Intelligent monitoring and early warning system based on genetic algorithm

Abstract

The invention relates to the technical field of intelligent monitoring and early warning, in particular to an intelligent monitoring and early warning system based on a genetic algorithm, which comprises the following steps: receiving video, current, vibration and other multi-dimensional monitoring signals, forming an original data cluster, and deconstructing the original data cluster into a structured semantic unit and an unstructured mode fragment; blind source separation and feature reconstruction are carried out on the unstructured fragments to generate a physical situation feature map, and knowledge map embedding and relation analysis are carried out on the structured units to generate a logic event feature chain. Performing co-evolution analysis on the two types of features through a genetic algorithm to generate an incidence matrix; and the system drives traceability exploration according to the matrix, locates an initial abnormal source and a propagation link, and forms a traceability report. And automatically generating and issuing an instruction for regulating and controlling the sampling frequency and the alarm threshold value of the front-end equipment according to the report. According to the method, deep fusion of multi-source heterogeneous information and closed-loop adaptive optimization of system parameters are realized, and the accuracy of complex anomaly early warning and the overall operation and maintenance efficiency of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and early warning technology, and in particular to an intelligent monitoring and early warning system based on genetic algorithms. Background Technology

[0002] Existing intelligent monitoring systems generally employ multiple independent technologies to process monitoring data from different sources. Video analytics, sensor alarms, and log parsing typically operate independently, each forming judgments based on specific data types. However, data from monitoring networks is multidimensional and heterogeneous, containing both continuous analog signals from sensors and structured text reports generated by the system. Existing methods fail to effectively achieve deep integration of the physical situation information and logical event information represented by these two types of data. This separate processing model relies on pre-defined simple rules or experience to establish correlations, failing to automatically uncover and characterize the complex, non-linear intrinsic relationships between the two types of information. This results in limited system capabilities for identifying complex anomalies caused by multi-factor coupling, and early warning analysis remains at the level of isolated phenomena.

[0003] Another drawback of existing technical solutions lies in their static and open-loop characteristics. Core parameters of monitoring systems, such as device sampling frequency and alarm thresholds, are typically set during deployment and fixed for a long period. When an anomaly is detected, the system only issues an alarm, lacking the ability to optimize its data collection and judgment mechanisms based on anomaly analysis results. Faced with dynamic changes in the operating environment and device status, fixed parameters easily lead to false alarms or missed alarms, preventing continuous improvement in overall system performance. A technical solution is needed that can automatically fuse heterogeneous information to discover deep correlations and thereby achieve closed-loop adaptive control of system parameters. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent monitoring and early warning system based on genetic algorithms.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent monitoring and early warning system based on genetic algorithms, comprising: The signal receiving module receives multi-dimensional monitoring signals collected from the monitoring network, forming a raw monitoring data cluster. This raw monitoring data cluster at least covers video image streams, equipment current waveform sequences, and environmental vibration maps. The signal deconstruction module performs multi-level signal deconstruction operations, decomposing the original monitoring data cluster into structured semantic units and unstructured pattern fragments. The structured semantic units come from the equipment operation report, and the unstructured pattern fragments come from the analog signals continuously collected by the sensors. The feature parsing and generation module performs blind source separation and feature reconstruction on the unstructured pattern fragments to generate a parsed physical situation feature map. At the same time, it performs knowledge graph embedding and relationship parsing on the structured semantic units to generate a parsed logical event feature chain. The co-evolutionary analysis module performs co-evolutionary analysis on the physical situation feature map and the logical event feature chain using a genetic algorithm architecture to generate an association matrix; The anomaly tracing and investigation module, based on the correlation matrix, drives the tracing and investigation mechanism to locate the initial abnormal disturbance source and its propagation link in the monitoring signal, and generates a tracing and investigation report. The control instruction generation module automatically derives and issues a set of control instructions for the sampling frequency and alarm threshold of the monitoring equipment based on the content of the source tracing report.

[0006] Preferably, the multi-level signal deconstruction operation includes: Initialize the signal deconstruction engine, which is configured with a video stream decoder, a text segmenter, and an analog signal sampler; The video stream decoder extracts keyframes and models the background of the video stream according to preset time slices, and extracts the motion trajectory sequence of the foreground target as the first type of unstructured mode segment. The environmental vibration spectrum is decomposed by wavelet packet transform using the analog signal sampler, and the energy entropy characteristics are calculated in multiple frequency bands to form a frequency domain energy distribution vector as a second type of unstructured mode fragment. The text segmenter performs dependency parsing on the device operation report to identify core event triples consisting of subject, predicate, and object, and converts each triple into a node with a timestamp to generate the structured semantic unit. The foreground target motion trajectory sequence, frequency domain energy distribution vector, and nodes with timestamps are synchronously sent to the feature alignment buffer.

[0007] Preferably, the step of performing blind source separation and feature reconstruction on the unstructured pattern fragment to generate a parsed physical situation feature map includes: Read the foreground target motion trajectory sequence and frequency domain energy distribution vector from the feature alignment buffer; An independent component analysis model is constructed to separate the intertwined multi-target motion components in the foreground target motion trajectory sequence, resulting in a set of independent motion sub-trajectories. A variational mode decomposition model is constructed to separate the mixed multi-source vibration signals in the frequency domain energy distribution vector, thereby obtaining a set of intrinsic mode functions. Design a graph convolutional network to map the set of motion sub-trajectories and the set of intrinsic mode functions into a dynamic spatiotemporal graph structure, where nodes represent moving targets or vibration sources and edges represent spatiotemporal correlations. By integrating the dynamic spatiotemporal graph structure, the importance weights between nodes are calculated through a graph attention mechanism to generate a comprehensive physical situation feature map that represents the dynamics of the physical environment.

[0008] Preferably, the step of implementing the knowledge graph embedding and relation parsing process for the structured semantic unit to generate the parsed logical event feature chain includes: Read the timestamp-marked node from the feature alignment buffer; Access a pre-built domain knowledge graph for device monitoring, which contains semantic relationships between device entities, failure modes, and maintenance actions; Each node is semantically similar to an entity in the knowledge graph. Successfully matched nodes are linked to the corresponding knowledge graph entity and inherit the attributes and relationships of the knowledge graph entity. Based on the linked nodes and their inheritance relationships, the causal or temporal logical relationships between the nodes are deduced using temporal reasoning rules, forming weighted logical relationship edges. All nodes with timestamps and logical relationship edges with weights are organized in chronological order into a chain structure with temporal dependencies and logical constraints, which is the parsed logical event feature chain.

[0009] Preferably, the step of performing co-evolutionary analysis on the physical situation feature map and the logical event feature chain using a genetic algorithm architecture to generate an association matrix includes: The correlation matrix describes the implicit dependencies between different dimensions of monitoring signals in time and space; Design a chromosome encoding scheme by cross-splitting the node feature vectors in the physical situation feature map with the node feature vectors in the logical event feature chain to form an initial chromosome population; Define a fitness function, which calculates chromosome quality based on the mutual information entropy of node states between the physical situation feature map and the logical event feature chain within the same time slice; Genetic iterative operations are performed, including selection based on roulette wheel strategy, crossover based on two-point crossover, and mutation based on random site flipping, to optimize the chromosome population generation by generation; After reaching the preset number of iterations, decode the chromosome with the highest fitness and extract its encoded node correspondence; Based on the node correspondence, a two-dimensional matrix is ​​constructed, where the rows of the matrix correspond to the nodes in the physical situation feature map, the columns correspond to the nodes in the logical event feature chain, and the matrix element values ​​are the correlation strength between the corresponding nodes, thus forming the correlation matrix.

[0010] Preferably, the step of locating the initial abnormal disturbance source and its propagation path in the monitoring signal based on the correlation matrix-driven source tracing mechanism includes: Singular value decomposition is performed on the correlation matrix to identify the left and right singular vectors that represent the strongest correlation pattern. Map the left singular vector back to the physical situation feature map to locate the set of nodes with the most significant abnormal responses in the physical feature space and mark them as potential physical anomaly sources. Map the right singular vector back to the logical event feature chain to locate the set of nodes with the highest degree of abnormality in the logical event chain and mark them as potential logical anomaly sources; A bidirectional propagation model is constructed, in which anomaly scores are propagated from the potential physical anomaly source along the edges of the physical situation feature graph, and simultaneously from the potential logical anomaly source along the edges of the logical event feature chain. During the propagation process, nodes with abnormal scores exceeding the threshold and the order in which they were triggered are recorded and summarized to form a set of abnormal propagation paths that start from physical or logical abnormal sources and gradually spread. The common origin node or the earliest node of all paths is found by backtracking from the set of abnormal propagation paths, which is determined as the initial abnormal disturbance source, and the relevant paths are integrated to form the complete propagation link.

[0011] Preferably, the process of generating the source tracing report includes: Enter the type, spatial location, or logical identifier information of the initial abnormal disturbance source into the report header; The propagation chain is visualized as a directed graph, with each node labeled with its anomaly type, intensity, and activation timestamp. Calculate the contribution of each node on the propagation link to the overall abnormal state, and generate a contribution distribution table; The report header, the directed graph of the visualization description, and the contribution distribution table are integrated to generate the structured source tracing report.

[0012] Preferably, the step of automatically deriving and issuing a set of control instructions for the sampling frequency and alarm threshold of the monitoring equipment based on the content of the source tracing report includes: Analyze the source tracing report to extract the key stages of the abnormal propagation and the types of sensors affected; Query the monitoring equipment configuration database to obtain the sampling frequency and alarm threshold setting value of the currently affected sensor type; Based on the speed and range data of abnormal propagation, the minimum sampling frequency required for the key stage is recalculated through a feedback control algorithm to ensure the capture of abnormal transient characteristics, and the dynamic alarm threshold of each sensor node is recalculated to filter false alarms. The recalculated sampling frequency value and dynamic alarm threshold are encapsulated into executable configuration instructions; The configuration command is sent to the corresponding monitoring device of the affected sensor type via a secure communication protocol.

[0013] Preferably, the recalculation of the minimum sampling frequency required for the critical stage based on the velocity and range data of anomaly propagation using a feedback control algorithm includes: An anomaly propagation dynamics model is established, with the speed and range data of the anomaly propagation used as model input; Set an objective function that requires the sampling system to be able to distinguish the movement details of the anomaly front during the critical phase; The objective function is solved using the gradient descent method, resulting in a set of sampling frequency sequences that theoretically meet the resolution requirements; Considering the hardware limitations and network bandwidth constraints of the monitoring equipment, the sampling frequency sequence is modified to obtain the minimum sampling frequency.

[0014] Preferably, the recalculation of the dynamic alarm threshold for each sensor node to filter false alarms includes: Collect the reading distribution of the affected sensor types under historical normal conditions and the reading distribution during this abnormal event; Gaussian mixture models are used to model the reading distribution under the historical normal state and the reading distribution during the current abnormal event, respectively. Calculate the difference in probability distribution between the two models and find the decision boundary that maximizes the distinction between normal and abnormal states; The decision boundary is transformed into a specific threshold value, which serves as the dynamic alarm threshold. For multi-parameter sensors, the corresponding dynamic alarm threshold is calculated and set independently for each key parameter.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By employing a genetic algorithm architecture to perform co-evolutionary analysis of physical situation feature maps and logical event feature chains, and leveraging the algorithm's global search and iterative optimization capabilities, the two types of heterogeneous features are forced to adapt and match each other during the evolutionary process. This process can automatically uncover implicit, non-preset, strong correlation rules between specific patterns in physical signals such as video and vibration, and logical events in device reports and operation logs. The resulting correlation matrix is ​​essentially a quantitative expression of the dynamic coupling relationship between the "physical world state" and the "event logical relationship" under the current monitoring system. This allows the system to determine anomalies no longer based on a single signal or simple rule, but rather on the synergistic verification of multi-dimensional evidence chains, improving the accuracy of complex anomaly diagnosis and the depth of early warning. It can identify potential risks caused by the slow coupling of multiple factors that are difficult to detect using traditional methods.

[0016] Driven by a dynamically generated correlation matrix, the system provides weighted guidance for reverse tracing, with the correlation strength within the matrix serving as the basis for analysis. The system can pinpoint the initial abnormal disturbance that triggers a chain reaction and clearly depict the propagation path of the anomaly from the physical layer to the information layer, or across devices and regions, generating a causal-logical source tracing report. Based on this report, the system automatically derives and generates a set of control instructions for the sampling frequency and alarm thresholds of front-end monitoring devices. This mechanism achieves a closed loop from intelligent analysis and precise source tracing to proactive optimization. The system can strengthen monitoring of identified weak points or adjust resource allocation to non-critical areas based on the source tracing results, enabling the resource configuration and alarm strategies of the entire monitoring network to continuously self-optimize as operational status and fault modes evolve, thus improving the overall adaptability and operational efficiency of the system. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the intelligent monitoring and early warning system based on genetic algorithms described in this invention; Figure 2 A flowchart illustrating the process of performing multi-level signal deconstruction operations; Figure 3 A flowchart for implementing knowledge graph embedding and relation parsing for structured semantic units; Figure 4 A comparison chart showing sensor sampling frequency adjustments during critical stages of abnormal propagation; Figure 5 This is a heatmap showing the correlation strength between the physical state feature map and the logical event feature chain. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 The system begins with the signal receiving module, which receives multi-dimensional monitoring signals aggregated from the monitoring network. These signals are integrated to form a raw monitoring data cluster, covering at least various heterogeneous data sources such as video streams, equipment current waveform sequences, and environmental vibration maps. The signal deconstruction module performs multi-level signal deconstruction operations on the raw monitoring data cluster, breaking it down into structured semantic units and unstructured pattern fragments. The structured semantic units mainly originate from text data such as equipment operation reports, while the unstructured pattern fragments originate from analog signals continuously acquired by sensors. The feature parsing and generation module processes these two types of data separately. For unstructured pattern fragments, it implements processes including blind source separation and feature reconstruction to generate a parsed physical situation feature map. For structured semantic units, it implements processes including knowledge graph embedding and relationship parsing to generate a parsed logical event feature chain. The co-evolutionary analysis module introduces a genetic algorithm architecture to perform co-evolutionary analysis on the physical situation feature map and the logical event feature chain, ultimately generating a correlation matrix characterizing the complex relationships within and between the two. The anomaly tracing and investigation module uses this correlation matrix to drive the tracing and investigation mechanism, locates the initial abnormal disturbance source in the monitoring signal and its propagation path in the entire monitoring network, and generates a structured tracing and investigation report accordingly. The control command generation module automatically derives and issues a set of control commands targeting the sampling frequency and alarm thresholds of the monitoring equipment based on the specific content of the tracing and investigation report, thus completing a closed loop from perception, analysis, tracing to control.

[0021] In one embodiment of the present invention, see [reference] Figure 2Consider a scenario where a factory assembly line is being monitored. The raw monitoring data cluster collected by the monitoring network includes video streams aligned with the assembly robot arm, equipment current waveform sequences attached to the robot arm's joint motors, and environmental vibration maps of the assembly base. The signal receiving module aggregates these multi-dimensional monitoring signals and sends them to the signal deconstruction module. The signal deconstruction module initializes its built-in signal deconstruction engine, which loads and configures the video stream decoder, text segmenter, and analog signal sampler. In practice, the video stream decoder slices the video stream at 100-millisecond intervals, extracts one frame from each slice as a keyframe, and uses a Gaussian mixture model to model the background of the continuous keyframe sequence. After background modeling, the foreground target motion trajectory sequence is extracted. Each trajectory point in the sequence contains the center coordinates (x, y) of the target bounding box and a timestamp t, forming a first-type unstructured pattern segment. The analog signal sampler processes the environmental vibration spectrum, which is a time-domain signal acquired by an accelerometer. The sampler performs a 5-level wavelet packet transform decomposition on the signal, dividing it into 32 frequency bands. Within each frequency band, the energy entropy characteristic is calculated based on the wavelet packet coefficients. The calculation of the energy entropy characteristic follows the formula:

[0022] in: This represents the energy entropy value of the j-th frequency band. This indicates the number of segments in the wavelet packet coefficients within that frequency band. This represents the proportion of energy of the k-th segment coefficient within the j-th frequency band to the total energy of that frequency band. After calculating the energy entropy of all frequency bands, they are arranged in frequency band order to form a frequency domain energy distribution vector [E1, E2, ..., E32], which serves as the second type of unstructured pattern fragment. The text segmenter analyzes the equipment operation reports generated within the same time period. The equipment operation report is a text record such as "robotic arm A completes part grabbing at time T". The text segmenter uses dependency parsing to identify the subject "robotic arm A", the predicate "completed", and the object "part grabbing", forming a core event triple (robotic arm A, completed, part grabbing). The triple is converted into a node with a timestamp T, becoming a component of the structured semantic unit. In specific implementation, the foreground target motion trajectory sequence, the frequency domain energy distribution vector, and the nodes with timestamps are synchronously written into the feature alignment buffer. The buffer aligns the data according to a unified time reference.

[0023] The feature parsing and generation module reads unstructured pattern fragments from the feature alignment buffer, namely the foreground target motion trajectory sequence and the frequency domain energy distribution vector. An independent component analysis (ICA) model is constructed, taking the coordinate time series of multiple moving targets intertwined in the foreground target motion trajectory sequence as input. Assuming the observed trajectory is a linear mixture of multiple independent motion source signals, an optimization algorithm estimates the unmixing matrix, thereby separating the sets of independent motion sub-trajectories, such as separating the trajectory of the robotic arm's body motion from the trajectory of parts moving on a conveyor belt. A variational mode decomposition (VM) model is constructed, taking the reconstructed segment of the original vibration signal corresponding to the frequency domain energy distribution vector as input. Through iterative search, a set of bandwidth-limited intrinsic mode functions (EMFs) with specific center frequencies is determined, separating the mixed multi-source vibration signals, such as the robotic arm motor vibration and background environmental vibration, into independent signal components, resulting in the set of EMFs.

[0024] In some embodiments, a graph convolutional network is designed to map the separated data. A set of motion sub-trajectories is input into the graph convolutional network. The network defines each independent moving target as a node, and the feature vector of a node consists of the velocity and acceleration sequence of that target's motion sub-trajectory. If the motion trajectories of two targets are significantly correlated in space and time, for example, if the distance between them is consistently less than a set threshold within three consecutive frames, an edge is established between the corresponding nodes, thereby constructing a dynamic spatiotemporal graph structure describing motion interaction. A set of intrinsic mode functions (IMFs) is also input into the graph convolutional network. Each vibration signal source is defined as a node, and the feature vector of a node consists of the instantaneous frequency and amplitude of the corresponding IMF. If the signals of two vibration sources exhibit synchronicity or causality in the time-frequency domain, an edge is established between the corresponding nodes, constructing a dynamic spatiotemporal graph structure describing vibration correlation. By integrating the two dynamic spatiotemporal graph structures mentioned above, the importance weights between nodes are calculated through a graph attention mechanism. The graph attention mechanism assigns a learnable attention coefficient to each node in the graph, which is used to aggregate the features of its neighboring nodes. Finally, a comprehensive physical situation feature map is generated that represents the dynamics of the physical environment. Each node in the graph contains information about its origin, type, and the strength of its association with other nodes.

[0025] Optionally, when constructing the independent component analysis model, the input foreground target motion trajectory sequence can be preprocessed with mean removal and whitening to improve the stability of the separation effect. Optionally, in the variational mode decomposition model, the preset value of the number of intrinsic mode functions can be adaptively adjusted based on the number of peaks in the vibration signal spectrum. It can be understood that the number of layers in the graph convolutional network can be configured according to the complexity of the moving target or vibration source. It can be understood that the multi-head attention setting in the graph attention mechanism can capture the dependencies between nodes from different subspaces, enhancing the representational ability of the physical situation feature map.

[0026] In one embodiment of the present invention, see [reference] Figure 3The system accesses a pre-built knowledge graph for equipment monitoring. The knowledge graph stores semantic relationships in triples, such as entities "Automated Guided Vehicle" (AGV), "robotic arm," "conveyor belt," "motor overheating," and relationships like "execution device," "potentially triggers," and "belongs to component." Each node with a timestamp is matched semantically with entities in the knowledge graph. The matching process calculates the cosine similarity between the node text and the entity name's word vector. A successful match is considered successful when the similarity exceeds a threshold of 0.85. For example, the node "AGV_01 arrived at station A at timestamp 1620000000" successfully matches the knowledge graph entity "Automated Guided Vehicle." The node is linked to the "Automated Guided Vehicle" entity and inherits its attributes, such as "speed range" and "rated load," as well as its associations, such as "execution device" pointing to "warehousing operations" and "belongs to component" pointing to "drive motor." Based on the linked nodes and their inheritance relationships, a pre-defined temporal reasoning rule is applied for deduction. This rule is defined as follows: if the timestamp of event A is earlier than the timestamp of event B, and there exists a "predecessor" or "potential trigger" relationship between entities of event A and entities of event B in the knowledge graph, then a weighted logical relationship edge is established between event A and event B. Based on this, a "trigger preparation" logical relationship edge with a weight of 0.7 is established from the "Automatic Guided Vehicle Arrival" node to the "Robotic Arm Grasping" node, and a "trigger" logical relationship edge with a weight of 0.9 is established from the "Robotic Arm Grasping" node to the "Conveyor Belt Start" node. All nodes with timestamps and weighted logical relationship edges are organized strictly in ascending order of timestamps, forming a chain structure with temporal dependencies and logical constraints—the parsed logical event feature chain. This logical event feature chain can be formally represented as an ordered set of nodes and edges.

[0027] The co-evolutionary analysis module performs co-evolutionary analysis on the physical situation feature map and the logical event feature chain using a genetic algorithm architecture. The physical situation feature map, derived from the output of the embodiment, contains physical nodes representing vibration and displacement. In a specific implementation, a chromosome encoding scheme is designed. This scheme cross-concatenates the 128-dimensional feature vector of each node in the physical situation feature map with the 64-dimensional feature vector of each node in the logical event feature chain, with the physical feature part first and the logical feature part second, forming a binary encoded string of total length (128 physical nodes + 64 logical nodes). This method is used to randomly generate an initial chromosome population containing 200 individuals. A fitness function is defined to evaluate the chromosome. The fitness function is calculated based on the mutual information entropy of the node states between the physical situation feature map and the logical event feature chain within the same time slice. In some embodiments, for a set of physical-logical node pairings decoded by the chromosome, the mutual information entropy fitness value F is calculated according to the following formula:

[0028] in: This represents the total number of nodes in the physical situation feature graph. This represents the total number of nodes in the logical event feature chain. It is an indicator function that, when chromosome encoding, will physical nodes With logical nodes The value is 1 if a pair is found, and 0 otherwise. It is a node State sequences and nodes The mutual information entropy value of the state sequence within a specific time window. A higher value indicates a better association pattern encoded by the chromosome. Genetic iteration operations are performed, including selection, crossover, and mutation. The selection operation is based on a roulette wheel strategy, where the probability of each chromosome being selected is related to its fitness value. The genes are proportional to the chromosome length; the crossover operation uses a two-point crossover method, exchanging partial gene segments of the parent chromosome between two randomly selected crossover points; the mutation operation is based on random site flipping, flipping any binary bit on the chromosome coding string with a probability of 0.005. A preset iteration count of 500 generations is set. After reaching the preset iteration count, fitness values ​​are selected from the final population. The highest-level chromosome is decoded, and the decoding process extracts the node correspondences of the chromosome encoding, i.e., which physical nodes are strongly associated with which logical nodes. Based on the extracted node correspondences, a two-dimensional matrix is ​​constructed. The rows of the matrix correspond to all nodes in the physical situation feature map, and the columns correspond to all nodes in the logical event feature chain. The matrix element values... For the corresponding row physical node AND column logical nodes The correlation strength between them is determined by the decoded pairing relationship and the mutual information entropy value, forming the final correlation matrix.

[0029] Optionally, the semantic similarity matching process can be combined with a domain dictionary for synonym expansion to improve the accuracy of linking nodes with entities in the knowledge graph. Optionally, the temporal reasoning rule base can include multiple logical relation templates to generate richer logical event feature chains. It is understandable that the concatenation order of feature vectors in the chromosome encoding scheme can be adjusted, with logical features preceding physical features. It is understandable that the calculation of mutual information entropy can employ a non-parametric method based on kernel density estimation to handle non-Gaussian distributed state sequence data.

[0030] In one embodiment of the present invention, considering an example scenario of wind turbine monitoring, the co-evolution analysis module has generated an association matrix. The rows of the association matrix correspond to nodes in the physical state feature graph, and the columns of the association matrix correspond to nodes in the logical event feature chain. The anomaly tracing and investigation module, based on the association matrix-driven tracing and investigation mechanism, first performs singular value decomposition on the association matrix. Singular value decomposition decomposes the association matrix into the product of three matrices, from which the left singular vector and right singular vector corresponding to the largest singular value are identified. The left and right singular vectors represent the strongest association pattern in the association matrix.

[0031] Mapping the left singular vector back to the physical situation feature map, each element in the left singular vector corresponds to a node in the physical situation feature map. Selecting the nodes corresponding to the three elements with the highest amplitude in the vector locates the set of nodes with the most significant abnormal responses in the physical feature space, marking them as potential physical anomaly sources. Mapping the right singular vector back to the logical event feature chain, each element in the right singular vector corresponds to a node in the logical event feature chain. Selecting the nodes corresponding to the two elements with the highest amplitude in the vector locates the set of nodes with the highest anomaly degree in the logical event chain, marking them as potential logical anomaly sources. Constructing a bidirectional propagation model, the bidirectional propagation model defines two independent score propagation processes. One process starts from the potential physical anomaly source and propagates the anomaly score along predefined edges in the physical situation feature map, such as the edge established by the co-occurrence relationship between vibration nodes and temperature nodes. The other process starts from the potential logical anomaly source and propagates along the weighted logical relationship edges in the logical event feature chain. In some embodiments, the propagation of anomaly scores along the edges follows a decay and superposition rule. For the increment of anomaly scores propagating from node u to adjacent node v... The calculation follows the formula:

[0032] in: It is the current anomaly score of the source node u. These are predefined association weights on the edge (u,v). It is a decay coefficient greater than zero. This is a distance metric between node u and node v, such as time difference or topology hop count. During propagation, the system records all nodes with anomaly scores exceeding a preset threshold of 0.6 and their exact timestamps when they were triggered. This data is then aggregated to form a set of anomaly propagation paths that start from a physical or logical anomaly source and gradually spread outwards. For example, paths like "Nacelle Vibration_Axial_Sensor 01->Gearbox Oil Temperature_Sensor 03" and "Pitch System Alarm_Log ID1001->Gearbox Overheat Warning_Log ID1002" are formed. A backtracking search is performed on this set of propagation paths to analyze the temporal and causal relationships of nodes in all paths. The earliest common origin node in terms of timestamps is found. For example, if the timestamp of node "Nacelle Vibration_Axial_Sensor 01" is earlier than all other related nodes, this node is identified as the initial anomaly source. All paths containing this origin node are then integrated to form a complete propagation link, which presents as a directed acyclic graph structure.

[0033] When generating the source tracing report, the type of the initial abnormal disturbance source, "abnormal mechanical vibration," and its spatial location, "axial position of wind turbine #05 nacelle," or logical identifier information, "sensor ID: VIB_05_AXIAL," are entered in the report header. The complete propagation chain is visualized as a directed graph, generated using a standard graph library. Each node in the graph is labeled with a different shape and color to indicate its anomaly type; for example, a red diamond represents "vibration anomaly," and an orange circle represents "temperature anomaly." The anomaly intensity value and the timestamp of its activation are also labeled next to the node. The contribution of each node in the propagation chain to the overall abnormal state is calculated. The contribution is based on the proportion of the total anomaly score received and forwarded by the node during propagation to the total score, generating a two-column contribution distribution table: one column for node identification and the other for contribution percentage. The report header, the directed graph visualization, and the contribution distribution table are integrated and encapsulated according to a predefined XML or JSON format to generate a structured source tracing report.

[0034] In some embodiments, after performing singular value decomposition on the correlation matrix, the top K largest singular values ​​and their corresponding vectors can be retained for analysis to focus on the most prominent correlation patterns. Optionally, when constructing a bidirectional propagation model, a coupling interface can be set between the physical situation feature graph and the logical event feature chain. When the anomaly score of one side exceeds a specific threshold, the relevant nodes on the other propagation path can be activated. Optionally, the directed graph of the visual description can support interactive expansion, and clicking on a node can display the original signal fragment or log details of that node. It is understood that contribution calculation can be performed iteratively until the sum of the contributions of all nodes converges to 100%. It is understood that the format of the source tracing investigation report can be adapted to a PDF document or database record according to the interface requirements of the downstream system.

[0035] In one embodiment of the present invention, considering an example scenario of monitoring a chemical pipeline leak, the source tracing report generated by the anomaly tracing module indicates that the initial abnormal disturbance source is "pipeline pressure anomaly_sensor P101". The propagation path shows that the anomaly affected "pipeline vibration anomaly_sensor V102" located 5 meters downstream and "ambient gas concentration anomaly_sensor C103" located 10 meters downstream within the following 3 seconds. The control command generation module automatically derives and issues a set of control commands based on the content of this source tracing report. First, it analyzes the source tracing report and extracts the key stage of anomaly propagation as 0 to 3 seconds after the initial disturbance occurs, and the types of sensors affected include pressure sensors, vibration sensors, and gas concentration sensors.

[0036] Query the monitoring equipment configuration database to obtain the sampling frequency and alarm threshold settings of the affected sensor types. See Table 1 for the query results of the monitoring equipment configuration database.

[0037] Table 1: Query Results of Monitoring Equipment Configuration Database

[0038] Based on the anomaly propagation speed and range data provided in the source tracing report, such as the anomaly front propagating along the pipeline at an average speed of 2 m / s, a feedback control algorithm recalculates the minimum sampling frequency required for the critical stage. The feedback control algorithm compares the current sampling frequency with the characteristic frequency of the anomaly signal, generates an error signal, and adjusts the sampling frequency command based on the error. In some embodiments, to ensure the capture of the transient characteristics of the anomaly, the minimum sampling frequency is recalculated. Follow the formula below:

[0039] in: Indicates the current sampling frequency of the sensor. This indicates the dominant frequency analyzed from the abnormal signal segment. Indicates abnormal propagation speed. This indicates the required spatial resolution of the system. Simultaneously, the dynamic alarm thresholds for each sensor node are recalculated to filter false alarms. The reading distributions of pressure sensors, vibration sensors, and gas concentration sensors under normal conditions over the past 24 hours, as well as the reading distribution within the first 3 seconds of this abnormal event, are collected. Gaussian mixture models are used to model both sets of distributions, calculating the distribution difference between the two probability models to find the decision boundary that maximizes the distinction between normal and abnormal states. This decision boundary is then transformed into specific threshold values ​​as the dynamic alarm thresholds. The recalculated sampling frequency values ​​and dynamic alarm thresholds are encapsulated into executable configuration instructions in JSON format, including fields such as device identifier, target parameter type, and target parameter value. These encapsulated configuration instructions are then sent to the corresponding monitoring devices of the affected sensor types via a secure communication protocol using TLS 1.3 with digital signatures to ensure the integrity and confidentiality of the transmitted instructions.

[0040] Optionally, when parsing the source tracing report, natural language processing technology can be used to automatically extract time window descriptions and sensor model descriptions for key stages. Optionally, the monitoring equipment configuration database can be designed with version management functionality to record the history of each configuration change for rollback purposes. In some embodiments, the feedback control algorithm can adopt the form of a proportional-integral-derivative controller, comparing the calculated minimum sampling frequency with the device's maximum allowed sampling frequency and taking the smaller value as the final command value. It is understood that for multi-parameter sensors, such as a sensor that simultaneously monitors temperature and humidity, it is necessary to independently calculate and set corresponding dynamic alarm thresholds for the two key parameters of temperature and humidity. It is understood that the issuance of configuration commands can be done via broadcast, multicast, or unicast, depending on the topology of the monitoring network.

[0041] See Figure 4 This chart compares the adjustment of sensor sampling frequencies during the critical stage of anomaly propagation, demonstrating the frequency adjustments of three types of sensors during the critical stage (0-3 seconds) of anomaly propagation in a chemical pipeline leak monitoring scenario. The sampling frequencies of all sensors significantly increased after adjustment, indicating that the system proactively increased the data acquisition density to capture transient anomaly characteristics. This chart visually reflects how the system dynamically optimizes sensor configuration based on the anomaly propagation speed and spatial resolution requirements after an anomaly occurs. By specifically increasing the sampling frequency, the system can more accurately capture transient anomaly characteristics, avoiding missed detections. This provides richer data support for subsequent anomaly tracing and propagation path analysis, effectively reducing the risk of misjudgment due to insufficient sampling frequency.

[0042] In one embodiment of the present invention, continuing the example scenario of monitoring chemical pipeline leaks, the control command generation module recalculates the minimum sampling frequency required for the critical stage based on the abnormal propagation speed and range data in the source tracing report. An abnormal propagation dynamics model is established, using the quantified abnormal propagation speed and range data from the source tracing report as model input. The abnormal propagation dynamics model treats the pipeline as a one-dimensional transmission medium and defines a partial differential equation for the variation of the abnormal signal intensity along the pipeline position x and time t. An objective function is set, requiring that during the critical stage of abnormal propagation, the sampling frequency configuration of the monitoring system enables the sampling system to distinguish the movement details of the abnormal front, i.e., the positional change of the abnormal signal captured between adjacent sampling points should be less than the spatial resolution accuracy required by the system. The objective function is solved using the gradient descent method. The gradient descent method iteratively adjusts the sampling frequency variable, calculating the theoretical resolution error of the sampling system for the movement of the abnormal front in each iteration, until the theoretical resolution error is less than or equal to a preset tolerance. At this point, a set of sampling frequency sequences that meet the resolution requirements is obtained, containing suggested frequency values ​​for different sensors in different time windows.

[0043] Considering the hardware limitations and network bandwidth constraints of the monitoring equipment, a feasibility correction is performed on the sampling frequency sequence obtained by the gradient descent method. The hardware limitations of the monitoring equipment include the maximum sampling rate of the sensor analog-to-digital converter and the upper limit of the processor's computational throughput; the network bandwidth constraint refers to the maximum number of data packets that a data transmission link can carry per unit time. The feasibility correction process involves multiplying each proposed frequency value in the sampling frequency sequence by a correction coefficient less than or equal to 1, where the correction coefficient is determined by the real-time load rate of the current hardware and network. In some embodiments, the minimum sampling frequency obtained after feasibility correction... The calculation follows the formula:

[0044] in: This indicates the highest sampling frequency allowed by the sensor's analog-to-digital converter. Indicates the currently available network bandwidth. This indicates the number of bits in a single sampled data packet. It is a constant coefficient related to protocol overhead. This is the target sampling frequency value obtained by solving the gradient descent method; the formula output is... This is the final determined minimum sampling frequency.

[0045] When recalculating the dynamic alarm thresholds for each sensor node to filter false alarms, the reading distributions of the affected pressure sensors, vibration sensors, and gas concentration sensors under historical normal conditions and during the current abnormal event are collected. In practice, the reading distribution under historical normal conditions refers to the cleaned data sequence collected by these sensors during the past 30 days without alarm records; the reading distribution during the current abnormal event refers to the original readings collected by the corresponding sensors during the initial abnormal phase identified in the source tracing report. Gaussian mixture models are used to model both sets of reading distributions. For historical normal data, the Gaussian mixture model fits a probability density function formed by the superposition of multiple Gaussian components, characterizing the reading fluctuation pattern under normal operating conditions; for the current abnormal data, the Gaussian mixture model fits another probability density function describing the reading characteristics under abnormal conditions. The probability distribution difference between the two Gaussian mixture models is calculated, and the Kullback-Leibler divergence is used to measure the probability distribution difference. A decision boundary that maximizes the distinction between normal and abnormal states is found; the decision boundary is the set of points or multiple points on the reading value axis that make the two probability density function values ​​equal. The decision boundary is transformed into a specific threshold value as a dynamic alarm threshold. For a pressure sensor, the decision boundary may correspond to a specific pressure value, and for a gas concentration sensor, the decision boundary may correspond to a concentration value.

[0046] See Figure 5 This is a heatmap showing the correlation strength between physical situation feature maps and logical event feature chains, illustrating the correlation strength between nodes in the physical situation feature map and nodes in the logical event feature chain. For combinations with weak correlation strength (such as "temperature change - equipment startup"), alarm rules can be optimized to avoid over-reliance on such indicators to trigger alarms, thereby reducing ineffective responses from maintenance personnel and improving work efficiency. For physical indicators with high correlation strength (such as pipeline pressure and gas concentration), redundant monitoring or increased sampling frequency can be added to the system design to enhance the ability to perceive risk events. When a system anomaly occurs, reverse tracing of correlation strength can quickly locate possible causes. For example, if a "leak detection" event is triggered, changes in "gas concentration" and "pipeline pressure" can be prioritized to accelerate fault location and repair. By identifying potential high-risk correlations, preventative maintenance of weak links can be carried out in advance, reducing the probability of accidents and minimizing downtime losses and safety risks.

[0047] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent monitoring and early warning system based on genetic algorithms, characterized in that, Includes the following steps: The signal receiving module receives multi-dimensional monitoring signals collected from the monitoring network, forming a raw monitoring data cluster. This raw monitoring data cluster at least covers video image streams, equipment current waveform sequences, and environmental vibration maps. The signal deconstruction module performs multi-level signal deconstruction operations, decomposing the original monitoring data cluster into structured semantic units and unstructured pattern fragments. The structured semantic units come from the equipment operation report, and the unstructured pattern fragments come from the analog signals continuously collected by the sensors. The feature parsing and generation module performs blind source separation and feature reconstruction on the unstructured pattern fragments to generate a parsed physical situation feature map. At the same time, it performs knowledge graph embedding and relationship parsing on the structured semantic units to generate a parsed logical event feature chain. The co-evolutionary analysis module performs co-evolutionary analysis on the physical situation feature map and the logical event feature chain using a genetic algorithm architecture to generate an association matrix; The anomaly tracing and investigation module, based on the correlation matrix, drives the tracing and investigation mechanism to locate the initial abnormal disturbance source and its propagation link in the monitoring signal, and generates a tracing and investigation report. The control instruction generation module automatically derives and issues a set of control instructions for the sampling frequency and alarm threshold of the monitoring equipment based on the content of the source tracing report.

2. The intelligent monitoring and early warning system based on genetic algorithm according to claim 1, characterized in that, The execution of multi-level signal deconstruction operations includes: Initialize the signal deconstruction engine, which is configured with a video stream decoder, a text segmenter, and an analog signal sampler; The video stream decoder extracts keyframes and models the background of the video stream according to preset time slices, and extracts the motion trajectory sequence of the foreground target as the first type of unstructured mode segment. The environmental vibration spectrum is decomposed by wavelet packet transform using the analog signal sampler, and the energy entropy characteristics are calculated in multiple frequency bands to form a frequency domain energy distribution vector as a second type of unstructured mode fragment. The text segmenter performs dependency parsing on the device operation report to identify core event triples consisting of subject, predicate, and object, and converts each triple into a node with a timestamp to generate the structured semantic unit. The foreground target motion trajectory sequence, frequency domain energy distribution vector, and nodes with timestamps are synchronously sent to the feature alignment buffer.

3. The intelligent monitoring and early warning system based on genetic algorithm according to claim 2, characterized in that, The process of performing blind source separation and feature reconstruction on the unstructured pattern fragments to generate a parsed physical situation feature map includes: Read the foreground target motion trajectory sequence and frequency domain energy distribution vector from the feature alignment buffer; An independent component analysis model is constructed to separate the intertwined multi-target motion components in the foreground target motion trajectory sequence, resulting in a set of independent motion sub-trajectories. A variational mode decomposition model is constructed to separate the mixed multi-source vibration signals in the frequency domain energy distribution vector, thereby obtaining a set of intrinsic mode functions. Design a graph convolutional network to map the set of motion sub-trajectories and the set of intrinsic mode functions into a dynamic spatiotemporal graph structure, where nodes represent moving targets or vibration sources and edges represent spatiotemporal correlations. By integrating the dynamic spatiotemporal graph structure, the importance weights between nodes are calculated through a graph attention mechanism to generate a comprehensive physical situation feature map that represents the dynamics of the physical environment.

4. The intelligent monitoring and early warning system based on genetic algorithm according to claim 2, characterized in that, The process of embedding knowledge graphs and parsing relationships for the structured semantic units to generate parsed logical event feature chains includes: Read the timestamp-marked node from the feature alignment buffer; Access a pre-built domain knowledge graph for device monitoring, which contains semantic relationships between device entities, failure modes, and maintenance actions; Each node is semantically similar to an entity in the knowledge graph. Successfully matched nodes are linked to the corresponding knowledge graph entity and inherit the attributes and relationships of the knowledge graph entity. Based on the linked nodes and their inheritance relationships, the causal or temporal logical relationships between the nodes are deduced using temporal reasoning rules, forming weighted logical relationship edges. All nodes with timestamps and logical relationship edges with weights are organized in chronological order into a chain structure with temporal dependencies and logical constraints, which is the parsed logical event feature chain.

5. The intelligent monitoring and early warning system based on genetic algorithm according to claim 1, characterized in that, The step involves performing co-evolutionary analysis of the physical situation feature map and the logical event feature chain using a genetic algorithm architecture to generate an association matrix, including: The correlation matrix describes the implicit dependencies between different dimensions of monitoring signals in time and space; Design a chromosome encoding scheme by cross-splitting the node feature vectors in the physical situation feature map with the node feature vectors in the logical event feature chain to form an initial chromosome population; Define a fitness function, which calculates chromosome quality based on the mutual information entropy of node states between the physical situation feature map and the logical event feature chain within the same time slice; Genetic iterative operations are performed, including selection based on roulette wheel strategy, crossover based on two-point crossover, and mutation based on random site flipping, to optimize the chromosome population generation by generation; After reaching the preset number of iterations, decode the chromosome with the highest fitness and extract its encoded node correspondence; Based on the node correspondence, a two-dimensional matrix is ​​constructed, where the rows of the matrix correspond to the nodes in the physical situation feature map, the columns correspond to the nodes in the logical event feature chain, and the matrix element values ​​are the correlation strength between the corresponding nodes, thus forming the correlation matrix.

6. The intelligent monitoring and early warning system based on genetic algorithm according to claim 5, characterized in that, The method of locating the initial abnormal disturbance source and its propagation path in the monitoring signal based on the correlation matrix-driven source tracing mechanism includes: Singular value decomposition is performed on the correlation matrix to identify the left and right singular vectors that represent the strongest correlation pattern. Map the left singular vector back to the physical situation feature map to locate the set of nodes with the most significant abnormal responses in the physical feature space and mark them as potential physical anomaly sources. Map the right singular vector back to the logical event feature chain to locate the set of nodes with the highest degree of abnormality in the logical event chain and mark them as potential logical anomaly sources; A bidirectional propagation model is constructed, in which anomaly scores are propagated from the potential physical anomaly source along the edges of the physical situation feature graph, and simultaneously from the potential logical anomaly source along the edges of the logical event feature chain. During the propagation process, nodes with abnormal scores exceeding the threshold and the order in which they were triggered are recorded and summarized to form a set of abnormal propagation paths that start from physical or logical abnormal sources and gradually spread. The common origin node or the earliest node of all paths is found by backtracking from the set of abnormal propagation paths, which is determined as the initial abnormal disturbance source, and the relevant paths are integrated to form the complete propagation link.

7. The intelligent monitoring and early warning system based on genetic algorithm according to claim 6, characterized in that, The formation of the source tracing investigation report includes: Enter the type, spatial location, or logical identifier information of the initial abnormal disturbance source into the report header; The propagation chain is visualized as a directed graph, with each node labeled with its anomaly type, intensity, and activation timestamp. Calculate the contribution of each node on the propagation link to the overall abnormal state, and generate a contribution distribution table; The report header, the directed graph of the visualization description, and the contribution distribution table are integrated to generate the structured source tracing report.

8. The intelligent monitoring and early warning system based on genetic algorithm according to claim 1, characterized in that, The automatic derivation and issuance of a set of control instructions for the sampling frequency and alarm threshold of the monitoring equipment based on the content of the source tracing report includes: Analyze the source tracing report to extract the key stages of the abnormal propagation and the types of sensors affected; Query the monitoring equipment configuration database to obtain the sampling frequency and alarm threshold setting value of the currently affected sensor type; Based on the speed and range data of abnormal propagation, the minimum sampling frequency required for the key stage is recalculated through a feedback control algorithm to ensure the capture of abnormal transient characteristics, and the dynamic alarm threshold of each sensor node is recalculated to filter false alarms. The recalculated sampling frequency value and dynamic alarm threshold are encapsulated into executable configuration instructions; The configuration command is sent to the corresponding monitoring device of the affected sensor type via a secure communication protocol.

9. The intelligent monitoring and early warning system based on genetic algorithm according to claim 8, characterized in that, The velocity and range data based on anomaly propagation are used to recalculate the minimum sampling frequency required for the critical stages through a feedback control algorithm, including: An anomaly propagation dynamics model is established, with the speed and range data of the anomaly propagation used as model input; Set an objective function that requires the sampling system to be able to distinguish the movement details of the anomaly front during the critical phase; The objective function is solved using the gradient descent method, resulting in a set of sampling frequency sequences that theoretically meet the resolution requirements; Considering the hardware limitations and network bandwidth constraints of the monitoring equipment, the sampling frequency sequence is modified to obtain the minimum sampling frequency.

10. The intelligent monitoring and early warning system based on genetic algorithm according to claim 8, characterized in that, The recalculation of the dynamic alarm thresholds for each sensor node to filter false alarms includes: Collect the reading distribution of the affected sensor types under historical normal conditions and the reading distribution during this abnormal event; Gaussian mixture models are used to model the reading distribution under the historical normal state and the reading distribution during the current abnormal event, respectively. Calculate the difference in probability distribution between the two models and find the decision boundary that maximizes the distinction between normal and abnormal states; The decision boundary is transformed into a specific threshold value, which serves as the dynamic alarm threshold. For multi-parameter sensors, the corresponding dynamic alarm threshold is calculated and set independently for each key parameter.

Images