Fault diagnosis method and system for networking type energy storage system based on multi-modal characteristics

Abstract

The invention relates to the technical field of power grids, and provides a multi-modal feature-based fault diagnosis method and a multi-modal feature-based fault diagnosis system for a network construction type energy storage system, which are used for sequentially carrying out normalization, multi-scale convolution, dynamic weighting of an attention mechanism, spatial-temporal feature extraction of a bidirectional long-short-term memory network and weighted pooling of the attention mechanism on acquired multi-modal data. According to the multi-modal-feature-based fault diagnosis method for the network-building type energy storage system, through multi-scale convolution, attention mechanism dynamic weighting, bidirectional long-short-term memory network spatial-temporal feature extraction and attention mechanism weighted pooling, each modal feature can be effectively extracted, the multi-layer network depth extension requirement of deep learning is avoided, and the fault diagnosis efficiency is improved. Model parameter quantity can be effectively reduced, dependence on cloud computing is reduced, and the method can be conveniently deployed on edge equipment; and transient faults and progressive faults in the multi-modal features can be effectively identified, and the fault diagnosis precision is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of power grids, and particularly to a fault diagnosis method and system for a grid-forming energy storage system based on multi-modal features. Background Art

[0002] With the accelerating transformation of the global energy structure, the high-proportion access of new energy to the power grid has become the core feature of the new power system. However, the dynamic reactive power support capabilities of new energy devices (such as wind power and photovoltaic) are significantly insufficient compared with conventional units, and they are usually connected to the end of the power grid, resulting in the system presenting "low inertia, weak damping" characteristics, which are prone to cause voltage stability problems.

[0003] GridForming energy storage technology can actively construct voltage / frequency by simulating the characteristics of synchronous machines, which can improve the system inertia and damping levels and effectively enhance the voltage support capabilities of new energy devices. However, the failure rate of its core component, the power conversion system (PCS), is still as high as 0.5% per year. To ensure the reliability of the power system, it is necessary to perform fault diagnosis on the power conversion system to perform maintenance in a timely manner according to the fault causes.

[0004] The fault causes of the power conversion system are diverse. To improve the accuracy of fault diagnosis, in the prior art, the diagnostic accuracy can be effectively improved by fusing multi-source heterogeneous data. For example, deep learning methods, which are processed by a multi-modal fusion deep learning model (such as CNNLSTM), can fuse vibration, temperature, and electrical signals for diagnosis. However, the number of model parameters is large, relying on cloud computing, and it is difficult to be deployed on edge devices. Summary of the Invention

[0005] Based on this, the objective of the present invention is to provide a fault diagnosis method and system for a grid-forming energy storage system based on multi-modal features, so as to solve the problem that the multi-modal feature fusion diagnosis method in the prior art has high requirements for computing resources and is difficult to be deployed on edge devices.

[0006] On the one hand, the present invention provides a fault diagnosis method for a grid-forming energy storage system based on multi-modal features, including: Collecting the working condition data of a target power conversion system, and obtaining the modal features of multiple target modes according to the working condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; Normalizing each of the modal features to obtain corresponding dimensionless modal features, and obtaining a first feature vector according to each of the dimensionless modal features; Extracting multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector, where the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; Obtain the first weights of the modal features according to the attention mechanism, and weight them into the second feature vector to obtain a third feature vector; Capture the temporal dependence relationship of the third feature vector according to the bidirectional long short-term memory network to obtain a hidden state sequence; Obtain the second weights of the hidden state sequence according to the attention mechanism, and weight them into the hidden state sequence to obtain global features; Map the global features to the fault category space according to the fully connected layer to obtain the fault values of each fault category, so as to obtain the final fault according to the fault values.

[0007] Optionally, the step of obtaining the final fault according to the fault values further includes: converting the fault values into a probability distribution to obtain the fault probabilities of each fault category, and outputting all fault categories whose fault probabilities exceed the preset fault threshold as the final fault.

[0008] Optionally, the step of obtaining the second feature vector by extracting the multi-scale features of the first feature vector according to the multi-scale convolution further includes: Obtain the instantaneous mutation features of the first feature vector according to the 1×1 convolution kernel; Obtain the short-term dependence relationship of the first feature vector according to the 3×3 convolution kernel; Obtain the long-term progressive trend of the first feature vector according to the 5×5 convolution kernel; Connect the instantaneous mutation features, the short-term dependence relationship, and the long-term progressive trend into the second feature vector through the CONCAT function.

[0009] Optionally, the step of obtaining the first weights of the modal features according to the attention mechanism obtains the first weights according to the following calculation formula: ; where, is the weight of the i-th modal feature, is the trainable weight matrix, is the bias vector, exp() is the exponential function, is the i-th modal feature, is the j-th modal feature, and N is the total number of modalities of the modal features.

[0010] Optionally, the second weights are obtained according to the following calculation formula: ; where, is the weight of the hidden state at the t-th time step, is the trainable weight matrix, is the bias vector, exp() is the exponential function, is the hidden state at the t-th time step, is the hidden state at the -th time step, and T is the total number of time steps in a single sampling period.

[0011] Optionally, the fault value is obtained according to the following calculation formula: ; where z is the fault value, is the global feature, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer, is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer, () is the non-linear activation function.

[0012] On the other hand, the present invention provides a fault diagnosis system for a networked energy storage system based on multi-modal features, including: An acquisition module, configured to acquire the operating condition data of a target energy storage converter, and obtain modal features of multiple target modes according to the operating condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; A normalization module, configured to normalize each of the modal features to obtain corresponding dimensionless modal features, and obtain a first feature vector according to each of the dimensionless modal features; A multi-scale feature enhancement module, configured to extract multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector, where the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; A first dynamic weighting module, configured to obtain first weights of each of the modal features according to an attention mechanism, and weight them into the second feature vector to obtain a third feature vector; A spatio-temporal feature extraction module, configured to capture the temporal dependence relationship of the third feature vector according to a bidirectional long short-term memory network to obtain a hidden state sequence; A second dynamic weighting module, configured to obtain second weights of the hidden state sequence according to an attention mechanism, and weight them into the hidden state sequence to obtain a global feature; A fault output module, configured to map the global feature to a fault category space according to a fully connected layer to obtain fault values of each fault class, so as to obtain a final fault according to the fault values.

[0013] Optionally, the fault output module is further configured to: convert each of the fault values into a probability distribution to obtain fault probabilities of each fault class, and output all fault classes whose fault probabilities exceed a preset fault threshold as the final fault.

[0014] The fault diagnosis method for a network-constructing energy storage system based on multi-modal features provided by the present invention can effectively enhance instantaneous faults and progressive faults in multi-modal features through multi-scale convolution; by using an attention mechanism to dynamically allocate the weights of each modal feature, the noise interference of a single modality can be reduced, the robustness can be improved, and the multi-modal fusion effect can be enhanced; by using a bidirectional long short-term memory network to capture long-term temporal dependencies, the recognition ability for progressive faults can be further improved; by weighting the hidden sequence through the attention mechanism and obtaining global features through weighted pooling, the sensitivity to sudden faults can be enhanced. The fault diagnosis method for a network-constructing energy storage system based on multi-modal features provided by the present invention can effectively extract each modal feature through multi-scale convolution, dynamic weighting by the attention mechanism, spatio-temporal feature extraction by the bidirectional long short-term memory network, and weighted pooling by the attention mechanism, avoiding the need for deep expansion of the multi-layer network in deep learning, effectively reducing the number of model parameters, reducing the dependence on cloud computing, and facilitating deployment on edge devices; and can effectively identify instantaneous faults and progressive faults in multi-modal features, improving the accuracy of fault diagnosis.

[0015] Multiple faults can be output through fault threshold comparison, achieving accurate detection of multi-modal composite faults, providing more accurate fault explanations, and facilitating the overall optimal design of the network-constructing energy storage system. Brief Description of the Drawings

[0016] Figure 1 It is the main flowchart of the fault diagnosis method for a network-constructing energy storage system based on multi-modal features in the embodiments of the present invention.

[0017] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. Specific Embodiments

[0018] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0019] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the related listed items.

[0021] To solve the problem that the existing multi-modal feature fusion diagnosis method has high requirements for computing resources and is difficult to be deployed on edge devices. The present invention provides a fault diagnosis method for a networked energy storage system based on multi-modal features. Through multi-scale convolution, instantaneous faults and progressive faults in multi-modal features can be effectively enhanced; through the attention mechanism, the weights of each modal feature can be dynamically allocated, which can reduce the noise interference of a single modality, improve robustness, and enhance the multi-modal fusion effect; through a bidirectional long short-term memory network to capture long-term temporal dependencies, the recognition ability for progressive faults can be further improved; through the attention mechanism to weight the hidden sequence and weighted pooling to obtain global features, the sensitivity to abrupt faults can be improved. Through multi-scale convolution, dynamic weighting by the attention mechanism, spatio-temporal feature extraction by a bidirectional long short-term memory network, and weighted pooling by the attention mechanism, the modal features can be effectively extracted, avoiding the need for deep expansion of the multi-layer network of deep learning, effectively reducing the number of model parameters, reducing the dependence on cloud computing, and facilitating deployment on edge devices; and it can effectively identify instantaneous faults and progressive faults in multi-modal features, improving the accuracy of fault diagnosis.

[0022] Specifically, as Figure 1 shown, it is the main flowchart of the fault diagnosis method for a networked energy storage system based on multi-modal features according to an embodiment of the present invention, including: Step S01: Collect the operating condition data of the target energy storage converter, and obtain the modal features of multiple target modalities according to the operating condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; Step S02: Normalize each of the modal features to obtain the corresponding dimensionless modal features, and obtain the first feature vector according to each of the dimensionless modal features; Step S03: Extract the multi-scale features of the first feature vector according to multi-scale convolution to obtain the second feature vector, where the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; Step S04: Obtain the first weight of each of the modal features according to the attention mechanism, and weight it into the second feature vector to obtain the third feature vector; Step S05: Capture the temporal dependence relationship of the third feature vector according to the bidirectional long short-term memory network to obtain the hidden state sequence; Step S06: Obtain the second weight of the hidden state sequence according to the attention mechanism, and weight it to the hidden state sequence to obtain the global feature; Step S07: Map the global feature to the fault category space according to the fully connected layer to obtain the fault values of each fault category, so as to obtain the final fault according to the fault values.

[0023] In step S01, the electromagnetic features include the total harmonic distortion rate, the temperature features include the temperature rise rate, and the mechanical features include the vibration spectrum energy entropy and the strain cumulative damage. Each feature is calculated through the corresponding sampling data.

[0024] Specifically, the electromagnetic data includes the current and voltage on the AC side of the PCS, as well as the current and voltage of the DC bus, which are obtained through the acquisition circuit. The total harmonic distortion rate and other electromagnetic features can be obtained through calculation of the data.

[0025] The temperature data is mainly the temperature of the power transistors in the PCS, which can be obtained through the temperature sensors set on the surface of the radiator of the power module of the PCS.

[0026] The mechanical data includes the fixed state of the DC bus, the vibration state of the transformer, and the vibration state of the cooling system. The fixed state of the DC bus can be obtained through the high-frequency strain gauges set on the busbar copper row for detection; the vibration state of the transformer can be obtained through the acceleration sensors set on the transformer casing for detecting the vibration condition of the transformer core and the loosening condition of the coil; the vibration state of the cooling system can be obtained through the MEMS (Micro-Electro-Mechanical System) vibration sensors set on the flange of the cooling system fan bracket for detecting the balance condition of the fan and the bearing wear condition.

[0027] In step S02, normalization can adopt, for example, Min-Max normalization: ; where is the original feature parameter, is the normalized feature parameter, is the minimum value of the feature parameter, is the maximum value of the feature parameter. The minimum value and the maximum value of the feature parameter can be obtained according to the historical data of the original feature parameter.

[0028] After normalization, the first feature vector is obtained, where is the first modal feature after normalization, is the second modal feature after normalization, is the Nth modal feature after normalization.

[0029] To achieve the detection of compound faults, step S07 further includes: converting each of the fault values into a probability distribution to obtain the fault probabilities of each fault class, and outputting all fault classes with fault probabilities exceeding a preset fault threshold as the final faults. When compound faults occur, multiple fault points can be effectively identified. For example, electro-thermal coupling faults such as IGBT overcurrent and heat dissipation module failure, electro-mechanical compound faults such as loose bus connection points and sudden increase in harmonic distortion rate, and control-electrical compound faults such as PLL (Phase-Locked Loop) unlocking and DC bus voltage fluctuation.

[0030] Step S03 specifically includes: Obtaining the instantaneous mutation characteristics of the first feature vector according to a 1×1 convolution kernel; Obtaining the short-term dependence relationship of the first feature vector according to a 3×3 convolution kernel; Obtaining the long-term progressive trend of the first feature vector according to a 5×5 convolution kernel; Connecting the instantaneous mutation characteristics, the short-term dependence relationship, and the long-term progressive trend into the second feature vector through the CONCAT function, and the splicing formula is as follows: ; Among them, is the second feature vector, is the feature vector enhanced by the first feature vector processed by a 1×1 convolution kernel, is the feature vector enhanced by the first feature vector processed by a 3×3 convolution kernel, is the feature vector enhanced by the first feature vector processed by a 5×5 convolution kernel, is the splicing function.

[0031] Among them, instantaneous faults include, for example, power transistor short circuits, short-term faults include, for example, voltage dips, and progressive faults include, for example, heat accumulation. Through multi-scale convolution, the corresponding mutation characteristics, short-term characteristics, and progressive characteristics can be effectively enhanced, avoiding the loss of mutation characteristics or progressive characteristics in subsequent processing, thereby effectively ensuring the comprehensive fault diagnosis effect including various different modal characteristics and ensuring the reliability of fault diagnosis.

[0032] In step S04, the first weight is obtained according to the following calculation formula: ; Among them, is the weight of the i-th modal feature, is the trainable weight matrix, is the bias vector, exp() is the exponential function, is the i-th modal feature, is the j-th modal feature, N is the total number of modalities of the modal features, and both i and j belong to [1~N].

[0033] In step S05, the bidirectional long short-term memory network is, for example, a BiLSTM (Bidirectional Long Short-Term Memory) recurrent neural network, which is mainly used to extract progressive fault features. Among them, through the pre-enhancement of the mutation features in step S03, the loss of the mutation features after the processing in step S05 can be avoided, thereby ensuring the reliability of the comprehensive extraction of various fault features.

[0034] In step S06, the second weight is obtained according to the following calculation formula: ; wherein, is the weight of the hidden state at the t-th time step, is the trainable weight matrix, is the bias vector, exp() is the exponential function, is the hidden state at the t-th time step, is the hidden state at the -th time step, and T is the total number of time steps in a single sampling period.

[0035] Through the attention mechanism weighted pooling in step S06, the sensitivity to mutation faults can be further improved. At the same time, through the enhancement of short-term dependence relationships and long-term progressive trends in step S03, the loss of short-term features and progressive features in the attention mechanism weighted pooling can be avoided, ensuring the reliability of the identification and diagnosis of short-term faults and long-term progressive faults.

[0036] In step S07, the fault value is obtained according to the following calculation formula: ; where z is the fault value, is the global feature, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer, is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer, () is the non-linear activation function.

[0037] The present invention also provides a fault diagnosis system for a grid-connected energy storage system based on multi-modal features, including: An acquisition module, configured to acquire the operating condition data of a target energy storage converter, and obtain modal features of multiple target modalities according to the operating condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; A normalization module, which is used to normalize each of the modal features to obtain corresponding dimensionless modal features, and obtain a first feature vector according to each of the dimensionless modal features; A multi-scale feature enhancement module, which is used to extract multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector, and the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; A first dynamic weighting module, which is used to obtain first weights of each of the modal features according to the attention mechanism and weight them into the second feature vector to obtain a third feature vector; A spatio-temporal feature extraction module, which is used to capture the temporal dependence relationship of the third feature vector according to a bidirectional long short-term memory network to obtain a hidden state sequence; A second dynamic weighting module, which is used to obtain second weights of the hidden state sequence according to the attention mechanism and weight them into the hidden state sequence to obtain global features; A fault output module, which is used to map the global features to a fault category space according to a fully connected layer to obtain fault values of each fault category, and obtain a final fault according to the fault values.

[0038] To improve the diagnosis of composite faults, the fault output module is further used to: convert each of the fault values into a probability distribution to obtain fault probabilities of each fault category, and output all fault categories whose fault probabilities exceed a preset fault threshold as the final fault.

[0039] The present invention combines multi-scale convolution, dynamic weighting by the attention mechanism, spatio-temporal feature extraction by a bidirectional long short-term memory network, and weighted pooling by the attention mechanism. And through multi-scale convolution, the transient, short-term, and cumulative features of each modality can be effectively enhanced, and thus the transient, short-term, and cumulative features of each modality can be effectively extracted, avoiding the need for deep expansion of the multi-layer network of deep learning, effectively reducing the number of model parameters, reducing the dependence on cloud computing, and facilitating deployment on edge devices; and can effectively identify instantaneous faults and progressive faults in multi-modal features, improving the accuracy of fault diagnosis.

[0040] In the description of this specification, the descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0041] The above-described embodiments merely represent several specific implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent shall be subject to the appended claims.

Claims

1. A fault diagnosis method for a network-forming energy storage system based on multi-modal features, characterized in that, Including: Collecting the working condition data of the target energy storage converter, and obtaining the modal features of multiple target modes according to the working condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; Normalizing each of the modal features to obtain corresponding dimensionless modal features, and obtaining a first feature vector according to each of the dimensionless modal features; Extracting multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector, where the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; Obtaining a first weight for each of the modal features according to the attention mechanism, and weighting it into the second feature vector to obtain a third feature vector; Capturing the temporal dependence relationship of the third feature vector according to a bidirectional long short-term memory network to obtain a hidden state sequence; Obtaining a second weight for the hidden state sequence according to the attention mechanism, and weighting it into the hidden state sequence to obtain a global feature; Mapping the global feature to a fault category space according to a fully connected layer to obtain the fault values of each fault category, so as to obtain the final fault according to the fault values.

2. The fault diagnosis method for a network-forming energy storage system based on multi-modal features according to claim 1, characterized in that, The step of obtaining the final fault according to the fault values further includes: converting each of the fault values into a probability distribution to obtain the fault probabilities of each fault category, and outputting all fault categories whose fault probabilities exceed a preset fault threshold as the final fault.

3. The fault diagnosis method for a network-forming energy storage system based on multi-modal features according to claim 2, characterized in that, The step of extracting multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector further includes: Obtaining the instantaneous mutation feature of the first feature vector according to a 1×1 convolution kernel; Obtaining the short-term dependence relationship of the first feature vector according to a 3×3 convolution kernel; Obtaining the long-term progressive trend of the first feature vector according to a 5×5 convolution kernel; Connecting the instantaneous mutation feature, the short-term dependence relationship, and the long-term progressive trend into the second feature vector through a CONCAT function.

4. The fault diagnosis method for a network-forming energy storage system based on multi-modal features according to claim 2, characterized in that, The step of obtaining a first weight for each of the modal features according to the attention mechanism obtains the first weight according to the following calculation formula: ； Among them, is the weight of the i-th modal feature, is the trainable weight matrix, is the bias vector, exp() is the exponential function, is the i-th modal feature, is the j-th modal feature, and N is the total number of modalities of the modal features.

5. The fault diagnosis method for a network-forming energy storage system based on multi-modal features according to claim 2, characterized in that, The second weight is obtained according to the following calculation formula: ； Among them, is the weight of the hidden state at the t-th time step, is the trainable weight matrix, is the bias vector, and exp() is the exponential function, is the hidden state at the t-th time step, is the hidden state at the -th time step, and T is the total number of time steps in a single sampling period.

6. The fault diagnosis method for a network-forming energy storage system based on multi-modal features according to claim 2, characterized in that, The fault value is obtained according to the following calculation formula: ； where z is the fault value, is the global feature, is the weight matrix of the first fully connected layer, is the bias term of the first fully connected layer, is the weight matrix of the second fully connected layer, is the bias term of the second fully connected layer, ( ) is the non-linear activation function.

7. A fault diagnosis system for a network-forming energy storage system based on multi-modal features, characterized in that, Including: A collection module for collecting the working condition data of the target energy storage converter, and obtaining the modal features of multiple target modes according to the working condition data, where the modal features include electromagnetic features, temperature features, and mechanical features; A normalization module for normalizing each of the modal features to obtain corresponding dimensionless modal features, and obtaining a first feature vector according to each of the dimensionless modal features; A multi-scale feature enhancement module for extracting multi-scale features of the first feature vector according to multi-scale convolution to obtain a second feature vector, where the multi-scale features include instantaneous mutation features, short-term dependence relationships, and long-term progressive trends; A first dynamic weighting module for obtaining a first weight for each of the modal features according to the attention mechanism, and weighting it into the second feature vector to obtain a third feature vector; A spatio-temporal feature extraction module for capturing the temporal dependence relationship of the third feature vector according to a bidirectional long short-term memory network to obtain a hidden state sequence; A second dynamic weighting module, which is used to obtain a second weight of the hidden state sequence according to the attention mechanism and weight it to the hidden state sequence to obtain global features; A fault output module, which is used to map the global features to a fault category space according to a fully connected layer to obtain fault values of each fault category, so as to obtain a final fault according to the fault values.

8. The fault diagnosis system for a grid-forming energy storage system based on multi-modal features according to claim 7, wherein, The fault output module is further used to: convert each of the fault values into a probability distribution to obtain a fault probability of each fault category, and output all fault categories whose fault probabilities exceed a preset fault threshold as the final fault.

Images