Equipment fault processing method and system based on double-closed-loop self-evolution intelligent agent

Abstract

The invention discloses an equipment fault processing method and system based on a double-closed-loop self-evolution intelligent agent. The method comprises the steps of performing iterative optimization on a preset multi-task fault detection model based on a preset training period according to operation data of multiple equipment; when an ampere event is triggered, acquiring an equipment ID, an original fault code and current operation data; inputting the current operation data into the optimized preset multi-task fault detection model, and outputting a fault prediction result of the equipment; extracting corresponding maintenance guidance information from the equipment maintenance knowledge base, and generating a fault maintenance work order; carrying out equipment fault processing according to the fault maintenance work order; and after the equipment fault is processed, carrying out iterative updating on the equipment maintenance knowledge base according to the fault maintenance work order. According to the method, model retraining is triggered through operation data, the maintenance conclusion automatically and reversely supplements the knowledge base, synchronous evolution of knowledge and the model is guaranteed, and then the fault diagnosis accuracy and the fault processing efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a device fault processing method and system based on a double closed-loop self-evolution intelligent agent. BACKGROUND

[0002] With the digital upgrading of the automobile manufacturing and equipment manufacturing industry cluster, enterprises have higher requirements for the reliable operation, safe production and rapid recovery capability of equipment. However, in the current equipment operation and maintenance, ANDON call and maintenance scene, the following outstanding technical pain points still exist: 1. Fragmentation of equipment operation and maintenance data: operation parameters are mostly stored in underlying systems, maintenance records are scattered in paper work orders or independent forms, and unstructured documents such as maintenance manuals are distributed on shared disks. The data sources are diverse and scattered in levels, and it is difficult for field personnel to quickly obtain complete knowledge of "symptoms-reasons-measures" in the same interface; 2. Lack of self-evolution diagnostic model: existing intelligent diagnostic solutions mostly use one-time training and static deployment, which cannot automatically update the model according to newly collected time series data such as vibration, temperature and current, and the diagnostic accuracy decreases over time; when the model deviates from the actual equipment state, the system lacks incremental learning or online adaptive mechanism; 3. Maintenance knowledge is difficult to deposit and reuse: after maintenance, relevant experience often only stays in personal notes or oral transmission, and does not form unified version management with maintenance manuals, lacks automatic analysis and structured storage of work order text, resulting in high repeat rate of similar faults and difficulty in expert experience inheritance. Four, ANDON call and intelligent diagnosis are disconnected: the ANDON system usually only reports line stop information and cannot call historical knowledge base and diagnostic model to generate real-time disposal scheme; maintenance personnel need to manually search information again after arriving on the scene, which prolongs the downtime, etc.

[0003] Therefore, how to improve the fault diagnosis accuracy and fault processing efficiency has become a problem to be solved.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY The main purpose of the present application is to provide a device fault processing method and system based on a double closed-loop self-evolution intelligent agent, which aims to solve the technical problem of how to improve the computing efficiency while realizing the high-precision classification of boundary samples.

[0005] To achieve the above purpose, the present application provides a device fault processing method based on a double closed-loop self-evolution intelligent agent, which comprises: S1, collecting operation data of multiple devices according to a preset sampling period; S2, iteratively optimize the preset multi-task fault detection model according to the running data of the plurality of devices based on a preset training period, the preset multi-task fault detection model being constructed based on a Transformer framework; S3, when a light-on event is triggered, obtain a device ID, an original fault code, and current running data; S4, input the current running data into the optimized preset multi-task fault detection model, and output a fault prediction result of the device; S5, based on the fault prediction result, extract corresponding maintenance guidance information from a device maintenance knowledge base, and generate a fault maintenance work order according to the device ID, the original fault code, the fault prediction result, and the maintenance guidance information; S6, perform device fault handling according to the fault maintenance work order; S7, after the device fault handling is completed, iteratively update the device maintenance knowledge base according to the fault maintenance work order.

[0006] Optionally, the S2 comprises: S2.1, data preprocessing and data labeling are performed on the running data of the plurality of devices to obtain running data in a preset format and corresponding category labels; S2.2, input the running data in the preset format and the corresponding category labels into the preset multi-task fault detection model for training based on a preset training period; S2.3, perform consistency checking on the trained preset multi-task fault detection model, the consistency including semantic consistency, logical consistency, and statistical consistency; S2.4, after the checking passes, iteratively optimize the preset multi-task fault detection model based on the trained preset multi-task fault detection model. Optionally, the semantic consistency checking on the trained preset multi-task fault detection model comprises: input historical fault running data into the trained preset multi-task fault detection model and the preset multi-task fault detection model respectively, and obtain final layer input vectors corresponding to the trained preset multi-task fault detection model and the preset multi-task fault detection model respectively; calculate a cosine similarity according to the final layer input vectors; perform semantic consistency checking on the trained preset multi-task fault detection model based on the cosine similarity.

[0007] Optionally, the logical consistency checking on the trained preset multi-task fault detection model comprises: inputting historical fault operation data into the trained preset multi-task fault detection model to obtain a fault detection sequence of each network layer on the historical operation data; performing logical consistency verification on the trained preset multi-task fault detection model based on the fault detection sequence of the historical operation data.

[0008] Optionally, the statistical consistency verification on the trained preset multi-task fault detection model comprises: inputting a verification data set into the preset multi-task fault detection model to output F1 scores of the preset multi-task fault detection model and confidences corresponding to each verification data, and inputting the verification data set into the trained preset multi-task fault detection model to output F1 scores of the trained preset multi-task fault detection model and confidences corresponding to each verification data; calculating a model score difference value and a confidence difference value of each verification data according to the F1 scores of the preset multi-task fault detection model and the confidences corresponding to each verification data and the F1 scores of the trained preset multi-task fault detection model and the confidences corresponding to each verification data, respectively; performing statistical consistency verification on the trained preset multi-task fault detection model according to the model score difference value and the confidence difference value of each verification data.

[0009] Optionally, the S7 comprises: S7.1, extracting fault symptom information, a fault category and a fault maintenance measure from the fault maintenance work order after completion of device fault processing; S7.2, generating a fault triple according to the fault symptom information, the fault category and the fault maintenance measure; S7.3, iteratively updating the device maintenance knowledge base according to the fault triple.

[0010] In addition, to achieve the above-mentioned purpose, the application further proposes a device fault processing system based on a double closed-loop self-evolution intelligent agent, which comprises: an edge gateway configured to collect operation data of multiple devices according to a preset sampling period; a model layer configured to iteratively optimize a preset multi-task fault detection model based on a preset training period according to the operation data of the multiple devices, wherein the preset multi-task fault detection model is constructed based on a Transformer framework; a light-on diagnosis engine configured to obtain a device ID, an original fault code and current operation data when a light-on event is triggered; The model layer is further configured to input the current operation data into the optimized preset multi-task fault detection model to output a fault prediction result of the device. a knowledge layer configured to extract corresponding maintenance guidance information from a device maintenance knowledge base based on the fault prediction result; The light-on diagnosis engine is further configured to generate a fault maintenance work order according to the device ID, the original fault code, the fault prediction result, and the maintenance guidance information. An MES module is configured to perform device fault processing according to the fault maintenance work order. The knowledge layer is further configured to iteratively update the device maintenance knowledge base according to the fault maintenance work order after the device fault processing is completed.

[0011] In addition, to achieve the above-mentioned purpose, the present application further provides a device fault processing device based on a double closed-loop self-evolution intelligent agent, which comprises a memory, a processor, and a device fault processing program based on a double closed-loop self-evolution intelligent agent stored on the memory and executable on the processor. The device fault processing program based on a double closed-loop self-evolution intelligent agent is configured to implement the steps of the device fault processing method based on a double closed-loop self-evolution intelligent agent as described above.

[0012] In addition, to achieve the above-mentioned purpose, the present application further provides a storage medium having a device fault processing program based on a double closed-loop self-evolution intelligent agent stored thereon. The device fault processing program based on a double closed-loop self-evolution intelligent agent, when executed by a processor, implements the steps of the device fault processing method based on a double closed-loop self-evolution intelligent agent as described above.

[0013] Firstly, the present application collects the running data of multiple devices according to a preset sampling period, iteratively optimizes a preset multi-task fault detection model based on a preset training period according to the running data of multiple devices, constructs the preset multi-task fault detection model based on a Transformer framework, then when a light-on event is triggered, acquires a device ID, an original fault code, and current running data, inputs the current running data into the optimized preset multi-task fault detection model, outputs a fault prediction result of the device, thereafter extracts corresponding maintenance guidance information from a device maintenance knowledge base based on the fault prediction result, and generates a fault maintenance work order according to the device ID, the original fault code, the fault prediction result, and the maintenance guidance information, performs device fault processing according to the fault maintenance work order, and finally iteratively updates the device maintenance knowledge base according to the fault maintenance work order after the device fault processing is completed. The present application re-trains the model through the running data (learning closed loop), automatically supplements the knowledge base in the reverse direction with the maintenance conclusion (knowledge closed loop), and couples the two closed loops through a consistency checking algorithm, thereby ensuring the synchronous evolution of knowledge and model, and further improving the fault diagnosis accuracy and fault processing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1A total flowchart of a first embodiment of the equipment fault processing method based on a double closed-loop self-evolution intelligent agent of the present application is shown in the figure. Figure 2 A learning closed-loop working time sequence diagram of the first embodiment of the equipment fault processing method based on a double closed-loop self-evolution intelligent agent of the present application is shown in the figure. Figure 3 A knowledge closed-loop working time sequence diagram of the first embodiment of the equipment fault processing method based on a double closed-loop self-evolution intelligent agent of the present application is shown in the figure. Figure 4 A structural block diagram of the first embodiment of the equipment fault processing system based on a double closed-loop self-evolution intelligent agent of the present application is shown in the figure.

[0015] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0016] It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0017] The embodiments of the present application provide an equipment fault processing method based on a double closed-loop self-evolution intelligent agent, which will be described with reference to Figure 1 , Figure 1 A total flowchart of a first embodiment of the equipment fault processing method based on a double closed-loop self-evolution intelligent agent of the present application is shown in the figure.

[0018] In the present embodiment, the equipment fault processing method based on a double closed-loop self-evolution intelligent agent comprises the following steps: S1, collecting the running data of multiple devices according to a preset sampling period.

[0019] It is easy to understand that the execution subject of the present embodiment can be a computer device with data processing, network communication and program running functions, and the present embodiment is not limited thereto.

[0020] The preset sampling period can be set by the user, for example, 1s, and the present embodiment is not limited thereto.

[0021] The multiple devices are multiple industrial devices of a factory, each of which has a corresponding ID number, and a sensor is installed on each large device for collecting running data.

[0022] The running data includes multiple-dimensional signals such as device vibration, temperature and current.

[0023] In the present embodiment, the edge gateway collects the device vibration, temperature data and current data collected by the sensor through an OPC-UA client (OPC Unified Architecture Client) at a period of 1s.

[0024] Edge gateway model: NVIDIA Jetson AGX Xavier, 8Core CPU + 512-core GPU; server: 2x Intel Xeon Silver 4314 + 2x NV A30; network: Gigabit Ethernet, live OPC-UA bus.

[0025] S2, based on the preset training period, the running data of the multi-device is used to iteratively optimize the preset multi-task fault detection model, and the preset multi-task fault detection model is constructed based on the Transformer framework.

[0026] The learning rate of the preset multi-task fault detection model is set to 1e-4.

[0027] Further, S2 includes: S2.1, data preprocessing and data labeling are performed on the running data of the multi-device to obtain running data in a preset format and corresponding category labels; S2.2, based on the preset training period, the running data in the preset format and the corresponding category labels are input into the preset multi-task fault detection model for training; S2.3, consistency verification is performed on the trained preset multi-task fault detection model, and the consistency includes semantic consistency, logical consistency and statistical consistency; S2.4, after the verification passes, the preset multi-task fault detection model is iteratively optimized according to the trained preset multi-task fault detection model.

[0028] The preset training period is the period for optimizing the preset multi-task fault detection model through incremental data (the running data collected between the last iteration of the model and before this iteration of the model), which can be 26h, for example, 26h of running data of the multi-device is collected, and the preset multi-task fault detection model is automatically iteratively optimized every 26h.

[0029] In a specific implementation, reference is made to Figure 2 , Figure 2For the learning closed loop working timing diagram of the first embodiment of the equipment fault processing method based on the double closed loop self-evolution intelligent agent of the present application, the collected operation data needs to be pushed to the Kafka topic sensor.topic, and then the data bus is based on Schema Registry, all operation data adopts Avro unified format, which is convenient for heterogeneous system subscription. If the model layer triggers the Airflow task at 02:00 every day, the abnormal data marking is performed in combination with the fault maintenance work order generated between 02:00 of the previous day and 02:00 of the current day, and whether the incremental data between 02:00 of the previous day and 02:00 of the current day exists the fault operation data of the fault maintenance work order is detected. If it exists, the fault category label is labeled for the fault operation data, and the normal category label is labeled for other operation data; if it does not exist, the normal category label is labeled for the incremental data. Finally, at 02:00 every day, the operation data in the preset format and the corresponding category label are automatically input into the preset multi-task fault detection model for training, and the double closed loop manager calls the consistency checking module to perform consistency checking on the preset multi-task fault detection model after training, and the consistency includes semantic consistency, logical consistency and statistical consistency.

[0030] It should be further pointed out that the preset multi-task fault detection model after training needs to be checked for consistency by a preset consistency checking sequence. After the semantic consistency checking is passed, the logical consistency checking is performed, and after the logical consistency checking is passed, the statistical consistency checking is performed.

[0031] It should be understood that if the semantic consistency, or the logical consistency, or the statistical consistency is not passed, manual review and processing are required.

[0032] After all the checking is passed, the preset multi-task fault detection model is iteratively optimized according to the preset multi-task fault detection model after training, and the optimized preset multi-task fault detection model is released through Canary, and the learning closed loop is completed.

[0033] The semantic consistency checking of the preset multi-task fault detection model after training includes: inputting the historical fault operation data into the preset multi-task fault detection model after training and the preset multi-task fault detection model respectively, and obtaining the input vectors of the final layer (the last layer in the model) corresponding to the preset multi-task fault detection model after training and the preset multi-task fault detection model respectively; calculating the cosine similarity according to the final layer input vectors; based on the cosine similarity, the semantic consistency of the preset multi-task fault detection model after training is checked, and when the cosine similarity is greater than a preset similarity threshold (for example, 0.85), it is determined that the semantic consistency checking is passed.

[0034] The historical fault operation data can be the historical operation data of a certain fault extracted from the equipment maintenance knowledge base.

[0035] The logical consistency verification of the trained preset multi-task fault detection model includes: inputting historical fault operation data into the trained preset multi-task fault detection model to obtain a fault detection sequence of each network layer on the historical operation data; and performing logical consistency verification on the trained preset multi-task fault detection model based on the fault detection sequence of the historical operation data.

[0036] The fault detection sequence can be understood as the sequence in which the model processes the historical operation data through multiple layers inside the model when the historical fault operation data is received, for example, from layer A to layer B, from layer B to layer C, and the like. It is necessary to determine whether the fault detection sequence is consistent when the same historical fault operation data is input into the preset multi-task fault detection model and the trained preset multi-task fault detection model, respectively. If consistent, it is determined that the logical consistency verification is passed.

[0037] The statistical consistency verification of the trained preset multi-task fault detection model includes: inputting a verification data set into the preset multi-task fault detection model to output an F1 score of the preset multi-task fault detection model and a confidence corresponding to each verification data, and inputting the verification data set into the trained preset multi-task fault detection model to output an F1 score of the trained preset multi-task fault detection model and a confidence corresponding to each verification data; calculating a model score difference and a confidence difference of each verification data according to the F1 score of the preset multi-task fault detection model and the confidence corresponding to each verification data, and the F1 score of the trained preset multi-task fault detection model and the confidence corresponding to each verification data; and performing statistical consistency verification on the trained preset multi-task fault detection model according to the model score difference and the confidence difference of each verification data.

[0038] When the model score difference is greater than or equal to 0, and the confidence difference of each verification data is less than or equal to a preset threshold (which can be user-defined), it is determined that the statistical consistency verification is passed.

[0039] S3, when the light-on event is triggered, the device ID, the original fault code and the current operation data are obtained.

[0040] It should be further noted that when the user presses the light-on button, the light-on event is triggered, and the light-on diagnosis engine (SlotFilling+GraphSAGE) can identify the device ID and the original fault code, and then obtain the current operation data of the device ID.

[0041] S4, inputting the current operation data into the optimized preset multi-task fault detection model to output a fault prediction result of the device.

[0042] It should be noted that the device corresponds to the device ID one-to-one.

[0043] S5, extracting corresponding maintenance guidance information from the equipment maintenance knowledge base based on the fault prediction result, and generating a fault maintenance work order according to the equipment ID, the original fault code, the fault prediction result and the maintenance guidance information.

[0044] In a specific implementation, relevant maintenance information is retrieved from the equipment maintenance knowledge base according to the fault prediction result, and the maintenance information is sorted to generate maintenance guidance information.

[0045] In this embodiment, a "reason (i.e. equipment ID and fault prediction result of the equipment) -> measure (i.e. maintenance guidance information) -> safety warning" Markdown card can be generated within 30 seconds and written back to the fault maintenance work order.

[0046] The fault operation data of the equipment can also be added to the fault maintenance work order.

[0047] S6, performing equipment fault handling according to the fault maintenance work order.

[0048] In a specific implementation, the fault maintenance work order needs to be pushed to the mobile terminal of the maintenance personnel through the MES model, so that the maintenance personnel can perform equipment fault handling. After the handling, the maintenance effect needs to be verified manually. If the effect is successful, it is determined that the model prediction is correct. If the effect is failed, the real cause needs to be diagnosed manually, and a new fault maintenance work order is added for handling.

[0049] S7, after the equipment fault handling is completed, iteratively updating the equipment maintenance knowledge base according to the fault maintenance work order.

[0050] Further, S7 includes: S7.1, extracting fault symptom information (fault operation data of the equipment), fault category and fault maintenance measure from the fault maintenance work order after the equipment fault handling is completed; S7.2, generating a fault triple according to the fault symptom information, the fault category and the fault maintenance measure; S7.3, iteratively updating the equipment maintenance knowledge base according to the fault triple.

[0051] In a specific implementation, the fault triple can be directly added to the equipment maintenance knowledge base.

[0052] Another implementation manner is described with reference to Figure 3 , Figure 3For the knowledge closed loop working time sequence diagram of the first embodiment of the equipment fault processing method of the double closed loop self-evolution intelligent agent based on the application, the knowledge patch is generated by Tarjan algorithm + mutual exclusion table according to the fault triplets, then the consistency check is performed on the knowledge patch, if it passes, the triplets are added to the equipment maintenance knowledge base for iterative update, if it does not pass, manual processing is required, after manual processing, the processed triplets are added to the equipment maintenance knowledge base for iterative update.

[0053] In the embodiment, first, the running data of the multiple devices is collected according to a preset sampling period, the preset multi-task fault detection model is iteratively optimized based on a preset training period and the running data of the multiple devices, the preset multi-task fault detection model is constructed based on a Transformer framework, then when a lamp-on event is triggered, the device ID, the original fault code and the current running data are acquired, the current running data is input into the optimized preset multi-task fault detection model, the fault prediction result of the device is output, then the corresponding maintenance guidance information is extracted from the equipment maintenance knowledge base based on the fault prediction result, and the fault maintenance work order is generated according to the original fault code, the fault prediction result and the maintenance guidance information, the equipment fault is processed according to the fault maintenance work order, finally, the equipment maintenance knowledge base is iteratively updated according to the fault maintenance work order after the equipment fault processing is completed. The embodiment realizes model retraining (learning closed loop) triggered by running data, automatically supplements the knowledge base in reverse with the maintenance conclusion (knowledge closed loop), and couples the two closed loops through a consistency check algorithm, so that the knowledge and the model are evolved synchronously, and the fault diagnosis accuracy and the fault processing efficiency are improved.

[0054] With reference to Figure 4 , Figure 4 The structural block diagram of the first embodiment of the equipment fault processing system of the double closed loop self-evolution intelligent agent based on the application.

[0055] The edge gateway 4001 is configured to collect the running data of the multiple devices according to a preset sampling period and a preset sampling frequency. The model layer 4002 is configured to iteratively optimize the preset multi-task fault detection model based on a preset training period and the running data of the multiple devices, and the preset multi-task fault detection model is constructed based on a Transformer framework. The lamp-on diagnosis engine 4003 is configured to acquire the device ID, the original fault code and the current running data when a lamp-on event is triggered. The model layer 4002 is further configured to input the current running data into the optimized preset multi-task fault detection model, and output the fault prediction result of the device. The knowledge layer 4004 is configured to extract the corresponding maintenance guidance information from the equipment maintenance knowledge base based on the fault prediction result. The lamp-on diagnosis engine 4003 is further configured to generate a fault maintenance work order according to the device ID, the original fault code, the fault prediction result and the maintenance guidance information. The MES module 4005 is configured to perform device fault processing according to the fault maintenance work order. The knowledge layer 4004 is further configured to iteratively update the device maintenance knowledge base according to the fault maintenance work order after the device fault processing is completed.

[0056] Other embodiments or specific implementations of the device fault processing system based on the double closed-loop self-evolution intelligent agent can refer to the above-mentioned method embodiments, which will not be described here.

[0057] It should be noted that in this document, the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such a process, method, article or system. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or system including the element.

[0058] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0059] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, an optical disk), and includes a plurality of instructions for making a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) execute the methods described in various embodiments of the present application.

[0060] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for device fault handling based on a double closed-loop self-evolving agent, characterized in that, The method comprises the following steps: S1, collecting the operation data of multiple devices according to a preset sampling period; S2, iteratively optimizing a preset multi-task fault detection model based on a preset training period according to the operation data of multiple devices, the preset multi-task fault detection model being constructed based on a Transformer framework; S3, when a lamp-on event is triggered, obtaining a device ID, an original fault code and current operation data; S4, inputting the current operation data into the optimized preset multi-task fault detection model to output a fault prediction result of the device; S5, extracting corresponding maintenance guidance information from a device maintenance knowledge base based on the fault prediction result, and generating a fault maintenance work order according to the device ID, the original fault code, the fault prediction result and the maintenance guidance information; S6, performing device fault processing according to the fault maintenance work order; S7, after the device fault processing is completed, iteratively updating the device maintenance knowledge base according to the fault maintenance work order.

2. The method of claim 1, wherein, The S2 comprises: S2.1, data preprocessing and data labeling are performed on the operation data of multiple devices to obtain operation data in a preset format and corresponding category labels; S2.2, the operation data in the preset format and the corresponding category labels are input into a preset multi-task fault detection model for training based on a preset training period; S2.3, consistency checking is performed on the trained preset multi-task fault detection model, the consistency including semantic consistency, logical consistency and statistical consistency; S2.4, after the checking passes, the trained preset multi-task fault detection model is iteratively optimized based on the trained preset multi-task fault detection model.

3. The method of claim 2, wherein, The semantic consistency checking of the trained preset multi-task fault detection model comprises: The historical fault operation data are respectively input into the trained preset multi-task fault detection model and the preset multi-task fault detection model to respectively obtain the final layer input vectors corresponding to the trained preset multi-task fault detection model and the preset multi-task fault detection model; The cosine similarity is calculated according to the final layer input vectors; The semantic consistency of the trained preset multi-task fault detection model is checked based on the cosine similarity.

4. The method of claim 2, wherein, The logical consistency checking of the trained preset multi-task fault detection model comprises: The historical fault operation data are input into the trained preset multi-task fault detection model to obtain the fault detection sequence of each network layer on the historical operation data; The logical consistency of the trained preset multi-task fault detection model is checked based on the fault detection sequence of the historical operation data.

5. The method of claim 2, wherein, The statistical consistency checking of the trained preset multi-task fault detection model comprises: The validation data set is input into the preset multi-task fault detection model to output the F1 score of the preset multi-task fault detection model and the confidence corresponding to each validation data, and the validation data set is input into the trained preset multi-task fault detection model to output the F1 score of the trained preset multi-task fault detection model and the confidence corresponding to each validation data; According to the F1 score of the preset multi-task fault detection model and the confidence corresponding to each verification data respectively, and the F1 score of the trained preset multi-task fault detection model and the confidence corresponding to each verification data, a model score difference value and a confidence difference value of each verification data are calculated. According to the model score difference value and the confidence difference value of each verification data, statistical consistency verification is performed on the trained preset multi-task fault detection model.

6. The method of claim 1, wherein, The S7 comprises: S7.1, after the device fault processing is completed, extracting fault symptom information, fault category and fault repair measures from the fault repair work order; S7.2, generating a fault triple according to the fault symptom information, the fault category and the fault repair measures; S7.3, iteratively updating the device repair knowledge base according to the fault triple.

7. A system for handling equipment failures based on a double closed loop self-evolving agent, characterized by, The system comprises: An edge gateway configured to collect operation data of multiple devices according to a preset sampling period; A model layer configured to iteratively optimize a preset multi-task fault detection model based on a preset training period according to the operation data of the multiple devices, the preset multi-task fault detection model being constructed based on a Transformer framework; A light-on diagnosis engine configured to, when a light-on event is triggered, acquire a device ID, an original fault code and current operation data; The model layer is further configured to input the current operation data into the optimized preset multi-task fault detection model, and output a fault prediction result of the device; A knowledge layer configured to extract corresponding repair guidance information from a device repair knowledge base based on the fault prediction result; The light-on diagnosis engine is further configured to generate a fault repair work order according to the device ID, the original fault code, the fault prediction result and the repair guidance information; An MES module configured to perform device fault processing according to the fault repair work order; The knowledge layer is further configured to, after the device fault processing is completed, iteratively update the device repair knowledge base according to the fault repair work order.

Images