Industrial internet of things based park multi-source heterogeneous data analysis method and system

Abstract

The application discloses a park multi-source heterogeneous data analysis method and system based on an industrial Internet of Things, and relates to the technical field of data analysis. The park multi-source heterogeneous data analysis method based on the industrial Internet of Things comprises the following steps: determining a suspicious target according to image information of a target area; predicting a moving path of the suspicious target according to a moving trend of the suspicious target, and acquiring characteristic information of the suspicious target through a supervision device on the moving path; predicting an abnormal probability of the suspicious target in an abnormal situation according to the characteristic information and in combination with work data of the suspicious target in a past preset time period, and judging whether the abnormal probability exceeds a preset threshold value; and tracking an abnormal target if the abnormal probability exceeds the preset threshold value. The application aims to solve the technical problem that, due to data islands and a lagging response mechanism, it is difficult to realize accurate risk identification and dynamic management and control in a complex industrial scene, and false positives or false negatives are prone to occur.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial internet of things park data analysis, and particularly relates to a park multi-source heterogeneous data analysis method and system based on industrial internet of things. BACKGROUND

[0002] In an industrial internet of things (IIoT) driven smart park, the deep interconnection of devices, personnel, logistics and production processes generates massive multi-source heterogeneous data such as video monitoring, device sensors, personnel positioning, business system logs, etc. However, due to data silos and lagging response mechanisms, it is difficult to achieve accurate risk identification and dynamic control in complex industrial scenarios, and false positives or false negatives may occur. SUMMARY

[0003] The main purpose of the present application is to provide a park multi-source heterogeneous data analysis method and system based on industrial internet of things, which aims to solve the technical problems that the existing technology is difficult to achieve accurate risk identification and dynamic control in complex industrial scenarios due to data silos and lagging response mechanisms, and false positives or false negatives may occur.

[0004] To achieve the above-mentioned purpose, in a first aspect, the present application provides a park multi-source heterogeneous data analysis method based on industrial internet of things, comprising:

[0005] According to the image information of the target area, determining a suspicious target based on an anomaly detection algorithm;

[0006] According to the moving trend of the suspicious target, predicting the moving path of the suspicious target, and obtaining the feature information of the suspicious target through the monitoring device on the moving path;

[0007] According to the feature information, and combining the work data of the suspicious target in the past preset time period, predicting the abnormal probability of the suspicious target appearing abnormal situation, and judging whether the abnormal probability exceeds a preset threshold;

[0008] If the abnormal probability exceeds the preset threshold, the suspicious target is marked as an abnormal target, and the abnormal target is tracked.

[0009] Optionally, the step of determining a suspicious target based on the image information of the target area according to an anomaly detection algorithm comprises:

[0010] Obtaining the image information of the target area, if the observation object is a worker, obtaining the behavior track and safety equipment wearing condition of the observation object;

[0011] Detecting abnormal actions through a spatio-temporal graph convolution network, and comparing the safety equipment wearing state with the preset specification in real time to judge whether the observation object is suspicious, and if so, marking it as a suspicious target.

[0012] Optionally, the step of determining the suspicious target according to the image information of the target area based on the anomaly detection algorithm comprises:

[0013] acquiring image information of the target area, if the observation object is an AGV device, acquiring a running posture and a surface state of the observation object;

[0014] determining whether the observation object is suspiciously abnormal through at least one of dynamic path planning comparison of the observation object, abnormal detection of the angle of the mechanical arm, sudden change detection of the speed, surface defect detection, and temperature anomaly detection, and if so, marking the observation object as a suspicious target.

[0015] Optionally, the step of determining the suspicious target according to the image information of the target area based on the anomaly detection algorithm comprises:

[0016] traversing all observation objects of the target area to acquire a number of suspicious targets in a preset time period;

[0017] if the number of suspicious targets exceeds a first target threshold, marking the target area as an abnormal area, and maintaining the facilities of the target area;

[0018] if the number of suspicious targets exceeds a second target threshold and is less than the first target threshold, marking the target area as a high-risk area, and increasing the inspection frequency of the target area.

[0019] Optionally, the step of predicting an abnormal probability of the suspicious target in an abnormal situation according to the feature information and in combination with work data of the suspicious target in a past preset time period, and determining whether the abnormal probability exceeds a preset threshold comprises:

[0020] acquiring work data of the suspicious target in a past preset time period, wherein the work data comprises at least one of product yield, work amount completion degree, continuous work time length, attendance record, and / or maintenance record.

[0021] Optionally, the step of predicting an abnormal probability of the suspicious target in an abnormal situation according to the feature information and in combination with work data of the suspicious target in a past preset time period, and determining whether the abnormal probability exceeds a preset threshold comprises:

[0022] an expression of the abnormal probability of the suspicious target in an abnormal situation is:

[0023]

[0024] wherein, , and are weight coefficients, P (s, t) represents the probability of the suspicious target passing through the monitoring point s, and z represents the total number of monitoring points, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, and respectively represent the mean and the standard deviation of the feature in the normal state, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xj represents the j-th historical work index, and respectively represent the mean and the standard deviation of the index in the normal state, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xj represents the j-th historical work index, Xi (s, t) represents the spatiotemporal coordinate of the actual trajectory of the suspicious target at the k-th moment, Xi (s, t) represents the spatiotemporal coordinate of the actual trajectory of the suspicious target at the k-th moment, Xi (s, t) represents the spatiotemporal coordinate of the actual trajectory of the suspicious target at the k-th moment, Xi (s, t) represents the spatiotemporal coordinate of the actual trajectory of the suspicious target at the k-th moment,

[0025] Optionally, the weight expression of the real-time feature is as follows:

[0026]

[0027] Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s, Xi (s, t) represents the i-th real-time feature collected at the monitoring point s,

[0028] Optionally, the step of judging whether the abnormal probability exceeds the preset threshold value comprises:

[0029] If the abnormal probability does not exceed the preset threshold value, the suspicious state of the suspicious target is released.

[0030] Optionally, the step of marking the suspicious target as an abnormal target and tracking the abnormal target if the abnormal probability exceeds the preset threshold value comprises:

[0031] When the abnormal probability exceeds one standard deviation of the preset threshold, a third-level alarm is set, and the abnormal event is recorded to a log system;

[0032] When the abnormal probability exceeds two standard deviations of the preset threshold, a second-level alarm is set, and the abnormal target is continuously recorded by a monitoring device;

[0033] When the abnormal probability exceeds three standard deviations of the preset threshold, a first-level alarm is set, an alarm is triggered, and the unmanned aerial vehicle continuously tracks the abnormal target.

[0034] In a second aspect, the application provides a park multi-source heterogeneous data analysis system based on an industrial Internet of Things, characterized in that the system comprises a management platform, a sensing network platform and an object platform which are sequentially established in communication:

[0035] The sensing network platform is configured to:

[0036] According to image information of a target area, a suspicious target is determined based on an abnormality detection algorithm;

[0037] According to a moving trend of the suspicious target, a moving path of the suspicious target is predicted, and feature information of the suspicious target is acquired through a monitoring device on the moving path;

[0038] According to the feature information, and in combination with work data of the suspicious target in a past preset time period, an abnormal probability of the suspicious target to appear an abnormal situation is predicted, and it is judged whether the abnormal probability exceeds a preset threshold;

[0039] The management platform is configured to:

[0040] If the abnormal probability exceeds the preset threshold, the suspicious target is marked as an abnormal target, and the abnormal target is tracked.

[0041] The application can achieve the following beneficial effects:

[0042] The application embodiment provides a park multi-source heterogeneous data analysis method and system based on an industrial Internet of Things. The application embodiment provides a park multi-source heterogeneous data analysis method based on an industrial Internet of Things, which comprises the following steps: determining a suspicious target according to image information of a target area based on an abnormality detection algorithm; predicting a moving path of the suspicious target according to a moving trend of the suspicious target, and acquiring feature information of the suspicious target through a supervision device on the moving path; predicting an abnormality probability of the suspicious target in an abnormal situation according to the feature information and in combination with work data of the suspicious target in a preset time period in the past, and judging whether the abnormality probability exceeds a preset threshold; if the abnormality probability exceeds the preset threshold, marking the suspicious target as an abnormal target, and tracking the abnormal target. Through image information and an abnormality detection algorithm, all-weather automatic monitoring is realized, potential abnormal targets can be identified, and the response time of traditional manual inspection is shortened; in combination with moving path prediction and supervision device linkage, target feature information is collected in real time by using existing sensors, cameras and other multi-source devices, data acquisition efficiency is improved, information island problems are reduced, real-time features and historical work data are fused, the misjudgment rate is reduced through multi-dimensional analysis, the scientific nature of abnormality judgment is ensured, and invalid alarms are reduced. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 FIG. 1 is a flowchart of a park multi-source heterogeneous data analysis method according to an application embodiment;

[0044] Figure 2 FIG. 4 is a framework diagram of a service platform according to an application embodiment;

[0045] Figure 3 FIG. 5 is a framework diagram of a management platform according to an application embodiment;

[0046] Figure 4 FIG. 6 is a framework diagram of a sensor network platform according to an application embodiment.

[0047] The application embodiment will be further described in combination with the drawings. DETAILED DESCRIPTION

[0048] The technical solutions in the application embodiments will be clearly and completely described in combination with the drawings. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0049] It should be noted that all directionality indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain specific posture (as shown in the drawings), and if the specific posture changes, the directionality indications will also change accordingly.

[0050] In the present application, unless otherwise explicitly specified and limited, the terms "connection", "fixation" and the like should be understood broadly, for example, "fixation" can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through an intermediate medium; can be internal communication of two elements or interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0051] In addition, if the present application embodiments involve "first", "second" and the like, the "first", "second" and the like are only for description purposes, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first" and "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel solutions, for example, "A and / or B" includes A solution, or B solution, or A and B solution. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that a person skilled in the art can realize it, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor is it within the scope of protection claimed by the present application.

[0052] Embodiment 1

[0053] Referring to Figure 1 The first embodiment of the present application provides a park multi-source heterogeneous data analysis method based on industrial Internet of Things, comprising the following operation steps:

[0054] S10, determining a suspicious target based on an abnormality detection algorithm according to image information of a target area.

[0055] Optionally, a plurality of cameras are reasonably deployed in the target area of the park to ensure coverage of key areas and passages. Through the industrial Internet of Things platform, image data captured by each camera is collected in real time to form a continuous image sequence. The collected images are preprocessed, the histogram equalization is used to enhance the contrast of the images and improve the image quality in dim environment; the Gaussian filter or median filter algorithm is used to remove noise in the images to improve the definition and purity of the images, thereby providing a reliable image basis for subsequent abnormality detection.

[0056] The pre-trained deep learning model (such as Faster R-CNN) is used for target recognition on the preprocessed image, and various objects such as workers, AGV equipment, and goods in the image are accurately recognized, and their positions and bounding boxes in the image are marked. AGV is the abbreviation of Automated Guided Vehicle.

[0057] In the industrial Internet of Things-based park scenario, multi-source heterogeneous data refers to data from different sources (multi-source) and with different data types and structures (heterogeneous). These data sources are extensive and cover image information of the target area (for abnormal detection to mark suspicious targets), feature information obtained by supervision equipment on the moving path of suspicious targets, work data of suspicious targets in the past preset time period (including product yield, work completion degree, continuous work time, attendance record, maintenance record, etc.), and data obtained through various detection methods (such as spatiotemporal graph convolution network to detect abnormal actions, real-time comparison of safety equipment wearing state, dynamic path planning comparison, mechanical arm angle abnormality detection, etc.). These different sources and types of data together constitute park multi-source heterogeneous data, which is used to analyze and judge the abnormal situations of objects (such as workers and AGV equipment) in the park.

[0058] The following is an illustrative description of suspicious target examples for different objects:

[0059] Abnormal behavior of workers: In the image information, if the worker is in a dangerous area (such as close to a heavy mechanical equipment that is running without protective measures), or makes actions inconsistent with the work process (such as touching the key equipment at non-operation time), after analyzing the behavior trajectory and action mode through the spatiotemporal graph convolution network, it may be determined as a suspicious target.

[0060] Workers not wearing safety equipment as required: If the image information shows that the worker does not wear safety helmets, safety shoes, safety glasses, and other necessary safety equipment, through real-time comparison of safety equipment wearing state with preset specifications, it will be marked as a suspicious target.

[0061] AGV equipment with abnormal running posture: In the image information, if the mechanical arm angle of the AGV equipment deviates significantly from the normal working angle, or the vehicle body is tilted, the wheels are off the ground, etc., through dynamic path planning comparison and mechanical arm angle abnormality detection, etc., it may be determined as a suspicious target.

[0062] AGV equipment with abnormal surface state: If the image information shows that the AGV equipment has obvious damage, oil leakage, deformation, etc. on the surface, through surface defect detection, it will be marked as a suspicious target.

[0063] AGV device with abnormal speed or position: Although image information is static, combined with historical data and preset path, if the AGV device should be in other position at the time when the picture is taken, or the current position is unreasonable according to its historical speed, there may be a sudden change in speed or path deviation problem. Through speed mutation detection and dynamic path planning comparison, it may be determined as a suspicious target.

[0064] S20, according to the moving trend of the suspicious target, predict the moving path of the suspicious target, and obtain the feature information of the suspicious target through the supervision device on the moving path.

[0065] Optionally, the historical moving data of the suspicious target is extracted from the database of the industrial Internet of Things platform, including moving time, position coordinates, moving speed, direction and other information. Clean and pretreat these data to remove outliers and noise data, and ensure the accuracy and reliability of the data. Time series analysis method (such as ARIMA model) or machine learning algorithm (such as LSTM neural network) is used to model the historical moving data of the suspicious target, analyze its moving trend and rule. For example, through the LSTM model, learn the moving mode of the suspicious target in different time periods and different environments, and predict its future moving direction and speed change trend. Based on the historical moving data as a reference, according to the current moving direction, moving speed, current task execution situation and carried goods of the suspicious target, the moving trend of the suspicious target is determined.

[0066] Combined with the digital map of the park, considering the obstacles in the map (walls, devices, goods stacking area, etc.), channel restrictions and traffic rules, the possible moving path of the suspicious target is predicted according to its moving trend. A* algorithm or Dijkstra algorithm is used for path planning to select the optimal moving path. In the prediction process, real-time changes in the environment around the suspicious target are obtained, such as the movement of other targets, the appearance or disappearance of obstacles, etc. According to these dynamic information, the predicted moving path is adjusted in time to ensure the accuracy and feasibility of the path.

[0067] On the predicted moving path, various supervision devices are reasonably deployed, such as high-definition cameras, temperature sensors, pressure sensors, vibration sensors, etc. When the suspicious target enters the monitoring range of the supervision device, its related feature information is collected in real time. For example, through the camera, the appearance image, color, shape and other information of the suspicious target are obtained; through the temperature sensor, the surface temperature is measured; through the pressure sensor and vibration sensor, the pressure and vibration during the operation of the suspicious target are detected. The feature information obtained from different supervision devices is fused to eliminate the redundancy and contradiction between the data, and improve the integrity and accuracy of the data. For example, the image information collected by the camera is associated with the physical quantity information collected by the sensor for correlation analysis, so as to more comprehensively understand the state of the suspicious target.

[0068] S30, according to the feature information, and in combination with the work data of the suspicious target in the past preset time period, an abnormal probability of the suspicious target appearing an abnormal situation is predicted, and it is judged whether the abnormal probability exceeds a preset threshold.

[0069] Optionally, a reasonable abnormal probability preset threshold is set according to the safety standard and actual demand of the park. The calculated abnormal probability is compared with the preset threshold. If the abnormal probability exceeds the preset threshold, it is determined that the suspicious target has a higher abnormal risk. If the abnormal probability does not exceed the preset threshold, it is considered that the suspicious target is in a normal range in the current state, and the suspicious state is released.

[0070] The abnormal situation is exemplarily illustrated as follows:

[0071] The abnormal situation of the worker is exemplarily illustrated as follows:

[0072] Abnormal action detection: In an industrial workshop, the behavior of the worker is analyzed by a spatio-temporal graph convolution network. Under normal circumstances, when the worker operates a machine tool, the action should be continuous and standard, such as slowly approaching the machine tool, accurately pressing the operation button, etc. If a worker suddenly rushes to the machine tool quickly and does not follow the operation process, the probability of the worker's abnormal action in the spatio-temporal graph convolution network model output exceeds the preset threshold, and it is determined that the worker's behavior is abnormal. For example, in an automobile parts production workshop, a worker suddenly touches a high-speed rotating mechanical part without wearing protective gloves. This dangerous action will be identified as an abnormality.

[0073] Illegal operation of equipment: after obtaining the behavior trajectory of the worker by using a deep learning target detection algorithm and a multi-target tracking algorithm, if it is found that the worker violates the preset operation specification when operating the equipment. For example, in the reaction kettle operation area of a chemical park, it is specified that the worker should operate according to a specific order and dosage when adding raw materials. If a worker does not add raw materials according to the order and the dosage exceeds the specified range, it is determined that the worker has an abnormal behavior of illegal operation of equipment by comparison with the preset specification.

[0074] Not wearing necessary equipment: at the entrance of the park or in the industrial workshop, the safety equipment wearing state of the worker is detected by using a trained safety equipment recognition model. If a worker enters a construction site without wearing a safety helmet, the model detects that he does not wear a safety helmet, and the record result is "not wearing". According to the preset specification, it is determined that the worker has an abnormal situation of safety equipment wearing.

[0075] Equipment wearing is not standardized: taking goggles as an example, although the staff wears goggles, the position is not adjusted, resulting in a large gap between the goggles and the face, which cannot play an effective protection role. Through image processing technology and deep learning feature extraction method analysis, it is found that the wearing state of the goggles does not meet the standard, and it is determined that the equipment wearing is not standardized.

[0076] Fatigue work: analyze the work data of the staff in the past preset time period, such as continuous working time. If a staff member works continuously for more than 12 hours, and shows signs of fatigue such as dull eyes and slow reaction in image information. Combined with its work data and real-time state, through the abnormal probability prediction formula, if the abnormal probability exceeds the preset threshold, it is determined that the staff is in the abnormal state of fatigue work.

[0077] AGV equipment abnormality examples are as follows:

[0078] Mechanical arm angle anomaly: use the posture recognition algorithm in computer vision technology to analyze the posture of the mechanical arm of the AGV equipment. According to the design parameters and normal working range of the mechanical arm of the AGV equipment, set the reasonable angle threshold of each joint of the mechanical arm. If the angle of a joint of the mechanical arm of a certain AGV equipment exceeds the set threshold range when transporting goods, for example, the normal working angle range of the joint is 30° - 120°, and the actual detected angle is 130°, it is determined that the mechanical arm angle is abnormal, and the AGV equipment is determined to be suspicious, and it is marked as a suspicious target.

[0079] Vehicle body inclination: calculate the inclination angle and other running posture parameters of the AGV equipment through the posture recognition algorithm. If the inclination angle of the vehicle body of a certain AGV equipment exceeds the normal range during driving, for example, the inclination angle of the vehicle body should be less than 5° during normal driving, but the actual detected inclination angle is 8°, indicating that the AGV equipment may have problems such as wheel failure or unbalanced loading of goods, which belongs to the running posture anomaly.

[0080] Speed mutation anomaly: use the speed sensor of the AGV equipment itself or through image analysis to obtain the real-time running speed of the AGV equipment. The time interval for collecting speed data is every 0.1 second. If the acceleration of the speed data of a certain AGV equipment exceeds the set threshold value during driving, for example, the absolute value of the acceleration exceeds for two consecutive times, it is determined that the AGV equipment has a speed mutation anomaly. For example, the AGV equipment originally travels at a constant speed of 1 m / s, suddenly accelerates to 3 m / s, and completes the acceleration in 0.1 seconds. This sudden change in speed will be identified as an anomaly.

[0081] Surface defect detection: image feature extraction algorithm is used to analyze the surface image of AGV equipment, and the texture, color, shape and other features of the surface are extracted. A surface defect detection model based on deep learning is constructed, and a large number of data sets containing normal and defective AGV equipment surface images are used to train the model. If there are scratches, pits, cracks and other defects on the surface of an AGV equipment, and the model detects that the severity of the defect is moderate or above, for example, the length of the scratch is more than 5 cm and the depth is deep, it is determined that the surface state of the AGV equipment is abnormal.

[0082] Temperature anomaly: temperature sensors are installed at key positions of AGV equipment (such as motors, batteries, controllers, etc.), and real-time temperature data of these positions are collected. According to the normal working temperature range of each component of the AGV equipment, the corresponding temperature threshold is set. If the motor working temperature threshold of an AGV equipment is 80℃, when the real-time monitoring of the motor temperature exceeds 80℃, it is determined that the AGV equipment has temperature anomaly, and there may be problems such as motor overload or poor heat dissipation.

[0083] S40, if the abnormal probability exceeds the preset threshold, the suspicious target is marked as an abnormal target, and the abnormal target is tracked.

[0084] Optionally, when the abnormal probability of the suspicious target exceeds the preset threshold, it is immediately marked as an abnormal target, and special identification is performed in the management system of the industrial Internet of Things platform, so that relevant personnel can quickly identify it. The relevant information of the abnormal target is recorded in detail, including the time, place, type of anomaly (such as equipment failure, personnel violation, etc.), abnormal probability value, feature information and working data. These information is stored in the database to form an abnormal event file, which provides a basis for subsequent analysis and processing.

[0085] Start multiple tracking means to track the abnormal target in real time. Use the camera network in the park to track the video, and use image recognition technology to lock the position and moving track of the abnormal target; use the positioning equipment carried by AGV equipment or unmanned aerial vehicle to accurately track the moving abnormal target; at the same time, combined with the sensor data in the industrial Internet of Things, analyze the running state of the abnormal target and the change of the surrounding environment.

[0086] Embodiment 2

[0087] Based on embodiment 1, the embodiment provides a park multi-source heterogeneous data analysis method based on industrial Internet of Things, including the following operation steps:

[0088] S10, determining a suspicious target based on an image information of a target area according to an abnormal detection algorithm.

[0089] Optionally, the step of determining a suspicious target based on an image information of a target area according to an abnormal detection algorithm comprises:

[0090] S101, acquire image information of the target area, if the observation object is a worker, acquire the behavior trajectory and safety equipment wearing situation of the observation object;

[0091] Optionally, according to actual needs and monitoring range, multiple cameras are reasonably deployed in the target area (such as an industrial workshop, a park entrance, an office area, etc.). For example, in an industrial workshop, cameras are arranged for different production lines and equipment operation areas to ensure that the working area of the worker is fully covered without monitoring dead angles; at the entrance of a park, a high-definition camera is installed to capture the entry and exit behavior and facial features of a person.

[0092] According to the light conditions, distance, and other factors of the monitoring scene, the parameters of the camera are adjusted, such as resolution, frame rate, exposure time, white balance, etc. In areas with relatively dark light, the infrared night vision function of the camera is turned on or additional lighting equipment is added to ensure that the acquired images are clear and accurate. The camera collects image data in real time and transmits the images to the back-end server or edge computing device through wired (such as Ethernet) or wireless (such as Wi-Fi, 5G) networks. To ensure the stability and real-time performance of data transmission, data compression technology is used to reduce bandwidth occupation, and a data caching mechanism is set up to prevent data loss caused by network fluctuations.

[0093] Deep learning target detection algorithms (such as YOLO series, Faster R-CNN, etc.) are used to detect workers in images in real time, and the positions and bounding boxes of the workers are identified. Based on the target detection results, multi-target tracking algorithms (such as DeepSORT, FairMOT, etc.) are used to continuously track the workers. These algorithms can accurately associate the same person in different frames by combining the appearance features and motion information of the target, and obtain the motion trajectory of the person.

[0094] For different safety equipment (such as safety helmets, safety shoes, goggles, protective gloves, etc.), extract their unique appearance features. For example, safety helmets usually have specific colors, shapes and signs; safety shoes have obvious protective structures on the toe part. Use image processing techniques (such as edge detection, color segmentation, etc.) and deep learning feature extraction methods (such as convolutional neural network features) to describe these features. Collect a large number of image data containing safety equipment and personnel without wearing safety equipment, and construct a data set. Use these data to train safety equipment recognition models, such as deep learning-based classification models (ResNet, EfficientNet, etc.). In the training process, use data augmentation techniques (such as rotation, flipping, adding noise, etc.) to expand the data set and improve the generalization ability of the model. Apply the trained safety equipment recognition model to the real-time collected images to detect the safety equipment wearing state of each worker. For each detected person, judge whether they are wearing various safety equipment, and record the detection results (wearing / not wearing). Considering the visibility of safety equipment of workers in different postures and angles, use multi-camera cooperative detection or image stitching technology to obtain personnel images from multiple angles, improve the accuracy of safety equipment wearing state detection.

[0095] S102, detect abnormal actions through a spatio-temporal graph convolution network, compare the safety equipment wearing state in real time with the preset specification, judge whether the observation object is suspicious or not, if yes, mark it as a suspicious target, if not, mark it as a normal target, and do not process it subsequently.

[0096] Optionally, each worker is regarded as a node in the graph, and the attributes of the node include the behavior characteristics (such as action posture, movement speed, acceleration, etc.) and appearance characteristics (such as clothing color, body shape, etc.) of the personnel. Behavior characteristics can be obtained by further analyzing personnel behavior trajectory data, such as using motion analysis algorithms to calculate personnel motion parameters; appearance characteristics can be extracted from the results of target detection and recognition.

[0097] Collect a large number of personnel behavior video data containing normal and abnormal actions, and label the video data to clearly define the category (normal / abnormal) of each action segment. Convert the video data into spatio-temporal graph data as the training set and test set of the model. The spatio-temporal graph convolution network (ST-GCN) is composed of multiple spatio-temporal graph convolution layers, pooling layers and fully connected layers. The spatio-temporal graph convolution layer is used to extract the features in the spatio-temporal graph, which captures the local and global features of personnel actions by convolution operation in the spatial and temporal dimensions. The pooling layer is used to reduce the dimension of the feature map and improve the computational efficiency of the model. The fully connected layer is used to classify the extracted features, outputting the probability that the action belongs to the normal or abnormal category.

[0098] The real-time collected personnel behavior trajectory data and appearance feature data are constructed into a space-time graph and input into the trained ST-GCN model. The model outputs the probability of the personnel action belonging to the normal or abnormal category. If the probability of the abnormal category exceeds a preset threshold, it is considered that the personnel has an abnormal action. Abnormal actions include sudden falling, staggering, frequent looking around, and violating operation equipment

[0099] Optionally, the step of determining the suspicious target according to the image information of the target region based on an anomaly detection algorithm comprises:

[0100] S110, acquiring image information of a target region, if the observation object is an AGV device, acquiring the running posture and surface state of the observation object;

[0101] Optionally, a posture recognition algorithm in computer vision technology, such as a posture estimation model based on deep learning (such as OpenPose, AlphaPose, etc.), is used to analyze the posture of the AGV device in the image. The key joints of the AGV device (such as wheels, body connections, mechanical arm joints, etc.) are identified by training the model, and the relative positions and angles between these joints are calculated to determine the running posture of the AGV device. Key running posture parameters are extracted from the posture recognition results, such as the inclination angle of the AGV device, the turning radius, the extension angle and rotation angle of the mechanical arm, etc. These parameters can directly reflect whether the running state of the AGV device is normal.

[0102] An image feature extraction algorithm (such as SIFT, SURF, HOG, etc.) is used to analyze the surface image of the AGV device and extract features such as texture, color, shape, etc. For example, the SIFT algorithm is used to detect the key points on the surface of the AGV device and calculate the feature descriptors of these key points, which are used for subsequent surface state analysis. A surface defect detection model based on deep learning is constructed, such as a convolutional neural network (CNN) model. A large number of data sets containing normal and defective AGV device surface images are used to train the model, so that the model can automatically learn and identify the defect features of the AGV device surface, such as scratches, dents, cracks, etc.

[0103] S120, by at least one of the dynamic path planning comparison of the observation object, the mechanical arm angle anomaly detection, the speed mutation detection, the surface defect detection and the temperature anomaly detection, it is judged whether the observation object is suspicious or not, if yes, it is marked as a suspicious target, if not, it is marked as a normal target, and it is not processed subsequently.

[0104] Optionally, AGV devices usually use preset path planning algorithms (such as A* algorithm, Dijkstra algorithm, etc.) for navigation. In normal operation, AGV devices should travel according to the planned path. The system obtains the actual travel path of the AGV device in real time and compares it with the preset planned path. The deviation value between the actual travel path and the planned path is calculated, such as the lateral deviation, longitudinal deviation and angle deviation of the path, etc. Set a reasonable deviation threshold, when the actual path deviation exceeds the threshold, it is determined that the AGV device may have an abnormality. For example, if the lateral deviation continuously exceeds 10 cm and the duration exceeds 5 seconds, it is determined to be a suspected abnormality, and it is marked as a suspected target.

[0105] According to the design parameters and normal working range of the mechanical arm of the AGV device, set reasonable angle threshold values for each joint of the mechanical arm. For example, the angle range of a certain joint of the mechanical arm during normal operation is 30°-120°, and angles less than 30° or greater than 120° are set as abnormal threshold values. Through the angle sensor installed at the joint of the mechanical arm or based on the angle data obtained by image analysis, the angle change of each joint of the mechanical arm is monitored in real time. When the monitored angle exceeds the set threshold range, it is determined that the angle of the mechanical arm is abnormal, the AGV device is determined to be a suspected abnormality, and it is marked as a suspected target.

[0106] The real-time running speed of the AGV device is obtained by using the speed sensor of the AGV device itself or through image analysis (such as based on the optical flow method). The time interval for collecting speed data should be small enough to accurately capture the mutation of speed, for example, collecting speed data every 0.1 second. Calculate the rate of change of adjacent time speed data, i.e. acceleration. Set a reasonable acceleration threshold, when the acceleration exceeds the threshold, it is determined that the AGV device has a speed mutation abnormality. For example, if the absolute value of acceleration continuously exceeds 10 m / s2 twice, it is considered that the AGV device has a speed abnormality, it is determined to be a suspected abnormality, and it is marked as a suspected target.

[0107] The surface features of the AGV device obtained by surface feature extraction are input into the trained surface defect detection model, and the model outputs whether there is a defect on the surface and the type of the defect (such as scratches, pits, cracks, etc.). At the same time, the severity of the defect is evaluated, for example, the defect is divided into mild, moderate and severe according to the area, depth and other indicators of the defect. Set the abnormality judgment rule of defect detection, such as when moderate and above degree defects are detected on the surface of the AGV device, it is determined that the surface state of the AGV device is abnormal, it is determined to be a suspected abnormality, and it is marked as a suspected target.

[0108] ​Install temperature sensors at key parts of the AGV device (such as motors, batteries, controllers, etc.), and collect real-time temperature data of these parts. The accuracy of the temperature sensor should meet the requirements of anomaly detection, for example, the accuracy is ±0.5℃. According to the normal working temperature range of each component of the AGV device, set the corresponding temperature threshold. Compare the real-time collected temperature data with the set threshold, when the temperature exceeds the threshold, it is determined that the AGV device has a temperature anomaly, and it is determined as a suspicious anomaly, and it is marked as a suspicious target. For example, if the working temperature threshold of the motor is 80℃, when the real-time monitoring of the motor temperature exceeds 80℃, the temperature anomaly alarm is triggered.

[0109] Optionally, the step of determining the suspicious target according to the image information of the target area based on the anomaly detection algorithm comprises:

[0110] S1001, traverse all observation objects of the target area, and obtain the number of suspicious targets in a preset time period;

[0111] Optionally, according to the actual scene requirement and the sensitivity requirement of anomaly detection, a reasonable time window (preset time period) is set, such as 1 hour, half a day, one day, etc. In each time window, the observation objects determined as suspicious targets are counted, and the number of suspicious targets is updated in real time. A counter can be used, and the counter is incremented by 1 each time a suspicious target is detected. At the same time, the number of suspicious targets currently counted is saved at the end of the time window, so as to be compared with the target threshold in the subsequent step.

[0112] S1002, if the number of suspicious targets exceeds the first target threshold, mark the target area as an abnormal area, and maintain the facilities in the target area;

[0113] Optionally, collect the number of suspicious targets of the target area in the normal running state, analyze the distribution rule and change trend. For example, the number of suspicious targets in different time periods of the past month is counted, and the statistical quantities such as mean value and standard deviation are calculated, which are used as a reference basis for setting the first target threshold. If the number of suspicious targets exceeds the first target threshold, it means that the number of suspicious targets obtained is inaccurate, which may represent the failure of the monitored facilities. Therefore, the facilities in the target area need to be maintained.

[0114] S1003, if the number of suspicious targets exceeds the second target threshold and is less than the first target threshold, mark the target area as a high-risk area, and increase the inspection frequency of the target area.

[0115] Optionally, the second target threshold should be less than the first target threshold, and the difference between the two should be reasonably set according to the actual situation of the target area and the accuracy requirement of anomaly detection. Generally speaking, the difference should not be too large to ensure that potential abnormal risks of the target area can be discovered in time. According to the risk level of the target area and the change situation of the number of suspicious targets, the inspection plan is dynamically adjusted. For example, under normal circumstances, the inspection frequency of the target area is once a day; when the number of suspicious targets exceeds the second target threshold, the inspection frequency is increased to four times a day, and inspection is carried out at different time periods such as morning, afternoon, evening, etc.

[0116] Optionally, the step of predicting the abnormal probability of the suspicious target under abnormal conditions according to the feature information and in combination with the work data of the suspicious target in the past preset time period, and judging whether the abnormal probability exceeds a preset threshold, comprises:

[0117] The work data of the suspicious target in the past preset time period is obtained, wherein the work data includes at least one of product yield, work amount completion degree, continuous work time length, attendance record and / or maintenance record.

[0118] S20, according to the moving trend of the suspicious target, the moving path of the suspicious target is predicted, and the feature information of the suspicious target is obtained through the monitoring equipment on the moving path.

[0119] S30, according to the feature information, and in combination with the work data of the suspicious target in the past preset time period, the abnormal probability of the suspicious target under abnormal conditions is predicted, and it is judged whether the abnormal probability exceeds a preset threshold.

[0120] Optionally, the expression of predicting the abnormal probability of the suspicious target under abnormal conditions is:

[0121]

[0122] In the formula, , and are weight coefficients, represents the probability of the suspicious target passing through the monitoring point s, and z represents the total number of monitoring points, represents the i-th real-time feature collected at the monitoring point s, and respectively represent the mean and standard deviation of the feature under normal state, represents the weight of the real-time feature i, and m represents the number of real-time features, represents the j-th historical work index, and respectively represent the index the mean and standard deviation in normal state, the weight of the index , n represents the total number of historical work indexes, represents the space-time coordinates of the suspicious target actual trajectory at time k, represents the theoretical coordinates based on the movement trend prediction, represents the measurement error tolerance of the positioning system, and q represents the number of collected coordinates, represents the normalization function.

[0123] is a Sigmoid normalization function that maps the calculation result of the formula to the interval (0, 1), making the value range of the anomaly probability P reasonable.

[0124] , and are weight coefficients, and a, b, and c represent the relative importance of the real-time feature deviation from normal, the historical work index deviation from normal, and the trajectory prediction deviation in the calculation of the anomaly probability. By adjusting these three weight coefficients, the influence of different factors in the anomaly judgment can be flexibly set according to the actual scene. For example, in a device running state monitoring scenario, more attention may be paid to real-time features, so the value of a can be appropriately increased, in a scenario that needs to consider the long-term running trend of the device, the value of b can be increased, and in a scenario that needs to accurately monitor the movement trajectory of the device, the value of c can be increased.

[0125] represents the probability of the suspicious target appearing at the monitoring point s. In the calculation of the weighted sum of the real-time feature deviation from normal, this probability is used to weight the real-time features collected by different monitoring points. If a monitoring point has a high probability of passing through the suspicious target, the real-time features collected by this monitoring point will have a greater contribution to the calculation of the anomaly probability, because these features are more likely to reflect the true state of the suspicious target.

[0126] Z represents the number of monitoring points set in the target area. As the upper limit of the sum, it determines the range of comprehensive analysis of the real-time features collected by all monitoring points. The more the number of monitoring points, the more comprehensive the monitoring of the suspicious target, but it will also increase the calculation complexity.

[0127] is a certain real-time feature value of the suspicious target collected at a specific monitoring point s, such as temperature, speed, attitude angle, etc. It is the basis data for calculating the deviation of the real-time feature from normal. By comparing with the mean and standard deviation in normal state, the abnormality of this feature at a specific monitoring point can be quantified. ​​​

[0128] represents the historical average of the i-th real-time feature under normal conditions. As a reference benchmark of normal conditions, it is used to measure whether the real-time feature deviates from the normal range. It is one of the key parameters to calculate the degree of feature deviation.

[0129] represents the historical standard deviation of the i-th real-time feature under normal conditions, reflecting the fluctuation range of the feature value. Together with the mean value, it is used to standardize the deviation degree of the real-time feature . By dividing the deviation by the standard deviation, the dimensions of different features can be unified, making the contributions of different features to the abnormal probability comparable.

[0130] represents the relative importance of the i-th real-time feature in the calculation of the abnormal probability. Different real-time features have different sensitivities to abnormalities, and the weight can be adjusted according to the importance of the feature and its ability to reflect abnormalities. For example, for the temperature feature of a device, if it is more sensitive to abnormal conditions, a larger weight can be given.

[0131] m represents the number of real-time features used to monitor the suspicious target. It determines the dimension of the comprehensive analysis of real-time features. The more the number of features, the more comprehensive the description of the suspicious target, but it also increases the calculation and data collection difficulty.

[0132] is a certain indicator value of the suspicious target in historical work, such as the cumulative running time of the device, the number of faults, energy consumption; the yield of the staff, the continuous working time, the attendance situation, etc. By comparing with the mean and standard deviation under normal conditions, the abnormal degree of the suspicious target in historical work indicators can be measured. Historical work indicators can reflect the long-term running state and potential problems of the device, providing additional information for abnormal probability calculation. And through the yield of the staff, the continuous working time, the attendance situation, it reflects the condition of the staff, when the yield is greatly reduced, the continuous working time is too long, and the recent frequent late and early leave, it indicates that the probability of staff abnormality is larger.

[0133] represents the historical average of the j-th historical work indicator under normal conditions. As a reference benchmark of normal conditions for historical work indicators, it is used to measure whether it deviates from the normal range.

[0134] represents the historical standard deviation of the j-th historical work indicator under normal conditions, reflecting the fluctuation range of the indicator value. Together with the mean together, the deviation degree of the standardized historical work indicators, so that the contribution of different indicators to the abnormal probability is comparable.

[0135] represents the relative importance of the jth historical work indicator in the calculation of the abnormal probability. The weight can be adjusted according to the importance of the indicator and the degree of association with the abnormality. For example, for the failure frequency indicator of the equipment, if it is closely related to the abnormal situation, a larger weight can be given.

[0136] n represents the number of historical work indicators used in the calculation of the abnormal probability. It determines the range of comprehensive analysis of the historical work indicators.

[0137] represents the specific position and time information of the suspicious target at time k in the actual operation process. By comparing with the theoretical coordinates predicted based on the movement trend, it can be measured whether the movement trajectory of the suspicious target deviates from the expectation, reflecting the abnormal situation of the suspicious target in the movement process.

[0138] represents the position and time information that the suspicious target should theoretically reach at time k according to the movement trend prediction algorithm of the suspicious target. As a reference benchmark for the actual trajectory, it is used to evaluate whether the movement of the suspicious target meets the expectation. If the actual trajectory deviates greatly from the theoretical coordinates, it may mean that the suspicious target has an abnormal situation.

[0139] represents the maximum error range that the positioning system may have when measuring the position of the suspicious target. It is used to standardize the trajectory prediction deviation , taking the measurement error of the positioning system into account. By squaring the deviation and dividing it by the square of the error tolerance, the influence of the measurement error on the calculation of the abnormal probability can be eliminated, so that the trajectory prediction deviation at different times has comparable contribution to the abnormal probability.

[0140] q represents the number of spatiotemporal coordinates of the suspicious target collected within a period of time. It determines the time span of the comprehensive analysis of the trajectory prediction deviation.

[0141] Optionally, the weight expression of the real-time feature is:

[0142]

[0143] In the formula, represents the dynamic weight of the ith real-time feature at time t, represents the probability of the suspicious target passing through the monitoring point s at time t, represents the ith real-time feature collected at the monitoring point s,​ represents the historical mean of the i-th feature in the normal state, represents the historical standard deviation of the i-th feature in the normal state, z represents the total number of monitoring points, and m represents the number of real-time features.

[0144] represents the relative importance of the i-th real-time feature in the abnormal probability calculation at a specific time t, which is dynamically adjusted with time and real-time features. By considering the time factor and the deviation degree of real-time features, the weight can more accurately reflect the sensitivity of the feature to the anomaly at the current time. For example, when a certain real-time feature deviates greatly at a certain time, its weight will increase accordingly, thereby playing a greater role in the abnormal probability calculation.

[0145] represents the likelihood of the suspicious target appearing at the monitoring point s at a specific time t. Similar to the abnormal probability prediction formula, in the calculation of the dynamic weight of real-time features, this probability is used to weight the real-time features collected by different monitoring points. The probability of the suspicious target passing through the monitoring point at different time points may be different, so considering the time factor can make the weight calculation more accurate.

[0146] S40, if the abnormal probability exceeds the preset threshold, the suspicious target is marked as an abnormal target, and the abnormal target is tracked.

[0147] Optionally, if the abnormal probability does not exceed the preset threshold, the suspicious state of the suspicious target is removed.

[0148] Specifically, when the suspicious state is removed, the system updates the state identifier of the target, changes it from the "suspicious" state to the "normal" state, and synchronously updates the related database record. At the same time, the system generates a state change log, records the time of removing the suspicious state, the target identifier, the removal reason (the abnormal probability does not exceed the threshold), and other information, and stores the log in the log system for subsequent query and audit.

[0149] Optionally, the step of marking the suspicious target as an abnormal target if the abnormal probability exceeds the preset threshold and tracking the abnormal target comprises:

[0150] S401, when the abnormal probability exceeds one standard deviation of the preset threshold, set the third level alarm, and record the abnormal event to the log system;

[0151] Optionally, the system calculates the standard deviation S of the abnormal probability according to the historical abnormal probability data. The standard deviation reflects the dispersion degree of the abnormal probability, and is used to measure the degree of deviation of the abnormal probability from the preset threshold.

[0152] The system records the information related to the abnormal event to the log system. The recorded content includes the abnormal target identification, the abnormal occurrence time, the abnormal probability value, the alarm level, the target current position, the moving speed, the direction and other key information. At the same time, the abnormal event is classified and coded, which is convenient for subsequent data analysis and statistics. After receiving the third-level alarm, the monitoring personnel needs to manually review the abnormal event within the specified time (such as 10 minutes). The review content includes checking the historical data of the target, monitoring video playback, communicating with relevant personnel, etc., to confirm whether there is a false alarm.

[0153] S402, when the abnormal probability exceeds two standard deviations of the preset threshold value, set as a second-level alarm, and continuously record the abnormal target through the monitoring device;

[0154] Optionally, the alarm information is displayed on the monitoring interface in the form of an orange icon or text prompt, and a relatively obvious prompt sound is emitted, attracting the attention of the monitoring personnel. The monitoring device continuously records the abnormal target. The monitoring device includes a camera, a sensor and the like, and real-time collects multi-dimensional data such as images, sounds, positions and speeds of the target, and transmits these data to the monitoring center for storage and analysis in real time. The system automatically analyzes and evaluates the continuously recorded data.

[0155] S403, when the abnormal probability exceeds three standard deviations of the preset threshold value, set as a first-level alarm, trigger the alarm, and the unmanned aerial vehicle continuously tracks the abnormal target.

[0156] Optionally, the alarm information is displayed on the monitoring interface in the form of a red icon or text prompt, and a high-decibel alarm is triggered, so as to ensure that the monitoring personnel and relevant personnel can notice it at the first time. The system automatically dispatches the unmanned aerial vehicle to continuously track the abnormal target. The unmanned aerial vehicle is equipped with a high-definition camera, a thermal imager and the like, and can obtain high-definition images, thermal imaging information and the like of the target in real time, and transmit these information to the monitoring center in real time.

[0157] Embodiment 3

[0158] On the basis of embodiment 1, the application discloses a park multi-source heterogeneous data analysis system based on an industrial Internet of Things, comprising a management platform, a sensing network platform and an object platform which are sequentially established in communication:

[0159] The sensing network platform is configured to:

[0160] According to the image information of the target area, determine a suspicious target based on an abnormality detection algorithm;

[0161] According to the moving trend of the suspicious target, predict a moving path of the suspicious target, and obtain feature information of the suspicious target through a supervision device on the moving path;

[0162] According to the feature information, and in combination with the work data of the suspicious target in a preset time period in the past, an abnormal probability of the suspicious target appearing an abnormal situation is predicted, and it is determined whether the abnormal probability exceeds a preset threshold value;

[0163] The management platform is configured to:

[0164] If the abnormal probability exceeds the preset threshold value, the suspicious target is marked as an abnormal target, and the abnormal target is tracked.

[0165] The user platform is configured to serve the functions of the user and the front end. The user obtains the required perception service information through the user platform, processes the perception service information, and converts it into user perception information. The user analyzes the user perception information and makes corresponding decisions in combination with their own intentions. The user perception information is converted into user control information through a corresponding information system and sent to the service platform, thereby expressing the user's corresponding service demand intentions.

[0166] The physical entity of the user platform includes various user terminals, such as mobile phones, computers, and special terminals, which realize the service of the user terminal through the combination with the user information system software.

[0167] The service platform is configured as an API server or other server for establishing communication between the management platform and the user platform to realize corresponding functions. The physical entity of the service platform includes various servers.

[0168] The management platform is configured to perform at least one of device running state monitoring management, data monitoring management, device parameter management, and life cycle management. The management platform is an operation and overall planning platform of the Internet of Things, which can include various management sub-platforms that perform different management businesses. The physical entity of the management platform includes various servers.

[0169] The sensing network platform is configured to perform at least one of network management, instruction management, device state management, data protocol management, data analysis, data classification, data transmission monitoring, and data transmission security management. The sensing network platform provides functions of communication transmission, analysis, identification, and classification of data, which avoids the direct aggregation of data of various object platforms in the management platform, causing the data of the management platform to be redundant and inefficient in data processing. The physical entity of the object platform includes various gateways and edge computing devices.

[0170] The object platform is configured to perform specific production control, detection, measurement, and other production work. The physical entity in the object platform includes various production devices and sensors.

[0171] Optionally, the sensing network platform includes a total database in communication with the management platform and at least two sensing network sub-platforms in communication with the total database.

[0172] Optionally, each sensing network sub-platform corresponds to an API function or an API server.

[0173] Embodiment 4

[0174] The embodiment provides a computer device, including a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements any method described above when executing the computer program.

[0175] Embodiment 5

[0176] The embodiment provides a computer readable storage medium, which stores a computer program capable of being loaded and executed by a processor to perform the method described above.

[0177] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings of the present application, or directly or indirectly applied to other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for analyzing multi-source heterogeneous data in a park based on the Industrial Internet of Things, characterized in that, include: Based on the image information of the target area, suspicious targets are identified using an anomaly detection algorithm; Based on the movement trend of the suspected target, predict the movement path of the suspected target, and obtain the characteristic information of the suspected target through the monitoring equipment along the movement path; Based on the aforementioned feature information and combined with the work data of the suspected target within a preset time period, the probability of an abnormal situation occurring in the suspected target is predicted, and it is determined whether the probability of an abnormal situation exceeds a preset threshold. The work data includes at least one of the following: product yield rate, workload completion rate, continuous working time, attendance records, and / or maintenance records. The expression for predicting the probability of an anomaly in the suspected target is as follows: In the formula, , and These are the weighting coefficients. Let represent the probability that a suspicious target passes through monitoring point s, and z represent the total number of monitoring points. This represents the i-th real-time feature collected at monitoring point s. and Representing features respectively The mean and standard deviation under normal conditions. The weight of real-time feature i is represented by m, and the number of real-time features is represented by m. This represents the j-th historical work indicator. and Representing indicators The mean and standard deviation under normal conditions. Indicators The weights, where n represents the total number of historical work indicators. This represents the spatiotemporal coordinates of the actual trajectory of the questionable target at time k. Represents theoretical coordinates based on movement trend prediction. This indicates the measurement error tolerance of the positioning system, and q represents the number of coordinates collected. Represents the normalization function; If the anomaly probability exceeds a preset threshold, the suspicious target is marked as an abnormal target and the abnormal target is tracked.

2. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of determining suspicious targets based on image information of the target area and an anomaly detection algorithm includes: Acquire image information of the target area; if the observed object is a staff member, acquire the observed object's behavior trajectory and safety equipment wearing status. Abnormal actions are detected by spatiotemporal graph convolutional networks, and the wearing status of safety equipment is compared with preset standards in real time to determine whether the observed object is suspected of being abnormal. If so, it is marked as a suspicious target.

3. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of determining suspicious targets based on image information of the target area and an anomaly detection algorithm includes: Acquire image information of the target area; if the observed object is an AGV device, acquire the running posture and surface state of the observed object. By comparing the dynamic path planning of the observed object, detecting abnormal robot arm angles, detecting sudden speed changes, detecting surface defects, and detecting abnormal temperatures, it is determined whether the observed object is suspected of being abnormal. If so, it is marked as a suspicious target.

4. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of determining suspicious targets based on image information of the target area and an anomaly detection algorithm includes: Traverse all observed objects in the target area to obtain the number of suspicious targets within a preset time period; If the number of suspicious targets exceeds the first target threshold, the target area will be marked as an abnormal area, and the facilities in the target area will be maintained. If the number of suspicious targets exceeds the second target threshold but is less than the first target threshold, the target area will be marked as a high-risk area, and the inspection frequency of the target area will be increased.

5. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The weight expression for real-time features is: In the formula, This represents the dynamic weight of the i-th real-time feature at time t. This represents the probability that a suspicious target passes through monitoring point s at time t. This represents the i-th real-time feature collected at monitoring point s. This represents the historical mean of the i-th feature under normal conditions. Let z represent the historical standard deviation of the i-th feature under normal conditions, z represent the total number of monitoring points, and m represent the number of real-time feature types.

6. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of predicting the probability of an anomaly in the suspected target's behavior based on the feature information and in conjunction with the suspected target's work data over a preset time period, and determining whether the anomaly probability exceeds a preset threshold, includes: If the anomaly probability does not exceed a preset threshold, the suspicious status of the target is lifted.

7. The method for analyzing multi-source heterogeneous data in industrial parks based on the Industrial Internet of Things as described in claim 1, characterized in that, The step of marking the suspicious target as an abnormal target and tracking the abnormal target if the anomaly probability exceeds a preset threshold includes: When the probability of an anomaly exceeds one standard deviation of a preset threshold, a level three alarm is set, and the anomaly event is recorded to the log system. When the probability of an anomaly exceeds twice the standard deviation of a preset threshold, a level two alarm is set, and the abnormal target is continuously recorded by the monitoring equipment. When the probability of an anomaly exceeds three times the standard deviation of a preset threshold, a Level 1 alarm is triggered, and the drone continues to track the abnormal target.

8. A multi-source heterogeneous data analysis system for industrial parks based on the Industrial Internet of Things, characterized in that, This includes establishing a communication management platform, a sensor network platform, and an object platform in sequence: The sensor network platform is configured as follows: Based on the image information of the target area, suspicious targets are identified using an anomaly detection algorithm; Based on the movement trend of the suspected target, predict the movement path of the suspected target, and obtain the characteristic information of the suspected target through the monitoring equipment along the movement path; Based on the aforementioned feature information and combined with the work data of the suspected target over a preset time period, the probability of an abnormal situation occurring in the suspected target is predicted, and it is determined whether the probability of an abnormal situation exceeds a preset threshold. The work data includes at least one of the following: product yield rate, workload completion rate, continuous working hours, attendance records and / or maintenance records; The expression for predicting the probability of an anomaly in the suspected target is as follows: In the formula, , and These are the weighting coefficients. Let represent the probability that a suspicious target passes through monitoring point s, and z represent the total number of monitoring points. This represents the i-th real-time feature collected at monitoring point s. and Representing features respectively The mean and standard deviation under normal conditions. The weight of real-time feature i is represented by m, and the number of real-time features is represented by m. This represents the j-th historical work indicator. and Representing indicators The mean and standard deviation under normal conditions. Indicators The weights, where n represents the total number of historical work indicators. This represents the spatiotemporal coordinates of the actual trajectory of the questionable target at time k. Represents theoretical coordinates based on movement trend prediction. This indicates the measurement error tolerance of the positioning system, and q represents the number of coordinates collected. Represents the normalization function; The management platform is configured as follows: If the anomaly probability exceeds a preset threshold, the suspicious target is marked as an abnormal target and the abnormal target is tracked.

Images