New energy power station equipment intelligent diagnosis system based on artificial intelligence

Abstract

The embodiment of the application relates to the technical field of equipment state monitoring, and particularly relates to a new energy power station equipment intelligent diagnosis system based on artificial intelligence; the system comprises a data acquisition and preprocessing module, which is used for acquiring multi-source sensing data and generating working condition fluctuation factors and decoupling characteristic data; a model library construction and updating module, which is used for constructing a historical characteristic distribution space and a diagnosis model library, and calculating characteristic density; a confidence degree calculation module, which is used for calculating characteristic confidence degrees in combination with working condition fluctuation factors and historical distribution relations; a dynamic routing matching module, which is used for matching a target diagnosis model from the diagnosis model library based on characteristic confidence degrees and characteristic density; a diagnosis execution and correction module, which is used for generating diagnosis results and performing weight dynamic correction according to verification time delay data. The application improves diagnosis accuracy and robustness, and realizes intelligent adaptive diagnosis and continuous optimization.

Description

Technical Field

[0001] This application relates to the field of equipment condition monitoring technology, and in particular to an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence. Background Technology

[0002] Existing intelligent diagnostic technologies for new energy power plant equipment cannot effectively distinguish between changes in the equipment's own health status and the impact of external environmental disturbances. When wind farms encounter sandstorms or extreme weather, fluctuations in sensor data caused by environmental factors are often misjudged as equipment failures, leading to numerous unnecessary shutdowns for maintenance. Meanwhile, genuine early fault signals may be masked by environmental noise, missing the optimal maintenance opportunity and ultimately resulting in major equipment damage. Secondly, existing systems generally use a single diagnostic model to handle all operating conditions. During the transition of photovoltaic power plants from weak sunlight in the early morning to strong sunlight at noon, or during the load change of wind turbines from light winds to strong winds, the models often show significant lack of adaptability. Under certain specific operating conditions, the diagnostic accuracy drops sharply, and even systemic misjudgments occur. More seriously, current technology lacks a quantitative evaluation mechanism for the reliability of diagnostic results. When faced with fault warnings from the system, maintenance personnel cannot judge their reliability and often fall into a dilemma of "over-maintaining if they trust the system, and taking on risks if they don't." Furthermore, existing diagnostic systems remain largely static after deployment, lacking the ability to self-optimize based on actual operational feedback. As equipment ages, operating conditions change, and new failure modes emerge, system performance gradually degrades, leading to increasingly prominent false alarms and missed alarms. Simultaneously, these systems have extremely low utilization efficiency in handling massive amounts of historical operational data and failure cases, failing to effectively identify and reuse similar historical operating conditions. This results in each diagnosis being relatively independent, hindering experience accumulation and knowledge transfer, severely restricting the improvement of intelligent operation and maintenance levels in new energy power plants, and increasing operating costs and safety risks.

[0003] In view of this, the present invention proposes an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence to solve the above problems. Summary of the Invention

[0004] To address this, the present invention provides an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, which overcomes the problems of low accuracy, poor adaptability, and inability to effectively utilize historical data for continuous learning in the existing technology for fault diagnosis of new energy power plant equipment under changing operating conditions.

[0005] To achieve the above objectives, this invention provides an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence. It includes:

[0006] The data acquisition and preprocessing module is used to acquire multi-source sensor data from new energy power plant equipment, and perform operating condition feature extraction and data decoupling processing on the multi-source sensor data to generate operating condition fluctuation factors and decoupled feature data.

[0007] The model library construction and update module is used to construct a historical feature distribution space and a diagnostic model library based on historically collected multi-source sensor data, and to calculate the historical local density of each data point in the historical feature distribution space.

[0008] The confidence calculation module is used to calculate the feature confidence level corresponding to the decoupled feature data based on the distribution relationship of the decoupled feature data in the historical feature distribution space and in combination with the operating condition fluctuation factor.

[0009] The dynamic routing matching module is used to dynamically route and match the target diagnostic model from the diagnostic model library based on the feature confidence and the real-time neighborhood density of the decoupled feature data in the historical feature distribution space.

[0010] The diagnostic execution and correction module is used to input the decoupled feature data into the target diagnostic model, generate the diagnostic results of the new energy power station equipment, and perform dynamic weight correction on the target diagnostic model based on the diagnostic results and the verification delay data obtained from the new energy power station equipment.

[0011] The storage module is used to store the historical feature distribution space, the diagnostic model library, the feature density, the diagnostic results, and the verification delay data.

[0012] Furthermore, the data acquisition and preprocessing module generates operating condition fluctuation factors and decoupling characteristic data, wherein,

[0013] The multi-source sensing data is decomposed into steady-state components and transient components, and the operating condition fluctuation factor is extracted from the transient components.

[0014] Calculate the mean value of the baseline characteristics of the steady-state component within the operating condition interval corresponding to the operating condition fluctuation factor;

[0015] The steady-state component is adaptively canceled using the operating condition fluctuation factor to eliminate feature drift caused by external environmental disturbances and generate the decoupled feature data.

[0016] Furthermore, the model library construction and update module constructs a historical feature distribution space and a diagnostic model library, wherein,

[0017] Multiple sets of historical sensor data collected from the new energy power station equipment within a historical time period are acquired, and the operating condition feature extraction and data decoupling processing are performed on each set of historical sensor data to obtain the corresponding historical operating condition fluctuation factor and historical decoupling feature data.

[0018] The historical feature distribution space is constructed based on all historical decoupling feature data, and the local density of each historical decoupling feature data in the historical feature distribution space is calculated as the feature density.

[0019] The historical decoupling feature data are clustered according to their corresponding historical operating condition fluctuation factors to obtain multiple operating condition categories. A corresponding diagnostic sub-model is trained or assigned to each operating condition category, and all diagnostic sub-models constitute the diagnostic model library.

[0020] Furthermore, the model library construction and update module calculates the historical local density of each data point in the historical feature distribution space, wherein,

[0021] For any candidate data point in the historical feature distribution space, a dynamic radius neighborhood is determined in the historical feature distribution space with the candidate data point as the center.

[0022] Determine the number of neighboring data points that fall within the dynamic radius neighborhood, and calculate the average kernel distance between each neighboring data point and the candidate data point;

[0023] The local reachability density of the candidate data points is determined based on the ratio of the number of neighboring data points to the average kernel distance, and is used as the feature density.

[0024] Further, the confidence calculation module calculates the feature confidence corresponding to the decoupled feature data, wherein,

[0025] Calculate the Mahalanobis distance from the decoupled feature data to the center point of the historical feature distribution space;

[0026] Obtain the intensity of the operating condition fluctuation factor at the current moment, and determine the confidence penalty coefficient based on the intensity of the operating condition fluctuation factor;

[0027] The Mahalanobis distance is corrected using the confidence penalty coefficient to obtain the corrected distance, and the corrected distance is input into the mapping function to output the feature confidence of the decoupled feature data.

[0028] Furthermore, the dynamic routing matching module dynamically routes and matches the target diagnostic model from the diagnostic model library, wherein,

[0029] Obtain the central feature density of the operating condition domain cluster corresponding to each candidate diagnostic model in the diagnostic model library;

[0030] Calculate the density deviation between the feature density of the decoupled feature data and the density of each central feature, and combine the feature confidence to calculate the fit score of each candidate diagnostic model.

[0031] The candidate diagnostic model with the highest fit score is selected as the target diagnostic model.

[0032] If the highest value of the fit score is lower than the preset routing threshold, the model fusion mechanism is triggered to generate a fused diagnostic model as the target diagnostic model.

[0033] Furthermore, the triggering model fusion mechanism uses the generated fusion diagnostic model as the target diagnostic model, wherein,

[0034] Select a predetermined number of candidate diagnostic models from the fit score;

[0035] Calculate the boundary distance from the decoupled feature data to the boundary of the working condition domain cluster corresponding to the top preset number of candidate diagnostic models;

[0036] The fusion weights of each candidate diagnostic model are determined based on the reciprocal of the boundary distance;

[0037] Based on the fusion weight, the output layer parameters of the top-ranked preset number of candidate diagnostic models are weighted and aggregated to generate the fusion diagnostic model.

[0038] Furthermore, the diagnostic execution and correction module performs dynamic weight correction on the target diagnostic model, wherein,

[0039] Obtain the predicted fault type corresponding to the diagnostic result, and monitor whether a real fault label consistent with the predicted fault type appears within a preset time window after the diagnostic result is output;

[0040] If this occurs, the time difference between the output time of the diagnostic result and the time when the actual fault label appears is calculated and used as the verification delay data.

[0041] Construct a custom loss function that includes a delay decay factor, wherein the delay decay factor is inversely proportional to the verification delay data;

[0042] Based on the customized loss function, the attention weights of each feature channel in the target diagnosis model are corrected by backpropagation.

[0043] Furthermore, after monitoring whether a real fault label consistent with the predicted fault type appears within the preset time window, the method further includes:

[0044] If no real fault label matching the predicted fault type appears within the preset time window, the diagnostic result will be marked as a false alarm sample.

[0045] Extract the feature channel activation distribution of the decoupled feature data corresponding to the false alarm sample, and mark the channels in the feature channel activation distribution whose activation intensity exceeds a preset activation threshold as interference channels;

[0046] A weight reduction correction operation is performed on the attention weights of the interference channels in the target diagnostic model.

[0047] Furthermore, the model library construction and update module is also used to incrementally update the diagnostic model library, wherein,

[0048] When the real-time neighborhood density of the newly added decoupled feature data in the historical feature distribution space is lower than the preset sparsity threshold, and the distance between the newly added decoupled feature data and the boundaries of all existing working condition domain clusters is greater than the preset boundary distance threshold, the newly added decoupled feature data is marked as a new working condition seed point.

[0049] Using the seed point of the new working condition as the core, new decoupling feature data that fall within its neighborhood range in the subsequent preset time period are collected to form a new working condition data cluster;

[0050] A new diagnostic sub-model is trained and generated based on the new working condition data cluster, and the new diagnostic sub-model is added to the diagnostic model library;

[0051] The feature distribution corresponding to the new operating condition data cluster is merged into the historical feature distribution space, and the feature density of the affected area is updated.

[0052] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention decomposes multi-source sensor data into steady-state and transient components using empirical mode decomposition (EMD). From the transient components, a weighted fusion of energy density and coefficient of variation (COP) is extracted as a condition fluctuation factor to quantify the intensity of external environmental disturbances. Then, based on the condition fluctuation factor, operating condition intervals are divided, and the mean value of the baseline characteristics of the steady-state components within each interval is calculated. Finally, the condition fluctuation factor is used to adaptively cancel the steady-state components, eliminating feature drift caused by external environmental disturbances from the original features, generating pure equipment state features decoupled from the operating conditions. This effectively solves the problems of severe environmental disturbances affecting sensor data, unstable feature representation, and easy masking of fault information in new energy power plant equipment under variable operating conditions. It ensures that the extracted decoupled feature data can truly reflect the health status of the equipment itself, eliminating the interference of external factors such as sudden wind speed changes, irradiance variations, and temperature fluctuations on diagnostic accuracy.

[0053] Furthermore, this invention first extracts and decouples historical sensor data from operating condition features consistent with real-time data, ensuring that historical and real-time data reside in the same feature representation space. Then, it constructs a historical feature distribution space based on all historical decoupled feature data and employs a local reachability density calculation method using dynamic radius neighborhood. This method determines the dynamic radius by calculating the distance between a candidate data point and its k-th nearest neighbor, counts the number of neighboring points within the neighborhood, calculates the average kernel distance, and uses the ratio of these two distances as the feature density. This achieves an adaptive measure of the local clustering degree in the feature space, avoiding the problem of underestimation in sparse regions and overestimation in dense regions using a fixed radius. Finally, it clusters the historical decoupled feature data according to historical operating condition fluctuation factors, naturally grouping the data according to the intensity of external disturbances. For each operating condition category, a corresponding diagnostic sub-model is trained or assigned, forming an adaptive diagnostic model library. This effectively solves the problem of insufficient generalization ability of a single diagnostic model for new energy power plant equipment under varying operating conditions, enabling the system to quickly locate the applicable diagnostic sub-model based on the current operating conditions, significantly improving the accuracy of fault identification under different environmental disturbance intensities.

[0054] Furthermore, this invention first constructs a multi-dimensional feature distribution space based on historical normal operation data as a benchmark model. By calculating the Mahalanobis distance from the current decoupled feature data to the center point of the distribution space, the inverse transformation of the covariance matrix is ​​used to quantify the degree of feature deviation and consider the correlation between features. Then, based on the intensity of the current operating condition fluctuation factor, a confidence penalty coefficient is dynamically determined through a negative exponential function to achieve an adaptive response to the degree of external environmental disturbance. Finally, the penalty coefficient is used to correct the Mahalanobis distance, comprehensively reflecting the credibility of the current equipment status. Based on this, the dynamic routing matching module uses a local density estimation algorithm to calculate the feature density of the current decoupled feature data, obtains the central feature density of the corresponding operating condition cluster for each candidate diagnostic model in the diagnostic model library, calculates the density deviation value, and weights the feature confidence to obtain the fitness score of each model. The candidate model with the highest score is selected as the target diagnostic model. When the highest score is lower than a preset routing threshold, a model fusion mechanism is triggered to improve diagnostic accuracy through multi-model fusion. It effectively solves the problems of traditional diagnostic systems, such as lack of confidence assessment under changing working conditions, rigid model selection, and poor reliability of diagnosis in boundary areas. It enables the system to dynamically adjust the diagnostic strategy according to data quality and environmental conditions, which significantly improves the credibility of diagnostic results and robustness under complex working conditions.

[0055] Furthermore, this invention first selects a predetermined number of candidate diagnostic models from the fit score, and uses support vector machine boundary recognition technology to calculate the boundary distance from the decoupled feature data to the boundary of each model's operating condition cluster. The fusion weights are determined based on the reciprocal of the boundary distance. Finally, the output layer parameters of the candidate diagnostic models are weighted and aggregated to generate a fused diagnostic model. This effectively solves the problem of poor diagnostic reliability of a single model in the operating condition boundary region and improves the diagnostic robustness under complex operating conditions. The dynamic weight correction mechanism establishes a diagnostic result tracking mechanism to obtain the predicted fault type and monitor the real fault labels within a predetermined time window. It calculates the time difference between the diagnostic output and the fault occurrence as verification delay data and constructs a customized loss function containing a delay attenuation factor. This attenuation factor is inversely proportional to the verification delay. Through backpropagation, the attention weights of each feature channel in the target diagnostic model are corrected, continuously improving the model's sensitivity to early, weak fault features and achieving self-optimization of diagnostic timeliness. The false alarm handling mechanism marks the diagnostic results as false alarm samples when no real fault labels appear within a preset time window. It extracts the feature channel activation distribution of the decoupled feature data, marks the channels with activation intensity exceeding a preset threshold as interference channels, and performs a weight reduction correction operation on the attention weights of the interference channels, thus forming a refined model tuning path driven by false alarm samples. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, according to an embodiment of this application.

[0057] Figure 2 This is a flowchart illustrating the method for determining whether to trigger the model fusion mechanism in the intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, as described in this application embodiment. Detailed Implementation

[0058] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0059] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0060] The method for determining the weighting coefficients α and β in the formula for calculating the operating condition fluctuation factor is as follows: An adaptive entropy weighting method is used, dynamically calculated based on the dispersion of transient component energy density and coefficient of variation in historical data. The information entropy of the energy density sequence and the coefficient of variation sequence is calculated; the smaller the entropy value, the greater the amount of information contained in the indicator, and thus a higher weight is assigned. Specifically, for an energy density sequence E and a coefficient of variation sequence CV containing N historical samples, their information entropy is calculated respectively: α = (1 - entropy_E) / [(1 - entropy_E) + (1 - entropy_CV)], β = 1 - α.

[0061] Methods for determining the number of levels and cluster boundaries for dividing operating condition intervals: The elbow rule combined with the silhouette coefficient is used to determine the optimal number of clusters. For the operating condition fluctuation factor sequence, the average silhouette coefficient is calculated for each cluster number k=2 to 10, and the k value corresponding to the largest silhouette coefficient is selected as the number of operating condition intervals; if multiple k values ​​correspond to similar silhouette coefficients, the k value corresponding to the elbow point is selected. The boundary threshold is the average value of the boundaries of each cluster.

[0062] The method for determining the historical time window length and the future short-term window length in the time attention mechanism is based on the signal attenuation time constant and sampling frequency of the sensor data. The autocorrelation function of each sensor data is calculated. The lag time corresponding to the first decay of the autocorrelation coefficient to 0.5 is used as the baseline value for the window length. The historical window length is taken as 1.5 times this baseline value, and the future window length is taken as 0.5 times the baseline value, rounded up to an integer multiple of the sampling period.

[0063] The method for determining the weight coefficients of each component in the quality mask calculation is as follows: the maximum likelihood estimation method is adopted. On the validation set labeled with missing real data events, the goal is to maximize the correlation coefficient between the repaired data and the real data. The combined coefficients of spatial attention weight, temporal attention weight and consistency weight are optimized by gradient descent.

[0064] The method for determining the value of k (number of nearest neighbors) in determining the dynamic radius is: k=min(30,floor(√N)), where N is the total number of sample points in the historical feature distribution space. When the sample size is less than 100, k=max(5,floor(N / 5)).

[0065] The upper limit constraint multiple for the dynamic radius is determined by taking three times the global average neighborhood radius. The global average neighborhood radius is obtained by calculating the distance to the k-th nearest neighbor of 10% of the sample points randomly selected from the entire feature space and then taking the arithmetic mean.

[0066] The method for determining the minimum positive number ε in the feature density calculation formula is as follows: take ε = 1e-6 × the average feature amplitude of the feature space, where the average feature amplitude is the average of the L2 norms of all historical decoupled feature data.

[0067] The sparsity threshold is determined by taking the 10th percentile of the feature density distribution of all data points in the historical feature distribution space. This quantile is recalculated after each incremental update.

[0068] Boundary distance threshold determination method: Take 1.5 times the average radius of all existing operating domain clusters. The average radius of each operating domain cluster is defined as the arithmetic mean of the Euclidean distances from all sample points within the cluster to the cluster center.

[0069] The method for determining the minimum sample size for collecting new working condition data clusters is as follows: take max(50, 10 × feature dimension), where the feature dimension is the dimension of the decoupled feature data. For multi-class tasks, if the number of classes is greater than 5, the minimum sample size is adjusted to max(100, 20 × feature dimension).

[0070] Method for determining the accuracy threshold for performance validation of new diagnostic sub-model training: Take the average accuracy of each diagnostic sub-model in the existing diagnostic model library on its respective validation set. If the average is lower than 80%, then take 80% as the threshold.

[0071] Method for determining the multiple of the neighborhood radius (seed point expansion): Take twice the distance from the seed point to its k-th nearest neighbor, where the value of k is the same as in method 5 above.

[0072] The method for determining the negative exponential function parameter of the confidence penalty coefficient is as follows: quantile mapping method is used. Historical operating condition fluctuation factors are arranged in ascending order and divided into 10 equal intervals. Within each interval, the false alarm rate is statistically analyzed, and the exponential decay relationship between the false alarm rate and the fluctuation factor is fitted to determine the decay coefficient. The penalty coefficient is calculated using the formula: γ = exp(-λ × F_wave), where λ is determined through fitting.

[0073] Method for determining the parameters of the Mahalanobis distance to confidence function: Based on the Mahalanobis distance distribution of historical normal samples, the parameters of the Sigmoid function are fitted by maximum likelihood estimation, so that the confidence of normal samples after mapping is concentrated in the range of 0.7 to 0.9, and the confidence of abnormal samples after mapping is lower than 0.3.

[0074] The method for determining the division-to-zero constant ϵ in the fit score calculation formula is as follows: take 1% of the variance of the central feature density of all candidate diagnostic models. If this value is less than 1e-6, then take 1e-6.

[0075] Routing threshold determination method: Based on cross-validation. A 20% sample is randomly selected from historical data as the validation set. For each validation sample, its fit score with all candidate models is calculated. The highest score is taken, and the 5th percentile of the highest scores across all validation samples is used as the routing threshold. Recalibration is performed annually.

[0076] The method for determining the preset number N of candidate models selected in model fusion is: N=min(5,max(2,floor(M / 3))), where M is the total number of diagnostic sub-models in the diagnostic model library.

[0077] The method for determining the preset time window length for verification delay monitoring is as follows: it is dynamically set based on the historical development cycle of each fault type. For known fault types, one-third of the time span from the first detectable feature to complete failure is taken as the window length; for unknown fault types, a default of 72 hours is used. The window length is also constrained by the equipment maintenance cycle and cannot exceed the time interval between two planned maintenance sessions.

[0078] The functional form and parameter determination method of the delay attenuation factor are as follows: An inverse proportional weighting function is used, with the delay attenuation factor = 1 / (1+t_delay / T_ref), where T_ref is the reference delay scale. T_ref is taken as the median of all historical verification delay data.

[0079] The method for determining the preset activation threshold in false alarm handling is as follows: based on the activation distribution statistics of normal diagnostic samples, the 95th percentile of the activation intensity distribution is taken as the threshold. This threshold is recalculated after processing every 1000 diagnostic samples.

[0080] The method for determining the exponential decay coefficient of the weight reduction correction operation is as follows: decay coefficient = min(0.5, multiple of the interference channel activation intensity exceeding the threshold × 0.1), ensuring that the single weight reduction amplitude does not exceed 50%.

[0081] The upper limit for the KL divergence regularization constraint is determined by taking twice the standard deviation of the output distribution of the original model on the validation set as the upper limit of the KL divergence. Specifically, the probability distribution of the output of the original model to each sample on the validation set is calculated, the average KL divergence among these distributions is calculated, and twice that value is taken as the upper limit.

[0082] The method for determining the local update range of feature density in incremental updates is as follows: take the newly added sample point as the center, and use 3 times the distance from it to the kth nearest neighbor as the radius to determine all data points in the hypersphere, where the value of k is the same as in method 5.

[0083] Method for determining the logarithmic baseline of the historical characteristic distribution space: For adaptive adjustment of the k value, when the total sample size N < 1000, k = floor(log2(N)); when N ≥ 1000, k = floor( ).

[0084] Methods for determining the neighborhood radius parameter in clustering algorithms: For DBSCAN clustering, the neighborhood radius is the arithmetic mean of the distances between all sample points and their fourth nearest neighbor; for OPTICS clustering, the neighborhood radius is twice the arithmetic mean.

[0085] Please see Figures 1-2 As shown, Figure 1 This is a schematic diagram of an intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, according to an embodiment of this application. Figure 2 This is a flowchart illustrating the method for determining whether to trigger the model fusion mechanism in the intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, as described in this application embodiment.

[0086] This application's embodiment of an artificial intelligence-based intelligent diagnostic system for new energy power plant equipment includes:

[0087] The data acquisition and preprocessing module is used to acquire multi-source sensor data from new energy power plant equipment, and perform operating condition feature extraction and data decoupling processing on the multi-source sensor data to generate operating condition fluctuation factors and decoupled feature data.

[0088] The model library construction and update module is used to construct a historical feature distribution space and a diagnostic model library based on historically collected multi-source sensor data, and to calculate the historical local density of each data point in the historical feature distribution space.

[0089] The confidence calculation module is used to calculate the feature confidence level corresponding to the decoupled feature data based on the distribution relationship of the decoupled feature data in the historical feature distribution space and in combination with the operating condition fluctuation factor.

[0090] The dynamic routing matching module is used to dynamically route and match the target diagnostic model from the diagnostic model library based on the feature confidence and the real-time neighborhood density of the decoupled feature data in the historical feature distribution space.

[0091] The diagnostic execution and correction module is used to input the decoupled feature data into the target diagnostic model, generate the diagnostic results of the new energy power station equipment, and perform dynamic weight correction on the target diagnostic model based on the diagnostic results and the verification delay data obtained from the new energy power station equipment.

[0092] The storage module is used to store the historical feature distribution space, the diagnostic model library, the feature density, the diagnostic results, and the verification delay data.

[0093] In this embodiment of the invention, the multi-source sensor data includes key information such as vibration signals, temperature curves, current and voltage parameters, power output characteristics, and environmental monitoring data related to the equipment's operating status, which are acquired in real time through a multi-channel data acquisition interface. Vibration signals reflect the health status of the equipment's mechanical components, temperature curves record the equipment's thermodynamic changes, current and voltage parameters reflect the operating characteristics of the electrical system, power output characteristics reflect the overall performance of the equipment, and environmental monitoring data provides information on external operating conditions.

[0094] In this embodiment of the invention, after the target diagnostic model outputs the diagnostic result and before the model undergoes dynamic weight correction, the verification delay data is actively collected or passively received from the new energy power plant equipment according to a preset verification time window. Specifically, the diagnostic execution and correction module first obtains the predicted fault type corresponding to the diagnostic result and monitors whether a real fault label consistent with the predicted fault type appears within the preset time window after the diagnostic result is output. If it appears, the time difference between the diagnostic result output time and the time when the real fault label appears is calculated, and this time difference is used as the verification delay data. At this time, the timing for obtaining the verification delay data is when the real fault label appears and is... The system records the time; if no real fault label matching the predicted fault type appears within the preset time window, the diagnosis result is marked as a false alarm sample. Although positive verification delay data is not obtained at this time, the false alarm judgment itself can be regarded as a special negative verification delay information for subsequent weight reduction correction operations. The length of the preset time window is dynamically set based on the historical development cycle of each fault type. For known fault types, 1 / 3 of the time span from the first detectable feature of the fault to complete failure is taken as the window length. For unknown fault types, the default is 72 hours. At the same time, the window length is constrained by the equipment maintenance cycle and does not exceed the time interval between two planned maintenance.

[0095] Understandably, the significance of this timing of data acquisition lies in: ensuring the causal order of model calibration, guaranteeing that the collection of validation latency data occurs after the diagnostic output and before calibration execution, avoiding the use of future data to correct logical errors in past models, and ensuring that the calibration basis comes from the time-series feedback of the device's actual response rather than the self-circulation of prediction results; achieving differentiated calibration based on latency awareness, by incorporating the time difference between the diagnostic output and the actual occurrence of the fault as a key parameter into the customized loss function as a latency decay factor. This factor is inversely proportional to the validation latency data, so that samples with shorter latency receive a larger gradient contribution in backpropagation calibration, thereby guiding the model to prioritize strengthening the attention weights of feature channels that can capture early, weak fault signs; and supporting continuous optimization of diagnostic timeliness. The explicit setting of this acquisition timing enables the system to distinguish between model diagnostic accuracy deviations and response latency fluctuations, achieving quantitative optimization of the diagnostic lead time during the dynamic weight calibration process, and gradually improving the model's early warning capability for progressive faults.

[0096] Specifically, the data acquisition and preprocessing module generates operating condition fluctuation factors and decoupling characteristic data, wherein,

[0097] The multi-source sensing data is decomposed into steady-state components and transient components, and the operating condition fluctuation factor is extracted from the transient components.

[0098] Calculate the mean value of the baseline characteristics of the steady-state component within the operating condition interval corresponding to the operating condition fluctuation factor;

[0099] The steady-state component is adaptively canceled using the operating condition fluctuation factor to eliminate feature drift caused by external environmental disturbances and generate the decoupled feature data.

[0100] In this embodiment of the invention, signal decomposition preprocessing is performed on multi-source sensor data. The Empirical Mode Decomposition (EMD) method is used to decompose the original signal into multiple Intrinsic Mode Functions (IMFs), identifying and separating steady-state and transient components. Signal decomposition is a fundamental step in feature extraction, separating the composite sensor signal into components of different frequencies and time scales, facilitating refined analysis. The decomposition process is based on the local features of the signal, requiring no prior assumptions about the signal form, and is highly adaptable. Steady-state components reflect the normal operating state and long-term trends of the equipment, while transient components contain information about external disturbances, changes in operating conditions, and potential faults. Through a screening process, oscillation modes at different scales are gradually extracted, ultimately yielding a set of IMF components reflecting the intrinsic characteristics of the signal.

[0101] The operating condition fluctuation factor is extracted from transient components. The intensity of external environmental disturbances is quantified by calculating the energy density and coefficient of variation of these transient components. Operating condition fluctuation factor extraction is a crucial step in identifying external disturbances. By quantitatively analyzing the statistical characteristics of transient components, the impact of environmental factors on equipment status is assessed. The extraction process first calculates the instantaneous energy of each transient component, then statistically analyzes the time-varying characteristics of the energy using a sliding window, and finally combines the energy density and coefficient of variation to derive the fluctuation factor. Energy density reflects the intensity of the disturbance, and the coefficient of variation reflects the degree of instability. The formula for calculating the operating condition fluctuation factor is:

[0102] ;

[0103] in, For operating condition fluctuation factor, The energy density of the transient component. The coefficient of variation of the transient component. and The weighting coefficients and .

[0104] The steady-state components are divided into operating condition intervals based on the intensity level of the operating condition fluctuation factor. Historical operating data are divided into different operating condition intervals, and the mean value of the baseline characteristics of the steady-state components within each interval is calculated. Operating condition interval division is a crucial step in data standardization, eliminating the influence of environmental factors by processing data under different external conditions separately. The division process uses clustering algorithms to determine reasonable interval boundaries based on the distribution characteristics of the operating condition fluctuation factor, typically dividing the data into four levels: normal, slight disturbance, moderate disturbance, and severe disturbance. Within each operating condition interval, the statistical mean value of each dimension of the steady-state components is calculated as the baseline reference value for that operating condition. These baseline values ​​reflect the normal operating status of the equipment under specific operating conditions, providing a standardized basis for subsequent data decoupling.

[0105] An adaptive cancellation process is employed to remove the steady-state components by utilizing the operating condition fluctuation factor. Based on the current operating condition fluctuation factor intensity, feature drift caused by external environmental disturbances is eliminated from the steady-state components, generating decoupled feature data. Adaptive cancellation is the core technology for removing environmental interference. Through a dynamic compensation mechanism, the original features containing environmental influences are transformed into pure features reflecting the internal state of the equipment. The cancellation process first determines the operating condition interval based on the current operating condition fluctuation factor, then obtains the corresponding baseline feature mean, estimates the impact of environmental disturbances on the features through difference calculation and proportional adjustment, and finally subtracts this impact from the original steady-state components to obtain the decoupled feature data.

[0106] This invention employs empirical mode decomposition (EMD) to decompose multi-source sensor data into steady-state and transient components. From the transient components, a condition fluctuation factor (WFD) weighted by energy density and coefficient of variation is extracted to quantify the intensity of external environmental disturbances. Then, based on the WFD, operating condition intervals are divided, and the mean value of the baseline features of the steady-state components within each interval is calculated. Finally, the WFD is used to adaptively cancel the steady-state components, eliminating feature drift caused by external environmental disturbances from the original features, generating pure equipment state features decoupled from the operating conditions. This effectively solves the problems of severe environmental disturbances affecting sensor data, unstable feature representation, and easy masking of fault information in new energy power plant equipment under variable operating conditions. It ensures that the extracted decoupled feature data can truly reflect the equipment's health status, eliminating the interference of external factors such as sudden wind speed changes, irradiance variations, and temperature fluctuations on diagnostic accuracy.

[0107] Specifically, after the data acquisition and preprocessing module and before the model library construction and update module, a spatial attention mechanism based on a graph convolutional network is first constructed. Each sensor is treated as a graph node, and spatial adjacency relationships are established by calculating the feature correlation matrix between different sensors at the same time. Graph convolution operations are used to aggregate information from adjacent sensors to infer the theoretical value of the missing channel. Simultaneously, a temporal attention mechanism based on a bidirectional long short-term memory network is constructed to extract the temporal dependencies of each sensor within historical and future short-term windows. Multi-scale temporal features are fused using attention weighting to predict missing data points. Then, a learnable gating fusion unit is designed to dynamically weight and combine the spatial inferred values ​​and temporal predicted values, outputting the repaired complete data tensor. Finally, this module also synchronously outputs a data quality mask matrix. The mask value ranges from 0 to 1. For the original valid data points, the mask value is 1. For the repaired data points, the mask value is calculated based on the repair confidence level. This mask value is jointly determined by the spatial attention weight, the temporal attention weight, and the consistency between the two prediction branches.

[0108] In this embodiment of the invention, when a portion of the sensor data channels experience a momentary interruption, the spatiotemporal repair module automatically collects data from other associated sensor channels at the same time. It constructs a spatial graph structure using each sensor as a graph node, determines spatial adjacency by calculating the feature correlation matrix between different sensors, and uses graph convolution operations to aggregate information from adjacent channels to infer the theoretical value of the missing channel. Simultaneously, a bidirectional temporal network is used to extract the changing trend of each sensor within a historical time window and the response features within a future short-term window. These are then weighted and fused using a temporal attention mechanism to form a temporal prediction value. A learnable gated fusion unit adaptively calculates the fusion weights of the spatial inference value and the temporal prediction value based on the intensity of fluctuations in the current operating condition, outputting the repaired complete data tensor. The spatial attention weight, temporal attention weight, and the degree of reconstruction consistency between the two prediction branches are calculated simultaneously to obtain the confidence level of each repair point as a quality mask output. This mask value is then passed to the confidence calculation module for confidence correction of the diagnostic features generated based on the repaired data.

[0109] Specifically, the model library construction and update module constructs a historical feature distribution space and a diagnostic model library, wherein,

[0110] Multiple sets of historical sensor data collected from the new energy power station equipment within a historical time period are acquired, and the operating condition feature extraction and data decoupling processing are performed on each set of historical sensor data to obtain the corresponding historical operating condition fluctuation factor and historical decoupling feature data.

[0111] The historical feature distribution space is constructed based on all historical decoupling feature data, and the local density of each historical decoupling feature data in the historical feature distribution space is calculated as the feature density.

[0112] The historical decoupling feature data are clustered according to their corresponding historical operating condition fluctuation factors to obtain multiple operating condition categories. A corresponding diagnostic sub-model is trained or assigned to each operating condition category, and all diagnostic sub-models constitute the diagnostic model library.

[0113] In this embodiment of the invention, the model library construction and update module first obtains multiple sets of historical sensor data collected from new energy power plant equipment within a preset time period (e.g., the past 12 months) from a historical database. These data cover complete operating records of the equipment under various normal operating conditions and known fault states. For each set of historical sensor data, the same processing flow of the aforementioned data acquisition and preprocessing module is invoked, sequentially performing empirical mode decomposition, steady-state and transient separation, operating condition fluctuation factor calculation, and adaptive cancellation processing to obtain the corresponding historical operating condition fluctuation factor and historical decoupling feature data. This process ensures that historical data and real-time data have a consistent feature representation space, laying the foundation for subsequent distribution space construction and model training.

[0114] Secondly, a historical feature distribution space is constructed based on all historical decoupling feature data. Specifically, all historical decoupling feature data are used as sample points in a high-dimensional space to form a historical feature distribution space describing the distribution of equipment normal state characteristics. This distribution space reflects the statistical regularity of equipment characteristics under multiple operating conditions and time periods, serving as a reference benchmark for subsequent confidence assessment and dynamic routing. Based on this, the local density of each historical decoupling feature data point in the historical feature distribution space is calculated as the feature density. The feature density employs a local reachability density calculation method based on dynamic radius neighborhood. For any candidate data point, a dynamic radius neighborhood is determined centered on it. The number of neighboring data points falling within this neighborhood is counted, and the average kernel distance between each neighboring data point and the candidate data point is calculated. The ratio of the number of neighboring data points to the average kernel distance is used as the local reachability density of that point. Feature density reflects the degree of clustering of historical data in the feature space. Regions with higher density indicate a higher frequency of occurrence of the feature combination, representing a common operating state of the equipment; regions with lower density indicate rare states or potential abnormal states.

[0115] Finally, the historical decoupling feature data is clustered according to its corresponding historical operating condition fluctuation factor. Specifically, the historical operating condition fluctuation factor is used as one of the clustering dimensions, and combined with the decoupling feature data, a density-based clustering algorithm (such as DBSCAN or OPTICS) is used to divide the historical samples into multiple operating condition categories. The operating condition fluctuation factor, as a quantitative indicator of external environmental disturbance, naturally groups the data according to the disturbance intensity, so that samples within the same operating condition category have similar environmental conditions and operating characteristics. After clustering, a corresponding diagnostic sub-model is trained or assigned to each operating condition category. For operating condition categories with sufficient samples, a dedicated diagnostic sub-model is trained using a supervised learning algorithm (such as support vector machine, random forest, or lightweight gradient booster); for operating condition categories with sparse samples, model parameters are borrowed from similar operating condition categories or a general baseline model is assigned through transfer learning. The diagnostic sub-models corresponding to all operating condition categories together constitute a diagnostic model library. This diagnostic model library features operating condition adaptability. During real-time diagnosis, the system can quickly locate the corresponding diagnostic sub-model based on the operating condition category to which the current operating condition fluctuation factor belongs. This avoids the problem of insufficient generalization ability of a single model under varying operating conditions and significantly improves the diagnostic accuracy under different environmental conditions.

[0116] Specifically, the model library construction and update module calculates the historical local density of each data point in the historical feature distribution space, wherein,

[0117] For any candidate data point in the historical feature distribution space, a dynamic radius neighborhood is determined in the historical feature distribution space with the candidate data point as the center.

[0118] Determine the number of neighboring data points that fall within the dynamic radius neighborhood, and calculate the average kernel distance between each neighboring data point and the candidate data point;

[0119] The local reachability density of the candidate data points is determined based on the ratio of the number of neighboring data points to the average kernel distance, and is used as the feature density.

[0120] In this embodiment of the invention, firstly, for any candidate data point in the historical feature distribution space, a dynamic radius neighborhood is determined in the historical feature distribution space centered on that candidate data point. Traditional fixed-radius neighborhood methods struggle to adapt to changes in the local density of the feature space. In sparse regions, a radius that is too small leads to insufficient samples within the neighborhood, while in dense regions, a radius that is too large introduces too many noise samples. To address this issue, this invention employs a dynamic radius determination strategy: calculating the Euclidean distance between the candidate data point and its k-th nearest neighbor data point, and using this distance as the dynamic radius. The value of k is adaptively determined based on the total sample size of the historical feature distribution space, typically set to the logarithm of the total sample size or a fixed empirical value (e.g., between 15 and 30). Furthermore, when a candidate data point is located at the edge of the feature space, the dynamic radius may abnormally increase due to a lack of sufficient neighbors. In this case, an upper limit constraint is introduced, limiting the dynamic radius to three times the global average neighborhood radius to ensure a reasonable range for the neighborhood.

[0121] Secondly, the number of neighboring data points falling within the dynamic radius neighborhood is determined, and the average kernel distance between each neighboring data point and the candidate data point is calculated. The number of all other historical decoupled feature data points contained within the spherical neighborhood defined by the dynamic radius is counted, denoted as the number of neighboring data points. This number directly reflects the density of samples around the candidate data point; a larger number indicates denser data in that region. Simultaneously, for each neighboring data point within the neighborhood, a Gaussian kernel function is used to calculate its kernel distance with the candidate data point. Compared to Euclidean distance, kernel distance better characterizes local similarity in high-dimensional feature space, and the Gaussian kernel function controls the rate of similarity decay through its kernel width parameter. After calculating the kernel distances between all neighboring data points and the candidate data point, they are summed and divided by the number of neighboring data points to obtain the average kernel distance. The smaller the average kernel distance, the more similar the candidate data point is to its neighboring data points, and the higher the consistency of local features.

[0122] Finally, the local reachability density of a candidate data point is determined based on the ratio of the number of neighboring data points to the average kernel distance, and this density is used as the feature density of that point. Specifically, the formula for calculating the feature density is:

[0123] ρ=(N_neighbor) / (D_kernel_avg+ε);

[0124] Where ρ is the feature density, N_neighbor is the number of neighboring data points falling within the dynamic radius neighborhood, D_kernel_avg is the average kernel distance, and ε is a very small positive number (e.g., 10⁻). 6This ratio is used to prevent computational overflow caused by a denominator of zero. Essentially, this ratio is a measure of local density: the more neighboring data points and the smaller the average kernel distance, the higher the feature density, indicating that the candidate data point is in a dense region of the historical feature distribution space, and the corresponding feature combination has a high frequency of occurrence in historical operation, representing a common state of the equipment. Conversely, when the number of neighboring data points is small or the average kernel distance is large, the feature density is low, indicating that the candidate data point is in a sparse or marginal region, and the corresponding feature combination is relatively rare, possibly corresponding to abnormal operating conditions or new fault modes.

[0125] This invention first extracts and decouples historical sensor data from real-time data to ensure they reside in the same feature representation space. Then, it constructs a historical feature distribution space based on all decoupled historical feature data and employs a local reachability density calculation method using dynamic radius neighborhood. This method determines the dynamic radius by calculating the distance between a candidate data point and its k-th nearest neighbor, counts the number of neighboring points within the neighborhood, calculates the average kernel distance, and uses the ratio of these two distances as the feature density. This achieves an adaptive measure of the local clustering degree in the feature space, avoiding the problems of underestimation in sparse regions and overestimation in dense regions using a fixed radius. Finally, it clusters the historical decoupled feature data according to historical operating condition fluctuation factors, naturally grouping the data based on the intensity of external disturbances. For each operating condition category, a corresponding diagnostic sub-model is trained or assigned, forming an adaptive diagnostic model library. This effectively solves the problem of insufficient generalization ability of a single diagnostic model for new energy power plant equipment under varying operating conditions, enabling the system to quickly locate the applicable diagnostic sub-model based on the current operating conditions, significantly improving the accuracy of fault identification under different environmental disturbance intensities.

[0126] Specifically, the confidence calculation module calculates the feature confidence corresponding to the decoupled feature data, wherein,

[0127] Calculate the Mahalanobis distance from the decoupled feature data to the center point of the historical feature distribution space;

[0128] Obtain the intensity of the operating condition fluctuation factor at the current moment, and determine the confidence penalty coefficient based on the intensity of the operating condition fluctuation factor;

[0129] The Mahalanobis distance is corrected using the confidence penalty coefficient to obtain the corrected distance, and the corrected distance is input into the mapping function to output the feature confidence of the decoupled feature data.

[0130] In this embodiment of the invention, a historical feature distribution space is constructed. By collecting decoupled feature data during normal equipment operation, principal component analysis (PCA) and kernel density estimation methods are used to establish a joint probability distribution model of multidimensional features. The historical feature distribution space serves as a benchmark model for confidence assessment, reflecting the statistical regularity of features under normal equipment conditions. The construction process first collects a large amount of decoupled feature data during normal operation to ensure the representativeness and completeness of the samples; then, PCA is used for dimensionality reduction to extract the main feature components and reduce computational complexity; finally, a Gaussian mixture model or kernel density estimation method is used to fit the joint probability distribution of the features, forming the probability density function of the multidimensional feature space. This distribution space serves as a "fingerprint" of the normal state, providing a quantitative standard for subsequent anomaly detection.

[0131] The Mahalanobis distance from the decoupled feature data to the center point of the historical feature distribution space is calculated. Through the inverse transformation of the covariance matrix, the degree to which the current feature deviates from the normal distribution is quantified. Mahalanobis distance is a classic method for multidimensional anomaly detection, considering both the correlation between features and the degree of variation among features. The calculation process first determines the mean vector of the historical feature distribution as the center point, then calculates the covariance matrix to describe the correlation between features, and finally quantifies the degree of deviation of the current feature using the Mahalanobis distance formula. The Mahalanobis distance calculation formula is:

[0132] ;

[0133] in, The Mahalanobis distance, This is the current decoupling feature vector. This is the vector of historical distribution mean. This is the inverse of the covariance matrix. The larger the Mahalanobis distance value, the further the current state deviates from the normal state, and the higher the probability of an anomaly.

[0134] Based on the intensity of the current operating condition fluctuation factor, a confidence penalty coefficient is determined to dynamically adjust the confidence assessment standard under environmental disturbance conditions. The confidence penalty coefficient is a key parameter of the adaptive adjustment mechanism, dynamically adjusting the diagnostic sensitivity according to the complexity of the external environment. When the operating condition fluctuation factor is high, it indicates strong external environmental disturbance; in this case, the confidence requirement should be appropriately lowered to avoid misjudgment. When the operating condition is stable, the confidence standard can be increased to enhance anomaly detection capabilities. The penalty coefficient maps the operating condition fluctuation factor through a negative exponential function, achieving a reasonable nonlinear adjustment relationship and ensuring that an appropriate confidence assessment standard can be obtained under various operating conditions.

[0135] The Mahalanobis distance is corrected using a confidence penalty coefficient to obtain a corrected distance, which is then converted into feature confidence scores using a sigmoid mapping function. The correction process first multiplies the Mahalanobis distance by the confidence penalty coefficient to obtain a corrected distance that considers environmental factors. Then, the sigmoid function maps the corrected distance to confidence values ​​in the [0,1] interval, providing an intuitive probabilistic interpretation. The sigmoid function exhibits good non-linear characteristics, smoothly converting distance metrics into confidence scores, while also possessing reasonable saturation characteristics to avoid the influence of extreme values. The final feature confidence score comprehensively reflects the credibility of the current device status and the reliability of the diagnostic results.

[0136] Specifically, the dynamic routing matching module dynamically routes and matches the target diagnostic model from the diagnostic model library, wherein,

[0137] Obtain the central feature density of the operating condition domain cluster corresponding to each candidate diagnostic model in the diagnostic model library;

[0138] Calculate the density deviation between the feature density of the decoupled feature data and the density of each central feature, and combine the feature confidence to calculate the fit score of each candidate diagnostic model.

[0139] The candidate diagnostic model with the highest fit score is selected as the target diagnostic model.

[0140] If the highest value of the fit score is lower than the preset routing threshold, the model fusion mechanism is triggered to generate a fused diagnostic model as the target diagnostic model.

[0141] In this embodiment of the invention, the feature density of decoupled feature data is calculated using a local density estimation algorithm. This algorithm quantifies the density characteristics of the current feature by statistically analyzing the distribution of neighboring samples in the feature space. Feature density reflects the distribution characteristics of the current device state in the feature space; high-density areas typically correspond to common operating states, while low-density areas may represent abnormal states or boundary conditions. Density estimation employs the k-nearest neighbor method, calculating the number of samples within a fixed radius around the current feature point and then normalizing the result to obtain the density value. This local density estimation method adaptively reflects the local characteristics of the feature distribution, providing an important reference for model selection.

[0142] The central feature density of each candidate diagnostic model in the diagnostic model library corresponding to its operating condition cluster is obtained to establish a density benchmark for model applicability. The diagnostic model library is the core knowledge base of the system, containing specialized diagnostic models trained for different operating conditions and fault types. Each model corresponds to a specific operating condition cluster, with a clearly defined scope of application and performance characteristics. The central feature density reflects the distribution center of the model training data and is an important indicator for evaluating model applicability. By pre-calculating the density characteristics of each model's corresponding operating condition cluster, a quantitative standard for model selection is established, providing foundational data for subsequent suitability evaluation.

[0143] The density deviation between the current feature density and the central feature densities of each candidate model is calculated, and the fitness score of each candidate diagnostic model is calculated by combining the feature confidence scores. The fitness score is a core indicator for model selection, comprehensively considering both data quality and model applicability. The calculation process first determines the density deviation between the current feature and the central features of each model; a smaller deviation indicates better fitness. Then, it combines the feature confidence scores, with data having higher confidence scores receiving greater weight in the fitness calculation. Finally, a weighted combination yields the comprehensive fitness score. The formula for calculating the fitness score is:

[0144] ;

[0145] in, For the model The compatibility rating For feature confidence, For the current feature density, For the model The central feature density, To prevent small constants from being divided by zero.

[0146] The candidate diagnostic model with the highest fitness score is selected as the target diagnostic model. If the highest score is lower than a preset routing threshold, a model fusion mechanism is triggered to generate a fused diagnostic model. The model selection strategy employs a competitive mechanism, determining the optimal model through score ranking. When the fitness of the optimal model is still insufficient, it indicates that the current state may be at the model boundary or in the intersection of multiple models. In this case, a single model is unlikely to provide a reliable diagnosis, and multi-model fusion is needed to improve diagnostic accuracy. The routing threshold is set based on statistical analysis of historical validation data to ensure that unnecessary model fusion is avoided while triggering the fusion mechanism in a timely manner when needed.

[0147] This invention first constructs a multi-dimensional feature distribution space based on historical normal operation data as a benchmark model. It calculates the Mahalanobis distance from the current decoupled feature data to the center point of the distribution space, quantifies the feature deviation using inverse covariance matrix transformation, and considers the correlation between features. Then, based on the intensity of the current operating condition fluctuation factor, it dynamically determines the confidence penalty coefficient using a negative exponential function to achieve an adaptive response to the degree of external environmental disturbance. Finally, it uses the penalty coefficient to correct the Mahalanobis distance, comprehensively reflecting the reliability of the current equipment status. The dynamic routing matching module, based on this, uses a local density estimation algorithm to calculate the feature density of the current decoupled feature data, obtains the central feature density of the corresponding operating condition cluster for each candidate diagnostic model in the diagnostic model library, calculates the density deviation value, and weights the feature confidence to obtain the fitness score of each model. The candidate model with the highest score is selected as the target diagnostic model. When the highest score is lower than a preset routing threshold, a model fusion mechanism is triggered to improve diagnostic accuracy through multi-model fusion. It effectively solves the problems of traditional diagnostic systems, such as lack of confidence assessment under changing working conditions, rigid model selection, and poor reliability of diagnosis in boundary areas. It enables the system to dynamically adjust the diagnostic strategy according to data quality and environmental conditions, which significantly improves the credibility of diagnostic results and robustness under complex working conditions.

[0148] Specifically, the triggering model fusion mechanism uses the generated fusion diagnostic model as the target diagnostic model, wherein,

[0149] Select a predetermined number of candidate diagnostic models from the fit score;

[0150] Calculate the boundary distance from the decoupled feature data to the boundary of the working condition domain cluster corresponding to the top preset number of candidate diagnostic models;

[0151] The fusion weights of each candidate diagnostic model are determined based on the reciprocal of the boundary distance;

[0152] Based on the fusion weight, the output layer parameters of the top-ranked preset number of candidate diagnostic models are weighted and aggregated to generate the fusion diagnostic model.

[0153] In this embodiment of the invention, the top N candidate diagnostic models are selected based on their fit scores. Typically, N is set to 3-5 models to ensure a balance between the diversity of the fusion models and computational efficiency. Model selection follows a diversity principle, considering not only fit scores but also ensuring complementarity in technical approaches and feature sensitivities. The selection process first sorts models by fit scores, then eliminates highly similar models through correlation analysis, ultimately determining a representative set of candidate models. These models will serve as the foundational components of the fusion, working collaboratively to provide more reliable diagnostic results.

[0154] The boundary distance from decoupled feature data to the boundary of the corresponding operating condition domain cluster of each candidate diagnostic model is calculated. Support Vector Machine (SVM) boundary recognition technology is used to accurately quantify the geometric distance from feature points to the boundary of the applicable domain of each model. Boundary distance is a key basis for weight allocation in fusion, reflecting the positional relationship of the current feature within each applicable domain. The calculation process first uses SVM technology to train the boundary recognizer for each operating condition domain, learning the complex geometry of the domain boundaries; then, the shortest distance from the current feature point to each domain boundary is calculated; the closer the distance, the stronger the applicability of the model; finally, the distance values ​​are normalized to eliminate the influence of scale differences between different model domains.

[0155] The fusion weights of each candidate diagnostic model are determined based on the reciprocal of the boundary distance, and softmax normalization is used to ensure the rationality of the weight distribution. Weight allocation is the core step in model fusion and directly affects the performance of the fused model. The weight calculation uses the reciprocal distance principle, giving higher weights to models that are closer in distance, reflecting the advantage of local experts. To avoid the numerical problem of zero distance, a small regularization term is added to the reciprocal calculation. Weight normalization is implemented using the softmax function, ensuring that all weights sum to 1, satisfying the probability distribution requirements, and providing good numerical stability.

[0156] Based on fusion weights, the output layer parameters of the top N candidate diagnostic models are weighted and aggregated to generate a fusion diagnostic model. Parameter aggregation is the final step in building the fusion model, achieving collaborative diagnosis of multiple models through intelligent parameter combination. The aggregation process employs a weighted averaging strategy, linearly combining the output layer weights and biases of each candidate model according to the fusion weights to generate new output layer parameters; simultaneously, the feature extraction layers of each model are preserved, enhancing the model's expressive power through multi-path feature fusion; finally, consistency verification ensures the stability and reliability of the fusion model. The fusion diagnostic model inherits the advantages of each component model, exhibiting stronger robustness and accuracy under complex conditions.

[0157] Specifically, after the dynamic routing matching module and before the diagnostic execution and correction module, a lightweight binary classifier is first connected in parallel at the back end of the feature extraction layer of the target diagnostic model as a domain discriminator. This classifier distinguishes whether the decoupled feature data of the current input comes from the feature distribution corresponding to the source domain (i.e., the historical model library) or from the feature distribution under the current device or operating conditions in the target domain. Then, a gradient reversal layer is embedded in the backpropagation path, so that the feature extractor updates its parameters in the direction of maximizing the loss of the domain discriminator during training, while the domain discriminator itself updates in the direction of minimizing the discrimination loss. This adversarial game forces the feature extractor to learn domain-invariant features. This module also calculates the output probability of the domain discriminator as a domain offset index. When this probability is lower than a preset threshold, a dynamic correction mechanism is triggered. A correction matrix is ​​generated based on the activation response of the intermediate layer of the domain discriminator. This correction matrix is ​​then multiplied element-wise with the attention weights of each feature channel of the target diagnostic model. Finally, a KL divergence regularization constraint is introduced to limit the divergence between the corrected model output distribution and the original model output distribution to not exceed a preset upper limit.

[0158] In this embodiment of the invention, when a matched target diagnostic model is applied to new energy power plant equipment of the same model but with different geographical distributions, the adversarial adaptive module first inputs the current decoupled feature data into the domain discriminator at the back end of the feature extraction layer. The domain discriminator calculates the probability value of the data belonging to the historical model library distribution of the source domain. If the probability value is lower than a preset threshold, it indicates that there is a significant deviation between the current equipment feature distribution and the model training distribution. At this time, the module triggers the adversarial correction mechanism. During the backpropagation process, the gradient inversion layer causes the feature extractor to update in the direction of increasing the domain discrimination error, forcing the feature extractor to discard the site-specific features and retain the domain-invariant features related to the essence of the fault. At the same time, a correction matrix is ​​generated according to the activation response intensity of each neuron in the intermediate layer of the domain discriminator. This matrix is ​​multiplied element-wise with the attention weights of each feature channel of the target diagnostic model to achieve dynamic adjustment of the model's focus. During the entire correction process, the module continuously calculates the KL divergence between the output distribution of the corrected model and the original model to ensure that the divergence does not exceed a preset upper limit to prevent catastrophic forgetting of the model. After the correction is completed, the adjusted model is passed to the diagnostic execution module for fault discrimination.

[0159] Specifically, the diagnosis execution and correction module performs dynamic weight correction on the target diagnosis model, wherein...

[0160] Obtain the predicted fault type corresponding to the diagnostic result, and monitor whether a real fault label consistent with the predicted fault type appears within a preset time window after the diagnostic result is output;

[0161] If this occurs, the time difference between the output time of the diagnostic result and the time when the actual fault label appears is calculated and used as the verification delay data.

[0162] Construct a custom loss function that includes a delay decay factor, wherein the delay decay factor is inversely proportional to the verification delay data;

[0163] Based on the customized loss function, the attention weights of each feature channel in the target diagnosis model are corrected by backpropagation.

[0164] In this embodiment of the invention, a diagnostic result tracking mechanism is established to obtain the predicted fault type corresponding to the diagnostic result and continuously monitor the equipment status within a preset time window, recording the actual fault occurrence. The tracking mechanism is the foundation for model validation and improvement, evaluating model performance by comparing the predicted results with the actual situation. The monitoring process employs a sliding time window strategy, with the window length dynamically adjusted according to the development cycle of different fault types, typically set to 2-3 times the prediction lead time. Monitoring content includes the changing trends of key equipment parameters, alarm signals, and maintenance records, ensuring accurate capture of the time node and type information of fault occurrence.

[0165] The time difference between the diagnostic result output time and the actual fault label occurrence time is calculated and used as verification latency data to evaluate the timeliness and accuracy of the prediction. Verification latency is an important indicator for measuring the value of prediction; the shorter the latency, the more timely the prediction and the higher its value for maintenance decisions. Latency calculation needs to consider the gradual nature of fault development and uses the fault confirmation time rather than the first anomaly time as the benchmark to ensure the objectivity of the assessment. Simultaneously, a latency distribution model is established to analyze the latency patterns under different fault types and severity levels, providing a statistical basis for subsequent corrections.

[0166] A custom loss function incorporating a latency decay factor is constructed, which is inversely proportional to the validation latency data. Dynamic adjustment of the loss function enhances the model's ability to learn early fault features. The custom loss function is a core tool for model optimization; by introducing latency sensitivity, it guides the model to focus more on features with early warning value. The loss function design uses a weighted cross-entropy form, with the latency decay factor serving as the sample weight. Samples with longer latency have lower weights, and samples with shorter latency have higher weights. This design allows the model to prioritize feature patterns that can detect faults early during training, gradually improving its early diagnostic capabilities.

[0167] Based on a customized loss function, backpropagation correction is performed on the attention weights of each feature channel in the target diagnostic model. A gradient update mechanism is used to adjust the model parameters, improving the model's sensitivity to early, subtle fault features. Weight correction is the core mechanism for model self-optimization, improving model performance through targeted parameter adjustments. The correction process employs a backpropagation algorithm based on the attention mechanism, calculating the gradient contribution of each feature channel to the latency-sensitive loss, and then updating the attention weights in the gradient direction. The correction strategy uses an adaptive learning rate, granting smaller adjustments to high-performing feature channels and larger update steps to channels requiring improvement. This refined weight adjustment mechanism maintains overall model stability while specifically enhancing the diagnostic capabilities of key features.

[0168] Specifically, after monitoring whether a real fault label consistent with the predicted fault type appears within the preset time window, the method further includes:

[0169] If no real fault label matching the predicted fault type appears within the preset time window, the diagnostic result will be marked as a false alarm sample.

[0170] Extract the feature channel activation distribution of the decoupled feature data corresponding to the false alarm sample, and mark the channels in the feature channel activation distribution whose activation intensity exceeds a preset activation threshold as interference channels;

[0171] A weight reduction correction operation is performed on the attention weights of the interference channels in the target diagnostic model.

[0172] In this embodiment of the invention, false alarm samples are identified. When no true fault label matching the predicted fault type appears within a preset time window, the corresponding diagnostic result is marked as a false alarm sample, and relevant feature data and diagnostic process information are recorded. False alarm identification is a crucial step in model quality control. Systematic analysis of false alarm samples helps identify weaknesses in the model and directions for improvement. The identification process not only marks false alarm events but also records detailed operating conditions, feature distribution, and model decision paths at the time, providing complete contextual information for subsequent analysis. Simultaneously, a false alarm sample library is established, and the type distribution and occurrence patterns of false alarms are periodically analyzed to identify systemic problems.

[0173] The activation distribution of feature channels corresponding to the decoupled feature data of false alarm samples is extracted. By analyzing the activation intensity of each feature channel in the false alarm decision-making process, key feature channels leading to misjudgments are identified. Activation distribution analysis is an important tool for diagnosing model interpretability. By visualizing the activation patterns of feature channels, the model's decision-making mechanism is revealed. The analysis process employs attention visualization technology to calculate the activation weight of each feature channel on the false alarm samples and identify abnormally high-activated channels. These high-activated channels are often the main cause of false alarms, possibly due to excessive sensitivity to certain interference signals or overreaction to normal changes.

[0174] Channels whose activation intensity exceeds a preset activation threshold in the characteristic channel activation distribution are marked as interference channels. These channels exhibit abnormal sensitivity in false alarm events and require weight adjustment. Interference channel identification is the foundation of accurate correction; quantitative analysis is used to determine the specific targets requiring adjustment. The threshold setting is based on the activation distribution statistics of normal diagnostic samples, typically selecting the 95th percentile of the activation intensity distribution as a benchmark. The marking process also considers the consistency of channel activation; a more conservative processing strategy is adopted for channels with occasional high activation, while a more aggressive correction measure is used for channels with persistent abnormal activation.

[0175] A weight reduction correction operation is performed on the attention weights of interfering channels in the target diagnostic model. This weight decay mechanism reduces the impact of these channels on model decisions while maintaining the stability of the overall model structure. Weight reduction correction is a key technique for refined model tuning, improving model performance through selective parameter adjustments. The correction operation employs an exponential decay strategy, with the decay coefficient dynamically determined based on the degree of interference in each channel; the more severe the interference, the greater the decay. The correction process also includes a cross-validation mechanism to ensure that weight adjustments do not impair the model's performance on normal samples. Through this gradual weight adjustment, the system can progressively improve diagnostic accuracy, reduce false alarm rates, and enhance overall reliability.

[0176] This invention first selects a predetermined number of candidate diagnostic models based on their fit score. It then uses support vector machine boundary recognition technology to calculate the boundary distance from the decoupled feature data to the boundary of each model's operating condition cluster. The fusion weights are determined based on the reciprocal of the boundary distance. Finally, the output layer parameters of the candidate diagnostic models are weighted and aggregated to generate a fused diagnostic model. This effectively solves the problem of poor diagnostic reliability of a single model in the operating condition boundary region and improves the diagnostic robustness under complex operating conditions. The dynamic weight correction mechanism establishes a diagnostic result tracking mechanism to obtain the predicted fault type and monitor the actual fault labels within a predetermined time window. It calculates the time difference between the diagnostic output and the fault occurrence as verification delay data and constructs a customized loss function containing a delay attenuation factor. This attenuation factor is inversely proportional to the verification delay. Through backpropagation, the attention weights of each feature channel in the target diagnostic model are corrected, continuously improving the model's sensitivity to early, weak fault features and achieving self-optimization of diagnostic timeliness. The false alarm handling mechanism marks the diagnostic results as false alarm samples when no real fault labels appear within a preset time window. It extracts the feature channel activation distribution of the decoupled feature data, marks the channels with activation intensity exceeding a preset threshold as interference channels, and performs a weight reduction correction operation on the attention weights of the interference channels, thus forming a refined model tuning path driven by false alarm samples.

[0177] Specifically, the communication interface module is used for data communication with new energy power plant equipment and information exchange with external systems, including:

[0178] A multi-protocol adaptive communication stack is constructed, supporting industrial fieldbus protocols, industrial Ethernet protocols, and wireless communication protocols. Through protocol identification and automatic switching mechanisms, it dynamically matches the communication protocols used by new energy power plant equipment without interrupting data transmission. A data aggregation and distribution unit is built on top of the communication stack. This unit performs protocol-independent transformation on multi-source sensor data collected from different sensor channels according to a unified data modeling standard, mapping data formats carried by different protocols to an internal standard data tensor while retaining the original protocol header information for subsequent traceability verification. A breakpoint resumption and data integrity verification mechanism is embedded within this module. When a communication link is momentarily interrupted, the unuploaded sensor data, along with its timestamp and sequence number, is cached in a local circular buffer. Once the link is restored, it is automatically resumed in a first-in, first-out order. Cyclic redundancy check and message digest dual verification ensure that the data has not been tampered with or lost during transmission. This module also exposes standardized application programming interfaces, including real-time data subscription interfaces, historical data query interfaces, diagnostic result push interfaces, and model configuration management interfaces. These interfaces support data exchange and control command issuance by upper-level monitoring systems, maintenance management systems, and decision support systems through secure authentication mechanisms.

[0179] In this embodiment of the invention, the communication interface module is deployed in the intelligent gateway device of the new energy power station. When the system starts, the adaptive communication stack automatically scans the available communication ports on the device side, identifies that the current wind farm uses the OPC-UA protocol while the photovoltaic area uses the Modbus TCP protocol, and the automatic protocol switching mechanism completes the dynamic loading of the protocol stack in milliseconds without manual configuration or system restart. During the process of the wind speed sensor collecting vibration data at a frequency of 100Hz, the communication link is momentarily interrupted for 1.2 seconds due to lightning interference. The communication interface module automatically writes the data packets during the interruption period with a timestamp accurate to the microsecond level and an incrementing sequence number into the local circular buffer. After the link is restored, it is automatically resumed to the data acquisition and preprocessing module according to the cache order, and the data integrity is ensured by double verification of CRC32 and MD5. The external maintenance management system subscribes to the real-time diagnostic result push interface by calling the RESTful API exposed by the communication interface module and after token authentication. When the diagnostic execution and correction module outputs an early bearing fault warning, the communication interface module encapsulates the diagnostic result in JSON format and actively pushes it to the maintenance management system, triggering the automatic work order generation process, thus realizing the seamless connection between the diagnostic system and the operation and maintenance business system.

[0180] Specifically, the model library construction and update module is also used to incrementally update the diagnostic model library, wherein,

[0181] When the real-time neighborhood density of the newly added decoupled feature data in the historical feature distribution space is lower than the preset sparsity threshold, and the distance between the newly added decoupled feature data and the boundaries of all existing working condition domain clusters is greater than the preset boundary distance threshold, the newly added decoupled feature data is marked as a new working condition seed point.

[0182] Using the seed point of the new working condition as the core, new decoupling feature data that fall within its neighborhood range in the subsequent preset time period are collected to form a new working condition data cluster;

[0183] A new diagnostic sub-model is trained and generated based on the new working condition data cluster, and the new diagnostic sub-model is added to the diagnostic model library;

[0184] The feature distribution corresponding to the new operating condition data cluster is merged into the historical feature distribution space, and the feature density of the affected area is updated.

[0185] In this embodiment of the invention, firstly, when the feature density of newly added decoupled feature data in the historical feature distribution space is lower than a preset sparsity threshold, and the distance between the newly added decoupled feature data and the boundaries of all existing operating condition clusters is greater than a preset boundary distance threshold, it is marked as a seed point for a new operating condition. A feature density lower than the sparsity threshold indicates that the data point is located in a sparse region of the historical feature distribution space, meaning that the feature combination rarely or never appears in historical operation records; a distance greater than the boundary distance threshold from the boundaries of all existing operating condition clusters indicates that the data point does not belong to the neighboring region of any existing operating condition category and cannot be covered by existing models. When both conditions are met simultaneously, the data point is determined to represent a new operating condition mode and is used as a seed point for the new operating condition. Both the sparsity threshold and the boundary distance threshold are dynamically set based on the statistical distribution of historical data. Typically, the sparsity threshold is taken as the 10th quantile of the feature density distribution, and the boundary distance threshold is taken as 1.5 times the average radius of each operating condition cluster.

[0186] Secondly, using the seed point of the new operating condition as the core, newly added decoupling feature data falling within its neighborhood range within a preset time period are collected to form a new operating condition data cluster. After the seed point is marked, the system continuously monitors the newly arriving decoupling feature data and uses a density-based spatial clustering method to include data points whose distance in the feature space is less than the neighborhood radius of the seed point into the same cluster. The neighborhood radius is adaptively determined based on the local sparsity around the seed point, and is usually set to twice the distance from the seed point to its k-th nearest neighbor. The length of the collection time window is dynamically adjusted according to the accumulation rate of the new operating condition data. Collection stops when the number of samples in the cluster reaches the minimum sample size required for training (e.g., 50 to 100). If the time window expires before the minimum sample size is reached, the window is extended or the sample size requirement is reduced to ensure that the new model can be generated in a timely manner.

[0187] Next, a new diagnostic sub-model is trained based on the new operating condition data cluster and added to the diagnostic model library. The training process employs a transfer learning strategy, using the existing operating condition category most similar to the current new operating condition as the pre-training foundation. This pre-trained model is then fine-tuned using the new operating condition data cluster. Transfer learning fully utilizes existing knowledge, significantly reducing the need for new operating condition data and accelerating model convergence. If the new operating condition data cluster is dissimilar to other existing operating condition categories, a supervised learning approach is used for training, employing semi-supervised training based on the label information of the new operating condition data (if available) or on clustering hypotheses. After training, the new diagnostic sub-model undergoes performance validation. Once it reaches a preset accuracy threshold (e.g., above 85%) on the validation set, it is officially added to the diagnostic model library and assigned a new operating condition category identifier.

[0188] Finally, the feature distributions corresponding to the new operating condition data clusters are merged into the historical feature distribution space, and the feature density of the affected areas is updated. The merging operation employs an incremental update strategy, directly adding the sample points of the new operating condition data clusters to the historical feature distribution space without requiring a global recalculation of the entire distribution space. The feature density update uses a local update strategy, recalculating only the feature density of the original data points within the neighborhood of the newly added sample points and the feature density of the newly added sample points themselves, avoiding the computational overhead of a global update. After the update is complete, the new operating condition pattern is fully integrated into the system's knowledge base, and subsequent encounters with similar data can be covered by the existing model without triggering repeated incremental updates.

[0189] Through the aforementioned incremental update mechanism, this invention ensures that the diagnostic model library can be continuously expanded and optimized as the equipment operates over time. New operating conditions and evolving fault types can be automatically identified and learned by the system without manual intervention or offline retraining, significantly improving the system's adaptability and knowledge accumulation capabilities in long-term operation scenarios.

[0190] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this application.

Claims

1. An intelligent diagnostic system for new energy power plant equipment based on artificial intelligence, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source sensor data from new energy power plant equipment, and perform operating condition feature extraction and data decoupling processing on the multi-source sensor data to generate operating condition fluctuation factors and decoupled feature data. The model library construction and update module is used to construct a historical feature distribution space and a diagnostic model library based on historically collected multi-source sensor data, and to calculate the historical local density of each data point in the historical feature distribution space. The confidence calculation module is used to calculate the feature confidence level corresponding to the decoupled feature data based on the distribution relationship of the decoupled feature data in the historical feature distribution space and in combination with the operating condition fluctuation factor. The dynamic routing matching module is used to dynamically route and match the target diagnostic model from the diagnostic model library based on the feature confidence and the real-time neighborhood density of the decoupled feature data in the historical feature distribution space. The diagnostic execution and correction module is used to input the decoupled feature data into the target diagnostic model, generate the diagnostic results of the new energy power station equipment, and perform dynamic weight correction on the target diagnostic model based on the diagnostic results and the verification delay data obtained from the new energy power station equipment. The storage module is used to store the historical feature distribution space, the diagnostic model library, the feature density, the diagnostic results, and the verification delay data.

2. The system according to claim 1, characterized in that, The data acquisition and preprocessing module generates operating condition fluctuation factors and decoupling characteristic data, wherein... The multi-source sensing data is decomposed into steady-state components and transient components, and the operating condition fluctuation factor is extracted from the transient components. Calculate the mean value of the baseline characteristics of the steady-state component within the operating condition interval corresponding to the operating condition fluctuation factor; The steady-state component is adaptively canceled using the operating condition fluctuation factor to eliminate feature drift caused by external environmental disturbances and generate the decoupled feature data.

3. The system according to claim 1, characterized in that, The model library construction and update module constructs a historical feature distribution space and a diagnostic model library, wherein... Multiple sets of historical sensor data collected from the new energy power station equipment within a historical time period are acquired, and the operating condition feature extraction and data decoupling processing are performed on each set of historical sensor data to obtain the corresponding historical operating condition fluctuation factor and historical decoupling feature data. The historical feature distribution space is constructed based on all historical decoupling feature data, and the local density of each historical decoupling feature data in the historical feature distribution space is calculated as the feature density. The historical decoupling feature data are clustered according to their corresponding historical operating condition fluctuation factors to obtain multiple operating condition categories. A corresponding diagnostic sub-model is trained or assigned to each operating condition category, and all diagnostic sub-models constitute the diagnostic model library.

4. The system according to claim 1, characterized in that, The model library construction and update module calculates the historical local density of each data point in the historical feature distribution space, wherein, For any candidate data point in the historical feature distribution space, a dynamic radius neighborhood is determined in the historical feature distribution space with the candidate data point as the center. Determine the number of neighboring data points that fall within the dynamic radius neighborhood, and calculate the average kernel distance between each neighboring data point and the candidate data point; The local reachability density of the candidate data points is determined based on the ratio of the number of neighboring data points to the average kernel distance, and is used as the feature density.

5. The system according to claim 1, characterized in that, The confidence calculation module calculates the feature confidence corresponding to the decoupled feature data, wherein... Calculate the Mahalanobis distance from the decoupled feature data to the center point of the historical feature distribution space; Obtain the intensity of the operating condition fluctuation factor at the current moment, and determine the confidence penalty coefficient based on the intensity of the operating condition fluctuation factor; The Mahalanobis distance is corrected using the confidence penalty coefficient to obtain the corrected distance, and the corrected distance is input into the mapping function to output the feature confidence of the decoupled feature data.

6. The system according to claim 1, characterized in that, The dynamic routing matching module dynamically routes and matches the target diagnostic model from the diagnostic model library, wherein... Obtain the central feature density of the operating condition domain cluster corresponding to each candidate diagnostic model in the diagnostic model library; Calculate the density deviation between the feature density of the decoupled feature data and the density of each central feature, and combine the feature confidence to calculate the fit score of each candidate diagnostic model. The candidate diagnostic model with the highest fit score is selected as the target diagnostic model. If the highest value of the fit score is lower than the preset routing threshold, the model fusion mechanism is triggered to generate a fused diagnostic model as the target diagnostic model.

7. The system according to claim 6, characterized in that, The triggering model fusion mechanism uses the generated fusion diagnostic model as the target diagnostic model, wherein... Select a predetermined number of candidate diagnostic models from the fit score; Calculate the boundary distance from the decoupled feature data to the boundary of the working condition domain cluster corresponding to the top preset number of candidate diagnostic models; The fusion weights of each candidate diagnostic model are determined based on the reciprocal of the boundary distance; Based on the fusion weight, the output layer parameters of the top-ranked preset number of candidate diagnostic models are weighted and aggregated to generate the fusion diagnostic model.

8. The system according to claim 1, characterized in that, The diagnostic execution and correction module performs dynamic weight correction on the target diagnostic model, wherein... Obtain the predicted fault type corresponding to the diagnostic result, and monitor whether a real fault label consistent with the predicted fault type appears within a preset time window after the diagnostic result is output; If this occurs, the time difference between the output time of the diagnostic result and the time when the actual fault label appears is calculated and used as the verification delay data. Construct a custom loss function that includes a delay decay factor, wherein the delay decay factor is inversely proportional to the verification delay data; Based on the customized loss function, the attention weights of each feature channel in the target diagnosis model are corrected by backpropagation.

9. The system according to claim 8, characterized in that, After monitoring whether a real fault label consistent with the predicted fault type appears within the preset time window, the method further includes: If no real fault label matching the predicted fault type appears within the preset time window, the diagnostic result will be marked as a false alarm sample. Extract the feature channel activation distribution of the decoupled feature data corresponding to the false alarm sample, and mark the channels in the feature channel activation distribution whose activation intensity exceeds a preset activation threshold as interference channels; A weight reduction correction operation is performed on the attention weights of the interference channels in the target diagnostic model.

10. The system according to claim 1, characterized in that, The model library construction and update module is also used to incrementally update the diagnostic model library, wherein... When the real-time neighborhood density of the newly added decoupled feature data in the historical feature distribution space is lower than the preset sparsity threshold, and the distance between the newly added decoupled feature data and the boundaries of all existing working condition domain clusters is greater than the preset boundary distance threshold, the newly added decoupled feature data is marked as a new working condition seed point. Using the seed point of the new working condition as the core, new decoupling feature data that fall within its neighborhood range in the subsequent preset time period are collected to form a new working condition data cluster; A new diagnostic sub-model is trained and generated based on the new working condition data cluster, and the new diagnostic sub-model is added to the diagnostic model library; The feature distribution corresponding to the new operating condition data cluster is merged into the historical feature distribution space, and the feature density of the affected area is updated.

Images