Wind power plant personnel positioning track deviation management and control method and system based on Internet of Things

Abstract

The invention relates to the technical field of wind power plant personnel management and control, and discloses a wind power plant personnel positioning track deviation management and control method and system based on the Internet of Things, and the method comprises the steps: firstly fusing the real-time positioning data, work order task data and environment state data of personnel, and constructing a time sequence data queue rich in context information; through dynamic time sequence feature engineering, the kinematics features and the behavior mode features of the trajectory are converted into personnel positioning trajectory time sequence feature vectors; and then an unsupervised learning model is used to carry out abnormal probability scoring on the personnel positioning track time sequence feature vector, and finally, dynamic grading alarm is realized according to a quantized abnormal score. Therefore, accurate identification, quantitative evaluation and intelligent early warning of personnel behavior risks are realized through deep fusion of multi-source data and application of a time sequence data analysis model, so that the defects in the prior art are overcome.

Description

Technical Field

[0001] This application relates to the field of personnel management technology in wind farms, and more specifically, to a method and system for controlling deviations in the positioning trajectory of personnel in wind farms based on the Internet of Things. Background Technology

[0002] Wind power, as a clean and renewable energy source, plays a significant role in the global energy transition. Wind farms are often located in vast areas with complex terrain and variable climates. Equipment is dispersed, and maintenance personnel frequently operate in high-risk environments such as high altitudes, high pressure, and strong winds, posing serious challenges to safety management. Traditional management methods rely on walkie-talkie reporting and paper work orders, which suffer from information delays, regulatory blind spots, and slow emergency response, making it difficult to achieve real-time and precise control over personnel operations. Once anomalies such as deviation from the path, entry into dangerous areas, or loss of contact occur, they are difficult to detect in a timely manner, easily leading to safety accidents.

[0003] To address this, the industry has introduced IoT technology to track personnel locations via location badges or smart terminals. However, existing solutions are mostly rudimentary, simply displaying location information on a map or triggering alarms using simple electronic fences. These methods lack fusion analysis of the operational context (such as work order tasks, inspection routes, and environmental conditions), failing to distinguish between normal detours and dangerous deviations, resulting in frequent false alarms and missed alarms. Furthermore, focusing only on static location points or area entry / exit behaviors lacks in-depth analysis of trajectory temporal characteristics (such as speed, dwell time, and movement patterns), making it difficult to identify complex risks such as "lingering on the edge of a danger zone." Alarm mechanisms are mostly based on fixed rules, lacking risk level quantification and dynamic classification, impacting emergency response efficiency. Therefore, a more intelligent and refined personnel trajectory monitoring and risk warning solution is urgently needed to improve the safety management level of wind farms. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application propose a method and system for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things (IoT), overcoming the technical shortcomings of existing wind farm personnel positioning schemes, such as lack of context awareness, inability to perform in-depth analysis of dynamic trajectories, rigid alarm mechanisms, and high false alarm rates.

[0005] According to one aspect of this application, a method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things is provided, comprising: acquiring raw positioning data, including personnel ID, timestamp, and coordinates; querying work order data and environmental status data based on the raw positioning data; constructing fused context data points based on the raw positioning data, work order data, and environmental status data; performing dynamic temporal feature engineering on the time queue of the fused context data points to obtain a temporal feature vector of personnel positioning trajectory; performing trajectory anomaly probability scoring on the temporal feature vector of personnel positioning trajectory to obtain anomaly scores; and performing dynamic hierarchical alarm decision-making based on the anomaly scores to obtain alarm events.

[0006] In one possible implementation, based on the original location data, work order data and environmental status data are queried, including: querying work order data based on personnel ID and timestamp, wherein the work order data includes the current personnel's task type and preset path; and querying environmental status data based on coordinates, wherein the environmental status data includes region ID and region type.

[0007] In one possible implementation, dynamic temporal feature engineering is performed on the time queue of fused context data points to obtain a temporal feature vector of the personnel positioning trajectory. This includes: extracting kinematic features from the time queue of fused context data points to obtain a kinematic feature set, which includes instantaneous velocity, instantaneous acceleration, and stationary dwell time; extracting context and behavioral features from the time queue of fused context data points to obtain a context feature set, which includes preset path distance, historical trajectory similarity, encoded task type, and encoded region type; and assembling feature vectors from the kinematic feature set and the context feature set to obtain the temporal feature vector of the personnel positioning trajectory.

[0008] In one possible implementation, kinematic feature extraction is performed on the time queue of fusion context data points to obtain a kinematic feature set, including: extracting the fusion context data point at time step t and the fusion context data point at time step (t-1) from the time queue of fusion context data points; calculating the Euclidean distance and time difference between the fusion context data point at time step t and the fusion context data point at time step (t-1), and calculating the instantaneous velocity at time step t based on the Euclidean distance and time difference; extracting the fusion context data point at time step (t-2) from the time queue of fusion context data points; calculating the instantaneous velocity at time step (t-1) based on the fusion context data point at time step (t-2) and the fusion context data point at time step (t-1); and calculating the instantaneous acceleration based on the instantaneous velocity at time step (t-1) and the instantaneous velocity at time step t.

[0009] In one possible implementation, kinematic feature extraction is performed on the time queue of fused context data points to obtain a kinematic feature set, including: comparing the time queue of instantaneous velocity with a stationary velocity threshold, and counting the total duration of continuous time below the stationary velocity threshold as the stationary dwell time.

[0010] In one possible implementation, context and behavioral features are extracted from the time queue of fused context data points to obtain a context feature set, including: extracting coordinates and a preset path from the fused context data points; calculating the shortest vertical distance from the coordinates to the preset path as the preset path distance; extracting a current trajectory segment composed of N coordinates from the time queue of fused context data points; retrieving a normal trajectory sample set from a historical normal trajectory database based on the task type; inputting the normal trajectory sample set and the current trajectory segment into a dynamic time warping module to obtain the historical trajectory similarity; and embedding and encoding the task type and region type to obtain the encoded task type and encoded region type.

[0011] In one possible implementation, the normal trajectory sample set and the current trajectory segment are input into the dynamic time warping module to obtain the historical trajectory similarity, including: calculating the DTW distance between the current trajectory segment and each normal trajectory sample in the normal trajectory sample set; and selecting the smallest DTW distance as the historical trajectory similarity.

[0012] In one possible implementation, an anomaly score is obtained by scoring the trajectory anomaly probability of the personnel positioning trajectory temporal feature vector, including: inputting the personnel positioning trajectory temporal feature vector into a pre-trained LSTM autoencoder to obtain a reconstruction error; and performing normalization processing on the reconstruction error based on the Sigmoid activation function to obtain the anomaly score.

[0013] In one possible implementation, the training data for the pre-trained LSTM autoencoder is a temporal feature vector of a person's location trajectory labeled as normal behavioral data.

[0014] According to another aspect of this application, an IoT-based personnel positioning trajectory deviation control system for wind farms is provided, comprising: a positioning data acquisition module for acquiring raw positioning data, including personnel ID, timestamp, and coordinates; a context data association query module for querying work order data and environmental status data based on the raw positioning data; a multi-source data fusion construction module for constructing fused context data points based on the raw positioning data, work order data, and environmental status data; a time-series feature engineering module for performing dynamic time-series feature engineering on the time queue of the fused context data points to obtain a personnel positioning trajectory time-series feature vector; a trajectory anomaly scoring module for performing trajectory anomaly probability scoring on the personnel positioning trajectory time-series feature vector to obtain anomaly scores; and a dynamic hierarchical alarm decision module for performing dynamic hierarchical alarm decisions based on the anomaly scores to obtain alarm events.

[0015] Compared with existing technologies, the IoT-based method and system for controlling personnel positioning trajectory deviations in wind farms provided in this application first integrates real-time personnel positioning data, work order task data, and environmental status data to construct a time-series data queue rich in contextual information. Then, through dynamic time-series feature engineering, the kinematic features and behavioral pattern features of the trajectory are transformed into a time-series feature vector of the personnel positioning trajectory. Next, an unsupervised learning model is used to score the anomaly probability of the personnel positioning trajectory time-series feature vector, and finally, dynamic hierarchical alarms are implemented based on the quantified anomaly scores. In this way, by deeply integrating multi-source data and using a time-series data analysis model, accurate identification, quantitative assessment, and intelligent early warning of personnel behavioral risks are achieved, thereby overcoming the shortcomings of existing technologies. Attached Figure Description

[0016] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain the application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 The illustration shows a schematic flowchart of an IoT-based method for controlling personnel location trajectory deviation in wind farms according to an embodiment of this application.

[0018] Figure 2 The figure shows a schematic flowchart of step S2 in the IoT-based wind farm personnel positioning trajectory deviation control method according to an embodiment of this application.

[0019] Figure 3 The figure shows a schematic flowchart of step S4 in the IoT-based wind farm personnel positioning trajectory deviation control method according to an embodiment of this application.

[0020] Figure 4 The figure shows a schematic flowchart of step S41 in the IoT-based wind farm personnel positioning trajectory deviation control method according to an embodiment of this application.

[0021] Figure 5 The figure shows a schematic flowchart of step S42 in the IoT-based wind farm personnel positioning trajectory deviation control method according to an embodiment of this application.

[0022] Figure 6 The figure shows a schematic flowchart of step S5 in the IoT-based wind farm personnel positioning trajectory deviation control method according to an embodiment of this application.

[0023] Figure 7 The figure shows a schematic block diagram of an IoT-based wind farm personnel positioning trajectory deviation control system according to an embodiment of this application. Detailed Implementation

[0024] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0025] Figure 1 The illustration shows a schematic flowchart of a wind farm personnel positioning trajectory deviation control method and system based on the Internet of Things according to an embodiment of this application. Figure 1 As shown, this application provides a method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things, including: S1, acquiring raw positioning data, which includes personnel ID, timestamp, and coordinates; S2, querying work order data and environmental status data based on the raw positioning data; S3, constructing fused context data points based on the raw positioning data, work order data, and environmental status data; S4, performing dynamic temporal feature engineering on the time queue of the fused context data points to obtain a temporal feature vector of personnel positioning trajectory; S5, performing trajectory anomaly probability scoring on the temporal feature vector of personnel positioning trajectory to obtain anomaly scores; S6, performing dynamic hierarchical alarm decision-making based on the anomaly scores to obtain alarm events.

[0026] For example, in step S1, raw location data is acquired, including personnel ID, timestamp, and coordinates. It should be understood that the personnel ID, as a unique identifier, binds an abstract geographical location to a specific worker entity. It is the sole index for subsequent association queries of the worker's work orders, historical behavior patterns, and responsibility tracing. Without this, the system cannot distinguish the trajectories of different personnel, making subsequent personalized and context-dependent analysis impossible. Simultaneously, the timestamp assigns a temporal attribute to discrete location points, precisely recording the moment the location information was generated, making it possible to string together a series of independent location points into a movement trajectory with a temporal order. This is a prerequisite for analyzing dynamic temporal characteristics such as speed, acceleration, and dwell time, and is also the foundation for understanding the continuity and dynamic changes in personnel behavior. Coordinates, as the core carrier of personnel location in the spatial dimension, directly describe the physical location of the personnel and are the direct basis for determining whether they have deviated from the preset path, entered a dangerous area, and for spatial relationship analysis with other geographical elements within the wind farm. Therefore, only by combining these three elements can a complete data point with identity, time, and spatial attributes be formed, providing an effective data source for subsequent context fusion and deep trajectory analysis.

[0027] In one embodiment, acquiring raw location data includes: firstly, deploying IoT positioning infrastructure within the wind farm. This infrastructure typically includes multiple positioning base stations installed in key areas (such as inside and around wind turbine towers, substations, control centers, and main maintenance access routes). These base stations can employ various positioning technologies such as Ultra-Wideband (UWB), Bluetooth (BLEAOA), LoRa, or a combination of GPS / BeiDou differential positioning. Simultaneously, each on-site worker must wear or be equipped with a smart positioning terminal with a built-in communication module, such as a positioning badge, a positioning module integrated into a safety helmet, or a smart bracelet. When the control method is activated, the smart positioning terminal worn by the personnel will actively send radio signals to one or more surrounding positioning base stations at a preset frequency (e.g., once per second), or the base stations will actively poll the terminal. After receiving these signals, the positioning base stations will transmit key signal parameters, such as Time of Arrival (TOA), Time Difference of Arrival (TDOA), Angle of Arrival (AOA), or Received Signal Strength Indication (RSSI), to the backend positioning engine server in real time. After receiving signal data from at least one or more base stations, the positioning engine server immediately runs a pre-set positioning algorithm, such as trilateration based on TDOA or triangulation based on AOA, to accurately calculate the three-dimensional or two-dimensional coordinates of the smart positioning terminal at the current moment. After the coordinate calculation is completed, the server immediately associates and encapsulates the coordinates with the unique personnel ID bound to the smart positioning terminal that sent the signal and the server's current system timestamp, ultimately forming a structured raw positioning data record. This record is then stored in a real-time database or pushed to a message queue for subsequent steps.

[0028] For example, in step S2, work order data and environmental status data are queried based on the original location data. It should be understood that raw location data alone can only answer the question of when and where a person is, but cannot explain the rationality of their behavior or potential risks. For example, it is impossible to determine whether a person is moving normally along a predetermined route or has dangerously deviated from a safety passage based solely on a single coordinate point. By querying work order data, the type of task the person is currently performing and the pre-planned path can be obtained, providing a benchmark and basis for subsequent judgments on whether their behavior is in line with expectations. Without work order context, any trajectory analysis would be blind, unable to distinguish between normal work detours and abnormal, unexplained deviations. Similarly, by querying environmental status data, this method can determine the type of area where the person's current coordinates are located, such as a safe maintenance passage, a high-voltage hazard zone requiring specific authorization, or a rest area where prolonged stay is permitted. This provides crucial environmental semantic information for assessing the risk level of personnel behavior. Without environmental context, it is impossible to determine whether staying in place is a normal rest or a potential accident symptom. Therefore, this step, by dynamically linking the real-time location of personnel with their work tasks and environment, constructs a more information-rich analysis scenario, which is an indispensable prerequisite for subsequent accurate trajectory deviation determination and intelligent risk assessment.

[0029] In one embodiment, such as Figure 2 As shown, based on the original location data, querying work order data and environmental status data includes: S21, querying work order data based on personnel ID and timestamp, the work order data including the current personnel's task type and preset path; S22, querying environmental status data based on coordinates, the environmental status data including area ID and area type.

[0030] Specifically, after the system obtains raw location data containing personnel ID and timestamp, it uses these two fields as composite query keywords to initiate a data request to the database of the backend work order management system or Enterprise Resource Planning (ERP) system. The query logic involves retrieving valid work order records from the work order data table where the personnel ID matches one in the current data, and the planned start time of the task is earlier than or equal to the current timestamp, and the planned end time is later than or equal to the current timestamp. Once a unique work order record meeting the conditions is found, the system parses and extracts predefined fields from the record, primarily the task type and preset path. Next, it queries environmental status data based on coordinates, including the area ID and area type. Specifically, using the coordinate fields from the raw location data as query input, it initiates a spatial query request to the backend Geographic Information System (GIS) database. The digital map of the wind farm has been pre-modeled in the GIS, where different functional areas are divided into geofences (polygons) with attributes. The core of this spatial query is to execute a "point-in-polygon" geometric judgment algorithm to determine which geofence polygon the current coordinate point falls within. Once the algorithm successfully matches a polygon, the system reads and extracts the corresponding environmental status information from the polygon's attribute table, mainly the region ID and region type.

[0031] For example, in step S3, a fused context data point is constructed based on the original location data, work order data, and environmental status data. It should be understood that the information held by the system is fragmented: a location record describing the physical state, a work order record describing the business instruction, and an environmental record describing spatial attributes. Although these data are all related to the same point in time and the same person, they are independent at the data level. Without fusion, subsequent feature engineering steps would require repeatedly and temporarily associating and matching these fragmented data, which not only greatly increases the complexity and computational overhead of subsequent processing but also introduces the risk of data being out of sync or incorrectly associated in time. By constructing a fused data point, it is ensured that at any given timestamp, all contextual information related to the person is encapsulated in an indivisible whole, forming a complete snapshot of the person's state at that moment.

[0032] In one embodiment, constructing a fusion context data point based on raw location data, work order data, and environmental status data includes: taking the raw location data, work order data, and environmental status data as input, and extracting the corresponding values ​​from each set of data according to predefined mapping rules, and filling them into the corresponding fields of a standard data structure. For example, the personnel ID value from the raw location data is assigned to the "personnel_id" field in the fusion data point structure, the preset path value from the work order data is assigned to the "preset_path" field, the area type value from the environmental status data is assigned to the "area_type" field, and so on. The entire process can be understood as creating and initializing a data object or record. Once all fields are successfully filled, a complete, self-contained fusion context data point is successfully constructed.

[0033] For example, in step S4, dynamic temporal feature engineering is performed on the time queue of the fused context data points to obtain the temporal feature vector of the personnel positioning trajectory. It should be understood that the time queue of the fused context data points is essentially a sequence of records. Although each record (i.e., the fused context data point) contains rich contextual information, they are still isolated snapshots of the state. Observing only a single or a few data points cannot reveal the dynamic attributes, inherent patterns, and complex evolutionary patterns of personnel behavior over time. For example, a single data point deviating from a preset path may be a harmless detour or the beginning of a dangerous detour; this cannot be distinguished by the point alone. A state of remaining stationary may be normal equipment operation or a sign of a fall or unconsciousness. Traditional methods typically judge these isolated points based on simple rules, which is extremely rigid and has a high false alarm rate because it completely ignores the process and continuity of behavior.

[0034] Based on this, dynamic temporal feature engineering is performed on the time queue of fused contextual data points to obtain the temporal feature vector of personnel positioning trajectory. Dynamic temporal feature engineering no longer treats data points as independent events, but rather places them within a time window and examines them as a continuous process. Through in-depth analysis and calculation of this time queue, a series of features that cannot be obtained from single-point information and can profoundly characterize behavioral dynamics can be extracted. These features quantify the personnel's behavioral trajectory from multiple dimensions: First, from a kinematic dimension, instantaneous velocity, acceleration, and other indicators can be calculated to determine whether the personnel's movement state is stable, rapid, or stationary; second, from a behavioral pattern dimension, stationary dwell time and similarity to historical normal trajectories can be calculated to determine whether the current behavior conforms to routine work habits; third, from a task compliance dimension, deviation distance from the preset path can be calculated to determine whether the personnel have followed the work instructions. More importantly, by numerically encoding contextual information such as task type and area type and incorporating them into the consideration, the interpretation of these features is no longer isolated, but closely coupled with specific business scenarios and environmental risks. Ultimately, all these carefully extracted, highly discriminative features are combined into a fixed-length numerical vector, namely the temporal feature vector of the personnel positioning trajectory.

[0035] In one embodiment, such as Figure 3 As shown, dynamic temporal feature engineering is performed on the time queue of fused context data points to obtain the temporal feature vector of the personnel positioning trajectory, including: S41, kinematic feature extraction is performed on the time queue of fused context data points to obtain a kinematic feature set, the kinematic feature set including instantaneous velocity, instantaneous acceleration and stationary dwell time; S42, context and behavioral feature extraction is performed on the time queue of fused context data points to obtain a context feature set, the context feature set including preset path distance, historical trajectory similarity, encoded task type and encoded region type; S43, feature vector assembly is performed on the kinematic feature set and the context feature set to obtain the temporal feature vector of the personnel positioning trajectory.

[0036] Specifically, kinematic feature extraction is first performed on the time queue of fused context data points to mine the physical motion state of personnel from the geometric and temporal attributes of the trajectory. In one embodiment, such as Figure 4As shown, kinematic feature extraction is performed on the time queue of fusion context data points to obtain a kinematic feature set, including: S411, extracting the fusion context data point at time step t and the fusion context data point at time step (t-1) from the time queue of fusion context data points; S412, calculating the Euclidean distance and time difference between the fusion context data point at time step t and the fusion context data point at time step (t-1), and calculating the instantaneous velocity at time step t based on the Euclidean distance and time difference; S413, extracting the fusion context data point at time step (t-2) from the time queue of fusion context data points; S414, calculating the instantaneous velocity at time step (t-1) based on the fusion context data point at time step (t-2) and the fusion context data point at time step (t-1); S415, calculating the instantaneous acceleration based on the instantaneous velocity at time step (t-1) and the instantaneous velocity at time step t.

[0037] Specifically, to calculate instantaneous velocity, the system extracts the latest (time step t) fusion context data point and its immediate predecessor (time step t-1) from the time queue. By reading the coordinates of these two data points, the system calculates the linear distance between them using the Euclidean distance formula, and simultaneously calculates the time difference by reading their timestamps. Finally, dividing the spatial distance by the time difference yields the instantaneous velocity at time step t. To calculate instantaneous acceleration, the system requires velocity values ​​from two consecutive time steps. After calculating the instantaneous velocity at time step t, the system backtracks to calculate the instantaneous velocity at time step t-1 (requiring data points from time steps t-1 and t-2), then divides the difference between the two velocity values ​​by the corresponding time interval to obtain the instantaneous acceleration at time step t. For calculating dwell time, the system analyzes the instantaneous velocity sequence within the entire time window. By setting a very small stationary speed threshold (e.g., 0.1 m / s) to filter out minute speed values ​​caused by positioning drift, the system iterates through the speed sequence and counts the duration of consecutive speeds below this threshold. If the speed at time t is below the threshold, the dwell time is incremented; if it is above the threshold, the dwell time counter is reset to zero. The final total duration of consecutive dwell times is the current dwell time characteristic. In other words, the instantaneous speed time sequence is compared with the stationary speed threshold, and the total duration of consecutive times below the stationary speed threshold is counted as the dwell time.

[0038] Next, contextual and behavioral features are extracted from the time queue of fused contextual data points to analyze the correlation between personnel behavior and preset norms, historical patterns, and environmental attributes. In one embodiment, such as... Figure 5As shown, context and behavioral features are extracted from the time queue of fused context data points to obtain a context feature set, including: S421, extracting coordinates and preset paths from the fused context data points; S422, calculating the shortest vertical distance from the coordinates to the preset path as the preset path distance; S423, extracting the current trajectory segment composed of N coordinates from the time queue of fused context data points; S424, retrieving a normal trajectory sample set from the historical normal trajectory database based on the task type; S425, inputting the normal trajectory sample set and the current trajectory segment into the dynamic time warping module to obtain the historical trajectory similarity; S426, embedding and encoding the task type and region type to obtain the encoded task type and encoded region type.

[0039] Specifically, to calculate the preset path distance, the system extracts the preset path information from the latest fusion context data points, including real-time coordinates and associated work order data. The preset path itself is a polyline segment composed of a series of coordinate points. The calculation process involves finding the shortest perpendicular distance from the current real-time coordinate point to this polyline. In computational geometry, this requires traversing each segment of the preset path, calculating the distance from each point to the segment, and taking the minimum value as the final preset path distance feature. Calculating historical trajectory similarity involves a more complex pattern matching process. The system first extracts the task type from the current fusion context data points and, based on this task type, retrieves a representative set of historical trajectory samples marked as normal from the background historical normal trajectory database. Simultaneously, the system extracts the current trajectory segment composed of the N most recent coordinate points from the current time queue. Then, the system uses the Dynamic Time Warping (DTW) algorithm to compare the current trajectory segment with each historical normal trajectory sample one by one. The DTW algorithm effectively calculates the similarity between two time sequences of unequal duration by finding an optimal alignment path to minimize the cumulative distance between corresponding points in the two sequences. After completing all alignments, the system obtains a set of DTW distance values ​​and selects the minimum value as the final historical trajectory similarity feature. The smaller this value, the more closely the current behavior pattern matches the known normal pattern. In other words, it calculates the DTW distance between the current trajectory segment and each normal trajectory sample in the normal trajectory sample set, and selects the smallest DTW distance as the historical trajectory similarity.

[0040] Then, since the task type and region type contained in the fusion context data points are text-based category data, they cannot be directly used by the mathematical model and therefore must be encoded. This application preferably employs embedding encoding technology. The system pre-trains or uses a pre-trained embedding layer that maps each discrete category label (such as "F03 wind turbine routine inspection" or "high-voltage danger zone") to a low-dimensional, dense floating-point vector. The advantage of this encoding method is that it can learn the semantic relationships between categories, making the vectors of categories with similar meanings (such as "troubleshooting" and "emergency repair") closer in the vector space. Through this process, textual information is transformed into numerical vectors with mathematical meaning.

[0041] Finally, after extracting and calculating all the above features, the system will assemble the kinematic feature set and the context feature set into feature vectors. This is a concatenation operation. The system concatenates all the obtained numerical features, including instantaneous velocity, instantaneous acceleration, dwell time, preset path distance, historical trajectory similarity, and all components of the embedded and encoded task type vector and region type vector, in a predefined, fixed order to form a single, higher-dimensional one-dimensional numerical vector, which is the temporal feature vector of the personnel positioning trajectory.

[0042] Here, considering that directly assembling the kinematic feature set and the context feature set into feature vectors to obtain the temporal feature vector of the personnel positioning trajectory, directly concatenating kinematic / behavioral features with clear numerical meaning (such as speed, distance) with one-hot encoded features (task type, region type) that do not have numerical meaning and only serve as category identifiers into a feature vector and feeding it into an anomaly detection model (such as a standard LSTM autoencoder) may lead to conceptual confusion and overfitting risks. That is, the model cannot distinguish which features describe behavior and which define context. For example, it may incorrectly learn that a certain context (such as task type = [0,0,1], representing high-altitude operations) is itself an abnormal pattern, because the sample size of high-altitude operations in the training data may be far less than that of regular inspections. Therefore, as long as it is a high-altitude operation, no matter how standard the personnel behavior is, the model may give an excessively high anomaly score. This is obviously not detecting abnormal behavior in a specific context, but rather treating the context itself as an anomaly.

[0043] Therefore, for task type and region type based on embedding encoding in context features, it is necessary to separate them from other features with clear numerical meaning, including kinematic features and context features. Considering the dynamic hierarchical alarm decision based on task type and region type to determine alarm threshold, the context features based on task type and region type are used as conditions for encoding numerical features.

[0044] That is, in another embodiment, assembling feature vectors from the kinematic feature set and the context feature set to obtain the temporal feature vector of the personnel positioning trajectory includes: encoding numerical features other than task type and region type into numerical feature vectors; encoding task type and region type into conditional feature vectors; using the conditional feature vectors as a conditional evidence log-likelihood representation for the numerical feature vectors, and calculating a conditional evidence log-likelihood vector; calculating the reconstruction difference between the numerical feature vectors and the conditional evidence log-likelihood vectors to obtain a reconstruction difference vector; and calculating the temporal feature vector of the personnel positioning trajectory based on the conditional evidence log-likelihood vector and the reconstruction difference vector.

[0045] Specifically, for example, numerical features other than task type and region type are encoded as numerical feature vectors. The task type and region type are encoded as conditional feature vectors. ,and and If they have the same length, then As a response to distribution The conditional evidence log-likelihood representation is used to calculate the conditional evidence log-likelihood vector, i.e.: ;in, Represents numerical eigenvectors. Represents the conditional eigenvector. This represents vector addition. This represents element-wise multiplication. express The element-wise inverse (i.e., the reciprocal of each element). Let represent the log-likelihood vector of conditional evidence.

[0046] Here, when using one-hot encoding for task type and region type, unified encoding is required, for example... This is used to convert it to a continuous positive value range. The conditional reconstruction probability score is represented as follows: As a condition, it directly acts on the distribution. The above serves as a probability constraint for distribution, used to measure the distribution within a given context. Down The conditional probability representation. And... Using KL divergence as a regularization term, that is, the distribution... Considered a prior distribution, based on evidence Posterior distribution probability representation Under the influence of the distribution The posterior effect.

[0047] Thus, based on conditions To make Show the distribution The conditional evidence is logarithmic, then, considering if the distribution... In context The more anomalous the input appears, the more it deviates from the normal behavioral manifold in that context, making it more difficult for the model to accurately reconstruct it. This means the difference between the reconstructed result and the original input will be greater. Therefore, the reconstruction difference can be directly used, i.e., calculating the reconstruction difference between the numerical feature vector and the conditional evidence log-likelihood vector to obtain the reconstruction difference vector. , This represents vector subtraction. Simultaneously, based on the above log-likelihood representation, spatial reconstruction is performed; that is, based on the conditional evidence log-likelihood vector and the reconstructed difference vector, the temporal feature vector of the personnel positioning trajectory is calculated, i.e.: ;in, Let represent the log-likelihood vector of conditional evidence. Represents the reconstructed difference vector. The L2 norm of a vector. This represents element-wise multiplication, which in the above formula applies to vectors. Each element is multiplied by the value of the exponential function. This represents the temporal feature vector of the personnel's location trajectory.

[0048] That is, by further utilizing the reconstructed probability distribution as the probability density response at the feature space level, the temporal feature vector of the personnel positioning trajectory is obtained, thereby clarifying its behavior-context probability matching characteristics. For example, the response deviation between the actually observed behavioral features (e.g., a speed of 0.1 m / s) and the behavioral features that the model believes should occur in the current context (e.g., a speed reconstructed by the model of 0.12 m / s) is considered. Simultaneously, for uncertainties based on reconstruction differences, if the model is inherently less confident in predicting behavior in a certain context (e.g., a complex maintenance task),... If the distribution is large, it will directly affect the distribution; conversely, if it is in a specific context (such as rest), the distribution will be amplified as a whole.

[0049] Therefore, assuming that the features corresponding to the behavioral data exist on a high-dimensional manifold, conditional context features are used as the conditionalized manifold. For example, the normal behavioral manifold in a rest area is low speed and stillness, while the normal behavioral manifold in an emergency evacuation route when an alarm sounds is high speed and directional movement. By injecting contextual features into the behavioral encoding in the form of conditions and prior evidence, the model can dynamically adjust the definition of normal behavior, avoiding misjudging the context itself as abnormal. In this way, the model can accurately distinguish between reasonable deviations (such as prolonged stays for ease of operation during maintenance tasks) and dangerous anomalies (such as sudden high-speed movement in a rest area), significantly improving accuracy. Furthermore, because the model understands the context, it will no longer easily issue alarms simply because the task type is rare or personnel enter infrequently visited areas, thus reducing the false alarm rate.

[0050] For example, in step S5, an anomaly probability score is obtained by performing a trajectory anomaly probability score on the personnel positioning trajectory temporal feature vector. It should be understood that while the personnel positioning trajectory temporal feature vector comprehensively describes the behavioral state of personnel over a period of time, it is itself a set of numerical values. Human managers cannot directly observe this vector to quickly and accurately determine whether the behavior it represents is abnormal and the severity of the anomaly. Traditional safety management methods may attempt to set static, hard rule thresholds for each dimension of this vector, such as "speed exceeding X meters / second" or "deviation from the path exceeding Y meters" triggering an alarm. The drawback of this method is obvious: it cannot capture the complex, non-linear relationships between multi-dimensional features. For example, high-speed movement is normal in maintenance channels but highly abnormal in narrow equipment areas. This rule-based system is extremely rigid and difficult to adapt to the changing operating scenarios of wind farms, resulting in a large number of false alarms and missed alarms.

[0051] Therefore, to address the aforementioned issues, this application employs trajectory anomaly probability scoring on the temporal feature vectors of personnel positioning trajectories to obtain an anomaly score. This trajectory anomaly probability scoring no longer relies on manually set, fragile rules, but instead utilizes a well-trained deep learning model to automatically and holistically assess the deviation between the behavioral pattern represented by the entire feature vector and the normal behavioral pattern. This model, through learning from massive amounts of normal operational data, has internalized the distribution patterns of normal behavior in the feature space. When a new behavioral feature vector is input, the model can provide a probabilistic score that accurately quantifies the likelihood that the current behavior is abnormal. This anomaly score, ranging from 0 to 1, not only solves the binary judgment problem of whether something is abnormal, but more importantly, it provides a continuous measure of the degree of anomaly, offering a solid and objective mathematical basis for subsequent refined, tiered alarm decisions. This transforms safety management from a passive, extensive response to a proactive, precise, and quantified risk-based intervention, thereby significantly improving the intelligence level and efficiency of wind farm operation and maintenance safety management.

[0052] In one embodiment, such as Figure 6 As shown, the abnormality score is obtained by scoring the trajectory anomaly probability of the personnel positioning trajectory temporal feature vector. The process includes: S51, inputting the personnel positioning trajectory temporal feature vector into a pre-trained LSTM autoencoder to obtain the reconstruction error, wherein the training data of the pre-trained LSTM autoencoder is the personnel positioning trajectory temporal feature vector labeled as normal behavior data; S52, normalizing the reconstruction error based on the Sigmoid activation function to obtain the abnormality score.

[0053] Specifically, the model architecture of the LSTM autoencoder is first described. The LSTM autoencoder is a neural network specifically designed for processing time-series data, structurally consisting of two symmetrical parts: an encoder and a decoder. The encoder's role is to compress the input sequence of temporal feature vectors representing personnel location trajectories, extracting its core temporal dynamic information and encoding it into a fixed-length, low-dimensional intermediate representation, commonly referred to as the context vector. Specifically, the encoder consists of one or more stacked LSTM layers. When a feature vector sequence representing the past N time steps is input, the first LSTM layer of the encoder processes the feature vectors sequentially, starting from the first time step. At each time step, the LSTM unit not only produces an output but also updates its internal hidden state and cell state. This cell state is crucial for the LSTM network to capture long-term dependencies; it acts like a conveyor belt, allowing information to be transmitted over long distances in the sequence without significant attenuation. The final hidden state and cell state output by the LSTM unit at the last time step in the sequence together constitute a condensed representation of the entire input sequence, i.e., the context vector.

[0054] The decoder part of the model performs the opposite function to the encoder. Its task is to receive the context vector generated by the encoder and, based on this, attempt to decompress the information and reconstruct an output sequence that is as consistent as possible with the original input sequence. The decoder structure also consists of one or more stacked LSTM layers. In implementation, a repeating vector layer is typically used to copy the single context vector output by the encoder N times, forming a sequence of the same length as the input sequence, which serves as the input to the first LSTM layer of the decoder. After receiving this repeated context information, the LSTM layer of the decoder begins to generate the reconstructed feature vector step by step. To ensure that the decoder outputs a vector with the same dimension as the original feature vector at each time step, a temporally distributed fully connected layer is usually connected after the output of the decoder's LSTM layer. This layer applies a standard fully connected layer to each time step of the sequence, thus mapping the output dimension of the LSTM back to the dimension of the original feature vector.

[0055] Before applying this model, it must be pre-trained. The training process follows the unsupervised learning paradigm, using only the feature vector sequences generated from labeled normal work trajectories as training data. The goal of training is to minimize the difference between the model input and the model output (i.e., the reconstructed sequence), which is typically measured using the mean squared error loss function. Through training with a large number of normal samples, the model is optimized into an expert in normal behavior patterns, proficient in how to efficiently compress and perfectly reconstruct the feature vector sequences corresponding to various normal behaviors.

[0056] Once the model is trained, it can be used for real-time trajectory anomaly probability scoring. The specific implementation process is as follows: First, the temporal feature vector of the personnel location trajectory is input into the pre-trained LSTM autoencoder. The model then performs forward propagation, where the encoder compresses the input sequence, and the decoder reconstructs the compressed representation, ultimately outputting a reconstructed feature vector sequence. Next, the system calculates the reconstruction error between the original input sequence and the reconstructed sequence output by the model. This error is calculated using the same loss function as during training, i.e., the mean of the sum of the squares of the differences between the two sequences at each time step and in each dimension. Since the model has only learned normal patterns, when the input sequence represents an abnormal behavior, its data distribution does not match the model's learned knowledge base, making it difficult for the model to effectively compress and reconstruct it. This results in a significant difference between the original input and the reconstructed output, i.e., a large reconstruction error value. Conversely, if the input is normal behavior, the reconstruction error will be very small. Finally, to transform this variable reconstruction error value into a standardized and easily understandable anomaly score, the system normalizes the reconstruction error. A preferred implementation is to use the Sigmoid activation function, which smoothly maps any real number to the interval between 0 and 1. By using the reconstruction error as input to the Sigmoid function, the output value is the final anomaly score.

[0057] For example, in step S6, dynamic hierarchical alarm decisions are made based on anomaly scores to obtain alarm events. It should be understood that the anomaly score itself does not directly guide the actions taken by wind farm safety managers. An event with a score of 0.75 and an event with a score of 0.95, while both indicating the existence of an anomaly, have significantly different potential risk levels and required response measures. Traditional alarm systems often employ a black-and-white model with a single threshold; any event exceeding the threshold triggers an alarm of the same level. This leads to two serious drawbacks: firstly, frequent alarms for low-risk, minor anomalies cause alarm fatigue, desensitizing managers to truly serious alarms; secondly, it fails to highlight extreme dangerous events requiring immediate, highest-priority intervention, potentially wasting valuable emergency response time on minor issues.

[0058] To address the aforementioned issues, this application employs a dynamic hierarchical alarm decision-making system based on anomaly scores. This system introduces a multi-layered decision-making mechanism, enabling the system to accurately differentiate and respond to varying risk levels based on anomaly scores. This not only enhances the value and guidance of alarm information but, more importantly, helps optimize the allocation of safety management resources. For low-risk events, the system may simply record the event or issue a mild alert; for medium-risk events, it may notify the team leader for verification; and for high-risk events, it may immediately trigger the highest-level emergency response plan, notifying all relevant parties to intervene. This hierarchical response mechanism, tailored to the risk level, ensures that management personnel's attention is focused on the most critical areas, thereby significantly improving the efficiency and intelligence of the entire wind farm's operation and maintenance management while ensuring personnel safety.

[0059] In one embodiment, dynamic hierarchical alarm decision-making based on anomaly scores to obtain alarm events includes: First, configuring a set of pre-configured alarm thresholds for classifying risk levels. These thresholds are not single values, but rather define multiple continuous and non-overlapping numerical ranges, each corresponding to a specific alarm level. For example, a low-risk range (anomaly score within [0.70, 0.85)), a medium-risk range (anomaly score within [0.85, 0.95)), a high-risk range (anomaly score greater than or equal to 0.95), and a no-risk range (anomaly score less than 0.70) can be preset. When a new anomaly score is input to the decision module, the system performs a comparison operation to determine which preset range the score falls into. This judgment process is typically implemented in software as a series of conditional branch statements. The system first checks whether the score is greater than or equal to the lower threshold of the highest-risk range; if not, it continues to check whether it is greater than or equal to the lower threshold of the medium-risk range; and so on, until the range to which it belongs is found.

[0060] Once the alarm level corresponding to the anomaly score is determined, the system does not simply issue a signal, but immediately begins constructing a complete and structured "alarm event." This alarm event is a standardized data record whose purpose is to comprehensively and clearly convey all relevant information about the anomaly, facilitating rapid understanding and handling by the recipient. A preferred implementation includes at least the following key fields: an event ID uniquely identifying the event; the personnel ID that triggered the event; the timestamp of the event; the alarm level determined through decision-making; the original anomaly score used as the basis for decision-making; and an automatically generated, human-readable alarm message. This message can be dynamically generated based on the alarm level and context information. For example, for a high-risk event, the message could be: "Personnel A007, while performing an inspection task on wind turbine F03, deviated significantly from the normal pattern, with a risk score of 0.97. Please verify immediately!"

[0061] After the alarm event is fully constructed, it is distributed and pushed. The system sends alarm events of different levels to different target channels according to pre-configured alarm rules. For example, low-risk events may only be written to the system's log database for subsequent auditing; medium-risk events may be sent to the person's direct supervisor via the company's instant messaging tool; while high-risk events will be immediately pushed to the large screen in the wind farm's central monitoring room and simultaneously notified to all relevant emergency response personnel via SMS, telephone, or a dedicated mobile application. Through this series of rigorous implementation steps, this invention successfully transforms a quantified risk score into a clearly categorized, information-rich, and precisely delivered alarm event, thus completing a full closed loop from data analysis to intelligent decision-making and action.

[0062] In summary, the IoT-based method for controlling personnel positioning trajectory deviation in wind farms provided in this application first integrates real-time personnel positioning data, work order task data, and environmental status data to construct a time-series data queue rich in contextual information. Then, through dynamic time-series feature engineering, the kinematic features and behavioral pattern features of the trajectory are transformed into a time-series feature vector of the personnel positioning trajectory. Next, an unsupervised learning model is used to score the anomaly probability of the personnel positioning trajectory time-series feature vector, and finally, dynamic hierarchical alarms are implemented based on the quantified anomaly scores. In this way, by deeply integrating multi-source data and applying a time-series data analysis model, accurate identification, quantitative assessment, and intelligent early warning of personnel behavioral risks are achieved, thereby overcoming the shortcomings of existing technologies.

[0063] This application also provides an IoT-based personnel positioning trajectory deviation control system for wind farms, used to execute the aforementioned IoT-based personnel positioning trajectory deviation control method for wind farms, such as... Figure 7As shown, the IoT-based wind farm personnel positioning trajectory deviation control system 700 includes: a positioning data acquisition module 710 for acquiring raw positioning data, which includes personnel ID, timestamp, and coordinates; a context data association query module 720 for querying work order data and environmental status data based on the raw positioning data; a multi-source data fusion construction module 730 for constructing fused context data points based on the raw positioning data, work order data, and environmental status data; a time-series feature engineering module 740 for performing dynamic time-series feature engineering on the time queue of the fused context data points to obtain a personnel positioning trajectory time-series feature vector; a trajectory anomaly scoring module 750 for performing trajectory anomaly probability scoring on the personnel positioning trajectory time-series feature vector to obtain anomaly scores; and a dynamic hierarchical alarm decision module 760 for performing dynamic hierarchical alarm decisions based on the anomaly scores to obtain alarm events.

[0064] This application also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-mentioned method steps to implement the IoT-based wind farm personnel positioning trajectory deviation control method provided in the above embodiments.

[0065] This application also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to achieve the IoT-based method for controlling personnel positioning trajectory deviation in wind farms provided in the above embodiments.

[0066] In this application, the system, computer-readable storage medium, or computer program product provided in the embodiments are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0067] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments.

[0068] The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things, characterized in that, include: Obtain raw location data, which includes personnel ID, timestamp, and coordinates; Based on the original location data, query work order data and environmental status data; Based on the original location data, work order data, and environmental status data, construct fused context data points; Dynamic temporal feature engineering is performed on the time queue of fused context data points to obtain temporal feature vectors of personnel positioning trajectories; Anomaly scores are obtained by scoring the trajectory anomaly probability of the temporal feature vector of the personnel positioning trajectory. Dynamic hierarchical alarm decisions are made based on anomaly scores to obtain alarm events.

2. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 1, characterized in that, Based on the original location data, query work order data and environmental status data, including: Work order data can be queried based on personnel ID and timestamp, where the work order data includes the current personnel's task type and preset path; Environmental status data is queried based on coordinates, and the environmental status data includes region ID and region type.

3. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 2, characterized in that, Dynamic temporal feature engineering is performed on the time queue of fused context data points to obtain temporal feature vectors of personnel location trajectories, including: Kinematic features are extracted from the time queue of fused context data points to obtain a kinematic feature set, which includes instantaneous velocity, instantaneous acceleration, and dwell time. Context and behavioral features are extracted from the time queue of fused context data points to obtain a context feature set, which includes preset path distance, historical trajectory similarity, encoded task type, and encoded region type. The kinematic feature set and the context feature set are assembled into feature vectors to obtain the temporal feature vector of the personnel positioning trajectory.

4. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 3, characterized in that, Kinematic feature extraction is performed on the time queue of fused context data points to obtain a kinematic feature set, including: Extract the fusion context data point at time step t and the fusion context data point at time step (t-1) from the time queue of the fusion context data points; Calculate the Euclidean distance and time difference between the fusion context data point at time step t and the fusion context data point at time step (t-1), and calculate the instantaneous velocity at time step t based on the Euclidean distance and time difference; Extract the fusion context data point at time step t-2 from the time queue of the fusion context data points; Calculate the instantaneous velocity at time step t-1 based on the fusion context data points at time step t-2 and t-1. Calculate the instantaneous acceleration based on the instantaneous velocity at time step (t-1) and time step t.

5. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 3, characterized in that, Kinematic feature extraction is performed on the time queue of fused context data points to obtain a kinematic feature set, including: comparing the time queue of instantaneous velocity with a stationary velocity threshold, and counting the total length of continuous time below the stationary velocity threshold as the stationary dwell time.

6. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 3, characterized in that, Context and behavioral features are extracted from the time queue of fused context data points to obtain a context feature set, including: Extract coordinates and preset paths from fused context data points; Calculate the shortest perpendicular distance from the coordinates to the preset path as the preset path distance; Extract the current trajectory segment consisting of N coordinates from the time queue of fused context data points; Based on task type, retrieve normal trajectory sample sets from historical normal trajectory database; The normal trajectory sample set and the current trajectory segment are input into the dynamic time warping module to obtain the historical trajectory similarity; The task type and region type are embedded and encoded to obtain the encoded task type and region type.

7. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 6, characterized in that, The normal trajectory sample set and the current trajectory segment are input into the dynamic time warping module to obtain the historical trajectory similarity, including: Calculate the DTW distance between the current trajectory segment and each normal trajectory sample in the normal trajectory sample set; The smallest DTW distance is selected as the historical trajectory similarity.

8. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 1, characterized in that, Anomaly scores are obtained by scoring the trajectory anomaly probability of the temporal feature vector of the personnel positioning trajectory, including: The temporal feature vector of the personnel positioning trajectory is input into a pre-trained LSTM autoencoder to obtain the reconstruction error; The reconstruction error is normalized using a Sigmoid activation function to obtain the anomaly score.

9. The method for controlling personnel positioning trajectory deviation in wind farms based on the Internet of Things according to claim 8, characterized in that, The training data for the pre-trained LSTM autoencoder consists of temporal feature vectors of human location trajectories labeled as normal behavioral data.

10. A wind farm personnel positioning trajectory deviation control system based on the Internet of Things, characterized in that, include: The location data acquisition module is used to acquire raw location data, which includes personnel ID, timestamp, and coordinates. The context data association query module is used to query work order data and environmental status data based on the original location data; The multi-source data fusion construction module is used to construct fusion context data points based on raw location data, work order data, and environmental status data; The temporal feature engineering module is used to perform dynamic temporal feature engineering on the time queue of fused context data points to obtain temporal feature vectors of personnel positioning trajectories. The trajectory anomaly scoring module is used to score the trajectory anomaly probability of the temporal feature vector of the personnel positioning trajectory in order to obtain an anomaly score. The dynamic hierarchical alarm decision module is used to make dynamic hierarchical alarm decisions based on anomaly scores to obtain alarm events.

Images