Analysis and diagnosis system for industrial big data acquisition

Abstract

The invention discloses an analysis and diagnosis system for industrial big data acquisition, belongs to the technical field of industrial big data analysis and diagnosis, and is used for solving the problems that an existing monitoring system generally can only detect parameter threshold over-limit alarms in a working procedure, causes of an upstream working procedure after abnormity are difficult to find, and time delay often exists due to cross-working-procedure influence. A cross-process directed causal association map is constructed, multi-process data such as ironmaking, continuous casting and steel rolling are integrated, a multi-layer attention network and physical constraint verification are combined, a node influence weight and an abnormal deviation degree are integrated, root cause confidence is calculated, a fault source is accurately positioned, a false correlation link is eliminated, the judgment credibility is improved, and the fault diagnosis accuracy is improved. Progressive production reduction or emergency production stop is adopted according to root cause confidence, safety and efficiency are balanced, yield loss is reduced, and a preset attention head ensures that all-channel abnormity monitoring is free of dead angles.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial big data analysis and diagnosis, more particularly, the present application relates to an analysis and diagnosis system for industrial big data collection. BACKGROUND

[0002] In large-scale industrial production processes, each process is usually undertaken by different devices and production lines, and traditional monitoring and diagnosis systems are often independent of each other and limited to data monitoring of a single process or a single device.

[0003] The prior art has the following disadvantages:

[0004] Existing monitoring systems can usually only detect parameter threshold overrun alarms within the process, and it is difficult to find the causes of abnormalities in the more upstream processes, and there is often a time delay in cross-process effects. Modern factories have accumulated a large amount of sensor data, but existing diagnostic methods are mainly based on single-point threshold alarms or simple correlation analysis, and industrial big data has not been fully utilized to mine complex multivariate correlation patterns.

[0005] In view of the above problems, the industrial field urgently needs a diagnosis system that can cross-process correlation analysis of industrial big data, which should be able to automatically identify the root cause link across production lines, consider the time delay and physical connection between signals, provide a reliable causal relationship judgment, and also have intelligent decision-making ability to take graded control measures according to the severity of the anomaly, minimize downtime and production loss under the premise of ensuring safety. The present application provides a solution to this problem. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an analysis and diagnosis system for industrial big data collection to solve the problems raised in the background art.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] An analysis and diagnosis system for industrial big data collection, comprising: a causal correlation graph construction module, a root cause positioning and link optimization module, a hierarchical intervention and dynamic control module, and a channel optimization and closed-loop correction module, the modules are connected by signals;

[0009] The causal correlation graph construction module integrates key signals of ironmaking, continuous casting and rolling processes to form an integrated data set across processes, uses time series similarity analysis and cross-correlation calculation to quantify the similarity and time delay between signals, and constructs a directed causal correlation graph, presets attention heads for various alarm detection devices, and defines the main channel and parallel channel of medium transmission and the corresponding dynamic coefficient;

[0010] The root cause positioning and link optimization module activates the corresponding attention head when the alarm detection device triggers an alarm, traces back the cause-effect correlation graph, filters out the local cause-effect edge sequence with high similarity to the alarm device and shared medium, calculates the root cause confidence of each node through a multi-layer attention network, identifies the most likely root cause, performs physical constraint verification on the identified cause-effect chain, and generates a root cause chain that meets the process constraints;

[0011] The hierarchical intervention and dynamic regulation module introduces a continuous monitoring mechanism. When the target node continuously triggers an exception and the root cause confidence is higher than the preset suspension threshold, it is determined to be a continuous alarm state. The gradual reduction of production is implemented by calculating the power coefficient decay ratio. When the root cause confidence falls below the recovery threshold and stabilizes, it is determined that the exception is resolved, and the production recovery process is started.

[0012] The channel optimization and closed-loop correction module determines whether to replace a single node or the entire channel according to the node change when the power coefficient of the main channel node decreases to the power threshold. The channel configuration is updated and returned to the monitoring link to observe whether the alarm continues to occur. The main channel state and the adjusted operation effect are continuously monitored to iteratively optimize the channel structure and power coefficient setting.

[0013] In a preferred embodiment, the cause-effect correlation graph construction module includes the following steps:

[0014] Deploy data collection devices in multiple process units, connect the key signals of each process to a unified data platform, and form an integrated data set across processes;

[0015] Time series similarity analysis combined with cross-correlation calculation is used to compare the change patterns of different device signals within a sliding time window, quantify the similarity between signals, and estimate the time delay;

[0016] For each pair of consecutive process signals, if the cross-correlation coefficient is higher than the relationship threshold and the delay is within the preset delay range in some period, it is determined that the two signals are related. The peak position of the cross-correlation is used to determine the order of the signals and mark the directed edge. It is defined as a candidate cause-effect relationship edge with related parameters, and a directed cause-effect correlation graph across processes is preliminarily constructed.

[0017] The integrated data set across processes uses time series similarity analysis combined with cross-correlation calculation to compare the change patterns of different device signals within a sliding time window, quantify the similarity between signals, and estimate the time delay;

[0018] For each pair of consecutive process signals, if the cross-correlation coefficient is higher than the relationship threshold and the delay is within the preset delay range in some period, it is determined that the two signals are related. The peak position of the cross-correlation is used to determine the order of the signals and mark the directed edge. It is defined as a candidate cause-effect relationship edge with related parameters, and a directed cause-effect correlation graph across processes is preliminarily constructed.

[0019] An attention head is preset for each type of alarm detection device on the causal graph, and each attention head only focuses on the nodes and edges of the channel to which it belongs;

[0020] Meanwhile, a preset main channel and multiple parallel channels are constructed, and different dynamic coefficients are set, wherein the channel refers to the transmission path of the medium between processes.

[0021] In a preferred embodiment, the root cause positioning and link optimization module comprises the following steps:

[0022] When any alarm detection device triggers an alarm, the causal correlation graph and the attention head setting activate the attention head corresponding to the alarm detection device that triggers the alarm, and analyze the local causal edge sequence of the alarm;

[0023] The attention head filters and retains the causal edges with high similarity to the alarm detection device and a shared medium label that is true as the local causal edge sequence of the alarm, wherein the shared medium label that is true means that there is a physical medium transmission relationship between the two devices at both ends of the causal chain and the collected data is statistically significantly correlated;

[0024] Based on the local causal edge sequence of the alarm, a multi-layer attention network is used to model and calculate the sequence, and the root cause confidence of each node is calculated comprehensively based on the external influence weight of the node, the received external influence weight and the abnormal deviation degree of the node itself;

[0025] The related nodes are sorted according to the root cause confidence, and the one with the highest confidence is identified as the most likely root cause of the cross-line fault;

[0026] The physical constraint verification and optimization are performed on each causal chain;

[0027] The delay distribution of each edge in the link is accumulated to obtain the total delay from the root cause to the final abnormality;

[0028] If the total delay exceeds the preset cross-line physical upper limit, the chain is determined to be a pseudo correlation and is pruned, and a root cause chain is generated.

[0029] In a preferred embodiment, the hierarchical intervention and dynamic control module comprises the following steps:

[0030] A continuous monitoring and alarm mechanism is introduced;

[0031] When a target node repeatedly triggers an abnormal alarm in consecutive time windows, and the corresponding root cause confidence is higher than the preset suspension threshold in each analysis, it is determined to be in a continuous alarm state;

[0032] The suspension threshold and the recovery threshold are set, and the suspension threshold is greater than the recovery threshold;

[0033] The set threshold and the continuous alarm state, if the confidence of the root cause node is higher than the suspension threshold, the control instruction is automatically sent to the upstream production control system, the emergency production reduction / stop intervention measure is taken, the supply or pressure of the medium to the downstream is reduced;

[0034] If the root cause confidence is between the suspension threshold and the recovery threshold, the dynamic adjustment of the production is calculated according to the position of the confidence relative to the two thresholds, and the gradual production reduction strategy is adopted.

[0035] In a preferred embodiment, the channel optimization and closed-loop correction module comprises the following steps:

[0036] The monitoring result after the channel adjustment, when the continuous alarm phenomenon disappears, that is, the reading of the target node returns to normal and the corresponding root cause confidence decreases to below the recovery threshold, and remains stable in multiple consecutive time windows, it is determined that the abnormality has been basically eliminated, and the limited production of the upstream enters the recovery stage;

[0037] The channel label of the causal graph is updated at the same time, and the monitoring is continued until the system runs stably.

[0038] The technical effect and advantages of the analysis and diagnosis system for industrial big data collection of the present application are:

[0039] By constructing a cross-process directed causal correlation graph, integrating multi-process data such as ironmaking, continuous casting and rolling, tracing the upstream root cause, identifying the cross-production line fault propagation link, combining a multi-layer attention network and physical constraint verification, calculating the root cause confidence by comprehensively considering the node influence weight, abnormal deviation degree, etc., accurately locating the fault source and eliminating false correlation links, improving the judgment credibility, taking a gradual reduction or emergency shutdown according to the root cause confidence, balancing safety and efficiency, reducing production loss, deeply mining the value of massive sensor data, breaking through the limitations of traditional analysis, presetting attention heads to ensure that there is no dead angle in the whole channel abnormality monitoring, and the causal graph supports multi-node and multi-link connection, adapts to complex process flow paths, has strong process adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0040] Fig. 1 The flowchart of the analysis and diagnosis system for industrial big data collection of the present application.

[0041] Fig. 2 The module diagram of the analysis and diagnosis system for industrial big data collection of the present application. DETAILED DESCRIPTION

[0042] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0043] Embodiments

[0044] Please refer to Figs. 1-2 As shown in the figure, the application discloses an analysis and diagnosis system for industrial big data collection, comprising: a cause-effect correlation graph construction module, a root cause positioning and link optimization module, a hierarchical intervention and dynamic regulation module, and a channel optimization and closed-loop correction module, and the modules are signal-connected;

[0045] The cause-effect correlation graph construction module: integrates key signals of ironmaking, continuous casting and rolling processes, forms a cross-process integrated data set, uses time series similarity analysis and cross-correlation calculation to quantify the similarity and time delay between signals, and constructs a directed cause-effect correlation graph, presets attention heads for various alarm detection devices, defines the main channel and parallel channel of medium transmission and the corresponding power coefficient;

[0046] The root cause positioning and link optimization module: when the alarm detection device triggers an alarm, the corresponding attention head is activated, the cause-effect correlation graph is traced back, and a local cause-effect edge sequence with high similarity to the alarm device and shared media is screened out. The root cause confidence of each node is calculated through a multi-layer attention network, the most likely root cause is identified, the identified cause-effect chain is physically constrained and verified, and a root cause chain that meets the process constraints is generated;

[0047] The hierarchical intervention and dynamic regulation module: a continuous monitoring mechanism is introduced. When the target node continuously triggers an exception and the root cause confidence is higher than the preset suspension threshold, it is determined to be a continuous alarm state. Gradual production reduction is implemented by calculating the power coefficient decay ratio. When the root cause confidence falls below the recovery threshold and stabilizes, it is determined that the exception is resolved, and the production recovery process is started;

[0048] The channel optimization and closed-loop correction module: when the power coefficient of the main channel node decreases to the power threshold, it is determined whether to replace a single node or the entire channel according to the node change, the channel configuration is updated, and the monitoring link is returned. Whether the alarm continues to occur is observed. The main channel state and the adjusted operation effect are monitored continuously, and the channel structure and power coefficient setting are iteratively optimized.

[0049] In the causal correlation graph construction module, the key signals of ironmaking, continuous casting and rolling processes are integrated to form a cross-process integrated data set, time series similarity analysis and cross-correlation calculation are used to quantify the similarity and time delay between signals, and a directed causal correlation graph is constructed, which presets the attention head for various alarm detection devices, defines the main channel and parallel channel of medium transmission and the corresponding dynamic coefficient, and the specific content includes:

[0050] In multiple process units including ironmaking blast furnace, continuous casting process and rolling mill unit, data acquisition devices are deployed, and key signals of ironmaking furnace temperature, continuous casting temperature and rolling mill vibration are connected to a unified data platform to form a cross-process integrated data set;

[0051] Time series similarity analysis combined with cross-correlation calculation is used to compare the change patterns of different equipment signals in a sliding time window, and the similarity between signals is quantified and the time delay is estimated, for each pair of process signals, the time series similarity is calculated in a sliding time window, and the Euclidean distance or dynamic time warping is used as the initial similarity index.

[0052] In order to quantify the time delay, the cross-correlation function is further calculated, the cross-correlation is a statistical quantity that measures the similarity of two signals with displacement, and the cross-correlation of continuous process signals and is defined as: The peak value of the cross-correlation function indicates that the similarity of the two signals is maximum at a certain delay , and this displacement corresponds to the time delay between the two signals, the cross-correlation and its normalized form can quantify the similarity between signals and estimate the time lag of medium transmission.

[0053] In addition, in order to learn the equipment operation mode from historical normal data and quantify the abnormal degree, the present application uses multivariate sensor estimation technology in the causal graph construction link, sets the dictionary matrix composed of historical normal operation data as , and the current observation vector as , then the weight vector is obtained by solving the weighted least squares equation, and the estimated value is calculated, the residual vector is , and the norm size measures the deviation degree of the current state from the normal mode, by comparing the residual norm with the standard deviation of the historical normal data, the abnormal deviation degree is defined, which will be used in the subsequent root cause positioning module.

[0054] If the cross-correlation coefficient is higher than the correlation threshold and the delay is within the preset delay range in some time period, it is considered that the two signals are related. According to the peak position of the cross-correlation, the order of the signals is determined. If the peak value of signal A appears in the lag of B, then A leads B, and the directed edge is recorded as A→B;

[0055] Each pair of related process signals is defined as a candidate causal relationship edge, and is accompanied by TSS similarity, delay distribution of high correlation time, and whether the two devices are real shared medium labels, thereby preliminarily constructing a cross-process directed causal correlation graph;

[0056] A preliminary directed causal graph is constructed on the set of device nodes across processes, where nodes represent ironmaking, continuous casting and rolling process, and edges represent causal relationships. Since the process flow path is not one-to-one, a certain node can be connected to multiple downstream nodes, and can also receive influence from multiple upstream nodes;

[0057] To ensure that the system can monitor each channel, a preset attention head is set for each type of alarm detection device on the causal graph. For example, the blast furnace outlet temperature alarm, the continuous casting mold temperature alarm and the rolling mill vibration alarm correspond to different attention heads. Each attention head only focuses on the nodes and edges of its own channel, thereby ensuring that any channel anomaly can be captured by the corresponding head. For complex cross-connections, some nodes can be mapped to multiple attention heads according to process knowledge.

[0058] A channel refers to the transmission path of the medium between processes, and a preset main channel and multiple parallel channels are constructed, and different dynamic coefficients are set.

[0059] In the root cause positioning and link optimization module, when the alarm detection device triggers an alarm, the corresponding attention head is activated, the causal correlation graph is traced back, the local causal edge sequence with high similarity to the alarm device and shared medium is screened out, the root cause confidence of each node is calculated through a multi-layer attention network, and the most likely root cause is identified. The physical constraint verification is performed on the identified causal chain to generate a root cause chain that meets the process constraints. The specific content includes:

[0060] When any alarm detection device triggers an alarm, the previously established causal correlation graph is automatically traced back, the alarm local causal edge sequence is analyzed by the attention head corresponding to the alarm detection device that triggers the alarm, and the attention head applies different criteria to filter edge relationships. Only the causal edges with high similarity to the alarm detection device and the shared medium label are retained as the alarm local causal edge sequence, and the shared medium label is true, which means that there is a physical medium transmission relationship between the devices at both ends of the causal chain and the collected data is statistically significantly correlated;

[0061] In the root cause confidence calculation process, the application first calculates the external influence weight and the received external influence weight of each node on the local causal graph. For a node , let its out-edge set be , and the weight of each edge be obtained by the aforementioned cross-correlation coefficient or attention mechanism, then the external influence weight of the node is defined as: ;

[0062] Similarly, the received external influence weight of the node is defined as: . The abnormal deviation degree is calculated based on the multivariate sensor estimation technology, as described above, the estimated value and the residual are solved by MSET, and the residual norm corresponding to the node is normalized, ;

[0063] A multi-layer attention network is applied to the local causal edge sequence, and the hidden feature representation of each node is spliced with the above three features and mapped to the intermediate representation : ; wherein, and are network weights and biases, and ReLU is a linear rectifier function;

[0064] The logistic function is used to normalize the intermediate representation to obtain the root cause confidence of the node: ; wherein, and are learnable parameters, is a sigmoid function, and the root cause confidence ranges from , and the larger the value, the higher the possibility of the node as the root cause;

[0065] When a node has a large external influence weight, a small received weight, and a significant abnormal deviation degree, the above formula will give a high confidence, and vice versa.

[0066] Accordingly, the related nodes are sorted, and the node with the highest confidence is identified as the most likely root cause of the cross-line fault. Since the downstream influence weight is large and the abnormality is significant, the node is given the highest root cause confidence, and is finally determined as the root cause of the cross-line fault;

[0067] Physical constraint verification and optimization are performed on each causal chain identified: the delay distribution of each edge in the chain is accumulated to obtain the total delay from the root cause to the final anomaly.

[0068] If the total delay of a link exceeds the pre-set cross-line physical upper limit, it is considered that the link does not conform to the real process constraint, is a pseudo correlation, and the link is directly pruned to generate a root cause chain.

[0069] In the hierarchical intervention and dynamic regulation module, a continuous monitoring mechanism is introduced, when the target node continuously triggers an exception and the root cause confidence is higher than the pre-set suspension threshold, it is determined that it is a continuous alarm state, and the production is gradually reduced by calculating the power coefficient decay ratio, when the root cause confidence decreases to below the recovery threshold and is stable, it is determined that the exception is removed, and the production recovery process is started, the specific content includes:

[0070] A continuous monitoring and alarm mechanism is introduced, when a target node repeatedly triggers an exception alarm in continuous multiple time windows, and the corresponding root cause confidence is always higher than the pre-set suspension threshold, for example 0.8, it is determined that it is a continuous alarm state, the exception is not accidental but continuous, and has reached the severity that needs intervention;

[0071] A two-level threshold strategy is set: a higher suspension threshold and a lower recovery threshold, the suspension threshold is greater than the recovery threshold, for example, the suspension threshold is set to 0.8 and the recovery threshold is set to 0.2;

[0072] The root cause confidence exceeding the suspension threshold indicates that the exception is serious and needs to take suspension measures; and decreasing to below the recovery threshold indicates that the conditions for system recovery are met;

[0073] Once it is detected that a root cause chain enters a continuous alarm state and the confidence of its root cause node is higher than the suspension threshold, a control instruction is automatically sent to the upstream production control system, and corresponding emergency production reduction / shutdown intervention measures are taken;

[0074] The supply or pressure of the medium to the downstream is correspondingly reduced, and the propagation path of the abnormal influence is cut off. In this way, the abnormality is suppressed in the first time to prevent equipment damage or safety accidents;

[0075] For the case that the root cause confidence C is between the suspension threshold and the recovery threshold , the invention adopts a gradual production reduction strategy instead of simple full stop or no stop, and the normal production load is , then the power coefficient decay ratio is calculated according to the position of the root cause confidence in the interval , and the calculation can be represented as: ; wherein ;

[0076] The power coefficient reflects the production capacity of the current channel, and its update formula is: That is, when the root cause confidence is closer to the suspension threshold, the production load is lower; when approaching the recovery threshold, the production load approaches the normal value, and through this linear mapping, the medium supply can be smoothly adjusted, avoiding unnecessary losses caused by full suspension and ensuring that the system has enough redundancy to suppress fault propagation;

[0077] For the case where the root cause confidence is between the suspension threshold and the recovery threshold, a gradual reduction strategy is adopted, rather than simply continuing full-speed production or complete shutdown;

[0078] According to the position of the confidence relative to the two thresholds, the power coefficient decay ratio is calculated to dynamically adjust the production, for example, if the root cause score is in the middle of the threshold interval, the system may reduce the production load to about 50% of the normal value; if the root cause score approaches the upper limit of the suspension threshold, the production is further reduced to ensure safety; conversely, if it approaches the lower limit of 0.2, only a slight reduction in production is required. Through proportional control, the system not only avoids unnecessary losses caused by shutdown, but also ensures that abnormal problems are controlled and do not develop into serious faults;

[0079] When the persistent alarm phenomenon disappears, specifically, the readings of the target node return to normal and the corresponding root cause confidence falls below the recovery threshold The following, and remains stable in multiple consecutive time windows, the system determines that the anomaly has been basically resolved, and the upstream restricted production can enter the recovery phase, its production load is gradually recovered to , the normal production load benchmark in reverse according to the above formula;

[0080] When the persistent alarm phenomenon disappears, specifically, the readings of the target node return to normal and the corresponding root cause confidence falls below the recovery threshold 0.2, and remains stable in multiple consecutive time windows, the system determines that the anomaly has been basically resolved.

[0081] In the channel optimization and closed-loop correction module, when the power coefficient of the main channel node decreases to the power threshold, it is determined whether to replace a single node or the entire channel according to the node change, and the channel configuration is updated and returned to the monitoring link to observe whether the alarm continues. Through continuous monitoring of the main channel state and the adjusted operation effect, the channel structure and power coefficient setting are iteratively optimized, including:

[0082] If the power coefficient of the main channel node decreases to a certain threshold, the main channel is set to be adjusted, and it is determined whether to replace a certain node or directly replace the entire channel according to the node change. After updating, it is returned to step two to observe whether the alarm continues;

[0083] Continuously monitor the power coefficient of the main channel, if the power coefficient of the main channel decreases below the preset power threshold due to long-term reduction in production, it means that the main channel cannot meet the production demand, and the channel or equipment needs to be adjusted;

[0084] Calculate the average root cause confidence and fault occurrence frequency for each node in the main channel, if the confidence of a certain node is high for a long time and the faults are concentrated, it is considered that the node has a serious fault, and the node should be repaired or replaced first;

[0085] If the confidence of multiple nodes is high and distributed throughout the channel, or the main channel still frequently alarms after replacing a single node and the power coefficient continues to decrease, it is determined that there is a systemic problem in the entire channel, and the parallel channel is enabled or the main channel is shut down for maintenance, the preset parallel channel can guide the medium into the standby line through the switch valve or scheduling; If there is no standby channel, the whole machine needs to be shut down for maintenance;

[0086] Once the node or channel is replaced, update the channel label of the causal graph and return to step two to continue monitoring, if the new channel still triggers an alarm, repeat the above process until it enters a stable running state;

[0087] In order to make the power coefficient change more smoothly after the channel adjustment, the application introduces a power coefficient update function with time, assuming that the power coefficient of the first channel is , the dynamic coefficient of the first channel is updated as follows during channel adjustment and continuous monitoring: ; Wherein, is the inertia factor of the first channel, which represents the frequency of the power coefficient update function, which is fitted based on historical process data, is the abnormal deviation degree, which is obtained by MSET residual normalization, is the discrete time index, is the phase shift constant of the first channel, which reflects the initial phase difference, and the sinusoidal update strategy provides smooth increase and decrease within a local time window, when the abnormal deviation degree increases, the fluctuation amplitude of

[0088] correspondingly increases, reflecting that the power coefficient of the channel needs to be adjusted by a larger amplitude;

[0089] When the abnormal deviation degree decreases, the fluctuation tends to be flat, so that the production gradually returns to normal. Through this closed-loop correction mechanism, the adjustment of the power coefficient of the channel is closely coupled with the process state, which effectively avoids overshoot or lag.

[0090] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product.

[0091] Those skilled in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and the constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0092] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0093] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0094] Finally, the above is merely preferred embodiments of the present application, and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An industrial big data acquisition, analysis, and diagnostic system, characterized in that, Signal connections between modules; The causal relationship graph construction module integrates key signals from ironmaking, continuous casting, and steel rolling processes to form a cross-process integrated dataset. It uses time series similarity analysis and cross-correlation calculation to quantify the similarity and time delay between signals, construct a directed causal relationship graph, preset attention heads for various alarm detection devices, and define the main and parallel channels of media transmission and their corresponding dynamic coefficients. The root cause localization and link optimization module activates the corresponding attention head when the alarm detection device triggers an alarm, traces back the causal relationship graph, filters out the local causal edge sequence that is highly similar to the alarm device and shares the same medium, calculates the root cause confidence of each node through a multi-layer attention network, identifies the most likely root cause, performs physical constraint verification on the identified causal chain, and generates a root cause chain that conforms to the process constraints. The graded intervention and dynamic control module introduces a continuous monitoring mechanism. When the target node continuously triggers anomalies and the root cause confidence level is higher than the preset pause threshold, it is determined to be a continuous alarm state. By calculating the power coefficient decay ratio, a gradual production reduction is implemented. When the root cause confidence level drops below the recovery threshold and stabilizes, the anomaly is determined to be resolved, and the production recovery process is started. The channel optimization and closed-loop correction module determines whether to replace a single node or the entire channel based on the changes in the node when the power coefficient of the main channel node drops to the power threshold. After updating the channel configuration, it returns to the monitoring stage to observe whether alarms continue to be generated. By continuously monitoring the status of the main channel and the operational effect after adjustment, it iteratively optimizes the channel structure and power coefficient settings.

2. The industrial big data acquisition, analysis, and diagnostic system according to claim 1, characterized in that, Data acquisition devices are deployed in multiple process units to connect key signals from each process to a unified data platform, forming an integrated dataset across processes; By combining time series similarity analysis with cross-correlation calculation, the change patterns of signals from different devices are compared within a sliding time window to quantify the similarity between signals and estimate time delay. For each pair of consecutive process signals, if the cross-correlation coefficient is higher than the correlation threshold and the delay is within the preset delay range during certain time periods, it is determined that the two signals are related. The order of the signals is determined according to the peak position of the cross-correlation and the directed edges are marked. These are defined as candidate causal relationship edges and are accompanied by relevant parameters. A preliminary directed causal relationship graph across processes is constructed.

3. The industrial big data acquisition, analysis, and diagnostic system according to claim 2, characterized in that, An attention head is preset for each type of alarm detection device on the causal graph, and each attention head only focuses on the nodes and edges of its own channel; Simultaneously, a preset main channel and multiple parallel channels are constructed, and different power coefficients are set. The power coefficient is a parameter that reflects the production capacity of the medium transmission path, where the channel refers to the transmission path of the medium between processes.

4. The industrial big data acquisition, analysis, and diagnostic system according to claim 1, characterized in that, When any alarm detection device triggers an alarm, the causal relationship graph and attention head settings are configured to activate the attention head corresponding to the alarm detection device that triggered the alarm, and to analyze the local causal edge sequence of the alarm. The attention-based screening retains causal edges that are highly similar to alarm detection devices and have true shared medium labels, as alarm local causal edge sequences. True shared medium labels mean that there is a physical medium transmission relationship between the devices at both ends of the causal chain and that their collected data are statistically significantly correlated.

5. The industrial big data acquisition, analysis, and diagnostic system according to claim 1, characterized in that, Based on the alarm local causal edge sequence, a multi-layer attention network is used to model and calculate the sequence. The root cause confidence of each node is calculated by combining the weight of the node's external influence, the weight of the external influence it receives, and its own abnormal deviation. Each relevant node is sorted by root cause confidence level, and the node with the highest confidence level is identified as the most likely root cause of the cross-line failure.

6. The industrial big data acquisition, analysis, and diagnostic system according to claim 4, characterized in that, Perform physical constraint verification and optimization on each causal chain; The total delay from the root cause to the final anomaly is obtained by summing the delay distributions of each side in the link. If the total delay exceeds the preset physical limit for cross-line connections, the chain is determined to be a spurious correlation and is pruned, generating a root cause chain.

7. The industrial big data acquisition, analysis, and diagnostic system according to claim 1, characterized in that, Introduce a continuous monitoring and alarm mechanism; When a target node repeatedly triggers abnormal alarms within multiple consecutive time windows, and the corresponding root cause confidence score is higher than the preset pause threshold in each analysis, it is determined to be in a continuous alarm state. Simultaneously set a pause threshold and a resume threshold, with the pause threshold being greater than the resume threshold.

8. The industrial big data acquisition, analysis, and diagnostic system according to claim 7, characterized in that, If the confidence level of the root cause node is higher than the pause threshold, the system will automatically send a control command to the upstream production control system to take emergency production reduction / shutdown intervention measures to reduce the supply or pressure of the medium to the downstream. If the root cause confidence level is between the pause threshold and the recovery threshold, the output is dynamically adjusted based on the position of the confidence level relative to the two thresholds, and a gradual production reduction strategy is adopted.

9. The industrial big data acquisition, analysis, and diagnostic system according to claim 1, characterized in that: After the channel adjustment, when the continuous alarm phenomenon disappears, that is, the reading of the target node returns to normal and the corresponding root cause confidence drops below the recovery threshold and remains stable within multiple consecutive time windows, it is determined that the anomaly has been resolved and the upstream restricted production has entered the recovery phase. Simultaneously update the channel labels of the causal graph and continuously monitor until the system is running stably.

Images