Wind power plant work ticket identification method and system based on multi-task learning

Abstract

The invention discloses a wind power plant work ticket identification method and system based on multi-task learning, and the system comprises a multi-task learning module, a signal fusion module and a real-time blocking module, and a task specific layer of the multi-task learning module comprises ticket face verification, operation step identification and process compliance inspection. The signal fusion module is used for associating equipment state monitoring signals in the work ticket based on the safety operation step sequence in the work ticket and judging whether all signals and change sequences meet the safety operation sequence in the work ticket or not; and the real-time blocking module outputs a real-time decision taking suggestion to centralized control personnel based on the equipment signal verification result, and meanwhile, the control interface directly transmits a blocking signal to the station equipment. According to the method, the work ticket text instruction and the real-time signal of the field equipment are brought into a unified verification framework for the first time, cross-modal fusion of multi-source signals is realized, the industrial problem that the ticket surface is compliant but is not actually executed is solved, and the method has the function of strong blocking of a safe operation chain.

Description

Technical Field

[0001] This invention relates to the field of wind farm operation and maintenance management technology, specifically to a wind farm work ticket recognition method and system based on multi-task learning. Background Technology

[0002] In the actual circulation of wind power work permits, there are serious blind spots in the supervision of compliance with work procedures. On the one hand, the necessary safety measures mentioned in the work permits rely heavily on manual verification, which is not only inefficient and difficult to handle high-frequency operations, but also prone to omissions for more complex processes (such as switching operations that must be performed in the order of "power off-voltage-grounding-tag-up"), posing a great threat to operational safety. On the other hand, it is impossible to verify the authenticity of the implementation of safety measures in the work permits. If staff fill in false safety records or only archive text records, there is a lack of process evidence, making it difficult to trace responsibility in the event of a safety accident. Current on-site operation supervision suffers from three disconnects: disconnect between text records and actual operation, disconnect between discrete inspections and continuous processes, and disconnect between formal compliance and inherent safety. As an important circulation medium for on-site operations, work permits need to be digitized to achieve closed-loop verification and process-penetrating supervision of work procedures.

[0003] Currently, there are two main methods for identifying wind farm work permits. One method is based on traditional Optical Character Recognition (OCR) technology to analyze the layout and extract the table structure and content. For example, patent CN119380364A discloses a method and system for identifying power work permit table data, which focuses on the extraction of ticket content and is not related to the actual operation process. The other method is derived from the recognition method, such as the work permit security measure recognition method and terminal based on multi-input LSTM-CNN disclosed in patent CN113870052B. However, this method only verifies the ticket content itself, which has a security vulnerability of "compliant ticket but not actually executed". Another example is the wind farm work permit inspection method and system based on image text recognition disclosed in patent CN114708597B. This method adds a mobile terminal check-in verification mechanism for personnel boarding, but it still lacks supervision of the actual operation security process and cannot verify whether the actual operation complies with safety regulations. If the manual review misses the "direct grounding without power verification" item, it may lead to fatal violations.

[0004] As shown above, the original work order data contains heterogeneous information such as structured text, unstructured text, and stamped images. Traditional single-task models extract features separately for each task, which leads to computational redundancy, as the same work order is processed repeatedly by different tasks; on the other hand, it results in information fragmentation, where the equipment number feature and the operation step feature lose their correlation. Summary of the Invention

[0005] The purpose of this invention is to provide a wind farm work ticket identification method and system based on multi-task learning. It introduces a multi-task learning approach and performs three major tasks—ticket verification, operation step identification, and process compliance inspection—through a shared feature extraction layer. This solves the pain point of consistency verification of people, tickets, and equipment, provides security assurance for wind farm operation and maintenance, and improves the efficiency of safety inspection, thereby overcoming the shortcomings of existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a wind farm work permit recognition method based on multi-task learning, comprising the following steps:

[0007] S1: Work order information extraction and multi-task collaborative processing, including:

[0008] S101: Input preprocessing: Denoising, contrast enhancement, and geometric transformation-related format standardization are performed on the original image data of the work order; redundant information is filtered out for text-based work orders to form standardized input data.

[0009] S102: Shared Feature Extraction: Extract common features strongly correlated with safe operations from standardized work order data to generate shared feature vectors. Let F be the set of real numbers, and d be the feature dimension, which includes structured features such as ticket number, equipment number, and timestamp, as well as semantic features such as operation keywords and security measure descriptions, i.e.: F = [F struct F sem ],in Here, d1 represents the structured feature vector, and d1 represents the dimension of the structured feature. d1 is a semantic feature sub-vector, d2 is the semantic feature dimension, and d1+d2=d;

[0010] S103: Multi-task parallel processing: Based on the shared feature vector, the following operations are performed synchronously through three independent task processing units, including: task A for ticket number-device association, task B for operation step parsing, and task C for basic compliance preliminary judgment.

[0011] S104: Output: Structured operation sequence, ticket number-device binding relationship, preliminary judgment result of basic compliance;

[0012] S2: Verification of the fusion of device status signals and operation sequences, including:

[0013] S201: Input the structured operation sequence, ticket number-device binding relationship, and real-time status signal of the associated device output in S104;

[0014] S202: Synchronous signal acquisition: Based on the ticket number-device binding relationship, the status signals of associated devices are acquired in real time through a preset communication interface to form a multi-source signal set;

[0015] S203: Signal-Operation Sequence Matching: Bidirectional mapping of multi-source signals to structured operation sequences.

[0016] Positive verification: Check whether the current device status meets the prerequisites for each step in the operation sequence "Step 1 → Step 2 → Step 3";

[0017] Reverse verification: Based on the timestamp of the device status signal: the "opening signal" should be after the "voltage test signal".

[0018] If the tripping signal is triggered after the voltage test, an abnormality is marked, and the sequence of signal changes is checked to see if it is consistent with the sequence of operations.

[0019] S204: Comprehensive Verification Decision: Combining the preliminary basic compliance judgment results output in S104 with the signal-operation sequence matching results in S203, the final verification conclusion is generated;

[0020] S205: Output: Multi-source signal verification report, including compliance conclusions and abnormal signal location;

[0021] S3: Real-time decision output and risk blocking based on the multi-source signal verification report output by S205, including:

[0022] S301: Decision Recommendation Generation: Based on the verification conclusion type, output the differential decision to the wind farm's central control center.

[0023] Alienation Decision Recommendations:

[0024] If "fully compliant": output "allow operation" suggestion and log the operation.

[0025] If "partially compliant": output a suggestion that "manual review is required," along with details of any abnormal signals;

[0026] If it is a "serious violation": output a "stop operation immediately" suggestion, along with the reason for the violation;

[0027] S302: Real-time risk blocking: For "serious violations" scenarios, a blocking signal is sent to the associated device through a preset control interface, i.e., the PLC communication protocol, to forcibly terminate the violation operation;

[0028] S303: Full-process data archiving: Work order information, verification process data, and decision records are archived in a unified database for subsequent auditing and algorithm optimization.

[0029] Furthermore, the processing procedure for task A, which associates ticket number with device in S103, is as follows:

[0030] Extract ticket number features V from shared feature vectors t With device number feature Ve The similarity between two features is defined as:

[0031] Sim(V t V e )=α·Sim struct +(1-α)·Sim sem

[0032] In the formula, α is the structured feature weight, with an initial value of 0.3; Sim struct For structured feature similarity; Sim sem For semantic feature similarity; where, Sim struct The following method is used to determine:

[0033] ① Normalized Euclidean distance complementarity: Calculated by normalizing the Euclidean distance between structured feature vectors and taking the complementarity value; ② Manhattan distance complementarity: Applicable to high-dimensional structured features, calculated by the normalized complementarity value of the sum of the absolute differences in each dimension; ③ Hamming distance matching degree: Applicable to binary encoded features, calculated by the proportion of identical bits; ④ Jaccard coefficient: Applicable to set-type features, calculated by the ratio of intersection to union.

[0034] Sim sem The following method is used to determine:

[0035] ① Cosine similarity: Measures the directional consistency of semantic vectors by the ratio of the vector dot product to the product of their magnitudes; ② Euclidean distance complementarity: Applicable to scenarios where semantic vectors are concentrated, calculated by normalizing the complementarity value of spatial distance; ③ Adjusted cosine similarity: Eliminates baseline bias by subtracting the mean, optimizing the matching accuracy of vectors generated by different encoding models; ④ WordMover's Distance matching degree: Applicable to direct text matching scenarios, calculated by the complementarity value of word vector transport costs.

[0036] The binding relationship between work tickets and corresponding wind farm equipment is established by defining the similarity of two features. When Sim(V t V e When )≥τ, where τ is a threshold of 0.7, it is determined that the ticket number is associated with the device, and the output is "Ticket number XXX is associated with device XXX"; otherwise, it is marked as "Pending manual review".

[0037] Furthermore, the processing procedure for task B in the operation step parsing of S103 is as follows:

[0038] Sequence decoding of the shared feature vector involves first identifying the core actions of "shutdown" and "voltage testing" in the operation steps from the shared feature vector, and then determining the execution order of the steps by combining the time markers or logical connectors in the features. Next, each operation step is associated with the device number feature in the shared feature vector to clarify the operation object corresponding to the step. Finally, the decoding results are organized into a structured sequence containing "step number, step type, execution order, and associated device".

[0039] Furthermore, the processing procedure for Task C, the initial assessment of basic compliance in S103, is as follows:

[0040] Based on the compliance operation templates in the compliance operation database, the basic logic of the operation sequence is verified, and the preliminary compliance judgment result is output.

[0041] This invention provides another technical solution: a wind farm work permit recognition system based on multi-task learning, including a multi-task learning module, a signal fusion module, and a real-time blocking module;

[0042] The multi-task learning module extracts shared features through a shared layer definition, including ticket verification task, operation step identification task, and process compliance inspection task. The ticket verification task is used to associate the ticket number with the wind turbine number, the operation step identification task is used to parse the safety operation step sequence in the work ticket, and the safety compliance inspection task is used to determine whether the safety process is compliant.

[0043] The signal fusion module associates the equipment status monitoring signals in the work order with the safety operation step sequence in the work order, and determines whether all signals and their change sequence meet the safety operation sequence in the work order.

[0044] The real-time blocking module outputs real-time decision-making suggestions to the central control personnel based on the equipment signal verification results, and at the same time transmits blocking signals directly to the station equipment through the control interface.

[0045] Furthermore, the multi-task learning module includes a work ticket acquisition device, a compliance database, an edge computing unit, and a multi-task processing server; wherein, the work ticket acquisition device and the compliance database are connected to the edge computing unit through a communication interface, the edge computing unit processes the data and sends it to the multi-task processing server, and the multi-task processing server sends the data signal to the signal fusion module based on the shared feature vector.

[0046] Furthermore, the signal fusion module includes an AI camera, a smart circuit breaker, and a SCADA system that establish a communication connection with the signal fusion engine; wherein, the warning sign recognition signal from the AI ​​camera, the tripping status signal from the smart circuit breaker, and the power-off status signal from the SCADA system are respectively sent to the signal fusion engine, and the signal fusion engine sends the verification report to the real-time blocking module.

[0047] Furthermore, the real-time blocking module includes a real-time decision controller, a central control center HMI, a PLC execution unit, a circuit breaker mechanism, and an emergency stop device. The real-time decision controller receives verification reports from the signal fusion engine and feeds these reports back to the central control center HMI and the PLC execution unit. The PLC execution unit is electrically connected to the circuit breaker mechanism and the emergency stop device, respectively. The central control center HMI sends blocking signals to the circuit breaker mechanism and the emergency stop device through the PLC execution unit.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. The present invention provides a wind farm work permit identification method and system based on multi-task learning. Due to the closed-loop supervision of the entire wind farm operation process, existing methods generally stop at work permit integrity checks. The present invention can not only identify work permit text, but also extract the operation chain sequence, understand the operation logic, and verify the completeness of operation and safety measures, thereby improving the transparency of the operation and maintenance process. It realizes penetrating supervision from text instructions to signal verification to physical execution, and strengthens the supervision of the process.

[0050] 2. The wind farm work ticket identification method and system based on multi-task learning of the present invention realizes the forward-looking prevention and control of site operations: existing safety prevention and control generally trace back after a safety accident occurs. The present invention provides operation chain blocking signals or instructions in real time after detecting illegal actions, which can block fatal operations such as "grounding before verification of power" in real time, actively avoid the occurrence of safety accidents, and ensure the safety and compliance of the operation process. Attached Figure Description

[0051] Figure 1 This is a flowchart of the verification method of the present invention;

[0052] Figure 2 This is a framework diagram of the verification system of the present invention;

[0053] Figure 3 This is a flowchart of the multi-task learning module of the present invention;

[0054] Figure 4 This is a flowchart of the signal fusion module of the present invention;

[0055] Figure 5This is a flowchart of the real-time blocking module of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1-5 The present invention provides a wind farm work permit recognition method based on multi-task learning, comprising the following steps:

[0058] Step 1: Work order information extraction and multi-task collaborative processing. The electronic work order for the wind farm contains images or structured text, as detailed below:

[0059] S101: Input preprocessing: The original image data of the work order is formatted (e.g., noise reduction, contrast enhancement, geometric transformation, etc.), and redundant information (e.g., irrelevant remarks fields) is filtered out for text-based work orders to form standardized input data.

[0060] S102: Shared Feature Extraction: Extract common features strongly correlated with safe operations from standardized work order data to generate shared feature vectors. ( Let be the set of real numbers, and d be the feature dimension. This set includes structured features (such as ticket number, equipment number, and timestamp) and semantic features (such as operation keywords and security measure descriptions), i.e.: F = [F... struct F sem ],in This is a structured feature vector (d1 is the dimension of the structured feature). Let d1 be a semantic feature sub-vector (d2 is the semantic feature dimension), and d1 + d2 = d;

[0061] S103: Multi-task parallel processing: Based on the shared feature vector, the following operations are performed synchronously through three independent task processing units, including: task A for ticket number-device association, task B for operation step parsing, and task C for basic compliance preliminary judgment.

[0062] The processing procedure for task A, which associates ticket number with device, is as follows:

[0063] Extract ticket number features V from shared feature vectors t With device number feature V e The similarity between two features is defined as:

[0064] Sim(Vt V e )=α·Sim struct +(1-α)·Sim sem

[0065] In the formula, α is the weight of the structured features (initial value is 0.3, which can be optimized through historical correlation data of wind farms); Sim struct For structured feature similarity; Sim se m represents the semantic feature similarity; where Sim struct The following method is used to determine:

[0066] ① Normalized Euclidean distance complementarity: Calculated by normalizing the Euclidean distance between structured feature vectors and taking the complementarity value; ② Manhattan distance complementarity: Applicable to high-dimensional structured features, calculated by the normalized complementarity value of the sum of the absolute differences in each dimension; ③ Hamming distance matching degree: Applicable to binary encoded features, calculated by the proportion of identical bits; ④ Jaccard coefficient: Applicable to set-type features, calculated by the ratio of intersection to union.

[0067] The choice of the above methods can be flexibly adjusted according to the type of structured features (such as numerical, binary, or set-based), and all of them can achieve similarity measurement of structured features.

[0068] Sim sem The following method is used to determine:

[0069] ① Cosine similarity: Measures the directional consistency of semantic vectors by the ratio of the vector dot product to the product of their magnitudes; ② Euclidean distance complementarity: Applicable to scenarios where semantic vectors are concentrated, calculated by normalizing the complementarity value of spatial distance; ③ Adjusted cosine similarity: Eliminates baseline bias by subtracting the mean, optimizing the matching accuracy of vectors generated by different encoding models; ④ WordMover's Distance matching degree: Applicable to direct text matching scenarios, calculated by the complementarity value of word vector transport costs.

[0070] The choice of the above methods can be flexibly adjusted according to the generation method of semantic features (such as BERT encoding vectors, bag-of-words vectors), and all of them can achieve the similarity measurement of semantic features.

[0071] The binding relationship between work tickets and corresponding wind farm equipment is established by defining the similarity of two features. When Sim(V t V e When )≥τ, τ is a threshold of 0.7, which can be adjusted according to the actual correlation accuracy of the wind field; determine the association between the ticket number and the equipment, and output "Ticket number XXX is associated with equipment XXX"; otherwise, mark it as "Pending manual review".

[0072] The processing procedure for Task B in this embodiment is as follows:

[0073] Sequence decoding of the shared feature vector involves first identifying the core actions of the operation steps (such as "circuit breaker tripping" and "voltage testing") from the shared feature vector. This is combined with time markers (such as "first step" and "after") or logical connectors (such as "must first...then...") in the features to determine the execution order of the steps. Then, each operation step is associated with the device number feature in the shared feature vector to clarify the operation object corresponding to the step. Finally, the decoding results are organized into a structured sequence containing "step number, step type, execution order, and associated device," for example: [{step number:1, step type:"circuit breaker tripping", execution order:1, associated device:"fan F03-12 circuit breaker"},

[0074] {Step number:2,Step type:"Electrical testing",Execution sequence:2,Associated device:"Fan F03-12 line"},...].

[0075] The processing procedure for Task C, the initial compliance assessment in this embodiment, is as follows:

[0076] Based on the compliance operation templates in the compliance operation database (such as "power off operation" being completed before "power testing operation"), the basic logical rationality of the operation sequence is verified (such as whether there are logical contradictions in the order of steps), and the initial compliance judgment result is output ("compliant" or "further verification required").

[0077] S104: Output: Structured operation sequence, ticket number-device binding relationship, preliminary judgment result of basic compliance.

[0078] Step 2: Based on the content of Step 1 above, Step 2 performs fusion verification of device status signals and operation sequences, as follows:

[0079] S201: Input the structured operation sequence, ticket number-equipment binding relationship, and real-time status signals of associated equipment (such as warning sign recognition signals from AI cameras, tripping status signals from smart circuit breakers, and power-off status signals from SCADA systems) output in S104.

[0080] S202: Synchronous signal acquisition: Based on the ticket number-equipment binding relationship, the status signals of associated equipment are collected in real time through preset communication interfaces (such as industrial Ethernet, 4G / 5G) to form a multi-source signal set (such as "circuit breaker tripping status of wind turbine F03-12 = tripped", "AI camera detects warning sign = present").

[0081] S203: Signal-Operation Sequence Matching: Bidirectional mapping of multi-source signals to structured operation sequences.

[0082] Forward verification: Check whether the current device status meets the prerequisites of the step according to the operation sequence (e.g., step 1 → step 2 → step 3). For example, step 1 "power test" requires "the device is powered off", so verify whether the SCADA system reports "power off status = yes".

[0083] Reverse verification: Check whether the sequence of signal changes is consistent with the sequence of operations based on the timestamp of the device status signal (e.g., the "opening signal" should be triggered before the "voltage test signal"; if the opening signal is triggered after the voltage test, it is marked as abnormal).

[0084] S204: Comprehensive Verification Decision: Combining the preliminary compliance judgment results output in S104 with the signal-operation sequence matching results in S203, the final verification conclusion is generated ("Fully Compliant", "Partially Compliant (Requires Manual Confirmation)" or "Seriously Non-compliant").

[0085] S205: Output: Multi-source signal verification report, including compliance conclusions and abnormal signal location;

[0086] Step 3: Based on the multi-source signal verification report output by S205, real-time decision output and risk blocking are carried out, as detailed below:

[0087] S301: Decision Recommendation Generation: Based on the verification conclusion type, output the differential decision to the wind farm's central control center.

[0088] Alienation Decision Recommendations:

[0089] If "fully compliant": output "allow operation" suggestion and log the operation.

[0090] If "partially compliant": output "manual review required" suggestion (with details of abnormal signals, such as "circuit breaker tripping time is 2 minutes later than the voltage testing step");

[0091] If it is a "serious violation": output a "stop operation immediately" suggestion (with the reason for the violation, such as "no warning sign was detected but voltage testing was performed");

[0092] S302: Real-time risk blocking: For "serious violations" scenarios, a blocking signal is sent to the associated equipment through a preset control interface (such as PLC communication protocol) to forcibly terminate the violation operation (such as locking the circuit breaker closing function or triggering the equipment emergency stop command).

[0093] S303: Full-process data archiving: Work order information, verification process data (such as signal timestamps and matching results), and decision records are uniformly archived into the database for subsequent auditing and algorithm optimization.

[0094] Based on the wind farm work ticket identification method based on multi-task learning provided in the above embodiments, this invention also provides a verification system applicable to the above verification method; its core modules include a multi-task learning module, a signal fusion module, and a real-time blocking module.

[0095] The multi-task learning module extracts shared features through a shared layer definition, including ticket verification tasks, operation step identification tasks, and process compliance verification tasks. The ticket verification task is used to associate the ticket number with the wind turbine number, the operation step identification task is used to parse the sequence of safe operation steps in the work order, and the safety compliance verification task is used to determine whether the safety process is compliant.

[0096] Specifically, such as Figure 3 As shown: The multi-task learning module includes a work ticket collection device, a compliance database, an edge computing unit, and a multi-task processing server; among them, the work ticket collection device and the compliance database are connected to the edge computing unit through a communication interface. The edge computing unit processes the data and sends it to the multi-task processing server. The multi-task processing server sends the data signal to the signal fusion module based on the shared feature vector.

[0097] In the above embodiment, the signal fusion module is based on the safety operation step sequence in the work order, and associates the equipment status monitoring (AI camera, smart circuit breaker, SCADA system, etc.) signals (whether the warning sign exists, the trip position, the power failure status) in the work order to determine whether all signals and their change sequence meet the safety operation sequence in the work order.

[0098] Specifically, such as Figure 4 As shown: The signal fusion module includes an AI camera, a smart circuit breaker, and a SCADA system that establish a communication connection with the signal fusion engine; wherein, the warning sign recognition signal of the AI ​​camera, the tripping status signal of the smart circuit breaker, and the equipment power-off status signal of the SCADA system are sent to the signal fusion engine, and the signal fusion engine sends the verification report to the real-time blocking module.

[0099] The real-time blocking module in the above embodiment outputs real-time decision-making suggestions to the central control personnel based on the equipment signal verification results, and at the same time transmits blocking signals directly to the station equipment through the control interface.

[0100] Specifically, such as Figure 5As shown: The real-time blocking module includes a real-time decision controller, a central control center HMI, a PLC execution unit, a circuit breaker mechanism, and an emergency stop device. The real-time decision controller receives verification reports from the signal fusion engine and feeds these reports back to the central control center HMI and the PLC execution unit. The PLC execution unit is electrically connected to the circuit breaker mechanism and the emergency stop device, respectively. The central control center HMI sends blocking signals to the circuit breaker mechanism and the emergency stop device through the PLC execution unit.

[0101] As can be seen from the above, the wind farm work permit recognition method and system based on multi-task learning provided by this invention achieves cross-modal fusion of multi-source signals for the first time: it incorporates work permit text instructions and real-time signals from field equipment (SCADA / camera / sensors, etc.) into a unified verification framework, solving the industry problem of "work permit compliance but actual non-execution"; on the other hand, it achieves strong blocking of the safe operation chain: it develops an instruction-signal binding rule base (such as grounding operations must simultaneously meet "circuit breaker tripping signal" and "warning sign visual confirmation"); and automatically locks the equipment when violations occur (sending a shutdown command to the PLC through an industrial communication protocol).

[0102] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A wind farm work permit recognition method based on multi-task learning, characterized in that, Includes the following steps: S1: Work order information extraction and multi-task collaborative processing, including: S101: Input preprocessing: Denoising, contrast enhancement, and geometric transformation-related format standardization are performed on the original image data of the work order; redundant information is filtered out for text-based work orders to form standardized input data. S102: Shared Feature Extraction: Extract common features strongly correlated with safe operations from standardized work order data to generate shared feature vectors. Let F be the set of real numbers, and d be the feature dimension, which includes structured features such as ticket number, equipment number, and timestamp, as well as semantic features such as operation keywords and security measure descriptions, i.e.: F = [F struct ,F sem ],in Here, d1 represents the structured feature vector, and d1 represents the dimension of the structured feature. d1 is a semantic feature sub-vector, d2 is the semantic feature dimension, and d1+d2=d; S103: Multi-task parallel processing: Based on the shared feature vector, the following operations are performed synchronously through three independent task processing units, including: task A for ticket number-device association, task B for operation step parsing, and task C for basic compliance preliminary judgment. S104: Output: Structured operation sequence, ticket number-device binding relationship, preliminary judgment result of basic compliance; S2: Verification of the fusion of device status signals and operation sequences, including: S201: Input the structured operation sequence, ticket number-device binding relationship, and real-time status signal of the associated device output in S104; S202: Synchronous signal acquisition: Based on the ticket number-device binding relationship, the status signals of associated devices are acquired in real time through a preset communication interface to form a multi-source signal set; S203: Signal-Operation Sequence Matching: Bidirectional mapping of multi-source signals to structured operation sequences. Positive verification: Check whether the current device status meets the prerequisites for each step in the operation sequence "Step 1 → Step 2 → Step 3"; Reverse verification: Based on the timestamp of the device status signal: "Opening signal" should be in the "Electrical test signal" sequence. If the tripping signal is triggered after the voltage test, an abnormality is marked, and the sequence of signal changes is checked to see if it is consistent with the sequence of operations. S204: Comprehensive Verification Decision: Combining the preliminary basic compliance judgment results output in S104 with the signal-operation sequence matching results in S203, the final verification conclusion is generated; S205: Output: Multi-source signal verification report, including compliance conclusions and abnormal signal location; S3: Real-time decision output and risk blocking based on the multi-source signal verification report output by S205, including: S301: Decision Recommendation Generation: Based on the type of verification conclusion, output differentiated decision recommendations to the wind farm's central control center: If "fully compliant": output "allow operation" suggestion and log the operation. If "partially compliant": output "manual review required" suggestion, along with details of any abnormal signals; If it is a "serious violation": output a "stop operation immediately" suggestion, along with the reason for the violation; S302: Real-time risk blocking: For "serious violations" scenarios, a blocking signal is sent to the associated device through a preset control interface, i.e., the PLC communication protocol, to forcibly terminate the violation operation; S303: Full-process data archiving: Work order information, verification process data, and decision records are archived in a unified database for subsequent auditing and algorithm optimization.

2. The wind farm work permit recognition method based on multi-task learning as described in claim 1, characterized in that: The processing procedure for Task A, which associates ticket number and device in S103, is as follows: Extract ticket number features V from shared feature vectors t With device number feature V e The similarity between two features is defined as: Yes(V) t V e ) = a·Yes struct +(1-α)·Yes sem In the formula, α is the structured feature weight, with an initial value of 0.3; Sim struct For structured feature similarity; Sim sem For semantic feature similarity; where, Sim struct The following method is used to determine: ① Normalized Euclidean distance complementarity: Calculated by normalizing the Euclidean distance between structured feature vectors and taking the complementarity value; ② Manhattan distance complementarity: Applicable to high-dimensional structured features, calculated by the normalized complementarity value of the sum of the absolute differences in each dimension; ③ Hamming distance matching degree: Applicable to binary encoded features, calculated by the proportion of identical bits; ④ Jaccard coefficient: Applicable to set-type features, calculated by the ratio of intersection to union. Sim sem The following method is used to determine: ① Cosine similarity: Measures the directional consistency of semantic vectors by the ratio of the vector dot product to the product of their magnitudes; ② Euclidean distance complementarity: Applicable to scenarios where semantic vectors are concentrated, calculated by normalizing the complementarity value of spatial distance; ③ Adjusted cosine similarity: Eliminates baseline bias by subtracting the mean, optimizing the matching accuracy of vectors generated by different encoding models; ④ WordMover's Distance matching degree: Applicable to direct text matching scenarios, calculated by the complementarity value of word vector transport costs. The binding relationship between work tickets and corresponding wind farm equipment is established by defining the similarity of two features. When Sim(V t V e When )≥τ, where τ is a threshold of 0.7, it is determined that the ticket number is associated with the device, and the message "Ticket number XXX is associated with device XXX" is output; otherwise, it is marked as "Pending manual review".

3. The wind farm work permit recognition method based on multi-task learning as described in claim 1, characterized in that: The processing procedure for Task B in the S103 operation step analysis is as follows: Sequence decoding of the shared feature vector involves first identifying the core actions of "shutdown" and "voltage testing" in the operation steps from the shared feature vector, and then determining the execution order of the steps by combining the time markers or logical connectors in the features. Next, each operation step is associated with the device number feature in the shared feature vector to clarify the operation object corresponding to the step. Finally, the decoding results are organized into a structured sequence containing "step number, step type, execution order, and associated device".

4. The wind farm work permit recognition method based on multi-task learning as described in claim 1, characterized in that: The processing procedure for Task C, the initial assessment of basic compliance in S103, is as follows: Based on the compliance operation templates in the compliance operation database, the basic logic of the operation sequence is verified, and the preliminary compliance judgment result is output.

5. A wind farm work permit recognition system based on multi-task learning, characterized in that: It includes a multi-task learning module, a signal fusion module, and a real-time blocking module; The multi-task learning module extracts shared features through a shared layer definition, including ticket verification task, operation step identification task, and process compliance inspection task. The ticket verification task is used to associate the ticket number with the wind turbine number, the operation step identification task is used to parse the safety operation step sequence in the work ticket, and the safety compliance inspection task is used to determine whether the safety process is compliant. The signal fusion module associates the equipment status monitoring signals in the work order with the safety operation step sequence in the work order, and determines whether all signals and their change sequence meet the safety operation sequence in the work order. The real-time blocking module outputs real-time decision-making suggestions to the central control personnel based on the equipment signal verification results, and at the same time transmits blocking signals directly to the station equipment through the control interface.

6. The wind farm work permit recognition system based on multi-task learning as described in claim 5, characterized in that: The multi-task learning module includes a work ticket collection device, a compliance database, an edge computing unit, and a multi-task processing server. The work ticket collection device and the compliance database are connected to the edge computing unit through a communication interface. The edge computing unit processes the data and sends it to the multi-task processing server. The multi-task processing server sends the data signal to the signal fusion module based on the shared feature vector.

7. The wind farm work permit recognition system based on multi-task learning as described in claim 6, characterized in that: The signal fusion module includes an AI camera, a smart circuit breaker, and a SCADA system that establish a communication connection with the signal fusion engine. The warning sign recognition signal from the AI ​​camera, the tripping status signal from the smart circuit breaker, and the power-off status signal from the SCADA system are sent to the signal fusion engine, which then sends a verification report to the real-time blocking module.

8. The wind farm work permit recognition system based on multi-task learning as described in claim 7, characterized in that: The real-time blocking module includes a real-time decision controller, a central control center HMI, a PLC execution unit, a circuit breaker mechanism, and an emergency stop device. The real-time decision controller receives verification reports from the signal fusion engine and feeds these reports back to the central control center HMI and the PLC execution unit. The PLC execution unit is electrically connected to both the circuit breaker mechanism and the emergency stop device. The central control center HMI sends blocking signals to the circuit breaker mechanism and the emergency stop device through the PLC execution unit.

Images