Alarm analysis method and device based on wind farm operation data and machine learning

Abstract

The application relates to an alarm analysis method and device based on wind farm operation data and machine learning, and relates to the technical field of power system data processing. The application comprises the following steps: obtaining original collection data of wind farm equipment; based on a pre-constructed alarm code information library, the original collection data is analyzed into alarm trigger records and reset records, and based on a preset dynamic anti-shake window and the alarm trigger records, first alarm analysis data is determined; based on a pre-constructed expert library, the first alarm analysis data is subjected to level matching and hierarchical processing, and second alarm analysis data is obtained; the second alarm analysis data is encapsulated, a push strategy is determined based on the alarm code level of the second alarm analysis data, and based on the push strategy, the encapsulated second alarm analysis data is sent to a corresponding downstream application system. The application has the technical effects of accurately analyzing wind farm equipment alarm information, filtering shaking, hierarchical processing, and reasonably pushing alarm information according to the level.

Description

Technical Field

[0001] This invention relates to the field of power system data processing technology, and in particular to an alarm analysis method and apparatus based on wind farm operation data and machine learning. Background Technology

[0002] Driven by the development of new energy power projects, wind power and other new energy sources have experienced rapid growth due to their advantages such as huge reserves and low carbon emissions, leading to a large-scale expansion of wind farm coverage. Efficient operation and monitoring of wind farms are crucial for ensuring the stability and reliability of power supply, and alarm analysis based on equipment operation data is a key factor in achieving this goal.

[0003] Currently, most wind farm alarm data analysis systems are based on a "unified judgment - single-condition triggering" model. This model primarily determines whether an alarm is triggered by monitoring whether the data at a single alarm measuring point changes. In actual operation, the system continuously collects equipment operating data and compares it with preset triggering conditions. When the data at a certain measuring point meets the triggering condition, the system immediately issues an alarm signal.

[0004] However, due to the massive number of monitoring points for wind turbines and electrical equipment, the collected data includes a large amount of high-frequency fluctuating data (high-frequency jitter data) and periodically changing data (frequent displacement data). If the traditional "univariate judgment - single-condition trigger" mode is used to analyze alarm data, the high-frequency displacement signals with millisecond-level reciprocating fluctuations can easily trigger continuous alarms, leading to an "alarm storm" and interfering with operation and maintenance decisions. For alarm points with second-level frequent displacement, the traditional analysis method does not distinguish between normal high-frequency actions and abnormal high-frequency actions, resulting in the problem of effective alarm information being overwhelmed or missed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an alarm analysis method and apparatus based on wind farm operation data and machine learning, aiming to solve at least one of the above-mentioned technical problems.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: Firstly, this application provides an alarm parsing method based on wind farm operation data and machine learning, employing the following technical solution: An alarm parsing method based on wind farm operation data and machine learning includes: Acquire raw data from wind farm equipment, including equipment identifier, model identifier, alarm code, acquired value, and timestamp; Based on a pre-built alarm code information database, the original collected data is parsed into alarm trigger records and revert records. Based on a preset dynamic anti-jitter window and alarm trigger records and revert records, the first alarm parsing data is determined. The first alarm parsing data represents the set of alarm trigger records and revert records after jitter filtering. The pre-built alarm code information database includes multiple alarm code configuration information. The alarm code configuration information includes device identifier, model identifier, alarm code, alarm trigger value, and measurement point data type. The alarm code is the identifier of the alarm measurement point. Based on a pre-built expert database, the first alarm parsing data is subjected to level matching and hierarchical processing to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier, and alarm trigger count after hierarchical processing. The pre-built expert database includes multiple alarm code hierarchical configuration information, which includes device identifier, model identifier, alarm code, alarm code level, and compression threshold. The second alarm parsing data is encapsulated, and a push strategy is determined based on the alarm code level of the second alarm parsing data. Based on the push strategy, the encapsulated second alarm parsing data is sent to the corresponding downstream application system.

[0007] The beneficial effects of this invention are as follows: By acquiring the original data collected from wind farm equipment, the first alarm parsing data is determined using an alarm code information database and a dynamic anti-jitter window, achieving standardized parsing of alarm data for different types of equipment, accurately filtering jitter data, and reducing the computational overhead and interference with operation and maintenance caused by invalid alarms; the second alarm parsing data is obtained by hierarchically processing the first alarm parsing data based on an expert database, achieving effective optimization of normal high-frequency action alarms and rapid response to abnormal high-frequency action alarms; the second alarm parsing data is encapsulated and sent to downstream application systems according to the alarm code level and a push strategy is determined, improving the system's integrability and compatibility.

[0008] Based on the above technical solution, the present invention can be further improved as follows.

[0009] Furthermore, the method for constructing the pre-built alarm code information database includes: Based on the device configuration information table and alarm measurement point mapping table in the configuration database, an alarm code information database is constructed. The measurement point data types include remote signaling type and telemetry type. The alarm trigger value is associated with the measurement point data type. The method for constructing the pre-built expert database includes: An expert database is constructed based on the displacement alarm code dictionary table in the configuration database. The alarm code levels include normal displacement level and abnormal displacement level. The compression threshold is used to define the minimum time window for data compression of normal displacement level alarms.

[0010] The beneficial effects of adopting the above-mentioned further solutions are: constructing an alarm code information database, and by configuring the measurement point data type and alarm trigger value, achieving standardized parsing of alarm data of different types of equipment, thereby improving compatibility; constructing an expert database and distinguishing between normal displacement and abnormal displacement levels, which can provide a basis for subsequent hierarchical processing of alarm data; optimizing data for normal displacement level alarms by compressing thresholds, thereby improving the fault response speed for abnormal displacement level alarms.

[0011] Furthermore, it also includes: Based on a preset period, alarm measurement points not in the expert database are selected from historical alarm data as target alarm measurement points, and the time series characteristics of the historical data collected for each target alarm measurement point are extracted. The time series characteristics include frequency characteristics and stability characteristics. The frequency characteristics represent the number of alarm changes per unit time, and the stability characteristics represent the regularity of alarm change behavior. Based on the historical data of the alarm measurement points marked in the expert database, the temporal features of the marked alarm measurement points are extracted to construct a training sample set. The training sample set includes positive samples and negative samples. The positive samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of normal change level, and the negative samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of abnormal change level. An isolated forest model is trained based on the training sample set. The isolated forest model contains multiple isolated trees, and each isolated tree is constructed using random sampling and random feature selection. The temporal features corresponding to each target alarm test point are input into the trained isolated forest model to obtain the anomaly score corresponding to each target alarm test point. The anomaly score represents the degree to which the target alarm test point deviates from the positive sample feature pattern. Based on the anomaly scores and time-series characteristics of each target alarm measurement point, multiple candidate alarm measurement points are selected; Each candidate alarm test point is pushed to the expert terminal, and the confirmation result from the expert terminal is received. The configuration information of the candidate alarm test points in the expert database is updated according to the confirmation result.

[0012] The beneficial effects of adopting the above-mentioned further solution are as follows: Target alarm points are selected from historical alarm data, and their frequency and stability features are extracted. A training sample set is constructed by combining this with the temporal features of alarm points already marked in the expert database, and the isolated forest model is trained. The temporal features of the target alarm points are input into the model to obtain anomaly scores. Candidate alarm points are then selected and pushed to experts for confirmation. The expert database is updated based on the confirmation results. This achieves adaptive learning, automatically classifying alarm points into the expert database, shifting from experience-driven to data-driven approaches, effectively managing alarm data, reducing manual point-to-point verification when new sites or equipment are added, and intelligently completing repetitive and tedious pipeline work. Simultaneously, the combination of algorithm training and expert confirmation reduces the impact of model misjudgment rate on system stability, balancing automation and reliability.

[0013] Furthermore, based on the pre-built alarm code information database, the original collected data is parsed into alarm trigger records and reset records. Based on a preset dynamic debouncing window and the alarm trigger records and reset records, the first alarm parsing data is determined, including: Based on the alarm codes in the original collected data, the corresponding alarm trigger values ​​and measurement point data types are matched from the alarm code information database; Based on the parsing rules corresponding to the matched measurement point data type and the alarm trigger value, determine whether the collected values ​​in the current raw collected data meet the alarm trigger conditions; If the collected values ​​in the current raw data meet the alarm triggering conditions, then the current raw data is determined as an alarm triggering record, and the collected values ​​after the current alarm triggering record until the next time the alarm triggering conditions are met are determined as the corresponding reset records of the alarm triggering record. Based on each alarm trigger record and reset record corresponding to the alarm measurement point, multiple consecutive initial alarm records are determined, wherein the initial alarm record includes the alarm start time and the alarm end time; For each pair of adjacent initial alarm records corresponding to the alarm measurement point, if the alarm interval time corresponding to the alarm trigger record between the two initial alarm records is not greater than the preset dynamic anti-shake window duration, then the two adjacent initial alarm records are compressed into one target alarm record. Based on each target alarm record, determine the first alarm parsing data.

[0014] The beneficial effects of adopting the above-mentioned further solution are as follows: by matching corresponding information from the alarm code information database based on the alarm code, judging the alarm triggering conditions according to the parsing rules and alarm triggering values, the alarm triggering record and the reverting record are accurately determined, and thus the initial alarm record is obtained; by using a preset dynamic anti-jitter window to compress adjacent initial alarm records, millisecond-level high-frequency jitter data can be effectively filtered out, reducing the computational overhead of invalid alarm data, reducing the interference of jitter data on operation and maintenance, and finally obtaining the first alarm parsing data after jitter filtering.

[0015] Furthermore, the step of determining whether the collected values ​​in the current raw collected data meet the alarm triggering conditions based on the parsing rules corresponding to the matched measurement point data type and the alarm triggering value includes: If the data type of the matched measurement point is remote signaling type, then determine whether the collected value in the current raw data is equal to the alarm trigger value. If the collected value is equal to the alarm trigger value, then it is determined that the collected value in the current raw collected data meets the alarm trigger condition. If the data type of the matched measurement point is telemetry type, then determine whether the collected value in the current raw data is greater than zero; If the current raw data collection value is greater than zero, then the current raw data collection value is determined to meet the alarm triggering condition.

[0016] The beneficial effects of adopting the above-mentioned further solution are as follows: Corresponding alarm triggering condition judgment rules are formulated for different types of measurement point data. For remote signaling type measurement points, the alarm triggering condition is determined by whether the collected value is equal to the alarm triggering value; for telemetry type measurement points, the alarm triggering condition is determined by whether the collected value is greater than zero. This can accurately identify different types of alarm triggering situations, effectively realize standardized parsing of alarm data from different types of equipment, and improve the compatibility of alarm parsing and the accuracy of alarm judgment.

[0017] Furthermore, based on the pre-built expert database, the first alarm parsing data is subjected to level matching and classification processing to obtain the second alarm parsing data, including: The device identifier, model identifier, and alarm code of each target alarm record in the first alarm parsing data are matched with the pre-built expert database to determine the alarm code level and compression threshold of the target alarm record; If the matched alarm code level is a normal shift level, then a time window is determined based on the compression threshold. Multiple target alarm records whose alarm start times fall within the same time window are aggregated and compressed to generate a summary alarm record. The second alarm parsing data is determined based on the summary alarm record. The summary alarm record includes the alarm start time, alarm end time, and alarm trigger count. The alarm start time of the summary alarm record is the alarm start time of the earliest target alarm record within the time window, the alarm end time is the alarm end time of the latest target alarm record within the time window, and the alarm trigger count is the total number of alarm triggers within the time window. If the matched alarm code level is an abnormal change level, or if no alarm code classification configuration information is matched in the expert database, then each target alarm record will remain unchanged, and a second alarm parsing data will be obtained.

[0018] The beneficial effects of adopting the above-mentioned further solutions are as follows: by matching the first alarm parsing data with the expert database, the alarm code level and compression threshold of the target alarm record can be determined; for alarms at the normal change level, by summarizing and compressing multiple target alarm records to generate a summary alarm record, the data of normal high-frequency action alarms can be effectively optimized and redundant alarm information can be reduced; for alarms at the abnormal change level or for which configuration information is not matched in the expert database, keeping the target alarm record in its original state can ensure the fault response speed and data accuracy of abnormal high-frequency action alarms.

[0019] Furthermore, the step of encapsulating the second alarm parsing data and determining the push strategy based on the alarm code level of the second alarm parsing data includes: The second alarm parsing data is standardized and packaged to generate structured data, which includes device identifier, alarm code, alarm start time, alarm end time, alarm code level, alarm trigger count, and processing identifier. Based on the alarm code level in the structured data, determine the push strategy; The step of determining the push strategy based on the alarm code level in the structured data includes: If the alarm code level is the normal change level, then the push will be delayed based on the period corresponding to the compression threshold configured in the expert database. If the alarm code level is an abnormal change level or no alarm code classification configuration information is found in the expert database, the alarm will be pushed immediately.

[0020] The beneficial effects of adopting the above-mentioned further solutions are as follows: by standardizing and encapsulating the second alarm parsing data into structured data, the standardization and readability of the data can be improved, which facilitates subsequent processing and analysis; the push strategy is determined according to the alarm code level in the structured data, and alarms at the normal change level are pushed with a periodic delay based on the compression threshold, which can reduce the interference of redundant alarm information and reduce network bandwidth consumption; alarms at the abnormal change level or those that do not match the configuration information are pushed immediately, which can ensure the timeliness of fault response and improve the monitoring and handling capabilities of equipment faults.

[0021] Secondly, this application provides an alarm analysis device based on wind farm operation data and machine learning, employing the following technical solution: An alarm analysis device based on wind farm operation data and machine learning mainly includes: The acquisition module is used to acquire raw data collected from wind farm equipment. The raw data collected includes equipment identifier, model identifier, alarm code, acquisition value, and timestamp. The first-stage parsing and anti-shake processing module is used to parse the original collected data into alarm trigger records and return records based on a pre-built alarm code information library, and determine the first alarm parsing data based on a preset dynamic anti-shake window and alarm trigger records and return records. The first alarm parsing data represents the set of alarm trigger records and return records after jitter filtering. The pre-built alarm code information library includes multiple alarm code configuration information, including device identifier, model identifier, alarm code, alarm trigger value and measurement point data type. The secondary classification processing module is used to perform level matching and classification processing on the first alarm parsing data based on a pre-built expert library to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier, and alarm trigger count after classification processing. The pre-built expert library includes multiple alarm code classification configuration information, which includes device identifier, model identifier, alarm code, alarm code level, and compression threshold. The encapsulation and push module is used to encapsulate the second alarm parsing data, determine the push strategy based on the alarm code level of the second alarm parsing data, and send the encapsulated second alarm parsing data to the corresponding downstream application system based on the push strategy.

[0022] Thirdly, this application provides an electronic device that adopts the following technical solution: An electronic device includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any of the first aspects, an alarm parsing method based on wind farm operation data and machine learning.

[0023] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in any one of the first aspects, an alarm parsing method based on wind farm operation data and machine learning.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an alarm parsing method based on wind farm operation data and machine learning, provided as an embodiment of the present invention. Figure 2 A flowchart illustrating another alarm parsing method based on wind farm operation data and machine learning, provided in one embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of performing an alarm analysis on the original collected data according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a process for performing secondary alarm analysis on the first alarm analysis data, provided as an embodiment of the present invention. Figure 5 A flowchart of an embodiment of the self-evolutionary incremental learning model based on isolated forest provided by the present invention; Figure 6 This is a data interaction diagram of an isolated forest-based self-evolutionary incremental learning model provided in one embodiment of the present invention.

[0026] Figure 7 A schematic diagram of an alarm analysis device based on wind farm operation data and machine learning, provided in one embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0029] This application provides an alarm parsing method based on wind farm operation data and machine learning. The method can be executed by an electronic device, which can be a server or a mobile terminal device. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The mobile terminal device can be a laptop, a desktop computer, etc., but is not limited to these.

[0030] like Figure 1 and Figure 2 As shown, an alarm parsing method based on wind farm operation data and machine learning mainly includes: S1, acquire the raw data collected from the wind farm equipment, the raw data collected includes equipment identifier, model identifier, alarm code, collected value and timestamp; In this embodiment, the data acquisition device can be various types of sensors, such as vibration sensors, temperature sensors, and pressure sensors. These sensors are installed on equipment such as wind turbines and substations in the wind farm to collect real-time operating data. The device identifier is used to uniquely identify each piece of equipment in the wind farm, the model identifier indicates the specific model of the equipment, the alarm code corresponds to the alarm measurement point, the collected value is the real-time operating value of the equipment, and the timestamp records the specific time of data acquisition.

[0031] The central control center uses an IoTDB database cluster to store the collected data from the wind farm. IoTDB adopts a multi-layer tree-structured data model, supports time-series organization of complex intelligent connected devices, and is suitable for storing massive amounts of data from various types of equipment in wind farms.

[0032] Data from alarm monitoring points of various equipment in the wind farm are collected in the manner described above, and examples of the data stored in IoTDB are shown in Tables 1 and 2.

[0033] Table 1: Table 1 contains the collected data of the electrical YX1 alarm codes of the booster station with equipment model ID 1 and equipment ID 1.

[0034] Table 2: Table 1 contains the collected alarm code data for fan YC1 with device model ID 2 and device ID 2.

[0035] S2, based on the pre-built alarm code information library, the original collected data is parsed into alarm trigger records and return records, and based on the preset dynamic anti-shake window and alarm trigger records and return records, the first alarm parsing data is determined. The first alarm parsing data represents the set of alarm trigger records and return records after jitter filtering. The pre-built alarm code information library includes multiple alarm code configuration information. The alarm code configuration information includes device identifier, model identifier, alarm code, alarm trigger value and measurement point data type. The alarm code is the identifier of the alarm measurement point. In this embodiment of the application, the method for constructing the pre-built alarm code information database includes: Based on the device configuration information table and alarm measurement point mapping table in the configuration database, an alarm code information database is constructed. The measurement point data types include remote signaling type and telemetry type, and the alarm trigger value is associated with the measurement point data type.

[0036] The alarm code information database is created based on the device configuration information table and alarm measurement point mapping table in the PostgreSQL database. Key fields in the database include device ID (device_id), model ID (model_id), alarm code (same as alarm measurement point (alarm_id)), measurement point data type (point_type), and alarm trigger value (target_value). Measurement point data types are divided into two categories: remote signaling and telemetry. Remote signaling type is commonly found in substation electrical alarms, with an alarm trigger value typically of 1. Telemetry type is commonly found in wind turbine alarms, with the alarm trigger value set according to the fault code table provided by the wind turbine manufacturer. The dynamic anti-jitter window is created based on user-preset anti-jitter parameters, with a default of 100 milliseconds, and the parameter is adjustable. The parsing rules differ for remote signaling and telemetry alarm codes. Taking the collected data of the YX1 electrical alarm code of the substation with model ID 1 and device ID 1 as an example, based on the parsing rules of the triggering and resetting of remote signaling type alarm codes, the alarm records after the original data mapping rules will be compressed according to the dynamic anti-shake window to reduce the pollution of the alarm results by invalid displacement signals.

[0037] The relationship between remote signaling type and telemetry type is shown in Table 3.

[0038] Table 3: In this embodiment of the application, specifically, S2 includes the following sub-steps: S21, based on the alarm codes in the original collected data, match the corresponding alarm trigger value and measurement point data type from the alarm code information database; S22, based on the parsing rules corresponding to the matched measurement point data type and the alarm trigger value, determine whether the collected value in the current raw collected data meets the alarm trigger condition; S23, if the collected value in the current raw collected data meets the alarm triggering condition, then the current raw collected data is determined as an alarm triggering record, and the collected value after the current alarm triggering record until the next time the alarm triggering condition is met is determined as the reset record corresponding to the alarm triggering record. S24, based on each alarm trigger record and reset record corresponding to the alarm measurement point, determine multiple consecutive initial alarm records, wherein the initial alarm record includes the alarm start time and the alarm end time; S25, for each pair of adjacent initial alarm records corresponding to the alarm measurement point, if the alarm interval time corresponding to the alarm trigger record between the two initial alarm records is not greater than the preset dynamic anti-shake window duration, then the two adjacent initial alarm records are compressed into one target alarm record. S26, Based on each target alarm record, determine the first alarm parsing data.

[0039] In this embodiment of the application, determining whether the collected values ​​in the current raw collected data meet the alarm triggering conditions based on the parsing rules corresponding to the matched measurement point data type and the alarm triggering value includes: If the data type of the matched measurement point is remote signaling type, then determine whether the collected value in the current raw data is equal to the alarm trigger value. If the collected value is equal to the alarm trigger value, then it is determined that the collected value in the current raw collected data meets the alarm trigger condition. If the data type of the matched measurement point is telemetry type, then determine whether the collected value in the current raw data is greater than zero; If the current raw data collection value is greater than zero, then the current raw data collection value is determined to meet the alarm triggering condition.

[0040] In the above embodiments, such as Figure 3 As shown, firstly, based on the alarm code information database, the configuration information related to the alarm code is read for the first parsing of the raw collected data. The parsing rules for alarm codes of remote signaling and telemetry types are different. Partial data from the alarm code information database is shown in Table 4 below.

[0041] Table 4: For remote signaling alarm codes, which are commonly found in substation electrical alarms, the alarm trigger value target_value=1; for telemetry alarm codes, which are commonly found in fan alarms, the alarm trigger value target_value is the value corresponding to the respective fault status word, according to the relevant descriptions in the preceding steps. Taking the collected data of the substation electrical alarm code YX1 with model ID 1 and device ID 1 in S1 as an example, the original collected data mapping rules are shown in Table 5 based on the parsing rules of the triggering and resetting of remote signaling type alarm codes.

[0042] Table 5: Taking the alarm code collection data of wind turbine YC1 with model ID 2 and device ID 2 in S1 as an example, based on the parsing rules of the triggering and resetting of telemetry type alarm codes described in step one, the original data mapping rules are shown in Table 6.

[0043] Table 6: Subsequently, a dynamic stabilization window duration is created based on the user-preset stabilization parameters (default 100 milliseconds, adjustable). For each device, the time interval is calculated based on two adjacent alarm records for the same alarm code. If the time interval is less than or equal to 100 milliseconds, compression processing is performed to reduce the contamination of alarm results by invalid displacement signals. Therefore, the YX1 and YC1 alarms mentioned above will be suppressed by the stabilization parameters during the parsing process. The parsing results after data stabilization processing are shown in Table 7.

[0044] Table 7: According to the analysis results, if there are N compressed sample data entries, the start time of the compressed alarm record is the start time of the first sample data entry, and the end time is the end time of the Nth sample data entry. In actual data acquisition at wind farms, this type of jitter frequently occurs. Without anti-jitter processing, over a hundred alarms may be reported within 10 seconds. Therefore, setting an appropriate anti-jitter window can effectively reduce the amount of such invalid alarm data.

[0045] S3, based on the pre-built expert library, performs level matching and hierarchical processing on the first alarm parsing data to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier and alarm trigger count after hierarchical processing. The pre-built expert library includes multiple alarm code hierarchical configuration information, including device identifier, model identifier, alarm code, alarm code level and compression threshold. In this embodiment of the application, the method for constructing the pre-built expert database includes: An expert database is constructed based on the displacement alarm code dictionary table in the configuration database. The alarm code levels include normal displacement level and abnormal displacement level. The compression threshold is used to define the minimum time window for data compression of normal displacement level alarms.

[0046] The expert database is created based on the displacement alarm code dictionary table in the PostgreSQL database. Key fields include device ID (device_id), model ID (model_id), alarm code (alarm_id), alarm code level (point_level), and compression threshold (threshold). Alarm code levels are divided into two categories: normal displacement and abnormal displacement. Normal displacement measurement points optimize alarm data based on the compression threshold (default 60 seconds) corresponding to the alarm code in the expert database, grouped by device_id and alarm_id. For the same group of data within the threshold period, only the start time of the first alarm and the end time of the last alarm are recorded, along with the total number of times the alarm occurs within the threshold period. Abnormal displacement measurement points record the triggering and recurrence of each alarm in real time, without data optimization. Taking alarm codes YX1 and YC1 as examples, according to the preset configuration information of the expert database, YX1 is an abnormal displacement level, recording each alarm in real time; YC1 is a normal displacement level, undergoing data optimization processing.

[0047] In this embodiment of the application, specifically, S3 includes the following sub-steps: S31, match the device identifier, model identifier and alarm code of each target alarm record in the first alarm parsing data with the pre-built expert database to determine the alarm code level and compression threshold of the target alarm record; S32, if the matched alarm code level is a normal shift level, then a time window is determined based on the compression threshold, and multiple target alarm records whose alarm start times fall within the same time window are aggregated and compressed to generate a summary alarm record. The second alarm parsing data is determined based on the summary alarm record. The summary alarm record includes the alarm start time, alarm end time, and alarm trigger count. The alarm start time of the summary alarm record is the alarm start time of the earliest target alarm record within the time window, the alarm end time is the alarm end time of the latest target alarm record within the time window, and the alarm trigger count is the total number of alarm triggers within the time window. S33, if the matched alarm code level is an abnormal change level, or if no alarm code classification configuration information is matched in the expert database, then each target alarm record is kept as is, and the second alarm parsing data is obtained.

[0048] In the above implementation, based on the expert database, preset alarm code hierarchical configuration information is read to perform secondary processing on the data parsed in the first step. The first step of parsing cleans up the impact of data jitter on the alarm results, while the secondary processing adaptively optimizes frequent alarms based on alarm code characteristics and business experience, preventing critical alarms from being buried among alarm information from routine action signals. Partial data from the expert database is shown in Table 8.

[0049] Table 8: like Figure 4 As shown, for normal displacement measurement points, the alarm data is optimized based on the compression threshold (threshold, default 60 seconds) corresponding to the alarm code in the expert database. The data is grouped by device_id and alarm_id. For the same group of data within the threshold period, only the start time of the first alarm and the end time of the last alarm are recorded, and the total number of times the alarm occurs within the threshold period is also recorded. For abnormal displacement measurement points, each alarm trigger and reset is recorded in real time, without data optimization. This is because the abnormal displacement measurement points preset in the expert database are alarm codes determined by fan and electrical professionals based on equipment characteristics and industry experience, indicating abnormal equipment operating conditions. These alarm codes are generally professionally recognized as fault, critical alarm, abnormality, accident, limit violation, or switch displacement level alarms. Their triggering is both warning and serious. Therefore, without data compression, it is recommended to record all data to promptly obtain information on equipment fault conditions.

[0050] Subsequently, based on the configuration information of the expert database regarding the preset alarm codes YX1 and YC1, it can be determined that YX1 is an abnormal displacement level, i.e., an abnormal high-frequency action; and YC1 is a normal displacement level, i.e., a normal high-frequency action. Therefore, the results of the above YX1 and YC1 alarms after secondary processing are shown in Table 9.

[0051] Table 9: According to the secondary processing results, for alarms with abnormal high-frequency actions (YX1), each valid trigger of the YX1 alarm was recorded, ensuring the accuracy and authenticity of the alarm data; for alarms with normal high-frequency actions (YC1), the data was effectively optimized, reducing the impact of redundant trigger records on the alarm results and optimizing the monitoring efficiency of monitoring personnel; if the expert database is not configured with the current alarm code, the data is recorded in full by default.

[0052] S4, the second alarm parsing data is encapsulated, and a push strategy is determined based on the alarm code level of the second alarm parsing data. Based on the push strategy, the encapsulated second alarm parsing data is sent to the corresponding downstream application system.

[0053] In this embodiment of the application, the step of encapsulating the second alarm parsing data and determining the push strategy based on the alarm code level of the second alarm parsing data includes: The second alarm parsing data is standardized and packaged to generate structured data, which includes device identifier, alarm code, alarm start time, alarm end time, alarm code level, alarm trigger count, and processing identifier. Based on the alarm code level in the structured data, determine the push strategy; The step of determining the push strategy based on the alarm code level in the structured data includes: If the alarm code level is the normal change level, then the push will be delayed based on the period corresponding to the compression threshold configured in the expert database. If the alarm code level is an abnormal change level or no alarm code classification configuration information is found in the expert database, the alarm will be pushed immediately.

[0054] In the above implementation, data standardization involves encapsulating the parsed alarm data into entity class objects in a unified format. The entity class objects contain core fields such as device ID, device code, model ID, model, alarm code, alarm code level, alarm start time, alarm end time, alarm trigger count, and processing identifier.

[0055] Multi-protocol distribution supports the transformation of standardized alarm data to different application systems, such as centralized monitoring platforms, operation and maintenance work order systems, and real-time monitoring dashboards, while being compatible with the interface requirements of different systems. The distribution strategy is based on the alarm code level: for normal displacement test points, one summary data point is pushed every compression cycle; for abnormal displacement test points, each parsed data point is pushed in real time; and for test points not configured in the expert database, the push logic is the same as for abnormal displacement test points. Distribution protocols include MQTT, HTTP / HTTPS, and WebSocket protocols.

[0056] Taking the YC1 alarm example data above as an example, it includes the core fields shown in Table 10.

[0057] Table 10: Distribution strategies include: Normal displacement measurement points: Push one summary data point every compression cycle (e.g., 60 seconds) to reduce interference from redundant alarm information and reduce network bandwidth usage; Abnormal displacement measurement points: Real-time push of each piece of parsed data (alarm triggering and reset records will not be merged into one piece of data) to ensure the timeliness of fault response; The expert database does not have any test points configured: the push logic is the same as that for abnormal displacement test points.

[0058] As an optional implementation of this application, an alarm parsing method based on wind farm operation data and machine learning further includes: Sa, based on a preset period, selects alarm measurement points not in the expert database from historical alarm data as target alarm measurement points, and extracts the time series characteristics of the historical data collected for each target alarm measurement point. The time series characteristics include frequency characteristics and stability characteristics. The frequency characteristics represent the number of alarm changes per unit time, and the stability characteristics represent the regularity of alarm change behavior. Sb, based on the historical data of the alarm measurement points marked in the expert database, extract the temporal features of the marked alarm measurement points and construct a training sample set. The training sample set includes positive samples and negative samples. The positive samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of normal change level, and the negative samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of abnormal change level. Sc, an isolated forest model is trained based on the training sample set. The isolated forest model contains multiple isolated trees, and each isolated tree is constructed using random sampling and random feature selection. Sd, input the temporal features corresponding to each target alarm test point into the trained isolated forest model to obtain the anomaly score corresponding to each target alarm test point. The anomaly score represents the degree to which the target alarm test point deviates from the positive sample feature pattern. Se, based on the anomaly score and time series characteristics of each target alarm measurement point, filters out multiple candidate alarm measurement points; Sf pushes each candidate alarm test point to the expert terminal and receives the confirmation result from the expert terminal. Based on the confirmation result, it updates the configuration information of the candidate alarm test points in the expert database.

[0059] In the above embodiments, such as Figure 5 and Figure 6 As shown, a self-evolutionary task is triggered once per hour (configurable) to extract the temporal features of the data corresponding to alarm measurement points not included in the expert database.

[0060] The frequency feature extraction process involves grouping historical data from the IoTDB database for the past hour (configurable) by device ID (device_id) and alarm code (alarm_id), counting the number of shifts in each group, filtering out data within a group that has experienced frequent shifts (e.g., ≥10 times within 10 minutes, threshold adjustable), and calculating the number of shifts within a threshold period. Record the displacement frequency and the corresponding alarm code information; The stability feature extraction process involves calculating the standard deviation of continuous displacement intervals based on the data selected from the frequency feature extraction, in order to quantify the periodic stability features.

[0061] Assume that the continuity changes to time: Then the displacement interval is: The mean interval is: The standard deviation is: The smaller the standard deviation here, the more regular the data displacement phenomenon, indicating stronger stability. Normal displacement measurement points typically conform to this pattern and are the alarm measurement points in the expert database that require data optimization. Here, based on statistical analysis of alarm data from over 2000 typhoon generators and electrical equipment, the standard deviation of over 90% of the normal displacement measurement point data... Therefore, set This is the periodic stability threshold.

[0062] Subsequently, a training sample set was constructed based on the historical data collected from the alarm measurement points already marked in the expert database.

[0063] Specifically, the positive sample construction involves: reading the alarm code information of the normal displacement level (point_level=1) in the expert database, querying the historical data of the corresponding alarm code in the IoTDB database, and extracting its frequency features and stability features based on the above feature extraction method. The construction of negative samples involves reading the alarm code information of the abnormal displacement level (point_level=2) in the expert database and extracting the same feature dimensions as the positive samples. The dataset is randomly divided into a training set (70%) and a validation set (30%) in a 7:3 ratio for model training and parameter optimization.

[0064] Next, the extracted features are standardized by performing Z-Score standardization on the input features (frequencies of displacement, standard deviation of intervals) to eliminate the influence of dimensions. ,in, The characteristic mean, The feature standard deviation is calculated and saved based on the training set for standardization of subsequent new data. Next, the isolated forest model is trained, specifically including: Initialize the isolated forest model and set up 100 isolated trees (iTree) (test 50~200 trees through grid search, stop increasing the number of trees when the accuracy of the validation set improves by <1%, and finally confirm 100 trees). Each tree uses random sampling (the sampling amount is 25% of the total samples in the training set) and random feature selection (randomly select 1 feature for node splitting). During training, the maximum tree depth was set to 10 to avoid overfitting and ensure that the model has the ability to generalize to normal displacement measurement points with different displacement frequencies. The model outputs anomaly scores for the input feature vector. This is used to measure the degree to which it deviates from the "normal displacement measurement point" pattern, and its anomaly score is: ,in for The path length in each isolated tree. for Average path length across all isolated trees Standardization constant ( (For sample size), the calculation formula is: ( For harmonic numbers, , (where Euler's constant is used).

[0065] Based on the ROC curve of the validation set derived from the expert database, when the false positive rate is <5%, the corresponding outlier score is 0.3, i.e. The point was judged to be a normal displacement measuring point.

[0066] Finally, candidate measurement point selection: For satisfying "frequency" Standard deviation of interval Abnormal scores The measurement points are pushed to the operation and maintenance experts for confirmation. After confirmation, they are automatically entered into the expert database, and their alarm code level point_level is set to 1. The compression threshold threshold is set to 60 seconds by default. For satisfying "frequency" However, it does not meet the "interval standard deviation" requirement. "or abnormal scores" "For any measurement point under any condition, the data will be pushed to the operation and maintenance expert for confirmation. Once confirmed, the data will be automatically entered into the expert database. The alarm code level point_level will be set to 2, and the compression threshold threshold will be set to empty by default." For the measurement points in the above two scenarios, if the operations and maintenance experts do not accept the training results, they can choose: Enter the expert database, and then manually set their alarm code level according to actual business requirements; The measured data will not be entered into the expert database and will continue to be used for training in subsequent self-evolutionary tasks.

[0067] In this embodiment, alarm data samples corresponding to new alarm codes confirmed by operation and maintenance experts are added to the training set every 24 hours, replacing the 10 worst-performing isolated trees in the model, while retaining historical learning results.

[0068] This method acquires raw data from wind farm equipment, uses an alarm code information database and a dynamic anti-jitter window to determine the first alarm parsing data, achieving standardized parsing of alarm data for different types of equipment, accurately filtering jitter data, and reducing the computational overhead and interference with operation and maintenance caused by invalid alarms. Based on an expert database, the first alarm parsing data is processed in a hierarchical manner to obtain the second alarm parsing data, enabling effective optimization of normal high-frequency action alarms and rapid response to abnormal high-frequency action alarms. The second alarm parsing data is encapsulated and sent to downstream application systems according to the alarm code level and a push strategy, improving the system's integrability and compatibility.

[0069] Figure 7 An alarm analysis device 200 based on wind farm operation data and machine learning is shown.

[0070] like Figure 7 As shown, an alarm analysis device 200 based on wind farm operation data and machine learning mainly includes: The acquisition module 201 is used to acquire the raw data collected by the wind farm equipment. The raw data collected includes equipment identifier, model identifier, alarm code, acquisition value and timestamp. The first-stage parsing and anti-shake processing module 202 is used to parse the original collected data into alarm trigger records and return records based on a pre-built alarm code information library, and determine the first alarm parsing data based on a preset dynamic anti-shake window and alarm trigger records and return records. The first alarm parsing data represents the set of alarm trigger records and return records after jitter filtering. The pre-built alarm code information library includes multiple alarm code configuration information, including device identifier, model identifier, alarm code, alarm trigger value and measurement point data type. The secondary classification processing module 203 is used to perform level matching and classification processing on the first alarm parsing data based on a pre-built expert library to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier and alarm trigger count after classification processing. The pre-built expert library includes multiple alarm code classification configuration information, which includes device identifier, model identifier, alarm code, alarm code level and compression threshold. The encapsulation and push module 204 is used to encapsulate the second alarm parsing data, determine the push strategy based on the alarm code level of the second alarm parsing data, and send the encapsulated second alarm parsing data to the corresponding downstream application system based on the push strategy.

[0071] In one example, the module in any of the above devices may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0072] For example, when modules in a device can be implemented via a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor capable of calling programs. Alternatively, these modules can be integrated together as a system-on-a-chip (SOC).

[0073] In this application, various objects such as messages / information / devices / network elements / systems / apparatus / actions / operations / processes / concepts may be named. It is understood that these specific names do not constitute a limitation on the relevant objects. The names may be changed depending on the scenario, context, or usage habits. The understanding of the technical meaning of the technical terms in this application should be mainly determined from their functions and technical effects embodied / performed in the technical solution.

[0074] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0075] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0076] Figure 8 This is a structural block diagram of an electronic device 300 according to an embodiment of this application.

[0077] like Figure 8As shown, the electronic device 300 includes a processor 301 and a memory 302, and may further include one or more of an information input / output (I / O) interface 303, a communication component 304, and a communication bus 305.

[0078] The processor 301 controls the overall operation of the electronic device 300 to complete all or part of the steps in the alarm parsing method based on wind farm operation data and machine learning described above. The memory 302 stores various types of data to support the operation of the electronic device 300. This data may include, for example, instructions for any application or method operating on the electronic device 300, as well as application-related data. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as one or more of Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0079] I / O interface 303 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 304 is used to test wired or wireless communication between electronic device 300 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 304 may include a Wi-Fi component, a Bluetooth component, and an NFC component.

[0080] The communication bus 305 may include a path for transmitting information between the aforementioned components. The communication bus 305 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 305 may be divided into an address bus, a data bus, a control bus, etc.

[0081] The electronic device 300 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the alarm parsing method based on wind farm operation data and machine learning given in the above embodiments.

[0082] The following describes the computer-readable storage medium provided in the embodiments of this application. The computer-readable storage medium described below can be referred to in correspondence with the alarm parsing method based on wind farm operation data and machine learning described above.

[0083] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the alarm parsing method based on wind farm operation data and machine learning described above.

[0084] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing application concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. An alarm parsing method based on wind farm operation data and machine learning, characterized in that, include: Acquire raw data from wind farm equipment, including equipment identifier, model identifier, alarm code, acquired value, and timestamp; Based on a pre-built alarm code information database, the original collected data is parsed into alarm trigger records and revert records. Based on a preset dynamic anti-jitter window and alarm trigger records and revert records, the first alarm parsing data is determined. The first alarm parsing data represents the set of alarm trigger records and revert records after jitter filtering. The pre-built alarm code information database includes multiple alarm code configuration information. The alarm code configuration information includes device identifier, model identifier, alarm code, alarm trigger value, and measurement point data type. The alarm code is the identifier of the alarm measurement point. Based on a pre-built expert database, the first alarm parsing data is subjected to level matching and hierarchical processing to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier, and alarm trigger count after hierarchical processing. The pre-built expert database includes multiple alarm code hierarchical configuration information, which includes device identifier, model identifier, alarm code, alarm code level, and compression threshold. The second alarm parsing data is encapsulated, and a push strategy is determined based on the alarm code level of the second alarm parsing data. Based on the push strategy, the encapsulated second alarm parsing data is sent to the corresponding downstream application system.

2. The alarm parsing method based on wind farm operation data and machine learning according to claim 1, characterized in that, The method for constructing the pre-built alarm code information database includes: Based on the device configuration information table and alarm measurement point mapping table in the configuration database, an alarm code information database is constructed. The measurement point data types include remote signaling type and telemetry type. The alarm trigger value is associated with the measurement point data type. The method for constructing the pre-built expert database includes: An expert database is constructed based on the displacement alarm code dictionary table in the configuration database. The alarm code levels include normal displacement level and abnormal displacement level. The compression threshold is used to define the minimum time window for data compression of normal displacement level alarms.

3. The alarm parsing method based on wind farm operation data and machine learning according to claim 1, characterized in that, Also includes: Based on a preset period, alarm measurement points not in the expert database are selected from historical alarm data as target alarm measurement points, and the time series characteristics of the historical data collected for each target alarm measurement point are extracted. The time series characteristics include frequency characteristics and stability characteristics. The frequency characteristics represent the number of alarm changes per unit time, and the stability characteristics represent the regularity of alarm change behavior. Based on the historical data of the alarm measurement points marked in the expert database, the temporal features of the marked alarm measurement points are extracted to construct a training sample set. The training sample set includes positive samples and negative samples. The positive samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of normal change level, and the negative samples are composed of the temporal features corresponding to the alarm measurement points with alarm code level of abnormal change level. An isolated forest model is trained based on the training sample set. The isolated forest model contains multiple isolated trees, and each isolated tree is constructed using random sampling and random feature selection. The temporal features corresponding to each target alarm test point are input into the trained isolated forest model to obtain the anomaly score corresponding to each target alarm test point. The anomaly score represents the degree to which the target alarm test point deviates from the positive sample feature pattern. Based on the anomaly scores and time-series characteristics of each target alarm measurement point, multiple candidate alarm measurement points are selected; Each candidate alarm test point is pushed to the expert terminal, and the confirmation result from the expert terminal is received. The configuration information of the candidate alarm test points in the expert database is updated according to the confirmation result.

4. The alarm parsing method based on wind farm operation data and machine learning according to claim 1, characterized in that, The pre-built alarm code information database parses the original collected data into alarm trigger records and reset records, and determines the first alarm parsing data based on the preset dynamic anti-shake window and alarm trigger records and reset records, including: Based on the alarm codes in the original collected data, the corresponding alarm trigger values ​​and measurement point data types are matched from the alarm code information database; Based on the parsing rules corresponding to the matched measurement point data type and the alarm trigger value, determine whether the collected values ​​in the current raw collected data meet the alarm trigger conditions; If the collected values ​​in the current raw data meet the alarm triggering conditions, then the current raw data is determined as an alarm triggering record, and the collected values ​​after the current alarm triggering record until the next time the alarm triggering conditions are met are determined as the corresponding reset records of the alarm triggering record. Based on each alarm trigger record and reset record corresponding to the alarm measurement point, multiple consecutive initial alarm records are determined, wherein the initial alarm record includes the alarm start time and the alarm end time; For each pair of adjacent initial alarm records corresponding to the alarm measurement point, if the alarm interval time corresponding to the alarm trigger record between the two initial alarm records is not greater than the preset dynamic anti-shake window duration, then the two adjacent initial alarm records are compressed into one target alarm record. Based on each target alarm record, determine the first alarm parsing data.

5. The alarm parsing method based on wind farm operation data and machine learning according to claim 4, characterized in that, The process of determining whether the collected values ​​in the current raw data meet the alarm triggering conditions based on the parsing rules corresponding to the matched measurement point data type and the alarm triggering value includes: If the data type of the matched measurement point is remote signaling type, then determine whether the collected value in the current raw data is equal to the alarm trigger value. If the collected value is equal to the alarm trigger value, then it is determined that the collected value in the current raw collected data meets the alarm trigger condition. If the data type of the matched measurement point is telemetry type, then determine whether the collected value in the current raw data is greater than zero; If the current raw data collection value is greater than zero, then the current raw data collection value is determined to meet the alarm triggering condition.

6. The alarm parsing method based on wind farm operation data and machine learning according to claim 4, characterized in that, Based on a pre-built expert database, the first alarm analysis data is subjected to level matching and classification processing to obtain the second alarm analysis data, including: The device identifier, model identifier, and alarm code of each target alarm record in the first alarm parsing data are matched with the pre-built expert database to determine the alarm code level and compression threshold of the target alarm record; If the matched alarm code level is a normal shift level, then a time window is determined based on the compression threshold. Multiple target alarm records whose alarm start times fall within the same time window are aggregated and compressed to generate a summary alarm record. The second alarm parsing data is determined based on the summary alarm record. The summary alarm record includes the alarm start time, alarm end time, and alarm trigger count. The alarm start time of the summary alarm record is the alarm start time of the earliest target alarm record within the time window, the alarm end time is the alarm end time of the latest target alarm record within the time window, and the alarm trigger count is the total number of alarm triggers within the time window. If the matched alarm code level is an abnormal change level, or if no alarm code classification configuration information is matched in the expert database, then each target alarm record will remain unchanged, and a second alarm parsing data will be obtained.

7. The alarm parsing method based on wind farm operation data and machine learning according to claim 5, characterized in that, The step of encapsulating the second alarm parsing data and determining the push strategy based on the alarm code level of the second alarm parsing data includes: The second alarm parsing data is standardized and packaged to generate structured data, which includes device identifier, alarm code, alarm start time, alarm end time, alarm code level, alarm trigger count, and processing identifier. Based on the alarm code level in the structured data, determine the push strategy; The step of determining the push strategy based on the alarm code level in the structured data includes: If the alarm code level is the normal change level, then the push will be delayed based on the period corresponding to the compression threshold configured in the expert database. If the alarm code level is an abnormal change level or no alarm code classification configuration information is found in the expert database, the alarm will be pushed immediately.

8. An alarm analysis device based on wind farm operation data and machine learning, characterized in that, include: The acquisition module is used to acquire raw data collected from wind farm equipment. The raw data collected includes equipment identifier, model identifier, alarm code, acquisition value, and timestamp. The first-stage parsing and anti-shake processing module is used to parse the original collected data into alarm trigger records and return records based on a pre-built alarm code information library, and determine the first alarm parsing data based on a preset dynamic anti-shake window and alarm trigger records and return records. The first alarm parsing data represents the set of alarm trigger records and return records after jitter filtering. The pre-built alarm code information library includes multiple alarm code configuration information, including device identifier, model identifier, alarm code, alarm trigger value and measurement point data type. The secondary classification processing module is used to perform level matching and classification processing on the first alarm parsing data based on a pre-built expert library to obtain the second alarm parsing data. The second alarm parsing data includes the alarm start and end time, processing identifier, and alarm trigger count after classification processing. The pre-built expert library includes multiple alarm code classification configuration information, which includes device identifier, model identifier, alarm code, alarm code level, and compression threshold. The encapsulation and push module is used to encapsulate the second alarm parsing data, determine the push strategy based on the alarm code level of the second alarm parsing data, and send the encapsulated second alarm parsing data to the corresponding downstream application system based on the push strategy.

9. An electronic device, characterized in that, Includes a processor, which is coupled to a memory; The processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes a computer program or instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1-7.

Images