Industrial Internet of Things sensor fault monitoring method, system, equipment and medium

Abstract

The invention discloses an industrial Internet of Things sensor fault monitoring method, system and device and a medium, and relates to the technical field of sensor fault monitoring. The invention provides an industrial Internet of Things sensor fault monitoring method, which comprises the steps of obtaining a target sensor and at least one associated sensor associated with the target sensor, and constructing a relation model based on the target sensor and the at least one associated sensor; according to the relation model and the actual reading sequence of the at least one associated sensor in the preset time period, obtaining at least one predicted value sequence of the target sensor in the preset time period, and calculating a fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor; and obtaining the relationship health degree of the target sensor according to each fault feature sequence, if at least one relationship health degree is lower than a preset threshold, generating a relationship abnormality report, and sending the relationship abnormality report to the target user.

Description

Technical Field

[0001] This application relates to the field of sensor fault monitoring technology, and in particular to a method, system, device and medium for industrial Internet of Things sensor fault monitoring. Background Technology

[0002] In fields where IoT technologies are widely used, such as industrial automation, smart buildings, and environmental monitoring, sensors, as the fundamental units for sensing the physical world, rely on the accuracy and reliability of their data as the cornerstone of the stable operation of the entire system. If a sensor drifts, is damaged, or experiences performance degradation, it will lead to erroneous control decisions based on its data, reduced energy efficiency, and even safety accidents. Therefore, fault monitoring and diagnosis of sensors themselves are of paramount importance.

[0003] However, existing technologies typically only report the abnormal data itself after detecting an anomaly in sensor data. They can only take action after the sensor readings have shown obvious abnormalities. This makes it impossible to provide early warnings in the early stages when sensor performance begins to decline slightly, resulting in problems such as high false alarm and false negative rates and insufficient accuracy. Summary of the Invention

[0004] The main purpose of this application is to provide a method, system, device and medium for industrial Internet of Things (IoT) sensor fault monitoring, which aims to solve the technical problems of existing technologies that, after detecting abnormal sensor data, only report the abnormal data itself, resulting in high false alarm and false negative rates and insufficient accuracy.

[0005] To achieve the above objectives, in a first aspect, this application provides a method for industrial Internet of Things (IoT) sensor fault monitoring, comprising: Obtain the target sensor and at least one associated sensor associated with the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor; Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor respectively. The relationship health of the target sensor is obtained based on each fault feature sequence. If at least one of the relationship health is lower than a preset threshold, a relationship anomaly report is generated and sent to the target user.

[0006] Optionally, the step of acquiring the target sensor and at least one associated sensor associated with the target sensor includes: Acquire the deployment location information of the target sensor and / or process flow information within a preset time period; Based on the first detection parameter of the target sensor, a second detection parameter associated with the first detection parameter is obtained, and the sensor used to detect the second detection parameter is marked as a candidate sensor. The candidate sensor located in the same equipment and / or the same process flow as the target sensor is marked as an associated sensor.

[0007] Optionally, the step of constructing a relationship model based on the first detection parameters of the target sensor and the second detection parameters of at least one associated sensor includes: Historical time-series data of the target sensor and at least one associated sensor under known normal operating conditions are obtained. The historical time-series data of at least one associated sensor is used as input, and the historical time-series data of the target sensor is used as output. A relationship model is obtained by training through a machine learning regression algorithm.

[0008] Optionally, the step of obtaining the relationship health of the target sensor based on the fault feature sequence includes: The statistical characteristics of the fault feature sequence within a preset time period are obtained, and the health of the relationship is characterized by the statistical characteristics. The statistical characteristics include one or more combinations of statistical distribution characteristics, trend change characteristics, and pattern consistency characteristics. The statistical distribution characteristics include calculating the standard deviation and absolute mean of the fault characteristic sequence; The trend-based feature includes fitting the trend of the fault feature sequence over time using linear regression. The pattern consistency feature includes calculating the difference between the distribution of the current fault feature sequence and the distribution of the fault feature sequence during historical normal periods using a divergence index.

[0009] Optionally, the step of generating a relationship anomaly report if at least one of the relationship health scores is below a preset threshold includes: When at least one of the relationship health values ​​is lower than a preset threshold, the target edge computing device acquires the actual value sequence and fault feature sequence of the target sensor within a preset time period; Check whether there is a complete actual reading sequence of associated devices in the target edge computing device within a preset time period. If there is no complete actual reading sequence of associated devices in the target edge computing device within a preset time period, send a request to the associated device to transmit the actual reading sequence of the preset time period. Based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, as well as the complete actual reading sequence of the associated sensor within the preset time period, a relationship anomaly report is generated.

[0010] Optionally, the step of generating a relationship anomaly report based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, and the complete actual reading sequence of the associated sensor within the preset time period, includes: Perform horizontal correlation and comparative analysis between the data from the target sensor and all associated sensors: Analyze whether the readings of the associated sensors conform to the expected physical or statistical relationship; The number of associated sensors that identify abnormal readings; Based on the process flow, the propagation path and temporal sequence of anomalies in the sensor network are analyzed.

[0011] Optionally, the relationship anomaly report shall include at least one of the following: anomaly summary, multi-sensor data snapshots and comparisons, description of collaborative anomaly patterns, preliminary root cause inferences, and evidence of correlation.

[0012] Secondly, this application provides an industrial Internet of Things (IoT) sensor fault monitoring system, comprising: a management platform, a sensor network platform, and an object platform that sequentially establish communication. The sensor network platform is configured as follows: Obtain the target sensor and at least one associated sensor associated with the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor; Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor respectively. The management platform is configured as follows: The relationship health of the target sensor is obtained based on each fault feature sequence. If at least one of the relationship health is lower than a preset threshold, a relationship anomaly report is generated and sent to the target user.

[0013] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the method described above.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above.

[0015] The beneficial effects that this application can achieve are: This application proposes an industrial IoT sensor fault monitoring method, system, device, and medium, comprising the following steps: acquiring a target sensor and at least one associated sensor, and constructing a relationship model based on the target sensor and the at least one associated sensor; acquiring at least one predicted value sequence of the target sensor within a preset time period based on the relationship model and the actual reading sequence of the at least one associated sensor within a preset time period, and calculating fault feature sequences between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor; acquiring the relationship health of the target sensor based on each fault feature sequence; if at least one of the relationship health is lower than a preset threshold, generating a relationship anomaly report and sending the relationship anomaly report to the target user. Even if all sensor readings increase synchronously due to changes in normal operating conditions, as long as the relationship between them conforms to the model, the residual is small, the health is high, and there will be no false alarms. An alarm will only be triggered when the behavior of the target sensor breaks the relationship with the associated sensor, thus naturally improving accuracy. Slow drift of the target sensor can cause a small but persistent and potentially increasing systematic deviation between its reading and the predicted value of the associated sensor. Before the target sensor reading exceeds the limit, its relationship health may have already declined and reached the threshold in advance, thus enabling early warning of the target sensor's failure state and providing a time window for predictive maintenance. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the sensor fault monitoring process according to an embodiment of this application; Figure 2 This is a schematic diagram of the sensor network platform involved in this application; Figure 3 This is a simplified flowchart illustrating the overall process of an embodiment of this application.

[0017] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0020] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0021] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0022] Example 1 Reference Figure 1 The first embodiment of this application provides a method for monitoring sensor faults in the industrial Internet of Things, including the following steps: S10. Obtain the target sensor and at least one associated sensor related to the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor.

[0023] Optionally, to facilitate understanding of the technical solution in this embodiment, a pump, a common component in industrial settings, is used as an example. The target sensor is an outlet temperature sensor installed on the pump, and the associated sensors are the pump's outlet pressure sensor, flow sensor, and motor power sensor. Under the pump's known normal operating conditions, historical data for a sufficiently long period is collected, including synchronous time-series data of outlet temperature, outlet pressure, flow rate, and motor power. In constructing the relationship model between the target sensor and the associated sensors, a white-box model or a black-box model can be selected. The white-box model is based on thermodynamics and pump principles to establish a physical relationship model. For example, the pump's temperature rise has a functional relationship with power, flow rate, and efficiency containing unknown parameters. Specific parameters are fitted using system identification techniques, such as the least squares method. For example, the mathematical expression of the trained model might be: ;in Let W represent the pump outlet temperature, F represent the motor power, F represent the pump flow rate, and P represent the outlet pressure. This model shows that, under normal conditions, temperature is positively correlated with the power-to-flow ratio and negatively correlated with pressure. It should be noted that the above mathematical expressions are only illustrative of the correlation between the sensors and do not represent the actual mathematical expression of the relationship model. The black-box model employs machine learning algorithms, such as Support Vector Machine Regression (SVR) or Random Forest. Outlet pressure, flow rate, and motor power are used as input features, and outlet temperature is used as the output label. Historical data is used for training, allowing the algorithm to automatically learn the complex nonlinear mapping relationship between the input features and the output label.

[0024] In this step, by utilizing the inherent physical relationships of the pump system, temperature is no longer viewed in isolation, but rather placed within a system composed of pressure, flow rate, and power, making the monitoring logic more consistent with the actual physical world.

[0025] S20. Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor.

[0026] Optionally, the preset time period is a time window set according to actual needs, such as the most recent 30 minutes. The actual reading sequences of pressure, flow rate, and power within the most recent 30 minutes are obtained. These actual reading sequences from the aforementioned associated sensors are input into the pre-trained relational model in S10 to calculate the predicted temperature value at each moment within these 30 minutes, forming a predicted value sequence. Simultaneously, the actual value sequence of the temperature sensor within the same time period is obtained, and then the residuals are calculated point by point to form a fault feature sequence, which is the residual sequence. For example: within 30 minutes, the model predicts the temperature should be 75℃, but the actual temperature sensor readings continuously fluctuate around 78℃. Then the fault feature sequence will show a series of stable positive residual values, such as +3℃, +2.8℃, +3.1℃..., indicating that the sensor readings are consistently too high.

[0027] This step enables early and quantitative characterization of faults. Even if the actual temperature reading remains within the safe threshold, the fault characteristic sequence can capture this weak deviation signal as soon as it begins to systematically deviate from the predicted value. Converting the problem from absolute values ​​to relative deviations allows the system to distinguish between normal operating condition fluctuations (predictable by the model with small residuals) and genuine anomalies (unexplainable by the model with large residuals).

[0028] S30. Obtain the relationship health of the target sensor based on each fault feature sequence. If at least one relationship health is lower than a preset threshold, generate a relationship anomaly report and send the relationship anomaly report to the target user.

[0029] Optionally, statistical analysis can be performed on the fault characteristic sequence obtained in S20. For example, the mean of the sequence can be calculated to reflect systematic bias and the standard deviation to reflect volatility. These two features can be combined into a relationship health index using a formula, such as a score from 0 to 1. The larger the residual and the more unstable the system, the lower the health score. A health threshold, such as 0.5, can be set. If the calculated health score is lower than this threshold, an alarm is triggered. A relationship anomaly report is generated. The report not only includes conclusions such as low health of the pump outlet temperature sensor, but more importantly, it needs to integrate data from related sensors for collaborative analysis to provide diagnostic conclusions. For example: The target temperature sensor reading is consistently higher than the model prediction value by about 3°C. At the same time, the readings of the related pressure and flow sensors are normal, and the relationship model between them is healthy. Based on the comprehensive assessment, the temperature sensor itself is likely to have experienced calibration drift, and on-site calibration is recommended.

[0030] In this step, through multi-sensor data fusion analysis, the report provides a preliminary diagnosis of the root cause of the fault, helping maintenance personnel distinguish between a faulty sensor and a faulty device, thus achieving precise fault isolation. Health is a continuously changing indicator; its downward trend can be used to predict remaining service life, allowing maintenance to be scheduled before the sensor completely fails, avoiding unexpected downtime. Maintenance personnel no longer receive a vague temperature anomaly alarm, but rather an action guide with clear diagnostic suggestions, enabling them to bring the correct tools and spare parts to the site for repair, shortening troubleshooting and repair time. For example: The system alarms, and the report indicates that the temperature sensor's health score is only 40 points. The maintenance engineer reviews the report's analysis, confirms that the pressure and flow are normal, and prioritizes checking the temperature sensor itself, discovering that a loose connection terminal caused increased contact resistance and a high reading. This quickly resolves the problem, avoiding a misdiagnosis as pump overheating and the need for extensive disassembly and repair.

[0031] Example 2 Based on Example 1, this example provides a method for industrial Internet of Things (IoT) sensor fault monitoring, including the following steps: S10. Obtain the target sensor and at least one associated sensor related to the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor.

[0032] Optionally, the step of acquiring the target sensor and at least one associated sensor associated with the target sensor includes: S101. Obtain the deployment location information of the target sensor and / or the process flow information within a preset time period; Specifically, the physical deployment location information of the target sensor is queried from the equipment asset management system or sensor metadata database. The physical deployment location information includes: the identification of the equipment to which it belongs, such as: pump P-101, reactor R-201; the specific location on the equipment, such as: outlet, inlet, bearing housing, shell; and spatial proximity, such as: installed on the same pipeline as pressure sensor / flow sensor.

[0033] From the manufacturing execution system or process flow diagram, query the process information in which the target sensor participated within a preset time period. This includes: the process unit / section it belongs to, such as: cooling water circulation system, feed pretreatment section; upstream raw material / energy source, such as: heating by boiler B-01; downstream product / energy destination, such as: feeding to distillation column T-01; current process formula or operating mode, such as: executing the production formula of "Product A" and being in the "heating" stage. The core logic of this step is to establish context for subsequent correlation analysis. The behavior of a sensor depends not only on itself, but also on its physical environment and logical flow. The deployment location defines the physical entity it monitors; sensors on the same equipment are inevitably affected by the operating status of that equipment, and there is an inherent physical correlation, such as the outlet pressure and outlet temperature of a pump. The process flow defines the logical process it monitors; there is a causal or functional correlation between sensors on upstream and downstream equipment, such as the preheater outlet temperature directly affecting the reactor inlet temperature.

[0034] In this step, it is ensured that the subsequently selected associated sensors and the target sensor are within the same functional domain, giving the constructed relationship model a clear physical or technological meaning, rather than a purely numerical association. Selecting associated sensors based on the actual physical layout and process flow allows the trained relationship model to better reflect the system's true operating mechanism, resulting in more accurate predictions and stronger adaptability to changes in operating conditions. When a fault occurs, deployment and process information helps to quickly understand the scope and propagation path of the fault, providing crucial clues for fault isolation.

[0035] S102. Based on the first detection parameter of the target sensor, obtain the second detection parameter associated with the first detection parameter, mark the sensor used to detect the second detection parameter as a candidate sensor, and mark the candidate sensor located in the same equipment and / or the same process flow as the target sensor as an associated sensor.

[0036] Specifically, based on physical laws, equipment principles, or process knowledge, analyze the first detection parameter of the target sensor, such as temperature; and the second detection parameter that is strongly correlated with it, such as pressure related to thermodynamics, flow rate related to heat transfer, power related to energy, and vibration related to equipment status. In the plant-wide sensor list, all sensors measuring these second detection parameters are marked as candidate sensors. This step ensures parameter-level correlation. Using the deployment location information and process flow information obtained in S101, candidate sensors are filtered. From the candidate sensors, those sensors installed on the same equipment as the target sensor are selected. For example, if the target sensor is a pump temperature sensor, then pressure, vibration, and flow sensors on the same pump are retained. From the candidate sensors, those sensors on equipment upstream and downstream of the target sensor in the same process flow are selected. For example, if the target sensor is a reactor temperature, then the outlet temperature sensor of the upstream preheater and the inlet temperature sensor of the downstream cooler are retained. Sensors selected through any one or two of the above criteria are finally marked as associated sensors.

[0037] In this step, precise selection limits the number of sensors involved in modeling to the minimum necessary set, reducing data processing volume and bandwidth consumption at the edge and in the cloud. Removing noise interference from irrelevant or weakly correlated sensors allows the relationship model to focus more on the strongest and most direct correlations, resulting in a simpler, more robust model with faster convergence and higher prediction accuracy.

[0038] Optionally, the step of constructing a relationship model based on the target sensor and at least one associated sensor includes: Historical time-series data of the target sensor and at least one associated sensor under known normal operating conditions are acquired. The historical time-series data of at least one associated sensor is used as input, and the historical time-series data of the target sensor is used as output. A relationship model is obtained by training a machine learning regression algorithm.

[0039] Specifically, ensure that the historical time-series data of the target sensor and associated sensors are strictly synchronized in timestamps. Handle missing values, such as through interpolation, and obvious outliers, such as spikes caused by transient interference. Use historical readings from at least one associated sensor as input features. Further derived features can be constructed to enhance model performance, such as: using moving averages to capture trends, using instantaneous rates of change to capture dynamics, and interaction terms between associated sensor readings, such as the product of pressure and flow, which may correspond to power. Divide the processed historical data into training, validation, and test sets. Select a suitable machine learning regression algorithm based on system characteristics and data volume. For example: Support Vector Machine Regression: suitable for small to medium-sized datasets, effectively handling nonlinear relationships; Random Forest Regression: strong anti-overfitting ability, providing feature importance assessment. Use the training set data to train the selected algorithm with the goal of minimizing prediction error, such as mean squared error, to determine model parameters. Use the validation set to find the optimal hyperparameter combination of the model through methods such as grid search or Bayesian optimization. The trained and validated relational model, including its structure and parameters, is solidified and deployed to edge computing devices or cloud analytics platforms for real-time or near-real-time prediction.

[0040] S20. Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor.

[0041] S30. Based on each fault characteristic sequence, obtain the relationship health of the target sensor. If at least one relationship health is lower than a preset threshold, generate a relationship anomaly report and send the relationship anomaly report to the target user.

[0042] Optionally, the step of obtaining the relationship health of the target sensor based on the fault feature sequence includes: Obtain the statistical characteristics of the fault feature sequence within a preset time period, and characterize the health of the relationship through the statistical characteristics. The statistical characteristics include one or more combinations of statistical distribution characteristics, trend change characteristics, and pattern consistency characteristics. Based on statistical distribution characteristics, this includes calculating the standard deviation and absolute mean of the fault characteristic sequence; Specifically, an increase in standard deviation directly indicates increased volatility in sensor readings around the predicted value, leading to decreased stability of the relationship. An increase in the absolute mean indicates a growing systematic deviation between the predicted and actual values. Standard deviation and absolute mean reveal problems from different perspectives: a high standard deviation and low absolute mean characterize stable drift, such as sensor calibration misalignment; a low standard deviation and high absolute mean characterize unstable jitter, such as signal interference or loose connections; and a high standard deviation and high absolute mean characterize severe inaccuracy and instability. Even before a significant change in the residual mean, a slight increase in the absolute mean may be the earliest signal of performance degradation, providing an early warning.

[0043] Based on trend change characteristics, this includes fitting the trend of fault characteristic sequences over time using linear regression; Specifically, the slope of the linear regression fit quantifies the rate of relationship deterioration, allowing for the prediction of remaining service life and enabling a leap from condition monitoring to trend prediction. A significant positive trend indicates that the failure is accelerating and requires immediate attention; while a negative trend may indicate that the system is self-healing or that maintenance measures are effective, providing directional guidance for decision-making.

[0044] Pattern consistency features include calculating the difference between the distribution of the current fault feature sequence and the distribution of fault feature sequences during historical normal periods using divergence indices.

[0045] Specifically, a large amount of residual data from historical normal periods is collected to construct a reference distribution, which can be modeled as a Gaussian distribution or directly using an empirical distribution. The distributional differences are calculated to obtain KL divergence or JS divergence. Divergence focuses on the overall shape of the probability distribution; even if the mean and variance do not change significantly, changes in the distribution shape, such as the appearance of bimodalities or changes in skewness, can be keenly detected, making it adept at identifying new or complex faults. It is not sensitive to isolated outliers, focusing on the overall comparison of the distribution, avoiding drastic fluctuations in health status due to a single extreme value, resulting in more stable and reliable assessment results.

[0046] In this step, volatility and deviation levels are assessed through statistical distribution characteristics, the rate and direction of deterioration are assessed through trend change characteristics, and fundamental changes in the overall distribution pattern are assessed through pattern consistency characteristics. Multiple dimensions of features are integrated into a single health index through weighted combination or minimum value averaging. This constructs a multi-layered monitoring system. While a single feature may fail, the probability of false alarms across all three dimensions is extremely low, significantly improving system robustness. Different types of faults exhibit varying characteristics across different dimensions. Slow drifts are captured by trend features, intermittent jitters are identified by statistical features, and complex behavioral distortions are detected by pattern features, achieving comprehensive fault detection. Different devices or sensors may focus on different health dimensions; adjusting the weights allows the assessment strategy to match the risk preferences of specific scenarios.

[0047] In this embodiment, the mathematical expression for relationship health can be:

[0048] Here, H represents the health of the relationship, which is used to score the overall health of the relationship. The range is usually 0-1, and the larger the value, the healthier the relationship. , , These are all weighting coefficients used to adjust the importance of each feature component, satisfying... . The absolute mean of the fault characteristic sequence is used to measure the magnitude of systematic deviation. It represents the standard deviation of the fault characteristic sequence and is used to measure volatility and instability. The slope represents the trend, obtained through linear regression, and reflects the rate of deterioration of the fault characteristic sequence. This represents the JS divergence. , , , These are all scaling parameters, used to adjust according to specific application scenarios and to normalize the dimensions of different features.

[0049] This component represents the statistical distribution characteristics. When the residual mean and standard deviation are small, the closer this component is to 1, the higher the health of the relationship. When there is systematic bias or drastic fluctuation, this component decays exponentially.

[0050] This represents the characteristic component of trend change. The closer the trend slope is to 0, the closer this component is to 1, indicating greater stability. When a significant deterioration trend occurs, this component decreases according to a hyperbolic law.

[0051] The characteristic component representing the pattern consistency property is used when the current distribution is similar to the historical normal distribution. When the value is close to 0, the component is close to 1. When the distribution pattern changes fundamentally, the component decays exponentially.

[0052] To further understand the above formula, regarding weight allocation, typically... , , It emphasizes the current state while also considering trends and pattern changes. Regarding the scaling parameters, and It can be set according to the sensor accuracy. It can be adjusted according to an acceptable rate of deterioration. It can be set to 1. Threshold setting: When the relationship health is less than 0.6, it can be considered a warning; when the relationship health is less than 0.3, it can be considered a serious anomaly.

[0053] Optionally, the step of generating a relationship anomaly report if at least one relationship health score is below a preset threshold includes: S301. When the health of at least one relationship is lower than a preset threshold, the target edge computing device acquires the actual value sequence and fault feature sequence of the target sensor within a preset time period. Specifically, when the relationship health monitoring system detects that at least one relationship health indicator is below a preset threshold, it immediately sends a trigger signal to the target edge computing device. Upon receiving the trigger signal, the target edge computing device reads data from its local cache or storage. This data includes: a sequence of actual values ​​from the target sensor over a preset time period, such as the past 30 minutes; and a sequence of fault characteristics from the same time period. The extracted data is timestamped and standardized to ensure data consistency. Utilizing the local storage capabilities of the edge device avoids the latency associated with requesting data from the cloud, ensuring high timeliness of fault response. Critical data is pre-cached at the edge to prevent loss of critical fault data due to network fluctuations.

[0054] S302. Check if there is a complete actual reading sequence of the associated device in the target edge computing device within a preset time period. If there is no complete actual reading sequence of the associated device in the preset time period in the target edge computing device, send a request to the associated device to transmit the actual reading sequence of the preset time period. Specifically, the target edge computing device checks whether it has a complete sequence of actual readings from associated sensors within the same preset time period. If the data is incomplete, it immediately sends a targeted data request to the edge devices where the associated sensors reside, including the time range, data type, and priority identifier. It receives the data returned by the associated devices, verifies its time range and completeness, and ensures synchronization with the target sensor data. Requesting additional data only when necessary avoids the bandwidth waste of continuously transmitting all associated data, demonstrating the advantages of intelligent collaboration between devices in an edge computing architecture, rather than complete reliance on a central node.

[0055] S303. Generate a relationship anomaly report based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, as well as the complete actual reading sequence of the associated sensor within the preset time period.

[0056] Optionally, the step of generating a relationship anomaly report based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, and the complete actual reading sequence of the associated sensor within the preset time period, includes: S3031. Perform horizontal correlation and comparative analysis on the data of the target sensor and all associated sensors: Specifically, the data from the target sensor and all associated sensors are precisely aligned and fused over time. Real-time correlation coefficients between the target sensor and the readings of each associated sensor are calculated, and multivariate analysis is used to identify anomalous patterns within the sensor population, such as isolated anomalies, group anomalies, or progressive anomalies.

[0057] S3032. Analyze whether the readings of the associated sensors conform to the expected physical or statistical relationship; Specifically, check whether the readings of associated sensors conform to known physical laws. For example, verify whether the pressure and flow rate in a pump system still conform to the pump characteristic curve. Verify whether the statistical relationships between associated sensors remain normal using a statistical model built based on historical normal data. Conduct independent health assessments of the relationships between associated sensors to form secondary health indicators. Establish a two-level verification system: target sensor-associated sensor and associated sensor-associated sensor, to improve diagnostic depth, fully utilize known physical constraints of the system, and provide a reliable theoretical basis for fault diagnosis. Clearly distinguish between sensor faults and equipment faults to enhance the persuasiveness of diagnostic conclusions and improve diagnostic accuracy and efficiency. Check whether the relationships between other associated sensors besides the target sensor are normal. For example, when the target temperature sensor is abnormal, analyze whether the relationship between the pressure sensor and flow sensor conforms to the pump characteristic curve. Determine whether the physical system itself is normal; this is the key logic for achieving fault isolation.

[0058] S3033, The number of associated sensors that identify abnormal readings; Specifically, a separate health assessment is applied to each associated sensor to identify sensors with abnormal readings. The proportion of abnormal sensors to the total number of associated sensors is statistically analyzed, and the distribution of abnormal sensors in the physical space or functional modules is examined. The number of associated sensors whose readings exhibit abnormal behavior, such as exceeding their reasonable range or rate of change, is counted within a preset time period. This is used to quantify the impact of the fault and assist in determining its severity and root cause. For example: isolated anomalies, such as only the target sensor being abnormal, indicate a fault in the sensor itself. Localized anomalies, such as some associated sensors being abnormal, may indicate a fault in a sub-component of the equipment. Global anomalies, such as most sensors being abnormal, strongly indicate a fault in the entire equipment or a major process disturbance.

[0059] S3034. Based on the process flow, analyze the propagation path and time sequence of anomalies in the sensor network.

[0060] Specifically, based on the system's process flow diagram, the temporal sequence of abnormal signals from each sensor is analyzed. For example, an abnormal pressure difference in the upstream filter occurs first, followed by a decrease in pump flow, and finally an abnormal reactor temperature. This facilitates root cause analysis and pinpoints the initial point of failure.

[0061] Optionally, the relationship anomaly report shall include at least one of the following: anomaly summary, multi-sensor data snapshots and comparisons, description of collaborative anomaly patterns, preliminary root cause inferences, and evidence of correlation.

[0062] Specifically, the exception summary mainly includes the following: Report Identifier: A unique ID used for tracking and tracing.

[0063] Trigger time: The specific point in time when the relationship health level falls below the threshold.

[0064] Target sensor information: sensor ID, physical location, such as pump P-101 outlet, detection parameters, such as temperature.

[0065] List of associated sensors: The IDs and parameters of the associated sensors involved in the analysis, such as: pressure sensor P-102, flow sensor F-101.

[0066] Severity level: A level based on health status values, such as: Critical, Warning, Caution.

[0067] Key findings: Brief conclusions, such as: The correlation between the temperature sensor T-101 readings and the pressure and flow rate readings deviates significantly from the normal model.

[0068] Multi-sensor data snapshots and comparisons mainly include the following: Time series comparison chart: For the target sensor, curves comparing the actual value series and the model predicted value series are plotted in the same coordinate system, visually demonstrating any discrepancies. For key correlated sensors, reading curves of their main correlated sensors are plotted within the same time period. This provides clear visual evidence of anomalies in the target sensor.

[0069] Fault Feature Sequence Plot: Plots the curve of residuals changing over time and labels its mean, standard deviation and trend line to quantitatively show the abnormal patterns, stability and deterioration trends of the relationship.

[0070] The description of collaborative anomaly patterns mainly includes the following: Summarize the patterns presented in the data snapshot using natural language. For example: Over the past 30 minutes, the reading of the target temperature sensor T-101 has consistently been approximately 3°C higher than the model's predicted value. Meanwhile, the readings of the associated pressure sensor P-102 and flow sensor F-101 fluctuate within normal ranges, and their correlation matches the normal characteristic curves of the pump. Indicate abnormal patterns such as: systematic positive deviations, increased volatility, and deteriorating trends.

[0071] Preliminary root cause inference mainly includes the following: Based on collaborative analysis, the most probable cause of the failure is given. Example 1: High confidence, such as 90% inference that the target temperature sensor T-101 has self-calibration drift. Example 2: Medium confidence, such as 75% inference that the pump P-101 has decreased internal efficiency or slight blockage, causing changes in its thermodynamic properties. Other possible causes can be listed according to confidence level.

[0072] Relevance evidence includes the following: List the key analytical evidence that supports the above inferences.

[0073] Evidence 1: The pressure-flow relationship model is in good health, such as 92%, indicating that the core hydraulic characteristics of the equipment have not changed, ruling out the possibility of a serious pump failure.

[0074] Evidence 2: The residual sequence of the temperature prediction model shows a significant systematic positive bias, such as a mean of +3.1℃, and the trend is stable, which is consistent with the typical characteristics of sensor drift.

[0075] Evidence 3: The vibration sensor data showed anomalies simultaneously, which were strongly correlated with temperature anomalies, both pointing to wear of mechanical parts.

[0076] To facilitate understanding of the technical solution in this embodiment, the following examples are provided: The application scenario is set as follows: Equipment: One centrifugal pump for conveying cooling water; Target sensor: Pump outlet temperature sensor; Related sensors: pump outlet pressure sensor, pump outlet flow sensor, pump motor power sensor; Preset time period: the past 60 minutes.

[0077] Building a Relationship Model: Based on deployment information, the system determines that the pump outlet temperature sensor, pump outlet pressure sensor, pump outlet flow sensor, and pump motor power sensor are all located on the same pump, conforming to the same equipment association rule. Historical data for several months under known normal operating conditions of the pump are collected. Using the historical data from the pump outlet pressure sensor, pump outlet flow sensor, and pump motor power sensor as input features, and the historical data from the pump outlet temperature sensor as the prediction target, a Support Vector Machine Regression (SVR) algorithm is used for training. The trained model yields a relationship model that learns the complex nonlinear relationship between the outlet temperature and pressure, flow rate, and power when the pump is operating normally. When the pump efficiency is normal, a specific combination of pressure, flow rate, and power will produce a expected temperature rise. This model quantifies this expectation.

[0078] Real-time monitoring and calculation of fault characteristic sequences: The system operates in real time, performing a check every 5 minutes. Within the last 60 minutes, the system acquires the actual reading sequences of the pump outlet pressure sensor, pump outlet flow sensor, and pump motor power sensor, inputting these sequences into the aforementioned support vector machine regression model to obtain the predicted value sequence of the pump outlet temperature sensor over the 60 minutes. Simultaneously, the system acquires the actual value sequence of the pump outlet temperature sensor within the same time period, calculates the fault characteristic sequence, and this fault characteristic sequence is the residual sequence.

[0079] Assess the health of the relationship: The system calculates health status based on a 60-minute fault characteristic sequence: Statistical distribution characteristics: The calculated absolute mean of the fault characteristic sequence is +2.8℃, and the standard deviation is 0.3℃. The absolute mean is significantly greater than zero, indicating the existence of a systematic positive bias.

[0080] Trend change characteristics: Linear regression of the residual series revealed a significant positive slope, indicating that the bias is slowly increasing, such as from +2.5℃ 60 minutes ago to +3.1℃ now.

[0081] Pattern consistency characteristics: Calculating the JS divergence between the current residual distribution and the historical normal residual distribution reveals a large divergence value, indicating that the current residual distribution pattern has undergone a fundamental change.

[0082] The system integrates the above features and calculates a comprehensive relationship health index of 0.35.

[0083] The preset alarm threshold is 0.6. Since 0.35 < 0.6, the system triggers an alarm.

[0084] Generate and send relationship anomaly report: The target edge computing device received a trigger command.

[0085] It locally acquires the actual value sequence and residual sequence of the pump outlet temperature sensor over the past 60 minutes.

[0086] It checks and confirms that the pump outlet pressure sensor, pump outlet flow sensor, and pump motor power sensor have complete data cached locally for the same time period.

[0087] The multi-sensor collaborative analysis will begin and a report will be generated. The relationship anomaly report mainly includes the following: Report ID: FDI-20231027-XXX; Trigger time: 2025-XX-XX 14:05:00; Target sensor: Temperature sensor T-101, located at the outlet of pump-A; Key finding: The readings of sensor T-101 show a significant deviation from the correlation between pressure P-101, flow rate F-101, and power W-101. Multi-sensor data snapshots and comparisons: legend omitted; Description of the cooperative anomaly mode: In the past 60 minutes, the readings of the target temperature sensor T-101 have shown a systematic positive deviation, and the deviation value has a slow increasing trend. At the same time, the readings of the associated pressure, flow and power sensors fluctuate normally, and the cooperative relationship model among the three is healthy.

[0088] Preliminary root cause inference: High confidence (90%): The target temperature sensor T-101 itself has experienced calibration drift or performance degradation. Low confidence (10%): There is an extremely minor heat exchange efficiency problem inside the pump that has not yet affected the hydraulic performance.

[0089] Related evidence: Evidence 1: The pressure-flow relationship model shows a healthy level of 92%. This indicates that the pump's core hydraulic characteristics and efficiency have not changed significantly, greatly eliminating the possibility of mechanical failure of the pump itself.

[0090] Evidence 2: The residual sequence of the temperature prediction model shows a stable and increasing systematic positive deviation, such as an absolute mean of +2.8℃. This is a typical characteristic of sensor drift, rather than a transient or drastic change pattern of equipment failure.

[0091] Evidence 3: The relationship between power readings and flow rate and pressure is consistent with normal motor characteristics, ruling out the possibility of temperature rise caused by motor overload.

[0092] Recommended action: It is recommended that temperature sensor T-101 be calibrated or replaced on-site during the next planned downtime window.

[0093] This complete example demonstrates how the entire system learns normal patterns, detects subtle anomalies, performs intelligent diagnosis, and finally generates a decision report. It goes beyond simply alarming for high temperature; it provides a data-driven, analytical, and conclusion-based anomaly report, clearly indicating that the problem likely lies with the sensor itself. This avoids the ineffective work of disassembling and repairing the pump, achieving accurate and efficient predictive maintenance.

[0094] Example 3 Based on Example 1, this example provides an industrial Internet of Things (IoT) sensor fault monitoring system, comprising: a management platform, a sensor network platform, and an object platform that sequentially establish communication. The sensor network platform is configured as follows: Obtain the target sensor and at least one associated sensor, and construct a relationship model based on the target sensor and at least one associated sensor; Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor respectively. The management platform is configured as follows: The relationship health of the target sensor is obtained based on each fault feature sequence. If at least one relationship health is lower than a preset threshold, a relationship anomaly report is generated and sent to the target user.

[0095] The user platform is configured to provide front-end services to users; users obtain the necessary perception service information through the user platform, process the perception service information, and transform it into user perception information; users analyze the user perception information and make corresponding decisions based on their own wishes, and transform the user perception information into user control information through the corresponding information system and send it to the service platform, thereby demonstrating the user's corresponding service needs and wishes.

[0096] The physical entities of the user platform include various user terminals, such as mobile phones, computers, and dedicated terminals, which provide user services through integration with user information system software.

[0097] The service platform is configured as an API server or other server used to establish communication between the management platform and the user platform to achieve corresponding functions; the physical entity of the service platform includes various servers.

[0098] The management platform is configured to perform at least one of the following: device operation status monitoring and management, data monitoring and management, device parameter management, and lifecycle management; the management platform is the overall operation platform for the Internet of Things, which may include various management sub-platforms, with different management sub-platforms performing different management tasks; the physical entities of the management platform include various servers.

[0099] The sensor network platform is configured to perform at least one of the following functions: network management, command management, device status management, data protocol management, data parsing, data classification, data transmission monitoring, and data transmission security management. The sensor network platform provides functions such as data communication, transmission, parsing, identification, and classification, avoiding the direct aggregation of data from various object platforms onto the management platform, which would otherwise result in data redundancy and low data processing efficiency. The physical entities of the object platforms include various gateways, edge computing devices, etc.

[0100] The object platform is configured to perform specific production control, detection, measurement and other production tasks; the physical entities in the object platform include various production equipment, sensors and so on.

[0101] Optionally, the sensor network platform includes a main database that communicates with the management platform and at least two sensor network sub-platforms that communicate with the main database. Optionally, each sensor network sub-platform may correspond to an API function or API server.

[0102] Example 4 This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program that can run on the processor, and the processor executes the computer program to implement any of the methods described above.

[0103] Example 5 This embodiment provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above.

[0104] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for fault monitoring of industrial Internet of Things (IoT) sensors, characterized in that, include: Obtain the target sensor and at least one associated sensor associated with the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor; Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor respectively. Based on each fault feature sequence, the relationship health of the target sensor is obtained. If at least one of the relationship health is lower than a preset threshold, a relationship anomaly report is generated and sent to the target user.

2. The industrial IoT sensor fault monitoring method as described in claim 1, characterized in that, The step of acquiring the target sensor and at least one associated sensor associated with the target sensor includes: Acquire the deployment location information of the target sensor and / or process flow information within a preset time period; Based on the first detection parameter of the target sensor, a second detection parameter associated with the first detection parameter is obtained, and the sensor used to detect the second detection parameter is marked as a candidate sensor. The candidate sensor located in the same equipment and / or the same process flow as the target sensor is marked as an associated sensor.

3. The industrial IoT sensor fault monitoring method as described in claim 1, characterized in that, The step of constructing a relationship model based on the target sensor and at least one associated sensor includes: Historical time-series data of the target sensor and at least one associated sensor under known normal operating conditions are obtained. The historical time-series data of at least one associated sensor is used as input, and the historical time-series data of the target sensor is used as output. A relationship model is obtained by training through a machine learning regression algorithm.

4. The industrial IoT sensor fault monitoring method as described in claim 1, characterized in that, The step of obtaining the relationship health of the target sensor based on the fault feature sequence includes: The statistical characteristics of the fault feature sequence within a preset time period are obtained, and the health of the relationship is characterized by the statistical characteristics. The statistical characteristics include one or more combinations of statistical distribution characteristics, trend change characteristics, and pattern consistency characteristics. The statistical distribution characteristics include calculating the standard deviation and absolute mean of the fault characteristic sequence; The trend-based feature includes fitting the trend of the fault feature sequence over time using linear regression. The pattern consistency feature includes calculating the difference between the distribution of the current fault feature sequence and the distribution of the fault feature sequence during historical normal periods using a divergence index.

5. The industrial IoT sensor fault monitoring method as described in claim 1, characterized in that, The step of generating a relationship anomaly report if at least one of the relationship health scores is below a preset threshold includes: When at least one of the relationship health values ​​is lower than a preset threshold, the target edge computing device acquires the actual value sequence and fault feature sequence of the target sensor within a preset time period; Check whether there is a complete actual reading sequence of associated devices in the target edge computing device within a preset time period. If there is no complete actual reading sequence of associated devices in the target edge computing device within a preset time period, send a request to the associated device to transmit the actual reading sequence of the preset time period. Based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, as well as the complete actual reading sequence of the associated sensor within the preset time period, a relationship anomaly report is generated.

6. The industrial IoT sensor fault monitoring method as described in claim 5, characterized in that, The step of generating a relationship anomaly report based on the actual value sequence and fault characteristic sequence of the target sensor within a preset time period, and the complete actual reading sequence of the associated sensor within the preset time period, includes: Perform horizontal correlation and comparative analysis between the data from the target sensor and all associated sensors: Analyze whether the readings of the associated sensors conform to the expected physical or statistical relationship; The number of associated sensors that identify abnormal readings; Based on the process flow, the propagation path and temporal sequence of anomalies in the sensor network are analyzed.

7. The industrial IoT sensor fault monitoring method as described in claim 1, characterized in that, An anomaly report should include at least one of the following: an anomaly summary, multi-sensor data snapshots and comparisons, description of collaborative anomaly patterns, preliminary root cause inferences, and evidence of correlation.

8. An industrial Internet of Things (IoT) sensor fault monitoring system, characterized in that, include: Establish the communication management platform, sensor network platform, and object platform in sequence: The sensor network platform is configured as follows: Obtain the target sensor and at least one associated sensor associated with the target sensor, and construct a relationship model based on the target sensor and at least one associated sensor; Based on the relationship model and the actual reading sequence of at least one associated sensor in a preset time period, obtain at least one predicted value sequence of the target sensor in the preset time period, and calculate the fault feature sequence between the at least one predicted value sequence of the target sensor and the actual value sequence of the target sensor respectively. The management platform is configured as follows: The relationship health of the target sensor is obtained based on each fault feature sequence. If at least one of the relationship health is lower than a preset threshold, a relationship anomaly report is generated and sent to the target user.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and execute the method of any one of claims 1 to 7.

Images