An equipment abnormality warning system based on AI intelligent analysis

Abstract

The present invention discloses an equipment anomaly warning system based on AI intelligent analysis, which specifically relates to the field of equipment group warning and is used to solve the problem of rigid warning strategies in existing group warning systems. The spatial deployment information of each equipment node in the collaborative perception network is collected through the adjacency association module to construct a spatial adjacency structure; the data acquisition module organizes the response data within the historical monitoring cycle and extracts a subset of local collaborative response sequences according to the adjacency structure; the model training module constructs a local collaborative response prediction model based on the response sequence; the drift identification module uses the model to identify drift anomaly nodes with a trend of residual growth in real-time responses; the diffusion evaluation module uses the residual gradient diffusion as the basis to evaluate the degree of collaborative interference to the adjacent domain; the warning adjustment module dynamically adjusts the warning participation weights of the drift nodes and the interfering adjacent nodes based on the disturbance impact results, thereby realizing adaptive optimization of the equipment group warning strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment group early warning technology, and more specifically, the present invention relates to an equipment abnormality early warning system based on AI intelligent analysis. Background Art

[0002] In high-risk environments such as forest fires, plateau freeze-thaw disasters, large reservoir flooding, and chemical leaks, early warning systems typically utilize a multi-point coordinated device sensing network to continuously monitor the region using a variety of sensors, including temperature, displacement, gas concentration, and pressure. These sensing nodes must maintain spatially consistent responses to form a comprehensive basis for assessing the evolution of the disaster. However, in real-world deployments, due to factors such as terrain variations, electromagnetic interference, aging equipment, unstable power supplies, and uneven microclimates, some nodes may experience non-hardware failures such as response delays, amplitude jitter, or signal fluctuations. These nodes may still operate and output data, but their response behavior significantly deviates from the main trend of their neighboring nodes, resulting in low-synchronization interference drift anomalies. These anomalies are often misinterpreted as data fluctuations or individual differences, making them difficult for existing early warning algorithms to detect. Once these interfering nodes contribute to the overall disaster assessment, the early warning system can easily misinterpret a small-scale emergency as a regional spread or misinterpret key early warning signs as noise, delaying emergency response. The consequences are particularly severe in situations such as the early stages of a wildfire, the initial occurrence of localized water seepage in a dam, or trace leaks of toxic gas.

[0003] Therefore, how to provide a highly reliable early warning system based on AI to achieve self-optimization, which can effectively identify and isolate such drifting interference device nodes and build an early warning credibility correction mechanism based on cross-node relationships, becomes the key to improving the accuracy of group early warning. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an equipment abnormality warning system based on AI intelligent analysis to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An AI-based equipment anomaly warning system includes an adjacency association module, a data acquisition module, a model training module, a drift identification module, a diffusion assessment module, and a warning adjustment module, wherein:

[0007] The adjacency association module obtains the location information of each device node in the collaborative perception network of group warning and establishes the node adjacency structure according to the spatial layout of the devices;

[0008] The data acquisition module collects the original response data of the device nodes within the historical early warning monitoring period, and generates a local coordinated response sequence subset centered on the target node according to the node adjacency structure;

[0009] The model training module inputs the response sequence subset and node adjacency structure into the graph neural network to train the local collaborative response prediction model;

[0010] The drift identification module obtains the real-time response data of all device nodes in the current early warning monitoring cycle, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes;

[0011] The diffusion assessment module performs disturbance diffusion calculation in the adjacent area for the drift abnormal node and identifies the coordinated interference to the adjacent domain;

[0012] The early warning adjustment module adjusts the early warning decision participation weights of the drift abnormal nodes and the corresponding collaborative interference adjacent nodes in the new group early warning monitoring cycle.

[0013] In a preferred embodiment, the adjacency association module obtains the location information of each device node in the collaborative perception network for group early warning, and establishes a node adjacency structure according to the spatial layout of the devices, specifically including:

[0014] Mark all device nodes in the collaborative sensing network for group warning with unique spatial identifiers;

[0015] Calculate the Euclidean distance between nodes based on the spatial deployment coordinates of the device nodes and establish the initial graph structure of the node spatial relationship;

[0016] In the initial graph structure, a set of adjacent node candidates whose distance to each node is lower than a set spatial threshold is screened, and a node adjacency structure including node spatial identification and adjacency relationship is established based on the adjacent node candidate set.

[0017] In a preferred embodiment, the data collection module collects the original response data of the device nodes within the historical early warning monitoring period, and generates a local coordinated response sequence subset centered on the target node according to the node adjacency structure, specifically including:

[0018] Screen historical monitoring periods whose response accuracy reaches the preset standard from archived historical warning events, and remove monitoring data with missing records;

[0019] Determine a response cooperative adjacency domain of each target node within a set number of hops based on the node adjacency structure, wherein the target node is a device node selected as a cooperative response center in the cooperative sensing network;

[0020] For each target node, extract the response data of all nodes in its response coordination adjacent domain during the historical monitoring period, and perform time synchronization correction on the response data;

[0021] The synchronous response data of each target node and its adjacent domain nodes are spliced ​​along the time axis to generate a complete local collaborative response sequence subset.

[0022] In a preferred embodiment, the model training module inputs the response sequence subset and the node adjacency structure into the graph neural network, and training the local collaborative response prediction model specifically includes:

[0023] The local collaborative response sequence subset corresponding to each target node is combined with its adjacent domain structure, converted into a node response value matrix according to a unified time window length, and a graph neural network training sample is constructed;

[0024] Perform graph convolution calculation on each training sample, perform adjacent response feature aggregation and propagation in the adjacency structure based on the node response value matrix, and generate the response prediction value of the target node through linear mapping;

[0025] The historical true response value of the target node at the same time is used as the supervision label, the mean square error function is used to define the residual loss and the batch training process is performed;

[0026] In each batch training, the response value matrix of the training sample is input, and after forward propagation and loss calculation, the gradient is back-propagated and the model parameters are updated. The iteration continues until the residual converges.

[0027] The model structure and parameters after training are retained to form a local collaborative response prediction model.

[0028] In a preferred embodiment, the drift identification module obtains real-time response data of all device nodes in the current early warning monitoring period, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes, specifically including:

[0029] Collect the real-time response value of each device node in the current early warning monitoring period, and organize it into a node response value matrix of the current early warning monitoring period according to the established node adjacency structure;

[0030] Taking the node response value matrix as the input basis, the local collaborative response prediction model is called to generate a set of real-time response prediction values ​​for each target node;

[0031] Calculate the difference residual between the predicted response value and the real-time response value of each target node and construct the residual matrix of the current early warning monitoring period;

[0032] The residual matrices of several consecutive early warning monitoring cycles are spliced ​​together, and the sliding growth slope of the residual is calculated according to the node dimension;

[0033] Identify and filter target nodes whose sliding growth slope shows a continuous upward trend, mark them as drift abnormal nodes and output the node identifier.

[0034] In a preferred embodiment, the diffusion assessment module performs disturbance diffusion calculation in adjacent areas on the drift abnormal node, and identifying the coordinated interference to the adjacent area specifically includes:

[0035] Obtain a set of neighboring nodes selected as collaborative response centers in the adjacent domain of each drift abnormal node, and define the set of neighboring nodes as the collaborative diffusion influence area;

[0036] The node residuals in the cooperative diffusion influence area are calculated in chronological order using the same calculation method as the difference residual of the target node. The sliding window difference is performed on the chronological node residual sequence to construct the residual growth gradient vector of each adjacent node.

[0037] The local average gradient is calculated for the residual gradient vectors of all adjacent nodes in the cooperative diffusion influence area to form the gradient diffusion coefficient affected by the current drift node disturbance;

[0038] The disturbance diffusion identification threshold of the drift abnormal node is set. If the gradient diffusion coefficient is higher than the disturbance diffusion identification threshold, it is judged that the drift abnormal node causes coordinated interference to the adjacent domain.

[0039] In a preferred embodiment, the warning adjustment module adjusts the warning decision participation weights of the drift abnormal node and the corresponding coordinated interference adjacent node in the new group warning monitoring cycle, specifically including:

[0040] Collect all drift anomaly nodes that cause coordinated interference to adjacent domains, extract their corresponding adjacent nodes to form interference adjacent nodes, build a warning weight adjustment list, and reset the warning decision-making participation rights of all drift anomaly nodes in the list to 0;

[0041] Calculate the weight adjustment range of the early warning decision-making participation of the interference adjacent node according to the preset weight adjustment function;

[0042] Based on the calculation result of the early warning decision participation weight adjustment range, perform weight reassignment operation on all device nodes in the early warning weight adjustment list to generate the updated early warning participation weight configuration;

[0043] The updated participation weight configuration is written into the group warning decision engine to weaken or exclude the influence of related device nodes in the subsequent group warning process.

[0044] In a preferred embodiment, the weight adjustment function is set as a weighted normalization function with the gradient diffusion coefficient as a penalty factor, and the corresponding adjustment amplitude calculation result is proportional to the gradient diffusion coefficient.

[0045] The technical effects and advantages of the equipment abnormality early warning system based on AI intelligent analysis of the present invention are as follows:

[0046] By constructing a collaborative perception network based on spatial adjacency structure and integrating a graph neural network model to model the collaborative response characteristics between local nodes, the system effectively improves the ability to perceive changes in the collaborative relationship between warning nodes. Compared with traditional solutions based on single-point anomaly identification, the system introduces collaborative response sequences of historical high-quality monitoring cycles during the model training phase, significantly enhancing the model's ability to express local dynamic patterns. During real-time monitoring, the model can output response prediction values ​​based on adjacency structure reasoning for each node and dynamically calculate collaborative deviation residuals based on real-time observations, thereby identifying drifting anomaly nodes with inconsistent response trends and avoiding single-point misjudgments. On this basis, by introducing a gradient diffusion evaluation mechanism, the interference trend of drifting nodes on adjacent nodes is quantitatively analyzed, and node participation weights distorted by collaborative disturbances are further eliminated. This overall optimizes the participation structure and credibility configuration of group warning decisions, providing a highly reliable early warning system based on AI self-optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a structural schematic diagram of an equipment abnormality early warning system based on AI intelligent analysis in the present invention. DETAILED DESCRIPTION

[0048] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention. Example 1

[0049] Figure 1 The present invention provides an equipment abnormality warning system based on AI intelligent analysis, including an adjacency association module, a data acquisition module, a model training module, a drift identification module, a diffusion assessment module, and a warning adjustment module, wherein:

[0050] The adjacency association module obtains the location information of each device node in the collaborative perception network of group warning and establishes the node adjacency structure according to the spatial layout of the devices;

[0051] The data acquisition module collects the original response data of the device nodes within the historical early warning monitoring period, and generates a local coordinated response sequence subset centered on the target node according to the node adjacency structure;

[0052] The model training module inputs the response sequence subset and node adjacency structure into the graph neural network to train the local collaborative response prediction model;

[0053] The drift identification module obtains the real-time response data of all device nodes in the current early warning monitoring cycle, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes;

[0054] The diffusion assessment module performs disturbance diffusion calculation in the adjacent area for the drift abnormal node and identifies the coordinated interference to the adjacent domain;

[0055] The early warning adjustment module adjusts the early warning decision participation weights of the drift abnormal nodes and the corresponding collaborative interference adjacent nodes in the new group early warning monitoring cycle.

[0056] The adjacency association module obtains the location information of each device node in the collaborative perception network of group early warning, and establishes a node adjacency structure according to the spatial layout of the devices.

[0057] All device nodes in the collaborative sensing network participating in the group early warning mission are labeled with unique spatial identifiers. During deployment, each device node is assigned specific spatial coordinate information. These coordinates are Cartesian coordinates in a three-dimensional coordinate system, including horizontal, vertical, and vertical height values, representing the device's spatial location in the actual monitoring scenario. To avoid ambiguity or overlap in node information in the subsequent graph structure, all device nodes are assigned unique spatial identifiers corresponding to their spatial coordinates. After the unique identifiers are labeled, a spatial relationship matrix between the device nodes is constructed based on the spatial coordinates of all device nodes. Specifically, the spatial relationship between any two nodes is measured using the Euclidean distance, calculated as the squared difference and square root of the three-dimensional coordinates. The calculated inter-node distances are stored in a symmetric distance matrix, where any entry in the matrix represents the actual physical distance between the two nodes. This distance matrix captures the spatial distribution relationships between all device nodes. The distance matrix is ​​then converted into a graph representation, where the device nodes are represented as nodes in the graph and the spatial distances between nodes are used as edge weights to construct an initial graph. The initial graph structure is an undirected graph, in which the edges between nodes represent the reachable or observable physical connection relationship in space. The edge weight is the normalized Euclidean distance between nodes. This graph structure retains the relative position relationship of device nodes in the real spatial layout.

[0058] After the initial graph structure is constructed, the adjacency relationship is screened and structured around each device node. First, based on the distance matrix, a traversal is performed on each node to extract all other nodes whose spatial distance to the node is less than the set spatial threshold. The setting of the spatial threshold is based on the perception range, communication distance or other physical action capabilities of the device in the collaborative perception network, for example, it is set to 5 meters, 8 meters or 10 meters. The selection of the spatial threshold must have an actual engineering basis and be specifically set under the constraints of the device type and perception scenario. In this embodiment, the default spatial threshold is the average value of the sum of the distances between a device and all other devices. When the distance between a device node and any other device node is less than this value, it is considered to have a potential collaborative relationship. All nodes with potential collaborative relationships are summarized to generate a node adjacency structure containing node spatial identification and adjacency relationships. This structure is a directed graph structure based on adjacency relationships.

[0059] The data acquisition module collects original response data of device nodes within a historical early warning monitoring period, and generates a local coordinated response sequence subset centered on a target node according to a node adjacency structure.

[0060] Archived warning event records in the collaborative sensing network are systematically processed to extract all event samples covering a complete monitoring cycle. Each warning event data entry includes the event occurrence time, device response record, actual fault confirmation time, and the final warning judgment result. Based on this information, evaluation metrics for warning response accuracy are defined, including response timeliness and response accuracy. Response timeliness refers to whether the device node responds within a set response window after the event occurs, for example, 30 seconds. Response accuracy measures whether the node's warning judgment is consistent with the final event category. For example, if the actual warning scenario is a high temperature, a record that is mistakenly classified as a medium or low temperature is considered an incorrect response. Statistics are performed on a per-event basis within the historical warning event database, and the response performance of all device nodes associated with each warning event is scored. Warning events with scores that meet set thresholds (e.g., response timeliness of at least 90% and response accuracy of at least 85%) covering the complete monitoring cycle are included in the valid sample set. At the same time, to ensure the integrity of the input data, the response data of each node in the sample is screened for integrity. If data is missing, abnormally interrupted, or timestamps are misaligned for some nodes in a monitoring cycle, the entire cycle is eliminated and not included in the subsequent analysis process. The missing data determination standard is: if the response data of a node within a monitoring cycle is missing for more than 10%, the warning monitoring cycle is deemed invalid.

[0061] The previously constructed node adjacency structure is called upon, containing the spatial identifiers of each device node and its adjacency relationships within the spatial structure. To implement local coordinated response modeling, a corresponding coordinated response adjacency domain is constructed for each target node. If key device nodes (such as high-precision master devices and key area devices) have been manually or rule-based designated as response centers during the alert deployment, the system generates adjacency domains for these nodes based on the designated node list. If no manual designation is made, the system defaults to treating all deployed nodes as target nodes one by one, forming a coordinated adjacency structure covering the entire network. When generating coordinated adjacency domains, the system limits the adjacency range based on the preset number of hops in the graph structure. The hop count is the maximum distance achievable along an adjacency edge from the target node. Setting the hop count to 2 includes the target node's direct neighboring nodes (one hop) and its neighboring nodes (two hops) within the adjacency domain. The hop count is determined based on the physical range or information perception boundary of the device coordinated response and can be set to an integer between 1 and 3 based on the actual alert scenario. A breadth-first search is performed on the target node with the number of hops as the search depth, and a complete collaborative adjacency domain is constructed without exceeding the hop limit.

[0062] After completing the collaborative neighboring domain construction, the system extracts response data from the list of neighboring domain nodes corresponding to each target node. Within the set of qualified historical monitoring periods, the system first extracts the target node's response data by time period. It then traverses each node in its neighboring domain, extracting its response value for the same time period. Response values ​​can be device status readings, event sensing levels, or digital signal values, depending on the sensor type of the devices deployed in the collaborative sensing network. To ensure temporal comparability of response data from different nodes, the system introduces a unified time synchronization mechanism. This synchronization process is accomplished through two steps: timestamp alignment and interpolation. First, the response data of all neighboring domain nodes is timestamp-aligned based on the network-wide master clock. Any node sampling times that are inconsistent or offset are aligned to the target node's master sampling time series. If any time points are missing, nearest neighbor interpolation or linear interpolation is used to fill in the missing data to ensure data continuity and completeness. After time synchronization, the system horizontally concatenates the response values ​​of the target node and all nodes in its neighboring domains in a unified timeline order, forming a two-dimensional local response series. Each row represents the response value of all nodes at the same moment, and the column dimension represents the topological order of the nodes in the adjacency domain. This response sequence constitutes the subset of the local coordinated response sequence of the target node, which serves as the core structural data of the input samples for graph neural network training.

[0063] The model training module inputs the response sequence subset and the node adjacency structure into the graph neural network to train the local collaborative response prediction model.

[0064] The time axis in the response sequence subset is truncated into a sliding time window of uniform length. The time window length is set to a fixed value, such as 120 time units, based on the historical sampling frequency of the node response data and the timeliness requirement of the early warning response. The response sequence of each adjacent node is clipped by sliding the time window to form multiple subsequence segments, and each subsequence segment corresponds to a set of temporally continuous response values. Taking each time window as a unit, the response subsequences of each node in the adjacent domain are horizontally spliced ​​to form a two-dimensional node response value matrix. The number of columns of the matrix is ​​equal to the number of adjacent nodes, and the number of rows is equal to the length of the time window. Without changing the previous topology, the time series response data and the adjacent topology information are integrated into a format of training samples that can be directly accepted by the graph neural network.

[0065] After constructing the graph neural network training samples, graph convolution is performed to extract collaborative response features in the local adjacency domain and predict the future response trends of the target node. Specifically, the adjacency structure in each training sample is first encoded and mapped into a sparse adjacency matrix based on the established node adjacency structure. This matrix is ​​used to identify the topological connection relationship between the target node and its adjacent nodes. This adjacency matrix serves as the structural input in the graph convolution calculation to ensure that the information propagation path is consistent with the actual physical deployment relationship. At the same time, the node response value matrix corresponding to the sample is input into the model. Each row in this matrix corresponds to the continuous response value sequence of a node in the adjacency domain within a fixed time window, and each column corresponds to the response observation of all nodes at a unified time. The above structural input and the specific element values ​​in the response value matrix together constitute the collaborative response feature input in the graph convolution process, which is used to describe the state evolution pattern of each adjacent node in the current time window. During the graph convolution propagation process, adjacent response feature aggregation centered on the target node is performed. Specifically, for the current target node, its adjacent node set is traversed, and the response value sequence of the corresponding adjacent node is extracted based on the connection relationship in the adjacency matrix, and the weight matrix is ​​applied to perform a linear transformation on its response characteristics. The weight matrix is ​​a learnable parameter, which is initially set to a random distribution and iteratively optimized during the training process. The response characteristics after linear transformation are processed by the activation function, and the ReLU function is selected to maintain the nonlinear expression ability. The activated features are then weighted aggregated. The aggregation method is designed according to the graph neural network model. Common methods include mean aggregation, weighted sum aggregation or maximum aggregation. The aggregation operation reflects the target node's ability to synergistically integrate adjacent response features. After the aggregation is completed, the target node representation vector obtained enters the linear mapping layer, performs a weight transformation and outputs a single response prediction value.

[0066] After completing the graph convolution calculation and response prediction output process for the training samples, a supervised learning mechanism is introduced to evaluate and optimize the model's predictive capabilities. The historical actual response value of each target node at the time corresponding to its predicted value is used as a supervisory label, forming the target output portion of the training data. These supervisory labels are temporally consistent with the response prediction value output by the model, ensuring that the model learns a precise response estimate at a specific moment, rather than fuzzy intervals or lagged data. To measure the degree of deviation between the model output value and the supervisory label, the mean squared error function is selected as the loss metric in this embodiment. This function provides continuous and differentiable residual feedback on the model prediction error, providing a clear path for gradient calculation in backpropagation. During the training process, a batch training mechanism is used to optimize the parameters of the graph neural network model. In batch training, multiple training samples are combined into a training batch according to a set batch size. The samples in each batch are uniformly input into the model for forward propagation. During the forward propagation process, the node response value matrix of the training samples is first input into the model structure. The graph convolution layer aggregates the response features of the adjacent domains and propagates them. The linear mapping layer then outputs the response prediction value for each target node. Subsequently, the residual of each target node's predicted value and its supervisory label is calculated, and the mean squared error loss for the current batch is calculated based on the error output across the entire batch. After the loss calculation is complete, the backpropagation mechanism is initiated, and the model parameters are gradient-derived and updated using the loss value. All weight matrices and bias terms involved in training in the model are numerically adjusted based on the error gradient of the batch. This update utilizes an adaptive gradient descent strategy to accelerate parameter convergence and avoid getting stuck in local optima. The training process is iteratively executed in rounds, with all training samples traversed in each round to complete a complete parameter update cycle. After each round of training, the model's residual performance on the validation samples is evaluated. If the validation residual shows no significant decrease over several consecutive rounds or reaches a preset residual convergence threshold, training is terminated. The residual convergence threshold is set based on the historical response fluctuation level; for example, setting the average predicted residual to less than 15% of the node response standard deviation effectively ensures the model's predictive accuracy in real-world applications.

[0067] After training, the model structure and parameter configuration output from the final round of training are retained to form a local collaborative response prediction model for subsequent real-time reasoning and anomaly identification. This model is capable of inferring target node response trends based on known node adjacency structures and recent response values, serving as a core prediction support module for subsequent early warning judgment and anomaly drift detection. Once the model structure and parameters are solidified, they can be deployed in edge processing nodes or central processing units, and can be regularly incrementally trained or migrated based on new data.

[0068] The drift identification module obtains the real-time response data of all device nodes in the current early warning monitoring period, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes.

[0069] During the current early warning monitoring cycle, real-time response data is first acquired from all device nodes in the collaborative sensing network. This response data consists of numerical status indicators collected by the nodes based on their local monitoring mechanisms during the current time period. These parameters, such as temperature, current, vibration amplitude, and flow rate, are determined based on the attributes of the monitored object. To ensure data temporal consistency, a unified time base is used for data synchronization to eliminate the impact of sampling time differences between nodes. Subsequently, based on the previously constructed node adjacency structure, the response values ​​of each node during the current cycle are arranged in columns according to the node number of the adjacent domain corresponding to each target node, forming a node response value matrix for the current early warning monitoring cycle. Each row in this matrix corresponds to a time point, and each column corresponds to the response value of a device node within a corresponding adjacent domain, thus fully reflecting the real-time response status of each target node within its collaborative adjacent domain. Based on the trained local collaborative response prediction model, this constructed node response value matrix is ​​used as input. The model automatically calls upon its internally stored graph neural network structure and parameters, aggregates the node response features in the adjacent structure of each target node through a graph convolutional propagation mechanism, and outputs a predicted value through a linear mapping layer.

[0070] It should be noted that the predicted value is the model's immediate estimation result of the target node response trend under the current input data structure, reflecting the node behavior expectations learned by the model based on historical collaborative relationships. It is used to compare with the real-time value to find drifting abnormal nodes, rather than being a real prediction value at a certain moment in the future.

[0071] The difference between the predicted response value and the actual real-time response value for each target node is calculated to obtain a set of node-wise residual values, which are then used to construct the residual matrix for the current early warning monitoring cycle. This residual matrix records the degree of deviation between the predicted and actual values ​​for each node along the time axis. Its matrix dimensions are consistent with the response value matrix to maintain time synchronization. To dynamically evaluate the changing trend of node prediction deviations over multiple consecutive monitoring cycles, the residual matrices for several consecutive cycles (for example, the last five cycles) are concatenated in chronological order to form a set of node residual time series. In this set, each node corresponds to a time-varying residual series. Based on this series, sliding window processing is performed and the time series slope is calculated. This is the rate of change of the residual value within the local time window, forming a sliding growth slope indicator for each node. This slope reflects whether the prediction residual of a node is continuously increasing. Target nodes with a continuously increasing slope indicator are screened according to the node dimension and marked as drifting anomalous nodes, with the node identifier output.

[0072] The diffusion assessment module performs disturbance diffusion calculation in adjacent areas on the drift abnormal nodes and identifies the coordinated interference to the adjacent areas.

[0073] For each device node marked as drifting anomaly, a pre-established node adjacency structure is invoked to identify the set of all device nodes within its adjacent range that exhibit collaborative response characteristics. In this scheme, the adjacent range is limited to a set of adjacent nodes with a topological hop count of no more than two, and the device nodes used for local collaborative prediction modeling are defined as collaborative response center nodes. This set is used as the collaborative diffusion influence area for the current drifting node. From consecutive historical early warning monitoring cycles, the response prediction residuals of all collaborative response center nodes within the collaborative diffusion influence area are extracted. To ensure the integrity of the analysis time series, the extracted data must cover at least five consecutive early warning monitoring cycles, and the length of each cycle must be consistent with the time window length used in graph neural network training. A sliding window difference operation is performed on the residual values ​​of each adjacent node in this time series, with a sliding window length of three cycles and a step size of one cycle. This generates a residual growth gradient vector for each adjacent node. This vector reflects the acceleration trend of the node's residual change within the local cycle and is used to characterize the incremental response anomaly during the anomaly diffusion process. The residual growth gradient vectors of all adjacent nodes in the cooperative diffusion influence area are uniformly processed. By taking the average of each gradient vector element by element position, a local average gradient value corresponding to the current drifting abnormal node is generated, which serves as an indicator of the perturbation diffusion intensity of this node in the adjacent domain. For the sake of uniform expression, this indicator is defined in this paper as the gradient diffusion coefficient. The higher the value, the more significant the tendency of the drifting node to cause cooperative response perturbations on adjacent nodes.

[0074] To determine whether the gradient diffusion coefficient constitutes coordinated interference, a disturbance diffusion identification threshold must be set. This threshold is derived from retrospective statistics of historical anomalies, and the recommended initial value is the mean of the normal residual growth gradient of adjacent nodes plus two standard deviations. If the gradient diffusion coefficient of a drifting node exceeds this threshold, the system determines that it has a positive coordinated interference effect, which constitutes coordinated interference with the prediction and judgment results in the adjacent domain.

[0075] The warning adjustment module adjusts the warning decision participation weights of the drift abnormal node and the corresponding collaborative interference adjacent node in a new group warning monitoring cycle.

[0076] Collect all drifting abnormal nodes that constitute collaborative interference to the adjacent domain and include them in the set of interference source nodes. Specifically, based on the nodes whose gradient diffusion coefficients have been identified above and exceed the disturbance diffusion identification threshold, determine the numbers of all interfered nodes in the adjacent domain in the spatial topological structure. By traversing the adjacent domain of each disturbance source node, extract the nodes marked as collaborative response centers in the domain and record them as interference adjacent nodes to ensure that the formed interference adjacent node set has a binding relationship with the corresponding mapping topology and the weight configuration participating in the current cycle. The set needs to be screened for uniqueness, remove possible duplicate nodes, and jointly construct an early warning weight adjustment list with the interference source node. This list will serve as the object set for subsequent weight correction operations and will directly participate in the weight parameter update.

[0077] In the warning weight adjustment list, all device nodes identified as drifting abnormal nodes must have their participation weights in this round of group warnings reset to 0. The technical intent of this operation is to completely eliminate the impact of nodes that continuously drift and cause significant interference on group warning outputs, preventing the spread of false alarms or misleading collective judgment. The reset operation should also synchronize with the update of the weight storage structure to ensure the consistency of the total weight of the current participating nodes. The original weight value should be recorded during the implementation as a basis for recovery judgment in subsequent cycles.

[0078] For adjacent nodes experiencing interference, the adjustment range for their warning participation weights must be determined based on the intensity of the disturbance from the source node. To this end, the system calls a preset weight adjustment function to perform weight decay calculations. This weight adjustment function uses a nonlinear decay function based on the normalized gradient diffusion coefficient. When setting the function, it is necessary to specify that the gradient diffusion coefficient is used as the penalty factor, and the calculated adjustment range is proportional to the gradient diffusion coefficient. The specific form can be set as a piecewise threshold, exponential decay, or a dynamic function model weighted by the frequency of historical periodic coordinated disturbances. Linear normalization is introduced to unify the numerical distribution, ensuring that the disturbance index results for all nodes in subsequent calculations fall within the [0, 1] range and avoid distortion of inter-node influence due to dimensional differences. When setting the function, the weight reduction threshold principle must be clearly defined. For example, nodes with a diffusion coefficient above 0.3 are assigned a high interference level, with a corresponding warning weight decay rate of no less than 60%. Nodes with a diffusion coefficient below 0.1 are assigned a weak interference level, with their warning weight retained at least 80%. After all weight decay ranges are calculated, the weights are reassigned for all nodes in the warning weight adjustment list. This operation requires overwriting the original weight configuration to ensure that the newly generated participation weight configuration can be directly used in the current or subsequent group warning cycle. The weight configuration is written to the warning inference engine. The specific write operation should write the new configuration to the memory mapping table and the periodic scheduling control table respectively to ensure that the group warning engine always uses the latest configuration when calling the weight participation calculation. Once the updated warning participation weight configuration is written and takes effect, the actual influence of each node will be automatically adjusted in the subsequent group warning process. All adjacent nodes that are in an abnormal drift state or are disturbed by it will lose their dominant role due to their weights being weakened or cleared, effectively improving the stability and anti-interference ability of the group warning judgment, while leaving room for the next stage of node recovery judgment and dynamic re-empowerment. This mechanism is highly adaptable in handling multi-node collaborative anomaly scenarios and is suitable for large-scale distributed warning network structures deployed in engineering.

[0079] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0080] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0081] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0084] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0085] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0086] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0087] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0088] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An equipment abnormality warning system based on AI intelligent analysis, characterized in that: It includes an adjacency association module, a data acquisition module, a model training module, a drift identification module, a diffusion assessment module, and an early warning adjustment module, among which: The adjacency association module obtains the location information of each device node in the collaborative perception network of group warning and establishes the node adjacency structure according to the spatial layout of the devices; The data acquisition module collects the original response data of the device nodes within the historical early warning monitoring period, and generates a local coordinated response sequence subset centered on the target node according to the node adjacency structure; The model training module inputs the response sequence subset and node adjacency structure into the graph neural network to train the local collaborative response prediction model; The drift identification module obtains the real-time response data of all device nodes in the current early warning monitoring cycle, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes; The diffusion assessment module performs disturbance diffusion calculation in the adjacent area for the drift abnormal node and identifies the coordinated interference to the adjacent domain; The warning adjustment module adjusts the warning decision-making participation weights of the drift abnormal nodes and the corresponding collaborative interference adjacent nodes in the new group warning monitoring cycle; The drift identification module obtains the real-time response data of all device nodes in the current early warning monitoring period, calls the local collaborative response prediction model to calculate the real-time residual, and identifies the drift abnormal nodes, specifically including: Collect the real-time response value of each device node in the current early warning monitoring cycle, and organize it into a node response value matrix of the current early warning monitoring cycle according to the established node adjacency structure; Taking the node response value matrix as the input basis, the local collaborative response prediction model is called to generate a set of real-time response prediction values ​​for each target node; Calculate the difference residual between the predicted response value and the real-time response value of each target node and construct the residual matrix of the current early warning monitoring period; The residual matrices of several consecutive early warning monitoring cycles are spliced ​​together, and the sliding growth slope of the residual is calculated according to the node dimension; Identify and select target nodes where the sliding growth slope shows a continuous upward trend, mark them as drift abnormal nodes, and output the node identifier; The diffusion assessment module performs disturbance diffusion calculation in adjacent areas on the drift abnormal node, and identifies the coordinated interference to the adjacent area, specifically including: Obtain a set of neighboring nodes selected as collaborative response centers in the adjacent domain of each drift abnormal node, and define the set of neighboring nodes as the collaborative diffusion influence area; The node residuals in the cooperative diffusion influence area are calculated in chronological order using the same calculation method as the difference residual of the target node. The sliding window difference is performed on the chronological node residual sequence to construct the residual growth gradient vector of each adjacent node. The local average gradient is calculated for the residual gradient vectors of all adjacent nodes in the cooperative diffusion influence area to form the gradient diffusion coefficient affected by the current drift node disturbance; The disturbance diffusion identification threshold of the drift abnormal node is set. If the gradient diffusion coefficient is higher than the disturbance diffusion identification threshold, it is judged that the drift abnormal node causes coordinated interference to the adjacent domain.

2. The device abnormality warning system based on AI intelligent analysis according to claim 1 is characterized in that: The adjacency association module obtains the location information of each device node in the collaborative perception network of group early warning and establishes the node adjacency structure according to the device spatial layout, specifically including: Mark all device nodes in the collaborative sensing network for group warning with unique spatial identifiers; Calculate the Euclidean distance between nodes based on the spatial deployment coordinates of the device nodes and establish the initial graph structure of the node spatial relationship; In the initial graph structure, a set of adjacent node candidates whose distance to each node is lower than a set spatial threshold is screened, and a node adjacency structure including node spatial identification and adjacency relationship is established based on the adjacent node candidate set.

3. The device abnormality warning system based on AI intelligent analysis according to claim 1 is characterized in that: The data acquisition module collects the original response data of the device nodes within the historical early warning monitoring period, and generates a local coordinated response sequence subset centered on the target node according to the node adjacency structure, specifically including: Screen historical monitoring periods whose response accuracy reaches the preset standard from archived historical warning events, and remove monitoring data with missing records; Determine a response cooperative adjacency domain of each target node within a set number of hops based on the node adjacency structure, wherein the target node is a device node selected as a cooperative response center in the cooperative sensing network; For each target node, extract the response data of all nodes in its response coordination adjacent domain during the historical monitoring period, and perform time synchronization correction on the response data; The synchronous response data of each target node and its adjacent domain nodes are spliced ​​along the time axis to generate a complete local collaborative response sequence subset.

4. The device abnormality warning system based on AI intelligent analysis according to claim 1 is characterized in that: The model training module inputs the response sequence subset and the node adjacency structure into the graph neural network, and training the local collaborative response prediction model specifically includes: The local collaborative response sequence subset corresponding to each target node is combined with its adjacent domain structure, converted into a node response value matrix according to a unified time window length, and a graph neural network training sample is constructed; Perform graph convolution calculation on each training sample, perform adjacent response feature aggregation and propagation in the adjacency structure based on the node response value matrix, and generate the response prediction value of the target node through linear mapping; The historical true response value of the target node at the same time is used as the supervision label, the mean square error function is used to define the residual loss and the batch training process is performed; In each batch training, the response value matrix of the training sample is input, and after forward propagation and loss calculation, the gradient is back-propagated and the model parameters are updated. The iteration continues until the residual converges. The model structure and parameters after training are retained to form a local collaborative response prediction model.

5. The device abnormality warning system based on AI intelligent analysis according to claim 1 is characterized in that: The warning adjustment module adjusts the warning decision participation weights of the drift abnormal node and the corresponding coordinated interference adjacent node in the new group warning monitoring cycle, specifically including: Collect all drift anomaly nodes that cause coordinated interference to adjacent domains, extract their corresponding adjacent nodes to form interference adjacent nodes, build a warning weight adjustment list, and reset the warning decision-making participation rights of all drift anomaly nodes in the list to 0; Calculate the weight adjustment range of the early warning decision-making participation of the interference adjacent node according to the preset weight adjustment function; Based on the calculation result of the early warning decision participation weight adjustment range, perform weight reassignment operation on all device nodes in the early warning weight adjustment list to generate the updated early warning participation weight configuration; The updated participation weight configuration is written into the group warning decision engine to weaken or exclude the influence of related device nodes in the subsequent group warning process.

6. The device abnormality warning system based on AI intelligent analysis according to claim 5 is characterized in that: The weight adjustment function is set as a weighted normalization function with the gradient diffusion coefficient as a penalty factor, and the corresponding adjustment amplitude calculation result is proportional to the gradient diffusion coefficient.

Images