Abnormal state detection method and system for wind driven generator system

Abstract

The invention relates to the technical field of wind power operation and maintenance management, and particularly provides an abnormal state detection method and system for a wind driven generator system. Joint feature extraction is carried out on the multi-source monitoring data flow through a feature coding network, a state evolution feature sequence and a cross-source association feature set are generated, a pre-trained anomaly diagnosis model is called to carry out fusion analysis on the features, and an anomaly diagnosis result containing anomaly confidence distribution information is generated; and after the abnormal type, the time dimension development track and the space dimension influence component identifier are analyzed, an equipment early warning instruction containing time and space positioning information is generated and sent to the target management terminal, so that the abnormal state of the wind driven generator system can be comprehensively and accurately detected, detailed abnormal information is provided in time, and the safety of the wind driven generator system is improved. And the equipment fault risk and the operation and maintenance cost are effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power operation and maintenance, in particular to an abnormal state detection method and system for a wind turbine system. BACKGROUND

[0002] In the field of wind power generation, the wind turbine system is the core equipment for converting wind energy into electrical energy, and its stable operation is crucial for the reliability and continuity of power supply. However, due to the long-term exposure of the wind turbine system to complex and variable natural environments, it faces various harsh conditions such as strong winds, sandstorms, and sudden temperature changes, resulting in a high probability of abnormal states of the equipment.

[0003] Currently, the abnormal state detection of the wind turbine system mainly relies on a single type of monitoring data, such as independent analysis of vibration signals or temperature signals. However, this single data source detection method has obvious limitations, as the abnormality of the wind turbine system is often caused by multiple factors, and a single data source cannot fully reflect the actual operating state of the equipment. At the same time, the existing technology lacks effective joint feature extraction methods when processing multi-source monitoring data, and cannot fully exploit the internal relationships between different monitoring signals, resulting in low accuracy and reliability of the abnormal diagnosis. In addition, the existing abnormal diagnosis results can only provide simple abnormality judgments, and cannot analyze the type of abnormality, the development trajectory in the time dimension, and the affected components in the spatial dimension, making it difficult for maintenance personnel to take targeted maintenance measures in a timely and accurate manner, increasing the risk of equipment failure and operation and maintenance costs. SUMMARY

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present application, the present application provides an abnormal state detection method for a wind turbine system, the method comprising: obtaining a multi-source monitoring data stream of the wind turbine system, the multi-source monitoring data stream containing three types of continuous monitoring records with timestamp alignment, i.e., vibration sensing signals, temperature sensing signals, and rotational speed sensing signals; performing joint feature extraction processing on the multi-source monitoring data stream through a feature encoding network to generate a state evolution feature sequence reflecting the operating state of the equipment and a cross-source correlation feature set, wherein the state evolution feature sequence contains pattern representations of changes in each monitoring signal over time, and the cross-source correlation feature set contains synchronous change relationship representations between different monitoring signals; calling a pre-trained abnormal diagnosis model to perform fusion analysis processing on the state evolution feature sequence and the cross-source correlation feature set to generate an abnormal diagnosis result of the equipment state, the abnormal diagnosis result containing confidence distribution information of abnormal occurrence; resolve the abnormal type of the wind turbine system and the development trajectory of the abnormality in the time dimension and the affected component identification in the spatial dimension based on the abnormal diagnosis result, wherein the development trajectory contains a time sequence description of the abnormality from initial occurrence to stable appearance, and the affected component identification contains specific component labels directly affected by the abnormality; generate a device warning instruction containing time and space positioning information according to the abnormal type, the development trajectory, and the affected component identification, and send the device warning instruction to a target management terminal to trigger a maintenance response operation.

[0005] In still another aspect, the present application also provides an abnormal state detection system for a wind turbine system, which comprises a processor and a machine-readable storage medium, the machine-readable storage medium is connected with the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to realize the above-mentioned method.

[0006] Based on the above aspects, the present application can comprehensively and accurately generate a state evolution feature sequence and a cross-source association feature set reflecting the running state of the device by obtaining a multi-source monitoring data stream containing three types of continuous monitoring records with timestamp alignment of vibration sensing signals, temperature sensing signals, and rotating speed sensing signals, and using a feature coding network to perform joint feature extraction processing. The internal relationship between different monitoring signals is fully considered, a pre-trained abnormal diagnosis model is called to perform fusion analysis processing on the state evolution feature sequence and the cross-source association feature set, an abnormal diagnosis result containing confidence distribution information of abnormal occurrence is generated, and the accuracy and reliability of abnormal diagnosis are greatly improved. Based on the abnormal diagnosis result, the abnormal type of the wind turbine system and the development trajectory of the abnormality in the time dimension and the affected component identification in the spatial dimension are resolved, so that the maintenance personnel can comprehensively and deeply understand the abnormal condition of the device. Finally, a device warning instruction containing time and space positioning information is generated according to these information, and is sent to a target management terminal to trigger a maintenance response operation, realizing timely and accurate early warning and maintenance of the abnormal state of the wind turbine system, effectively reducing the fault risk and operation and maintenance cost of the device, and improving the running efficiency and reliability of the wind power generation system. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 is the execution flow diagram of the abnormal state detection method for the wind turbine system provided by the embodiment of the present application.

[0008] Figure 2 is the schematic diagram of exemplary hardware and software components of the abnormal state detection system for the wind turbine system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0009] The application will be described in detail below with reference to the accompanying drawings, Figure 1 is a flowchart of an abnormal state detection method for a wind turbine system provided by an embodiment of the application. The abnormal state detection method for a wind turbine system will be described in detail below.

[0010] Step S110: Obtain a multi-source monitoring data stream of the wind turbine system, which contains three types of continuous monitoring records with timestamp alignment, including vibration sensing signals, temperature sensing signals, and rotational speed sensing signals.

[0011] In this embodiment, the vibration sensing signals are mainly used to monitor the mechanical vibration of each component in the wind turbine system. In detail, during the operation of the wind turbine, the normal operation and abnormal conditions of the mechanical components will produce different patterns of vibration. For example, problems such as gear wear and bearing looseness of the gearbox will cause changes in the vibration signal. Vibration sensors are usually installed on key components of the wind turbine, such as the gearbox, generator bearing, etc., to reflect the operating state of these components by collecting vibration signals in real time.

[0012] The temperature sensing signals reflect the temperature changes of each component in the wind turbine system. In detail, both excessively high or low temperatures can indicate that the system is abnormal. For example, the generator will generate heat during long-term operation, and if the heat dissipation system fails, the temperature of the generator will rise. Temperature sensors will be placed on components that are prone to heat generation, such as generators, gearboxes, transformers, etc., to monitor the temperature of these components in real time.

[0013] The rotational speed sensing signal records the rotational speed of the wind wheel. The rotational speed of the wind wheel is closely related to the wind strength, the output power of the generator, and other factors. Under different wind conditions, the wind wheel needs to operate at an appropriate rotational speed to ensure efficient power generation of the generator. The rotational speed sensor is generally installed on the main shaft of the wind wheel to obtain the rotational speed information of the wind wheel by measuring the rotational speed of the main shaft.

[0014] In order to ensure the quality of the multi-source monitoring data stream, the collected signals need to be timestamp aligned. In detail, due to the differences in sampling frequency and data transmission time of different sensors, the timestamps of these signals need to be calibrated to ensure their consistency in the time dimension. This can be achieved through hardware synchronization or software algorithms. For example, a high-precision clock source can be used to synchronize the sampling times of each sensor, or an algorithm can be used to adjust the timestamps after data collection, so that different types of signals have accurate corresponding relationships at the same time point. After timestamp alignment, a multi-source monitoring data stream containing vibration sensing signals, temperature sensing signals, and rotational speed sensing signals is obtained.

[0015] Step S120: Joint feature extraction processing of the multi-source monitoring data stream is performed through a feature coding network to generate a state evolution feature sequence reflecting the running state of the equipment and a cross-source association feature set, wherein the state evolution feature sequence contains mode representations of each monitoring signal changing over time, and the cross-source association feature set contains synchronous change relationship representations between different monitoring signals.

[0016] After obtaining the multi-source monitoring data stream, joint feature extraction processing needs to be performed on it through a feature coding network to generate a state evolution feature sequence and a cross-source association feature set that can reflect the running state of the equipment. The feature coding network can convert the original multi-source monitoring data stream into more representative and analyzable features.

[0017] Step S121: The vibration sensing signal is input into a vibration encoder of the feature coding network, and the vibration encoder performs time series pattern learning on vibration signal values of consecutive timestamps through multi-layer time convolution operations to generate vibration time series features containing local fluctuation patterns and global trend patterns.

[0018] In this step, the main task of the vibration encoder is to perform time series pattern learning on vibration signal values of consecutive timestamps to extract local fluctuation patterns and global trend patterns in the vibration signal.

[0019] The multi-layer time convolution operation can be understood as a sliding window calculation method, which performs convolution operations on the vibration signal in the time dimension to capture local features in the signal. Through multi-layer convolution operations, features at different levels can be gradually extracted. For example, the first convolution layer can extract short-term fluctuation features in the vibration signal, and subsequent convolution layers can further extract more complex features such as long-term trend changes and periodic features based on this.

[0020] When performing time convolution operations, the size and step of the convolution kernel need to be determined. The size of the convolution kernel determines the range of local features that can be captured by the convolution operation, while the step affects the sliding speed of the convolution operation in the time dimension. Different convolution kernel size and step settings will have different effects on the extracted features. For example, a larger convolution kernel can capture more extensive local features, but may lose some detailed information; while a smaller convolution kernel can extract more fine-grained local features, but the computational complexity may be greater.

[0021] During the multi-layer time convolution operation, an activation function can also be introduced to increase the non-linear representation ability of the model. For example, a common activation function is the ReLU (Rectified Linear Unit) function, which can convert negative inputs to zero while keeping positive inputs unchanged. Through the action of the activation function, the model can learn more complex feature patterns.

[0022] After multiple layers of time convolution operations, the vibration encoder generates vibration time series features that contain both local fluctuation patterns and global trend patterns. Local fluctuation patterns reflect changes in the vibration signal over a short period of time, such as sudden vibration peaks or fluctuations. Global trend patterns represent the changing trend of the vibration signal over a longer period of time, such as gradual increases or decreases in vibration amplitude.

[0023] In detail, before performing multiple layers of time convolution operations, the parameters of the convolution layer need to be determined, including the number, size, and step of the convolution kernel. The number of convolution kernels determines the number of feature maps output by the convolution layer, and each convolution kernel can extract different feature patterns. For example, setting multiple convolution kernels can simultaneously extract multiple local features in the vibration signal.

[0024] The size of the convolution kernel affects the range of local features that can be captured by the convolution operation. In this embodiment, the appropriate size of the convolution kernel can be selected according to the characteristics of the vibration signal and the analysis requirements. For example, if the vibration signal changes dramatically, a smaller convolution kernel may be selected to capture more detailed local features; if the vibration signal changes relatively smoothly, a larger convolution kernel can be selected to extract more extensive local features.

[0025] The step determines the sliding speed of the convolution operation in the time dimension. A larger step can reduce the amount of calculation, but may lose some detailed information; a smaller step can capture the changes in the signal more meticulously, but the amount of calculation will increase accordingly. Therefore, a balance needs to be struck between computational efficiency and feature extraction effectiveness to select an appropriate step.

[0026] After determining the parameters of the convolution layer, the multiple layers of time convolution operations are started. Starting from the input vibration sensing signal, the first convolution layer uses the preset convolution kernel to perform convolution operation on the signal to generate the first layer of feature maps. Each convolution kernel slides on the input signal, calculates the convolution result, and stores the result in the corresponding feature map position.

[0027] The output of the first convolution layer is used as the input of the second convolution layer, and the second convolution layer uses different convolution kernels to perform convolution operation on the feature maps of the first layer to generate the second layer of feature maps. In this way, through multiple layers of convolution operations, more complex features can be gradually extracted.

[0028] After each layer of convolution operation, an activation function can be introduced to increase the non-linear expression ability of the model. For example, using the ReLU function to process the convolution result, converting negative inputs to zero and keeping positive inputs unchanged. This way the model can learn more complex feature patterns.

[0029] After multiple layers of time convolution operations and activation function processing, the output of the last convolution layer is the vibration time series feature. These features contain both local fluctuation patterns and global trend patterns of the vibration signal. Local fluctuation patterns can be reflected by observing short-term changes in the feature map, such as sudden peaks or fluctuations. Global trend patterns can be determined by analyzing the change trend of the feature map over a longer period of time, such as gradual increase or decrease of feature values.

[0030] Step S122: input the temperature sensing signal into the temperature encoder of the feature encoding network, and the temperature encoder learns the time series correlation of the temperature signal values under the same operating condition through self-attention mechanism to generate temperature time series features containing steady-state temperature change patterns and dynamic temperature change patterns.

[0031] In this step, the temperature sensing signal is input into the temperature encoder of the feature encoding network. The temperature encoder uses self-attention mechanism to learn the time series correlation of the temperature signal values under the same operating condition to generate temperature time series features containing steady-state temperature change patterns and dynamic temperature change patterns.

[0032] Self-attention mechanism is a mechanism that can automatically focus on the relationship between different positions in the input sequence. In temperature signal processing, self-attention mechanism can help the model capture the correlation between temperature signals at different time points. For example, under a certain operating condition, the temperature signal may be affected by the temperature value at the previous time point, the ambient temperature, etc. Self-attention mechanism can determine the correlation between these time points by calculating the attention weights between different time points of temperature signal.

[0033] Under the same operating condition, the temperature signal has certain regularity. The steady-state temperature change pattern reflects the slow change of temperature under normal operating conditions, for example, when the wind turbine is stable, the temperature of the generator will remain in a relatively stable range, and will have a slow rise or fall over time. The dynamic temperature change pattern reflects the rapid change of temperature when the operating condition changes, for example, when the wind suddenly increases or the generator load changes, the temperature will rise or fall rapidly.

[0034] The temperature encoder learns the time series correlation of the temperature signal through self-attention mechanism, first converts the input temperature signal value into query (Query), key (Key) and value (Value) three vectors. Query vector is used to represent the temperature signal at the current time point, key vector is used to represent the temperature signal at other time points, and value vector contains the specific information of the temperature signal.

[0035] Then, the attention weight is obtained by calculating the similarity between the query vector and the key vector. The attention weight represents the degree of association between the temperature signal at the current time point and the temperature signals at other time points. According to the attention weight, the value vector is weighted and summed to obtain the attention output at the current time point.

[0036] By repeating this process multiple times, the entire temperature signal sequence is processed, and temperature time series features containing steady-state temperature change patterns and dynamic temperature change patterns can be generated. These features will help subsequent analysis of the change rule of the temperature signal and determine whether the device is abnormal.

[0037] In a specific example, before using the self-attention mechanism, the input temperature signal value needs to be converted into query, key, and value vectors. This can be achieved through linear transformation. Specifically, three different weight matrices can be used to linearly transform the input temperature signal to obtain the query vector, key vector, and value vector.

[0038] The dimensions of each vector can be set according to actual needs. Generally speaking, the dimensions of the vector will affect the computational complexity and feature expression ability of the self-attention mechanism. Higher dimensions can enable the model to learn more complex features, but the computational load will also increase accordingly; lower dimensions can reduce the computational load, but some information may be lost.

[0039] After obtaining the query, key, and value vectors, the attention weight is obtained by calculating the similarity between the query vector and the key vector. A common similarity calculation method is dot product operation, that is, the query vector and the key vector are dot multiplied to obtain a similarity score.

[0040] In order to make the similarity score comparable, it will also be scaled, usually divided by the square root of the dimension of the query vector. Then, the similarity score is converted into a probability distribution using the softmax function to obtain the attention weight. The attention weight represents the degree of association between the temperature signal at the current time point and the temperature signals at other time points.

[0041] According to the calculated attention weight, the value vector is weighted and summed. Specifically, the attention weight of each time point is multiplied by the corresponding value vector, and then all the results are aggregated to obtain the attention output at the current time point.

[0042] By repeating this process multiple times, the entire temperature signal sequence is processed, and the attention output at each time point can be obtained. These attention outputs will reflect the association between the temperature signals at different time points.

[0043] After processing by the self-attention mechanism, the attention output sequence obtained is the temperature time sequence feature, which contains the steady-state temperature change pattern and the dynamic temperature change pattern. The steady-state temperature change pattern can be reflected by observing the stable change of the attention output in a long time, and the dynamic temperature change pattern can be determined by analyzing the rapid change of the attention output in a short time.

[0044] Step S123: input the rotation speed sensing signal into the rotation speed encoder of the feature encoding network, and the rotation speed encoder models the synchronous change relationship between the rotation speed signal value and the corresponding wind turbine state by a recurrent neural network to generate rotation speed time sequence features containing state transition patterns and stable operation patterns.

[0045] In this step, the rotation speed sensing signal is input into the rotation speed encoder of the feature encoding network. The rotation speed encoder uses a recurrent neural network (RNN) to model the synchronous change relationship between the rotation speed signal value and the corresponding wind turbine state to generate rotation speed time sequence features containing state transition patterns and stable operation patterns.

[0046] The recurrent neural network is a neural network model specially used for processing sequence data, which can process the current input by remembering the previous input information. In the rotation speed signal processing, the recurrent neural network can capture the dependency relationship between different time points of the rotation speed signal and the synchronous change relationship between the rotation speed signal and the wind turbine state.

[0047] The state of the wind turbine can be divided into different modes, such as starting, accelerating, stable operation, decelerating and stopping, etc. In different states, the rotation speed signal will show different change rules. For example, in the starting stage, the rotation speed will gradually increase from zero; in the stable operation stage, the rotation speed will remain in a relatively stable range.

[0048] The rotation speed encoder processes the rotation speed signal by a recurrent neural network, first inputting the input rotation speed signal value into the neurons of the recurrent neural network in turn. Each neuron will update its own hidden state according to the current input and the previous hidden state, and output a result. This process will be repeated until the entire rotation speed signal sequence is processed.

[0049] During processing, the recurrent neural network learns the synchronous change relationship between the rotation speed signal and the wind turbine state. For example, when the wind turbine transitions from the starting state to the stable operation state, the rotation speed signal will have a corresponding change, and the recurrent neural network can capture this change pattern.

[0050] After processing by the recurrent neural network, the rotational speed time series features including state transition patterns and stable operation patterns can be generated. The state transition patterns reflect the changes in the rotational speed signal when the wind turbine transitions between different states, while the stable operation patterns represent the characteristics of the rotational speed signal in the stable operation state of the wind turbine.

[0051] In one specific example, when using a recurrent neural network for sequence modeling, a suitable network structure needs to be selected. Common recurrent neural network structures include Simple RNN, LSTM, and GRU, etc.

[0052] The Simple RNN structure is relatively simple, but it has the problem of gradient vanishing or gradient explosion, and its performance may not be good when processing long sequence data. LSTM and GRU can better handle long sequence data by introducing a gating mechanism.

[0053] In this embodiment, a suitable recurrent neural network structure can be selected according to the characteristics of the rotational speed signal and the analysis requirements. If the rotational speed signal changes are complex and long sequence data needs to be processed, LSTM or GRU can be selected; if the rotational speed signal changes are relatively simple, Simple RNN can be selected.

[0054] After determining the structure of the recurrent neural network, the parameters of the network, including the weight matrix and the bias vector, need to be initialized. The weight matrix is used to connect the neurons of the input layer, the hidden layer, and the output layer, and the bias vector is used to adjust the output of the neurons.

[0055] The way of initializing the parameters will affect the training effect of the recurrent neural network. Common initialization methods include random initialization and Xavier initialization. Random initialization randomly assigns values to the parameters, while Xavier initialization initializes the parameters according to the dimensions of the input and output to ensure that the input and output of the network have similar variances.

[0056] The input rotational speed signal values are sequentially input into the recurrent neural network, and at each time step, the hidden state is updated according to the current input and the previous hidden state, and a result is output. During training, the parameters of the network can be adjusted according to the error between the output result and the true label, so that the network can learn the synchronous change relationship between the rotational speed signal and the wind turbine state.

[0057] The backpropagation algorithm can be used to calculate the gradient of the error, and the parameters of the network can be updated according to the gradient. The training process will be iterated until the performance of the network reaches a satisfactory level.

[0058] After the training and processing of the recurrent neural network, the hidden state sequence at the last time step is the rotational speed time series feature. These features contain state transition patterns and stable operation patterns. The state transition pattern can be reflected by observing the changes in the hidden state at different time points, for example, when the wind turbine state changes, the hidden state will change significantly. The stable operation pattern can be determined by analyzing the characteristics of the hidden state in the stable time period.

[0059] Step S124: concatenating the vibration time series feature, the temperature time series feature, and the rotational speed time series feature along the time dimension to generate the state evolution feature sequence.

[0060] After obtaining the vibration time series feature, the temperature time series feature, and the rotational speed time series feature, these three feature sequences are concatenated along the time dimension to generate the state evolution feature sequence. The state evolution feature sequence contains the pattern representation of the change of each monitoring signal over time, which integrates the information of vibration, temperature, and rotational speed, and can more comprehensively reflect the operating state of the wind turbine system.

[0061] Concatenating along the time dimension means combining the three feature sequences on the time axis. For example, at a certain time t, the vibration time series feature, the temperature time series feature, and the rotational speed time series feature at that time are concatenated in order to form a new feature vector. By performing the above concatenation operation on the entire time sequence, the state evolution feature sequence is obtained.

[0062] The above concatenation method can preserve the time order and change pattern of each monitoring signal, so that subsequent analysis can consider the temporal correlation between different signals. For example, when analyzing the abnormal state of the equipment, the change of the state evolution feature sequence at different time points can be observed to determine which signal change caused the abnormality and whether there is a correlation between different signal changes.

[0063] Step S125: constructing a cross-source association module of the feature encoding network, and using the cross-source association module to calculate the association weight between the vibration time series feature and the temperature time series feature through a cross-attention mechanism to generate a first association feature reflecting the coupling relationship between mechanical vibration and thermal effect.

[0064] In order to further explore the association relationship between different monitoring signals, a cross-source association module of the feature encoding network is constructed. The cross-source association module uses a cross-attention mechanism to calculate the association weight between the vibration time series feature and the temperature time series feature to generate a first association feature reflecting the coupling relationship between mechanical vibration and thermal effect.

[0065] The cross-attention mechanism is similar to the self-attention mechanism, but it focuses on the relationship between different input sequences. In this embodiment, the cross-attention mechanism calculates the degree of association between the vibration time series features and the temperature time series features, and determines the importance of each feature in the other feature.

[0066] There is a close coupling relationship between mechanical vibration and thermal effect. For example, the vibration of a mechanical component can generate friction, leading to an increase in temperature; and an excessively high temperature can also affect the vibration characteristics of the mechanical component. Through the cross-attention mechanism, the strength of this coupling relationship can be quantified.

[0067] First, the vibration time series features and the temperature time series features are input into the cross-attention mechanism as query sequences and key-value sequences, respectively. The query sequence represents the target to be focused on, while the key-value sequence provides information that can be matched. In calculating the association weights, the cross-attention mechanism measures the similarity between each element in the vibration time series features and the elements in the temperature time series features through a series of operations.

[0068] The specific calculation process is as follows: first, linearly transform the query sequence and the key-value sequence to map them into the same feature space, ensuring that their dimensions match. Then, by calculating the dot product between each element in the query sequence and the elements in the key-value sequence, a similarity score matrix is obtained. To make the score matrix comparable, it can be scaled, usually by dividing by the square root of the query sequence dimension. Next, use the softmax function to convert the scaled score matrix into a probability distribution, which is the association weight. The association weight represents the degree of association between each element in the vibration time series features and the elements in the temperature time series features.

[0069] According to the calculated association weights, the temperature time series features are weighted and summed. Specifically, each temperature time series feature element is multiplied by the corresponding association weight, and then all the results are aggregated to obtain the first association feature, which reflects the association between the vibration time series features and the temperature time series features, that is, the coupling relationship between mechanical vibration and thermal effect. For example, if the vibration time series feature element at a certain time has a larger association weight with some elements in the temperature time series features, it means that the coupling relationship between mechanical vibration and thermal effect at that time is stronger.

[0070] Step S126: using the cross-source association module to calculate the association weights of the rotation speed time series features and the temperature time series features through the cross-attention mechanism, and generating a second association feature reflecting the coupling relationship between power transmission and heat dissipation.

[0071] Similarly, the cross-source association module continues to use the cross-attention mechanism to calculate the association weight between the rotational speed time series feature and the temperature time series feature to generate the second association feature reflecting the coupling relationship between power transmission and heat dissipation. During power transmission, the change in the rotational speed of the wind wheel will affect the working state of the generator, thereby producing different degrees of heat dissipation, and the heat dissipation may have a certain impact on the rotational speed, so there is a coupling relationship between the rotational speed time series feature and the temperature time series feature.

[0072] The rotational speed time series feature is input into the cross-attention mechanism as a query sequence, and the temperature time series feature is input as a key-value sequence. Similar to the calculation of the first association feature, the query sequence and the key-value sequence are first linearly transformed to map them to the same feature space, ensuring dimension matching. Then, the dot product between each element in the query sequence and the elements in the key-value sequence is calculated to obtain a similarity score matrix. In order to make the scores comparable, scaling processing is performed, usually dividing by the square root of the query sequence dimension. Then, the scaled score matrix is converted into a probability distribution using the softmax function to obtain the association weight.

[0073] The association weight represents the degree of association between each element in the rotational speed time series feature and the elements in the temperature time series feature. According to the association weight, the temperature time series feature is weighted and summed. Each temperature time series feature element is multiplied by the corresponding association weight, and then all the results are aggregated to obtain the second association feature reflecting the coupling relationship between power transmission and heat dissipation. For example, when the rotational speed of the wind wheel changes, if the association weight with the temperature time series feature is large, it indicates that the coupling relationship between the change in rotational speed during power transmission and heat dissipation is relatively close.

[0074] Step S127: Concatenate the first association feature and the second association feature along the feature dimension to generate the cross-source association feature set.

[0075] After obtaining the first association feature reflecting the coupling relationship between mechanical vibration and thermal effect and the second association feature reflecting the coupling relationship between power transmission and heat dissipation, the two association features are concatenated along the feature dimension to generate the cross-source association feature set. The cross-source association feature set contains the representation of the synchronous change relationship between different monitoring signals, which integrates the two different types of coupling relationships between mechanical vibration and thermal effect, and power transmission and heat dissipation.

[0076] Concatenating along the feature dimension means combining the first association feature and the second association feature in the dimension direction of the feature. For example, at a certain time, the first association feature vector and the second association feature vector at that time are concatenated in order to form a new feature vector. By performing the above concatenation operation on the entire time series, the cross-source association feature set is obtained.

[0077] This splicing method can preserve the information of different correlation features, so that subsequent analysis can consider different types of coupling relationships at the same time. For example, when analyzing the abnormal state of the device, it can be judged whether the abnormality is caused by the change of the coupling relationship between mechanical vibration and thermal effect, or the coupling relationship between power transmission and heat dissipation, or the result of their joint action, by observing the changes of the cross-source correlation feature set at different time points.

[0078] Step S130: calling a pre-trained anomaly diagnosis model to perform fusion analysis processing on the state evolution feature sequence and the cross-source correlation feature set, and generating an abnormal diagnosis result of the device state, the abnormal diagnosis result containing confidence distribution information of abnormal occurrence.

[0079] After obtaining the state evolution feature sequence and the cross-source correlation feature set, a pre-trained anomaly diagnosis model is called to perform fusion analysis processing. The anomaly diagnosis model is trained by a large amount of historical data, which can learn the feature patterns under normal and abnormal states.

[0080] The anomaly diagnosis model mainly consists of a state evolution analysis layer, a correlation analysis layer, a feature fusion layer and an abnormality discrimination layer. The state evolution analysis layer is used to process the state evolution feature sequence and mine the time sequence dependency relationship therein; the correlation analysis layer is used to analyze the cross-source correlation feature set and learn the collaborative relationship between multiple source signals; the feature fusion layer fuses the outputs of the state evolution analysis layer and the correlation analysis layer to obtain features containing time sequence information and correlation information; and the abnormality discrimination layer judges according to the fused features to generate an abnormal diagnosis result of the device state, which contains confidence distribution information of abnormal occurrence, i.e. the possibility of occurrence of different abnormal types.

[0081] Step S131: inputting the state evolution feature sequence into the state evolution analysis layer of the anomaly diagnosis model, and using the state evolution analysis layer to perform time sequence dependency learning on the state evolution feature sequence through a bidirectional long short-term memory network to generate time sequence dependency features containing historical state information and future trend information.

[0082] The state evolution feature sequence is input into the state evolution analysis layer of the anomaly diagnosis model. The state evolution analysis layer uses a bidirectional long short-term memory network (Bi-LSTM) to perform time sequence dependency learning on the state evolution feature sequence. The bidirectional long short-term memory network is a special recurrent neural network that can consider both forward and reverse information of the sequence, thereby better capturing the time sequence dependency relationship in the sequence.

[0083] First, the state evolution feature sequence is divided into continuous time window units. Each time window unit contains a preset length of time sequence feature subsequence. This is done to facilitate the processing of the bidirectional long short-term memory network. Then, each time window unit is processed.

[0084] For each time window unit, forward long short-term memory processing and backward long short-term memory processing are performed respectively. In forward long short-term memory processing, the time sequence feature subsequence is input into the neurons of the long short-term memory network in the forward order of time. Each neuron updates its own hidden state according to the current input and the previous hidden state, and outputs a result. This process is repeated until the entire forward time sequence feature subsequence is processed, and finally a forward hidden state containing historical cumulative information is obtained.

[0085] In backward long short-term memory processing, the time sequence feature subsequence is input into the neurons of the long short-term memory network in the reverse order of time. Similarly, each neuron updates its own hidden state according to the current input and the previous hidden state, and outputs a result. Through reverse processing, future trend information can be captured, and finally a backward hidden state containing future prediction information is obtained.

[0086] The forward hidden state and the backward hidden state are spliced step by step in time. At each time step, the forward hidden state vector and the backward hidden state vector of that time step are spliced together in order to form a new vector. By performing the above splicing operation on the entire time window unit, a joint hidden state containing bidirectional time sequence dependence is obtained.

[0087] Finally, the joint hidden state is processed by global average pooling in the time dimension. Global average pooling processing is to average the joint hidden state in the time dimension, sum the feature vectors of each time step, and then divide by the number of time steps to obtain a new feature vector. The new feature vector reflects the long-term dependence of the entire state evolution feature sequence, i.e., a time sequence dependence feature containing historical state information and future trend information.

[0088] In a specific example, before the state evolution feature sequence is input into the bidirectional long short-term memory network, it needs to be divided into continuous time window units. The purpose of division is to enable the bidirectional long short-term memory network to better process long sequence data and avoid the problem of gradient disappearance or gradient explosion.

[0089] A length of a time window is preset, and the state evolution feature sequence is divided according to the length. For example, a subsequence with a preset length is intercepted from the starting position of the state evolution feature sequence as a first time window unit, and then the subsequent subsequences are sequentially intercepted as other time window units by moving backward. If the length of the state evolution feature sequence cannot be divided by the length of the time window, the length of the last time window unit may be less than the preset length.

[0090] For each divided time window unit, forward long short-term memory processing is performed. In a forward order of time, the time sequence feature subsequences in the time window unit are sequentially input into neurons of the long short-term memory network.

[0091] The neurons of the long short-term memory network have three gating mechanisms inside, namely, an input gate, a forget gate, and an output gate. The input gate controls the entry of new input information, the forget gate determines which information in the hidden state at the last time needs to be forgotten, and the output gate controls how much information in the hidden state at the current time needs to be output.

[0092] At each time step, the neurons update their own hidden state and cell state through the gating mechanism according to the current input, the hidden state at the last time, and the cell state. Specifically, the forget gate calculates a forgetting factor according to the current input and the hidden state at the last time to determine how much information in the cell state at the last time needs to be forgotten; the input gate calculates an input factor to determine how much new input information needs to be added to the cell state; then the cell state is updated to combine the forgotten cell state at the last time and the newly added information. Finally, the output gate calculates an output factor according to the updated cell state and the current input to determine how much information in the hidden state at the current time needs to be output.

[0093] By continuously repeating the process, the forward hidden state containing historical accumulated information is finally obtained until the entire forward time sequence feature subsequence is processed.

[0094] Similarly, for each time window unit, backward long short-term memory processing is performed. Unlike the forward long short-term memory processing, here the time sequence feature subsequences in the time window unit are sequentially input into the neurons of the long short-term memory network in a reverse order of time.

[0095] The neurons of the backward long short-term memory network also have the three gating mechanisms of the input gate, the forget gate, and the output gate inside. At each time step, the neurons update their own hidden state and cell state through the gating mechanism according to the current input, the hidden state at the last time, and the cell state. The specific calculation process is similar to that of the forward long short-term memory processing, except that the input order is reversed.

[0096] By reverse processing, future trend information can be captured. Because in the reverse processing process, neurons can adjust the hidden state at the current time according to the subsequent input information, so as to obtain the backward hidden state containing future prediction information.

[0097] The forward hidden state and the backward hidden state are spliced time step by time step. At each time step, the forward hidden state vector and the backward hidden state vector of the time step are spliced in order to form a new vector.

[0098] For example, at a time t, the forward hidden state vector is A, and the backward hidden state vector is B. They are spliced to obtain a new vector [A, B]. By performing the above splicing operation on the entire time window unit, a joint hidden state containing bidirectional time sequence dependence is obtained.

[0099] The joint hidden state is subjected to global average pooling processing in the time dimension. The purpose of the global average pooling processing is to compress the joint hidden state in the time dimension and extract the feature information of the entire time window unit.

[0100] The specific calculation process is to sum the joint hidden state in the time dimension, and then divide by the number of time steps. For example, for a joint hidden state matrix containing multiple time steps, the feature vectors of each time step are aggregated to obtain a sum vector, and then the sum vector is divided by the number of time steps to obtain a new feature vector. The new feature vector is a time sequence dependence feature reflecting the long-term dependence relationship of the entire state evolution feature sequence.

[0101] Step S132: input the cross-source association feature set into the association analysis layer of the anomaly diagnosis model, and use the association analysis layer to learn the topological relationship of the feature nodes and the association edges in the cross-source association feature set through a graph neural network, to generate an association relationship feature containing a multi-source signal collaborative relationship.

[0102] The cross-source association feature set is input into the association analysis layer of the anomaly diagnosis model. The association analysis layer uses a graph neural network (GNN) to learn the topological relationship of the feature nodes and the association edges in the cross-source association feature set, to generate an association relationship feature containing a multi-source signal collaborative relationship.

[0103] The graph neural network is a neural network model specially designed for processing graph-structured data, which can learn the topology of data through the connection relationship between nodes. In the cross-source association feature set, the first association feature and the second association feature are respectively mapped as the node features of the graph neural network, where the first association feature corresponds to the coupling nodes of mechanical vibration and thermal effect, and the second association feature corresponds to the coupling nodes of power transmission and heat dissipation. Then, the edge features of the graph neural network are constructed according to the synchronous change relationship of the first association feature and the second association feature.

[0104] In a specific example, the first association feature and the second association feature in the cross-source association feature set are respectively mapped as the node features of the graph neural network. The mapping process is to convert the first association feature and the second association feature into a form suitable for processing by the graph neural network. Specifically, the first association feature vector and the second association feature vector are respectively taken as the feature vectors of different nodes in the graph neural network.

[0105] For example, the first association feature vector is taken as the feature representing the coupling nodes of mechanical vibration and thermal effect, and the second association feature vector is taken as the feature representing the coupling nodes of power transmission and heat dissipation. In this way, each node in the graph neural network has corresponding feature information, which reflects different types of coupling relationships.

[0106] The edge features of the graph neural network are constructed according to the synchronous change relationship of the first association feature and the second association feature. The edge features represent the connection relationship and the association degree between nodes.

[0107] In constructing the edge features, it is necessary to analyze the synchronous change of the first association feature and the second association feature at different time points. For example, if the first association feature and the second association feature change simultaneously at some time points, it indicates that the association degree between the two nodes is strong; if their changes have no obvious synchronicity, it indicates that the association degree is weak.

[0108] The value of the edge feature can be determined by calculating the similarity or correlation between the first association feature and the second association feature. For example, the correlation coefficient method is used to measure the association degree between them, and the calculated correlation coefficient is taken as the value of the edge feature. In this way, the edge feature can reflect the association relationship between nodes.

[0109] The node features and edge features are processed by the graph convolution operation to aggregate the neighborhood information, generating the local aggregation features of each node. The graph convolution operation is the core operation in the graph neural network, which can pass and aggregate the information of adjacent nodes to the current node through the connection relationship between nodes.

[0110] The specific calculation process is that, for each node in the graph, the features of the adjacent nodes are weighted and summed according to the features of the adjacent nodes and the edge features. The edge features serve as weights, determining the importance of the features of the adjacent nodes in the aggregation process. Through the above aggregation operation, each node can obtain the information in its neighborhood, thereby updating its own features.

[0111] For example, for a node with multiple adjacent nodes, the feature vector of each adjacent node is multiplied by the corresponding edge feature value, and then all the results are aggregated to obtain a new feature vector. The new feature vector is the local aggregation feature of the node, which contains the information of the node and its neighborhood.

[0112] The global attention feature of each node is generated by performing global importance weighting processing on the local aggregation feature through the graph attention mechanism. The graph attention mechanism can automatically learn the importance weights between nodes, enabling the model to pay more attention to important nodes and information.

[0113] In the graph attention mechanism, the attention weight between each node and other nodes is first calculated. The attention weight represents the degree of attention of one node to another node. The process of calculating the attention weight is to perform linear transformation on the local aggregation feature of the node, then measure the similarity between nodes through operations such as dot product, and finally use the softmax function to convert the similarity into a probability distribution, which is the attention weight.

[0114] According to the calculated attention weight, the local aggregation feature is weighted and summed. Specifically, the local aggregation feature of each node is multiplied by the corresponding attention weight, and then all the results are aggregated to obtain a new feature vector. The new feature vector is the global attention feature of the node, which considers the importance of the node in the global graph structure.

[0115] The global attention features are concatenated along the node dimension to generate the correlation relationship feature reflecting the collaborative relationship of multi-source signals. Concatenating along the node dimension means combining the global attention features of different nodes in the dimension direction of the node.

[0116] For example, the global attention feature vector representing the mechanical vibration and thermal effect coupling node and the global attention feature vector representing the power transmission and heat dissipation coupling node are sequentially concatenated to form a new feature vector. Through the above concatenation operation on the entire graph structure, the correlation relationship feature reflecting the collaborative relationship of multi-source signals is obtained.

[0117] This splicing method can preserve the information of different nodes, so that subsequent analysis can consider different types of coupling relationships at the same time. For example, when analyzing the abnormal state of the device, by observing the changes of the association relationship features at different time points, it can be judged whether the abnormality is caused by the change of the coupling relationship between mechanical vibration and thermal effect, or by the change of the coupling relationship between power transmission and heat dissipation, or by the joint action of the two.

[0118] Step S133: The feature fusion layer of the anomaly diagnosis model fuses the time-dependent features and the association relationship features through a gating mechanism to generate a fusion feature vector containing both time sequence information and association information.

[0119] The feature fusion layer of the anomaly diagnosis model fuses the time-dependent features and the association relationship features through a gating mechanism. The gating mechanism can flexibly control the weight and contribution of the time-dependent features and the association relationship features in the fusion process, so that the fused feature vector can contain both time sequence information and association information.

[0120] The gating mechanism mainly includes an input gate, a forget gate and an output gate. The input gate is used to determine how much information in the time-dependent features and the association relationship features needs to be input into the fusion process; the forget gate is used to determine how much information in the previous fusion state needs to be forgotten; and the output gate is used to determine how much information in the current fusion state needs to be output as the fusion feature vector.

[0121] The specific fusion process is that first, the input gate will calculate an input factor through linear transformation according to the time-dependent features and the association relationship features. The input factor represents how much information in the time-dependent features and the association relationship features needs to be input into the fusion process in the current fusion step. For example, if the part corresponding to the time-dependent features in the input factor has a larger value, it means that the information of the time-dependent features has a greater impact on the fusion result in this step.

[0122] At the same time, the forget gate will also calculate a forgetting factor through linear transformation according to the time-dependent features and the association relationship features. The forgetting factor determines how much information in the previous fusion state needs to be forgotten. In each fusion process, the previous fusion state may contain some irrelevant or outdated information, and the forget gate can help the model to remove these information, so that the fusion result is more accurate.

[0123] Next, the time-dependent features and the association features are combined according to the input factor. Specifically, the time-dependent feature vector is multiplied by the part of the input factor corresponding to the time-dependent feature, and the association feature vector is multiplied by the part of the input factor corresponding to the association feature. Then, the two results are aggregated to obtain a new feature vector. The new feature vector contains the information needed to be input into the fusion process in the current step.

[0124] Then, the previous fusion state is updated according to the forgetting factor. The previous fusion state vector is multiplied by the forgetting factor to obtain an updated fusion state vector. The updated fusion state vector removes some information that is no longer relevant.

[0125] The newly obtained feature vector and the updated fusion state vector are merged to obtain the current fusion state. Finally, the output gate calculates an output factor through linear transformation according to the current fusion state. The output factor determines how much information in the current fusion state needs to be output as the fusion feature vector. The current fusion state vector is multiplied by the output factor to obtain the final fusion feature vector. The fusion feature vector contains both the time sequence information carried by the time-dependent feature and the association information contained in the association feature.

[0126] Step S134: input the fusion feature vector into the anomaly discrimination layer of the anomaly diagnosis model, and use the anomaly discrimination layer to perform non-linear transformation on the fusion feature vector through a fully connected neural network to generate an anomaly diagnosis result containing confidence of different abnormal types.

[0127] The fusion feature vector is input into the anomaly discrimination layer of the anomaly diagnosis model. The anomaly discrimination layer uses a fully connected neural network to perform non-linear transformation on the fusion feature vector to generate an anomaly diagnosis result containing confidence of different abnormal types.

[0128] The fully connected neural network is composed of multiple neuron layers, and each layer of neurons is connected to all neurons of the next layer. In the anomaly discrimination layer, the fusion feature vector is first input into the first layer of neurons. The first layer of neurons will perform linear transformation on the input fusion feature vector, i.e., each neuron will multiply each element of the fusion feature vector by the corresponding weight, then aggregate all results and add a bias term. Then, the linear transformation result is non-linearly transformed through an activation function. The activation function can increase the non-linear expression ability of the model, so that the model can learn more complex feature patterns. Common activation functions include ReLU function, etc.

[0129] After the processing of the first layer of neurons, the resulting output is passed as input to the second layer of neurons. The second layer of neurons also performs linear and non-linear transformations. This process is repeated for each layer of the fully connected neural network until the last layer of neurons.

[0130] The output of the last layer of neurons is the anomaly diagnosis result. In this layer, the number of neurons is usually the same as the number of abnormal types. The output value of each neuron represents the confidence of the occurrence of the corresponding abnormal type. For example, if there are three abnormal types, then the last layer has three neurons, and the output value of each neuron represents the likelihood of the occurrence of the three abnormal types. In order to make these output values have the meaning of probability, the output of the last layer of neurons is usually processed using the softmax function. The softmax function converts the output value of each neuron into a probability value, so that the sum of all probability values is 1. In this way, an abnormal diagnosis result containing the confidence of different abnormal types is obtained, which can help determine the current abnormal state of the device and the likelihood of the occurrence of the abnormality.

[0131] Step S140: Based on the anomaly diagnosis result, the abnormal type of the wind turbine system and the development trajectory of the abnormality in the time dimension and the affected component identification in the spatial dimension are analyzed, wherein the development trajectory contains a time sequence description of the abnormality from the initial appearance to the stable appearance, and the affected component identification contains the specific component label directly affected by the abnormality.

[0132] After obtaining the abnormal diagnosis result, it needs to be analyzed to determine the abnormal type of the wind turbine system, the development trajectory of the abnormality in the time dimension, and the affected component identification in the spatial dimension.

[0133] Step S141: Extract the abnormal type label with a confidence greater than a set confidence from the abnormal diagnosis result as the abnormal type.

[0134] First, the abnormal type is extracted from the abnormal diagnosis result. The abnormal diagnosis result contains the confidence of different abnormal types. By setting a confidence threshold, the abnormal type label with a confidence greater than the threshold is determined as the final abnormal type. For example, assuming that the abnormal diagnosis result contains the confidence of multiple abnormal types, such as the confidence of abnormal type A is 0.8, the confidence of abnormal type B is 0.3, and the set confidence threshold is 0.5, then abnormal type A is determined as the abnormal type. The set confidence threshold can be adjusted according to the actual application scenario and the accuracy requirement of abnormal judgment.

[0135] Step S142: Extract the original monitoring signal segment corresponding to the abnormal occurrence timestamp from the state evolution feature sequence, and determine the starting time point of the initial appearance of the abnormality through timestamp backtracking.

[0136] In order to determine the development trajectory of the anomaly in the time dimension, it is necessary to find the starting time point of the initial appearance of the anomaly. From the state evolution feature sequence, according to the time stamp in the anomaly diagnosis result showing the occurrence of the anomaly, the corresponding original monitoring signal segment is extracted. These original monitoring signal segments contain the specific values of the vibration sensing signal, the temperature sensing signal and the rotating speed sensing signal near the time point of the occurrence of the anomaly.

[0137] Through the method of time stamp backtracking, the original monitoring signal segments are analyzed. The change trend of the signal in time is observed, and the earliest time point at which the signal starts to appear abnormal change is found, which is taken as the starting time point of the initial appearance of the anomaly. For example, in the vibration sensing signal, if it is found that the fluctuation amplitude of the signal suddenly increases after a certain time point, and such increasing trend continues to exist, then the time point may be the starting time point of the initial appearance of the anomaly.

[0138] Step S143: Extracting the original monitoring signal segment corresponding to the anomaly stable appearance time stamp from the state evolution feature sequence, and determining the end time point of the stable existence of the anomaly through time stamp backtracking.

[0139] Similarly, from the state evolution feature sequence, the original monitoring signal segment corresponding to the anomaly stable appearance time stamp is extracted. The anomaly stable appearance time stamp can be determined according to the time period with higher confidence and lasting for a period of time in the anomaly diagnosis result. The original monitoring signal segments are analyzed through time stamp backtracking, the change of the signal in time is observed, and the latest time point at which the signal returns to the normal state or the anomaly feature is no longer obvious is found, which is taken as the end time point of the stable existence of the anomaly. For example, in the temperature sensing signal, if it is found that the temperature signal returns to the normal fluctuation range after a certain time point, then the time point can be taken as the end time point of the stable existence of the anomaly.

[0140] Step S144: Taking the time sequence between the starting time point and the end time point as the time range of the development trajectory.

[0141] The time sequence between the determined starting time point of the initial appearance of the anomaly and the end time point of the stable existence of the anomaly is taken as the time range of the development trajectory of the anomaly in the time dimension. The time range describes the entire process from the initial appearance to the stable appearance and then possibly to the end of the anomaly. For example, the starting time point is t1, and the end time point is t2, then the time range is the time period from t1 to t2, and in the time period, the development trend and change of the anomaly can be further analyzed.

[0142] Step S145: Extract the correlation features within the abnormal occurrence time range from the cross-source correlation feature set, and determine the monitoring signal type with the highest correlation degree to the abnormality through feature importance analysis.

[0143] In order to determine the influence of the abnormality on the spatial dimension of the component identification, it is necessary to find the monitoring signal type with the highest correlation degree to the abnormality. From the cross-source correlation feature set, extract the correlation features within the abnormal occurrence time range. These correlation features contain information about different types of coupling relationships between mechanical vibration and thermal effects, power transmission and heat dissipation during the occurrence of the abnormality.

[0144] Step S1451: Extract the correlation feature subsequence in the cross-source correlation feature set whose time range is aligned with the development trajectory time range.

[0145] According to the development trajectory time range of the abnormality in the time dimension, extract the correlation feature subsequence in the cross-source correlation feature set whose time range is aligned with the development trajectory time range. This can ensure that the extracted correlation features are within the time period of the abnormality and can accurately reflect the correlation between different monitoring signals during the occurrence of the abnormality. For example, if the development trajectory time range is from t1 to t2, then extract the correlation features within the time period from t1 to t2 from the cross-source correlation feature set to form a correlation feature subsequence.

[0146] Step S1452: Call the feature importance evaluation module of the abnormality diagnosis model, and use the feature importance evaluation module to calculate the gradient value of the correlation feature subsequence to the abnormality diagnosis result through gradient backpropagation.

[0147] Call the feature importance evaluation module of the abnormality diagnosis model, which calculates the gradient value of the correlation feature subsequence to the abnormality diagnosis result through gradient backpropagation. Gradient backpropagation is a method for calculating the derivative of a function, which can be used in neural networks to determine the degree of influence of input features on output results.

[0148] Specifically, input the correlation feature subsequence into the abnormality diagnosis model, and calculate the gradient value of each feature element in the correlation feature subsequence to the abnormality diagnosis result through the gradient backpropagation algorithm based on the error between the abnormality diagnosis result and the true label. The gradient value represents how much a small change in the feature element will affect the abnormality diagnosis result. If the gradient value of a certain feature element is large, it means that the feature element has a large influence on the abnormality diagnosis result and has a high correlation degree to the abnormality.

[0149] Step S1453: Take the gradient value as the feature importance score, and select the correlation feature with the highest feature importance score as the key correlation feature.

[0150] The calculated gradient value is taken as a feature importance score. The higher the feature importance score, the higher the correlation between the feature and the anomaly. The feature importance scores of all feature elements in the correlation feature subsequence are compared, and the correlation feature with the highest score is selected as the key correlation feature. The key correlation feature can most directly reflect the correlation between different monitoring signals when the anomaly occurs.

[0151] Step S1454: determining the corresponding monitoring signal type according to the source identification of the key correlation feature, wherein the first correlation feature corresponds to the combination of the vibration sensing signal and the temperature sensing signal, and the second correlation feature corresponds to the combination of the rotating speed sensing signal and the temperature sensing signal.

[0152] The corresponding monitoring signal type is determined according to the source identification of the key correlation feature. In the foregoing processing, the first correlation feature reflects the coupling relationship between mechanical vibration and thermal effect, corresponding to the combination of the vibration sensing signal and the temperature sensing signal; the second correlation feature reflects the coupling relationship between power transmission and heat dissipation, corresponding to the combination of the rotating speed sensing signal and the temperature sensing signal.

[0153] It is checked whether the key correlation feature belongs to the first correlation feature or the second correlation feature. If the key correlation feature belongs to the first correlation feature, the corresponding monitoring signal type is the vibration sensing signal and the temperature sensing signal; if the key correlation feature belongs to the second correlation feature, the corresponding monitoring signal type is the rotating speed sensing signal and the temperature sensing signal.

[0154] Step S1455: taking the monitoring signal type corresponding to the key correlation feature as the monitoring signal type with the highest correlation degree to the anomaly.

[0155] The monitoring signal type corresponding to the key correlation feature determined through the foregoing steps is taken as the monitoring signal type with the highest correlation degree to the anomaly. This monitoring signal type can indicate which monitoring signals are most related to the occurrence of the anomaly when the anomaly occurs.

[0156] Step S146: determining the specific component directly affected by the anomaly as the impact component identification according to the physical sensor position information corresponding to the monitoring signal type.

[0157] After determining the monitoring signal type with the highest correlation degree to the anomaly, the specific component directly affected by the anomaly is determined according to the physical sensor position information corresponding to the monitoring signal type. Different types of monitoring signals are collected by physical sensors installed at different positions, and these sensors are usually installed near the key components of the wind turbine system.

[0158] For example, if the monitoring signal type with the highest correlation degree to the anomaly is a vibration sensor signal and a temperature sensor signal, and the vibration sensor is installed on the gearbox and the temperature sensor is also installed near the gearbox, it can be inferred that the gearbox is likely to be the specific component directly affected by the anomaly, and the label of the gearbox is identified as the affected component. In this way, the affected component in the spatial dimension of the anomaly can be determined.

[0159] Step S150: generating a device warning instruction containing time and space positioning information according to the anomaly type, the development trajectory, and the affected component identification, and sending the device warning instruction to the target management terminal to trigger a maintenance response operation.

[0160] After determining the anomaly type, the development trajectory of the anomaly in the time dimension, and the affected component identification in the spatial dimension, a device warning instruction containing time and space positioning information needs to be generated based on these information, and the instruction is sent to the target management terminal to trigger a maintenance response operation.

[0161] Step S151: obtaining a preset mapping relationship table of anomaly type and warning level, and searching for the corresponding warning level identification according to the anomaly type.

[0162] First, a mapping relationship table of preset anomaly type and warning level is obtained. The mapping relationship table is preset in the system design stage, which specifies the corresponding warning level of different anomaly types. According to the determined anomaly type, the corresponding warning level identification is searched in the mapping relationship table. For example, if the anomaly type is a certain serious fault that seriously affects the operation of the device, the corresponding warning level may be high-level warning; if the anomaly type is some minor anomaly, the corresponding warning level may be low-level warning. The warning level identification can be represented by different symbols or codes.

[0163] Step S152: extracting the start time point and the end time point of the anomaly from the time range of the development trajectory, and generating time positioning information containing time interval description.

[0164] The start time point and the end time point of the anomaly are extracted from the time range of the development trajectory in the time dimension. The two time points are combined to generate time positioning information containing time interval description. For example, if the start time point is t1 and the end time point is t2, the time positioning information can be represented as the time period from t1 to t2. The time positioning information can clearly inform the maintenance personnel of the time range of the anomaly, providing a time reference for formulating a maintenance plan.

[0165] Step S153: searching for a preset component location information table according to the affected component identification, and obtaining the spatial coordinates of the affected component as spatial positioning information.

[0166] According to the identified influence component identifier, a preset component position information table is searched. The component position information table records the spatial coordinate information of each component in the wind turbine system. The spatial coordinate information of the component corresponding to the influence component identifier is searched in the table and is taken as the spatial positioning information. The spatial positioning information can enable the maintenance personnel to accurately know the specific position of the component affected by the anomaly in the wind turbine system, thereby facilitating rapid arrival at the fault site for maintenance.

[0167] Step S154: constructing an information structure of the device warning instruction, the information structure including the warning level identifier, the time positioning information, and the spatial positioning information.

[0168] The information structure of the device warning instruction is constructed, and the warning level identifier, the time positioning information, and the spatial positioning information are combined together. The information structure can adopt a specific format to organize these information. For example, the warning level identifier can be placed in the front, and then the time positioning information and the spatial positioning information are sequentially arranged. The above information structure can clearly convey the key information such as the severity of the anomaly, the occurrence time, and the position of the affected component.

[0169] Step S155: converting the information structure into a communication protocol format recognizable by the target management terminal to generate the device warning instruction.

[0170] The constructed information structure is converted into a communication protocol format recognizable by the target management terminal. Different target management terminals can support different communication protocols, and thus the information structure needs to be converted accordingly. For example, if the target management terminal supports a certain specific network communication protocol, the information structure needs to be encoded according to the format of the protocol, including adding a protocol header, a check code, and other information. After the conversion, the final device warning instruction is generated.

[0171] Finally, the device warning instruction is sent to the target management terminal. After receiving the device warning instruction, the target management terminal can trigger a corresponding maintenance response operation according to the warning level identifier and other information. For example, if it is a high-level warning, the maintenance personnel can be immediately notified to go to the site for inspection and repair; if it is a low-level warning, the maintenance can be arranged at a suitable time. In this way, the entire process of wind turbine system anomaly state detection and warning is completed.

[0172] Further, the method can further include: Step S210: acquiring historical normal operation data and historical anomaly operation data of the wind turbine system, the historical normal operation data and the historical anomaly operation data being in a format consistent with the multi-source monitoring data stream.

[0173] To train the feature encoding network and the anomaly diagnosis model, historical normal operation data and historical abnormal operation data of the wind turbine system need to be obtained. The format of these data is consistent with the previously obtained multi-source monitoring data stream, and they all contain vibration sensor signals, temperature sensor signals, and rotation speed sensor signals, and have time-stamped continuous monitoring records.

[0174] The historical normal operation data is collected during the normal operation of the wind turbine system, which reflects the operating characteristics of the equipment in the normal state. For example, during normal operation, the fluctuation amplitude of the vibration sensor signal is within a certain range, the temperature sensor signal remains relatively stable, and the rotation speed sensor signal also conforms to the normal operation rule.

[0175] The historical abnormal operation data is collected when the wind turbine system has an abnormal situation, which contains monitoring signal data when various types of abnormalities occur. For example, when the gearbox fails, the vibration sensor signal will have abnormal fluctuations, the temperature sensor signal may rise, and the rotation speed sensor signal may also change.

[0176] By collecting a large amount of historical normal operation data and historical abnormal operation data, rich samples can be provided for model training, allowing the model to learn the feature patterns in normal and abnormal states.

[0177] Step S220: Perform time-stamped alignment processing on the historical normal operation data and the historical abnormal operation data to generate a training sample set.

[0178] After obtaining the historical normal operation data and the historical abnormal operation data, time-stamped alignment processing needs to be performed on these data. Since the sampling frequency and data transmission time of different sensors may differ, the timestamps of these data need to be calibrated to ensure consistency of different types of signals in the time dimension.

[0179] Time-stamped alignment processing can be achieved through hardware synchronization or software algorithms. For example, a high-precision clock source can be used to synchronize the sampling time of each sensor, or the timestamp can be adjusted through an algorithm after data collection. After time-stamped alignment processing, the historical normal operation data and the historical abnormal operation data are combined to generate a training sample set. The training sample set contains a large amount of multi-source monitoring data with time-stamped alignment, which will be used for training of the feature encoding network and the anomaly diagnosis model.

[0180] Step S230: Initialize the network parameters of the feature encoding network and the anomaly diagnosis model.

[0181] Before starting the training, the network parameters of the feature encoding network and the anomaly diagnosis model need to be initialized. The network parameters include weight matrices and bias vectors, etc., which determine the calculation method and feature learning ability of the neural network.

[0182] There are many methods to initialize network parameters, and the common ones are random initialization and Xavier initialization, etc. Random initialization is to randomly assign network parameters, so that each parameter takes a random value within a small range. Xavier initialization is to initialize parameters according to the dimensions of input and output, so as to ensure that the input and output of the network have similar variances, which can make the model more stable during training.

[0183] By initializing the network parameters, an initial state is provided for the feature encoding network and the anomaly diagnosis model, so that the model can start learning the data features in the training sample set. For the feature encoding network, it includes vibration encoder, temperature encoder, speed encoder, cross-source correlation module and other parts, each of which has corresponding weight matrices and bias vectors to be initialized. For example, the convolution kernel weights and biases of each layer in the multi-layer time convolution layer of the vibration encoder need to be initialized; the linear transformation matrix and bias for converting the input into query, key and value vectors in the self-attention mechanism of the temperature encoder also need to be initialized; the weights and biases of the recurrent neural network structure (such as LSTM or GRU) of the speed encoder also need to be initialized.

[0184] The anomaly diagnosis model is no exception. The bidirectional long short-term memory network of the state evolution analysis layer, the graph neural network of the correlation analysis layer, the gating mechanism of the feature fusion layer, and the fully connected neural network of the anomaly discrimination layer all have their weight matrices and bias vectors initialized with initial values in the initialization step. The choice of these initial values has an important influence on the training speed and final performance of the model. For example, if the initial value is too large or too small, it may cause the problem of gradient vanishing or gradient explosion during the training of the model, making the training unable to proceed normally. A suitable initialization method, such as Xavier initialization, can alleviate these problems to some extent, allowing the model to learn data features more stably in the early stage of training.

[0185] Step S240: inputting the multi-source monitoring data in the training sample set into the feature encoding network to generate state evolution feature sequences and cross-source correlation feature sets for training.

[0186] After the network parameters of the feature encoding network and the anomaly diagnosis model are initialized, the multi-source monitoring data in the training sample set is input into the feature encoding network to generate state evolution feature sequences and cross-source correlation feature sets for training. The specific execution process can be referred to the description of the related embodiments of the foregoing step S120.

[0187] Step S250: inputting the state evolution feature sequence for training and the cross-source correlation feature set into the anomaly diagnosis model to generate an anomaly diagnosis result for training.

[0188] After obtaining the state evolution feature sequence for training and the cross-source correlation feature set, they are input into the anomaly diagnosis model. The specific implementation process can refer to the description of the related embodiments of step S130.

[0189] Step S260: calculating a loss value between the anomaly diagnosis result for training and a sample true label.

[0190] After obtaining the anomaly diagnosis result for training, the loss value between the result and the sample true label needs to be calculated. The sample true label is pre-labeled in the training sample set, which explicitly indicates the true abnormal type corresponding to each sample. The loss value is used to measure the difference between the output of the anomaly diagnosis model and the true situation. By continuously reducing the loss value, the model can gradually learn more accurate feature patterns and improve the accuracy of anomaly diagnosis.

[0191] Common loss functions include cross-entropy loss function, etc. For multi-classification problems (such as abnormal type discrimination in this embodiment), the cross-entropy loss function can well measure the difference between the probability distribution of the model output and the true label. The specific calculation process is that, for each sample in the training sample set, according to its true label and the confidence of different abnormal types output by the anomaly diagnosis model, the calculation rule of the cross-entropy loss function is calculated.

[0192] Suppose the true label of a sample is a certain abnormal type, and the corresponding true probability distribution is 1 at the abnormal type and 0 at other abnormal types; while the anomaly diagnosis model outputs the confidence of each abnormal type (probability value after softmax function processing). The cross-entropy loss function compares the true probability distribution and the probability distribution output by the model, and calculates the difference between them. The loss values calculated for each sample are summed up to obtain the loss value of the entire training sample set. The loss value reflects the overall performance of the model in the current training state. The smaller the loss value, the closer the output of the model to the true label, and the better the performance of the model.

[0193] Step S270: updating the network parameters of the feature encoding network and the anomaly diagnosis model through a back propagation algorithm until the loss value converges to a preset threshold.

[0194] After calculating the loss value of the training anomaly diagnosis result and the sample true label, the back propagation algorithm is used to update the network parameters of the feature encoding network and the anomaly diagnosis model. The back propagation algorithm is an optimization algorithm based on the gradient descent principle, which can adjust the network parameters according to the gradient information of the loss value, so that the loss value gradually decreases.

[0195] First, starting from the last layer of the anomaly diagnosis model (anomaly discrimination layer), the gradient of the output of this layer neurons to the loss value is calculated according to the loss value. This gradient represents the rate of change of the loss value with respect to the output of this layer. Then, according to the gradient information, the gradients of the weight matrix and the bias vector of this layer are calculated. The gradients of the weight matrix and the bias vector represent the rate of change of the loss value with respect to them, and through these gradients, it can be known how to adjust the values of the weight matrix and the bias vector to make the loss value decrease.

[0196] Then, the gradient information is propagated back to the previous layer (feature fusion layer). In the feature fusion layer, according to the received gradient information, the gradients of each parameter (such as the weights and biases of the input gate, the forgetting gate, and the output gate) in the gating mechanism of this layer are calculated. Similarly, these gradient information is used to guide the adjustment of the parameters of this layer. In the above manner, the gradient information is constantly propagated back to the state evolution analysis layer and the correlation analysis layer, and the gradients of the weight matrix and the bias vector of each module (such as the bidirectional long short-term memory network and the graph neural network) in these two layers are calculated.

[0197] For the feature encoding network, the same back propagation method is also used. Starting from the cross-source correlation module, according to the gradient information feedback from the anomaly diagnosis model, the parameter gradients of the cross-attention mechanism in this module are calculated; then the gradient information is propagated to the vibration encoder, the temperature encoder and the speed encoder, and the gradients of their respective parameters (such as convolution layer parameters, self-attention mechanism parameters, recurrent neural network parameters, etc.) are calculated.

[0198] According to the calculated gradient information, the optimization algorithm (such as stochastic gradient descent, Adam optimizer, etc.) is used to update the network parameters of the feature encoding network and the anomaly diagnosis model. The optimization algorithm will adjust the values of the parameters according to the size and direction of the gradient, according to the set learning rate. The learning rate controls the step size of parameter update. If the learning rate is too large, the model may skip the optimal solution, resulting in failure to converge; if the learning rate is too small, the training speed of the model will be very slow.

[0199] After each update of the network parameters, the multi-source monitoring data in the training sample set is input into the feature encoding network and the anomaly diagnosis model again, and steps S240-S260 are repeated to calculate a new loss value. This process is repeated until the loss value converges to a preset threshold. The preset threshold is a value set in advance before the training starts. When the loss value is less than the threshold, it means that the model has learned enough feature patterns, and the training can be stopped. At this time, the parameters of the feature encoding network and the anomaly diagnosis model have been adjusted to a relatively optimal state, and the abnormal state of the wind turbine system can be detected and diagnosed more accurately.

[0200] Figure 2 An exemplary hardware and software components of the abnormal state detection system 100 for a wind turbine system according to some embodiments of the present application are shown in the schematic diagram. For example, the processor 120 can be used in the abnormal state detection system 100 for a wind turbine system and used to perform the functions in the present application.

[0201] The abnormal state detection system 100 for a wind turbine system can be a general-purpose server or a special-purpose server, both of which can be used to implement the abnormal state detection method for a wind turbine system according to the present application. Although only one server is shown in the present application, for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0202] For example, the abnormal state detection system 100 for a wind turbine system can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. For example, the abnormal state detection system 100 for a wind turbine system can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The abnormal state detection system 100 for a wind turbine system also includes an I / O interface 150 between the computer and other input / output devices.

[0203] For the convenience of description, only one processor is described in the abnormal state detection system 100 for the wind power generator system. However, it should be noted that the abnormal state detection system 100 for the wind power generator system in the present application can also include multiple processors, and thus the steps performed by one processor described in the present application can also be performed jointly by multiple processors or individually. For example, if the processor of the abnormal state detection system 100 for the wind power generator system performs steps A and B, it should be understood that steps A and B can also be performed jointly by two different processors or individually in one processor. For example, a first processor performs step A, a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0204] In addition, the embodiment of the present application further provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the abnormal state detection method for the wind power generator system is realized.

[0205] It should be noted that, in order to simplify the description of the present application and to help the understanding of one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for detecting abnormal conditions of a wind turbine system, characterized in that: The method comprises: Acquire a multi-source monitoring data stream of a wind turbine system, wherein the multi-source monitoring data stream includes three types of continuous monitoring records with aligned timestamps: vibration sensor signals, temperature sensor signals, and speed sensor signals; Performing joint feature extraction processing on the multi-source monitoring data stream through a feature encoding network to generate a state evolution feature sequence and a cross-source correlation feature set reflecting the operating state of the equipment, wherein the state evolution feature sequence includes a pattern representation of each monitoring signal changing over time, and the cross-source correlation feature set includes a representation of the synchronous change relationship between different monitoring signals; Calling a pre-trained anomaly diagnosis model to perform fusion analysis on the state evolution feature sequence and the cross-source correlation feature set to generate an anomaly diagnosis result of the device state, wherein the anomaly diagnosis result includes confidence distribution information of the anomaly occurrence; Analyze the abnormality type of the wind turbine system and the development trajectory of the abnormality in the time dimension and the identification of affected components in the spatial dimension based on the abnormality diagnosis result, wherein the development trajectory includes a time series description from the initial occurrence of the abnormality to the stable manifestation, and the affected component identification includes the labels of specific components directly affected by the abnormality; An equipment warning instruction including time and space positioning information is generated according to the abnormality type, the development trajectory, and the affected component identifier, and the equipment warning instruction is sent to a target management terminal to trigger a maintenance response operation.

2. The method according to claim 1, characterized in that The joint feature extraction processing of the multi-source monitoring data stream by the feature coding network to generate a state evolution feature sequence and a cross-source correlation feature set reflecting the operation state of the equipment includes: Inputting the vibration sensor signal into the vibration encoder of the feature encoding network, the vibration encoder performs time series pattern learning on the vibration signal values ​​of continuous time stamps through multi-layer time convolution operations to generate vibration time series features including local fluctuation patterns and global trend patterns; Inputting the temperature sensor signal into the temperature encoder of the feature encoding network, the temperature encoder performs time series association learning on the temperature signal values ​​under the same operating conditions through a self-attention mechanism to generate temperature time series features including steady-state temperature change patterns and dynamic temperature change patterns; The speed sensor signal is input into the speed encoder of the feature encoding network. The speed encoder performs sequence modeling on the synchronous change relationship between the speed signal value and the corresponding wind rotor state through a recurrent neural network to generate a speed time series feature including a state transition mode and a stable operation mode; splicing the vibration time series features, the temperature time series features, and the speed time series features along the time dimension to generate the state evolution feature sequence; Constructing a cross-source association module of the feature encoding network, and using the cross-source association module to calculate the association weights of the vibration time series features and the temperature time series features through a cross-attention mechanism to generate a first association feature reflecting the coupling relationship between mechanical vibration and thermal effect; Utilizing the cross-source association module to calculate the association weights of the speed time series features and the temperature time series features through a cross-attention mechanism, and generating a second association feature reflecting the coupling relationship between power transmission and heat dissipation; The first correlation feature and the second correlation feature are spliced ​​along the feature dimension to generate the cross-source correlation feature set.

3. The method according to claim 1, characterized in that The calling of the pre-trained abnormality diagnosis model performs fusion analysis on the state evolution feature sequence and the cross-source correlation feature set to generate an abnormality diagnosis result of the device state, including: Inputting the state evolution feature sequence into the state evolution analysis layer of the abnormality diagnosis model, using the state evolution analysis layer to perform temporal dependency learning on the state evolution feature sequence through a bidirectional long short-term memory network to generate a temporal dependency feature containing historical state information and future trend information; Inputting the cross-source correlation feature set into the correlation analysis layer of the abnormality diagnosis model, using the correlation analysis layer to perform topological relationship learning on the feature nodes and correlation edges in the cross-source correlation feature set through a graph neural network to generate correlation relationship features containing multi-source signal collaborative relationships; Utilizing the feature fusion layer of the abnormality diagnosis model to fuse the temporal dependency feature and the association relationship feature through a gating mechanism to generate a fused feature vector containing both temporal information and association information; The fused feature vector is input into the abnormality discrimination layer of the abnormality diagnosis model, and the abnormality discrimination layer is used to perform nonlinear transformation on the fused feature vector through a fully connected neural network to generate an abnormality diagnosis result containing confidence levels of different abnormality types.

4. The method according to claim 3, characterized in that The state evolution feature sequence is input into the state evolution analysis layer of the abnormality diagnosis model, and the state evolution analysis layer is used to perform temporal dependency learning on the state evolution feature sequence through a bidirectional long short-term memory network to generate a temporal dependency feature containing historical state information and future trend information, including: Dividing the state evolution feature sequence into continuous time window units, each time window unit contains a time series feature subsequence of a preset length; Perform forward long-short-term memory processing on the forward time sequence feature subsequence of each time window unit to generate a forward hidden state containing historical accumulated information; Perform backward long short-term memory processing on the reverse time order feature subsequence of each time window unit to generate a backward hidden state containing future prediction information; Concatenating the forward hidden state and the backward hidden state time-step by time step to generate a joint hidden state containing bidirectional temporal dependency; A global average pooling process is performed on the joint hidden state in the time dimension to generate a temporal dependency feature that reflects the long-term dependency relationship of the entire state evolution feature sequence.

5. The method according to claim 3, characterized in that The step of inputting the cross-source correlation feature set into the correlation analysis layer of the abnormality diagnosis model, utilizing the correlation analysis layer to perform topological relationship learning on feature nodes and correlation edges in the cross-source correlation feature set through a graph neural network, and generating correlation relationship features containing multi-source signal collaborative relationships, includes: Mapping the first correlation feature and the second correlation feature in the cross-source correlation feature set into node features of a graph neural network, respectively, wherein the first correlation feature corresponds to a coupling node between mechanical vibration and thermal effect, and the second correlation feature corresponds to a coupling node between power transmission and heat dissipation; Constructing an edge feature of a graph neural network based on the synchronous change relationship between the first correlation feature and the second correlation feature; Performing neighborhood information aggregation processing on the node features and edge features through graph convolution operations to generate local aggregate features for each node; Performing global importance weighting on the local aggregated features through the graph attention mechanism to generate a global attention feature for each node; The global attention features are spliced ​​along the node dimension to generate correlation features that reflect the collaborative relationship of multi-source signals.

6. The method according to claim 2, characterized in that The analyzing the abnormality type of the wind turbine system and the development trajectory of the abnormality in the time dimension and the identification of the affected components in the space dimension based on the abnormality diagnosis result includes: Extracting the abnormality type label with a confidence level greater than a set confidence level in the abnormality diagnosis result as the abnormality type; Extracting the original monitoring signal segment corresponding to the timestamp of the abnormality occurrence from the state evolution feature sequence, and determining the starting time point of the initial occurrence of the abnormality by backtracking the timestamp; Extracting the original monitoring signal segment corresponding to the abnormal stable appearance timestamp from the state evolution feature sequence, and determining the end time point of the abnormal stable existence through backtracking the timestamp; The time series between the starting time point and the ending time point is used as the time range of the development trajectory; Extracting correlation features within the time range of the anomaly occurrence from the cross-source correlation feature set, and determining the monitoring signal type with the highest correlation with the anomaly through feature importance analysis; According to the physical sensor position information corresponding to the monitoring signal type, a specific component directly affected by the abnormality is determined as the affected component identifier.

7. The method according to claim 6, characterized in that Extracting correlation features within the time range of the anomaly occurrence from the cross-source correlation feature set, and determining the monitoring signal type with the highest correlation with the anomaly through feature importance analysis, includes: intercepting a correlation feature subsequence whose time range in the cross-source correlation feature set is aligned with the time range of the development trajectory; Calling a feature importance evaluation module of the abnormality diagnosis model, and using the feature importance evaluation module to calculate a gradient value of the associated feature subsequence to the abnormality diagnosis result through gradient back propagation; The gradient value is used as the feature importance score, and the associated feature with the highest feature importance score is selected as the key associated feature; Determining the corresponding monitoring signal type according to the source identifier of the key correlation feature, wherein the first correlation feature corresponds to a combination of a vibration sensor signal and a temperature sensor signal, and the second correlation feature corresponds to a combination of a speed sensor signal and a temperature sensor signal; The monitoring signal type corresponding to the key correlation feature is taken as the monitoring signal type with the highest correlation degree with the abnormality.

8. The method according to claim 1, characterized in that The generating of the equipment warning instruction including the time and space positioning information according to the abnormality type, the development trajectory, and the affected component identifier includes: Obtain a preset mapping relationship table between abnormality types and warning levels, and search for the corresponding warning level identifier according to the abnormality type; Extracting the start time point and the end time point of the anomaly from the time range of the development trajectory to generate time location information including a description of the time interval; Searching a preset component position information table according to the affected component identifier to obtain the spatial coordinates of the affected component as spatial positioning information; Constructing an information structure of the device warning instruction, wherein the information structure includes the warning level identifier, the time positioning information, and the space positioning information; The information structure is converted into a communication protocol format recognizable by the target management terminal to generate the device early warning instruction.

9. The method according to claim 1, characterized in that The training process of the feature encoding network and the abnormality diagnosis model includes: Acquire historical normal operation data and historical abnormal operation data of the wind turbine system, wherein the formats of the historical normal operation data and the historical abnormal operation data are consistent with the multi-source monitoring data stream; Performing timestamp alignment processing on the historical normal operation data and the historical abnormal operation data to generate a training sample set; Initializing network parameters of the feature encoding network and the abnormality diagnosis model; Inputting the multi-source monitoring data in the training sample set into the feature encoding network to generate a state evolution feature sequence and a cross-source correlation feature set for training; Inputting the state evolution feature sequence and the cross-source correlation feature set for training into the anomaly diagnosis model to generate an anomaly diagnosis result for training; Calculating the loss value between the abnormal diagnosis result used for training and the true label of the sample; The network parameters of the feature encoding network and the abnormality diagnosis model are updated by a back propagation algorithm until the loss value converges to a preset threshold.

10. An abnormal state detection system for a wind turbine system, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the abnormal state detection method for a wind turbine system as described in any one of claims 1 to 9.

Images