A method and system for intelligent hazard detection based on image recognition

Abstract

This invention relates to the field of intelligent inspection and image analysis technology, and discloses an intelligent hazard investigation method and system based on image recognition. The method includes establishing a hazard knowledge network representing the correlation between hazard types, features, and causes. Hazard images are collected during inspections and input into the network. Hazard entities are extracted, and their features are matched with hazard features in the network. Network reasoning is used to obtain the associated hazard types and causes. Hazard probabilities are calculated based on the correlation strength of causes. Intelligent investigation tasks bound to locations are generated and output based on a comparison of the probability and a threshold to guide targeted investigations. Finally, the correlation relationships in the hazard knowledge network are optimized based on feedback from on-site verification of the tasks. This invention achieves intelligent hazard identification and cause reasoning through a knowledge network, and optimizes the network using a closed-loop business feedback mechanism, improving the intelligence depth of hazard investigation and the system's adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and image analysis technology, specifically to an intelligent hazard investigation method and system based on image recognition. Background Technology

[0002] In routine inspections across industries such as power, chemical, and construction, using cameras or drones to capture images and then automatically identifying potential hazards through computer vision models has become a common technique for improving inspection efficiency. This conventional approach typically treats hazard identification as a typical object detection or image classification task. The technical path involves collecting labeled data, training a recognition model, and deploying the model for prediction. The model's output usually includes the specific hazard objects present in the image and their location information.

[0003] These conventional technical solutions have limitations. Their function is limited to the level of "discovery" and "what is" of potential hazards, merely informing the site of a phenomenon or entity that does not comply with safety regulations. They lack the ability to analyze the underlying causes of hazards, failing to answer crucial questions for prevention and eradication, such as "why did this hazard occur?" or "what deeper faults, management loopholes, or environmental factors might have caused this phenomenon?" This means that the investigation work still heavily relies on secondary analysis based on expert experience, resulting in insufficient depth of intelligent technology.

[0004] The existing solution is an open-loop system. The optimization and iteration of the identification model mainly relies on labeling more new samples to improve its identification accuracy. However, the model's updates are disconnected from the actual results of on-site verification within the business closed loop. The system cannot utilize feedback from frontline personnel after on-site verification to optimize and correct deeper knowledge such as the "logic between different hazard characteristics and specific causes." This results in the system's analytical capabilities being static, unable to continuously learn from actual business feedback, and its reasoning accuracy and adaptability failing to autonomously improve with use. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent hazard investigation method and system based on image recognition, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent hazard detection method based on image recognition, the method comprising:

[0007] Establish a hazard knowledge network to represent the relationship between hazard types, hazard characteristics, and hazard causes;

[0008] Images of potential hazards in the target area are collected along the designated inspection route, and these images are then input into the hazard knowledge network.

[0009] From the initial analysis results output by the hazard knowledge network, extract the hazard entities present in the hazard image;

[0010] The entity features of the hidden danger entity are matched with the hidden danger features in the hidden danger knowledge network, and reasoning is performed through the hidden danger knowledge network to obtain the hidden danger type and hidden danger cause associated with the hidden danger entity.

[0011] Based on the correlation strength of the causes of the hidden dangers, the probability of hidden dangers in the corresponding area of ​​the hidden danger image is calculated;

[0012] Based on the comparison between the probability of the potential hazard and a preset threshold, an intelligent investigation task is generated that is bound to the location of the potential hazard image.

[0013] The intelligent investigation task is output to guide the targeted investigation of potential hazards in the target area;

[0014] Based on the on-site verification feedback for the intelligent investigation task, the relationships in the hidden danger knowledge network are optimized.

[0015] Preferably, the step of establishing a hazard knowledge network to represent the correlation between hazard types, hazard characteristics, and hazard causes includes:

[0016] Obtain historical hazard cases, and analyze the types of historical hazards, historical image features, and historical causes from these cases;

[0017] The historical hazard type is defined as a node type in the hazard knowledge network, the historical image features are abstracted as attribute features in the hazard knowledge network, and the historical causes are established as associated nodes in the hazard knowledge network.

[0018] A directed connection is established between the node type, the attribute feature, and the associated node. The weight of the directed connection is initialized based on the co-occurrence frequency of the corresponding item in the historical hidden danger cases.

[0019] Preferably, extracting the hazard entities present in the hazard image from the initial analysis results output from the hazard knowledge network includes:

[0020] Receive the initial analysis results of the hidden danger knowledge network on the hidden danger image, the initial analysis results including multiple image regions in the hidden danger image that have a potential matching degree with the attribute features;

[0021] Entity boundaries are defined for each of the multiple image regions, and local feature enhancement is performed within the regions defined by the entity boundaries.

[0022] Identify whether there is an entity feature that matches any of the attribute features within the region after the local feature enhancement. If the match is successful, define the region bounded by the entity boundary as a potential entity.

[0023] Preferably, the step of matching the entity features of the potential hazard entity with the potential hazard features in the potential hazard knowledge network, and then using the potential hazard knowledge network to infer the potential hazard type and cause associated with the potential hazard entity, includes:

[0024] The similarity between the entity features and all attribute features in the hidden danger knowledge network is calculated, and the attribute feature with the highest similarity is selected as the matching result.

[0025] Locate the attribute feature node corresponding to the matching result in the hidden danger knowledge network, and activate all directed connections originating from the attribute feature node;

[0026] Traverse the path along the activated directed connection, gather all reachable node types and associated nodes, take the gathered node types as the hazard types associated with the hazard entity, and take the gathered associated nodes as the hazard causes associated with the hazard entity.

[0027] Preferably, calculating the probability of a hazard in the region corresponding to the hazard image based on the correlation strength of the hazard causes includes:

[0028] For each hazard cause associated with the hazard entity, the path weight from the attribute feature node corresponding to the matching result to the associated node corresponding to the hazard cause is obtained in the hazard knowledge network, and the path weight is defined as the association strength of the hazard cause.

[0029] The entity risk value of the entity is obtained by weighted fusion of the correlation strength of all the causes of the hidden danger associated with the entity.

[0030] The sum of the entity risk values ​​of all the hazard entities in the hazard image is calculated, and the sum is normalized and mapped to a value between zero and one hundred. This value is then used as the hazard probability of the corresponding area in the hazard image.

[0031] Preferably, after acquiring the hazard image of the target area, the method further includes:

[0032] Collect non-image dimension information at the same spatiotemporal point as the hazard image, wherein the non-image dimension information includes at least one of ambient temperature information, equipment vibration information, and gas concentration information;

[0033] In the aforementioned hidden danger knowledge network, auxiliary decision-making nodes are established for each type of non-image dimension information;

[0034] During network inference, the quantized value of the non-image dimension information is used as the input of the auxiliary decision node to adjust the effective weight of the directed connection from the attribute feature node to the associated node.

[0035] Preferably, the step of generating an intelligent investigation task bound to the location of the hazard image based on the comparison result between the numerical value of the hazard probability and a preset threshold includes:

[0036] When the probability of the potential hazard is greater than a preset threshold, a task generation instruction is triggered;

[0037] In response to the task generation instruction, the hazard type, the cause of the hazard, and the non-image dimension information associated with the hazard entity in the region corresponding to the hazard image are integrated to form a structured description;

[0038] Based on the location information of the hazard image and the structured description, an intelligent investigation task is generated that includes the target location, target object, and key points of concern.

[0039] Preferably, after outputting the intelligent screening task, the method further includes:

[0040] During the on-site investigation of the aforementioned intelligent investigation task, the on-site handling measures and post-handling status for the potential hazards are recorded.

[0041] The on-site handling measures are compared with the hazard causes associated in the hazard knowledge network, and the root cause nodes targeted by the on-site handling measures are marked.

[0042] The root cause node is associated with the corresponding hidden danger entity to form a verified knowledge association pair.

[0043] Preferably, the step of optimizing the relationships in the hazard knowledge network based on on-site verification feedback for the intelligent investigation task includes:

[0044] Receive the verified knowledge association pair and identify the original directed connection between the root cause node and the attribute feature node;

[0045] Based on the effectiveness verification results of the aforementioned on-site handling measures, the weights of the original directed connections may be increased or decreased.

[0046] For redundant connections that exist in the hidden danger knowledge network but have never been associated with the verified knowledge association pairs, their weights are gradually reduced until they are removed.

[0047] Preferably, the present invention also includes an intelligent hazard investigation system based on image recognition, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent hazard investigation method based on image recognition as described above.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] By constructing a reasonable knowledge network that deeply correlates hazard types, characteristics, and causes, the entity features identified from images are matched with network nodes, and logical reasoning is performed along a pre-defined association path. This achieves a leap from identifying surface phenomena to inferring deep causes. The method outputs no longer isolated hazard labels, but rather analytical results containing potential causal chains, enabling hazard investigation to directly focus on the root causes of problems and improving the depth and relevance of analytical decisions.

[0050] By using on-site verification feedback from intelligent screening tasks as key data, the system continuously optimizes the relationships and weights between nodes in the knowledge network, constructing a complete closed loop from perception and decision-making to verification and learning. This mechanism enables the system's reasoning logic to no longer be static and unchanging, but to continuously self-correct and enhance itself based on feedback from real-world scenarios, thereby gradually improving its accuracy and adaptability in inferring causes in different complex environments.

[0051] The reasoning mechanism of the hazard knowledge network transforms discrete identification results into interpretable analytical conclusions, providing on-site personnel with clear directions for investigation and reducing their absolute reliance on professional experience. Meanwhile, the closed-loop optimization mechanism ensures that the system's analytical capabilities continuously evolve with the accumulation of business data, making it a dynamic intelligent tool that ultimately achieves a dual improvement in the efficiency and accuracy of hazard investigation. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent hazard detection method based on image recognition described in this invention.

[0053] Figure 2 A flowchart for establishing a hazard knowledge network;

[0054] Figure 3 A flowchart illustrating the process of matching entity features and performing network inference.

[0055] Figure 4 A comparison chart of effective weights in a knowledge network;

[0056] Figure 5 This is a diagram showing the initial knowledge network connection weight distribution. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 This invention provides an intelligent hazard identification method based on image recognition. The method includes: first, constructing a hazard knowledge network, which formally represents the inherent relationship between hazard type, hazard characteristics, and hazard causes. In the application phase, hazard images of the target area are collected along a preset inspection path, and the collected images are input into the constructed hazard knowledge network. The system extracts the actual hazard entities in the images from the initial analysis results generated after preliminary analysis of the input images by the hazard knowledge network. Subsequently, the entity features exhibited by these hazard entities are matched with standardized hazard features stored in the hazard knowledge network, and reasoning is performed using the internal relationships within the knowledge network to obtain the specific hazard type associated with the hazard entity and its possible causes.

[0059] Based on the correlation strength between the inferred causes of various hazards and the current features, the system calculates the comprehensive hazard probability of the area corresponding to the hazard image. This hazard probability value is compared with a preset threshold, and based on the comparison result, an intelligent investigation task bound to the geographic location information of the hazard image is automatically generated. The generated intelligent investigation task is output to guide inspection personnel to conduct targeted on-site verification of the target area. Finally, based on feedback information obtained after on-site verification of the intelligent investigation task, the system iteratively optimizes the internal relationships of the hazard knowledge network to make it more accurate.

[0060] Example 1: See Figure 2In practical implementation, historical hazard cases are acquired, and from these cases, the types of historical hazards, historical image features, and historical causes are analyzed. Specifically, historical hazard types are defined as node types in the hazard knowledge network, historical image features are abstracted into attribute feature nodes in the hazard knowledge network, and historical causes are established as association nodes in the hazard knowledge network. After analyzing node types, attribute feature nodes, and association nodes based on historical hazard cases, the abstraction process of attribute features needs to combine feature engineering methods from the field of computer vision to perform grayscale conversion, edge detection, and feature point extraction on historical image features. The extracted low-dimensional visual features are combined into high-dimensional feature vectors, serving as the core data form of attribute feature nodes in the hazard knowledge network. In addition to being based on co-occurrence frequency, the initial weights of directed connections also need to be normalized to map to the [0,1] interval to ensure the comparability of connection weights between different types of nodes. For example, when "cable damage" and "insulation layer cracking image features" co-occur 80 times in historical hazard cases, and "cable damage" appears a total of 100 times, the initial weight of the directed connection between them is calculated to be 0.8, which is on the same quantification dimension as the weights of other related items. In some embodiments, directed connections are established between node types, attribute feature nodes, and related nodes based on the logical relationships reflected in historical hazard cases. The initial weight value of the directed connection is initialized based on the co-occurrence frequency of the corresponding items in historical hazard cases. It can be understood that the co-occurrence frequency refers to the number of times node types, attribute feature nodes, and related nodes appear simultaneously in historical hazard cases. Optionally, the initial weight of the directed connection is determined by the formula:

[0061] ;

[0062] in: This represents the initial weights of a directed connection. This indicates the number of times the node type, attribute feature node, or associated node of a directed connection co-occurs in historical potential hazard cases. This represents the total number of times the starting node of a directed connection appears in historical hazard cases. In specific implementations, the starting node of a directed connection is a node type or attribute feature node, and the ending node of a directed connection is an attribute feature node or an associated node. In some embodiments, the process of parsing historical hazard cases includes reading text records and image data from the case database, extracting historical hazard types, historical image features, and historical causes. It can be understood that the construction of the hazard knowledge network depends on the collection and parsing of historical hazard cases. The directed connections between node types, attribute feature nodes, and associated nodes form a network of relationships between hazard types, hazard features, and hazard causes.

[0063] Example 2: See Figure 3In specific implementation, the system receives the initial analysis results generated by the hidden danger knowledge network after processing the input hidden danger image. These initial analysis results include multiple image regions in the hidden danger image that have a potential matching degree with the attribute features stored in the hidden danger knowledge network. In practice, entity bounding operations are performed on each of the multiple potentially matching image regions, and local feature enhancement processing is applied to each region determined by the entity bounding operation. This local feature enhancement processing aims to highlight the detailed texture and structural information within the image region. When performing entity bounding operations on multiple image regions, the YOLO object detection algorithm is used to generate initial bounding boxes. Non-maximum suppression is used to remove redundant boxes with an overlap greater than 0.5, ensuring that the bounding box of each potential hidden danger region is unique. The local feature enhancement stage employs an adaptive histogram equalization algorithm to optimize for uneven brightness and blurred details within the bounding box, improving the matching accuracy between entity features and attribute features. Similarity calculation employs a weighted approach combining cosine similarity and Euclidean distance, with cosine similarity accounting for 0.6 and Euclidean distance for 0.4. A successful match is considered achieved when the overall similarity score exceeds 0.7. For example, the cosine similarity between the "metal corrosion feature vector" of a potential hazard entity and the "rust attribute feature vector" in the knowledge network is 0.75, and the corresponding Euclidean distance similarity is 0.68, resulting in an overall score of 0.722, which meets the matching threshold requirement. The weighting is based on statistical analysis of the effectiveness of feature matching in historical hazard cases. Cosine similarity is better suited for judging the directional consistency of high-dimensional features, while Euclidean distance is better suited for measuring the difference in feature values. The weight allocation is determined by combining the performance of both in past matching tasks. In some embodiments, it identifies whether there is an entity feature within the region after local feature enhancement that matches any attribute feature in the hazard knowledge network. If a match is successful, the region defined by the entity boundary delineation operation is defined as a hazard entity. In specific implementations, a successful match is determined when the similarity between the entity feature and the attribute feature exceeds a set threshold.

[0064] The process of matching the entity features of a potential hazard with the hazard features in a hazard knowledge network and then reasoning through the hazard knowledge network involves, in practice, calculating the similarity between the extracted entity features of the potential hazard and all attribute features stored in the hazard knowledge network. The attribute feature with the highest similarity score is selected as the final matching result. In some embodiments, the similarity calculation can use the cosine similarity method, expressed by the formula:

[0065] ;

[0066] in: This indicates the similarity calculation result. This represents the entity feature vector extracted from the entity with potential hazards. This represents a feature vector of a certain attribute stored in the hazard knowledge network. The process involves locating the attribute feature node corresponding to the matching result within the hazard knowledge network and activating all directed connections originating from that node. This activation operation implies considering the relationships represented by these directed connections during subsequent reasoning. Optionally, a path traversal is performed along the activated directed connections to aggregate all node types and associated nodes reachable through these connections. The aggregated node types are considered as the hazard types associated with the hazard entity, and the aggregated associated nodes are considered as the hazard causes associated with the hazard entity. In practice, a breadth-first search algorithm is used for node search during path traversal. In practice, the path traversal uses a breadth-first search algorithm to search for nodes. The breadth-first search algorithm starts with the attribute feature node corresponding to the matching result. First, the node is visited and its type is checked. If it is a node type or an associated node, it is added to the aggregation set. Then, all nodes directly connected to the node through directed connections are visited and marked as visited. Then, the next level of connected nodes of these directly connected nodes are visited in the same way to ensure that each node is processed only once, until all nodes reachable through directed connections have been traversed. Finally, the node types accumulated in the aggregation set are used as the hazard types associated with the hazard entity, and the associated nodes are used as the causes of the hazard.

[0067] Example 3: In specific implementation, for each hazard cause associated with a hazard entity, the weight of the path traversed from the attribute feature node corresponding to the matching result to the associated node corresponding to the hazard cause is obtained in the hazard knowledge network. This path weight is defined as the association strength of the hazard cause relative to the current hazard entity. In specific implementation, if multiple paths exist, the maximum value among the path weights or their weighted average is selected as the association strength. Optionally, the association strengths of all hazard causes associated with the hazard entity are weighted and fused to obtain the entity risk value of the hazard entity. It can be understood that the weights for weighted fusion can be set according to the type of hazard cause or historical statistical data. When calculating the entity risk value of the hazard entity, the weighted summation formula is used as follows:

[0068] ;

[0069] in: This represents the risk value of the i-th potential hazard entity. This represents the correlation strength weight of the j-th hazard cause. This indicates the correlation strength of the cause of the j-th hidden danger.

[0070] The normalization process uses the min-max standardization method, which is the maximum sum of historical entity risk values. Minimum value The total risk value of current image entities At that time, the probability of hidden dangers When fusing non-image dimensional information, the auxiliary decision node uses a linear regression model to process the quantized values. For example, the quantized value of ambient temperature information is... (Unit: °C) When The output adjustment factor is 1.2. The output coefficient is 1.0. The output coefficient is 0.8. The effective weight of the directed connection from the attribute feature node to the associated node is adjusted. For example, if the original weight is 0.7, the effective weight at a temperature of 65℃ is 0.7 × 1.2 = 0.84.

[0071] The sum of the entity risk values ​​of all identified hazard entities in the hazard image is calculated, and the sum is normalized and mapped to a value between zero and one hundred. This value is used as the hazard probability of the corresponding area in the hazard image. In some embodiments, the normalization process is based on a linear mapping between the maximum and minimum sum of entity risk values ​​in historical data, expressed by the following formula:

[0072] ;

[0073] in: This represents the calculated probability value of the potential hazard. This represents the sum of entity risk values ​​in the current hazard image. This represents the maximum sum of historical entity risk values. This represents the minimum sum of entity risk values ​​based on historical statistics.

[0074] After acquiring images of potential hazards in the target area, the process also includes acquiring non-image dimension information at the same spatiotemporal point as the hazard images. In specific implementations, non-image dimension information includes at least one of ambient temperature, equipment vibration, and gas concentration information. It can be understood that non-image dimension information is acquired synchronously through corresponding sensors deployed along the inspection path. In the hazard knowledge network, an auxiliary decision-making node is established for each type of non-image dimension information. The auxiliary decision-making node is connected to relevant attribute feature nodes and associated nodes. In some embodiments, the auxiliary decision-making node receives quantized non-image dimension information values ​​as input. When performing network inference to obtain the hazard type and hazard cause associated with the hazard entity, the quantized value of the non-image dimension information is used as the input to the corresponding auxiliary decision-making node. The auxiliary decision-making node outputs an adjustment coefficient based on the input value. This adjustment coefficient is used to dynamically adjust the effective weight of the directed connections from relevant attribute feature nodes to associated nodes in the current inference process. In specific implementations, the effective weight is the product of the original directed connection weight and the adjustment coefficient.

[0075] See Figure 4This is a comparison chart showing the adjustment of connection weights in a knowledge network within an intelligent hazard identification system. It illustrates the dynamic optimization of network connection weights by non-image-level information and is a core analytical chart in the hazard knowledge network inference stage. Gray represents the original weights, and blue represents the effective weights after adjustment by non-image information. Increased temperature sensor data strengthens the connection weight. Temperature sensor data is associated with the cause of "excessive temperature," increasing the weight. Stable gas sensor data decreases the connection weight. This chart demonstrates the dynamic optimization effect of non-image-level information on the hazard knowledge network: by adjusting the effective weights of network connections using real-time data from auxiliary decision nodes (such as sensors), the network inference results are made more closely aligned with the actual environment, improving the accuracy of hazard identification.

[0076] Example 4: In specific implementation, when the probability of a potential hazard exceeds a preset threshold, a task generation command is automatically triggered. In this implementation, the preset threshold is set based on historical investigation records and risk level classification. The triggering condition for the task generation command can be expressed by the formula:

[0077] ;

[0078] in: This indicates that the trigger status of the task generation command is 1 for triggered and 0 for not triggered. This represents the calculated probability value of the potential hazard. This represents a system-preset threshold. In response to a task generation command, the system integrates the hazard type, hazard cause, and related non-image dimension information associated with the hazard entities identified in the corresponding area of ​​the hazard image, forming a structured task description. This structured task description organizes information using predefined fields and formats. In some embodiments, based on the location information carried by the hazard image and the generated structured description, an intelligent investigation task is generated, containing the target location, target object, and key points of concern. The target location is obtained from the spatiotemporal metadata of the hazard image, the target object is the identified hazard entity type, and the key points of concern are the associated hazard causes and non-image dimension anomalies.

[0079] After outputting the intelligent investigation task, during the on-site investigation of the reported hazard entity, the on-site handling measures and the status after handling are recorded. In specific implementation, on-site handling measures include repair, replacement, isolation, and cleaning, and the status after handling is recorded through review images or text descriptions. Optionally, the recorded on-site handling measures are compared with the hazard causes associated with the hazard entity in the hazard knowledge network to mark the root cause node directly targeted by the on-site handling measures. In some embodiments, the root cause node is the associated node that the on-site handling measures aim to eliminate or control. The root cause node is associated with the corresponding hazard entity to form a verified knowledge association pair. The verified knowledge association pair includes the attribute characteristics of the hazard entity, the root cause node, and the handling effectiveness marker. In specific implementation, referring to Table 1, the verified knowledge association pair can be recorded as a data record containing specific fields.

[0080] Table 1: Verified Knowledge Association Pair Record Table

[0081] ;

[0082] The process of forming verified knowledge association pairs between root cause nodes and corresponding hazard entities involves storing the structured information shown in the table above as a complete record in a knowledge verification database. This database is used for subsequent optimization of the hazard knowledge network. In some embodiments, when the same hazard entity is associated with multiple hazard causes, multiple verified knowledge association pairs are generated. Each knowledge association pair records the correspondence between a specific root cause node and a corresponding remedial measure.

[0083] Example 5: In specific implementation, the verified knowledge association pairs are received from the field verification stage. The verified knowledge association pairs include root cause nodes and corresponding hazard entity identifiers and matching attribute features. In specific implementation, the verified knowledge association pair records are read from the knowledge verification database. The existing directed connections between the root cause nodes and attribute feature nodes of the verified knowledge association pair records are identified in the hazard knowledge network. It can be understood that the identification process searches in the topology of the hazard knowledge network based on the node name or unique identifier.

[0084] Based on the effectiveness verification results of the on-site treatment measures, the weights of the original directed connections are strengthened or weakened. When adjusting the weights of the original directed connections, the learning rate coefficient is... The reward factor is set to 0.1 based on the validity verification results. When the treatment is effective When the treatment is ineffective Penalty coefficient The average weight of the network is obtained by calculating the arithmetic mean of the weights of all directed connections. The statistical period for redundant connections is set to 30 days, the weight decay factor is set to 0.8, and the removal threshold is set to 0.05. For example, if a directed connection initially has a weight of 0.3 and is not associated for three consecutive statistical periods, the weight decays to 0.24, 0.192, and 0.1536 respectively. If it is still higher than the threshold of 0.05, it is retained. If it decays to 0.0983 and 0.0786 for another two periods and is still not associated, it continues to decay until it falls below 0.05, at which point it is removed. After each weight adjustment, L2 regularization is used to constrain the weights of all connections in the entire hidden danger knowledge network to prevent excessively large weights of individual connections from causing bias in network inference. The parameter settings are determined based on the association patterns of historical hidden danger cases and field verification feedback data, combined with the stability requirements of network inference. Weight adjustments are based on the effectiveness verification results; effective weights are strengthened, and ineffective weights are weakened. Redundant connections decay gradually at fixed intervals. In some embodiments, the validity verification result is recorded in the validity tag field of the verified knowledge association pair. If the validity tag confirms that the action measure is effective, the weight of the corresponding original directed connection is increased; if the validity tag confirms that the action measure is ineffective, the weight of the corresponding original directed connection is decreased. Optionally, the weight adjustment is calculated according to an incremental formula, expressed as:

[0085] ;

[0086] in: This represents the adjusted directed connection weights. This represents the original directed connection weights before adjustment. This represents the learning rate coefficient. This represents the reward factor based on the validity verification results. Indicates the penalty coefficient. This represents the absolute value of the deviation between the current directed connection weights and the network average weights. In practice, each time a new validity verification result is received, the relevant original directed connection weights are updated and recalculated.

[0087] For redundant connections that exist in the hazard knowledge network but have never been verified, their weights are gradually reduced until they are removed. In some embodiments, the system periodically scans all directed connections in the hazard knowledge network for usage records. A usage record refers to the number of times a directed connection is activated and subsequently verified during network inference. Optionally, for directed connections with zero usage records within a statistical period, their weights are gradually reduced by multiplying them by a weight decay factor less than 1. When the weight of a directed connection falls below a set removal threshold after continuous decay, the directed connection is removed from the hazard knowledge network.

[0088] See Figure 5 This is a bar chart showing the connection weight distribution of the intelligent hazard investigation knowledge network. It displays the weight values ​​of each connection in the hazard knowledge network and is a core analytical chart in the knowledge network construction phase. The red dashed line represents the average weight of the network, used to compare the weight levels of each connection. The highest weight indicates the strongest correlation between the auxiliary node and the feature. The gas leak feature has a strong correlation with the cause of pressure anomalies. The lowest weight indicates a weak correlation between equipment cracks and improper installation. This chart visually presents the initial weight distribution of the hazard knowledge network: connections with higher weights represent a closer correlation between the corresponding feature and the cause / auxiliary information, which is the core basis for subsequent network reasoning (identifying hazard types and causes). At the same time, the average weight line can help identify "strong correlation" and "weak correlation" connections, providing a benchmark for subsequent weight optimization.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent hazard identification based on image recognition, characterized in that, Includes the following steps: Establish a hazard knowledge network to represent the relationship between hazard types, hazard characteristics, and hazard causes; Images of potential hazards in the target area are collected along the designated inspection route, and these images are then input into the hazard knowledge network. From the initial analysis results output by the hazard knowledge network, extract the hazard entities present in the hazard image; The entity features of the hidden danger entity are matched with the hidden danger features in the hidden danger knowledge network, and reasoning is performed through the hidden danger knowledge network to obtain the hidden danger type and hidden danger cause associated with the hidden danger entity. Based on the correlation strength of the causes of the hidden dangers, the probability of hidden dangers in the corresponding area of ​​the hidden danger image is calculated; Based on the comparison between the probability of the potential hazard and a preset threshold, an intelligent investigation task is generated that is bound to the location of the potential hazard image. The intelligent investigation task is output to guide the targeted investigation of potential hazards in the target area; Based on the on-site verification feedback for the intelligent investigation task, the relationships in the hidden danger knowledge network are optimized; The step of matching the entity features of the potential hazard with the potential hazard features in the potential hazard knowledge network, and then using the potential hazard knowledge network to infer the potential hazard type and cause associated with the potential hazard, includes: The similarity between the entity features and all attribute features in the hidden danger knowledge network is calculated, and the attribute feature with the highest similarity is selected as the matching result. Locate the attribute feature node corresponding to the matching result in the hidden danger knowledge network, and activate all directed connections originating from the attribute feature node; Traverse the path along the activated directed connection, gather all reachable node types and associated nodes, take the gathered node types as the hidden danger types associated with the hidden danger entity, and take the gathered associated nodes as the hidden danger causes associated with the hidden danger entity. The step of calculating the probability of a hazard in the region corresponding to the hazard image based on the correlation strength of the hazard causes includes: For each hazard cause associated with the hazard entity, the path weight from the attribute feature node corresponding to the matching result to the associated node corresponding to the hazard cause is obtained in the hazard knowledge network, and the path weight is defined as the association strength of the hazard cause. The entity risk value of the entity is obtained by weighted fusion of the correlation strength of all the causes of the hidden danger associated with the entity. The sum of the entity risk values ​​of all the hazard entities in the hazard image is calculated, and the sum is normalized and mapped to a value between zero and one hundred. This value is then used as the hazard probability of the corresponding area in the hazard image.

2. The intelligent hazard investigation method based on image recognition according to claim 1, characterized in that, The establishment of a hazard knowledge network to represent the relationship between hazard types, hazard characteristics, and hazard causes includes: Obtain historical hazard cases, and analyze the types of historical hazards, historical image features, and historical causes from these cases; The historical hazard type is defined as a node type in the hazard knowledge network, the historical image features are abstracted as attribute features in the hazard knowledge network, and the historical causes are established as associated nodes in the hazard knowledge network. A directed connection is established between the node type, the attribute feature, and the associated node. The weight of the directed connection is initialized based on the co-occurrence frequency of the corresponding item in the historical hidden danger cases.

3. The intelligent hazard investigation method based on image recognition according to claim 2, characterized in that, The step of extracting hazard entities from the hazard image from the initial analysis results output from the hazard knowledge network includes: Receive the initial analysis results of the hidden danger knowledge network on the hidden danger image, the initial analysis results including multiple image regions in the hidden danger image that have a potential matching degree with the attribute features; Entity boundaries are defined for each of the multiple image regions, and local feature enhancement is performed within the regions defined by the entity boundaries. Identify whether there is an entity feature that matches any of the attribute features within the region after the local feature enhancement. If the match is successful, define the region bounded by the entity boundary as a potential entity.

4. The intelligent hazard investigation method based on image recognition according to claim 1, characterized in that, After acquiring images of potential hazards in the target area, the method further includes: Collect non-image dimension information at the same spatiotemporal point as the hazard image, wherein the non-image dimension information includes at least one of ambient temperature information, equipment vibration information, and gas concentration information; In the aforementioned hidden danger knowledge network, auxiliary decision-making nodes are established for each type of non-image dimension information; During network inference, the quantized value of the non-image dimension information is used as the input of the auxiliary decision node to adjust the effective weight of the directed connection from the attribute feature node to the associated node.

5. The intelligent hazard investigation method based on image recognition according to claim 4, characterized in that, The step of generating an intelligent investigation task bound to the location of the hazard image based on the comparison result between the numerical value of the hazard probability and a preset threshold includes: When the probability of the potential hazard is greater than a preset threshold, a task generation instruction is triggered; In response to the task generation instruction, the hazard type associated with the hazard entity in the region corresponding to the hazard image, the hazard cause, and the non-image dimension information are integrated to form a structured description; Based on the location information of the hazard image and the structured description, an intelligent investigation task is generated that includes the target location, target object, and key points of concern.

6. The intelligent hazard investigation method based on image recognition according to claim 5, characterized in that, After outputting the intelligent screening task, the method further includes: During the on-site investigation of the aforementioned intelligent investigation task, the on-site handling measures and post-handling status for the potential hazards are recorded. The on-site handling measures are compared with the hazard causes associated with the hazard knowledge network, and the root cause nodes targeted by the on-site handling measures are marked. The root cause node is associated with the corresponding hidden danger entity to form a verified knowledge association pair.

7. The intelligent hazard investigation method based on image recognition according to claim 6, characterized in that, The optimization of the relationships in the hazard knowledge network based on on-site verification feedback for the intelligent investigation task includes: Receive the verified knowledge association pair and identify the original directed connection between the root cause node and the attribute feature node; Based on the effectiveness verification results of the aforementioned on-site handling measures, the weights of the original directed connections may be increased or decreased. For redundant connections that exist in the hidden danger knowledge network but have never been associated with the verified knowledge association pairs, their weights are gradually reduced until they are removed.

8. An intelligent hazard detection system based on image recognition, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the image recognition-based intelligent hazard investigation method as described in any one of claims 1 to 7.

Images