Abnormal Analysis Method and System Based on Blasting Data

Abstract

The present invention relates to the technical field of anomaly detection, and discloses an anomaly analysis method and system based on blasting data, including collecting blasting data of each monitoring node, dividing the monitoring nodes into multiple regions, statistically calculating the anomaly density of each region within a preset historical period, and dynamically setting the regional weight of the training samples of the autoencoder according to the anomaly density; reconstructing the current round of data by using the weighted-trained autoencoder, calculating the residual of each node, and extracting the activation path information of the nodes with residuals higher than the initial screening threshold of the reconstruction error; constructing a graph structure model between the monitoring nodes, and calculating the anomaly propagation path according to the correlation between the nodes; matching the activation path with the propagation path to determine the anomaly state of the target node. It significantly improves the accuracy, stability and intelligent feedback ability of blasting data anomaly analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of anomaly detection, and specifically to an anomaly analysis method and system based on blasting data. Background Art

[0002] Currently, with the continuous improvement of industrial automation and digitalization levels, data processing systems have been widely applied in various fields. Especially in complex engineering environments, the acquisition, processing, and anomaly detection of real-time data have become key technologies to ensure the safe and stable operation of the system. In recent years, information processing methods implemented based on digital computers have been widely adopted in fields such as blasting engineering and geotechnical engineering for real-time monitoring and anomaly detection of multi-dimensional data such as vibration, acoustic emission, stress, and temperature collected during the blasting process.

[0003] In the prior art, most methods rely on traditional signal processing algorithms and statistical analysis techniques to preprocess data and perform threshold judgment. Although anomalies can be detected to a certain extent, when dealing with complex and variable blasting data environments, there are problems such as low detection accuracy, sensitivity to noise and data fluctuations, and insufficient analysis of the internal structure and spatio-temporal correlation of data. On the other hand, with the development of artificial intelligence and machine learning algorithms, deep learning models, especially autoencoders, have shown great advantages in data compression and reconstruction. However, their application in blasting data anomaly detection still faces challenges such as insufficient model training, unclear regional focus, and failure to capture the anomaly propagation law. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the technical problems of inaccurate detection of structural anomalies, insufficient recognition of regional characteristics, and insufficient interpretability of anomaly propagation paths in the existing blasting data anomaly analysis methods.

[0006] To solve the above technical problems, the present invention provides the following technical solution: An anomaly analysis method based on blasting data, including collecting blasting data of each monitoring node and dividing the monitoring nodes into multiple regions;

[0007] Statistical analysis of the anomaly density in each region within a preset historical period, and dynamically setting the regional weight of the training samples of the autoencoder according to the anomaly density;

[0008] Reconstructing the current round of data using the weighted-trained autoencoder, calculating the reconstruction error of each monitoring node, and extracting the activation path information of the nodes with a reconstruction error higher than the initial screening threshold of the reconstruction error;

[0009] Constructing a graph structure model between monitoring nodes and calculating the anomaly propagation path according to the correlation between nodes;

[0010] Match the activation path with the propagation path to determine the abnormal state of the target node.

[0011] As a preferred solution of the abnormal analysis method based on blasting data according to the present invention, wherein: the blasting data includes vibration signals, acoustic emission signals, stress data and temperature;

[0012] Dividing the monitoring nodes into multiple regions includes extracting temporal features from the blasting data collected by each monitoring node to form a feature sequence corresponding to each node; calculating the similarity between any two monitoring nodes using the Pearson correlation coefficient, and constructing a similarity matrix between nodes;

[0013] Take the similarity matrix as an undirected graph, where each monitoring node is used as a node in the graph, and the edges between nodes represent the degree of similarity of data features; use the spectral clustering algorithm to partition the undirected graph to form several sub-regions, so that the monitoring nodes within each sub-region are correlated in data behavior.

[0014] As a preferred solution of the abnormal analysis method based on blasting data according to the present invention, wherein: setting the regional weight of the autoencoder training samples includes counting the number of abnormal nodes in each sub-region during the historical period, and calculating the abnormal density of each sub-region; based on the difference between the abnormal density of each sub-region and the average abnormal density of all sub-regions, determine the sample weight coefficient corresponding to each sub-region;

[0015] During the training process of the autoencoder, apply the sample weight coefficients to the training samples in different sub-regions respectively.

[0016] As a preferred solution of the abnormal analysis method based on blasting data according to the present invention, wherein: reconstructing the current round of data includes using the trained autoencoder to reconstruct the blasting data collected in the current round, and calculating the reconstruction error e i , where the reconstruction error is the difference between the original data and the reconstructed data of the monitoring node.

[0017] As a preferred solution of the abnormal analysis method based on blasting data according to the present invention, wherein: extracting activation path information includes determining whether the reconstruction error of each node exceeds the reconstruction error initial screening threshold T1, and when it exceeds the reconstruction error initial screening threshold T1, determining it as a high reconstruction error node; extracting the activation path information H of the high reconstruction error node from the input layer to the hidden layer in the autoencoder structure i , where the activation path information is the activation value sequence of each hidden layer of the high reconstruction error node during the encoding stage.

[0018] As a preferred solution of the anomaly analysis method based on blasting data according to the present invention, wherein: constructing a graph structure model between monitoring nodes includes that the edge between nodes represents the feature similarity value between any two monitoring nodes;

[0019] The calculation of the anomaly propagation path according to the correlation between nodes includes that for each node i determined to have a high reconstruction error, its adjacent node set N in the graph structure is extracted i ; In N i , nodes with a reconstruction error exceeding the initial screening threshold T1 of the reconstruction error are screened out to form the initial anomaly propagation path node set S of the high reconstruction error node i i ;

[0020] When the high reconstruction error node i, all adjacent nodes N of the high reconstruction error node i i , and the node set together form a closed connected subgraph, it is determined that the propagation path has stable consistency in the graph structure, and the propagation path is marked as a high-confidence propagation path;

[0021] The node set is: a set composed of nodes that have a direct connection relationship with any node in N i but do not have a direct connection with the high reconstruction error node i itself;

[0022] Nodes in the initial anomaly propagation path node set S i but not included in the final high-confidence propagation path node set are set as the low-confidence propagation path candidate set L i ;

[0023] The final anomaly propagation path node set S' i includes, from the high-confidence propagation path node set, sorting in descending order based on the feature similarity value |r ij | between each propagation node j and the high reconstruction error node i, and selecting the top n nodes with the highest feature similarity value; and latent induced anomaly nodes identified from the low-confidence propagation path candidate set L i that meet the inducing factor ; wherein, the inducing factor G m is composed of neighborhood anomaly density, reconstruction error volatility, and temporal drift degree of graph embedding features, represents the mean value of G m , represents the standard deviation of G m .

[0024] As a preferred solution of the anomaly analysis method based on blasting data according to the present invention, wherein: determining the anomaly state of the target node includes calculating the Jaccard similarity to calculate Hi Similarity with S' i Sim i ;

[0025] When Sim i ≥ T2 and e i ≥ T1, it is determined as a structurally abnormal node, indicating that the structurally abnormal node shows a consistent abnormal trend in both internal feature response and external structure propagation of the model;

[0026] When Sim i < T2 and e i ≥ T1, it is determined as a sudden abnormal node, indicating that although the sudden abnormal node shows significant abnormality in the model, it does not form a coordinated diffusion in the structure diagram;

[0027] When e i < T1 and multiple adjacent nodes of the node are the initial abnormal propagation path nodes, it is determined as a latent abnormal node, indicating that the latent abnormal node is in an environment with a high incidence of abnormalities and needs to be focused on.

[0028] Anomaly analysis system based on blasting data, wherein:

[0029] Data module, collecting blasting data of each monitoring node and dividing the monitoring nodes into multiple regions;

[0030] Training module, counting the anomaly density in each region within a preset historical period and dynamically setting the regional weights of the training samples of the autoencoder according to the anomaly density;

[0031] Reconstruction module, reconstructing the current round of data using the weighted-trained autoencoder, calculating the reconstruction error of each monitoring node, and extracting the activation path information of the nodes whose reconstruction error is higher than the reconstruction error initial screening threshold;

[0032] Propagation module, constructing a graph structure model between monitoring nodes and calculating the abnormal propagation path according to the correlation between nodes;

[0033] Confirmation module, matching the activation path and the propagation path to calculate the similarity and determining the abnormal state of the target node.

[0034] Computer device, including a memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of any method described in the present invention are implemented.

[0035] Computer-readable storage medium, on which a computer program is stored, characterized in that: when the computer program is executed by a processor, the steps of any method described in the present invention are implemented.

[0036] Advantages of the present invention: By constructing a graph structure model based on blasting data and introducing a regional division and abnormal density guiding mechanism, the self-encoder is adaptively trained and optimized for key regions, improving the model's recognition ability for high-risk regions. Combining the judgment of reconstruction error and the extraction of activation paths can accurately mine the internal response characteristics of the model. Further, through the recognition of propagation paths and the matching of Jaccard similarity, the linkage judgment of the structural consistency and propagation trend of abnormal nodes is realized. Finally, a classification mechanism for structural, sudden, and latent anomalies is established, making the anomaly detection results more interpretable and refined, and significantly improving the accuracy, stability, and intelligent feedback ability of blasting data anomaly analysis. Description of the Drawings

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0038] Figure 1 It is the overall flowchart of the anomaly analysis method based on blasting data provided by the first embodiment of the present invention. Detailed Embodiments

[0039] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed embodiments of the present invention with reference to the drawings of the specification. Obviously, the described embodiments are some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0040] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides an anomaly analysis method based on blasting data, including,

[0041] S1: Collect the blasting data of each monitoring node and divide the monitoring nodes into multiple regions.

[0042] The blasting data includes vibration signals, acoustic emission signals, stress data, and temperature.

[0043] First, analyze and process the blasting data collected by each monitoring node, extract the time series features, and form the feature sequence of each monitoring node, denoted as:

[0044]

[0045] where X iDenote the data sequence of the \(i\)-th monitoring node over \(T\) time steps, which may include vibration amplitude, acoustic emission amplitude, stress change, or temperature.

[0046] Subsequently, calculate the similarity between any two monitoring nodes based on the feature sequences of each node. The Pearson correlation coefficient is defined as follows:

[0047]

[0048] where and are the feature means of nodes \(i\) and \(j\) respectively. The calculation results of \(\rho\) between all node pairs form a symmetric similarity matrix \(R\). ij The calculation results are composed into a symmetric similarity matrix \(R\).

[0049] Next, take the similarity matrix \(R\) as the adjacency matrix of an undirected graph, and construct the graph structure \(G=(V, E)\), where each monitoring node is a node \(v\) in the graph i \(\in V\). If \(|\rho\) ij |\geq0\), then an edge \(e\) i is established between nodes \(v\) j and \(v\) ij \(\in E\), and the edge weight is taken as the absolute value or the original value of \(\rho\) ij . \(V\) represents the node set, and \(E\) represents the edge set.

[0050] Finally, based on the graph structure, use the spectral clustering algorithm to cluster and partition the monitoring nodes to form several sub-regions. The monitoring nodes within each sub-region have strong consistency in historical behavior patterns, providing a regional basis for subsequent training sample weighting.

[0051] Furthermore, by quantifying the temporal features of the blasting data and modeling the similarity, representing the behavioral correlation between monitoring nodes with a graph structure, and introducing the spectral clustering algorithm to complete the intelligent regional division in space, it no longer relies on manual geographical partitioning.

[0052] Even further, by constructing a behavioral similarity graph and performing spectral clustering division, the model can perform adaptive optimization at the granularity of a more data-characteristic structure unit (sub-region) during the training phase, significantly improving the sensitivity and learning efficiency for high-risk regions. It lays a foundation for improving the overall anomaly detection accuracy, especially having stronger robustness and interpretability under complex working conditions.

[0053] S2: Statistically calculate the anomaly density of each region within a preset historical period, and dynamically set the regional weights of the training samples of the autoencoder according to the anomaly density.

[0054] In each round of detection cycle, the reconstruction error of each monitoring node is first calculated according to the reconstruction result of the autoencoder. Let the original feature vector collected by the i-th monitoring node in this round be X i , the original features include the time series indicators such as vibration, acoustic emission stress, temperature, etc. recorded at multiple time points of the node. i Input into the autoencoder to obtain the corresponding reconstructed output The representation model reconstructs the current state of the node based on historical learning results.

[0055] Assume the reconstruction error of the i-th node is:

[0056]

[0057] Among them, ||·|| is the selected distance measurement method. i Represents the reconstruction error of each monitoring node.

[0058] Set the initial screening threshold T1 of the reconstruction error. The initial screening threshold T1 can be set to the 85% quantile of the reconstruction error according to the model training error distribution; if e i ≥T1, the node is considered an abnormal node. All monitoring nodes are classified according to the area they belong to, and the area number is R (k) , whose total number of nodes is |R (k) |. In a sliding time window of length T, count the number of abnormal nodes that appear in each round in the kth region And calculate the anomaly density of the area in the current time window:

[0059]

[0060] The above abnormal density D k It indicates the frequency of abnormality in the region in the past T cycles, reflecting its risk level. The average abnormal density of all regions is further calculated:

[0061]

[0062] Where K represents the total number of regions.

[0063] In order to achieve adaptive focusing during regional sample training, a nonlinear function is introduced to set the weighting coefficient α for the regional samples. k , which is expressed as follows:

[0064]

[0065] Among them, β is the adjustment factor, which is used to control the sensitivity of weight change. This function has the characteristic of center compression. When the abnormal density of a certain area is higher than the average level, α kThe value increases significantly, thereby enhancing the influence of the samples in this region during the training process. Conversely, when the anomaly density is low, the α k value decreases, reducing its interference and ensuring that the model focuses on learning the features of the key regions.

[0066] Furthermore, by introducing a region-weighting mechanism based on anomaly density, the model training process can dynamically perceive the risk distribution of the monitoring region, thereby achieving spatial adaptive adjustment of sample learning. In the blasting monitoring scenario, due to differences in geological structure, operation frequency, etc. in different regions, the probability of generating anomalies is significantly different. This design can guide the model to focus on modeling the features of high-risk regions and improve the overall model's characterization ability and response accuracy for abnormal behaviors in key regions.

[0067] Even further, the anomaly density at the regional level is statistically calculated based on data from multiple cycles to avoid misjudgment caused by single perturbations, and the training intensity of regional samples is flexibly adjusted through a non-linear function, making the model training distribution more conform to the actual risk distribution. This strategy not only improves the model's adaptability to diverse regional structures but also enhances the recognition accuracy of weak anomalies and latent anomalies, providing a more accurate and stable basis for subsequent anomaly type determination.

[0068] S3: After reconstructing the current round of data using the weighted-trained autoencoder and calculating the reconstruction error of each node, extract the activation path information of the nodes whose reconstruction error is higher than the initial screening threshold of the reconstruction error.

[0069] For all nodes with high reconstruction errors, extract their activation path information from the structure of the trained autoencoder. Specifically, assume that the encoding part of the autoencoder consists of L layers, and the activation value of the l-th layer is denoted as Then the activation path of the i-th node is:

[0070]

[0071] This activation path information H i represents the set of feature neurons activated during the transfer of this node from the input layer to the hidden layer, and can reflect the response mode of this node during the encoding stage. The activation path information will be compared with the anomaly propagation path in subsequent steps to determine the consistency matching degree of structural anomalies.

[0072] Activation value is output using the ReLU activation function.

[0073] Furthermore, by extracting the activation path information from the nodes with high reconstruction error, an accurate characterization of the response mechanism of the input samples can be formed inside the neural network structure, so that instead of relying on a single numerical judgment, the hierarchical response process of the model is introduced as the basis for anomaly recognition. This path reflects the trajectory of the propagation and activation of node features in each hidden layer of the autoencoder, which helps to trace the transmission link of anomaly information in the model and provides an interpretable basis for subsequent path-level similarity comparison.

[0074] Furthermore, through the matching analysis of the activation path and the propagation path, not only the accuracy of anomaly recognition is enhanced, but also the ability of the model to distinguish complex anomaly structures is improved. At the same time, the activation value is output by the ReLU function, which further ensures the sparsity and stability of the path expression and avoids the influence of interference features on the final judgment result.

[0075] S4: Construct a graph structure model between the monitoring nodes, and calculate the anomaly propagation path according to the correlation between the nodes.

[0076] Calculate the behavioral similarity between any two nodes i and j according to the historical behavior data of each monitoring node (such as time series features such as vibration, acoustic emission, stress, temperature, etc.). Let the data sequences of nodes i and j be x i 、x j , then their similarity can be expressed by the Pearson correlation coefficient as:

[0077]

[0078] Construct a graph structure G=(V, E), where: V represents all monitoring nodes. If the correlation |r ij | between nodes i and j is greater than the similarity threshold θ, then an undirected edge e ij is established between them, and the edge represents the feature similarity value |r ij |.

[0079] This graph is used to depict the potential anomaly propagation relationship at the data level.

[0080] For the node i determined to be a node with high reconstruction error in the current round, extract the set of adjacent nodes in its graph structure:

[0081] N i ={j∈V|e ij ∈E}

[0082] Then, select the nodes in N i whose reconstruction error values are also greater than or equal to the initial reconstruction error threshold T1 as the set of anomaly propagation path nodes of node i:

[0083] S i ={j∈N i |ej ≥ T1}

[0084] Among them, S i represents the set of nodes that are strongly correlated with the high reconstruction error node i and whose reconstruction errors also exceed T1 around the high reconstruction error node i.

[0085] To improve the structural rationality of the abnormal propagation path recognition, after constructing S i when the high reconstruction error node i, all adjacent nodes N of the high reconstruction error node i i , and the node set together form a closed connected subgraph, it is determined that the propagation path has stable consistency in the graph structure, and the said propagation path is marked as a high-confidence propagation path;

[0086] The said node set is: the set composed of nodes that have a direct connection relationship with any node in N i but do not have a direct connection with the high reconstruction error node i itself.

[0087] Among the nodes in the initial abnormal propagation path node set S i but not included in the final high-confidence propagation path node set, the nodes are called the low-confidence propagation path candidate set L i .

[0088] Introduce the composite induction factor G m to characterize the potential risk of abnormal diffusion caused by the node m ∈ L i . The induction factor is constructed based on the following three dimensions: Neighborhood anomaly density: referring to the proportion of nodes in the adjacent nodes of the node m that are in the high reconstruction error state; Reconstruction error volatility: measuring the change range of the reconstruction error of the node m in different rounds; Temporal drift degree of the graph embedding feature: evaluating the dynamic change degree of the embedding vector of the node m between different time periods.

[0089] In the actual blasting monitoring scenario, although some nodes do not form a significant collaborative propagation path, they may still cause latent diffusion behavior. Relying solely on edge weights or reconstruction errors for judgment is likely to result in missing the identification of high-sensitivity areas. Therefore, in this embodiment, the construction of the composite induction factor is used to achieve earlier discovery of the abnormal source.

[0090] The construction of the composite induction factor G m adopts a hierarchical structure constraint method and is defined as follows:

[0091]

[0092] Among them, δ m represents the neighborhood anomaly density of the node m, that is, the proportion of the adjacent nodes whose reconstruction errors exceed the threshold T1; Represents the reconstruction error volatility of node m, ξ m Represents the degree of temporal drift of the graph embedding features of node m.

[0093] For all nodes in the candidate set L of low-confidence propagation paths in the current round i Calculate their induced factor values G m , and then based on the overall distribution characteristics of the induced factors within this set, obtain its mean value and standard deviation

[0094] If a certain node meets the following conditions:

[0095]

[0096] It is determined as a latent induced abnormal node, with local abnormal driving characteristics,

[0097] In the process of selecting nodes in the high-confidence propagation path, based on the feature similarity value |r ij | between each node and the central node i, perform a descending order sorting, and retain the adjacent nodes with the top n feature similarity values;

[0098] The final abnormal propagation path node set S' i Includes, from the high-confidence propagation path node set, based on the feature similarity value |r ij | between each propagation node j and the high-reconstruction error node i, perform a descending order sorting, and select the nodes with the top n feature similarity values; and from the candidate set L of low-confidence propagation paths i Identify the latent induced abnormal nodes that meet the induced factor .

[0099] Compared with the strategy of using a fixed threshold, the above determination conditions have stronger scenario adaptability and statistical interpretability, can dynamically reflect the relative outlier degree of induced factors in different rounds and different regions, so as to more effectively identify latent abnormal sources and avoid the occurrence of missed detections and misjudgments.

[0100] In the process of constructing the monitoring node graph structure and identifying the abnormal propagation path, further introduce the structural consistency determination to improve the stability and recognition accuracy of the propagation path. The graph structure effectively expresses the potential abnormal associations of each monitoring node in the blasting site at the time series data level, laying a foundation for the construction of the propagation path.

[0101] For each node i determined as a high-reconstruction error node in the current round, extract its adjacent node set N in the graph structure i ; in N iAmong them, nodes with reconstruction errors that also exceed the initial screening threshold T1 of the reconstruction error are selected to form the initial abnormal propagation path node set S of this node. i This step can quickly identify the neighborhood propagation relationships that are strongly correlated with high-reconstruction-error nodes and have the same abnormal trend, and achieve a preliminary lock on the potential risk diffusion area.

[0102] Furthermore, to improve the structural rationality and credibility of the propagation path, this embodiment introduces a structural consistency determination mechanism: If node i, its adjacent nodes, and some second-order adjacent nodes form a closed subgraph structure in the graph structure (such as a triangular closed loop or a four-ring structure), it is determined that this propagation path has topological consistency and is marked as a high-confidence propagation path. This mechanism helps to eliminate misconnections caused by accidental noise or isolated perturbations and improve the path stability. After being marked as a high-confidence path, for the propagation path set S i The nodes in are sorted in descending order according to their feature similarity value |r ij | with the central node i, and only the first n adjacent nodes with the largest propagation intensity are retained as the final propagation path set S' i . This mechanism improves the recognition accuracy while reducing the computational complexity, and provides high-quality data support for subsequent activation path comparison and abnormal type determination.

[0103] In the process of constructing the abnormal propagation path, the path relationship is not only constructed based on the correlation between high-reconstruction-error nodes and their adjacent nodes, but also a structural consistency judgment mechanism and an edge weight optimization screening mechanism for propagation path nodes are further introduced. On the basis of the initially constructed propagation path, if the target node, its adjacent nodes, and the second-order adjacent nodes together form a closed subgraph structure, it is determined that this path has topological consistency and is marked as a high-confidence propagation path accordingly. By introducing the above dual mechanisms, the propagation path is superior to the existing method of constructing paths only based on adjacent reconstruction errors in terms of structural rationality, node representativeness, and propagation trend expression, significantly improving the accuracy and stability of abnormal propagation recognition and providing a more reliable data basis for subsequent abnormal state determination.

[0104] S5: Match the activation path with the propagation path to determine the abnormal state of the target node.

[0105] The matching judgment between the activation path and the propagation path includes: extracting the activation path set of the target node i and the abnormal propagation path set formed by it in the graph structure, and calculating the similarity Sim i indicating the matching degree between the internal response process and the external structure diffusion process of this node in the model.

[0106] The similarity Sim i The formula is expressed as: The activation path of the i-th node is H i, the final set of abnormal propagation paths is S' i , calculate the coincidence degree between the two through Jaccard similarity, which is defined as follows:

[0107]

[0108] where, |·| represents the cardinality of the set, ∩ represents the intersection, ∪ represents the union, and Sim i ranges from [0, 1], and the closer it is to 1, the more consistent the two paths are.

[0109] The matching judgment of the activation path and the propagation path also includes: setting a similarity judgment threshold T2, and combining the reconstruction error initial screening threshold T1 to classify and determine the abnormal type of node i.

[0110] When Sim i ≥ T2 and e i ≥ T1, it is determined as a structural abnormal node, indicating that the node shows a consistent abnormal trend in both the internal feature response and the external structure propagation of the model.

[0111] When Sim i < T2 and e i ≥ T1, it is determined as a sudden abnormal node, indicating that although the node shows significant abnormalities in the model, it does not form a coordinated diffusion in the structure diagram.

[0112] When e i < T1 and multiple adjacent nodes of this node are abnormal propagation path nodes (i.e., |S i | ≥ τ),

[0113] it is determined as a latent abnormal node, indicating that the node is in an environment with a high incidence of abnormalities and needs to be focused on. The number threshold τ of adjacent abnormal nodes represents the minimum number of abnormal neighbors required to determine latent abnormalities, which is set to 2. The similarity judgment threshold T2 is set to the upper limit of empirical similarity.

[0114] Further, by introducing a similarity matching mechanism for the activation path and the propagation path, when the model determines the abnormal state of a node, it not only depends on its reconstruction error, but also comprehensively judges based on its abnormal diffusion relationship in the structure association graph. This design realizes the collaborative verification of the internal response of the model (i.e., the activation behavior of the encoder) and the external structure propagation trend, thereby distinguishing different types of abnormal manifestation forms, and improving the classification accuracy of abnormal recognition and the logical interpretability of model judgment.

[0115] Furthermore, through the matching judgment of the activation path and the propagation path, three typical abnormal types can be effectively identified: structural abnormality, sudden abnormality, and latent abnormality, enabling the system to implement differential processing for different risk forms during the abnormal recognition process.

[0116] In practical applications, this method not only improves the accuracy of anomaly detection, but also enhances the model's adaptive perception ability for misjudgments of high reconstruction errors, anomaly diffusion trends, and risk marginal areas in complex scenarios, thus providing more forward-looking and robust decision-making support for the monitoring system.

[0117] Example 2 is an embodiment of the present invention, which provides an anomaly analysis system based on blasting data, including

[0118] A data module that collects blasting data from each monitoring node and divides the monitoring nodes into multiple regions.

[0119] A training module that statistically calculates the anomaly density in each region within a preset historical period and dynamically sets the regional weights of the training samples of the autoencoder according to the anomaly density.

[0120] A reconstruction module that reconstructs the current round of data using the weighted-trained autoencoder, calculates the reconstruction error of each node, and extracts the activation path information of the nodes whose reconstruction error is higher than the initial screening threshold of the reconstruction error.

[0121] A propagation module that constructs a graph structure model between the monitoring nodes and calculates the anomaly propagation path according to the correlation between the nodes.

[0122] A confirmation module that matches the activation path and the propagation path to calculate the similarity and determines the anomaly status of the target node.

[0123] Example 3 is an embodiment of the present invention, which is different from the previous two embodiments in that

[0124] If the above functions are implemented in the form of software function units 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 invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes various media such as USB flash drives, mobile hard disks, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0125] Logic and / or steps represented in a flowchart or otherwise described herein, for example, can be considered as a definite ordered list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in conjunction with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0126] More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, a computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then stored in a computer memory.

[0127] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0128] Example 4, an embodiment of the present invention, provides an abnormal analysis method and system based on blasting data. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0129] Based on the blasting operation scenario in an open-pit mine, a distributed blasting data acquisition system consisting of 36 sensor nodes was constructed, covering an area of ​​about 900 square meters. The data collected by each node include vibration signals (unit: mm / s), acoustic emission signals (unit: dB), stress data (unit: MPa) and temperature information (unit: °C), which are recorded at a frequency of 10 Hz per second to form a multi-dimensional time series data sequence. In order to ensure the breadth and objectivity of the experiment, 30 consecutive days of operation data were selected as the historical reference window, and a total of about 259,200 valid data records were monitored during the test period.

[0130] First, the Pearson correlation coefficient is used to calculate the data feature similarity between all node pairs, and a symmetric similarity matrix is ​​constructed based on this. Then, the 36 nodes are divided into 4 regions through the spectral clustering method. Subsequently, the number of abnormal nodes in each region in the historical period is counted, and the abnormal judgment standard is that the reconstruction error is higher than the historical average reconstruction error mean + 1.5 times the standard deviation. According to the abnormal density of each region, the sample weight coefficient is designed using a nonlinear function to adjust the contribution of samples in each region during the autoencoder training process.

[0131] After completing the weighted training, the autoencoder is used to reconstruct the current round of data. For each node, the reconstruction error between its original feature and the model output is calculated, and the high reconstruction error nodes with reconstruction errors higher than the initial screening threshold T1=0.85T_1=0.85T1=0.85 are extracted, and their activation paths are further analyzed. The adjacent nodes of each high reconstruction error node are extracted through the graph structure model, and the initial propagation path set is constructed. If the propagation path forms a closed subgraph structure with its adjacent nodes and second-order adjacent nodes, it is identified as a high-confidence propagation path.

[0132] In the process of propagation path screening, the inducing factor GmG_mGm is constructed at the same time, including the neighborhood anomaly density, error volatility and embedding drift degree. The standard deviation outlier detection is performed on the inducing factor to identify the latent induced abnormal nodes and merge them with the high-confidence path nodes to form the final propagation path set. Finally, the matching degree between the activation path and the propagation path is calculated based on the Jaccard similarity, and the abnormal state of the target node is classified and determined.

[0133] In the continuous sampling period, the simulation generated 36 nodes with a total of 3,600 sets of blasting time series data. In the current detection round, the system identified 9 nodes with high reconstruction errors and screened out 21 candidate nodes when constructing the initial propagation path. After structural connectivity analysis, a high-confidence propagation path set of 12 nodes was formed, of which nodes with similarity values ​​greater than 0.7 accounted for about 66.7%. In the low-confidence path, based on the distribution of inducing factors, 4 potential induced abnormal nodes were identified, bringing the number of final propagation path nodes to 16.

[0134] By comparing with the average propagation pattern in the previous 30 days, it is found that the final set of propagation paths is more concentrated in high-density areas, and shows stronger connectivity and activation response overlap on local subgraphs. At the same time, the recognition results of three types of abnormal nodes are: 4 structural abnormalities, 2 sudden abnormalities, and 3 latent abnormalities, and the coverage rate is improved by about 18% compared with traditional reconstruction error detection.

[0135] Judging from the experimental data, the proposed autoencoder structure based on dynamically weighted training of abnormal density significantly enhances the model's learning ability for high-frequency abnormal areas. Through the regional weight mechanism, the nodes in areas with higher abnormal density receive more attention during the training process, significantly improving the recognition sensitivity of the model in the current round of detection. Compared with the traditional autoencoder with fixed sample weights, this method improves the average reconstruction error by about 12%, but reduces the false alarm rate by about 15%, demonstrating good convergence and discriminative ability.

[0136] At the same time, the determination of closed subgraph structure is introduced in graph structure construction and propagation path evaluation, which not only ensures the structural consistency of propagation paths, but also avoids the contamination of propagation paths by "single-point isolation error" in the node screening stage. By further introducing the construction mechanism of composite inducing factors and using dynamic statistical determination criteria, the limitation of traditional methods relying on hard thresholds to identify abnormalities is broken through, enabling the system to identify edge nodes with latent risks, especially showing good foresight in areas with multi-point high-density distribution.

[0137] Jaccard matching is used for collaborative analysis between activation paths and propagation paths, which not only improves the interpretability of abnormal classification, but also provides a semantic basis for subsequent evolutionary prediction based on distribution trends. Generally speaking, this method takes into account accuracy, stability and interpretability, the technical chain is compact, and the innovation points cover multiple levels such as regional dynamic modeling, graph structure propagation construction and intelligent determination driven by inducing factors, showing obvious technical advantages and application potential in real-time safety assessment of blasting operation scenarios.

[0138] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. The abnormal analysis method based on blasting data is characterized by: include, Collecting blasting data of each monitoring node, and dividing the monitoring node into multiple areas according to the blasting data; the blasting data includes vibration signals, acoustic emission signals, stress data and temperature; According to the blasting data, the abnormal density of each region in the preset historical period is counted, and the regional weight of the autoencoder training sample is dynamically set according to the abnormal density; The current round of data is reconstructed using the weighted trained autoencoder, the reconstruction error of each monitoring node is calculated, and the activation path information of the nodes whose reconstruction error is higher than the reconstruction error initial screening threshold is extracted; Build a graph structure model between monitoring nodes, calculate the correlation between nodes based on the burst data of each monitoring node, and calculate the abnormal propagation path based on the correlation between nodes; Match the activation path with the propagation path to determine the abnormal state of the target node, including using Jaccard similarity to calculate the activation path information H i and the final abnormal propagation path node set S' i SimilaritySim i ; When Sim i ≥T2 and reconstruction error e i When ≥T1, it is determined to be a structural abnormal node, indicating that the structural abnormal node presents a consistent abnormal trend in both the internal characteristic response of the model and the external structural propagation; When Sim i <T2 and the reconstruction error e i ≥ T1, it is determined as a burst-type abnormal node, indicating that although the burst-type abnormal node shows significant abnormalities in the model, it does not form a collaborative diffusion in the structure diagram; When the reconstruction error e i <T1 and multiple adjacent nodes of the node are initial abnormal propagation path nodes, it is determined as a latent abnormal node, indicating that the latent abnormal node is in an environment with a high incidence of abnormalities and needs to be focused on; where T1 represents the initial screening threshold of the reconstruction error, and T2 represents the similarity judgment threshold.

2. The abnormality analysis method based on blasting data according to claim 1, characterized in that: Dividing the monitoring nodes into multiple areas includes extracting time series features from the blasting data collected by each monitoring node to form a feature sequence corresponding to each node; using the Pearson correlation coefficient to calculate the similarity between any two monitoring nodes and constructing a similarity matrix between nodes; The similarity matrix is ​​used as an undirected graph, in which each monitoring node is a node in the graph, and the edges between nodes represent the similarity degree of data features; the undirected graph is divided into several sub-regions by using a spectral clustering algorithm, so that the monitoring nodes in each sub-region have correlation in data behavior.

3. The abnormality analysis method based on blasting data according to claim 2, characterized in that: The setting of the regional weights of the autoencoder training samples includes counting the number of abnormal nodes in each sub-region in the historical period and calculating the abnormal density of each sub-region; determining the sample weight coefficient corresponding to each sub-region based on the difference between the abnormal density of each sub-region and the average abnormal density of all sub-regions; During the training process of the autoencoder, the sample weight coefficients are applied to the training samples in different sub-regions respectively.

4. The abnormality analysis method based on blasting data according to claim 3, characterized in that: The reconstructing of the current round of data includes reconstructing the blasting data collected in the current round by using the trained autoencoder, and calculating the reconstruction error e of each monitoring node. i , the reconstruction error is the difference between the original data and the reconstructed data of the monitoring node.

5. The abnormality analysis method based on blasting data according to claim 4, characterized in that: Extracting activation path information includes determining whether the reconstruction error of each node exceeds the reconstruction error initial screening threshold T1, and determining it as a high reconstruction error node when it exceeds the reconstruction error initial screening threshold T1; extracting activation path information H of the high reconstruction error node from the input layer to the hidden layer in the autoencoder structure; i , the activation path information is the activation value sequence of each hidden layer of the high reconstruction error node in the encoding stage.

6. The abnormality analysis method based on blasting data according to claim 5, characterized in that: Constructing a graph structure model between monitoring nodes includes: the edges between nodes represent the feature similarity values ​​between any two monitoring nodes; The calculation of the abnormal propagation path according to the correlation between nodes includes, for each node i determined to have a high reconstruction error, extracting its adjacent node set N in the graph structure: i ; In N i In the process, the nodes whose reconstruction errors also exceed the reconstruction error initial screening threshold T1 are screened out to form the initial abnormal propagation path node set S of the high reconstruction error node i. i ; When the high reconstruction error node i and all adjacent nodes N of the high reconstruction error node i i , node collection When a closed connected subgraph is formed together, it is determined that the propagation path has stable consistency in the graph structure, and the propagation path is marked as a high-confidence propagation path; The node set For: with N i The set of nodes that have a direct connection relationship with any node in but are not directly connected to the high reconstruction error node i itself; In the initial abnormal propagation path node set S i The nodes in the node set that are not included in the final high-confidence propagation path are set as the low-confidence propagation path candidate set L i ; The final abnormal propagation path node set S' i Including, from the high confidence propagation path node set, based on the feature similarity value between each propagation node j and the high reconstruction error node i | r ij | Sort in descending order and select the top n nodes with the feature similarity value; and select the low-confidence propagation path candidate set L i Satisfaction-inducing factors identified in The latent inducing abnormal node; wherein the inducing factor G m It is composed of the neighborhood anomaly density, reconstruction error volatility, and the temporal drift of graph embedding features. Represents G m The mean of Represents G m The standard deviation of .

7. An abnormality analysis system based on blasting data using the method according to any one of claims 1 to 6, characterized in that: The data module collects the blasting data of each monitoring node and divides the monitoring nodes into multiple areas; The training module counts the anomaly density of each region within a preset historical period and dynamically sets the regional weights of the autoencoder training samples based on the anomaly density; The reconstruction module reconstructs the current round of data using the weighted trained autoencoder, calculates the reconstruction error of each monitoring node, and extracts the activation path information of the nodes whose reconstruction error is higher than the reconstruction error initial screening threshold; The propagation module builds a graph structure model between monitoring nodes and calculates the abnormal propagation path based on the correlation between nodes; The confirmation module matches the activation path with the propagation path and calculates the similarity to determine the abnormal state of the target node.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the abnormality analysis method based on burst data according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the abnormality analysis method based on blasting data according to any one of claims 1 to 6 are implemented.

Images