A method and system for predicting electrical equipment failures

Abstract

The present invention relates to a method and system for predicting electrical equipment failures, which improves the accuracy and timeliness of failure prediction through dynamic optimization of models. First, the device identifier, real-time data information, and historical prediction errors of the target electrical equipment are obtained. The target device category is determined based on the device identifier, and the corresponding target optimization module is called to dynamically optimize the preset initial fusion model, and the target prediction fusion model is output. The method for determining the initial fusion model includes classifying electrical equipment, collecting historical data and performing feature extraction, selecting a matching basic prediction model for training and calculating error metrics, and finally selecting the optimal combination method for application in actual failure prediction, which can provide strong support for the maintenance and management of equipment. The present invention also provides an electrical equipment failure prediction system, which provides a comprehensive, accurate, and real-time solution for the failure prediction of electrical equipment through efficient data management and model optimization strategies.

Description

Technical Field

[0001] This application relates to the technical field of fault prediction, and particularly to a method and system for predicting faults of electrical equipment. Background Art

[0002] There are various types of electrical equipment in the robotic manipulator. Different types of electrical equipment, such as sensors, motors, and power systems, have significantly different fault characteristics and prediction requirements due to differences in their working principles, operating environments, and usage scenarios. Even for the same type of electrical equipment, such as temperature sensors and position sensors, there are obvious differences in their fault modes, data characteristics, and requirements for prediction accuracy. This diversity and complexity make the fault prediction of electrical equipment a challenging technical problem.

[0003] When the existing technologies perform fault prediction on various types of electrical equipment, they usually adopt a separate monitoring method, that is, a prediction model is constructed separately for each device. Although this method can perform customized prediction for the fault characteristics of specific devices, it also leads to a sharp increase in the number of models, which not only increases the complexity and maintenance cost of the system, but also makes it difficult to achieve collaborative optimization between devices and comprehensive evaluation of the overall health status. Summary of the Invention

[0004] The inventors of this application found that most of the existing fault prediction models rely on historical data for training and prediction, and the data sources are usually relatively single, resulting in poor timeliness and accuracy of the models. Real-time data can reflect the changes in the current operating state of the device, which is of great significance for capturing sudden faults and dynamically adjusting the prediction model. However, in the process of model update and optimization, the existing methods usually adopt a fixed model structure or a static weight allocation strategy, lacking the ability to dynamically respond to real-time data, thus limiting the prediction accuracy and adaptability of the model. The fixed model design is difficult to adapt to the dynamic changes in the operating state of the robotic manipulator. Especially in the high-load and high-precision operation scenarios of the manipulator, the operating conditions and workload of the device may change at any time, further reducing the reliability and accuracy of the prediction model. And because the association between data may be ignored or the association between data is not clear, the application of the model is limited in terms of generality.

[0005] According to the first aspect of the present invention, there is provided a method for predicting faults of electrical equipment, which is used to predict faults of the electrical equipment of a robotic manipulator, and includes the following steps:

[0006] Obtain the device identifier, real-time data information X, and historical prediction error H(X) of the target electrical equipment;

[0007] Determine the target device category according to the device identifier, call the target optimization module corresponding to the target device category, and enable the target optimization module to dynamically optimize the preset initial fusion model F based on the real-time data information X and the historical prediction error H(X) according to the following formula base (X) to output the target prediction fusion model F opt (X), F opt (X)=F base (X)+(γ×e -β|H(X)| +λ s )×W s ×X;

[0008] where γ and β are both coefficients, λ s represents the adaptive correction term, and W s represents the weight parameter of the target electrical device, and one target optimization module corresponds to one device category;

[0009] Based on the device identifier and the real-time data information, make the target prediction fusion model output the prediction result of the target electrical device.

[0010] Optionally, the method for obtaining the adaptive correction term λ s includes the following steps:

[0011] Obtain the historical prediction errors of all electrical devices in the target device category and the preset basic correction term λ s (0) ;

[0012] Calculate the average error H s ,

[0013] ;

[0014] where N represents the number of electrical devices in the target device category, and H i represents the historical prediction error of the i-th electrical device in the target device category;

[0015] Define the influence factor α s of the target device category and the influence factor of the target electrical device, α s =e -k|Hs| , , and k and p both represent adjustment coefficients;

[0016] Calculate the adaptive correction term λ according to the formula s .

[0017] Optionally, the adjustment coefficient k is obtained according to the following formula:

[0018] ;

[0019] ;

[0020] The adjustment coefficient p is obtained according to the following formula:

[0021] ;

[0022] ;

[0023] where k0 and p0 are both initial adjustment coefficients, and t k and t p are both adjustment parameters.

[0024] Optionally, the method for obtaining the historical prediction error H(X) of the target electrical equipment includes the following steps:

[0025] Obtain the prediction error H t-1 (X) of the target electrical equipment at time t-1, the prediction error H t-2 (X) at time t-2, and the mean value of the prediction errors H m (X) at M moments before time t-1;

[0026] Calculate the first historical prediction error H EWMA,t-1 (X) based on the weighted moving average method according to the following formula,

[0027] ;

[0028] where q represents the smoothing factor, and 0 < q < 1;

[0029] Calculate the second historical prediction error T t (X) according to the following formula,

[0030] T t-1 (X)=r×(H t-1 (X)-H t-2 (X));

[0031] where r represents the trend adjustment coefficient, and 0.5 < r < 1.5;

[0032] Calculate the historical prediction error H(X) according to the formula H(X)=θ1×H EWMA,t-1 (X)+θ2×T t-1 (X), where θ1 and θ2 are both constants, and the sum of the two is equal to 1.

[0033] Optionally, the method for determining the initial fusion model F base (X) includes the following steps:

[0034] Classify the electrical devices of the robotic arm to obtain multiple device categories;

[0035] For each device category, collect historical data and perform feature extraction;

[0036] Select at least two basic prediction models that match the feature type from the model database according to the extracted feature type;

[0037] Use the historical data of the device category to train the selected basic prediction models, calculate the error metrics, and select the optimal combination method for the device category as the initial fusion model F base (X).

[0038] Optionally, the step of using the historical data of the device category to train the selected basic prediction models, calculate the error metrics, and select the optimal combination method for the device category as the initial fusion model F base (X) includes the following steps:

[0039] Segment the historical data of each device category by time period, and use the historical data of each time period to train the selected basic models;

[0040] Calculate the error of each basic model in each time period according to the following formula

[0041] ;

[0042] where represents the error of the i-th basic model f i in the time period T i , represents the number of samples in the time period T i , represents the actual value, represents the predicted value;

[0043] For each time period, select the basic model with the smallest error as the optimal model in that time period;

[0044] Statistically calculate the frequency of each basic model being selected as the optimal model in all time periods according to the following formula ,

[0045] ;

[0046] where represents the j-th basic model is the optimal model within time period T i inside; represents an indicator function;

[0047] Select at least two base models with the highest frequency for fusion as the target base model. If multiple base models have the same frequency, select at least two base models with the smallest error as the target base model;

[0048] Perform weighted fusion on the selected target base models to obtain the initial fusion model F base (X).

[0049] Optionally, the performing weighted fusion on the selected target base models to obtain the initial fusion model F base (X) includes the following steps:

[0050] Calculate the average error of each target base model over all time periods;

[0051] Calculate the weight of the i-th base model according to the target base model and the error using the following formula ; ,

[0052] ;

[0053] where S represents the set of target base models, ε represents the adjustment coefficient, represents the average error of the i-th base model ;

[0054] Obtain the formula for the initial fusion model F base (X),

[0055] ;

[0056] where , and k represents the number of selected target base models.

[0057] Optionally, after the formula for the initial fusion model F base (X) is determined, the following steps are further included:

[0058] Calculate the average error of each base model on the latest historical data at every preset time interval;

[0059] After the average error of the latest historical data exceeds the threshold error, trigger model update to re-select the optimal base model or adjust the weights of the target base models, so as to obtain an updated initial fusion model.

[0060] Optionally, for different device categories, the corresponding initial fusion model F baseThe weight parameter W in (X) s satisfies the following formula:

[0061] ;

[0062] where, W s (0) represents the initial weight parameter corresponding to the device category.

[0063] According to the second aspect of the present invention, there is provided an electrical equipment fault prediction system, including at least one processor, and the at least one processor is configured to execute a computer program or instruction to perform operations corresponding to the electrical equipment fault prediction method as described above.

[0064] According to the solution of the present application, by obtaining the device identifier, real-time data information X, and historical prediction error H(X) of the target electrical equipment, and calling the target optimization module corresponding to the target device category based on the device identifier, the initial fusion model F base (X) is dynamically optimized, and the target prediction fusion model F opt (X) is output, thereby solving the problems in the prior art that the fault prediction model overly relies on historical data, the utilization of real-time data is insufficient, and the fixed model leads to poor prediction accuracy and timeliness. Moreover, for different device categories (such as sensors, motors, power systems, etc.), the embodiments of the present invention extract a general formula for the prediction fusion model F opt (X) of the key electrical equipment applicable to the robotic manipulator. This formula can adapt to the fault characteristics and prediction requirements of different device categories by dynamically adjusting the adaptive correction term λ s and the weight parameter W s , thereby significantly improving the prediction accuracy and reliability while maintaining the model consistency. In addition, the unified prediction fusion model F opt (X) can not only avoid separately constructing and maintaining multiple models for each device category, improve the generality and scalability of the model, but also continuously adjust the model, correct the model with real-time data assistance, so that its accuracy is continuously improved, and at the same time, it can reduce the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a schematic flowchart of the electrical equipment fault prediction method according to an embodiment of the present invention;

[0066] Figure 2 is a schematic flowchart of the method for determining the initial fusion model in the fault prediction method of the present invention;

[0067] Figure 3It is a schematic flowchart of the specific acquisition method of the initial fusion model in the fault prediction method of the present invention. Specific embodiments

[0068] To make the above objects, features, and advantages of the present application more obvious and understandable, the following will describe in detail the specific embodiments of the present application with reference to the accompanying drawings. It can be understood that the specific embodiments described herein are only used to explain the present application, rather than limiting the present application. Additionally, it should be noted that for the sake of description, only the parts related to the present application are shown in the drawings, not all the structures. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0069] The terms "including" and "having" and any variations thereof in the present application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these methods, products, or devices.

[0070] Referring to "embodiments" herein means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The phrase appears at various positions in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0071] The embodiments of the present invention provide a method for predicting faults of electrical equipment, which is used to predict faults of the electrical equipment of a robotic arm. The electrical equipment is a key electrical equipment of the robotic arm and includes multiple categories, such as sensor category, motor category, controller category, power system category, etc. Each category includes multiple electrical equipment. The sensor category can include, for example, temperature sensors, position sensors, force sensors, and vision sensors. The motor category can include, for example, servo motors, stepper motors, and DC motors. The controller category can include, for example, logic controllers, microprogram controllers, etc. The power system category can include, for example, power modules, battery packs, and voltage regulators.

[0072] Before this application, it was generally believed in the prior art that due to the differences in the categories of electrical equipment of robotic arms and the complexity of data, it was impossible to design a general prediction model to meet the needs of all equipment categories or all electrical equipment. This technical prejudice led to the existing methods usually adopting a separate monitoring method, that is, constructing a fault prediction model separately for each piece of equipment. Based on such a technical prejudice, the inventors of this application have tried various ways to solve the fault prediction problem of the electrical equipment of robotic arms, but none of them have achieved ideal results. The following are the main directions tried by the inventors under this technical prejudice and the problems existing therein. For example, on the basis of the existing model, increasing the feature dimension and adding real-time data to improve the timeliness and accuracy of the prediction results. Although this method improves the prediction accuracy to a certain extent, it cannot fundamentally solve the problems of model immobilization, model complexity, and management difficulties. Another example is that the inventors tried to divide the same type of electrical equipment into the same equipment category (for example, dividing temperature sensors, position sensors, etc. into the sensor category, and dividing servo motors, stepper motors, etc. into the motor category), so as to divide the electrical equipment of the robotic arm into multiple equipment categories. For each equipment category, a separate integrated prediction model is set, and these prediction models are independent of each other, without relevance and generality. This method reduces the calculation cost to a certain extent and improves the pertinence of the prediction model. However, this division method still fails to fundamentally solve the problems of a large number of models, high system complexity, and scalability of other electrical categories.

[0073] In the process of exploring solutions to the above problems, the inventors tried to use a general model to meet the fault prediction requirements of multiple electrical equipment. After verification, this method significantly improved the prediction performance, and the general model can take into account the characteristics of different types of electrical equipment, improve the scalability and maintenance efficiency of the system, reduce the requirement of equipping a single equipment model for different types of equipment, and continuously use the data obtained during operation for model optimization, improving generality while also improving accuracy. Thus, the solution of the embodiment of the present invention is proposed. Figure 1 FIG. shows a schematic flowchart of an electrical equipment fault prediction method according to an embodiment of the present invention. As Figure 1 shown, the electrical equipment fault prediction method includes:

[0074] Step S100, obtaining the equipment identifier, real-time data information X, and historical prediction error H(X) of the target electrical equipment;

[0075] Step S200, determining the target equipment category according to the equipment identifier, calling the target optimization module corresponding to the target equipment category, and enabling the target optimization module to optimize the preset initial integrated model F based on the real-time data information X and the historical prediction error H(X) according to the following formula base(X) is dynamically optimized to output the target prediction fusion model F opt (X),

[0076] F opt (X) = F base (X) + (γ × e -β|H(X)| + λ s ) × W s × X;

[0077] where γ and β are both coefficients, λ s represents the adaptive correction term, and W s represents the weight parameter of the target electrical device, and one target optimization module corresponds to one device category;

[0078] Step S300, based on the device identifier and real-time data information, enables the target prediction fusion model to output the fault prediction result of the target electrical device.

[0079] According to the solution of the embodiment of the present invention, by obtaining the device identifier, real-time data information X and historical prediction error H(X) of the target electrical device, and calling the target optimization module corresponding to the target device category based on the device identifier, the initial fusion model F base (X) is dynamically adjusted to output the target prediction fusion model F opt (X), thus solving the problems in the prior art that the fault prediction model overly relies on historical data, insufficient utilization of real-time data, and poor prediction accuracy and timeliness caused by model immobilization. Moreover, for different device categories (such as sensors, motors, controllers, power systems, etc.), the embodiment of the present invention extracts the general formula of the prediction fusion model F opt (X) applicable to the key electrical devices of the robot manipulator. This formula can adapt to the fault characteristics and prediction requirements of different device categories by dynamically adjusting the adaptive correction term λ s and the weight parameter W s , thereby significantly improving the prediction accuracy and reliability while maintaining model consistency. In addition, the unified prediction fusion model F opt (X) can avoid separately constructing and maintaining multiple models for each device category.

[0080] In this step S100, the device identifier, real-time data information X, and historical prediction error H(X) of the target electrical device are obtained from the storage query database. The device identifier is a code or name used to uniquely identify the target electrical device. In the storage query database, the device identifier is usually associated with the basic information of the device, such as model, manufacturer, installation location, etc. The way to obtain the device identifier can be, for example, to directly query the corresponding device identifier in the storage query database by inputting specific information of the device, such as serial number, model code, etc., or to select the target device from the device list in the storage query database, and the system automatically extracts its device identifier.

[0081] The real-time data information X refers to the real-time changing data such as the current working state and operating parameters of the target electrical device. These data are usually collected in real time through sensors, monitoring devices, etc., or can also be the data during the interaction between electrical devices. These are stored in the storage query database. The way to obtain the real-time data information X can be the synchronous mode of the storage query database. The storage query database system is synchronized with the monitoring system of the electrical device in real time to ensure that the real-time data information X in the storage query database is the latest. In addition, it can also be through API interface calls. By calling the API interface provided by the storage query database, the model can obtain the real-time data information X of the target electrical device in real time. Other ways can also be to set up a scheduled task to regularly query and update the real-time data information X of the target electrical device from the storage query database.

[0082] The historical prediction error H(X) refers to the error records generated during past fault predictions of the target electrical device. These error data are crucial for optimizing the prediction model and improving the prediction accuracy. During fault prediction, real-time data is obtained, and the next working condition is predicted through the model. Specifically, it can be the predicted data of the next working condition. The difference between the predicted data of the next working condition and the actual data is judged, and the fault is predicted by combining the data change trend. The difference between each prediction and the actual data is recorded as the prediction error and used as the historical prediction error. The real-time data information X is updated in real time to ensure that the prediction model makes predictions based on the latest data. At the same time, the historical prediction error H(X) can be set to be updated regularly or in real time as needed to reflect the changes in the model performance.

[0083] In one embodiment, the storage query database is constructed using MySQL technology. The construction method of this storage query database includes:

[0084] Step 1: Design the data model.

[0085] The data model includes a device information table, a real-time data table, a historical prediction error table, and a model parameter table. The device information table stores the basic information of the devices, including device identifiers, device categories, and electrical devices, etc. Among them, the device categories can be, for example, sensor categories, motor categories, and power system categories, etc. The electrical devices are independent electrical devices under each device category, such as temperature sensors and position sensors, etc. The real-time data table stores the real-time data of the devices, such as sensor readings, motor currents, etc., and supports fast writing and querying. The historical prediction error table stores the historical prediction errors of the devices. The model parameter table stores the parameters of the prediction model, such as the weight parameter W s , the adaptive correction term λ s .

[0086] Step two, partition or shard the data.

[0087] Data partitioning is to divide a large table into multiple smaller sub-tables (i.e., partitions) according to certain rules such as time, region, etc. Each partition can be stored and queried independently. In one embodiment, partition the time data. For example, partition the real-time data table and the historical prediction error table by time, such as by day or by month. Thus, the amount of data for a single query can be reduced. That is, when querying the data for a certain time period, only the corresponding partition needs to be scanned instead of the entire table. And after partitioning, the storage query database can process the query requests of multiple partitions in parallel.

[0088] Data sharding is to distribute the data of a large table to multiple physical storage nodes, and each node only stores part of the data. In one embodiment, shard the device data. For example, shard and store the data according to the device identifier, and distribute the data of different devices to different physical storage nodes. Thus, the query concurrency ability can be improved. That is, the queries of different devices can be distributed to different nodes, avoiding a single node becoming a performance bottleneck. And it can also support horizontal expansion. That is, by adding storage nodes, the storage and computing capabilities of the storage query database can be easily expanded.

[0089] Step three, adopt a hierarchical storage architecture and achieve dynamic migration between hot data and cold data through an automated policy.

[0090] Store high-frequency query data such as the latest real-time data and historical prediction errors of electrical equipment in Redis, while store cold data such as historical data in MySQL. Through automation strategies such as the TTL mechanism and cache warming, achieve dynamic migration between hot data and cold data, ensure that data with high-frequency access can be quickly responded to, and at the same time reduce the pressure on persistent storage. In this architecture, Redis serves as an in-memory storage query database, responsible for storing and quickly accessing hot data such as the latest real-time data, while MySQL serves as a disk storage query database for storing cold data such as historical data. Through a reasonable data migration strategy, the system can optimize the use of storage resources while ensuring query efficiency.

[0091] Step four, perform performance optimization design on the storage query database.

[0092] Performing performance optimization design on the storage query database includes index optimization, query caching, and batch writing and asynchronous processing. In the index optimization design, create composite indexes for high-frequency query fields to accelerate queries, regularly analyze query logs, dynamically adjust the index strategy, and delete indexes for low-frequency query fields to reduce storage overhead. In the query caching design, use Redis to cache high-frequency query results, such as the latest real-time data and historical prediction errors of devices, and set a reasonable cache expiration time to ensure the timeliness of cached data. In the batch writing and asynchronous processing design, adopt batch processing for the writing operations of real-time data to reduce the writing frequency of the storage query database, and asynchronize the update operations of historical prediction errors to avoid blocking the writing of real-time data.

[0093] When setting up the database, adopt a dual-database system, including a cloud database and a local database. The cloud database is used to collect data of scattered electrical equipment, and the local database can interact with the cloud database. In this way, local electrical equipment can call the data of other electrical equipment. When building and optimizing the model, the large amount of data in the cloud database can ensure the accuracy of the model, while the data in the local database can make the adjustment of the specific electrical equipment model by the target optimization module more precise. Through this method, a large amount of historical data of the same type of electrical equipment can be borrowed, so that the initial fusion model can have a relatively high compliance at the beginning. When the number of specific electrical devices used is large, the unit usage cost is greatly reduced.

[0094] In step S200, each device category has different operating characteristics and failure modes. Therefore, it is necessary to adjust and optimize the prediction model for different device categories so that the model can more accurately reflect the devices of that category, that is, add some specific adjustments to the general model. Each device category corresponds to a target optimization module, which is responsible for dynamically optimizing the prediction model of that category of devices. The target optimization module adjusts model parameters such as γ, β, , W s and the corresponding initial fusion model F base (X) to adapt to the operating states of different devices. However, regardless of the device category, the same formula can be used to output the target prediction fusion model F opt (X).

[0095] It can be understood that the initial fusion model F base (X) is a basic prediction model trained based on historical data. The F base (X) of different device categories are usually different because different categories of devices have different operating characteristics. γ×e -β|H(X)| is a dynamic correction term based on the historical prediction error H(X). γ and β are used to control the amplitude and decay rate of the correction term. e -β|H(X)| indicates that the influence of the error on the correction term decreases as the absolute value of the error increases. The adaptive correction term λ s can be dynamically adjusted according to the characteristics of the device category and the changes in real-time data, and is used to further adjust the model output. By introducing the historical prediction error H(X) and real-time data information X, the model output is dynamically adjusted so that the model can adapt to the changes in the device state. Different initial fusion models F base (X) and optimization parameters are designed for different device categories to improve the pertinence of the model. The bias of the model is corrected using the historical prediction error H(X) to ensure that the model is more accurate in future predictions.

[0096] Figure 2 FIG. shows a schematic flowchart of a method for determining the initial fusion model F base (X) according to an embodiment of the present invention. As Figure 2 shown, the method for determining the initial fusion model F base (X) includes:

[0097] Step S201, classifying the electrical devices of the robot manipulator to obtain multiple device categories;

[0098] Step S202, collecting historical data and performing feature extraction for each device category;

[0099] Step S203: Select at least two basic prediction models that match the feature type from the model database according to the extracted feature type;

[0100] Step S204: Use the historical data of the device category to train the selected basic prediction models, calculate the error metrics, and select the optimal combination method for this device category as the initial fusion model F base (X).

[0101] In step S201, the device categories include sensor categories, motor categories, driver categories, etc. The initial fusion models of the electrical devices in each category can be the same and use the same model parameters. Further, γ, β, λ s and W s can be set for each specific electrical device. In this solution, different specific electrical devices can be distinguished, and the prediction will be more detailed and accurate. However, in general, the correlation coefficient, etc., specific to a device category can already fully meet the requirements.

[0102] In step S202, collect the historical operation data of the electrical device, such as sensor data, motor data, and driver data, etc. Among them, the sensor data can include temperature, position, pressure, current, voltage, etc. The motor data can include current, voltage, speed, torque, etc. The driver data can include voltage, current, and power, etc. The key features extracted from the historical data can include statistical features, time-domain features, and frequency-domain features. Among them, the statistical features include mean, variance, maximum value, and minimum value, etc. The time-domain features include peak value and waveform factor, etc. The frequency-domain features include spectral energy and main frequency, etc.

[0103] In step S203, the model database stores various models, such as linear regression models, support vector machines, random forests, neural networks, etc. For different device categories, the selected basic prediction models may be different or the same. To improve efficiency, it is preferred to use the same basic prediction models. For example, for the sensor category, the extracted features are mean, variance, maximum value, minimum value, peak value, waveform factor, spectral energy, main frequency, skewness, kurtosis, etc., and the selectable basic models are linear regression models, support vector machines, and random forests. Another example is the motor category, the extracted features are current, speed, torque, vibration amplitude, vibration frequency, motor temperature, ambient temperature, etc., and the selectable basic models are linear regression models, support vector machines, and random forests. For the power supply category, the extracted features are voltage, current, power, voltage fluctuation, current fluctuation, power supply temperature, radiator temperature, etc., and the selectable basic models are linear regression models, support vector machines, and neural networks.

[0104] Figure 3Shows Figure 2 The initial fusion model F of the shown step S204 base (X). The schematic flowchart of the acquisition method is as follows Figure 3 As shown, this step S204 includes:

[0105] Step S2041: Segment the historical data of each device category according to time periods, and use the historical data of each time period to train the selected basic models;

[0106] Step S2042: Calculate the error of each basic model in each time period according to the following formula

[0107] ;

[0108] Wherein, Represents the error of the j-th basic model f j In the time period T i Inside, Represents the time period T i Inside the number of samples, Represents the actual value, Represents the predicted value;

[0109] Step S2043: For each time period, select the basic model with the smallest error as the optimal model in this time period;

[0110] Step S2044: Statistically count the frequency of each basic model being selected as the optimal model in all time periods according to the following formula ,

[0111] ;

[0112] Wherein, Represents the j-th basic model, Is the optimal model in the time period T i Inside, Represents the indicator function;

[0113] Step S2045: Select at least two basic models with the highest frequency for fusion as the target basic model. If the frequencies of multiple basic models are the same, select at least two basic models with the smallest error as the target basic model;

[0114] Step S2046: Perform weighted fusion on the selected target basic models to obtain the initial fusion model F base (X).

[0115] In step S4021, assume that the divided time periods are T1, T2,..., T m , and the basic models are f1, f2,..., fn , calculate the error in each time period through the formula in step S2042 to obtain E 11 , E 21 , E 31 ,..., E m1 , E 12 , E 22 , E 32 ,..., E mn , compare these errors, determine the base model corresponding to the minimum error value, and use the base model with the minimum error value as the optimal model in this time period. On this basis, count the frequencies of each base model that can be used as the optimal model in all time periods. For example, if the h-th and k-th models are the optimal models in the time period, then use f h , f k as the target base models. Further, the models f h , f k can be weighted and fused, and the corresponding weights of each model are selected for the weighted fusion method. Usually, it is sufficient to select 2 optimal models for weighted fusion, or the number of models used for weighted fusion can be increased according to needs. Further, perform a secondary determination on the data obtained at different times. If the optimal models are regularly distributed in the divided time periods, select the corresponding models with smaller errors in different time periods. For example, select different models for use during the startup stage, normal operation stage, emergency stage, and before the end of the device.

[0116] To further optimize the accuracy of the initial fusion model, the present application further refines the weighted fusion method, making the model more accurate by introducing more possible models. The specific weighted fusion in this solution includes the following steps:

[0117] Calculate the average error of each target base model in all time periods;

[0118] Calculate the weight of the i-th base model according to the target base model and the error according to the following formula ,

[0119] ;

[0120] where S represents the set of target base models, ε represents the adjustment coefficient, represents the average error of the i-th base model , and j is the number of all base models to be used; obtain the formula for the initial fusion model F base (X),

[0121] ;

[0122] Among them, , k represents the number of selected target base models, and the initial fusion model F base (X) is obtained. In this way, according to the average error, the weights of the base models to be used can be obtained. By different weight distributions, multiple models are taken into account, and further, the influence of errors is reduced.

[0123] To improve the ability of the model to continuously adapt to changes in device components, such as the replacement of specific electrical equipment, normal changes during the usage period, etc., and eliminate the influence brought by these normal changes. After the formula of the initial fusion model F base (X) is determined, every preset time interval, the average error of each base model on the latest historical data is calculated. After the average error of the latest historical data exceeds the threshold error, the model update is triggered to reselect the optimal base model or adjust the weights of the target base models, so as to obtain the updated initial fusion model. In this scenario, since the data used by the base model is the data information collected in the recent period, it is closer to the actual situation, so that the data information with reduced relevance in the historical data used to form the model before can be eliminated, and thus a prediction model more in line with the existing device can be obtained.

[0124] In the specific usage process, in order to accurately predict fault-related information, a large amount of data of electrical equipment is used to construct a general prediction model. When the relevant model is applied to a specific electrical device, the target electrical device is optimized. As the electrical device operates, it can be considered to reduce the proportion of data of other electrical equipment in the construction of the previous general model, so that the prediction model is more in line with the prediction of the current electronic device. Therefore, it is necessary to optimize the data range used. At the same time of optimization, it is also necessary to avoid the adverse effects brought by data fluctuations, so as to achieve a smooth transition. In the technical solution of this application, the above purpose is achieved by using the historical data of the latest period. In some other cases, when the electrical equipment is updated, this specific method can also be used to update the data, so as to optimize the model.

[0125] As the target prediction fusion model is used, the weight parameters can also be adjusted according to the historical prediction errors of device categories. Based on the initial weight parameters corresponding to the device categories, the historical prediction errors are considered. The weight parameters W base in the initial fusion model F s (X) corresponding to different device categories satisfy the following formula:

[0126] ;

[0127] Among them, W s (0)Represents the initial weight parameters of the corresponding device category. In this scheme, the system can adjust the weight parameters in the target prediction fusion model according to the initial weight parameters, and can optimize the weight parameters as the usage process changes, thereby making the target prediction fusion model more effective.

[0128] In order to further accurately adjust the error of the prediction model, the adaptive correction term λ s The acquisition method of is further refined, including the following steps: First, the system comprehensively collects the historical prediction error data of all electrical equipment in the target equipment category, and pre-sets a basic correction term λ s (0) , as the basis for building an adaptive correction mechanism to ensure the accuracy and reliability of subsequent calculations; then, in order to evaluate the overall prediction error of the entire target device category, the system adopts the average error calculation method. Specifically, the average error H of all electrical devices in the target device category is calculated according to the following formula s ,

[0129] ;

[0130] Where N is the number of electrical devices in the target device category, and H i represents the historical prediction error of the i-th electrical device in the target device category. This step helps to determine the prediction error level of the target device category. Subsequently, in order to fine-tune the correction term, the impact factor α of the target device category is defined. s and the influencing factors of the target electrical equipment , are calculated as follows: α s =e -k|Hs| , , k and p are adjustment coefficients, which are used to adjust the sensitivity of the above two influencing factors to the average error Hs. According to the formula Calculate the adaptive correction term λ s .

[0131] In the above steps, the historical prediction error data is introduced, and the average error of all electrical equipment is obtained through the average error calculation method. Then, the relationship (sensitivity) between the target electrical equipment and the errors of all electrical equipment is determined. The adaptive correction item in the prediction model of the target equipment is selected through the sensitivity relationship to reduce the difference between the target electrical equipment and all electrical equipment, so that the corrected prediction model is more relevant to the target electrical equipment and the impact of other electrical equipment on the error is reduced.

[0132] Optionally, the adjustment coefficient k is obtained according to the following formula:

[0133] ;

[0134] ;

[0135] The adjustment coefficient p is obtained according to the following formula:

[0136] ;

[0137] ;

[0138] wherein, both k0 and p0 are initial adjustment coefficients, and t k and t p are both adjustment parameters. By putting historical data into the above formula, corresponding adjustment parameters are determined for different target devices, and by borrowing historical data, the adaptive correction term is made closer to the actual situation, and the sensitivities of different electrical devices relative to all electrical devices are determined more accurately.

[0139] In order to more accurately reflect the historical prediction error, the data at different times are smoothed and the change trend is adjusted. The historical prediction error H(X) of the target electrical device is obtained by the following method, which specifically includes the following steps: Obtain the prediction error H t-1 (X) of the target electrical device at time t-1, the prediction error H t-2 (X) at time t-2, and the mean value H m (X) of the prediction errors at M times before time t-1; Calculate and obtain the first historical prediction error H EWMA,t-1 (X) based on the weighted moving average method according to the following formula,

[0140] ;

[0141] wherein, q represents the smoothing factor, and 0 < q < 1;

[0142] Calculate and obtain the second historical prediction error T t (X) according to the following formula,

[0143] T t-1 (X)=r × (H t-1 (X)-H t-2 (X));

[0144] wherein, r represents the trend adjustment coefficient, and 0.5 < r < 1.5;

[0145] Calculate according to the formula H(X)=θ1 × H EWMA,t-1 (X)+θ2 × T t-1(X) Calculate the historical prediction error H(X), where both θ1 and θ2 are constants, and the sum of the two is equal to 1. Through the above steps, it is possible to eliminate the influence brought by the change trend that should appear in some design situations according to the change trend of device data. Specifically, in some cases, there will be a normal and reasonable change trend during the use process. If the relevant change trend is not considered, the expansion or contraction trend within the normal range will be ignored, which will instead bring an unrealistic deviation to the prediction model. On the basis of introducing the smoothing factor and the trend adjustment coefficient, if there is no change, the smoothing factor is selected as 0.5 and the trend adjustment coefficient is selected as 1. When there is a change trend, the balance factor starts to change from 0.5, and the trend adjustment coefficient starts to change from 1. According to historical data, determine the most accurate and practical balance factor and trend adjustment coefficient for different electrical devices.

[0146] In step S300, the prediction fusion model of the target electrical device is corrected. According to the device identifier and real-time data information, predict whether the target electrical device will fail or the probability of failure. Specifically, it can be to obtain the possible data parameters of the next working condition of the target electrical device according to the prediction fusion model, and judge whether a failure will occur based on the obtained possible data parameters of the next working condition, and provide early warning of relevant information to the user. In some specific applications, the output result of the model can be directly converted into a failure probability for display.

[0147] In addition, the present application also provides an electrical device failure prediction system, which includes at least one processor. The at least one processor is used to execute computer programs or instructions to perform the operations corresponding to the aforementioned electrical device failure prediction method. The entire system is also provided with a database. The processor has a relevant model. During the specific use process, obtain the device identifier, real-time data information and historical prediction error of the target electrical device, determine the target device category according to the device identifier, call the target optimization module corresponding to the target device category, and enable the target optimization module to dynamically optimize the preset initial fusion model based on the real-time data information and historical prediction error to output the target prediction fusion model. Based on the device identifier and real-time data information, enable the target prediction fusion model to output the prediction result of the target electrical device.

[0148] The prediction system further includes a database and a communication unit. The database adopts the method of a local database and a cloud database. The cloud database is designed as the core hub for data collection and sharing. It can widely collect data from scattered electrical devices in various places. These data cover multi-dimensional information such as the operating status, performance parameters, and environmental conditions of the devices. The distributed storage and powerful processing capabilities of the cloud database enable it to efficiently process massive data, providing a basis for subsequent data analysis, model establishment, and optimization. Through the cloud database, electrical devices in different locations can break through geographical restrictions and achieve data interconnection and interoperability. Therefore, local electrical devices can easily call the data of other electrical devices, promoting the sharing and utilization of information resources.

[0149] The local database, as a supplement to the cloud database, stores the real-time data of local electrical devices and can also obtain the required data from the cloud database according to needs. This enables the local database to maintain data independence while making full use of the resources in the cloud database, providing strong support for the model optimization and fault diagnosis of local electrical devices.

[0150] In the process of model establishment and optimization, the large amount of data in the cloud database plays an important role. These data cover the extensive operating experience of the same type of electrical devices, providing rich samples for the training and verification of the model. By deeply mining the laws and characteristics in these data, a more accurate and reliable electrical device model can be established. The data in the local database enables the target optimization module to more precisely adjust the model for specific electrical devices. This way of combining global data with local data not only ensures the universality of the model but also takes into account the particularity of the model, improving the adaptability and accuracy of the model.

[0151] Through the setting of this dual-database system, a large amount of historical data of the same type of electrical devices can be borrowed at the beginning of model establishment, enabling the initial fusion model to have relatively high compliance and accuracy. This not only reduces the time and cost required for model establishment but also improves the effect of the model in practical applications. At the same time, when the number of specific electrical devices in use is large, due to the collaborative effect of the cloud database and the local database, the unit usage cost is greatly reduced, and the scale effect can further highlight the economy and practicality.

[0152] In summary, the present invention proposes a method and system for predicting electrical equipment failures, which improve the accuracy and timeliness of failure prediction through a dynamic optimization model. First, the device identifier, real-time data information, and historical prediction errors of the target electrical equipment are obtained. After obtaining the necessary data, the present invention determines the target device category according to the device identifier and calls the corresponding target optimization module. The target optimization module dynamically optimizes the preset initial fusion model based on the real-time data information and historical prediction errors according to a specific formula, and outputs the target prediction fusion model. Considering the influence of historical prediction errors, the model can adapt to the failure characteristics and prediction requirements of different device categories by dynamically adjusting the adaptive correction term and weight parameters.

[0153] The method for determining the initial fusion model is also one of the innovations of the present invention. By classifying the electrical equipment of the robot manipulator, collecting historical data and extracting features, and then selecting a basic prediction model that matches the feature type from the model database. The selected basic prediction model is trained using historical data, and the error index is calculated. Finally, the optimal combination method is selected as the initial fusion model to ensure that the initial fusion model has high accuracy and generalization ability.

[0154] To further optimize the accuracy of the model, the present invention also proposes a refinement of the weighted fusion method. By introducing more possible models and calculating the weights of each model according to the average error, a more accurate initial fusion model is obtained. At the same time, the present invention also considers the continuous adaptation and update of the model. By regularly calculating the average error of each basic model on the latest historical data and triggering model update when the error exceeds the threshold, the timeliness and accuracy of the model are ensured.

[0155] The present invention applies the target prediction fusion model to actual failure prediction. Based on the device identifier and real-time data information, the failure prediction result of the target electrical equipment is output, which can provide strong support for the maintenance and management of the equipment and reduce the possibility and impact of failures.

[0156] In addition, the present invention also provides an electrical equipment failure prediction system, which includes at least one processor for executing computer programs or instructions to implement the above electrical equipment failure prediction method. The entire system provides a comprehensive, accurate, and real-time solution for the failure prediction of electrical equipment through efficient data management and model optimization strategies.

[0157] The above is only a specific embodiment of the present application. Any improvement made on the premise of the present application's concept is regarded as the protection scope of the present application.

Claims

1. A method for predicting electrical equipment faults, characterized in that, For fault prediction of the electrical equipment of a robotic manipulator, the following steps are included: Obtain the device identifier, real-time data information X, and historical prediction error H(X) of the target electrical equipment; Determine the target device category according to the device identifier, call the target optimization module corresponding to the target device category, and enable the target optimization module to dynamically optimize the preset initial fusion model F based on the real-time data information X and the historical prediction error H(X) according to the following formula base (X), so as to output the target prediction fusion model F opt (X). F opt F(X) = base F(X)+(γ × e -β|H(X)| + λ s ) × W s × X where γ and β are both coefficients, and λ s represents an adaptive correction term, and W s represents the weight parameter of the target electrical device, and one target optimization module corresponds to one device category; Based on the device identifier and the real-time data information, enable the target prediction fusion model to output the prediction result of the target electrical equipment; The determination method of the initial fusion model F base (X) includes the following steps: Classify the electrical equipment of the robotic manipulator to obtain multiple device categories; For each device category, collect historical data and perform feature extraction; Select at least two basic prediction models that match the feature type from the model database according to the extracted feature type; Segment the historical data of each device category by time period, and use the historical data of each time period to train the selected basic models; Calculate the error of each basic model in each time period according to the following formula Among them, represents the error of the i-th basic model f i within the time period T i and represents the number of samples within the time period T i , represents the actual value, and represents the predicted value; For each time period, select the basic model with the smallest error as the optimal model in that time period; Count the frequency of each basic model being selected as the optimal model within all time periods according to the following formula , Among them, represents the j-th base model, is the optimal model within the time period T i and (·) represents the indicator function;​ Select at least two basic models with the highest frequency for fusion as the target basic model. If the frequencies of multiple basic models are the same, select at least two basic models with the smallest error as the target basic model; Perform weighted fusion on the selected target base model to obtain the initial fusion model F base (X).

2. The electrical equipment fault prediction method according to claim 1, characterized in that, The adaptive correction term λ s is obtained by the following steps: Obtain the historical prediction errors of all electrical devices in the target device category and the preset basic correction term λ s (0) ; Calculate the average error H of all electrical devices in the target device category according to the following formula s , where N represents the number of electrical devices in the target device category, and H i represents the historical prediction error of the i-th electrical device in the target device category; Define the influence factor α for the target device category s and the influence factor of the target electrical device , α s =e -k|Hs| , , where both k and p represent adjustment coefficients; According to the formula λ s =λ s (0) ×α s × calculate to obtain the adaptive correction term λ s .

3. The electrical equipment fault prediction method according to claim 2, characterized in that The adjustment coefficient k is obtained according to the following formula: The adjustment coefficient p is obtained according to the following formula: where k0 and p0 are both initial adjustment coefficients, and t k and t p are both adjustment parameters.

4. The electrical equipment fault prediction method according to any one of claims 1-3, characterized in that The method for obtaining the historical prediction error H(X) of the target electrical equipment includes the following steps: Obtain the prediction error H of the target electrical device at time t-1 t-1 (X), the prediction error H at time t-2 t-2 (X), and the average value of the prediction errors H at M moments before time t-1 m (X); The first historical prediction error H EWMA,t-1 (X) is calculated based on the weighted moving average method according to the following formula Where q represents the smoothing factor, and 0 < q < 1; The second historical prediction error T is calculated according to the following formula t (X), T t-1 (X) = r×(H t-1 (X) - H t-2 (X)); Where r represents the trend adjustment coefficient, and 0.5 < r < 1.5; Calculate the historical prediction error H(X) according to the formula H(X)=θ1×H EWMA,t-1 (X)+θ2×T t-1 (X), where θ1 and θ2 are both constants, and the sum of the two is equal to 1.

5. The electrical equipment fault prediction method according to claim 4, characterized in that, Performing weighted fusion on the selected target base model to obtain an initial fusion model F base (X), which includes the following steps: Calculate the average error of each target basic model in all time periods; Calculate the weight of the i-th base model according to the target base model and the error according to the following formula of , Among them, S represents the set of target base models, and ε represents the adjustment coefficient. represents the i-th base model of the average error; Obtain the initial fusion model F base (X)'s formula Among them, , k represents the number of selected target base models.

6. The electrical equipment fault prediction method according to claim 5, wherein The initial fusion model F base After the formula of (X) is determined, the following steps are further included: At every preset time interval, calculate the average error of each basic model on the latest historical data; After the average error of the latest historical data exceeds the threshold error, trigger model update to re-select the optimal basic model or adjust the weights of the target basic model, so as to obtain the updated initial fusion model.

7. The electrical equipment fault prediction method according to any one of claims 1-3 and 5-6, characterized in that Initial fusion model F corresponding to different device categories base Weight parameter W in (X) s Satisfy the following formula: W s =W s (0) ×e -β|H(X)| Among them, W s (0) represents the initial weight parameter corresponding to the device category.

8. An electrical equipment fault prediction system, characterized in that, Includes at least one processor, and the at least one processor is configured to execute a computer program or instruction to perform the operations corresponding to the electrical equipment fault prediction method according to any one of claims 1-7.

Images