threshold comparison logic, and the feature uncertainty index of the target feature is compared with the stored uncertainty threshold. When the feature uncertainty index is higher than the uncertainty threshold, an adaptive sampling control command is generated to instruct the edge sensing node to reduce the sliding window parameter; The payload of the adaptive sampling control command includes a unique identifier for the target edge device, the identifier of the sensor channel to be adjusted, and the corresponding control parameter type identifier.

7. The method for intelligent monitoring and collaborative response to performance risks according to claim 6, characterized in that, The edge sensing node receives the adaptive sampling control command and updates the sliding window parameters. Updating the sliding window parameters specifically includes: The edge-aware node parses the adaptive sampling control command and reads the system baseline window size and normalization factor stored locally. The updated sliding window parameters are calculated using the exponential decay formula. The calculation process includes: The ratio of the characteristic uncertainty index to the normalization factor is calculated using a division operation; The product of the ratio and the negative adjustment sensitivity coefficient is calculated using multiplication. The exponential decay coefficient of the product is calculated using the natural exponential function; The integer part of the product of the system baseline window size and the exponential decay coefficient is calculated using multiplication and rounding operations and used as the updated sliding window parameter.

8. The method for intelligent monitoring and collaborative response to performance risks according to claim 1, characterized in that, The calculation of the response priority of risk events using multidimensional weighted logic includes: When the probability of performance risk exceeds the alarm trigger threshold, a risk pending work order is created; the total number of downstream orders and the total value of goods affected by the risk node are counted, and the score of the affected business scope is calculated using a segmented mapping function; The remaining time from the current moment to the contract delivery deadline is calculated using the time difference formula, and the task urgency score is calculated using the time decay function. The response priority is calculated using a linear weighted summation formula. The calculation process is as follows: the product of the affected business scope score and the scope weight coefficient is calculated using multiplication, and the product of the task urgency score and the urgency weight coefficient is calculated using addition.

9. The intelligent monitoring and collaborative response method for performance risk according to claim 1, characterized in that, The step of calculating de-identified data using a differential privacy algorithm includes: Identify sensitive data fields in the data to be transmitted and determine the allocated privacy budget parameters; Based on the privacy budget parameters and the global sensitivity of the query function, the perturbation noise is calculated using the Laplace mechanism; The process of calculating the disturbance noise is as follows: the random noise value that conforms to the privacy budget parameter is calculated using the Laplace distribution probability density function, wherein the location parameter is set to zero and the scale parameter is determined by dividing the global sensitivity by the privacy budget parameter; The desensitized data value is calculated using the additive perturbation formula. The calculation process involves superimposing the random noise value onto the original data of the sensitive data field.

10. An intelligent monitoring and collaborative response system applied to the intelligent monitoring and collaborative response method for performance risk as described in any one of claims 1-9, characterized in that, This includes edge sensing nodes, communication networks, and cloud-based decision-making platforms. The edge sensing node is used to continuously collect physical environment data, calculate compressed feature data using the mean aggregation algorithm based on the currently effective sliding window parameters, and dynamically adjust the sliding window parameters in response to remote commands. The communication network is used to provide an uplink data transmission channel and a downlink command transmission channel between the edge sensing node and the cloud decision-making platform; The cloud-based decision-making platform includes a data access and governance module, a performance risk prediction module, a feedback control strategy module, and a collaborative response module. The data access and governance module is used to calculate standardized feature vectors using data cleaning and fusion methods; The performance risk prediction module is used to calculate the performance risk probability using model inference algorithms and to calculate the characteristic uncertainty index using variance statistics methods. The feedback control strategy module is used to generate adaptive sampling control instructions for the edge sensing node based on the feature uncertainty index. The collaborative response module is used to calculate the response priority of risk events using multidimensional weighted logic and to calculate de-identified data using differential privacy algorithms.

Images