Industrial Internet of Things intelligent decision-making method, system and equipment based on large model

Abstract

The invention belongs to the technical field of industrial Internet of Things, and particularly relates to an industrial Internet of Things intelligent decision-making method, system and device based on a large model, and the method comprises the steps: generating a self-adaptive dynamic threshold value through employing a residual compensation algorithm based on the real-time data, working condition parameters and historical data of the device; calculating an abnormal score based on the dynamic threshold value and the equipment real-time data; inputting the abnormal score and the multi-modal data structured context into a large model, generating a decision instruction with confidence evaluation, and recording an audit log; the decision instruction comprises an operation action, an operation parameter and a constraint condition; when the confidence coefficient of the decision instruction reaches a confidence coefficient threshold value or the abnormal score exceeds a set threshold value, automatic operation is triggered, closed-loop feedback is carried out based on the execution effect of the automatic operation, and the confidence coefficient, the confidence coefficient threshold value, the residual compensation coefficient and the abnormal score weight are dynamically updated. And through a residual error compensation algorithm and working condition self-adaptive dynamic threshold adjustment, the false alarm rate of the system is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial Internet of Things, and specifically relates to an industrial Internet of Things intelligent decision-making method, system and equipment based on a large model. Background Art

[0002] In the Industrial Internet of Things (IIoT), industrial equipment alarm mechanisms generally rely on fixed thresholds. This static setting cannot be dynamically adjusted to accommodate changes in actual operating conditions. For example, the normal operating parameter range of equipment varies under different operating conditions such as load rate, ambient temperature, and operating time. Fixed thresholds cannot adapt to these changes, resulting in reduced alarm accuracy and the possibility of false alarms or missed alarms, affecting normal equipment operation and production efficiency.

[0003] With the rapid development of artificial intelligence and big model technologies, many companies are actively exploring solutions for intelligent decision-making. However, in industrial scenarios, big model technology is currently limited to data analysis and fails to form an effective closed loop with automated operations. While big models can perform in-depth analysis of large amounts of industrial data, the results often cannot be directly translated into specific automated operation instructions, resulting in a disconnect between decision-making and execution. Furthermore, the lack of standardized big model instruction interaction templates makes effective information exchange and collaboration between different systems difficult, further hindering the widespread application of intelligent decision-making technology in the Industrial Internet of Things. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an industrial Internet of Things intelligent decision-making method, system and device based on a large model to solve the above-mentioned technical problems.

[0005] In a first aspect, the technical solution of the present invention provides an industrial Internet of Things intelligent decision-making method based on a large model, comprising: S1. Based on the equipment's real-time data, operating parameters, and historical data, a residual compensation algorithm is used to generate adaptive dynamic thresholds. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. S2. Calculate an anomaly score based on dynamic thresholds and real-time device data. The anomaly score is a weighted sum of the single parameter exceeding the standard ratio, multi-parameter associated anomalies, and historical similarity multiplied by sensitivity. S3. Input real-time equipment data, anomaly scores, and multimodal structured context into the big model to generate decision instructions with confidence assessments and record audit logs. The decision instructions include operation actions, operation parameters, and constraints. The structured context includes maintenance records, production plans, spare parts inventory status, and safety operating specifications. S4. When the confidence level of the decision instruction reaches the confidence threshold, the automated operation is triggered, and closed-loop feedback is performed based on the execution effect of the automated operation, and the confidence level, confidence threshold, residual compensation coefficient and anomaly score weight are dynamically updated; when the anomaly score exceeds the set threshold but the confidence level does not reach the confidence threshold, an early warning notification is triggered.

[0006] Based on real-time equipment data, operating parameters, and historical data, a residual compensation algorithm is used to generate adaptive dynamic thresholds. This fully considers the equipment's operating characteristics under varying load rates, ambient temperatures, operating hours, and other conditions, dynamically adjusting the thresholds to reflect the operating conditions. This effectively avoids the problem of false alarms or missed alarms caused by traditional fixed thresholds' inability to adapt to changing operating conditions, thereby improving the accuracy of equipment alarms. By calculating anomaly scores, comprehensively considering the proportion of single parameter out-of-spec values, multi-parameter correlated outliers, and historical similarity, and multiplying them by sensitivity, the system can comprehensively and accurately assess the degree of equipment anomaly. This not only focuses on whether a single parameter exceeds the standard, but also considers the correlation between parameters and the similarity to historical faults, providing a more reliable basis for subsequent decision-making.

[0007] Anomaly scores and the structured context of multimodal data are fed into a large model to generate decision instructions with confidence assessments and record audit logs. When conditions are met, automated actions are triggered. Closed-loop feedback is provided based on the execution results, dynamically updating relevant parameters. This closed-loop mechanism tightly integrates decision-making and execution, automating the entire process from data collection and analysis to decision execution, improving the intelligence and responsiveness of industrial IoT systems.

[0008] As a further limitation of the technical solution of the present invention, the dynamic threshold includes a dynamic threshold upper limit and a dynamic threshold lower limit, wherein:

[0009]

[0010] is the load compensation coefficient, is the environmental compensation coefficient, Refers to 3 times the standard deviation of the threshold baseline, is the frequency band energy weight.

[0011] The formula introduces parameters such as load compensation coefficient, environment compensation coefficient, three times the standard deviation of the threshold baseline and frequency band energy weight, and comprehensively considers the impact of multiple factors such as load, environment, historical data fluctuations and vibration signal characteristics on the threshold, making the generated dynamic threshold more scientific and reasonable and able to better adapt to complex industrial environments.

[0012] As a further limitation of the technical solution of the present invention, the steps for calculating the baseline value are as follows: Get the rated power of the device and real-time power , and calculate the load rate; Eliminate abnormal values from the current, vibration amplitude, and temperature data collected by the equipment under normal conditions, and exclude data during equipment maintenance or sensor failure; The data is divided into different working condition segments according to the load rate. For the normal data in each working condition segment, a time window is set to calculate the arithmetic mean of the data in the time window to obtain the segmented baseline value. , among which, k The segment baseline value of each working condition segment is ; The weighted sum of different segment baseline values is used to obtain the global baseline value, that is, the baseline value ; in, , , , ; Where, N is the number of data points in the window, This is the real-time data of the normal equipment within the window. M is the total number of working condition segments, Under the current working condition k The weight of each working condition segment, r is the current load rate, For the k The average value of the load factor of each working condition segment, is the sharpening factor.

[0013] This series of steps fully considers the equipment's operating conditions under different operating conditions, eliminating the influence of abnormal data and irrelevant factors, making the baseline value more accurate and reliable, and providing a solid foundation for the generation of dynamic thresholds. Dividing the operating conditions into sections according to load rate and calculating baseline values for each section better adapts to the equipment's operating characteristics under different load conditions, allowing the dynamic threshold to more accurately reflect the equipment's normal operating range under different conditions, thereby improving the accuracy of anomaly detection.

[0014] As a further limitation of the technical solution of the present invention, the method for calculating the frequency band energy includes: Vibration signal After adding the Hanning window, perform FFT transformation to obtain the spectrum ; Extract the spectrum of the target frequency band and calculate the energy ratio of the target frequency band to obtain the frequency band energy; ; .

[0015] The vibration signal is subjected to a Hanning window and then an FFT transform is performed to obtain a spectrum. The spectrum of the target frequency band is extracted and the energy percentage of the target frequency band is calculated to obtain the frequency band energy. This method effectively extracts energy information from specific frequency bands in the vibration signal, assisting in analyzing the vibration characteristics of the equipment and providing an important reference for anomaly detection and fault diagnosis. The calculation of frequency band energy takes into account the frequency characteristics of the vibration signal and can more comprehensively reflect the operating status of the equipment. Compared with traditional detection methods based solely on time domain parameters, it can more accurately detect potential equipment faults, improving the sensitivity and accuracy of anomaly detection.

[0016] As a further limitation of the technical solution of the present invention, the step of calculating the anomaly score based on the dynamic threshold and the real-time data of the device includes: For each parameter, count the number of times it exceeds the dynamic threshold within the sliding time window; take the ratio of the number of times exceeding the limit to the total number of sampling points as the parameter's exceeding standard ratio; calculate the arithmetic mean of the exceeding standard ratios of all parameters to obtain the single parameter exceeding standard ratio ;Parameters include temperature, vibration amplitude and current; Calculate the Pearson correlation coefficient matrix between each parameter; take the absolute mean of the off-diagonal elements as the synchronization score ; 1 minus the synchronization score is used as the correlation outlier value ;

[0017] ,

[0018] Where n is the number of monitoring parameters, are the elements of the correlation coefficient matrix, For parameters and The covariance of is the standard deviation; for the time series data in the sliding time window, each column is a parameter, and each row is the sampling value at the same time point to construct the parameter data matrix X; Extract the multidimensional feature vector of the current device; calculate the normalized Euclidean distance between it and the historical fault feature vector; and use 1 minus the normalized distance as the historical similarity; Extract the current device feature vector (including temperature mean, vibration frequency band energy, current fluctuation range, etc.), retrieve the most similar fault feature vector from the historical fault library ; Calculate normalized Euclidean distance

[0019] Historical similarity

[0020] are the maximum value vector and minimum value vector of all feature vectors in the historical fault feature library respectively; The anomaly score is obtained by multiplying the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity by the sensitivity; The calculation formula of the abnormality score is:

[0021] Where, is the weight factor, The sensitivity is adjusted dynamically according to the equipment operation time and environmental conditions.

[0022] By calculating the percentage of single-parameter exceedances, multi-parameter correlated outliers, and historical similarities, and then weighting and multiplying these by sensitivity to derive anomaly scores, the system comprehensively assesses the degree of device anomalies from multiple perspectives. It not only focuses on individual parameter anomalies but also considers correlations between parameters and similarities to historical failures, making anomaly scores more accurate and reliable. Sensitivity dynamically adjusts based on device operating time and environmental conditions, adapting to changes in the device during different operating stages and environments. This allows the anomaly score to more accurately reflect the device's actual anomaly, improving the adaptability and accuracy of anomaly detection.

[0023] As a further limitation of the technical solution of the present invention, the generation of the large model decision instruction in S3 includes: Receive anomaly scores, real-time equipment data, maintenance records, production plans, and spare parts inventory status, and output decision options with confidence, including operation actions, confidence levels, operation parameters, and constraints. The confidence threshold is initialized based on the equipment type and dynamically updated based on the execution effect of the automated operation.

[0024] The large model inputs structured contextual data, including maintenance records, production plans, spare parts inventory status, and safe operating procedures. It comprehensively considers multiple factors, providing a comprehensive basis for generating decision instructions and making decisions more scientific and reasonable. It outputs decision options with confidence levels, including operation actions, confidence levels, operating parameters, and constraints, providing operators with clearer and more reliable decision recommendations. Confidence thresholds are initially set based on the equipment type and dynamically updated based on execution results, continuously improving the accuracy and reliability of decisions.

[0025] As a further limitation of the technical solution of the present invention, the content of the automated operation includes: triggering one of the following automated operations according to the type of decision instruction: Emergency shutdown, load reduction operation, automatic generation of maintenance work orders and push to MES system, connection to ERP system to reserve spare parts inventory, and sending early warning notifications to remind responsible persons.

[0026] Different automated actions are triggered based on the type of decision instruction, such as emergency shutdown, load reduction, automatic maintenance work order generation and push to the MES system, connection to the ERP system to reserve spare parts inventory, and early warning notifications to alert responsible personnel. These actions can be tailored to different equipment anomalies, improving the automation level and emergency response capabilities of the Industrial Internet of Things system. Automated operations enable rapid response to equipment anomalies, reducing the time and errors of manual intervention, improving production efficiency, and taking timely measures to prevent further escalation of equipment failures, thus ensuring production safety.

[0027] As a further limitation of the technical solution of the present invention, in S4, after the automated operation is triggered, the step of performing closed-loop feedback based on the execution effect of the automated operation and dynamically updating the instruction parameters includes: Collect data on the actual operating status of the equipment after executing the decision instructions, including equipment failure repair time, downtime, maintenance cost, and operation success rate; Update the confidence of the single decision instruction of the large model according to the success rate of the recent N operations; ; are all constants; The execution effect evaluation index of the closed-loop feedback mechanism is calculated based on the collected data; the execution effect evaluation index is a reward function:

[0028] in,

[0029]

[0030] Where, is the MTBF improvement rate, is the cost saving rate, is the number of incorrect operations, are weight coefficients respectively; like If the value is lower than the set threshold for T consecutive times, the confidence threshold of the large model decision process, as well as the residual compensation coefficient and anomaly score weight, will be dynamically adjusted.

[0031] The system collects actual operational status data of devices after executing decision instructions, including fault repair time, downtime, maintenance costs, and operation success rate. Based on this data, it calculates the execution effectiveness evaluation indicators of the closed-loop feedback mechanism. This objectively assesses the execution effectiveness of automated operations and provides a basis for subsequent parameter adjustments. A dynamic confidence update strategy is used to update the confidence level of a single decision instruction in the large model based on the success rate of the last N operations. If the confidence level falls below the set threshold for T consecutive times, the confidence threshold of the large model's decision process, as well as the residual compensation coefficient and anomaly scoring weight, are dynamically adjusted. This closed-loop feedback mechanism continuously optimizes the decision-making process, improves the accuracy and reliability of decisions, and enables the Industrial Internet of Things system to better adapt to the ever-changing industrial environment.

[0032] In a second aspect, the technical solution of the present invention further provides an industrial Internet of Things intelligent decision-making method based on a large model, comprising: The dynamic threshold calculation module is used to generate adaptive dynamic thresholds using a residual compensation algorithm based on real-time equipment data, operating parameters, and historical data. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. Anomaly detection and scoring module, which is used to calculate anomaly scores based on dynamic thresholds and real-time device data. The anomaly score is the weighted sum of the single parameter exceedance ratio, multi-parameter associated anomalies, and historical similarity multiplied by sensitivity; The large model decision instruction generation module is used to input real-time equipment data, anomaly scores, and multimodal structured context into the large model, generate decision instructions with confidence assessments, and record audit logs. The decision instructions include operation actions, operation parameters, and constraints. The structured context includes maintenance records, production plans, spare parts inventory status, and safety operating specifications. The automated execution engine module is used to trigger automated operations when the confidence level of a decision instruction reaches a confidence threshold, and to provide closed-loop feedback based on the execution effect of the automated operation, dynamically updating the confidence level, confidence threshold, residual compensation coefficient, and anomaly score weight; when the anomaly score exceeds the set threshold but the confidence level does not reach the confidence threshold, an early warning notification is triggered.

[0033] As a further limitation of the technical solution of the present invention, the dynamic threshold includes a dynamic threshold upper limit and a dynamic threshold lower limit, wherein:

[0034]

[0035] is the load compensation coefficient, is the environmental compensation coefficient, Refers to 3 times the standard deviation of the threshold baseline, is the frequency band energy weight.

[0036] As a further limitation of the technical solution of the present invention, the dynamic threshold calculation module performs the following steps to calculate the baseline value: Get the rated power of the device and real-time power , and calculate the load rate; Eliminate abnormal values from the current, vibration amplitude, and temperature data collected by the equipment under normal conditions, and exclude data during equipment maintenance or sensor failure; The data is divided into different working condition segments according to the load rate. For the normal data in each working condition segment, a time window is set to calculate the arithmetic mean of the data in the time window to obtain the segmented baseline value. , among which, k The segment baseline value of each working condition segment is ; The weighted sum of different segment baseline values is used to obtain the global baseline value, that is, the baseline value ; in, , , , ; Where, N is the number of data points in the window, This is the real-time data of the normal equipment within the window. M is the total number of working condition segments, Under the current working condition k The weight of each working condition segment, r is the current load rate, For the k The average value of the load factor of each working condition segment, is the sharpening factor.

[0037] As a further limitation of the technical solution of the present invention, the method for calculating the frequency band energy by the dynamic threshold calculation module includes: Vibration signal After adding the Hanning window, perform FFT transformation to obtain the spectrum ; Extract the spectrum of the target frequency band and calculate the energy ratio of the target frequency band to obtain the frequency band energy; ; .

[0038] As a further limitation of the technical solution of the present invention, the anomaly detection and scoring module calculates an anomaly score based on a dynamic threshold and real-time device data, including the following steps: For each parameter, count the number of times it exceeds the dynamic threshold within the sliding time window; take the ratio of the number of times exceeding the limit to the total number of sampling points as the parameter's exceeding standard ratio; calculate the arithmetic mean of the exceeding standard ratios of all parameters to obtain the single parameter exceeding standard ratio ;Parameters include temperature, vibration amplitude and current; Calculate the Pearson correlation coefficient matrix between each parameter; take the absolute mean of the off-diagonal elements as the synchronization score ; 1 minus the synchronization score is used as the correlation outlier value ;

[0039] ,

[0040] Where n is the number of monitoring parameters, are the elements of the correlation coefficient matrix, For parameters and The covariance of is the standard deviation; for the time series data in the sliding time window, each column is a parameter, and each row is the sampling value at the same time point to construct the parameter data matrix X; Extract the multidimensional feature vector of the current device; calculate the normalized Euclidean distance between it and the historical fault feature vector; and use 1 minus the normalized distance as the historical similarity; Extract the current device feature vector (including temperature mean, vibration frequency band energy, current fluctuation range, etc.), retrieve the most similar fault feature vector from the historical fault library ; Calculate normalized Euclidean distance

[0041] Historical similarity

[0042] are the maximum value vector and minimum value vector of all feature vectors in the historical fault feature library respectively; The anomaly score is obtained by multiplying the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity by the sensitivity; The calculation formula of the abnormality score is:

[0043] Where, is the weight factor, The sensitivity is dynamically adjusted according to the equipment operation time and environmental conditions.

[0044] As a further limitation of the technical solution of the present invention, the large model decision instruction generation module is specifically used to receive input structured context data, including maintenance records, production plans, spare parts inventory status and safety operating specifications, and output decision options with confidence including operation actions, confidence levels, operation parameters and constraints.

[0045] As a further limitation of the technical solution of the present invention, the content of the automated operation of the automated execution engine module includes: triggering one of the following automated operations according to the type of decision instruction: Emergency shutdown, load reduction operation, automatic generation of maintenance work orders and push to MES system, connection to ERP system to reserve spare parts inventory, and sending early warning notifications to remind responsible persons.

[0046] As a further limitation of the technical solution of the present invention, after the automated operation is triggered, the step of performing closed-loop feedback based on the execution effect of the automated operation and dynamically updating the instruction parameters includes: Collect data on the actual operating status of the equipment after executing the decision instructions, including equipment failure repair time, downtime, maintenance cost, and operation success rate; Update the confidence of the single decision instruction of the large model according to the success rate of the recent N operations; ; are all constants; The execution effect evaluation index of the closed-loop feedback mechanism is calculated based on the collected data; the execution effect evaluation index is a reward function:

[0047] in,

[0048]

[0049] Where, is the MTBF improvement rate, is the cost saving rate, is the number of incorrect operations, are weight coefficients respectively; like If the value is lower than the set threshold for T consecutive times, the confidence threshold of the large model decision process, as well as the residual compensation coefficient and anomaly score weight, will be dynamically adjusted.

[0050] In a third aspect, the technical solution of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores computer program instructions that can be executed by the at least one processor, and the computer program instructions are executed by the at least one processor so that the at least one processor can execute the large model-based industrial Internet of Things intelligent decision-making method as described in the first aspect.

[0051] The beneficial effect of this invention lies in achieving multi-dimensional collaborative perception of equipment status through a residual compensation algorithm and dynamic threshold calculation that adapts to operating conditions. Compared to traditional fixed threshold methods, this method reduces the system's false alarm rate. By establishing an intelligent closed-loop system from perception to decision-making to execution and feedback, the mean time to troubleshoot (MTTR) is shortened, thereby reducing equipment downtime costs. Based on this closed-loop feedback mechanism, the system develops continuous evolutionary capabilities, with the proportion of autonomous decision-making steadily increasing. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 A schematic flow chart of a method according to an embodiment of the present invention.

[0054] Figure 2 A schematic block diagram of a system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the specific embodiments. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0056] like Figure 1 As shown, an embodiment of the present invention provides an industrial Internet of Things intelligent decision-making method based on a large model, including: S1. Based on the equipment's real-time data, operating parameters, and historical data, a residual compensation algorithm is used to generate adaptive dynamic thresholds. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. S2. Calculate an anomaly score based on dynamic thresholds and real-time device data. The anomaly score is a weighted sum of the single parameter exceeding the standard ratio, multi-parameter associated anomalies, and historical similarity multiplied by sensitivity. S3. Input real-time equipment data, anomaly scores, and multimodal structured context into the big model to generate decision instructions with confidence assessments and record audit logs. The decision instructions include operation actions, operation parameters, and constraints. The structured context includes maintenance records, production plans, spare parts inventory status, and safety operating specifications. S4. When the confidence level of the decision instruction reaches the confidence threshold, the automated operation is triggered, and closed-loop feedback is performed based on the execution effect of the automated operation, and the confidence level, confidence threshold, residual compensation coefficient and anomaly score weight are dynamically updated; when the anomaly score exceeds the set threshold but the confidence level does not reach the confidence threshold, an early warning notification is triggered.

[0057] In the embodiment of the present invention, the dynamic threshold includes a dynamic upper threshold and a dynamic lower threshold, wherein:

[0058]

[0059] It should be noted that is the load compensation coefficient, the initial value is 0.2, is the environmental compensation coefficient, with an initial value of 0.05. Refers to 3 times the standard deviation of the threshold baseline, is the frequency band energy weight, which is fitted through historical fault data and has a value of 0.1.

[0060] The steps for calculating the baseline value are as follows: Get the rated power of the device and real-time power , and calculate the load rate; Eliminate abnormal values from the current, vibration amplitude, and temperature data collected by the equipment under normal conditions, and exclude data during equipment maintenance or sensor failure; The data is divided into different working condition segments according to the load rate. For the normal data in each working condition segment, a time window is set to calculate the arithmetic mean of the data in the time window to obtain the segmented baseline value. , among which, k The segment baseline value of each working condition segment is In the embodiment of the present invention, the time window size is 1 hour, and the average value of the data in each operating condition window is calculated to obtain the segmented baseline value; The data is divided into three sections according to the load rate: low load (0-30%), medium load (30-70%), and high load (70-100%); auxiliary sectioning is performed according to the ambient temperature: normal temperature working condition: temperature ≤ 50°C; high temperature working condition: temperature > 50°C; The weighted sum of different segment baseline values is used to obtain the global baseline value, that is, the baseline value ; in, , , , For example, if the rated power of the motor is 10kW and the real-time power is 7kW, the load rate is 70%.

[0061] Where, N is the number of data points in the window, This is the normal real-time data of the device within the time window. M is the total number of working condition segments, For the current working condition k The weight of each working condition segment, r is the current load rate, For the k The average value of the load factor of each working condition segment, is the sharpening factor.

[0062] The baseline value update mechanism updates the baseline value every hour in normal operation mode; in special maintenance mode, the baseline value is updated every 15 minutes within 72 hours after equipment maintenance; the exception handling mechanism suspends the baseline value update and issues an alarm when data anomalies are detected.

[0063] In an embodiment of the present invention, a method for calculating frequency band energy includes: Vibration signal After adding the Hanning window, perform FFT transformation to obtain the spectrum ; Extract the spectrum of the target frequency band and calculate the energy ratio of the target frequency band to obtain the frequency band energy; ; .

[0064] In some embodiments, the step of calculating an anomaly score based on a dynamic threshold and real-time device data includes: For each parameter, count the number of times it exceeds the dynamic threshold within the sliding time window; take the ratio of the number of times exceeding the limit to the total number of sampling points as the parameter's exceeding standard ratio; calculate the arithmetic mean of the exceeding standard ratios of all parameters to obtain the single parameter exceeding standard ratio ; Parameters include temperature, vibration amplitude, and current; for example, if there are 600 temperature sampling points (100 Hz × 60 seconds × 10 minutes), and 120 of them exceed the upper limit of the dynamic threshold, the excess ratio is 20%.

[0065] Calculate the Pearson correlation coefficient matrix between each parameter; take the absolute mean of the off-diagonal elements as the synchronization score ; 1 minus the synchronization score is used as the correlation outlier value ;

[0066] ,

[0067] Where n is the number of parameters, are the elements of the correlation coefficient matrix, For parameters and The covariance of is the standard deviation; for the time series data in the sliding time window, each column is a parameter, and each row is the sampling value at the same time point to construct the parameter data matrix X; for example, the correlation coefficient between temperature and vibration is normal , when abnormal, it drops to , the synchronization score decreases and the associated outlier value increases.

[0068] Extract the multidimensional feature vector of the current device; calculate the normalized Euclidean distance between it and the historical fault feature vector; and use 1 minus the normalized distance as the historical similarity; Extract the current device feature vector (including temperature mean, vibration frequency band energy, current fluctuation range, etc.), retrieve the most similar fault feature vector from the historical fault library ; Calculate normalized Euclidean distance

[0069] Historical similarity

[0070] are the maximum value vector and minimum value vector of all feature vectors in the historical fault feature library respectively; for example, the Euclidean distance between the current vibration spectrum and the historical bearing fault spectrum is 0.3, and the maximum historical distance is 1.5, then the historical similarity is .

[0071] The anomaly score is obtained by multiplying the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity by the sensitivity; The calculation formula of the abnormality score is:

[0072] Where, is the weight factor, The sensitivity is dynamically adjusted according to the device operation time and environmental conditions. , the sensitivity of new equipment (operation time < 100 hours) , used to reduce false alarms, extreme environments (temperature>50℃ or humidity>90%) Used to increase sensitivity. In other cases, the sensitivity is the same as the indicated sensitivity.

[0073] The output range of the anomaly score is 0-100, and different levels of alarms are triggered according to the score threshold: score > 90: emergency shutdown; 80 < score ≤ 90: planned maintenance; score ≤ 80: continuous monitoring.

[0074] In some embodiments, generating the large model decision instruction in S3 includes: The system inputs structured contextual data, including maintenance records, production plans, spare parts inventory status, and safe operating specifications, and outputs decision options with confidence, including operation actions, confidence levels, operation parameters, and constraints. The confidence threshold is initialized based on the equipment type and dynamically updated based on the execution effect.

[0075] Structured data template: { "context": { "abnormal_score": 85, "temperature_threshold": [15, 100], "maintenance_history": ["2025-02-20 Bearing Replacement"], "production_phase": "peak" } } Model decision output format (decision option with confidence): { "options": [{ "action": "emergency_shutdown", "confidence": 0.91, "params": {"delay_min": 5}, "constraints": ["require_digital_signature"] }] } Confidence threshold strategy: Critical equipment: Shutdown threshold ≥ 0.9, warning threshold ≥ 0.75, other cases are ignored. Ordinary equipment: Shutdown threshold ≥ 0.85, warning threshold ≥ 0.65, other cases are ignored.

[0076] In some embodiments, the content of the automated operation includes: triggering one of the following automated operations based on the type of the decision instruction: Emergency shutdown, load reduction operation, automatic generation of maintenance work orders and push to MES system, connection to ERP system to reserve spare parts inventory, and sending early warning notifications to remind responsible persons.

[0077] If anomaly score > 90 and associated device status = "normal": Generate emergency repair orders and automatically reserve spare parts warehouses else if 80<Abnormality Score<=90and Production Plan = "Idle Period": Generate planned maintenance orders and push them to the maintenance queue else: Record early warnings and continuously monitor In some embodiments, in S4, after the automated operation is triggered, the step of performing closed-loop feedback based on the execution effect of the automated operation and dynamically updating the instruction parameters includes: Collect data on the actual operating status of the equipment after executing the decision instructions, including equipment failure repair time, downtime, maintenance cost, and operation success rate; Update the confidence of the single decision instruction of the large model according to the success rate of the recent N operations; In the embodiment of the present invention, , .

[0078] The execution effect evaluation index of the closed-loop feedback mechanism is calculated based on the collected data; the execution effect evaluation index is a reward function:

[0079] in,

[0080]

[0081] Where, is the MTBF improvement rate, is the cost saving rate, is the number of incorrect operations, are weight coefficients respectively; like If the value is lower than the set threshold for T consecutive times, the confidence threshold of the large model decision process, as well as the residual compensation coefficient and anomaly score weight, will be dynamically adjusted.

[0082] The confidence threshold is adjusted in steps of ±0.05 and does not exceed the safety range of the device type; The anomaly score weights are redistributed every 24 hours, with priority given to reducing the historical similarity weights.

[0083] For example, a pump frequently shuts down due to sudden load changes ( <0.4 for 4 consecutive times); Adjustment process: Increase the shutdown confidence threshold (0.9 → 0.93); Reduce the load compensation factor (0.2→0.18); Reduce the historical similarity weight from 0.3 to 0.25. After the adjustment, the number of false stops was reduced by 35%. Rebounded to 0.65.

[0084] In addition, for example, in high temperature environments, there are many false alarms of vibration ( <0.5 for 3 consecutive times); Adjustment process: Increase environmental compensation coefficient (0.05→0.06); Sensitivity reduced from 1.2 to 1.0; optimized vibration frequency range (500-1000Hz→600-900Hz).

[0085] like Figure 2 As shown, an embodiment of the present invention further provides an industrial Internet of Things intelligent decision-making method based on a large model, including: The dynamic threshold calculation module is used to generate adaptive dynamic thresholds using a residual compensation algorithm based on real-time equipment data, operating parameters, and historical data. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. Anomaly detection and scoring module, which is used to calculate anomaly scores based on dynamic thresholds and real-time device data. The anomaly score is a weighted sum of the single parameter exceedance ratio, multi-parameter associated anomalies, and historical similarity, multiplied by sensitivity; A large model decision instruction generation module is used to receive anomaly scores and structure multimodal data into context data, generate decision instructions with confidence assessments, and record audit logs; the decision instructions include operation actions, operation parameters, and constraints; The automated execution engine module is used to trigger automated operations when the confidence of a decision instruction reaches a preset confidence threshold or the anomaly score exceeds a set threshold, and to provide closed-loop feedback based on the execution effect of the automated operation, dynamically updating the confidence, confidence threshold, residual compensation coefficient, and anomaly score weight.

[0086] The dynamic threshold includes a dynamic upper threshold and a dynamic lower threshold, wherein:

[0087]

[0088] is the load compensation coefficient, is the environmental compensation coefficient, Refers to 3 times the standard deviation of the threshold baseline, is the frequency band energy weight.

[0089] In some embodiments, the dynamic threshold calculation module calculates the baseline value in the following steps: Get the rated power of the device and real-time power , and calculate the load rate; Eliminate abnormal values from the current, vibration amplitude, and temperature data collected by the equipment under normal conditions, and exclude data during equipment maintenance or sensor failure; Divide the data into different working condition segments according to the load rate, and set a time window for the normal data in each working condition segment. Calculate the arithmetic mean of the data in the window to obtain the segmented baseline value. , among which, k The segment baseline value of each working condition segment is ; The weighted sum of different segment baseline values is used to obtain the global baseline value, that is, the baseline value ; in, , , , ; Where, N is the number of data points in the window, This is the real-time data of the normal equipment within the window. M is the total number of working condition segments, For the current working condition k The weight of each working condition segment, r is the current load rate, For the k The average value of the load factor of each working condition segment, is the sharpening factor.

[0090] In some embodiments, the method for calculating the frequency band energy by the dynamic threshold calculation module includes: Vibration signal After adding the Hanning window, perform FFT transformation to obtain the spectrum ; Extract the spectrum of the target frequency band and calculate the energy ratio of the target frequency band to obtain the frequency band energy; ; .

[0091] In some embodiments, the anomaly detection and scoring module calculates an anomaly score based on a dynamic threshold and real-time device data, including the following steps: For each parameter, count the number of times it exceeds the dynamic threshold within the sliding time window; take the ratio of the number of times exceeding the limit to the total number of sampling points as the parameter's exceeding standard ratio; calculate the arithmetic mean of the exceeding standard ratios of all parameters to obtain the single parameter exceeding standard ratio ;Parameters include temperature, vibration amplitude and current; Calculate the Pearson correlation coefficient matrix between each parameter; take the absolute mean of the off-diagonal elements as the synchronization score ; 1 minus the synchronization score is used as the correlation outlier value ;

[0092] ,

[0093] Where n is the number of monitoring parameters, are the elements of the correlation coefficient matrix, For parameters and The covariance of is the standard deviation; for the time series data in the sliding time window, each column is a parameter, and each row is the sampling value at the same time point to construct the parameter data matrix X; Extract the multidimensional feature vector of the current device; calculate the normalized Euclidean distance between it and the historical fault feature vector; and use 1 minus the normalized distance as the historical similarity; Extract the current device feature vector (including temperature mean, vibration frequency band energy, current fluctuation range, etc.), retrieve the most similar fault feature vector from the historical fault library ; Calculate normalized Euclidean distance

[0094] Historical similarity

[0095] are the maximum value vector and minimum value vector of all feature vectors in the historical fault feature library respectively; The anomaly score is obtained by multiplying the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity by the sensitivity; The calculation formula of the abnormality score is:

[0096] Where, is the weight factor, The sensitivity is dynamically adjusted according to the equipment operation time and environmental conditions.

[0097] In some embodiments, the large model decision instruction generation module is specifically used to receive input structured context data, including maintenance records, production plans, spare parts inventory status and safety operating specifications, and output decision options with confidence including operation actions, confidence levels, operation parameters and constraints.

[0098] In some embodiments, the automation operation of the automation execution engine module includes: triggering one of the following automation operations based on the type of the decision instruction: Emergency shutdown, load reduction operation, automatic generation of maintenance work orders and push to MES system, connection to ERP system to reserve spare parts inventory, and sending early warning notifications to remind responsible persons.

[0099] In some embodiments, after the automated operation is triggered, closed-loop feedback is performed based on the execution effect of the automated operation, and the steps of dynamically updating the instruction parameters include: Collect data on the actual operating status of the equipment after executing the decision instructions, including equipment failure repair time, downtime, maintenance cost, and operation success rate; Update the confidence of the single decision instruction of the large model according to the success rate of the recent N operations; In the embodiment of the present invention, , .

[0100] The execution effect evaluation index of the closed-loop feedback mechanism is calculated based on the collected data; the execution effect evaluation index is a reward function:

[0101] in,

[0102]

[0103] Where, is the MTBF improvement rate, is the cost saving rate, is the number of incorrect operations, are weight coefficients respectively; like If the value is lower than the set threshold for T consecutive times, the confidence threshold of the large model decision process, as well as the residual compensation coefficient and anomaly score weight, will be dynamically adjusted.

[0104] An embodiment of the present invention further provides an electronic device comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. The communication bus can be used to transmit information between the electronic device and a sensor. The processor can call the logic instructions in the memory to execute the following method: S1. Based on the real-time data of the equipment, operating parameters and historical data, a residual compensation algorithm is used to generate an adaptive dynamic threshold; wherein the real-time data of the equipment includes current, vibration amplitude and temperature; the operating parameters include load rate, ambient temperature, and operating time; the historical data is normal / abnormal time series data of a set time in the past; S2. Based on the dynamic threshold and the real-time data of the equipment, the anomaly score is calculated, and the anomaly score is the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity, and multiplied by the sensitivity; S3. The anomaly score and the structured context of the multimodal data are input into the large model to generate a decision instruction with confidence assessment, and an audit log is recorded; the decision instruction includes an operation action, operation parameters and constraints; S4. When the confidence of the decision instruction reaches a preset confidence threshold or the anomaly score exceeds the set threshold, an automated operation is triggered, and closed-loop feedback is performed based on the execution effect of the automated operation to dynamically update the confidence, confidence threshold, residual compensation coefficient and anomaly score weight.

[0105] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0106] Although the present invention has been described in detail with reference to the accompanying drawings and in conjunction with preferred embodiments, the present invention is not limited thereto. Without departing from the spirit and essence of the present invention, persons of ordinary skill in the art may make various equivalent modifications or substitutions to the embodiments of the present invention, and such modifications or substitutions shall be within the scope of the present invention. Any changes or substitutions that can be easily conceived by persons skilled in the art within the technical scope disclosed in the present invention shall be within the scope of protection of the present invention.

Claims

1. An intelligent decision-making method for industrial Internet of Things based on a large model, characterized in that: include: S1. Based on the equipment's real-time data, operating parameters, and historical data, a residual compensation algorithm is used to generate adaptive dynamic thresholds. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. S2. Calculate an anomaly score based on dynamic thresholds and real-time device data. The anomaly score is a weighted sum of the single parameter exceeding the standard ratio, multi-parameter associated anomalies, and historical similarity multiplied by sensitivity. S3. Input real-time equipment data, anomaly scores, and multimodal structured context into the big model to generate decision instructions with confidence assessments and record audit logs. The decision instructions include operation actions, operation parameters, and constraints. The structured context includes maintenance records, production plans, spare parts inventory status, and safety operating specifications. S4. When the confidence level of the decision instruction reaches the confidence threshold, the automated operation is triggered, and closed-loop feedback is performed based on the execution effect of the automated operation, and the confidence level, confidence threshold, residual compensation coefficient and anomaly score weight are dynamically updated; when the anomaly score exceeds the set threshold but the confidence level does not reach the confidence threshold, an early warning notification is triggered.

2. The large-scale model-based intelligent decision-making method for industrial Internet of Things according to claim 1 is characterized in that: The dynamic threshold includes a dynamic upper threshold and a dynamic lower threshold, wherein: Where, is the load compensation coefficient, is the environmental compensation coefficient, Refers to 3 times the standard deviation of the threshold baseline, is the frequency band energy weight.

3. The large-scale model-based intelligent decision-making method for industrial Internet of Things according to claim 2 is characterized in that: The steps for calculating the baseline value are as follows: Obtain the rated power and real-time power of the equipment and calculate the load factor; Eliminate abnormal values from the current, vibration amplitude, and temperature data collected by the equipment under normal conditions, and exclude data during equipment maintenance or sensor failure; The data is divided into different working condition segments according to the load rate. For the normal data in each working condition segment, a time window is set to calculate the arithmetic mean of the data in the time window to obtain the segmented baseline value; The weighted sum of different segment baseline values is used to obtain the global baseline value, i.e. the baseline value.

4. The large-scale model-based intelligent decision-making method for industrial Internet of Things according to claim 3 is characterized in that: The calculation methods of frequency band energy include: Vibration signal After adding the Hanning window, perform FFT transformation to obtain the spectrum ; Extract the spectrum of the target frequency band and calculate the energy ratio of the target frequency band to obtain the frequency band energy; ; 。 5. The large-scale model-based intelligent decision-making method for industrial Internet of Things according to claim 4 is characterized in that: Based on dynamic thresholds and real-time device data, the steps for calculating anomaly scores include: For each parameter, count the number of times it exceeds the dynamic threshold within the time window; take the ratio of the number of times exceeding the limit to the total number of sampling points as the parameter's exceeding the standard ratio; calculate the arithmetic mean of the exceeding standard ratios of all parameters to obtain the single parameter exceeding standard ratio ;Parameters include temperature, vibration amplitude and current; Calculate the Pearson correlation coefficient matrix between each parameter; take the absolute mean of the off-diagonal elements as the synchronization score ; 1 minus the synchronization score is used as the correlation outlier value ; Extract the multidimensional feature vector of the current device; calculate the normalized Euclidean distance between it and the historical fault feature vector; subtract the normalized distance from 1 as the historical similarity ; The anomaly score is obtained by multiplying the weighted sum of the single parameter exceeding the standard ratio, the multi-parameter associated anomaly value and the historical similarity by the sensitivity; The calculation formula of the abnormality score is: Where, is the weight factor, The sensitivity is set according to the equipment operation time and environmental conditions.

6. The large model-based intelligent decision-making method for industrial Internet of Things according to claim 5 is characterized in that: The generation of the large model decision instruction in S3 includes: Receive anomaly scores, real-time equipment data, maintenance records, production plans, and spare parts inventory status, and output decision options with confidence, including operation actions, confidence levels, operation parameters, and constraints. The confidence threshold is initialized based on the equipment type and dynamically updated based on the execution effect of the automated operation.

7. The large-scale model-based intelligent decision-making method for industrial Internet of Things according to claim 6 is characterized in that: The content of the automated operation includes: triggering one of the following automated operations based on the type of decision instruction: Emergency shutdown, load reduction operation, automatic generation of maintenance work orders and push to MES system, connection to ERP system to reserve spare parts inventory, and sending early warning notifications to remind responsible persons.

8. The large model-based intelligent decision-making method for industrial Internet of Things according to claim 7 is characterized in that: In S4, after the automated operation is triggered, closed-loop feedback is performed based on the execution effect of the automated operation, and the steps of dynamically updating the calculation parameters include: Collect data on the actual operating status of the equipment after executing the decision instructions, including equipment failure repair time, downtime, maintenance cost, and operation success rate; Update the confidence of the single decision instruction of the large model according to the success rate of the recent N operations; ; are all constants; The execution effect evaluation index of the closed-loop feedback mechanism is calculated based on the collected data; the execution effect evaluation index is a reward function: in, Where, is the MTBF improvement rate, is the cost saving rate, is the number of incorrect operations, are weight coefficients respectively; like If the value is lower than the set threshold for T consecutive times, the confidence threshold of the large model decision process, as well as the residual compensation coefficient and anomaly score weight, will be dynamically adjusted.

9. An industrial Internet of Things intelligent decision-making system based on a large model, characterized in that: include: The dynamic threshold calculation module is used to generate adaptive dynamic thresholds using a residual compensation algorithm based on real-time equipment data, operating parameters, and historical data. Real-time equipment data includes current, vibration amplitude, and temperature; operating parameters include load rate, ambient temperature, and operating time; and historical data includes normal / abnormal time series data from a set period in the past. Anomaly detection and scoring module, which is used to calculate anomaly scores based on dynamic thresholds and real-time device data. The anomaly score is the weighted sum of the single parameter exceedance ratio, multi-parameter associated anomalies, and historical similarity multiplied by sensitivity; The large model decision instruction generation module is used to input real-time equipment data, anomaly scores, and multimodal structured context into the large model, generate decision instructions with confidence assessments, and record audit logs. The decision instructions include operation actions, operation parameters, and constraints. The structured context includes maintenance records, production plans, spare parts inventory status, and safety operating specifications. The automated execution engine module is used to trigger automated operations when the confidence level of a decision instruction reaches a confidence threshold, and to provide closed-loop feedback based on the execution effect of the automated operation, dynamically updating the confidence level, confidence threshold, residual compensation coefficient, and anomaly score weight; when the anomaly score exceeds the set threshold but the confidence level does not reach the confidence threshold, an early warning notification is triggered.

10. An electronic device, characterized in that: The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores computer program instructions that can be executed by the at least one processor, and the computer program instructions are executed by the at least one processor to enable the at least one processor to execute the large model-based industrial Internet of Things intelligent decision-making method as described in any one of claims 1 to 8.

Images