A Device Fault Detection Method Based on an Instantaneous Learning Local Model

Abstract

The present invention discloses a device fault detection method based on an instant learning local model, which relates to the field of device fault detection and includes: real-time monitoring of multi-dimensional data of a device to establish dynamic fuzzy boundaries for different device operating states; establishing an equal number of local models according to specific fault types in each operating state, and dynamically aggregating multiple local models based on the fuzzy boundaries to form a fault recognition composite model; based on the fault recognition results of the composite model, analyzing the real-time data of the device operation to evaluate the fluctuation of the device operating state; when it is detected that the main operating state of the device fluctuates insignificantly, an unsupervised instant learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model; by establishing local models for specific fault types in each device operating state, combining dynamic division of fuzzy boundaries and dynamic adjustment of fault detection sensitivity, the problems of sensitivity decline and missed detection in traditional methods are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of equipment fault detection, and more particularly to a method for detecting equipment faults based on an instant learning local model. Background Art

[0002] During the operation of industrial production equipment, the reliability and stability of the equipment directly affect production efficiency and safety. Traditional equipment fault detection methods usually rely on preset thresholds or rule sets, lacking sufficient flexibility for fluctuations and changes in the equipment operation state. These methods can only detect in a static operation environment and are difficult to adapt to diverse fault types under complex dynamic conditions of the equipment. Especially when the equipment operation state ages or the fluctuations are not obvious for a long time, the sensitivity of traditional fault detection methods will be significantly reduced, easily leading to missed or misjudged faults.

[0003] To improve the accuracy of fault detection, fuzzy logic technology is adopted to divide the equipment operation state using fuzzy boundaries. Fuzzy boundaries can achieve smooth transitions between different equipment states, reducing false alarms and missed detections in the detection.

[0004] However, simply relying on the division of fuzzy boundaries cannot achieve precise detection of multiple complex faults. For this reason, a fault detection method based on a local model is proposed. By establishing local models for specific fault types under each equipment operation state and combining dynamic division of fuzzy boundaries with dynamic adjustment of fault detection sensitivity, the problems of decreased sensitivity and missed detections in traditional methods are effectively solved, while meeting the diverse detection requirements under the complex environment of industrial equipment. Summary of the Invention

[0005] To solve the above technical problems, a method for detecting equipment faults based on an instant learning local model is provided, and this technical solution solves the problems raised in the above background art.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for detecting equipment faults based on an instant learning local model, comprising:

[0008] Real-time monitoring of multi-dimensional data of the equipment to establish dynamic fuzzy boundaries for different equipment operation states;

[0009] According to the specific fault types under each operation state, establish an equal number of local models, and dynamically aggregate multiple local models based on the fuzzy boundaries to form a fault recognition composite model;

[0010] Based on the fault recognition results of the composite model, evaluate the fluctuation of the equipment operation state by analyzing the real-time data of the equipment operation;

[0011] When it is detected that the fluctuations in the main operating state of the device are not obvious, an unsupervised instant learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model.

[0012] Preferably, the multi-dimensional data of the real-time monitoring device to establish dynamic fuzzy boundaries for different device operating states specifically includes:

[0013] Divide the device operation into multiple operating states, and define a set of fuzzy intervals of operating data for each operating state where S j is the jth operating state;

[0014] Establish a Gaussian membership function relationship based on the correspondence between historical operating data and historical operating states used to represent the possibility that the value of the ith operating parameter X i of the device is in the jth operating state S j ;

[0015] Real-time monitor the operating data of the device, and establish a multi-dimensional operating data time series data set X(t), X(t)=[X1(t), X2(t),…X n (t)], X n (t) represents the operating data of the nth measurement dimension monitored at time t, and n is the total number of measurement dimensions;

[0016] Substitute the monitored real-time data into the membership function Calculate the membership degree of the real-time value of the ith operating parameter X i in the fuzzy set , and select the fuzzy interval boundary with the largest membership degree as the fuzzy boundary B(t) of the current device operating state at time t. Dynamically adjust the fuzzy boundary through multi-dimensional cross-validation. The specific expression for the dynamic adjustment of the fuzzy boundary is:

[0017]

[0018] In the formula, B(t + 1) is the adjusted fuzzy boundary at time t, η is the set adjustment scale, q i is the importance weight of the ith dimension i, X i (t) represents the collected data of the ith measurement dimension monitored at time t, is the standard specification operating parameter value set for the ith measurement dimension in the current state;

[0019] Use the dynamically adjusted fuzzy boundary as the condition for delimiting the fuzzy boundary at the next moment and the basis for judging the category of the real-time operating state of the device.

[0020] Preferably, according to specific fault types in each operating state, an equal number of local models are established, and multiple local models are dynamically aggregated based on fuzzy boundaries to form a fault recognition composite model, which specifically includes:

[0021] Set corresponding fault types for each device operating state, establish an equal number of local models for different fault types, and each local model is used to identify a fixed fault type. The local model is represented by M r where r is the model index, representing the r-th fault type;

[0022] According to the historical fault characteristics of the device, use machine learning algorithms to train each local model, and use the trained local models to identify their corresponding real-time data;

[0023] Dynamically allocate the weights of each local model according to the real-time operating state of the device, specifically:

[0024] Obtain the real-time data recognition results and recognition result confidence levels of the local models;

[0025] Based on the recognition result confidence level and the membership degree of the real-time data in this operating state, comprehensively calculate the confidence interval of the local model at the current moment;

[0026] Dynamically allocate the weights of the local models according to the confidence interval calculation results. The calculation expression for the local model weights is:

[0027]

[0028] where w r (t + 1) is the weight of the local model with model index r at time t + 1, CI r (t) is the confidence interval of the local model with index r at time t, b is the number of local models, and CI a (t) is the confidence interval of the local model with index a at time t;

[0029] According to the calculated local model weights, perform weighted aggregation on all local models to establish a composite model for comprehensive identification of device faults.

[0030] Preferably, based on the fault recognition results of the composite model, by analyzing the real-time operating data of the device, evaluating the fluctuations in the device operating state specifically includes:

[0031] Obtain the recognition results of the composite model. When the recognition result is a device fault, record and generate a detection result report for the corresponding fault;

[0032] When no device failure is detected in the recognition result, select the local model with the largest local model weight in the composite model at the current moment to be responsible for recognizing the operating state of the device corresponding to the failure, mark it as the main operating state, and evaluate the abnormal fluctuation of the main operating state of the device. Specifically:

[0033] Set the time window for evaluating the fluctuation of the main operating state. Obtain the calculation results of the time series membership degree of the main operating state within the time window, and calculate the evaluation value of the fluctuation of the main operating state of the device according to the change of the time series membership degree. The calculation formula is:

[0034]

[0035] In the formula, Ce is the evaluation value of the fluctuation of the main operating state of the device, T is the size of the set time window for evaluating the fluctuation of the main operating state, is the change rate of the membership degree of the main operating state from time t to time t + 1, and δ CF is the peak factor of the membership degree change. The calculation method of the peak factor is the ratio of the maximum change rate of the membership degree to the root mean square of the membership degree change;

[0036] Set the threshold of the evaluation value of the fluctuation of the main operating state. When the evaluation value of the fluctuation of the main operating state of the device is less than or equal to the threshold of the evaluation value of the fluctuation of the main operating state, generate a signal indicating that the fluctuation of the main operating state of the device is not obvious.

[0037] Preferably, when it is monitored that the fluctuation of the main operating state of the device is not obvious, the method of dynamically adjusting the fault detection sensitivity threshold of the local model by using unsupervised online learning specifically includes:

[0038] When a signal indicating that the fluctuation of the main operating state of the device is not obvious is monitored, retrieve all local model sets corresponding to the main operating state and mark them as insensitive local model sets;

[0039] Set the sensitivity adjustment time window, and perform clustering analysis on the historical fault data of the device through unsupervised learning within the sensitivity adjustment time window to calculate the clustering center of each fault, which is used to describe the typical parameter values of the device in each fault state;

[0040] Set the initial fault detection sensitivity threshold for all insensitive local models according to the dispersion degree of the training data. The initial fault detection sensitivity threshold is set based on the maximum distance between the clustering center and the data points of the training data;

[0041] Obtain the real-time operation data of the device, calculate the real-time distance between the current operation data and the clustering center of each insensitive local model, and dynamically adjust the detection sensitivity threshold of the local model. The update formula of the clustering center is:

[0042]

[0043] where tar y (t) is the clustering center of the y-th insensitive local model at time t, and tar(t + 1) is the adjusted detection sensitivity threshold, is the set learning rate, and H(t) is the real-time data at time t within the sensitivity adjustment time window;

[0044] After the sensitivity adjustment time window ends, obtain the historical data of the main operating state fluctuation evaluation time window and the sensitivity adjustment time window, mark it as the window adjustment historical data, and re-identify faults for the window adjustment historical data through the composite model;

[0045] When the recognition result is a device fault, record and generate the corresponding fault detection result report.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] Monitor the multi-dimensional data of the device in real time, and establish dynamic fuzzy boundaries for different device operating states; according to the specific fault types under each operating state, establish an equal number of local models, and dynamically aggregate multiple local models based on the fuzzy boundaries to form a fault recognition composite model; based on the fault recognition results of the composite model, evaluate the fluctuation of the device operating state by analyzing the real-time data of the device operation; when it is detected that the main operating state fluctuation of the device is not obvious, adopt an unsupervised instant learning method to dynamically adjust the fault detection sensitivity threshold of the local model.

[0048] By establishing local models for specific fault types under each device operating state, combining dynamic division of fuzzy boundaries and dynamic adjustment of fault detection sensitivity, the problems of sensitivity decline and missed detection in traditional methods are effectively solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of a device fault detection method based on an instant learning local model of the present invention;

[0050] Figure 2 is a flowchart of monitoring the multi-dimensional data of the device in real time and establishing dynamic fuzzy boundaries for different device operating states of the present invention;

[0051] Figure 3 is a flowchart of establishing an equal number of local models according to the specific fault types under each operating state and dynamically aggregating multiple local models based on the fuzzy boundaries to form a fault recognition composite model of the present invention;

[0052] Figure 4 is a flowchart of evaluating the fluctuation of the device operating state by analyzing the real-time data of the device operation based on the fault recognition results of the composite model of the present invention;

[0053] Figure 5 This is a flow chart of the present invention for dynamically adjusting the fault detection sensitivity threshold of a local model using an unsupervised real-time learning method when it is monitored that the fluctuation of the main operating state of the equipment is not obvious. DETAILED DESCRIPTION

[0054] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0055] Reference Figure 1 As shown, a device fault detection method based on instant learning local model includes:

[0056] Monitor multi-dimensional data of devices in real time and establish dynamic fuzzy boundaries for different device operating states;

[0057] According to the specific fault type under each operating state, an equal number of local models are established, and multiple local models are dynamically aggregated based on fuzzy boundaries to form a composite fault identification model;

[0058] Based on the fault identification results of the composite model, the fluctuation of equipment operation status is evaluated by analyzing the real-time data of equipment operation;

[0059] When it is detected that the fluctuation of the main operating status of the equipment is not obvious, an unsupervised real-time learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model.

[0060] Reference Figure 2 As shown, the multi-dimensional data of the equipment is monitored in real time, and dynamic fuzzy boundaries are established for different equipment operating states.

[0061] Divide the equipment operation into multiple operating states and define a set of operating data fuzzy intervals for each operating state. Among them S j is the jth operating state;

[0062] Establish a Gaussian membership function relationship based on the correspondence between historical operation data and historical operation status Used to represent the i-th operating parameter X of the equipment i The value is in the jth operating state S j The possibility of, where the membership function The expression is:

[0063]

[0064] Where μ is the operating parameter X i The central value of the time series data set, σ 2is the variance of the data set;

[0065] Real-time monitor the operation data of the device, establish a multi-dimensional operation data time series data set X(t), X(t) = [X1(t), X2(t), … X n (t)], X n (t) represents the operation data of the nth measurement dimension monitored at time t, and n is the total number of measurement dimensions;

[0066] Substitute the monitored real-time data into the membership function Calculate the membership degree of the real-time value of the ith operation parameter X i in the fuzzy set Select the boundary of the fuzzy interval with the largest membership degree as the fuzzy boundary B(t) of the current device operation state at time t, and dynamically adjust the fuzzy boundary through multi-dimensional cross-validation joint. The specific expression of the fuzzy boundary dynamic adjustment is:

[0067]

[0068] In the formula, B(t + 1) is the adjusted fuzzy boundary at time t, η is the set adjustment scale, q i is the importance weight of the ith dimension i, X i (t) represents the collected data of the ith measurement dimension monitored at time t, is the standard specification operation parameter value set for the ith measurement dimension in the current state;

[0069] Take the dynamically adjusted fuzzy boundary as the condition for delimiting the fuzzy boundary at the next moment and the judgment basis for the category of the device's real-time operation state. In addition, the fuzzy boundary can also establish a flexible and highly adaptable composite model for different operation states and fault types of the device.

[0070] Refer to Figure 3 As shown, according to the specific fault types in each operation state, establish an equal number of local models, and dynamically aggregate multiple local models based on the fuzzy boundary to form a fault identification composite model.

[0071] Each local model is constructed according to a specific fault type and realizes real-time aggregation through membership degree and dynamic weight allocation, so that the composite model can adapt to the real-time state changes of the device. Combining the input of real-time data, the composite model can provide high-precision fault detection and diagnosis results in a variety of complex environments.

[0072] Set the corresponding fault types for each device operation state, establish an equal number of local models for different fault types, each local model is used to identify a fixed fault type, and the local model is denoted as M rIt is indicated that r is the model index, representing the r-th type of fault;

[0073] Based on the historical fault characteristics of the equipment, machine learning algorithms are used to train each local model, and the trained local models are used to identify their corresponding real-time data;

[0074] The weights of each local model are dynamically allocated according to the real-time operating state of the equipment, specifically:

[0075] Obtain the real-time data recognition results and recognition result confidence levels of the local models;

[0076] Based on the recognition result confidence level and the membership degree of the real-time data under this operating state, comprehensively calculate the confidence interval of the local model at the current moment;

[0077] Dynamically allocate the weights of the local models according to the calculated confidence interval results. The calculation expression for the local model weights is:

[0078]

[0079] In the formula, w r (t + 1) is the weight of the local model with model index r at time t + 1, CI r (t) is the confidence interval of the local model with index r at time t, b is the number of local models, CI a (t) is the confidence interval of the local model with index a at time t;

[0080] According to the calculated local model weights, all local models are weighted and aggregated to establish a composite model for comprehensive identification of equipment faults. The formation formula of the composite model is:

[0081]

[0082] M mix is the composite model after weighted aggregation. The composite model can be adaptively adjusted according to the real-time operating state of the equipment. It includes the fault detection capabilities of all local models. Through the method of dynamic weight aggregation, the model most suitable for the current state contributes the most to fault detection.

[0083] Refer to Figure 4 As shown, based on the fault recognition results of the composite model, by analyzing the real-time operating data of the equipment, evaluate the fluctuation of the equipment operating state.

[0084] Although the local model-based method has improved the fault detection ability to a certain extent, in the case of insignificant fluctuations in the device state or device aging, the fault detection sensitivity of the local model is prone to decline, resulting in a weakened ability to detect potential faults. At this time, the operating parameters of the device may slowly deviate from the normal state but do not reach the preset fault threshold, and traditional detection models are difficult to capture these changes in a timely manner.

[0085] Obtain the recognition result of the composite model. When the recognition result is a device fault, record and generate a detection result report corresponding to the fault.

[0086] When the recognition result does not detect a device fault, select the local model with the largest local model weight in the composite model at the current moment to be responsible for recognizing the operating state of the device corresponding to the fault, mark it as the main operating state, and evaluate the abnormal fluctuations in the main operating state of the device. Specifically:

[0087] Set the time window for evaluating the fluctuations in the main operating state. Obtain the calculation results of the time series membership degrees of the main operating state within the time window, and calculate the evaluation value of the fluctuations in the main operating state of the device according to the change of the time series membership degrees. The calculation expression is:

[0088]

[0089] In the formula, Ce is the evaluation value of the fluctuations in the main operating state of the device, T is the size of the set time window for evaluating the fluctuations in the main operating state, is the change rate of the membership degree of the main operating state from time t to time t + 1, and δ CF is the peak factor of the membership degree change. The calculation method of the peak factor is the ratio of the maximum change rate of the membership degree to the root mean square of the membership degree change;

[0090] Set the threshold of the evaluation value of the fluctuations in the main operating state. When the evaluation value of the fluctuations in the main operating state of the device is less than or equal to the threshold of the evaluation value of the fluctuations in the main operating state, generate a signal indicating that the fluctuations in the main operating state of the device are not obvious.

[0091] Refer to Figure 5 As shown, when it is detected that the fluctuations in the main operating state of the device are not obvious, an unsupervised online learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model.

[0092] As the operating time of the device increases, the operating state and parameters of the device may change. To maintain the accuracy of the local model, the system continuously updates the local model based on the online learning method. Whenever new data is collected, through the online learning method of the unsupervised local model, combined with real-time data analysis and dynamic threshold adjustment, efficient fault detection under the device operating state is achieved, enabling the system to dynamically respond to changes during device operation. Especially in the case of device aging or insignificant state fluctuations, the system can still maintain a high-sensitivity fault identification ability.

[0093] When a signal indicating insignificant fluctuations in the main operating state of the device is detected, retrieve all local model sets corresponding to this main operating state and mark them as insensitive local model sets;

[0094] Set a sensitivity adjustment time window. Within the sensitivity adjustment time window, perform clustering analysis on the historical fault data of the device through the unsupervised learning method, and calculate the clustering center of each fault through the K-means clustering algorithm to describe the typical parameter values of the device in each fault state;

[0095] Set an initial fault detection sensitivity threshold for all insensitive local models according to the dispersion degree of the training data. The initial fault detection sensitivity threshold is set based on the maximum distance (Euclidean distance) between the clustering center and the data points of the training data;

[0096] Obtain the real-time operating data of the device, calculate the real-time distance (Euclidean distance) between the current operating data and the clustering center of each insensitive local model, and dynamically adjust the local model detection sensitivity threshold. The update formula for the clustering center is:

[0097]

[0098] In the formula, tar y (t) is the clustering center of the yth insensitive local model at time t, and tar(t + 1) is the adjusted detection sensitivity threshold, is the set learning rate, and H(t) is the real-time data at time t within the sensitivity adjustment time window;

[0099] Assume that the operating characteristics of the device change slowly, causing its parameter values to gradually deviate from the original clustering center. The system then automatically reduces the fault detection threshold to capture device anomalies more sensitively.

[0100] After the sensitivity adjustment time window ends, obtain the historical data of the main operating state fluctuation evaluation time window and the sensitivity adjustment time window, mark it as window adjustment historical data, and perform re-fault identification on the window adjustment historical data through the composite model;

[0101] Re-identifying and detecting the historical data in the model adjustment stage is to prevent missed detections or false fault judgments caused by equipment model adjustments, improve the accuracy and integrity of the detection results. By analyzing the equipment status data and fault detection trends, it can help enterprises predict potential problems, formulate preventive maintenance plans, and reduce the risk of sudden equipment failures.

[0102] When the recognition result is an equipment fault, record and generate a corresponding fault detection result report, systematically record the operating conditions of the equipment, the fault detection process and results, help managers comprehensively understand the current operating health of the equipment, provide an accurate analysis of equipment faults, provide data support for equipment maintenance, update and operation strategies, and help the management make timely and scientific decisions.

[0103] Furthermore, this solution also proposes a storage medium for an equipment fault detection method based on an instant learning local model, on which a computer-readable program is stored. When the computer-readable program is called, it executes the above-mentioned equipment fault detection method based on an instant learning local model.

[0104] It can be understood that the storage medium can be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid-state drive (SSD).

[0105] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A device fault detection method based on real-time learning local model, characterized in that: include: Monitor multi-dimensional data of devices in real time and establish dynamic fuzzy boundaries for different device operating states; According to the specific fault type under each operating state, an equal number of local models are established, and multiple local models are dynamically aggregated based on fuzzy boundaries to form a composite fault identification model; Based on the fault identification results of the composite model, the fluctuation of equipment operation status is evaluated by analyzing the real-time data of equipment operation; When the fluctuation of the main operating status of the equipment is not obvious, the unsupervised real-time learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model; The real-time monitoring of multi-dimensional data of the equipment to establish dynamic fuzzy boundaries for different equipment operating states specifically includes: Divide the equipment operation into multiple operating states and define a set of operating data fuzzy intervals for each operating state. ,in is the jth operating state; Establish a Gaussian membership function relationship based on the correspondence between historical operation data and historical operation status , used to represent the i-th operating parameter of the device The value of is in the jth operating state possibility; Monitor the equipment's operating data in real time and establish a multi-dimensional operating data time series data set , , Represents the operating data of the nth measurement dimension monitored at time t, where n is the total number of measurement dimensions; Bring the monitored real-time data into the membership function , calculate the i-th operating parameter The real-time value of the fuzzy set The fuzzy interval boundary with the largest membership degree is selected as the fuzzy boundary of the current equipment operation state at time t. , the fuzzy boundary is dynamically adjusted through multi-dimensional cross-validation. The specific expression of dynamic adjustment of the fuzzy boundary is: Where, is the fuzzy boundary adjusted at time t, To set the adjustment scale, is the importance weight of the i-th dimension i, represents the collected data of the i-th measurement dimension monitored at time t, The standard operating parameter value set for the i-th measurement dimension in the current state; The dynamically adjusted fuzzy boundary is used as the basis for determining the fuzzy boundary demarcation condition at the next moment and the real-time operating status category of the equipment.

2. The device fault detection method based on real-time learning local model according to claim 1 is characterized in that: The method of establishing an equal number of local models according to the specific fault type in each operating state and dynamically aggregating multiple local models based on fuzzy boundaries to form a composite fault identification model specifically includes: Set the corresponding fault type for each equipment operation state, and establish an equal number of local models for different fault types. Each local model is used to identify a fixed fault type. The local model is used to Indicates that r is the model index, indicating the rth fault type; Based on the historical fault characteristics of the equipment, a machine learning algorithm is used to train each local model, and the trained local model is used to identify the corresponding real-time data; The weights of each local model are dynamically allocated according to the real-time operating status of the device, specifically: Obtain real-time data recognition results and recognition result confidence of the local model; Based on the confidence of the recognition result and the membership of the real-time data in the operating state, the confidence interval of the local model at the current moment is comprehensively calculated; The local model weight is dynamically allocated according to the confidence interval calculation results. The local model weight calculation expression is: Where, is the weight of the local model with model index r at time t+1, is the confidence interval of the local model with index r at time t, b is the number of local models, is the confidence interval of the local model with index a at time t; According to the calculated local model weights, all local models are weighted and aggregated to establish a composite model for comprehensive identification of equipment faults.

3. The device fault detection method based on real-time learning local model according to claim 2 is characterized in that: The fault identification results based on the composite model, by analyzing the real-time data of equipment operation and evaluating the fluctuation of equipment operation status, specifically include: Obtain the composite model identification results. If the identification result is a device fault, record and generate a corresponding fault detection result report. If the identification result does not detect an equipment fault, the local model with the largest weight in the composite model at the current moment is selected to identify the equipment operating state corresponding to the fault, mark it as the main operating state, and perform an abnormal evaluation of the main operating state fluctuation of the equipment, specifically: Set the main operating status fluctuation assessment time window, obtain the time series membership calculation results of the main operating status within the time window, and calculate the main operating status fluctuation assessment value of the equipment according to the change of the time series membership. The calculation expression is: ； Where, is the fluctuation assessment value of the main operating status of the equipment, The size of the time window for evaluating the main operating status fluctuations is set. is the membership change rate of the main operating state from time t to time t+1, is the membership change peak factor, which is calculated as the ratio of the maximum membership change rate to the root mean square of the membership change; A main operation state fluctuation evaluation value threshold is set. When the main operation state fluctuation evaluation value of the equipment is less than or equal to the main operation state fluctuation evaluation value threshold, a main operation state fluctuation insignificant signal of the equipment is generated.

4. The device fault detection method based on real-time learning local model according to claim 3 is characterized in that: When it is detected that the fluctuation of the main operating state of the equipment is not obvious, the unsupervised real-time learning method is used to dynamically adjust the fault detection sensitivity threshold of the local model, specifically including: When a signal indicating that the main operating state of the equipment is not fluctuating significantly is detected, all local model sets corresponding to the main operating state are retrieved and marked as insensitive local model sets; Set a sensitivity adjustment time window, perform cluster analysis on the device's historical fault data within the sensitivity adjustment time window using unsupervised learning methods, and calculate the cluster center of each fault to describe the typical parameter values of the device in each fault state; An initial fault detection sensitivity threshold is set for all insensitive local models according to the degree of discreteness of the training data. The initial fault detection sensitivity threshold is set based on the maximum distance between the cluster center and the data point of the training data. Obtain the real-time operating data of the device, calculate the real-time distance between the current operating data and the cluster center of each insensitive local model, and dynamically adjust the local model detection sensitivity threshold. The cluster center update formula is: Where, is the cluster center of the y-th insensitive local model at time t, Adjusted detection sensitivity threshold, To set the learning rate, Real-time data at time t within the sensitivity adjustment time window; After the sensitivity adjustment time window ends, the historical data of the main operating status fluctuation assessment time window and the sensitivity adjustment time window are obtained and marked as window adjustment historical data. The window adjustment historical data is re-identified for faults using the composite model; When the identification result is a device failure, a corresponding fault detection result report is recorded and generated.

Images