High-precision fault diagnosis method and system for energy storage converter signals

Abstract

The invention specifically discloses an energy storage converter signal high-precision fault diagnosis method and system, and the method comprises the steps: firstly carrying out the synchronous processing of three-phase current and DC voltage of an energy storage converter, and constructing a combined signal view which can reflect the instantaneous dynamic state of a system; on the basis, an adaptive signal decomposition technology is introduced, and an original signal which is highly coupled and full of noise is deeply deconstructed into a series of intrinsic mode components with more physical significance. And through a global context coding mechanism, internal relevance and co-evolution laws of the modal components on a time sequence are captured and quantified, so that a depth representation vector capable of amplifying fault weak feature differences is constructed. And finally, based on the depth representation, driving an intelligent classifier to realize accurate identification and positioning of different IGBT open-circuit faults. Therefore, weak differences between different fault modes can be effectively amplified, and reliable technical guarantee is provided for safe operation of the energy storage converter.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis, specifically to a high-precision fault diagnosis method and system for energy storage converter signals. Background Technology

[0002] With the steady progress towards the carbon peaking and carbon neutrality strategic goals, the penetration rate of renewable energy sources, represented by photovoltaics and wind power, in the power system is climbing at an unprecedented rate. To address the impact of their inherent intermittency and volatility on the power grid, large-scale energy storage systems have become a key supporting technology for building new power systems. As the core hub for bidirectional energy exchange between energy storage units and the grid, the energy storage converter operates under harsh conditions of high voltage, high current, and high-frequency switching. Its internal power semiconductor devices (such as IGBTs) and other critical components are weak points in the system, with a high risk of failure. If a fault occurs and is not identified and located in a timely and accurate manner, it will not only interrupt the normal function of the energy storage system but may also trigger a chain reaction, threatening the safety and stability of the entire power grid. Therefore, developing a high-precision, high-reliability energy storage converter fault diagnosis system is of vital practical significance for ensuring the safe operation of modern power systems.

[0003] Against this backdrop, academia and industry have conducted extensive research on fault diagnosis technology for energy storage converters. Most existing solutions rely on the analysis of system-level measurement signals (such as three-phase output current and DC bus voltage). While these methods are effective in handling certain faults with obvious characteristics, their inherent limitations become increasingly apparent as the requirements for refined diagnosis and accurate location increase. Particularly when dealing with faults such as open circuits in IGBTs located in different physical positions, the weak disturbances generated by a single fault source can quickly couple to the entire system due to the closed-loop control characteristics and topological symmetry of the power electronic system. This results in a high degree of similarity in the characteristic responses of different faulty units in the macroscopic output signal, with subtle differences easily masked by strong background noise and harmonic components. Traditional signal processing methods lack adaptability and depth mining capabilities in feature extraction, making it difficult to effectively decouple and capture differentiated information sufficient to distinguish fault sources from such highly similar signals mixed with strong noise. This limits diagnostic accuracy and makes it difficult to accurately trace the source of a specific faulty IGBT unit.

[0004] Therefore, there is a need for an optimized method and system for high-precision fault diagnosis of energy storage converter signals. Summary of the Invention

[0005] The embodiments of this application aim to at least solve one of the technical problems existing in the prior art, and provide a method and system for high-precision fault diagnosis of energy storage converter signals.

[0006] On one hand, embodiments of this application provide a high-precision fault diagnosis method for energy storage converter signals, including: Acquire raw sequence data of three-phase output current and raw sequence data of DC bus voltage; Synchronization windowing is performed on the raw sequence data of three-phase output current and raw sequence data of DC bus voltage to obtain a synchronization signal window matrix; The synchronization signal window matrix is ​​subjected to adaptive variational mode decomposition to obtain the signal eigenmode function matrix; After converting the signal eigenmode function matrix into a time series of eigenmode signal vectors, global context IMF encoding is performed on it to obtain the global signal eigenmode context encoded vector. Fault classification is performed on the global signal intrinsic mode context encoding vector to obtain diagnostic results.

[0007] Optionally, adaptive variational mode decomposition of the synchronization signal window matrix is ​​performed to obtain the signal eigenmode function matrix, including: VMD parameters are initialized for each channel signal in the synchronization signal window matrix to obtain the initial state set of the channel; The constrained variational model is used to solve and iteratively update each channel signal and its corresponding initial channel state to obtain the final set of channel modal components. The final set of channel modal components is arranged into the signal eigenmode function matrix.

[0008] Optionally, after converting the signal eigenmode function matrix into a time series of eigenmode signal vectors, global context IMF encoding is performed on it to obtain a global signal eigenmode context encoded vector, including: Transform the signal eigenmode function matrix into a time series of eigenmode signal vectors; Add position encoding to each intrinsic mode signal vector in the time series of intrinsic mode signal vectors to obtain a time series of intrinsic mode signal vectors containing position information; The time series of the intrinsic mode signal vector containing location information is input into the global context IMF coding network to obtain the global signal intrinsic mode context coding vector.

[0009] Optionally, the time series of the intrinsic mode signal vector containing location information is input into the global context IMF coding network to obtain the global signal intrinsic mode context coding vector, including: The node state randomness of each intrinsic mode signal vector in the time series of intrinsic mode signal vectors is measured to obtain the sequence distribution of the temporal information gain of intrinsic mode signals; Based on the sequence distribution of the temporal information gain of the intrinsic mode signal, the sequence distribution of the temporal adjustment coefficient of the intrinsic mode signal is determined; Based on the sequence distribution of the timing adjustment coefficients of the intrinsic mode signal, the sequence distribution of the intrinsic mode signal vector is weighted to obtain the sequence distribution of the eigenvector of the adjusted intrinsic mode signal. The sequence distribution of the eigenmode signal feature vector after adjustment is input into the forward long short-term memory sequence coding model to obtain the global signal eigenmode context coding vector.

[0010] Optionally, fault classification is performed on the global signal intrinsic mode context encoding vector to obtain diagnostic results, including: The global signal intrinsic mode context encoding vector is input into the classifier to obtain the fault probability vector; Diagnostic decisions are made based on the fault probability vector to obtain the diagnostic results.

[0011] Optionally, the classifier includes one or more fully connected layers, wherein the activation function of the last fully connected layer is: Activation function.

[0012] Optionally, the diagnostic decision is made based on the fault probability vector to obtain the diagnostic result, including: Find the category index with the highest probability and its confidence level from the fault probability vector; Map the category index to a predefined fault label table to obtain the fault label; In response to a confidence level exceeding a preset confidence threshold, the diagnostic result is generated based on the fault label.

[0013] On the other hand, embodiments of this application also provide a high-precision fault diagnosis system for energy storage converter signals, comprising: The raw sequence data acquisition module is used to acquire the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage; The synchronization windowing module is used to synchronize and window the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage to obtain the synchronization signal window matrix. The signal intrinsic mode decomposition module is used to perform adaptive variational mode decomposition on the synchronization signal window matrix to obtain the signal intrinsic mode function matrix. The global context encoding module is used to perform global context IMF encoding on the time series of the signal eigenmode function matrix after converting it into eigenmode signal vectors to obtain the global signal eigenmode context encoded vector. The fault classification and diagnosis module is used to classify faults in the global signal intrinsic mode context encoding vector to obtain diagnostic results.

[0014] Compared with existing technologies, the high-precision fault diagnosis method and system for energy storage converters provided in this application firstly constructs a joint signal view that reflects the instantaneous dynamics of the system by synchronously processing the three-phase current and DC voltage of the energy storage converter. Based on this, an adaptive signal decomposition technique is introduced to deeply deconstruct the highly coupled and noisy original signal into a series of more physically meaningful intrinsic mode components. Furthermore, through a global context encoding mechanism, the intrinsic correlation and co-evolution law of these mode components in the time series are captured and quantified, thereby constructing a deep representation vector that can amplify subtle differences in fault characteristics. Finally, based on this deep representation, an intelligent classifier is driven to accurately identify and locate different IGBT open-circuit faults. This effectively amplifies subtle differences between different fault modes, thus providing a reliable technical guarantee for the safe operation of energy storage converters. Attached Figure Description

[0015] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0016] Figure 1 This is a flowchart of a high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0017] Figure 2 This is a data flow diagram of a high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0018] Figure 3 This is a flowchart of sub-step S3 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0019] Figure 4 This is a flowchart of sub-step S4 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0020] Figure 5 This is a flowchart of sub-step S43 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0021] Figure 6 This is a flowchart of sub-step S5 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0022] Figure 7This is a flowchart of sub-step S52 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application.

[0023] Figure 8 This is a block diagram of a high-precision fault diagnosis system for energy storage converter signals according to an embodiment of this application. Detailed Implementation

[0024] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0025] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0026] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0027] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0028] It is worth noting that all data acquisition actions in this application were carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0029] To address the technical problems described in the background, this application proposes a high-precision fault diagnosis method for energy storage converter signals. First, it synchronously processes the three-phase current and DC voltage of the energy storage converter to construct a joint signal view reflecting the instantaneous dynamics of the system. Based on this, an adaptive signal decomposition technique is introduced to deeply deconstruct the highly coupled and noisy original signal into a series of more physically meaningful intrinsic mode components. Then, through a global context encoding mechanism, the inherent correlation and co-evolution law of these mode components in the time series are captured and quantified, thereby constructing a deep representation vector that can amplify subtle differences in fault characteristics. Finally, based on this deep representation, an intelligent classifier is driven to accurately identify and locate open-circuit faults in different IGBTs. This effectively amplifies subtle differences between different fault modes, thus providing a reliable technical guarantee for the safe operation of the energy storage converter.

[0030] Figure 1 This is a flowchart of a high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 2 This is a data flow diagram of a high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 1 and Figure 2 As shown, the high-precision fault diagnosis method for energy storage converter signals includes the following steps: S1, acquiring the original sequence data of three-phase output current and the original sequence data of DC bus voltage. S2, performing synchronous windowing on the original sequence data of three-phase output current and the original sequence data of DC bus voltage to obtain a synchronous signal window matrix. S3, performing adaptive variational mode decomposition on the synchronous signal window matrix to obtain the signal intrinsic mode function matrix. S4, after converting the signal intrinsic mode function matrix into a time series of intrinsic mode signal vectors, performing global context IMF encoding on it to obtain a global signal intrinsic mode context encoding vector. S5, performing fault classification on the global signal intrinsic mode context encoding vector to obtain a diagnostic result.

[0031] In the aforementioned high-precision fault diagnosis method for energy storage converter signals, step S1 involves acquiring the original sequence data of the three-phase output current and the original sequence data of the DC bus voltage. It should be understood that since the fault characteristics of the energy storage converter (such as IGBT open-circuit faults) are directly reflected in the waveform distortion of the three-phase output current and the ripple changes of the DC bus voltage, and the original electrical signals are the only carrier for capturing fault information, this application uses dedicated sensors to collect continuous time-series data of the three-phase output current and the DC bus voltage to obtain the original data source containing potential fault information. This provides accurate and complete basic data for subsequent signal processing, ensuring that fault characteristics are not lost due to data omissions or distortion, thus laying a data foundation for high-precision diagnosis.

[0032] In the specific implementation process, the three-phase output current data is acquired through Hall current sensors, which are connected in series in the three-phase output circuit of the energy storage converter. These sensors can sense current changes in real time and output corresponding analog signals. The DC bus voltage data is acquired through a high-precision voltage divider resistor network, which is connected in parallel across the DC bus to attenuate the high-voltage signal to a low-voltage signal that can be acquired at a fixed ratio. The two acquired analog signals are synchronously converted into digital sequence data by an analog-to-digital converter triggered by the same clock source. The converted digital sequence data is stored in real time in the data buffer, forming the original sequence data of the three-phase output current and the original sequence data of the DC bus voltage.

[0033] In the aforementioned high-precision fault diagnosis method for energy storage converter signals, step S2 involves synchronizing the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage to obtain a synchronization signal window matrix. It should be understood that since the acquired raw current and DC bus voltage data are continuous data streams, their data volume is enormous and lacks clear analysis boundaries. Furthermore, if the three-phase current and DC voltage signals are out of sync in time, it will lead to distortion of the correlation between fault characteristics across multiple channels, failing to accurately reflect the instantaneous dynamics of the system at the time of the fault. Therefore, this application further performs time alignment processing on the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage, and cuts them into multiple data windows of fixed length to transform the continuous data stream into a structured synchronization signal window matrix. This ensures that the multi-channel signals within each data window are completely matched in the time dimension, avoiding misjudgments of fault characteristics due to time deviations, and simultaneously transforms the data into a matrix form suitable for subsequent adaptive variational mode decomposition, improving the efficiency and accuracy of subsequent processing.

[0034] In the specific implementation process, the window parameters are first set. The window length corresponds to a fixed number of sampling points, which must cover at least two power frequency cycles to ensure that the complete fault characteristic cycle is included. The sliding step size is set to 1 / 2 of the window length to ensure the continuity of data between adjacent windows and avoid missing fault characteristics. Next, a window trigger signal is generated by the system timing control module. When the trigger signal arrives, the original sequence segments of the three-phase output current (including the sampled values ​​of each of the three phases A, B, and C corresponding to the window time) and the original sequence segments of the DC bus voltage (including the sampled values ​​of the corresponding window time) within the same time interval are synchronously extracted from the data buffer. Then, the extracted segments are arranged according to the channel dimension to construct a synchronization signal window matrix. The rows of the matrix correspond to the four signal channels: phase A current, phase B current, phase C current, and DC bus voltage. The columns of the matrix correspond to the sampling points within the window. The data in the same column are the sampled values ​​of the four channels at the same sampling time, forming a structured synchronization signal window matrix for subsequent adaptive variational mode decomposition of the signal.

[0035] In the aforementioned high-precision fault diagnosis method for energy storage converter signals, step S3 involves performing adaptive variational mode decomposition on the synchronization signal window matrix to obtain the signal eigenmode function matrix. It should be understood that because the signal in the synchronization signal window matrix contains fundamental components, high-frequency noise, harmonic interference, and weak fault features, these components are highly coupled, and traditional signal processing methods cannot effectively separate them, leading to the masking of fault features. Therefore, this application further performs adaptive variational mode decomposition on the synchronization signal window matrix to deconstruct the coupled signal into a series of eigenmode components with clear physical meaning. This allows subtle signal distortions caused by faults (such as specific harmonics caused by IGBT open circuits) to be separated from noise and the fundamental frequency, amplifying the characteristic differences between different fault modes, providing a high-purity feature data source for subsequent global context coding, and avoiding a decrease in diagnostic accuracy due to feature mixing. Figure 3 This is a flowchart of sub-step S3 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 3 As shown, step S3 includes the following steps: S31, initializing the VMD parameters of each channel signal in the synchronization signal window matrix to obtain the initial state set of the channels; S32, solving and iteratively updating the constrained variational model for each channel signal and its corresponding initial state to obtain the final mode component set of the channels; and S33, arranging the final mode component set of the channels into the signal eigenmode function matrix.

[0036] Specifically, in step S31, VMD parameters are initialized for each channel signal in the synchronization signal window matrix to obtain an initial state set for each channel. It should be understood that since Variational Mode Decomposition (VMD) relies on iterative optimization, a lack of reasonable initial parameters can lead to slow convergence, getting trapped in local optima, or distorted decomposition results, failing to effectively separate the target modes. Therefore, this application initializes VMD parameters for each channel signal in the synchronization signal window matrix to construct an initial state set for each channel at the start of iteration. This provides a stable starting point for subsequent constraint variational model solving, ensuring that the iterative process of each channel converges quickly to the global optimum, avoiding mode aliasing or decomposition failure due to improper initial parameters, and guaranteeing the separation accuracy of the intrinsic mode components of each channel.

[0037] In the specific implementation process, the four channels of the synchronization signal window matrix are first traversed, and the time series of each channel is processed sequentially. For the currently processed channel, the frequency domain representations of K intrinsic mode functions are initialized to zero, where K is a preset number of modes to ensure that there is no initial interference at the start of the iteration. Next, the center frequency of each mode is initialized and evenly distributed within the normalized frequency range according to frequency bands (covering DC to Nyquist frequencies), so that each mode corresponds to a different signal frequency band. Then, the frequency domain representation of the Lagrange multipliers is initialized to zero to satisfy the initial constraints of the constrained variational model. Finally, the initial mode functions, initial center frequencies, and initial Lagrange multipliers of each channel are integrated to form the initial state set of that channel, with four independent initial state sets corresponding to the four channels.

[0038] Specifically, in step S32, a constrained variational model is solved and iteratively updated for each channel signal and its corresponding initial state to obtain the final set of channel modal components. More specifically, this application further performs constrained variational model solving and iterative updates on each channel signal and its corresponding initial state to gradually optimize the modal components, center frequency, and Lagrange multipliers. This allows each modal component to continuously approach the optimal solution through iteration, satisfying the dual constraints of signal reconstruction and bandwidth minimization, ultimately obtaining a stable, non-aliased set of intrinsic modal components. This effectively separates the hidden weak fault features in the channel, providing highly identifiable feature data for subsequent feature encoding.

[0039] In the specific implementation process, for a single channel, starting with its initial state set, an iterative loop is entered. First, the mode function is updated. In the frequency domain, based on the original signal, other mode components, and Lagrange multipliers, the frequency domain representation of the current mode is adjusted according to the variational model formula. Next, the center frequency is updated. The center frequency is calculated based on the power spectrum of the updated mode function to ensure that the mode corresponds to the target frequency band. Then, the Lagrange multipliers are updated. The multipliers are adjusted according to the signal reconstruction error to strengthen the constraints. After each iteration, the iterative difference of each mode component is calculated. Iteration stops when the difference is less than a preset convergence threshold. Finally, the final frequency domain mode function is subjected to an inverse Fourier transform to convert it into time domain intrinsic mode components, which are integrated into the final mode component set of that channel. The iterative update of the four channels is completed in sequence.

[0040] Specifically, step S33 involves arranging the final modal component sets of each channel into the signal intrinsic mode function matrix. It should be understood that because the final modal component sets of each channel are scattered and lack a unified structured organization, subsequent global context encoding cannot efficiently extract cross-channel modal correlation information, resulting in the loss of collaborative fault feature relationships between different channels. Therefore, this application further arranges the final modal component sets of all channels according to preset rules to construct a unified signal intrinsic mode function matrix. This integrates the scattered modal components into structured data, preserving the temporal sequence and channel correlation of each channel's modal components. This facilitates the subsequent global context encoding module in capturing cross-channel and cross-modal collaborative evolution patterns, amplifying the differentiated features of different fault modes, and providing complete feature matrix support for accurate fault classification.

[0041] In the specific implementation process, the arrangement rules are first determined, and the final mode components of each channel are stacked sequentially according to the channel order of phase A current, phase B current, phase C current, and DC bus voltage. All final mode components of the phase A current channel are taken and sequentially used as the first N rows of the signal intrinsic mode function matrix, where N is the number of modes per channel. Next, all final mode components of the phase B current channel are taken as rows N+1 to 2N of the matrix. Then, all final mode components of the phase C current channel are taken as rows 2N+1 to 3N of the matrix. Finally, all final mode components of the DC bus voltage channel are taken as rows 3N+1 to 4N of the matrix. The number of columns in the matrix remains consistent with the number of sampling points in the synchronization signal window matrix, ultimately forming a complete signal intrinsic mode function matrix.

[0042] In the aforementioned high-precision fault diagnosis method for energy storage converter signals, step S4 involves converting the signal intrinsic mode function matrix into a time series of intrinsic mode signal vectors, followed by global context IMF encoding to obtain a global signal intrinsic mode context encoded vector. It should be understood that the signal intrinsic mode function matrix only presents the static distribution of each modal component, failing to capture the intrinsic correlation and co-evolutionary laws of the modal components in the time dimension. The differences in modal changes caused by open-circuit faults in different IGBTs are reflected in the temporal co-evolutionary features, and relying solely on the static matrix cannot amplify these subtle differences. Therefore, this application further converts the signal intrinsic mode function matrix into a temporal form and then performs global context IMF encoding to quantify the temporal correlation and co-evolutionary laws of the modal components. This constructs a deep representation vector containing global temporal information, effectively amplifying the subtle differences between different fault modes, avoiding insufficient fault feature identification due to a lack of temporal correlation analysis, and laying the foundation for accurate fault identification by subsequent intelligent classifiers. Figure 4 This is a flowchart of sub-step S4 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 4As shown, step S4 includes the following steps: S41, converting the signal eigenmode function matrix into a time series of eigenmode signal vectors; S42, adding position encoding to each eigenmode signal vector in the time series of eigenmode signal vectors to obtain a time series of eigenmode signal vectors containing position information; S43, inputting the time series of eigenmode signal vectors containing position information into a global context IMF coding network to obtain the global signal eigenmode context encoded vector.

[0043] Specifically, step S41 converts the signal intrinsic mode function matrix into a time series of intrinsic mode signal vectors. It should be understood that because the signal intrinsic mode function matrix uses modal components as rows and sampling points as columns, its dimensional structure focuses on the spatial distribution of modes, without considering the time dimension corresponding to the sampling points as the core dimension. This makes it impossible to directly input into a time-series coding network to capture temporal correlations. Therefore, this application further performs dimensional transformation on the signal intrinsic mode function matrix, reconstructing it into a time series of intrinsic mode signal vectors, thereby establishing a time-series data structure centered on time steps. This allows each time step to correspond to the feature values ​​of all modal components at that moment, forming a complete time-series data chain. This provides a suitable data format for subsequent addition of positional encoding and global context encoding, avoiding the loss of time-series information due to data dimensional mismatch.

[0044] In the specific implementation process, the dimensions of the signal intrinsic mode function matrix are first determined. Rows represent the intrinsic mode components of all channels (e.g., 6 modes per channel, 24 rows in total), and columns represent the sampling points of the synchronization window (e.g., 1024 columns). Next, each sampling point is defined as a time step, and data is extracted sequentially from the first column to the last column of the matrix. Each column vector (containing the sampled values ​​of the 24 mode components) is vector-normalized to eliminate dimensional differences. Then, the processed column vectors are arranged in the order of the sampling points, forming a time series of intrinsic mode signal vectors with 1024 time steps and a dimension of 24 for each time step vector. Each vector in this series fully reflects the characteristic states of all modes at the corresponding time point.

[0045] Specifically, step S42 involves adding positional encoding to each intrinsic mode signal vector in the time series of the intrinsic mode signal vector to obtain a time series of intrinsic mode signal vectors containing positional information. It should be understood that, in order to ensure that the global context IMF encoding network (such as a long short-term memory network) can accurately distinguish the order of different time steps when processing time series data, avoid misjudging the temporal pattern of fault evolution as disordered features, and ensure the accuracy of the encoding results, this application further adds positional encoding to each vector in the time series of the intrinsic mode signal vector to inject temporal positional information. This ensures that the encoding network accurately identifies the temporal sequence relationship of each mode vector, correctly captures the evolution of fault features over time, avoids global context encoding distortion due to temporal position confusion, and ensures that subsequent encoding results reflect the true temporal coordination patterns.

[0046] In the specific implementation process, the key parameters of the time series of intrinsic mode signal vectors are first determined, including the total number of time steps (consistent with the number of sampling points in the synchronization signal window matrix) and the dimension of each intrinsic mode signal vector (consistent with the number of rows in the signal intrinsic mode function matrix). Based on these parameters, a position coding matrix is ​​generated. The number of rows in this matrix equals the total number of time steps, and the number of columns equals the vector dimension. Each element in the matrix is ​​calculated using sine and cosine functions of different frequencies to ensure that the position coding at different time steps is unique and reflects the temporal sequence. Subsequently, for each intrinsic mode signal vector in the time series, element-wise addition is performed with the position vector of the corresponding time step in the position coding matrix. After the operation, a time series of intrinsic mode signal vectors incorporating time and position information is obtained. This series can be directly input into the subsequent global context IMF coding network for processing.

[0047] Specifically, in step S43, the time series of the intrinsic mode signal vector containing location information is input into a global context IMF encoding network to obtain the global signal intrinsic mode context encoding vector. It should be understood that since the time series containing location information only possesses temporal identifiers and modal features, it does not explore the global correlation between modal vectors at different time steps, such as the abrupt correlation of modes before and after a fault. Such global correlations are crucial for distinguishing different IGBT open-circuit faults. Therefore, this application further inputs the time series into a global context IMF encoding network to extract and quantify the global temporal correlation and co-evolutionary laws of the modal vectors. This compresses high-dimensional time series data into low-dimensional, highly discriminative deep representation vectors, amplifying the differences in global temporal correlations between different fault modes, avoiding incomplete fault feature extraction due to focusing only on local temporal information, and providing accurate feature support for subsequent fault classification. Figure 5This is a flowchart of sub-step S43 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 5 As shown, step S43 includes the following steps: S431, performing a node state randomness measurement on each intrinsic mode signal vector in the time series of the intrinsic mode signal vector to obtain the sequence distribution of the intrinsic mode signal temporal information gain. S432, determining the sequence distribution of the intrinsic mode signal temporal adjustment coefficients based on the sequence distribution of the intrinsic mode signal temporal information gain. S433, performing weight adjustment processing on the sequence distribution of the intrinsic mode signal vector based on the sequence distribution of the intrinsic mode signal temporal adjustment coefficients to obtain the sequence distribution of the adjusted intrinsic mode signal feature vector. S434, inputting the sequence distribution of the adjusted intrinsic mode signal feature vector into a forward long short-term memory sequence coding model to obtain the global signal intrinsic mode context coding vector.

[0048] More specifically, step S431 involves measuring the node state randomness of each intrinsic mode signal vector in the time series of the intrinsic mode signal vector to obtain the sequence distribution of the temporal information gain of the intrinsic mode signal, expressed by the formula:

[0049] in, It is the th in the time series of the intrinsic mode signal vector The first eigenmode signal vector at time node Location feature value and The first The mean and standard deviation of the intrinsic mode signal vector at each time point. For the first The length of the intrinsic mode signal vector at the time node, Let e ​​be the value of the logarithmic function with the natural constant e as the base. for The probability distribution, for The state entropy of the intrinsic mode signal, for The state entropy of the intrinsic mode signal, To obtain the maximum value, It is the numerical stability constant. In the sequence distribution of the temporal information gain of the intrinsic mode signal and The time-series information gain of the intrinsic mode signals between.

[0050] In other words, the intrinsic mode signal state entropy is obtained by calculating the Gaussian probability distribution of the intrinsic mode signal vector at each time point. Then, the difference in state entropy between adjacent nodes is calculated and normalized to obtain the temporal information gain of the intrinsic mode signal, so as to accurately capture the random mutations caused by faults. During normal operation, the modal signal distribution is stable and the state entropy fluctuates little. During a fault, the disorder of the distribution increases sharply, and the information gain can quantify this change, providing a quantifiable basis for subsequent fault identification and avoiding subjective threshold misjudgment. The sequence distribution of the temporal information gain of the formed intrinsic mode signal can clearly mark the fault time interval. The gain remains at a low level in the normal interval and the gain increases significantly in the fault interval, effectively preventing background noise from masking the fault characteristics.

[0051] More specifically, step S432, based on the sequence distribution of the temporal information gain of the intrinsic mode signal, determines the sequence distribution of the temporal adjustment coefficients of the intrinsic mode signal, expressed by the formula:

[0052] in, This refers to the size of a local window. and For learnable weight parameters, For the sigmoid function, The sequence distribution of the timing adjustment coefficients of the intrinsic mode signal The corresponding intrinsic mode signal timing adjustment coefficient.

[0053] That is, the mean gain of the intrinsic mode signal temporal information within the local window is introduced, and the gain of the current node is compared with the mean gain of the current node through a learnable parameter. , The weighted fusion input to the sigmoid function not only utilizes local means to smooth out the effects of instantaneous noise, but also adapts to the fault characteristics under different loads and grid disturbance conditions in the PCS. This ensures that the intrinsic mode signal timing adjustment coefficients respond only to the continuous changes in the information of the actual fault, suppressing noise interference. Specifically, the sequence distribution of the intrinsic mode signal timing adjustment coefficients maintains low values ​​in the normal and noise ranges, and stably maintains high values ​​in the fault range, eliminating spurious peaks. Furthermore, the trend of change perfectly matches the fault duration, providing a reliable basis for subsequent weight adjustment.

[0054] More specifically, step S433 involves performing weighted adjustment processing on the sequence distribution of the intrinsic mode signal vector based on the sequence distribution of the intrinsic mode signal timing adjustment coefficients to obtain the adjusted sequence distribution of the intrinsic mode signal feature vector, expressed by the formula:

[0055] in, These are the first two sequences of the eigenvectors of the adjusted intrinsic mode signals. The eigenvectors of the adjusted intrinsic mode signals.

[0056] In other words, by element-wise multiplication of each intrinsic mode signal vector with its corresponding temporal adjustment coefficient, the amplitudes of fault features such as harmonics are amplified in the fault region, while the amplitudes of redundant information such as the fundamental mode are reduced in the normal region. Simultaneously, the vector dimension and temporal structure remain unchanged, ensuring that the coding model can completely capture the evolution of the fault from occurrence to stabilization. In the sequence distribution of the adjusted intrinsic mode signal feature vectors, the feature amplitudes in the fault region are significantly higher than those in the normal region, greatly improving fault feature identification. Redundant information and noise amplitudes are compressed to extremely low levels, and the temporal structure is completely preserved, providing a high-purity fault feature data source for the coding model.

[0057] More specifically, in step S434, the sequence distribution of the adjusted intrinsic mode signal feature vector is input into a forward long short-term memory sequence coding model to obtain the global signal intrinsic mode context coding vector, expressed by the formula:

[0058] in, These are the 1st, 2nd, and 3rd elements in the sequence distribution of the eigenvectors of the adjusted intrinsic mode signals, respectively. The and the first The eigenvectors of the adjusted intrinsic mode signals For a forward long short-term memory sequence coding model, This is the global signal intrinsic mode context encoding vector.

[0059] In other words, the input gate, forget gate, and output gate mechanism of a forward LSTM are used to process the regulated sequence node by node. The forget gate ignores redundant features in normal intervals, the input gate focuses on learning fault interval features, and the output gate integrates the temporal dependencies of each node. Ultimately, the sequence distribution of the high-dimensional regulated intrinsic mode signal feature vector is compressed into a low-dimensional global signal intrinsic mode context encoding vector. This vector contains fault occurrence intervals and feature amplitude information, while also preserving evolutionary patterns, providing a compact feature representation for accurate fault differentiation. In this way, the dimensionality of the global signal intrinsic mode context encoding vector is effectively compressed, freeing it from the redundant structure constraints of the original high-dimensional time series data. Simultaneously, the vector contains a very high proportion of global information closely related to the fault, maximizing the preservation of key content such as fault occurrence intervals and feature evolutionary patterns, while significantly eliminating redundant information. For IGBT open-circuit faults at different physical locations in the energy storage converter, the corresponding global signal intrinsic mode context encoding vectors have clear and distinguishable differentiated features, which can effectively distinguish the fault types of IGBTs in different phase bridge arms. This provides highly identifiable feature support for the subsequent fault classifier, significantly improving the classifier's accuracy in identifying IGBT open-circuit faults and meeting the requirements of actual operation and maintenance for diagnostic accuracy.

[0060] In the aforementioned high-precision fault diagnosis method for energy storage converter signals, step S5 involves classifying the global signal intrinsic mode context encoding vector to obtain diagnostic results. It should be understood that although the global signal intrinsic mode context encoding vector integrates the global temporal correlation of modal components with weak fault features, it is only a high-dimensional feature representation and does not establish a mapping relationship with specific fault categories, thus failing to directly output fault information usable for operation and maintenance. Therefore, this application further performs fault classification processing on the global signal intrinsic mode context encoding vector to transform deep feature representation into clear fault identification and location results. This allows for the establishment of a correspondence between features and fault types (such as open-circuit faults in different IGBTs) through classification logic, accurately distinguishing between normal states and various fault modes, while ensuring that the diagnostic results have clear physical meaning, such as the location of the specific faulty IGBT. This provides a direct and reliable decision-making basis for the operation and maintenance of energy storage converters, avoiding the inability to apply diagnostic results due to a disconnect between features and faults. Figure 6 This is a flowchart of sub-step S5 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 6 As shown, step S5 includes the following steps: S51, inputting the global signal intrinsic mode context encoding vector into a classifier to obtain a fault probability vector; S52, performing a diagnostic decision based on the fault probability vector to obtain the diagnostic result.

[0061] Specifically, in step S51, the global signal intrinsic mode context encoding vector is input into the classifier to obtain the fault probability vector. In a specific example of this application, the classifier includes one or more fully connected layers, and the activation function of the last fully connected layer in the one or more fully connected layers is... Activation function. It should be understood that since the global signal intrinsic mode context encoding vector is a deep vector integrating global temporal features, its numerical distribution cannot directly reflect the probability of each fault category, lacking a quantifiable basis for category attribution, making it difficult to support subsequent decision-making. Therefore, this application further inputs this encoding vector into a classifier, thereby using the classifier's nonlinear mapping capability to transform the deep features into probability distributions for each fault category and normal state. This allows for quantifying the degree of matching between the encoding vector and each fault category, providing a quantifiable basis for subsequent diagnostic decisions, and avoiding subjective or misjudgment due to a lack of probability support. Simultaneously, the fully connected layer of the classifier can achieve feature dimensionality compression and focus on key information, strengthening the feature weights with high fault category discrimination and improving the accuracy of the probability distribution.

[0062] In the specific implementation process, the classifier structure is first constructed: the input layer dimension is consistent with the dimension of the global signal intrinsic mode context encoding vector, and two fully connected hidden layers are set in the middle, with the number of nodes being 1 / 2 and 1 / 4 of the dimension of the encoding vector, respectively. A ReLU activation function is used after each hidden layer to achieve nonlinear transformation. The number of nodes in the output layer is consistent with the number of target fault categories (including normal state), and a Softmax activation function is used in the output layer. Next, the global signal intrinsic mode context encoding vector is input into the classifier input layer. The vector undergoes feature dimension compression and key information extraction in the first hidden layer, and then further optimizes the feature representation in the second hidden layer. Finally, the output layer calculates the probability value corresponding to each node using the Softmax function. The probability values ​​of all nodes together constitute a fault probability vector. Each element in this vector corresponds to the probability of a preset category, such as normal state, open circuit of the upper IGBT of phase A, and open circuit of the lower IGBT of phase A, and the sum of all elements is 1.

[0063] Specifically, step S52 involves making a diagnostic decision based on the fault probability vector to obtain the diagnostic result. It should be understood that since the fault probability vector only presents the numerical distribution of probabilities for each category, it does not form a clear fault conclusion, and there may be cases where the confidence of the highest probability category is insufficient. For example, if noise interference causes probability dispersion, directly using this vector can easily lead to misjudgment or ambiguous decision-making. Therefore, this application further conducts diagnostic decisions based on the fault probability vector to extract reliable fault information from the probability distribution and transform it into a structured diagnostic result. This ensures the reliability of the diagnostic result through confidence threshold screening, avoiding the use of low-confidence probability results as the basis for fault assessment. Simultaneously, through the mapping between category indexes and fault labels, abstract probability values ​​are transformed into concrete fault type and location information, making the diagnostic result interpretable and providing operational guidance, ensuring that energy storage converter faults can be accurately located and quickly handled. Figure 7 This is a flowchart of sub-step S52 of the high-precision fault diagnosis method for energy storage converter signals according to an embodiment of this application. Figure 7 As shown, step S52 includes the following steps: S521, finding the category index with the highest probability and its confidence level from the fault probability vector; S522, mapping the category index to a predefined fault label table to obtain a fault label; and S523, generating the diagnostic result based on the fault label in response to the confidence level exceeding a preset confidence threshold.

[0064] More specifically, step S521 involves finding the category index with the highest probability and its confidence level from the fault probability vector. It should be understood that since the fault probability vector contains the probability distribution of normal states and various IGBT open-circuit faults, each probability value corresponds to a different category but lacks a clear primary / secondary identifier. If the category index with the highest probability and its confidence level are not extracted, it will be impossible to subsequently pinpoint the core fault candidate direction, potentially leading to multi-category confusion in decision-making. Therefore, this application further filters the category index with the highest probability value from the fault probability vector and defines this highest probability value as the confidence level, thereby determining the most likely fault category candidate. This provides a unique and clear target object for subsequent fault label mapping, avoiding ambiguity in decision-making due to dispersed probability distributions. Simultaneously, the confidence level quantifies the reliability of the candidate category, providing key parameters for subsequent threshold judgment and ensuring that the initial basis for diagnostic decisions is targeted and quantitatively supported.

[0065] In the specific implementation process, the fault probability vector is first obtained. The dimension of this vector is consistent with the number of preset fault categories. For example, if it includes normal state and 6 types of IGBT open circuit faults, the vector dimension is 7. Then, the vector traversal program is started. Starting from the first element of the vector, the probability values ​​are compared element by element, and the current maximum probability value, the corresponding category index, and the probability value are recorded simultaneously. If multiple probability values ​​are found to be the same and are all the maximum during the traversal, the item with the smallest category index is selected as the target index according to the preset rule. For example, if index 2 and index 3 both have the maximum probability, index 2 is selected to ensure the uniqueness of the result. After the traversal is completed, the determined category index (such as index 3) and its corresponding confidence score, such as 0.98, are output, and this set of data is temporarily stored in the system cache as input data for subsequent fault label mapping.

[0066] More specifically, step S522 maps the category index to a predefined fault label table to obtain fault labels. It should be understood that since the category index is an abstract numerical identifier, such as index 1, index 4, etc., it does not carry physical meaning information such as specific fault type and fault location, and cannot be directly understood or used by maintenance personnel for fault handling. Without mapping, this would lead to a break in the transmission of fault information. Therefore, this application further matches the extracted category index with the predefined fault label table to transform the abstract numerical index into fault labels containing concrete fault characteristics. This allows fault categories to be transformed from meaningless numbers into structured fault descriptions, such as an open-circuit fault of the IGBT in the upper arm of phase B, possessing clear physical meaning and maintenance guidance value. Simultaneously, the predefined label table ensures a one-to-one correspondence between the index and fault information, avoiding mapping deviations and laying the foundation for generating executable diagnostic results.

[0067] In the specific implementation process, a predefined fault label table is first invoked. This table uses a key-value pair structure for storage. The key is the category index, which is completely consistent with the category order of the fault probability vector. The value is the fault label, which contains two dimensions of information: fault type and fault location. For example, index 0 corresponds to the normal state, i.e., no fault; index 1 corresponds to an open circuit fault of the IGBT in the upper arm of phase A, i.e., the upper arm of phase A; index 2 corresponds to an open circuit fault of the IGBT in the lower arm of phase A, i.e., the lower arm of phase A, and so on up to 6 types of faults. Next, a hash query mechanism is established for the label table. The category index obtained in the previous steps is used as the query key and input into the hash table for fast matching. During the query process, the validity of the index is verified. If the index exceeds the range of 0-6, an internal prompt of invalid index and abnormal signal is triggered. After a successful match, the corresponding fault label is output. For example, category index 3 corresponds to an open circuit fault of the IGBT in the upper arm of phase B, i.e., the upper arm of phase B. This label is directly passed to the subsequent diagnostic result generation stage.

[0068] More specifically, in step S523, in response to the confidence level exceeding a preset confidence threshold, the diagnostic result is generated based on the fault label. It should be understood that even the category index with the highest probability may have a low confidence level due to factors such as signal noise and operating condition fluctuations, such as only 0.6. Generating a diagnosis based on this result could easily lead to misjudgment, causing unnecessary maintenance costs or missing real faults. Therefore, this application further determines whether the confidence level exceeds the preset confidence threshold, and only generates a diagnostic result based on the fault label when the condition is met, thereby filtering out fault information that meets reliability standards. This effectively eliminates uncertain results with low confidence, ensuring the accuracy and reliability of the diagnostic results. Simultaneously, the fault label is converted into a structured diagnostic report containing a timestamp, fault details, and confidence level, enabling maintenance personnel to directly locate and handle faults based on the results, avoiding maintenance delays caused by incomplete information, and ensuring the timeliness and accuracy of energy storage converter fault handling.

[0069] In the specific implementation process, a pre-set confidence threshold is first invoked. This threshold is determined through classification and verification of more than 1,000 sets of energy storage converter fault samples. For example, it is set to 0.95 to ensure that the diagnostic misjudgment rate is less than 0.1% under this threshold. The threshold is stored in the system configuration file and can be adjusted as needed. Next, the confidence level obtained in the previous steps is compared with this threshold. If the confidence level exceeds the threshold, such as 0.98 > 0.95, a structured diagnostic result is generated based on the fault label. The result includes four core pieces of information: diagnostic timestamp (accurate to milliseconds), fault type (e.g., IGBT open circuit fault), fault location (e.g., upper arm of phase B), and confidence level (e.g., 0.98). The result is synchronized to the display interface and local log module of the energy storage converter monitoring system. If the confidence level does not exceed the threshold, such as 0.85 < 0.95, a prompt message "The current fault feature confidence level is insufficient. It is recommended to continue monitoring for one power frequency cycle and then re-diagnose" is generated. No fault label-related content is output; only the current fault probability vector and signal acquisition time are recorded for subsequent backtracking analysis.

[0070] In summary, a high-precision fault diagnosis method for energy storage converter signals based on embodiments of this application is elucidated. First, it synchronously processes the three-phase current and DC voltage of the energy storage converter to construct a joint signal view reflecting the instantaneous dynamics of the system. Based on this, an adaptive signal decomposition technique is introduced to deeply deconstruct the highly coupled and noisy original signal into a series of more physically meaningful intrinsic mode components. Then, through a global context encoding mechanism, the intrinsic correlation and co-evolutionary laws of these mode components over time are captured and quantified, thereby constructing a deep representation vector that can amplify subtle differences in fault characteristics. Finally, based on this deep representation, an intelligent classifier is driven to accurately identify and locate different IGBT open-circuit faults. This effectively amplifies subtle differences between different fault modes, thus providing a reliable technical guarantee for the safe operation of the energy storage converter.

[0071] Furthermore, a high-precision fault diagnosis system for energy storage converter signals is also provided.

[0072] Figure 8 This is a block diagram of a high-precision fault diagnosis system for energy storage converter signals according to an embodiment of this application. Figure 8 As shown, the high-precision fault diagnosis system 100 for energy storage converter signals according to an embodiment of this application includes: a raw sequence data acquisition module 110, used to acquire raw sequence data of three-phase output current and raw sequence data of DC bus voltage; a synchronization windowing module 120, used to perform synchronization windowing on the raw sequence data of three-phase output current and raw sequence data of DC bus voltage to obtain a synchronization signal window matrix; a signal intrinsic mode decomposition module 130, used to perform adaptive variational mode decomposition on the synchronization signal window matrix to obtain a signal intrinsic mode function matrix; a global context encoding module 140, used to perform global context IMF encoding on the time series of the signal intrinsic mode function matrix converted into intrinsic mode signal vectors to obtain a global signal intrinsic mode context encoding vector; and a fault classification diagnosis module 150, used to perform fault classification on the global signal intrinsic mode context encoding vector to obtain a diagnosis result.

[0073] Here, those skilled in the art will understand that the specific operation of each module in the above-mentioned high-precision fault diagnosis system for energy storage converter signals has been referenced above. Figures 1 to 7 The high-precision fault diagnosis method for energy storage converter signals has been described in detail, and therefore, its repeated description will be omitted.

[0074] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details of the above embodiments are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the specific details described above.

[0075] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the unit division is only a logical functional division, and other division methods may exist in actual implementation. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the present invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0077] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in a system claim may also be implemented by a single unit through software or hardware.

[0078] Finally, it should be noted that the above description has been given for illustrative and descriptive purposes. Furthermore, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although modifications or equivalent substitutions may be made to the technical solutions with reference to preferred embodiments, they will not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-precision fault diagnosis method for energy storage converter signals, characterized in that, include: Acquire raw sequence data of three-phase output current and raw sequence data of DC bus voltage; Synchronization windowing is performed on the raw sequence data of three-phase output current and raw sequence data of DC bus voltage to obtain a synchronization signal window matrix; The synchronization signal window matrix is ​​subjected to adaptive variational mode decomposition to obtain the signal eigenmode function matrix; After converting the signal eigenmode function matrix into a time series of eigenmode signal vectors, global context IMF encoding is performed on it to obtain the global signal eigenmode context encoded vector. Fault classification is performed on the global signal intrinsic mode context encoding vector to obtain diagnostic results.

2. The high-precision fault diagnosis method for energy storage converter signals according to claim 1, characterized in that, The synchronization signal window matrix is ​​subjected to adaptive variational mode decomposition to obtain the signal eigenmode function matrix, including: VMD parameters are initialized for each channel signal in the synchronization signal window matrix to obtain the initial state set of the channel; The constrained variational model is used to solve and iteratively update each channel signal and its corresponding initial channel state to obtain the final set of channel modal components. The final set of channel modal components is arranged into the signal eigenmode function matrix.

3. The high-precision fault diagnosis method for energy storage converter signals according to claim 1, characterized in that, After converting the signal's intrinsic mode function matrix into a time series of intrinsic mode signal vectors, global context IMF encoding is performed to obtain the global signal intrinsic mode context encoded vector, including: Transform the signal eigenmode function matrix into a time series of eigenmode signal vectors; Add position encoding to each intrinsic mode signal vector in the time series of intrinsic mode signal vectors to obtain a time series of intrinsic mode signal vectors containing position information; The time series of the intrinsic mode signal vector containing location information is input into the global context IMF coding network to obtain the global signal intrinsic mode context coding vector.

4. The high-precision fault diagnosis method for energy storage converter signals according to claim 3, characterized in that, The time series of the intrinsic mode signal vector containing location information is input into the global context IMF coding network to obtain the global signal intrinsic mode context coding vector, including: The node state randomness of each intrinsic mode signal vector in the time series of intrinsic mode signal vectors is measured to obtain the sequence distribution of the temporal information gain of intrinsic mode signals; Based on the sequence distribution of the temporal information gain of the intrinsic mode signal, the sequence distribution of the temporal adjustment coefficient of the intrinsic mode signal is determined; Based on the sequence distribution of the timing adjustment coefficients of the intrinsic mode signal, the sequence distribution of the intrinsic mode signal vector is weighted to obtain the sequence distribution of the eigenvector of the adjusted intrinsic mode signal. The sequence distribution of the eigenmode signal feature vector after adjustment is input into the forward long short-term memory sequence coding model to obtain the global signal eigenmode context coding vector.

5. The high-precision fault diagnosis method for energy storage converter signals according to claim 1, characterized in that, Fault classification is performed on the global signal intrinsic mode context encoding vector to obtain diagnostic results, including: The global signal intrinsic mode context encoding vector is input into the classifier to obtain the fault probability vector; Diagnostic decisions are made based on the fault probability vector to obtain the diagnostic results.

6. The high-precision fault diagnosis method for energy storage converter signals according to claim 5, characterized in that, The classifier comprises one or more fully connected layers, wherein the activation function of the last fully connected layer is: Activation function.

7. The high-precision fault diagnosis method for energy storage converter signals according to claim 5, characterized in that, Diagnostic decisions are made based on fault probability vectors to obtain the diagnostic results, including: Find the category index with the highest probability and its confidence level from the fault probability vector; Map the category index to a predefined fault label table to obtain the fault label; In response to a confidence level exceeding a preset confidence threshold, the diagnostic result is generated based on the fault label.

8. A high-precision fault diagnosis system for energy storage converter signals, characterized in that, include: The raw sequence data acquisition module is used to acquire the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage; The synchronization windowing module is used to synchronize and window the raw sequence data of the three-phase output current and the raw sequence data of the DC bus voltage to obtain the synchronization signal window matrix. The signal intrinsic mode decomposition module is used to perform adaptive variational mode decomposition on the synchronization signal window matrix to obtain the signal intrinsic mode function matrix. The global context encoding module is used to perform global context IMF encoding on the time series of the signal eigenmode function matrix after converting it into eigenmode signal vectors to obtain the global signal eigenmode context encoded vector. The fault classification and diagnosis module is used to classify faults in the global signal intrinsic mode context encoding vector to obtain diagnostic results.

Images