Production line abnormity real-time diagnosis system based on industrial internet of things

Abstract

The invention belongs to the technical field of fault prediction and management, and discloses a production line abnormity real-time diagnosis system based on industrial Internet of Things, which comprises a data acquisition and processing module, a distributed sensor network covering key equipment of a production line is constructed, multi-dimensional production line data is acquired, and the data acquisition and processing module is used for acquiring data of the production line; a self-adaptive sampling technology is adopted to dynamically adjust the multi-dimensional production line data acquisition frequency according to the equipment state, and preliminary multi-dimensional production line data processing is executed at the edge end; and the multi-scale time sequence management module adopts a hot, warm and cold three-level hierarchical storage architecture, compulsively switches sampling frequencies of key equipment parameters in combination with a multi-level safety threshold mechanism, performs resource allocation through a hierarchical calculation architecture, and introduces an abnormal sensitive new mode detection and double-track system template updating mechanism to identify a novel abnormal mode. The state change of the equipment is continuously monitored; it is ensured that resources can be efficiently scheduled in normal, early warning and abnormal states, and the anti-risk capacity of the system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction and management, and more specifically, to a real-time diagnosis system for production line anomalies based on the industrial Internet of Things. Background Art

[0002] Patent Publication No. CN118915566A discloses an online monitoring system for anomalies in HVAC equipment based on the Internet of Things, belonging to the technical field of HVAC control. The system includes: a data acquisition module for real-time acquisition of multi-dimensional operation data of HVAC equipment; an edge computing module responsible for performing real-time data processing, preliminary anomaly detection, and local caching at the device end; a communication module to ensure the reliability, security, and efficiency of data transmission; a central processing module for performing in-depth data analysis using cloud computing resources, maintaining a multi-dimensional dynamic benchmark model, and conducting comprehensive anomaly diagnosis in combination with a knowledge graph; an adaptive threshold management module for dynamically optimizing the anomaly detection threshold to balance detection accuracy and efficiency; a predictive maintenance module for predicting equipment performance trends and potential faults based on historical fault data and real-time data, formulating personalized maintenance plans, and optimizing resource scheduling; and a human-computer interaction module providing a multi-platform, visual user interface.

[0003] Existing real-time diagnosis systems for production line anomalies in industrial Internet of Things mainly have the following problems: Many existing elastic scaling technologies, such as the automatic scaling method with fixed thresholds, often over-allocate resources during peak loads, resulting in resource waste. Even when the load is low, the system still maintains a large resource pool and fails to reduce resources in a timely manner, which increases the waste of computing resources and costs. In traditional elastic resource management, resource adjustment is often based on fixed time intervals or preset rules, and this method is slow to respond to load fluctuations and cannot quickly adapt to sudden load changes. When the load changes, the system may experience resource shortages or over-allocation, affecting performance or causing resource waste.

[0004] Existing technologies usually use the same resource allocation strategy for all tasks, ignoring the importance and priority of tasks. This results in high-priority tasks may not be able to obtain sufficient resources in a timely manner, affecting the performance of critical tasks, and at the same time may also cause low-priority tasks to occupy too many resources, affecting the overall performance of the system. Traditional elastic scaling methods have the problem of a fixed resource pool size and are difficult to make refined adjustments according to real-time load conditions, resulting in the system being unable to achieve optimal resource allocation in the case of large load fluctuations. The elastic scaling mechanisms in existing technologies usually rely on simple rules, which are too single in response to changes in system load and lack flexibility. The resource scheduling and task allocation in existing technologies usually adopt a fixed mode and lack adaptability, and cannot be flexibly adjusted according to real-time workloads and task requirements.

[0005] In view of this, the present invention proposes a real-time production line anomaly diagnosis system based on the industrial Internet of Things to solve the above problems. Summary of the Invention

[0006] To overcome the above defects of the prior art and to achieve the above object, the present invention provides the following technical solutions: A real-time production line anomaly diagnosis system based on the industrial Internet of Things, including: An acquisition and processing module, which constructs a distributed sensor network covering key production line equipment, acquires multi-dimensional production line data, predicts the equipment status based on the multi-dimensional production line data, dynamically adjusts the multi-dimensional production line data acquisition frequency according to the equipment status by using an adaptive sampling technology, and performs preliminary multi-dimensional production line data processing at the edge side; A multi-scale time series management module, which adopts a three-level hierarchical storage architecture of hot, warm, and cold, combines a multi-level security threshold mechanism to forcibly switch the sampling frequency of key equipment parameters, allocates resources through a hierarchical computing architecture, introduces an anomaly-sensitive new mode detection and a dual-track template update mechanism to identify new anomaly modes, and continuously monitors changes in equipment status; An anomaly detection module, which constructs a context-aware anomaly detection framework to adjust the detection strategy, and adopts a hybrid detection method combining deep learning and detection rules to perform full coverage detection of point anomalies, pattern anomalies, and system anomalies; A root cause analysis module, which constructs a multi-dimensional industrial knowledge graph based on the preliminarily processed multi-dimensional production line data, visualizes the fault propagation path through the multi-dimensional industrial knowledge graph, and diagnoses the root cause of different anomalies in real time; A multi-terminal collaborative interaction module, which supports access by multiple terminals and provides role-based differential information display and intelligent interaction, and automatically adjusts the anomaly information presentation method according to the root cause of different anomalies.

[0007] Preferably, the method for constructing a distributed sensor network covering key production line equipment includes: Comprehensively investigate all key equipment on the production line to determine the key equipment, which includes production equipment, auxiliary equipment, energy and power equipment, material handling and conveying equipment, detection equipment, and process control equipment; For each key equipment, determine the core parameters to be monitored, which include temperature parameters, vibration parameters, pressure parameters, current parameters, rotational speed parameters, and noise parameters; obtain historical key equipment failure data, use a clustering algorithm to determine the anomaly modes that each key equipment has historically had, and obtain the corresponding core parameter change data; based on the core parameter change data and the historically existing anomaly modes of each key equipment, plan the optimal monitoring points for each key equipment according to a preset evaluation standard; The preset evaluation criteria include the strongest signal strength criterion, the shortest installation time criterion, the shortest maintenance time criterion, and the minimum interference to the equipment criterion; corresponding sensor types are selected according to the monitored core parameters, and three installation methods are used for classification implementation, and the three installation methods include magnetic installation, adhesive fixing installation, and mechanical fixing installation.

[0008] Preferably, the multi-dimensional production line data includes equipment physical state data, electrical parameter data, process parameter data, product information data, and production line environment data.

[0009] Preferably, the method for dynamically adjusting the multi-dimensional production line data acquisition frequency includes: Construct an equipment state prediction model, use historical multi-dimensional production line data as the input of the equipment state prediction model, and output three equipment states. The equipment states include stable state, transitional state, and abnormal state. The equipment state prediction model is a fully connected neural network model; generate a dynamic adjustment mechanism according to the three equipment states, preset a reference acquisition frequency and a high acquisition frequency. If the equipment state is in a stable state, reduce the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the equipment state is in a transitional state, increase the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the equipment state is in an abnormal state, trigger an event-driven mechanism and increase the multi-dimensional production line data to the high acquisition frequency.

[0010] Preferably, the method for performing preliminary multi-dimensional production line data processing at the edge end includes: At the edge end, identify and remove outliers in the multi-dimensional production line data based on the box plot method, and fill in the missing values in the multi-dimensional production line data by the interpolation method; denoise the multi-dimensional production line data after filling in the missing values by mean filtering, and perform timestamp alignment and standard deviation normalization processing on the denoised multi-dimensional production line data.

[0011] Preferably, the method for forcibly switching the sampling frequency of key equipment parameters includes: The hot, warm, and cold three-level hierarchical storage architecture includes a hot data layer, a warm data layer, and a cold data layer. Preset multi-level security thresholds include five-level thresholds of normal operation threshold, warning threshold, alarm threshold, danger threshold, and emergency threshold; Preset the default sampling frequency of key equipment, allocate key equipment parameters to the hot data layer, warm data layer, and cold data layer for storage according to the preset default sampling frequency of key equipment, and define the trigger logic of the five-level thresholds. The trigger logic includes single-point trigger, duration trigger, change rate trigger, and comprehensive trigger; Single-point trigger means that it is triggered when the key device parameters exceed the preset warning threshold. Duration trigger means that it is triggered only when the key device parameters exceed the threshold for n periods of time continuously. Rate-of-change trigger means that the change speed of the key device parameters exceeds the normal operation threshold. Comprehensive trigger means that the number of trigger logics is greater than 1. When any one of the trigger logics in the five-level threshold is triggered, the sampling frequency is forced to switch to the preset default sampling frequency of the key device.

[0012] Preferably, the method for resource allocation through a hierarchical computing architecture includes: Define the hierarchical computing architecture as consisting of an edge layer, a fog computing layer, and a cloud computing layer. According to different types of key devices and the number of key devices, evaluate the basic load requirements of the edge layer, fog computing layer, and cloud computing layer. Create resource pools in the edge layer, fog computing layer, and cloud computing layer respectively, and divide the resource pools into core resource pools, regular resource pools, and elastic resource pools according to the basic load requirements of different layers. Among them, the core resource pool is used to process tasks corresponding to key devices, the regular resource pool is used to process tasks corresponding to non-key devices, and the elastic resource pool is used as reserved resources during the peak load. The size of the core resource pool is: ; where represents the size of the core resource pool; represents the number of key devices; the size of the regular resource pool is: ; where represents the size of the regular resource pool; represents the adjusted computing requirement according to the importance of key devices; The size of the elastic resource pool is: ; where represents the elasticity coefficient, which is used to determine the proportion of the elastic resource pool; Dynamically adjust and constrain the elasticity coefficient through the elasticity coefficient adjustment formula. The elasticity coefficient adjustment formula is: ; where represents the maximum value of the elasticity coefficient; represents the minimum value of the elasticity coefficient; represents the current resource utilization rate; represents the preset load threshold; According to the basic load requirements of the edge layer, fog computing layer, and cloud computing layer, allocate corresponding computing resources respectively. The computing resources allocated to the edge layer are equal to those allocated to the cloud computing layer, and both are less than the computing resources allocated to the fog computing layer. Logically isolate the computing resources through virtualization technology. Introduce an elastic scaling strategy, which includes horizontal expansion strategy, vertical expansion strategy, and resource recycling strategy to dynamically schedule the computing resources. Automatically adjust the resource allocation ratio according to the preset multi-level security threshold and the computing task migration mechanism. The computing task migration mechanism includes downward migration, upward migration, and horizontal migration. Downward migration means that when the status of a critical device is abnormal, the analysis tasks related to the critical device executed at the cloud data layer are migrated to the fog data layer or the edge layer. Upward migration means that when the computing resources are insufficient, the tasks corresponding to non-critical devices are migrated upward to a higher level. Horizontal migration means that within the same layer, the tasks corresponding to critical devices are migrated from the highest load node to the lowest load node.

[0013] Preferably, the method for identifying a new abnormal pattern includes: Simultaneously monitor the multi-dimensional deviations of critical device parameters. The dimensional deviations include amplitude deviation, frequency deviation, correlation deviation, and time series pattern deviation. Construct an abnormal metric index system, which includes three abnormal metric indexes: the absolute deviation of critical device parameters, the relative deviation of critical device parameters, and the Z-score. Use the DBSCAN clustering algorithm to detect abnormal data in the multi-dimensional production line data after preliminary processing at the edge, and construct an abnormal feature vector. Identify a new abnormal pattern through a dual-track template update mechanism. The dual-track template update mechanism includes a fast response track and a robust verification track. When conflicts occur between the fast response track and the verification track templates, start a conflict arbitration program, save the conflict record to a preset future abnormal template, and update the identification of the new abnormal pattern.

[0014] Preferably, the method for adjusting the detection strategy includes: The context-aware anomaly detection framework includes context modeling, policy dynamic adjustment, detection method fusion, and anomaly level classification. Context modeling includes device status context, production condition context, environmental factor context, and sensor parameter quality context. Use a sliding time window to store context information less than the preset context information threshold, and store context information greater than or equal to the preset context information threshold in a time series database. Policy dynamic adjustment includes setting detection strategy trigger conditions and adaptive thresholds. The detection strategy trigger conditions include policy adjustment based on working conditions and policy adjustment based on sensor parameter quality. Policy adjustment based on working conditions includes three policy adjustments, which are device stable operation, device load mutation, and production line environment change adjustment. Policy adjustment based on sensor parameter quality is that when the quality of sensor data deteriorates, redundant sensors are used for compensation. Adaptive threshold adjustment includes using an adaptive threshold and adjusting the adaptive threshold based on the changes in historical critical device parameters. Detection method fusion includes point anomaly detection, pattern anomaly detection, and system anomaly detection. Anomaly level classification includes low-level anomalies, medium-level anomalies, and high-level anomalies. Adjust the detection strategy according to the constructed context-aware anomaly detection framework.

[0015] Preferably, the method for performing full-coverage detection of point anomalies, pattern anomalies, and system anomalies includes: Using a deep learning model to predict and detect point data, pattern associations, and system-level anomalies in key devices; for point anomalies, combining the Z-score statistical method with LSTM prediction residual analysis to perform single-point deviation detection; for pattern anomalies, extracting multi-dimensional data patterns based on time-series models such as CNN-LSTM, and combining with the Apriori algorithm to identify anomaly patterns in any production line environment; for system anomalies, using a graph neural network to analyze the associated fault propagation between key devices to perform full-coverage detection of point anomalies, pattern anomalies, and system anomalies.

[0016] The technical effects and advantages of the real-time production line anomaly diagnosis system based on industrial Internet of Things of the present invention: Through a hierarchical computing architecture of the edge layer, fog computing layer, and cloud computing layer, reasonably allocate computing resources according to different types of key devices and the number of key devices, realize distributed processing of computing tasks, and improve system stability. Adopt a core resource pool, a regular resource pool, and an elastic resource pool to ensure that key device tasks are given priority to obtain computing resource support, and at the same time dynamically expand resources during peak loads to avoid resource bottlenecks. Improve resource utilization efficiency: In traditional resource scheduling methods, many systems may use a fixed resource pool size or a relatively rough elastic resource pool allocation strategy, which may lead to over-allocation (wasting computing resources) or resource shortage (unable to meet sudden demands). By dynamically adjusting the elastic coefficient, the system can adjust the size of the resource pool according to the real-time load situation, improve the utilization efficiency of resources, avoid resource waste, and ensure that the system performance will not decline under high load conditions.

[0017] Compared with traditional static resource allocation or single-strategy scaling methods (such as resource expansion based on fixed time intervals), the elastic coefficient adjustment based on real-time load perception can quickly respond to load fluctuations. The system can quickly increase the capacity of the resource pool when the load increases, and appropriately shrink the resource pool when the load decreases, improving the flexibility and response speed of the system, and ensuring that it can adapt to rapidly changing workloads.

[0018] For cloud computing or distributed computing systems, resources are costly. In most cloud platforms, resources are paid on demand (such as CPU, memory, storage, etc.), so if the resource allocation is unreasonable, it may lead to unnecessary cost increases. By dynamically adjusting the elastic coefficient, the system can reduce the elastic resource pool when the load is low, reducing unnecessary resource overhead. When the load increases, the system will automatically allocate more resources, which helps to optimize the cost management of cloud computing.

[0019] In the case where high-priority tasks or critical tasks require more computing resources, the dynamic expansion of the elastic resource pool can ensure that these tasks obtain sufficient resources, ensure that the tasks are completed on time, and avoid performance degradation or task delays caused by insufficient resources. Therefore, this dynamic adjustment method is particularly suitable for systems with multi-task parallel processing and task priority differential scheduling. Traditional elastic scaling strategies are usually based on fixed rules or coarser-grained adjustment methods and cannot very finely adapt to changes in system load. This method dynamically adjusts the size of the elastic resource pool based on the current utilization rate, making resource allocation more refined, being able to quickly respond when the load increases, and automatically reducing resource usage when the load decreases, thus avoiding problems of over-expansion or under-expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the structure of the production line anomaly real-time diagnosis system based on the industrial Internet of Things according to the present invention; Figure 2 Schematic diagram of the flow of the production line anomaly real-time diagnosis method based on the industrial Internet of Things according to the present invention; Figure 3 Flowchart of the method for performing full-coverage detection of point anomalies, pattern anomalies, and system anomalies provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Embodiment 1 Please refer to Figure 1 and Figure 3 , this embodiment proposes a production line anomaly real-time diagnosis system based on the industrial Internet of Things, including: An acquisition and processing module, which constructs a distributed sensor network covering key production line equipment, acquires multi-dimensional production line data, predicts the equipment status based on the multi-dimensional production line data, dynamically adjusts the multi-dimensional production line data acquisition frequency according to the equipment status using an adaptive sampling technique, and performs preliminary multi-dimensional production line data processing at the edge; A multi-scale time series management module, which adopts a three-level hierarchical storage architecture of hot, warm, and cold, combines a multi-level security threshold mechanism to forcibly switch the sampling frequency of key equipment parameters, allocates resources through a hierarchical computing architecture, introduces an anomaly-sensitive new mode detection and a dual-track template update mechanism to identify new anomaly patterns, and continuously monitors changes in equipment status; Anomaly detection module, which constructs a context-aware anomaly detection framework to adjust detection strategies, and adopts a hybrid detection method combining deep learning with detection rules to conduct full-coverage detection of point anomalies, pattern anomalies and system anomalies; Root cause analysis module, which constructs a multi-dimensional industrial knowledge graph based on the multi-dimensional production line data after preliminary processing, visualizes the fault propagation path through the multi-dimensional industrial knowledge graph, and diagnoses the root causes of different anomalies in real time; Multi-terminal collaborative interaction module, which supports access by multiple terminals, provides role-based differentiated information display and conducts intelligent interaction, and automatically adjusts the presentation method of anomaly information according to the root causes of different anomalies.

[0023] The method for constructing a distributed sensor network covering key production line equipment includes: Comprehensively investigate all key equipment on the production line to determine key equipment, which includes production equipment, auxiliary equipment, energy and power equipment, material handling and conveying equipment, detection equipment and process control equipment; Production equipment includes numerically controlled machine tools, lathes, milling machines, injection molding machines, die casting machines, extrusion machines, stamping equipment, forging equipment, presses, welding equipment (laser welding, spot welding and arc welding), heat treatment equipment (furnaces, drying equipment) and surface treatment equipment (electroplating and spraying); Auxiliary equipment includes industrial cooling equipment, vacuum equipment, waste gas treatment equipment and waste water treatment equipment; Energy and power equipment includes cooling towers, chillers, generator sets, boilers and steam equipment; Material handling and conveying equipment includes industrial robots, robotic arms, AGV automatic guided vehicles and cranes; Detection equipment includes automatic optical inspection equipment, coordinate measuring machines, X-ray inspection equipment and material analysis instruments; Process control equipment includes PLC controllers and industrial control gateway equipment. For each key equipment, determine the core parameters to be monitored, which include temperature parameters, vibration parameters, pressure parameters, current parameters, rotational speed parameters and noise parameters; Obtain historical key equipment failure data, use clustering algorithms to determine the abnormal patterns that have existed in the history of each key equipment, and obtain the corresponding core parameter change data; Based on the core parameter change data and the abnormal patterns that have existed in the history of each key equipment, plan the optimal monitoring points for each key equipment according to the preset evaluation criteria. The preset evaluation criteria include the strongest signal strength standard, the shortest installation time standard, the shortest maintenance time standard and the minimum interference to the equipment standard; Select the corresponding sensor types (temperature, vibration, acoustics, electricity and rotational speed) according to the monitored core parameters, and implement them in three installation methods, which include magnetic installation, adhesive fixed installation and mechanical fixed installation.

[0024] Multi-dimensional production line data includes equipment physical state data, electrical parameter data, process parameter data, product information data, and production line environment data; Equipment physical state data includes equipment surface temperature, internal temperature, bearing temperature, equipment amplitude, equipment vibration speed, equipment noise level, equipment abnormal sound, hydraulic equipment pressure, pneumatic equipment pressure, coolant flow rate, lubricating oil flow rate, and process gas flow rate; Electrical parameter data includes motor current, supply voltage, active power, reactive power, insulation resistance, power grid harmonics, relay switch status, and contactor switch status; Process parameter data includes production line speed, unit time output, processing temperature, cooling temperature, forming pressure, extrusion pressure, injection pressure, raw material density, viscosity, cutting speed, feed rate, and processing depth; Product information data includes product dimensions and product appearance images; production line environment data includes production line temperature, production line humidity, and electromagnetic interference intensity.

[0025] The method for dynamically adjusting the multi-dimensional production line data acquisition frequency includes: The acquisition and processing module constructs a distributed sensor network covering key equipment on the production line, acquires multi-dimensional production line data, and uses adaptive sampling technology to dynamically adjust the multi-dimensional production line data acquisition frequency according to the equipment status, and performs preliminary multi-dimensional production line data processing at the edge; Construct an equipment status prediction model, use historical multi-dimensional production line data as the input of the equipment status prediction model, and output three equipment statuses. The equipment statuses include stable status, transition status, and abnormal status. The equipment status prediction model is a fully connected neural network model; the stable status indicates that the equipment is running smoothly without abnormal changes. The transition status indicates that the fluctuation in the change stage of the equipment from stable to abnormal is greater than or equal to the preset fluctuation threshold for the change stage of the equipment from stable to abnormal. The abnormal status indicates that the equipment has a fault or anomaly.

[0026] Generate a dynamic adjustment mechanism according to the three equipment statuses, preset a reference acquisition frequency and a high acquisition frequency. If the equipment status is in the stable status, reduce the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the equipment status is in the transition status, increase the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the equipment status is in the abnormal status, trigger an event-driven mechanism and increase the multi-dimensional production line data to the high acquisition frequency.

[0027] The method for performing preliminary multi-dimensional production line data processing at the edge includes: At the edge, identify and remove outliers in the multi-dimensional production line data based on the box plot method, and fill in missing values in the multi-dimensional production line data by interpolation; denoise the multi-dimensional production line data after filling in the missing values by mean filtering, and perform timestamp alignment and standard deviation normalization processing on the denoised multi-dimensional production line data.

[0028] The method for forcibly switching the sampling frequency of key device parameters includes: The three - level hierarchical storage architecture of hot, warm, and cold includes a hot data layer, a warm data layer, and a cold data layer. The preset multi - level security thresholds include five - level thresholds: normal operation threshold, warning threshold, alarm threshold, danger threshold, and emergency threshold. The normal operation threshold indicates that the key device parameters are within the normal working range. The warning threshold indicates that the key device parameters start to deviate from the normal working range and attention needs to be paid. The alarm threshold indicates that the key device parameters are significantly deviated from the normal working range and may affect the device performance. The danger threshold indicates that the key device parameters are close to the limit value of deviation from the normal working range and there is a risk of device damage. The emergency threshold indicates that the key device parameters exceed the normal working range and may lead to system crashes or safety accidents. Preset the default sampling frequency of key devices. Allocate key device parameters to the hot data layer, warm data layer, and cold data layer for storage according to the preset default sampling frequency of key devices. Define the trigger logics for the five - level thresholds. The trigger logics include single - point trigger, duration trigger, rate - of - change trigger, and comprehensive trigger. The single - point trigger means that when the key device parameters exceed the preset warning threshold, it is triggered. The duration trigger means that the key device parameters need to exceed the threshold for n periods of time to be triggered. The rate - of - change trigger means that the change speed of the key device parameters exceeds the normal operation threshold. The comprehensive trigger means that the number of trigger logics is greater than 1. When any one of the trigger logics in the five - level thresholds is triggered, the sampling frequency is forcibly switched to the preset default sampling frequency of key devices.

[0029] The method for resource allocation through a hierarchical computing architecture includes: Define the hierarchical computing architecture as consisting of an edge layer, a fog computing layer, and a cloud computing layer. The edge layer is directly deployed on computing nodes near production equipment, including edge servers, industrial PCs, programmable logic controllers (PLCs), and intelligent gateways, etc., and is mainly responsible for real - time data collection and preliminary processing. Edge - layer devices usually have 4 - 16 - core CPUs, 4 - 32GB of memory, and 128GB - 1TB of storage space.

[0030] The fog computing layer is deployed on medium - scale computing nodes at the workshop or regional level, including factory server clusters, industrial control centers, etc. Fog - layer devices usually have 16 - 64 - core CPUs, 64 - 256GB of memory, and several TB of storage space, and support multi - thread parallel computing and medium - scale data processing.

[0031] The cloud computing layer represents the industrial cloud platform, which has a large - scale server cluster, a distributed storage system, and high - performance computing resources. The cloud layer can provide computing power of several hundred to thousands of cores, TB - level memory, and PB - level storage space, and support large - scale data analysis and complex algorithm operations.

[0032] Evaluate the basic load requirements of the edge layer, fog computing layer, and cloud computing layer based on different types of key devices and the number of key devices. For each type of key device , evaluate its basic computing load requirements , representing the computing resources required by the key device during normal operation; calculate the total basic load requirements: ; where represents the total basic load, indicating the total computing requirements of all devices; represents the number of the th key device; represents the number of types of key devices; Use the importance weight of the key device to adjust the allocated load: ; where represents the adjusted computing requirements based on the importance of the key device; for example, a typical automobile assembly line may require the edge layer to process 5,000 - 10,000 data points per second, and the fog layer to aggregate and process 50,000 - 100,000 data points per second.

[0033] Create resource pools in the edge layer, fog computing layer, and cloud computing layer respectively, and divide the resource pools into core resource pools, regular resource pools, and elastic resource pools according to the basic load requirements of different layers; among them, the core resource pool is used to process tasks corresponding to key devices, the regular resource pool is used to process tasks corresponding to non - key devices, and the elastic resource pool is used as reserved resources during peak loads; The size of the core resource pool is: ; where represents the size of the core resource pool; represents the number of key devices; the size of the regular resource pool is: ; where represents the size of the regular resource pool, that is, the total computing resources required by non - key devices; The size of the elastic resource pool is: ; where represents the size of the elastic resource pool; represents the elasticity coefficient, which is used to determine the proportion of the elastic resource pool. According to the expert experience method, ; However, in the existing technology, the load and resource requirements are dynamically changing. For example, the computing requirements are high during some periods and low during others, or the system cannot respond quickly to sudden loads, resulting in resource waste or performance degradation. By adjusting the elasticity coefficient, the size of the elastic resource pool can be dynamically adjusted to ensure that the system can quickly respond to load fluctuations, allocate more resources during peak loads, save resources during low loads, and avoid waste.

[0034] Dynamically adjust and constrain the elasticity coefficient through the elasticity coefficient adjustment formula, and the elasticity coefficient adjustment formula is as follows: where represents the maximum value of the elasticity coefficient; represents the minimum value of the elasticity coefficient; represents the current resource utilization rate (such as CPU usage or memory usage), reflecting the current load situation; represents the preset load threshold, indicating that the dynamic adjustment of the elastic resource pool is triggered under this load; It should be noted that the design concept of this elasticity coefficient adjustment formula is based on load-aware dynamic resource adjustment, that is, the system dynamically adjusts the size of the elastic resource pool according to the current load situation ( ). When the load is high, the size of the elastic resource pool should be increased to ensure that the computing requirements of the system are met; when the load is low, the size of the elastic resource pool should be reduced to avoid resource waste.

[0035] The formula uses a linear function to adjust the elasticity coefficient. When the resource utilization rate increases, the elasticity coefficient will also increase accordingly, so that the elastic resource pool increases and the system can obtain more resources. Conversely, when the load is low, the elasticity coefficient decreases, reducing the scale of the elastic resource pool and saving resources.

[0036] When , the elasticity coefficient , that is, the system has no additional elastic resources and is in the lowest resource state.

[0037] When reaches the maximum load (close to 100%), the elasticity coefficient reaches the maximum value , and at this time the scale of the elastic resource pool is the largest, and the system can make full use of resources to cope with the load peak.

[0038] is a key parameter that determines the trigger point for the adjustment of the elastic resource pool. When the load is lower than this threshold, the elastic resource pool remains small to avoid over-allocation of resources; when the load exceeds this threshold, the elastic resource pool begins to gradually increase, thereby improving the computing power of the system. This threshold-based adjustment method helps to avoid resource waste and at the same time enhance the system's ability to cope with sudden increases in load.

[0039] Beneficial effects compared with the prior art: Improve resource utilization efficiency: In traditional resource scheduling methods, many systems may use a fixed resource pool size or a relatively rough elastic resource pool allocation strategy, which may lead to over-allocation (wasting computing resources) or resource shortage (unable to meet sudden demands). By dynamically adjusting the elasticity coefficient, the system can adjust the size of the resource pool according to the real-time load conditions, improve the utilization efficiency of resources, avoid resource waste and ensure that the system performance will not decline under high load conditions.

[0040] Respond to load fluctuations in real time: Compared with traditional static resource allocation or scaling methods with a single strategy (such as resource expansion based on fixed time intervals), the adjustment of the elasticity coefficient based on real-time load perception can quickly respond to load fluctuations. The system can quickly increase the capacity of the resource pool when the load increases, and appropriately shrink the resource pool when the load decreases, improving the flexibility and response speed of the system, and ensuring that it can adapt to rapidly changing workloads.

[0041] Optimize cost management: For cloud computing or distributed computing systems, resources have costs. In most cloud platforms, resources are paid on demand (such as CPU, memory, storage, etc.). Therefore, if the resource allocation is unreasonable, it may lead to unnecessary cost increases. By dynamically adjusting the elasticity coefficient, the system can reduce the elastic resource pool when the load is low, reducing unnecessary resource overheads. When the load increases, the system will automatically allocate more resources, which helps to optimize the cost management of cloud computing.

[0042] Guarantee task performance: In the case where high-priority tasks or critical tasks require more computing resources, the dynamic expansion of the elastic resource pool can guarantee that these tasks obtain sufficient resources, ensure that the tasks are completed on time, and avoid performance degradation or task delays caused by resource shortages. Therefore, this dynamic adjustment method is particularly suitable for systems with multi-task parallel processing and task priority differential scheduling.

[0043] Avoid over- or under-allocation of resources: Traditional elastic scaling strategies usually rely on fixed rules or relatively coarse-grained adjustment methods and cannot very finely adapt to changes in system load. This method dynamically adjusts the size of the elastic resource pool based on the current utilization rate, making resource allocation more refined, being able to quickly respond when the load increases, and automatically reducing resource usage when the load decreases, thus avoiding the problems of over-expansion or under-expansion.

[0044] According to the basic load requirements of the edge layer, fog computing layer, and cloud computing layer, corresponding computing resources are allocated respectively. The computing resources allocated to the edge layer are equal to those allocated to the cloud computing layer, and both are less than the computing resources allocated to the fog computing layer. Through virtualization technology, logical isolation of computing resources is achieved to ensure that critical production tasks are not interfered with by other tasks. For example, a separate resource space is divided for spindle temperature monitoring, and an independent CPU core and memory area are configured.

[0045] An elastic scaling strategy is introduced. The elastic scaling strategy includes horizontal scaling, vertical scaling, and resource recycling strategies for dynamic scheduling of computing resources. Horizontal scaling strategy: When the node load exceeds the warning line and lasts for more than 15 minutes, the system automatically starts additional computing nodes within the same layer. For example, when the CPU utilization rate of the edge server continuously exceeds 75%, the standby edge server is activated to share the load.

[0046] Vertical scaling strategy: When the resources of a single node are approaching the bottleneck but cannot be horizontally scaled, the available resource quota of this node is dynamically increased. For example, the available memory of a certain fog computing node is temporarily increased from 64GB to 96GB.

[0047] Resource recycling strategy: When the node load continuously remains below 40% for more than 30 minutes, the system gradually recovers excess resources or shuts down redundant nodes, releasing resources to the common pool.

[0048] The resource allocation ratio is automatically adjusted according to the preset multi-level security thresholds and the computing task migration mechanism. The computing task migration mechanism includes downward migration, upward migration, and horizontal migration. Downward migration means that when the status of a critical device is abnormal, the analysis tasks related to this critical device executed at the cloud data layer are migrated to the fog data layer or the edge layer; reducing network latency and improving response speed. For example, the bearing vibration analysis task is migrated from the cloud to the edge server for direct processing. Upward migration means that when computing resources are insufficient, the tasks corresponding to non-critical devices are migrated upward to a higher level; for example, when multiple devices are abnormal simultaneously, the historical data processing task is migrated from the fog layer to the cloud layer, releasing fog layer resources for anomaly handling. Horizontal migration means that within the same layer, the tasks corresponding to critical devices are migrated from the highest load node to the lowest load node; for example, the non-critical tasks on a certain fog computing node are migrated to a node with lower load in the same layer.

[0049] The methods for identifying new abnormal patterns include: Monitor deviations in multiple dimensions of key equipment parameters simultaneously. The dimensional deviations include amplitude deviation (the degree to which the key equipment parameter value exceeds the normal range), frequency deviation (abnormal change frequency of the key equipment parameter), correlation deviation (abnormal relationship between multiple parameters of the key equipment), and time-series pattern deviation (abnormal time-series pattern); construct an abnormal metric index system, which includes three abnormal metric indexes: absolute deviation of key equipment parameters, relative deviation of key equipment parameters, and Z-score; use the DBSCAN clustering algorithm to detect abnormal data in the multi-dimensional production line data after preliminary processing at the edge side, and construct an abnormal feature vector. Identify new abnormal patterns through a two-track template update mechanism. The two-track template update mechanism includes a fast response track and a robust verification track. The identification method of the fast response track includes: Initial detection stage: The system detects an abnormal pattern that cannot match the known abnormal template, calculates its abnormal score. When the abnormal score exceeds the preset threshold (usually 0.85) and the duration exceeds the preset window (such as 5 consecutive minutes), trigger the new pattern candidate confirmation process. The system automatically extracts the feature vector and context information of this abnormality. Fast template generation: Based on the detected new abnormal features, the system automatically generates a temporary abnormal template. The temporary abnormal template contains abnormal feature descriptions, detection conditions, and credibility scores; assign a unique identifier and version number to the temporary abnormal template, and record the generation time and conditions. Temporary deployment and monitoring: The temporary abnormal template is deployed to the template library of the fast response track and starts to participate in real-time abnormal detection. The system sets a higher false alarm tolerance for the temporary abnormal template to avoid excessive interference with production. At the same time, start monitoring the performance of the abnormal template, and record indicators such as matching situations, accuracy rates, and recall rates. Feedback collection and adjustment: The system automatically collects the detection results and operation data of the temporary abnormal template, dynamically adjusts the parameters of the temporary template based on real-time feedback, and optimizes the detection performance; collect the confirmation or negation feedback from on-site operators on the detection results. The identification method of the robust verification track includes: Strict verification stage: Submit the temporary template generated by the fast response track to the verification environment, use historical data and simulation data to backtest the template, and evaluate its reliability; conduct cross-verification to ensure the stability of the template under different working conditions. Expert review process: The system automatically generates an analysis report of the new pattern and submits it to domain experts. The expert team evaluates the physical rationality and industrial significance of the new pattern, and modifies and improves the template according to the expert opinions. Formal abnormal template construction: Based on the verification results and expert opinions, construct a formal abnormal template, add detailed semantic explanations and processing suggestions to the template, and set the applicable conditions and reliability levels of the formal abnormal template. Template Knowledge Base Integration: Integrate the formally verified exception templates into the global exception template library, establish the association relationship between the new exception templates and the existing exception templates, update the exception knowledge graph, and reflect the newly identified exception patterns; When conflicts occur between the quick response track and the verification track templates, initiate the conflict arbitration process, save the conflict records to the preset future exception templates, and update the identification of new exception patterns.

[0050] The methods for adjusting the detection strategy include: The context-aware anomaly detection framework includes context modeling, dynamic policy adjustment, detection method fusion, and anomaly level classification. Context modeling includes device status context, production condition context, environmental factor context, and sensor parameter quality context; Use a sliding time window to store context information smaller than the preset context information threshold, and store context information greater than or equal to the preset context information threshold in the time series database; Dynamic policy adjustment includes setting detection strategy trigger conditions and adaptive thresholds. The detection strategy trigger conditions include policy adjustment based on working conditions and policy adjustment based on sensor parameter quality; Policy adjustment based on working conditions includes three types of policy adjustments. The three types of policy adjustments include equipment stable operation, sudden change in equipment load, and production line environment change adjustment; The specific adjustment methods are: Equipment stable operation adjustment: The equipment stable operation strategy mainly focuses on the performance of the equipment under normal working conditions. By monitoring the equipment operation status (such as temperature, vibration, running speed, etc.), if the equipment is in a normal and stable working state, the detection strategy should maintain a low sensitivity to avoid false alarms.

[0051] Adjustment method: Under the condition of stable operation of the equipment, the detection threshold and alarm conditions can be appropriately relaxed to reduce unnecessary alarms and consumption of system resources. If fluctuations occur in the equipment (such as sudden temperature changes), a more sensitive detection strategy should be initiated.

[0052] Sudden change in equipment load adjustment: When the equipment load suddenly changes (such as startup, shutdown, or load overload, etc.), it will cause fluctuations in equipment performance or abnormal behaviors. In this case, the original detection strategy may no longer be applicable, so rapid adjustments are needed.

[0053] Adjustment method: Once the load changes suddenly, the detection strategy should immediately increase the sensitivity and strengthen the detection of load anomalies and equipment failures. This may include quickly collecting and analyzing more frequent data, lowering the threshold to capture more subtle abnormal signals, and even enabling backup sensors for cross-verification.

[0054] Production line environment change adjustment: The production line environment (such as temperature, humidity, air pressure, light, etc.) has a significant impact on the operation of the equipment. If the environment changes, it may lead to performance changes or errors in the equipment, so the detection strategy needs to be adjusted.

[0055] Adjustment method: When the production environment changes (such as a temperature rise or humidity increase), the detection strategy should reset the thresholds and standards according to these environmental changes. This may include dynamically adjusting the detection thresholds, enhancing the anomaly detection model related to the environment, and reducing the interference caused by environmental changes through multi-sensor fusion.

[0056] The strategy adjustment based on the quality of sensor parameters is that when the quality of sensor data deteriorates (such as data drift), redundant sensors are used for compensation; the adaptive threshold adjustment includes using an adaptive threshold and adjusting the adaptive threshold based on the changes in historical key equipment parameters; The fusion of detection methods includes point anomaly detection, pattern anomaly detection, and system anomaly detection; the classification of anomaly levels includes low-level anomalies, medium-level anomalies, and high-level anomalies; the method of classifying anomaly levels includes obtaining the deviation values between all abnormal key equipment and normal key equipment through sensors, presetting the first deviation threshold and the second deviation threshold. If the deviation value between the abnormal key equipment and the normal key equipment is less than or equal to the preset first deviation threshold, the anomaly level is classified as a low-level anomaly; if the deviation value between the abnormal key equipment and the normal key equipment is greater than the preset first deviation threshold and less than the preset second deviation threshold, the anomaly level is classified as a medium-level anomaly; if the deviation value between the abnormal key equipment and the normal key equipment is greater than or equal to the preset second deviation threshold, the anomaly level is classified as a high-level anomaly; the detection strategy is adjusted according to the constructed context-aware anomaly detection framework.

[0057] The methods for conducting full-coverage detection of point anomalies, pattern anomalies, and system anomalies include:

[0058] Using a deep learning model to predict and detect point data, pattern associations, and system-level anomalies in key equipment; for point anomalies, combining the Z-score statistical method with LSTM prediction residual analysis for single-point deviation detection; for pattern anomalies, extracting multi-dimensional data patterns based on time-series models such as CNN-LSTM, and combining with the Apriori algorithm to identify anomaly patterns in any production line environment; for system anomalies, using graph neural networks to analyze the associated fault propagation between key equipment for full-coverage detection of point anomalies, pattern anomalies, and system anomalies.

[0059] In this embodiment, through the hierarchical computing architecture of the edge layer, fog computing layer, and cloud computing layer, computing resources are reasonably allocated according to different types of key devices and the number of key devices, realizing the distributed processing of computing tasks and improving system stability. The core resource pool, conventional resource pool, and elastic resource pool are adopted to ensure that the tasks of key devices are given priority to obtain computing resource support, and at the same time, resources are dynamically expanded during peak loads to avoid resource bottlenecks. Improve resource utilization efficiency: In traditional resource scheduling methods, many systems may use a fixed resource pool size or a relatively rough elastic resource pool allocation strategy, which may lead to over-allocation (wasting computing resources) or resource shortage (unable to meet sudden demands). By dynamically adjusting the elasticity coefficient, the system can adjust the size of the resource pool according to the real-time load conditions, improve the utilization efficiency of resources, avoid resource waste, and ensure that the system performance will not decline under high load conditions.

[0060] Compared with traditional static resource allocation or single-strategy scaling methods (such as resource expansion based on fixed time intervals), the adjustment of the elasticity coefficient based on real-time load perception can quickly respond to load fluctuations. The system can quickly increase the capacity of the resource pool when the load increases, and appropriately shrink the resource pool when the load decreases, improving the flexibility and response speed of the system and ensuring that it can adapt to rapidly changing workloads.

[0061] For cloud computing or distributed computing systems, resources are costly. In most cloud platforms, resources are paid on demand (such as CPU, memory, storage, etc.). Therefore, if the resource allocation is unreasonable, it may lead to unnecessary cost increases. By dynamically adjusting the elasticity coefficient, the system can reduce the elastic resource pool when the load is low, reducing unnecessary resource overhead. When the load increases, the system will automatically allocate more resources, which helps to optimize the cost management of cloud computing.

[0062] In the case where high-priority tasks or critical tasks require more computing resources, the dynamic expansion of the elastic resource pool can ensure that these tasks obtain sufficient resources, ensure that the tasks are completed on time, and avoid performance degradation or task delay caused by resource shortage. Therefore, this dynamic adjustment method is particularly suitable for systems with multi-task parallel processing and task priority differential scheduling. Traditional elastic scaling strategies are usually based on fixed rules or relatively coarse-grained adjustment methods and cannot very finely adapt to the changes in system load. This method dynamically adjusts the size of the elastic resource pool based on the current utilization rate, making the resource allocation more refined, being able to quickly respond when the load increases, and automatically reducing resource usage when the load decreases, thus avoiding the problems of over-expansion or under-expansion.

[0063] Embodiment 2 Please refer to Figure 2As shown in the figure, for the parts not described in detail in this embodiment, refer to the description in Embodiment 1. A collaborative management method for an Internet of Things gateway based on edge computing is provided, including: S1. Construct a distributed sensor network covering key production line equipment, collect multi-dimensional production line data, predict the equipment status based on the multi-dimensional production line data, dynamically adjust the multi-dimensional production line data collection frequency according to the equipment status using an adaptive sampling technique, and perform preliminary multi-dimensional production line data processing at the edge side; S2. Adopt a three-level hierarchical storage architecture of hot, warm, and cold, combine a multi-level security threshold mechanism to forcibly switch the sampling frequency of key equipment parameters, allocate resources through a hierarchical computing architecture, introduce an anomaly-sensitive new mode detection and a dual-track template update mechanism to identify new anomaly modes, and continuously monitor changes in equipment status; S3. Construct a context-aware anomaly detection framework to adjust the detection strategy, and adopt a hybrid detection method combining deep learning and detection rules to perform full-coverage detection of point anomalies, pattern anomalies, and system anomalies; S4. Construct a multi-dimensional industrial knowledge graph based on the preliminarily processed multi-dimensional production line data, visually display the fault propagation path through the multi-dimensional industrial knowledge graph, and diagnose the root causes of different anomalies in real time; S5. Support access by multiple terminals, provide role-based differential information display, and perform intelligent interaction, and automatically adjust the anomaly information presentation method according to the root causes of different anomalies.

[0064] Since the electronic device introduced in this embodiment is the electronic device adopted by the production line anomaly real-time diagnosis system based on the industrial Internet of Things in the embodiments of the present application, based on the production line anomaly real-time diagnosis system based on the industrial Internet of Things introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manners and various change forms of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device adopted by the production line anomaly real-time diagnosis system based on the industrial Internet of Things in the embodiments of the present application, it falls within the protection scope of the present application.

[0065] The above formulas are all dimensionless and only take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0066] The above are only the preferred implementation manners of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for ordinary technical users in the technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A real-time diagnosis system for production line anomalies based on the Industrial Internet of Things, characterized in that, Including: The acquisition and processing module constructs a distributed sensor network covering the key equipment on the production line, acquires multi-dimensional production line data, predicts the equipment status based on the multi-dimensional production line data, dynamically adjusts the multi-dimensional production line data acquisition frequency according to the equipment status by using the adaptive sampling technology, and performs preliminary multi-dimensional production line data processing at the edge side; The multi-scale time series management module adopts a three-level hierarchical storage architecture of hot, warm, and cold, combines a multi-level security threshold mechanism to forcibly switch the sampling frequency of key equipment parameters, allocates resources through a hierarchical computing architecture, introduces an anomaly-sensitive new mode detection and a two-track template update mechanism to identify new anomaly modes, and continuously monitors the changes in equipment status; The anomaly detection module constructs a context-aware anomaly detection framework to adjust the detection strategy, and adopts a hybrid detection method combining deep learning and detection rules to perform full-coverage detection of point anomalies, pattern anomalies, and system anomalies; The root cause analysis module constructs a multi-dimensional industrial knowledge graph based on the preliminarily processed multi-dimensional production line data, visualizes the fault propagation path through the multi-dimensional industrial knowledge graph, and diagnoses the root cause of different anomalies in real time; The multi-terminal collaborative interaction module supports access by multiple terminals, provides role-based differential information display and performs intelligent interaction, and automatically adjusts the presentation method of anomaly information according to the root cause of different anomalies.

2. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 1, wherein The method for constructing the distributed sensor network covering the key equipment on the production line includes: Comprehensively investigate all key equipment on the production line to determine the key equipment, which includes production equipment, auxiliary equipment, energy and power equipment, material handling and conveying equipment, detection equipment, and process control equipment; For each key equipment, determine the core parameters to be monitored, which include temperature parameters, vibration parameters, pressure parameters, current parameters, rotational speed parameters, and noise parameters; obtain the historical key equipment failure data, use the clustering algorithm to determine the anomaly modes that each key equipment has historically, and obtain the corresponding core parameter change data; based on the core parameter change data and the historically existing anomaly modes of each key equipment, plan the optimal monitoring points for each key equipment according to the preset evaluation criteria; The preset evaluation criteria include the strongest signal intensity standard, the shortest installation time standard, the shortest maintenance time standard, and the minimum interference to the equipment standard; select the corresponding sensor type according to the monitored core parameters, and implement classification using three installation methods, which include magnetic installation, adhesive fixation installation, and mechanical fixation installation.

3. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 2, wherein, The multi-dimensional production line data includes equipment physical state data, electrical parameter data, process parameter data, product information data, and production line environment data.

4. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 3, characterized in that, The method for dynamically adjusting the multi-dimensional production line data acquisition frequency includes: Construct an equipment status prediction model, use the historical multi-dimensional production line data as the input of the equipment status prediction model, and output three equipment statuses, which include stable status, transition status, and abnormal status. The equipment status prediction model is a fully connected neural network model; Generate a dynamic adjustment mechanism according to three device states, preset a reference acquisition frequency and a high acquisition frequency. If the device state is in a stable state, reduce the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the device state is in a transition state, increase the multi-dimensional production line data acquisition frequency to the reference acquisition frequency; if the device state is in an abnormal state, trigger an event-driven mechanism and increase the multi-dimensional production line data to the high acquisition frequency.

5. The real-time diagnosis system for production line anomalies based on industrial Internet of Things according to claim 4, characterized in that, The method for performing preliminary multi-dimensional production line data processing at the edge side includes: At the edge side, identify and remove outliers in the multi-dimensional production line data based on the box plot method, and fill in missing values in the multi-dimensional production line data by interpolation; denoise the multi-dimensional production line data after filling in the missing values by mean filtering, and perform timestamp alignment and standard deviation normalization processing on the denoised multi-dimensional production line data.

6. The real-time production line anomaly diagnosis system based on industrial Internet of Things according to claim 5, characterized in that, The method for forcibly switching the sampling frequency of key device parameters includes: The hot, warm, and cold three-level hierarchical storage architecture includes a hot data layer, a warm data layer, and a cold data layer, and the preset multi-level security thresholds include five-level thresholds: normal operation threshold, early warning threshold, warning threshold, danger threshold, and emergency threshold; Preset the default sampling frequency of key devices, allocate key device parameters to the hot data layer, the warm data layer, and the cold data layer for storage according to the preset default sampling frequency of key devices, define the trigger logics for the five-level thresholds, and the trigger logics include single-point trigger, duration trigger, change rate trigger, and comprehensive trigger; Single-point trigger means that when the key device parameter exceeds the preset early warning threshold, it is triggered. Duration trigger means that the key device parameter needs to exceed the threshold for n consecutive time periods to be triggered; Change rate trigger means that the change speed of the key device parameter exceeds the normal operation threshold; Comprehensive trigger means that the number of trigger logics is greater than 1; When any one of the trigger logics in the five-level thresholds is triggered, the sampling frequency is forcibly switched to the preset default sampling frequency of key devices.

7. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 6, characterized in that The method for resource allocation through a hierarchical computing architecture includes: Define the hierarchical computing architecture as consisting of an edge layer, a fog computing layer, and a cloud computing layer. According to different types of key devices and the number of key devices, evaluate the basic load requirements of the edge layer, the fog computing layer, and the cloud computing layer, create resource pools in the edge layer, the fog computing layer, and the cloud computing layer respectively, and divide the resource pools into core resource pools, regular resource pools, and elastic resource pools according to the basic load requirements of different layers; among them, the core resource pool is used to process tasks corresponding to key devices, the regular resource pool is used to process tasks corresponding to non-key devices, and the elastic resource pool is used as reserved resources during the peak load; The size of the core resource pool is: ; where represents the size of the core resource pool; represents the number of key devices; The size of the regular resource pool is: ; where represents the size of the regular resource pool; represents the calculated requirements adjusted according to the importance of key devices; The size of the elastic resource pool is: ; where represents the size of the elastic resource pool; represents the elasticity coefficient, which is used to determine the proportion of the elastic resource pool; Dynamically adjust and constrain the elasticity coefficient through the elasticity coefficient adjustment formula, and the elasticity coefficient adjustment formula is: ; where represents the maximum value of the elasticity coefficient; represents the minimum value of the elasticity coefficient; represents the current resource utilization rate; represents the preset load threshold; According to the basic load requirements of the edge layer, the fog computing layer, and the cloud computing layer, allocate corresponding computing resources respectively. The computing resources allocated to the edge layer are equal to the computing resources allocated to the cloud computing layer, and both are less than the computing resources allocated to the fog computing layer; perform logical isolation on the computing resources through virtualization technology, Introduce an elastic scaling strategy, and the elastic scaling strategy includes horizontal expansion strategy, vertical expansion strategy, and resource recycling strategy to perform dynamic scheduling on the computing resources; Automatically adjust the resource allocation ratio according to the preset multi-level security threshold and the computing task migration mechanism, where the computing task migration mechanism includes downward migration, upward migration, and horizontal migration; downward migration means that when the key device status is abnormal, the analysis tasks related to the key device executed at the cloud data layer are migrated to the fog data layer or the edge layer; upward migration means that when the computing resources are insufficient, the tasks corresponding to non-critical devices are migrated upward to a higher level; horizontal migration means that within the same layer, the tasks corresponding to the key device are migrated from the highest load node to the lowest load node.

8. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 7, wherein, The method for identifying new abnormal patterns includes: Simultaneously monitor the multi-dimensional deviations of key device parameters, where the dimensional deviations include amplitude deviation, frequency deviation, correlation deviation, and time series pattern deviation; construct an abnormal metric index system, which includes three abnormal metric indexes: the absolute deviation of key device parameters, the relative deviation of key device parameters, and the Z-score; use the DBSCAN clustering algorithm to detect abnormal data in the multi-dimensional production line data after preliminary processing at the edge, and construct an abnormal feature vector. Identify new abnormal patterns through a dual-track template update mechanism, which includes a fast response track and a robust verification track. When conflicts occur between the templates of the fast response track and the verification track, start a conflict arbitration program, save the conflict record to the preset future abnormal template, and update the identification of new abnormal patterns.

9. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 8, characterized in that, The method for adjusting the detection strategy includes: The context-aware anomaly detection framework includes context modeling, policy dynamic adjustment, detection method fusion, and anomaly level classification. Context modeling includes device status context, production condition context, environmental factor context, and sensor parameter quality context. Adopt a sliding time window to store context information less than the preset context information threshold, and store context information greater than or equal to the preset context information threshold in the time series database. Policy dynamic adjustment includes setting detection strategy trigger conditions and adaptive thresholds. The detection strategy trigger conditions include policy adjustment based on working conditions and policy adjustment based on sensor parameter quality; policy adjustment based on working conditions includes three policy adjustments, which are device stable operation, device load mutation, and production line environment change adjustment; policy adjustment based on sensor parameter quality means that when the sensor data quality deteriorates, redundant sensors are used for compensation; adaptive threshold adjustment includes using an adaptive threshold and adjusting the adaptive threshold based on the changes in historical key device parameters. Detection method fusion includes point anomaly detection, pattern anomaly detection, and system anomaly detection; anomaly level classification includes low-level anomalies, medium-level anomalies, and high-level anomalies; adjust the detection strategy according to the constructed context-aware anomaly detection framework.

10. The real-time production line anomaly diagnosis system based on the industrial Internet of Things according to claim 9, characterized in that, The method for performing full-coverage detection of point anomalies, pattern anomalies, and system anomalies includes: Use deep learning models to predict and detect point data, pattern associations, and system-level anomalies in critical equipment; for point anomalies, combine the Z-score statistical method with LSTM prediction residual analysis to perform single-point deviation detection; for pattern anomalies, extract multi-dimensional data patterns based on time series models such as CNN-LSTM, and combine with the Apriori algorithm to identify abnormal patterns in any production line environment; for system anomalies, use graph neural networks to analyze the associated fault propagation between critical equipment to perform full-coverage detection of point anomalies, pattern anomalies, and system anomalies.

Images