Flat wire common mode inductor optimization method and system based on high frequency loss suppression

Abstract

The application provides a flat wire common mode inductance optimization method and system based on high-frequency loss suppression, and relates to the technical field of inductance optimization. The method comprises the following steps: performing multi-dimensional running state sensing collection; performing loss suppression adjustment to accumulate historical state feature sequences and loss evolution feature sequences; performing running law mining to construct a loss evolution law model; performing loss suppression adjustment to obtain real-time loss state feature data; performing law matching judgment, if the preset law triggering condition is met, performing loss trend prediction to obtain predicted loss state data; performing collaborative decision processing to obtain optimization adjustment instructions, and performing loss suppression optimization control. The application solves the technical problem that in the prior art, the flat wire common mode inductance is usually monitored for loss or adjusted for fixed parameters according to a small amount of real-time collected running parameters, the loss state is difficult to be comprehensively and accurately online represented, the actual loss change cannot be timely identified, and the accuracy of loss control is reduced.

Description

Technical Field

[0001] This invention relates to the field of inductor optimization technology, specifically to a method and system for optimizing flat wire common-mode inductors based on high-frequency loss suppression. Background Technology

[0002] Flat wire common-mode inductors are widely used in electromagnetic interference (EMI) suppression applications due to their advantages such as high conductor utilization, low parasitic parameters, and strong current carrying capacity. During high-frequency operation, the loss level of flat wire common-mode inductors is affected by various operating factors, including current, voltage, operating frequency, temperature, and load conditions. The loss state dynamically changes with operating conditions, affecting not only the energy conversion efficiency of the device but also reducing EMI suppression performance and long-term operational reliability. Therefore, continuous optimization and control of losses based on operating conditions is necessary.

[0003] Existing technologies typically monitor the loss of flat wire common-mode inductors or adjust fixed parameters based on a small number of real-time collected operating parameters. Since the operating status is affected by a variety of factors and changes rapidly, the loss status is difficult to be fully and accurately characterized online, resulting in the inability to identify actual loss changes in a timely manner and reducing the accuracy of loss control. Summary of the Invention

[0004] This application provides a method and system for optimizing flat wire common-mode inductors based on high-frequency loss suppression. It aims to solve the technical problem that existing technologies typically rely on a small number of real-time acquired operating parameters to monitor the loss of flat wire common-mode inductors or adjust fixed parameters. This makes it difficult to comprehensively and accurately characterize the loss state online, resulting in the inability to identify actual loss changes in a timely manner and reducing the accuracy of loss control.

[0005] The first aspect disclosed in this application provides a method for optimizing a flat-wire common-mode inductor based on high-frequency loss suppression. The method includes: synchronously performing multi-dimensional operational state sensing and acquisition during the operation of the flat-wire common-mode inductor to generate an online multi-source state data stream; performing loss suppression adjustment on the online multi-source state data stream based on a real-time feedback control mechanism to accumulate historical state feature sequences and loss evolution feature sequences; mining operational patterns based on the historical state feature sequences and loss evolution feature sequences to construct a loss evolution pattern model; performing loss suppression adjustment based on the incremental multi-source state data stream to obtain real-time loss state feature data; performing pattern matching judgment on the incremental multi-source state data; if the incremental multi-source state data meets a preset pattern triggering condition, inputting the incremental multi-source state data into the loss evolution pattern model for loss trend prediction to obtain predicted loss state data; performing collaborative decision processing on the real-time loss state feature data and the predicted loss state data to obtain an optimization adjustment command, and using the optimization adjustment command for loss suppression optimization control.

[0006] The second aspect of this application discloses a flat wire common-mode inductor optimization system based on high-frequency loss suppression. The system is used in the aforementioned flat wire common-mode inductor optimization method based on high-frequency loss suppression. The system includes: a state sensing module for synchronously performing multi-dimensional operational state sensing and acquisition during the operation of the flat wire common-mode inductor, generating an online multi-source state data stream; a first suppression adjustment module for performing loss suppression adjustment on the online multi-source state data stream based on a real-time feedback control mechanism to accumulate historical state feature sequences and loss evolution feature sequences; and a pattern mining module for performing operational pattern mining based on the historical state feature sequences and loss evolution feature sequences, constructing... A loss evolution law model is established; a second suppression and adjustment module is used to perform loss suppression and adjustment based on incremental multi-source state data stream to obtain real-time loss state characteristic data; a loss trend prediction module is used to perform pattern matching judgment on the incremental multi-source state data, and if the incremental multi-source state data meets the preset pattern triggering condition, the incremental multi-source state data is input into the loss evolution law model to predict the loss trend and obtain predicted loss state data; an optimization control module is used to perform collaborative decision processing on the real-time loss state characteristic data and the predicted loss state data to obtain optimization adjustment instructions, and use the optimization adjustment instructions to perform loss suppression optimization control.

[0007] One or more technical solutions provided in this application have at least the following beneficial effects: By synchronously acquiring multi-dimensional operating status information such as current, voltage, temperature, and frequency, a continuous, complete, and time-consistent online multi-source status data stream is formed. This provides a comprehensive data foundation for loss status identification, operating pattern analysis, and subsequent optimization control, improving the completeness and accuracy of operating status perception. Real-time feedback control responds promptly to current loss changes and implements dynamic adjustments, while continuously accumulating historical data on operating status and loss evolution. This ensures stable real-time system operation and provides reliable data support for subsequent pattern mining, improving the efficiency of historical information utilization. By analyzing the correlation between historical operating states and loss changes, recurring loss evolution patterns are extracted and a pattern model is established, enabling the system to systematically represent loss change trends. This approach enhances the accuracy of loss prediction under different operating conditions. It utilizes newly added operating data to continuously perform real-time loss state estimation and feedback adjustment, dynamically acquiring current loss state characteristics to achieve rapid response to changes in operating conditions, ensuring high timeliness and stability of the real-time control process. By matching the current operating state with historical patterns, trend prediction is initiated only when conditions conform to historical patterns, effectively avoiding prediction deviations caused by pattern mismatch, improving the targeting of prediction model calls, and enhancing the reliability of loss trend prediction under complex operating conditions. Furthermore, it integrates real-time feedback results with predicted trend information to form an optimized adjustment strategy, achieving synergistic complementarity between current response capabilities and forward-looking adjustment capabilities, thereby improving the response speed, adjustment accuracy, and adaptive optimization capabilities of loss suppression control under complex operating conditions.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the flat wire common-mode inductor optimization method based on high-frequency loss suppression provided in the embodiments of this application.

[0010] Figure 2 The graph shows the variation curve of the common-mode inductor loss in the flat wire common-mode inductor optimization method based on high-frequency loss suppression provided in the embodiments of this application.

[0011] Figure 3 This is a schematic diagram of a flat wire common-mode inductor optimization system based on high-frequency loss suppression, provided in an embodiment of this application.

[0012] Explanation of reference numerals in the attached diagram: State perception module 10, first suppression and adjustment module 20, pattern mining module 30, second suppression and adjustment module 40, loss trend prediction module 50, and optimization control module 60. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] Example 1, as Figure 1 As shown in the figure, this application provides a method for optimizing flat wire common-mode inductors based on high-frequency loss suppression. The method includes: During the operation of the flat wire common mode inductor, multi-dimensional operation status sensing and acquisition are performed synchronously to generate online multi-source status data streams.

[0015] After the flat-wire common-mode inductor enters high-frequency operation, various operating parameters are synchronously acquired through acquisition modules located at the current input terminal, voltage acquisition terminal, temperature monitoring position, and frequency detection unit. These data are then timestamped according to a unified time base, forming corresponding operating status data. The multi-dimensional operating status includes current amplitude, voltage change, device temperature, operating frequency, and other status parameters reflecting operating conditions. The multi-source status data stream refers to a data set continuously generated in chronological order, containing information from multiple operating dimensions. The status data of each dimension are synchronously arranged according to the acquisition sequence to form a continuously updated online multi-source status data stream, providing a unified data input basis for subsequent loss status analysis, historical pattern accumulation, and online adjustment.

[0016] The online multi-source state data stream is subjected to loss suppression adjustment based on a real-time feedback control mechanism to accumulate historical state feature sequences and loss evolution feature sequences.

[0017] The online multi-source state data stream is continuously segmented according to a preset time window to generate a multi-dimensional state feature sequence for the corresponding time period. This sequence is then input into a multi-task loss state representation network for processing to obtain the loss state features within the corresponding time window. The multi-task loss state representation network employs a shared feature extraction structure and multiple loss estimation task heads for collaborative computation, simultaneously estimating different loss representation dimensions. These estimations are then fused through correlation verification to form a unified loss state feature. The current loss state is compared with a historical state space matching model to obtain the corresponding real-time loss evolution features. Based on these evolution features, the current loss suppression strategy is adjusted to complete feedback regulation. Simultaneously, the process of changing the operating state and the corresponding loss changes are continuously saved, forming historical state feature sequences and loss evolution feature sequences in chronological order, providing continuous historical samples for subsequent pattern mining.

[0018] Based on the historical state feature sequence and loss evolution feature sequence, operational patterns are mined to construct a loss evolution pattern model.

[0019] After accumulating data over a preset period, the historical state feature sequences and loss evolution feature sequences are correlated and matched based on their time correspondence, and state change features for each time period are extracted. Historical data are aggregated according to the similarity of operating states, forming multiple state evolution windows, each corresponding to a continuous operating process. Further correlation analysis is performed between state change features and each state evolution window to filter out state change patterns with recurring characteristics and their corresponding loss evolution patterns, establishing a correspondence between the two. The loss evolution law model is a correlation description model between state changes and loss changes established based on recurring operating patterns. It is used to express the development law of loss changes under different operating states, rather than a fixed parameter model.

[0020] Loss suppression adjustment is performed based on incremental multi-source state data streams to obtain real-time loss state characteristic data.

[0021] The system continuously receives new incremental multi-source state data streams and performs real-time analysis using a data processing workflow consistent with historical phases. First, it divides the incremental data into time windows, extracts multi-dimensional state features of the current operating phase, and then inputs these features into a multi-task loss state representation network to obtain the current loss state estimate. It then combines this with a historical state space matching model to obtain the real-time loss evolution features corresponding to the current operating condition and completes loss suppression control for the current cycle based on a feedback adjustment strategy. Simultaneously, the fused loss state output is saved as real-time loss state feature data. This real-time loss state feature data characterizes the actual loss state and degree of change at the current operating moment, providing a real-time reference for subsequent pattern prediction result fusion.

[0022] The incremental multi-source state data is subjected to pattern matching and judgment. If the incremental multi-source state data meets the preset pattern triggering condition, the incremental multi-source state data is input into the loss evolution law model to predict the loss trend and obtain the predicted loss state data.

[0023] State-related features are extracted from the current incremental multi-source state data and matched with historical state change patterns stored in the loss evolution law model. The law triggering condition refers to the current state-related features and historical repeated state change patterns reaching a preset matching degree, indicating that the current operation process has a referable historical evolution law. When this condition is met, the current incremental state data is input into the loss evolution law model, and the development trend of the loss state in the future period is inferred based on the corresponding historical evolution relationship, outputting predicted loss state data. If the matching result does not meet the preset condition, law prediction is not performed, but only the real-time feedback adjustment method continues, and the current data continues to be included in historical samples for law update.

[0024] The real-time loss state characteristic data and the predicted loss state data are processed collaboratively to obtain an optimized adjustment command, which is then used for loss suppression optimization control.

[0025] Real-time loss status characteristic data and predicted loss status data are mapped to each other according to a unified state dimension, and collaborative analysis is performed by combining the degree of difference, trend of change, and consistency between the two types of data. Real-time loss status characteristic data reflects the actual loss situation under the current operating state, while predicted loss status data reflects the possible loss change trend in future operating stages. Based on the comprehensive analysis results of the two types of data, the current adjustment amount and the advance adjustment amount are determined, corresponding optimized adjustment commands are generated, and the optimized adjustment commands are sent to the flat wire common mode inductor operation control unit to dynamically adjust the operating parameters. This achieves continuous loss suppression control that combines real-time response and trend prediction, while providing new feedback data for the next cycle of operating status acquisition.

[0026] Furthermore, the method for loss suppression adjustment of the online multi-source state data stream based on a real-time feedback control mechanism includes: S1: Perform sliding window time-series processing based on a preset time window on the online multi-source state data stream to construct a multi-dimensional state feature sequence; S2: Input the multi-dimensional state feature sequence into a pre-constructed multi-task loss state representation network for loss state estimation and output a loss state feature sequence; S3: Input the loss state feature sequence into a pre-constructed historical state space matching model to retrieve real-time loss evolution features; S4: Construct a feedback adjustment strategy based on the real-time loss evolution features, perform loss suppression adjustment under real-time operating conditions, and return to steps S1 to S4 to track loss evolution under continuous operation until a preset data accumulation period is completed.

[0027] The online multi-source state data stream is continuously segmented into sliding segments according to a preset time window length. Each slide updates the window range with a fixed time step, preserving some overlapping data between adjacent windows to maintain the continuity of operational state changes. For each sliding window, the current, voltage, temperature, frequency, and other operational state parameters within the window are rearranged in a uniform time order, and time synchronization processing of data from different sampling frequencies is performed to generate a state data set for the corresponding time period. A multi-dimensional state feature sequence is constructed based on the changes of each operational parameter within the window. This multi-dimensional state feature sequence reflects the continuous changes of multiple operational parameters over time, providing unified input data for subsequent loss state estimation.

[0028] A multidimensional state feature sequence is input into a pre-constructed multi-task loss state representation network in chronological order. After data normalization by the input processing layer, it is sent to a shared feature extraction layer to extract unified features of the correlation changes between different operating states, forming shared loss feature data. This shared loss feature data is then input into multiple loss estimation task heads. Each task head generates a single-dimensional loss estimation result based on its corresponding loss representation dimension. Inconsistencies between different estimation results are then eliminated through correlation verification and fusion, forming loss state feature frames for corresponding time windows. These loss state feature sequences are then continuously output in chronological order. The loss state feature sequences are used to describe the evolution of loss states during continuous operation.

[0029] The system maps the currently obtained loss state feature sequence to a historical state space matching model for state retrieval. This model is a state index space established based on the relationship between state features and loss changes during historical operation, with each historical operating state corresponding to specific loss evolution information. The system performs matching calculations based on the similarity between the current loss state features and historical state nodes, filters historical evolution trajectories corresponding to similar operating states, and extracts loss change information highly correlated with the current operating condition to form real-time loss evolution features. These real-time loss evolution features reflect the direction and degree of loss change corresponding to the current operating state during historical operation.

[0030] Based on real-time loss evolution characteristics, the loss change trend corresponding to the current operating state is determined. Combined with the currently collected operating state parameters, a corresponding feedback adjustment strategy is generated to adjust the operating parameters of the flat wire common-mode inductor in real time, maintaining the current operating state within the preset loss control range. After completing the adjustment for this cycle, new online multi-source state data streams are received, and the sliding window processing, loss state estimation, historical state space matching, and feedback adjustment processes are re-executed to achieve loss evolution tracking during continuous operation. Simultaneously, the multi-dimensional state feature sequences, loss state feature sequences, and real-time loss evolution characteristics generated in each cycle are continuously stored in chronological order until the preset data accumulation period is reached, providing a historical data foundation for subsequent operation pattern mining and loss evolution pattern model construction.

[0031] Furthermore, it also includes: A baseline estimation task head is pre-constructed based on the loss state estimation objective. The baseline estimation task head is then trained and optimized using various sample loss label data to obtain multiple loss estimation task heads. The input of each loss estimation task head is shared loss feature data, and the output is a single-dimensional loss estimation result. A multi-task loss state representation network is constructed based on the shared feature extraction structure and the multiple loss estimation task heads. The multi-task loss state representation network includes an input processing layer, a shared feature extraction layer, and multiple loss estimation task heads configured in parallel at the output of the shared feature extraction layer.

[0032] Based on the different loss states that need to be characterized during the high-frequency operation of flat wire common-mode inductors, multiple loss state estimation targets are pre-determined, and corresponding benchmark estimation task heads are designed for each target. The loss state estimation target refers to the estimation object used to describe different manifestations of loss changes, such as overall loss level, temperature-related loss, loss caused by frequency changes, and the degree of magnetic saturation influence. Each benchmark estimation task head adopts a unified data input interface and output structure, and establishes the mapping relationship between shared features and corresponding loss estimation results through pre-defined network layers. Each benchmark estimation task head only retains the output differences for different loss dimensions, providing a unified basic structure for subsequent independent training and multi-task collaborative construction.

[0033] Sample operation data generated under different operating conditions are collected, and various sample loss label data are constructed based on the corresponding measurement results. Shared loss feature data is then input into each benchmark estimation task head for independent training. During training, each task head calculates the estimation error using only the corresponding loss label and continuously adjusts its internal parameters based on the error results, gradually establishing a mapping relationship between shared loss features and the target loss dimension. After training, multiple loss estimation task heads are formed. Each task head receives the same shared loss feature data as input but outputs the estimation result for a single loss dimension. The single-dimensional loss estimation result represents independent estimation data corresponding to a specific loss characterization target, and different task heads remain independent of each other.

[0034] Multiple trained loss estimation task heads are uniformly connected to the output of a shared feature extraction structure to construct a complete multi-task loss state representation network. The shared feature extraction structure, located in the middle of the network, is used to uniformly extract common change features from the input runtime data, avoiding multiple tasks repeatedly learning the same runtime information. The multiple loss estimation task heads each receive the shared loss feature data output by the shared feature extraction structure and generate independent estimation results based on their respective loss estimation objectives. Because each task head shares the same feature base, consistency between different loss dimensions is maintained, while redundant computation is reduced, resulting in a unified structure that combines shared feature extraction and multi-task parallel estimation.

[0035] The multi-task loss state representation network, following a data processing flow, comprises an input processing layer, a shared feature extraction layer, and multiple loss estimation task heads. The input processing layer receives the multidimensional state feature sequence generated by the sliding window and performs data format unification, feature arrangement, and numerical normalization to ensure data from different sources meets unified input requirements. The shared feature extraction layer extracts the correlation and change information between different operating parameters based on the processed multidimensional state features, generating shared loss feature data. Multiple loss estimation task heads are set in parallel at the output of the shared feature extraction layer. Each task head calculates its corresponding single-dimensional loss estimation result based on the shared loss feature data, and these multiple estimation results are used together as input for subsequent loss state correlation and fusion processing, achieving synchronous representation of multidimensional loss states.

[0036] Furthermore, the method involves inputting the multidimensional state feature sequence into a pre-constructed multi-task loss state representation network for loss state estimation, and outputting a loss state feature sequence. The first multidimensional loss feature slice is extracted sequentially from the multidimensional state feature sequence; the first multidimensional loss feature slice is input to the multi-task loss state representation network and transmitted to the shared feature extraction layer via the input processing layer; loss correlation feature extraction is performed in the shared feature extraction layer to obtain shared loss feature data, and the shared loss feature data is transmitted to the multiple loss estimation task heads respectively; loss state estimation based on differentiated loss representation targets is performed in the multiple loss estimation task heads to obtain multiple loss estimation results; correlation verification and fusion are performed on the multiple loss estimation results to output the first loss state feature frame.

[0037] The constructed multidimensional state feature sequence is divided into slices according to time sequence. Each slice corresponds to the operating state data within a continuous time window, maintaining the time correspondence between various operating parameters. The multidimensional loss feature slice refers to a data unit containing information on the changes of current, voltage, temperature, frequency, and other operating state parameters within the same time window, used to describe the operating state corresponding to the current stage of loss formation. First, the first multidimensional loss feature slice located at the beginning of the multidimensional state feature sequence is extracted as the current calculation object, and its time index information in the overall sequence is retained, providing a unified data input unit for subsequent continuous slice processing and loss state estimation.

[0038] The first multi-dimensional loss feature slice is input into the multi-task loss state representation network. It first undergoes data format conversion, feature arrangement, numerical normalization, and dimensionality unification in the input processing layer, transforming operating parameters from different sources into a unified data representation. After processing, the normalized data is transmitted to the shared feature extraction layer. This shared feature extraction layer is a processing structure in the network used to extract common variation patterns among different operating parameters. By calculating the correlations between multi-dimensional state features, it retains common feature information related to loss changes and forms the data foundation for multiple subsequent loss estimation tasks, avoiding the repeated extraction of the same operating features by different task heads.

[0039] The shared feature extraction layer continuously extracts common information reflecting the loss formation process based on the correlation and change relationships between various operating parameters in the first multi-dimensional loss feature slice, generating shared loss feature data. This shared loss feature data is a unified feature expression formed by the combined effects of multiple operating states, simultaneously reflecting the comprehensive influence of factors such as current, temperature, frequency, and voltage on loss changes. After generating the shared loss feature data, it is copied and transmitted to multiple loss estimation task heads. Each task head uses the same shared data as input but performs independent estimations based on different loss characterization targets, providing a unified feature basis for the subsequent generation of multiple loss estimation results.

[0040] Multiple loss estimation task heads receive shared loss feature data and independently complete loss state estimation based on pre-defined differentiated loss characterization targets. These differentiated loss characterization targets represent different dimensions of loss change for each task head, such as overall loss level, thermal loss variation, electromagnetic loss variation, and frequency-related loss. Each task head calculates the shared loss feature data based on its trained mapping relationship and outputs a corresponding single-dimensional loss estimation result. Each estimation result reflects the change in a specific loss dimension under the current operating state. Multiple loss estimation results collectively constitute a multi-dimensional loss state description for the current time window, providing foundational data for subsequent correlation and fusion.

[0041] After obtaining the single-dimensional loss estimation results output from multiple loss estimation task heads, consistency verification is performed based on the historical correlation between each loss dimension. Estimation results that significantly deviate from the overall trend are corrected, and fusion processing is completed by considering the correlation between the estimation results. The loss state feature frame refers to the unified state expression after fusing all loss estimation results within a single time window, which can completely reflect the comprehensive loss state of the flat wire common-mode inductor at the current moment. After generating the first loss state feature frame, it is written into the loss state feature sequence in chronological order, and the above process is repeated to extract the next multi-dimensional loss feature slice, forming a continuously updated loss state feature sequence.

[0042] Furthermore, the method involves correlating, verifying, and fusing the multiple loss estimation results to output a first loss state feature frame. Multiple sample state feature frames are generated based on the single-dimensional estimation results of multiple sets of samples. Based on these multiple sample state feature frames, state correlation analysis is performed on the single-dimensional estimation results of the multiple sets of samples to obtain multiple correlation relationships between multiple loss state characterization dimensions. Using the multiple loss state characterization dimensions as topology nodes, node connection relationships are constructed based on the multiple correlation relationships to form a loss state correlation topology. The correlation weights of the loss state correlation topology are adjusted using the single-dimensional estimation results of the multiple sets of samples and the multiple sample state feature frames. The multiple loss estimation results are mapped at the node level to the loss state correlation topology, and consistency verification and fusion processing are performed based on the topology correlation relationships to output the first loss state feature frame.

[0043] The single-dimensional estimation results of multiple sets of samples obtained during the training phase are combined according to their respective sample and time order. Each set of samples contains estimation data corresponding to multiple loss characterization dimensions. Following a unified data arrangement method, the estimation results of each loss dimension for the same sample within the same time window are integrated to form a corresponding sample state feature frame. The sample state feature frame is a data structure used to describe the comprehensive loss state of a single sample at a certain moment. It stores the estimation results of multiple loss characterization dimensions and their corresponding relationships, and retains time identifiers and sample numbers, enabling different sample states to undergo subsequent correlation analysis in a unified format, providing standard data units for establishing correlations between loss states.

[0044] Based on the data of the changes of each loss representation dimension over time in multiple sample state feature frames, statistical analysis is performed on the degree of synchronous change and the order of change between different dimensions, and the common change patterns of different loss representation dimensions in multiple samples are calculated. For loss dimensions that maintain synchronous, lagging, or reverse change relationships, their corresponding correlation forms and correlation strengths are recorded to form the correlation relationships between multiple loss state representation dimensions. These correlation relationships describe the changing connections formed between different loss representation dimensions during long-term operation, and are not limited to data correspondence at a single moment, but rather a stable correlation structure jointly established based on a large number of historical samples.

[0045] Each loss state representation dimension is defined as an independent topological node, and connections are established between corresponding nodes based on the obtained association relationships. Here, a node represents a loss state representation dimension, and a connection between nodes indicates a stable correlation between two loss dimensions; if there is no stable correlation between two loss dimensions, no connection is established. After completing all node connections, a loss state association topology is formed. This association topology uses a network structure to express the overall association relationships between multiple loss dimensions. Different nodes form a complete association network through connections, providing a unified topological foundation for subsequent loss state consistency verification.

[0046] The single-dimensional estimation results of multiple sets of samples are continuously mapped to the loss state association topology. The stability of the connections between nodes is statistically analyzed, and the association weights of each connection are adjusted based on the long-term common changes of corresponding nodes in multiple sample state feature frames. The association weight represents the degree of correlation between two loss state representation dimensions. When two nodes consistently change in most samples, the association weight of the corresponding connection is increased; if the correlation fluctuates significantly across different samples, the corresponding connection weight is decreased. After continuous adjustment, the loss state association topology gradually forms an association structure that conforms to the historical sample change patterns, improving the reference reliability in subsequent state fusion processes.

[0047] Multiple loss estimates obtained in the current time window are mapped to various topological nodes in the loss state association topology. Node-level mapping means that each loss estimate corresponds to only one loss state representation dimension node, maintaining the same node position as in the historical association topology. Based on the connection relationships and corresponding association weights between nodes, consistency checks are performed on the current node estimates. When a node deviates significantly from historical association patterns with its associated nodes, it is corrected based on the association relationships with adjacent nodes. After completing the checks on all nodes, all node estimates are fused to form a unique first loss state feature frame corresponding to the current time window. This first loss state feature frame comprehensively reflects the overall loss state after coordination and consistency among multiple loss dimensions, and can serve as data input for subsequent historical state space matching and feedback adjustment.

[0048] Furthermore, based on the historical state feature sequence and the loss evolution feature sequence, operational patterns are mined to construct a loss evolution pattern model. The method includes: After performing time-related matching on the historical state feature sequence and the loss evolution feature sequence, state change correlation features are obtained by performing time-series change feature extraction. State aggregation is then performed on the historical state feature sequence to obtain multiple state evolution window sequences for various sample operating states. Based on these multiple state evolution window sequences, window correlation matching is performed on the state change correlation features to obtain multiple state correlation feature sequences. Loss evolution pattern analysis under state constraints is then performed on these multiple state correlation feature sequences to obtain repeating state change patterns and repeating loss evolution patterns. Based on these repeating state change patterns and repeating loss evolution patterns, a loss evolution law correlation representation is constructed to obtain the loss evolution law model.

[0049] Historical state feature sequences and loss evolution feature sequences are temporally correlated and matched according to a unified time benchmark, establishing a one-to-one correspondence between operational state data and loss evolution data at the same operational moment, forming a continuous state-loss correspondence dataset. The changes, directions, and rates of change of each operational state parameter and loss state between adjacent moments are calculated in chronological order, and the synchronization and sequential relationships between state changes and loss changes are further extracted. Among these, the state change correlation feature refers to the data set describing the degree of correlation between the operational state change process and the loss evolution process. It not only preserves the changing trends of each operational parameter but also records the temporal correspondence between state changes and loss changes, providing a unified correlation basis for subsequent operational pattern analysis.

[0050] Based on the variation characteristics of each operating parameter in the historical state feature sequence, data with similar operating states are aggregated and divided into multiple state evolution windows according to continuous operating time. A state evolution window refers to a data segment that maintains similar operating state variation characteristics within a continuous operating time period; each window preserves the complete state change process within its corresponding time range. For data from different historical samples but with similar operating state variation processes, they are divided into different state evolution windows and arranged into multiple state evolution window sequences according to time order. This allows for comparative analysis of different historical operating processes using a unified data organization method, thereby reducing the impact of random fluctuations on the subsequent pattern extraction process.

[0051] State change correlation features are mapped to corresponding state evolution window sequences. Continuous matching of state change correlation features at each time point within the same state evolution window is performed, and window correspondences are established based on the similarity between state evolution windows. Correspondence features belonging to the same type of operating state window are reorganized chronologically to form corresponding state correlation feature sequences. These state correlation feature sequences are data sequences reflecting the continuous evolution of the correlation between state changes and loss changes during the evolution of the same operating state. They retain complete temporal evolution information and can describe the development process of the correlation at different operating stages, providing a continuous data foundation for subsequent loss evolution pattern analysis.

[0052] Based on multiple state-related feature sequences, statistical analysis is performed on the continuous state change processes in each sequence, and associated segments with consistent change patterns are selected according to preset state constraints. State constraints refer to the restrictions that the range, order, and duration of changes in operating state parameters meet preset consistency requirements; only data conforming to the same operating state evolution characteristics are retained for pattern analysis. Repeated state change processes and corresponding loss change processes that satisfy the state constraints are statistically analyzed. State change processes that occur repeatedly with consistent trends are defined as repetitive state change patterns, and the corresponding loss change processes are defined as repetitive loss evolution patterns, thus establishing a correspondence between stable operating states and loss evolution.

[0053] A correspondence is established between repetitive state change patterns and corresponding repetitive loss evolution patterns, and a mapping from state change to loss evolution is constructed according to the order of operational state evolution. The loss evolution pattern association representation is a data description structure formed based on recurring operational patterns from multiple historical samples. It records the loss evolution process, direction of change, duration, and associated conditions corresponding to different state change patterns. Multiple pattern association representations are uniformly organized into a loss evolution pattern model, with each state change pattern corresponding to a set of loss evolution patterns. When a state change process consistent with historical patterns occurs during subsequent online operation, the corresponding historical evolution patterns can be retrieved based on this loss evolution pattern model, providing a pattern basis for subsequent loss trend prediction, rather than relying on estimation using fixed control parameters.

[0054] Furthermore, the method for performing pattern matching and judgment on the incremental multi-source state data includes: State features are extracted from the incremental multi-source state data to obtain incremental state association features; the incremental state association features are matched with the loss evolution law association representation to calculate the state similarity; when the state similarity meets the preset matching conditions, it is determined that the law triggering condition is met, and the incremental multi-source state data is input into the loss evolution law model for loss trend prediction.

[0055] The system acquires newly collected incremental multi-source state data during real-time operation and performs time synchronization, data arrangement, and window division according to the same data processing methods as historical state characteristics to form state data fragments for the current operation phase. Based on the changing trends, magnitudes, and collaborative relationships between various operating parameters within the current time window, the system performs correlation analysis on the incremental multi-source state data, extracts data that reflects the evolutionary characteristics of the current operating state, and generates incremental state correlation features.

[0056] Incremental state correlation features are input into the loss evolution law correlation representation, and compared with each historical recurring state change pattern to calculate the degree of matching between the current operating state and historical states. During the matching process, the order of state changes, the correlation between various operating parameters, the magnitude of change, and the duration are used as comparison criteria to calculate the differences between the current state and each historical operating pattern, and to synthesize the corresponding state similarity. The state similarity represents the degree of consistency between the current operating state and historical patterns; a higher value indicates that the current operating state is closer to the historical recurring operating pattern. When multiple historical patterns meet the matching requirements, the historical pattern with the highest state similarity is selected as the matching result corresponding to the current operating state, providing a corresponding pattern basis for subsequent loss trend prediction.

[0057] The obtained state similarity is compared with preset matching conditions. When the state similarity reaches a preset threshold, the current operating state is determined to have consistent state evolution characteristics with the historical repetitive operating pattern, i.e., the pattern triggering condition is met. The loss evolution pattern model is invoked, using the current incremental multi-source state data as the current state input. Based on the historical loss evolution pattern obtained through matching, the loss change process in the current operating state over a subsequent period is deduced, generating corresponding predicted loss state data. The predicted loss state data includes the direction, magnitude, and continuous evolution process of loss change in the future operating stage, which is used for subsequent collaborative decision-making with real-time loss state feature data. If the state similarity does not reach the preset matching condition, the invocation of the loss evolution pattern model is stopped, real-time feedback adjustment continues, and the current incremental operating data is continuously accumulated into the historical state feature sequence and the loss evolution feature sequence to participate in subsequent pattern updates.

[0058] Furthermore, the method also includes: During the continuous operation and loss evolution tracking process in cyclic steps S1 to S4, historical operation data is accumulated to obtain the historical state feature sequence and the loss evolution feature sequence. The historical state feature sequence and the loss evolution feature sequence are used to extract regular features based on the time series correlation analysis method to obtain the loss evolution correlation feature set. If the loss evolution correlation feature set meets the preset regularity judgment conditions, the operation regularity is mined based on the historical state feature sequence and the loss evolution feature sequence to construct the loss evolution regularity model.

[0059] During the continuous cyclic execution of steps S1 to S4, multidimensional state feature sequences, loss state feature sequences, and real-time loss evolution features corresponding to each time window are continuously acquired according to a fixed sampling period and stored continuously in a unified time order. For each operating cycle, the current operating state, the corresponding loss state, and the loss change result after feedback adjustment are recorded, and a time correspondence between the three is established. Historical operating datasets are formed by sequentially appending data according to the operating time. Data reflecting the continuous change process of the operating state constitutes the historical state feature sequence, and data reflecting the corresponding loss change process constitutes the loss evolution feature sequence. Both sequences maintain a complete time evolution order and state correspondence, providing continuous and traceable historical sample data for subsequent operational pattern analysis.

[0060] Once the accumulated historical operational data reaches a preset scale, the historical state characteristic sequence and the loss evolution characteristic sequence are analyzed synchronously according to their temporal correspondence. Furthermore, time-series correlation analysis is performed based on the continuous correlation between state changes and loss changes at each time point. During the analysis, the recurring state change processes and their corresponding loss evolution processes in different operational stages are statistically compared. Correlation segments with consistent trends, identical order of changes, and similar durations are selected, and each correlated segment is uniformly coded and categorized. The loss evolution correlation feature set is a dataset composed of multiple state change features and loss change features with stable correspondences. It describes the typical state-loss correlation patterns formed during historical operation, providing a direct data source for subsequent pattern model construction.

[0061] The validity of patterns in the loss evolution correlation feature set is determined. When the repeated state change processes in the correlation feature set reach a preset number of occurrences, and the corresponding loss evolution processes remain stable and consistent, the preset pattern determination condition is met. Corresponding features that meet the determination condition are used as the objects for pattern extraction. Combining historical state feature sequences and loss evolution feature sequences, operational pattern mining is performed to establish a correspondence between repeated state change patterns and repeated loss evolution patterns, forming pattern correlation representations according to the state evolution sequence. Finally, all correlation representations are uniformly organized into a loss evolution pattern model, which internally stores historical loss evolution patterns corresponding to different operational states, providing a callable pattern foundation for pattern matching and loss trend prediction during subsequent online operation.

[0062] Furthermore, after incrementally updating the historical state feature sequence and loss evolution feature sequence based on a preset update cycle, regular feature extraction is performed to obtain an update evolution correlation feature set. If the update evolution correlation feature set satisfies the regularity judgment condition, the loss evolution regularity correlation representation of the loss evolution regularity model is updated based on the operational regularity mining of incremental data. If the update evolution correlation feature set does not satisfy the regularity judgment condition, the call to the loss evolution regularity model is stopped, and loss suppression adjustment is performed based on the real-time feedback control mechanism to reconstruct the loss evolution regularity model under the re-accumulation of state feature sequence and loss feature sequence.

[0063] Once the system reaches the preset update cycle, the newly formed historical state feature sequences and loss evolution feature sequences are incrementally written into the existing historical data. Sequence expansion and correspondence correction are then performed according to chronological order to maintain a consistent data organization structure between the new data and the original historical data. After the incremental update, time-series correlation analysis is re-executed on the updated historical state feature sequences and loss evolution feature sequences. The correlation between state change processes and loss change processes in the newly added operating phases is re-statistically analyzed, and correlation segments with continuous change characteristics and recurring characteristics are extracted from the new operating data to form an updated evolution correlation feature set. This updated evolution correlation feature set is a set of correlation features formed by integrating the latest operating data with the original patterns, used to reflect the development and changes of patterns under the current operating environment.

[0064] The updated evolutionary correlation feature set is compared with preset rule judgment conditions. When the repetitive state change pattern and corresponding loss evolution pattern in the newly added correlation features meet the preset stability requirements, it is determined that the newly added data has formed a new effective operating rule. Operating rule mining is performed based on the newly added operating data, establishing correlations between the newly added repetitive state change patterns and corresponding repetitive loss evolution patterns, and comparing them with the original loss evolution rule correlation representations. For newly formed correlations, they are added to the original correlation representations; for existing correlations, their correlation content and corresponding relationships are corrected and improved based on the new samples. After all correlation updates are completed, the loss evolution rule model is updated synchronously to ensure that the model continuously maintains its ability to express rules consistent with the current operating conditions.

[0065] When the updated evolutionary correlation feature set fails to meet the preset rule judgment conditions, it indicates that the state changes and loss changes in the newly added operating data have not yet formed a stable and repeatable corresponding rule, or the original rule can no longer accurately represent the current operating state. At this time, the use of the loss evolution rule model for loss trend prediction is suspended, and only the real-time feedback control mechanism is retained to perform online loss suppression adjustment for the current operating state. The operating state and loss change process are continuously tracked and recorded according to steps S1 to S4. At the same time, the historical data of the subsequently collected state feature sequence and loss evolution feature sequence are re-accumulated. When the re-accumulated data reaches the preset scale, the rule feature extraction and operating rule mining process is re-executed to establish a new loss evolution rule correlation representation. Based on this, the loss evolution rule model is reconstructed, so that the rule model can readjust to the changes in the current operating environment and operating state.

[0066] To verify the practical application effect of this invention, a flat-wire common-mode inductor in the EMI filter module of a new energy vehicle inverter was used as the test object for continuous operation verification. During the test, operating status information such as current, voltage, operating frequency, and temperature were collected, with a sampling period of 100ms. The system operated continuously for 30 minutes under different load variations. First, online status sensing, real-time feedback adjustment, and historical data accumulation were completed. When the historical data met the conditions for pattern modeling, the system determined whether the historical pattern triggering conditions were met based on the current operating status. When the triggering conditions were met, the loss evolution law model was invoked to predict future loss change trends, and the prediction results were coordinated with the real-time feedback results for adjustment. When the triggering conditions were not met, only real-time feedback control was used. Test results for some typical operating moments are shown in Table 1. Table 1: Operational Data

[0067] Drawing from Table 1 Figure 2 As can be seen, as the running time increases, after the regularity model is established, the predicted loss and the actual loss maintain a high degree of consistency, and adjustments can be made in advance to gradually reduce the magnitude of loss changes, which verifies that the present invention has good loss prediction ability and dynamic optimization ability.

[0068] Example 2, based on the same inventive concept as the flat wire common-mode inductor optimization method based on high-frequency loss suppression in the previous examples, such as... Figure 3 As shown in the figure, this application provides a flat wire common-mode inductor optimization system based on high-frequency loss suppression, the system comprising: The system comprises the following modules: a state perception module 10, which synchronously performs multi-dimensional operational state perception and acquisition during the operation of the flat wire common-mode inductor, generating an online multi-source state data stream; a first suppression and adjustment module 20, which performs loss suppression and adjustment on the online multi-source state data stream based on a real-time feedback control mechanism to accumulate historical state feature sequences and loss evolution feature sequences; a pattern mining module 30, which performs operational pattern mining based on the historical state feature sequences and loss evolution feature sequences to construct a loss evolution pattern model; a second suppression and adjustment module 40, which performs loss suppression and adjustment based on the incremental multi-source state data stream to obtain real-time loss state feature data; a loss trend prediction module 50, which performs pattern matching judgment on the incremental multi-source state data, and if the incremental multi-source state data meets a preset pattern triggering condition, inputs the incremental multi-source state data into the loss evolution pattern model for loss trend prediction to obtain predicted loss state data; and an optimization control module 60, which performs collaborative decision processing on the real-time loss state feature data and the predicted loss state data to obtain optimization adjustment instructions, and uses the optimization adjustment instructions for loss suppression optimization control.

[0069] Furthermore, the first suppression adjustment module 20 is used to perform the following operation steps: S1: Perform sliding window time-series processing based on a preset time window on the online multi-source state data stream to construct a multi-dimensional state feature sequence; S2: Input the multi-dimensional state feature sequence into a pre-constructed multi-task loss state representation network for loss state estimation and output a loss state feature sequence; S3: Input the loss state feature sequence into a pre-constructed historical state space matching model to retrieve real-time loss evolution features; S4: Construct a feedback adjustment strategy based on the real-time loss evolution features, perform loss suppression adjustment under real-time operating conditions, and return to steps S1 to S4 to track loss evolution under continuous operation until a preset data accumulation period is completed.

[0070] Furthermore, the first suppression adjustment module 20 is used to perform the following operation steps: A baseline estimation task head is pre-constructed based on the loss state estimation objective. The baseline estimation task head is then trained and optimized using various sample loss label data to obtain multiple loss estimation task heads. The input of each loss estimation task head is shared loss feature data, and the output is a single-dimensional loss estimation result. A multi-task loss state representation network is constructed based on the shared feature extraction structure and the multiple loss estimation task heads. The multi-task loss state representation network includes an input processing layer, a shared feature extraction layer, and multiple loss estimation task heads configured in parallel at the output of the shared feature extraction layer.

[0071] Furthermore, the first suppression adjustment module 20 is used to perform the following operation steps: The first multidimensional loss feature slice is extracted sequentially from the multidimensional state feature sequence; the first multidimensional loss feature slice is input to the multi-task loss state representation network and transmitted to the shared feature extraction layer via the input processing layer; loss correlation feature extraction is performed in the shared feature extraction layer to obtain shared loss feature data, and the shared loss feature data is transmitted to the multiple loss estimation task heads respectively; loss state estimation based on differentiated loss representation targets is performed in the multiple loss estimation task heads to obtain multiple loss estimation results; correlation verification and fusion are performed on the multiple loss estimation results to output the first loss state feature frame.

[0072] Furthermore, the first suppression adjustment module 20 is used to perform the following operation steps: Multiple sample state feature frames are generated based on the single-dimensional estimation results of multiple sets of samples. Based on these multiple sample state feature frames, state correlation analysis is performed on the single-dimensional estimation results of the multiple sets of samples to obtain multiple correlation relationships between multiple loss state characterization dimensions. Using the multiple loss state characterization dimensions as topology nodes, node connection relationships are constructed based on the multiple correlation relationships to form a loss state correlation topology. The correlation weights of the loss state correlation topology are adjusted using the single-dimensional estimation results of the multiple sets of samples and the multiple sample state feature frames. The multiple loss estimation results are mapped at the node level to the loss state correlation topology, and consistency verification and fusion processing are performed based on the topology correlation relationships to output the first loss state feature frame.

[0073] Furthermore, the pattern mining module 30 is used to perform the following operation steps: After performing time-related matching on the historical state feature sequence and the loss evolution feature sequence, state change correlation features are obtained by performing time-series change feature extraction. State aggregation is then performed on the historical state feature sequence to obtain multiple state evolution window sequences for various sample operating states. Based on these multiple state evolution window sequences, window correlation matching is performed on the state change correlation features to obtain multiple state correlation feature sequences. Loss evolution pattern analysis under state constraints is then performed on these multiple state correlation feature sequences to obtain repeating state change patterns and repeating loss evolution patterns. Based on these repeating state change patterns and repeating loss evolution patterns, a loss evolution law correlation representation is constructed to obtain the loss evolution law model.

[0074] Furthermore, the loss trend prediction module 50 is used to perform the following operation steps: State features are extracted from the incremental multi-source state data to obtain incremental state association features; the incremental state association features are matched with the loss evolution law association representation to calculate the state similarity; when the state similarity meets the preset matching conditions, it is determined that the law triggering condition is met, and the incremental multi-source state data is input into the loss evolution law model for loss trend prediction.

[0075] Furthermore, the first suppression adjustment module 20 is used to perform the following operation steps: During the continuous operation and loss evolution tracking process in cyclic steps S1 to S4, historical operation data is accumulated to obtain the historical state feature sequence and the loss evolution feature sequence. The historical state feature sequence and the loss evolution feature sequence are used to extract regular features based on the time series correlation analysis method to obtain the loss evolution correlation feature set. If the loss evolution correlation feature set meets the preset regularity judgment conditions, the operation regularity is mined based on the historical state feature sequence and the loss evolution feature sequence to construct the loss evolution regularity model.

[0076] Furthermore, after incrementally updating the historical state feature sequence and loss evolution feature sequence based on a preset update cycle, regular feature extraction is performed to obtain an update evolution correlation feature set. If the update evolution correlation feature set satisfies the regularity judgment condition, the loss evolution regularity correlation representation of the loss evolution regularity model is updated based on the operational regularity mining of incremental data. If the update evolution correlation feature set does not satisfy the regularity judgment condition, the call to the loss evolution regularity model is stopped, and loss suppression adjustment is performed based on the real-time feedback control mechanism to reconstruct the loss evolution regularity model under the re-accumulation of state feature sequence and loss feature sequence.

[0077] Through the foregoing detailed description of the flat wire common-mode inductor optimization method based on high-frequency loss suppression, those skilled in the art can clearly understand the flat wire common-mode inductor optimization system based on high-frequency loss suppression in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for optimizing flat wire common-mode inductors based on high-frequency loss suppression, characterized in that, The method includes: During the operation of the flat wire common mode inductor, multi-dimensional operation status sensing and acquisition are performed synchronously to generate online multi-source status data streams; The online multi-source state data stream is subjected to loss suppression adjustment based on a real-time feedback control mechanism to accumulate historical state feature sequences and loss evolution feature sequences; Based on the historical state feature sequence and loss evolution feature sequence, operational patterns are mined to construct a loss evolution pattern model; Loss suppression adjustment is performed based on incremental multi-source state data stream to obtain real-time loss state characteristic data; The incremental multi-source state data is subjected to pattern matching and judgment. If the incremental multi-source state data meets the preset pattern triggering condition, the incremental multi-source state data is input into the loss evolution law model to predict the loss trend and obtain the predicted loss state data. The real-time loss state characteristic data and the predicted loss state data are processed collaboratively to obtain an optimized adjustment command, which is then used for loss suppression optimization control.

2. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 1, characterized in that, The method for loss suppression regulation of the online multi-source state data stream based on a real-time feedback control mechanism includes: S1: Perform sliding window timing processing based on a preset time window on the online multi-source state data stream to construct a multi-dimensional state feature sequence; S2: Input the multidimensional state feature sequence into the pre-constructed multi-task loss state representation network to estimate the loss state and output the loss state feature sequence. S3: Input the loss state feature sequence into the pre-constructed historical state space matching model to retrieve real-time loss evolution features; S4: Based on the real-time loss evolution characteristics, construct a feedback adjustment strategy, perform loss suppression adjustment under real-time operating conditions, and then return to steps S1 to S4 to track loss evolution under continuous operation until the preset data accumulation period is completed.

3. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 2, characterized in that, Also includes: A baseline estimation task header is pre-constructed based on the loss state estimation objective; The baseline estimation task head is trained and optimized using multiple sample loss label data to obtain multiple loss estimation task heads. The input of each loss estimation task head is shared loss feature data, and the output is a single-dimensional loss estimation result. The multi-task loss state representation network is constructed based on the shared feature extraction structure and the multiple loss estimation task heads; The multi-task loss state characterization network includes an input processing layer, a shared feature extraction layer, and multiple loss estimation task heads configured in parallel at the output of the shared feature extraction layer.

4. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 3, characterized in that, The method involves inputting the multidimensional state feature sequence into a pre-constructed multi-task loss state representation network for loss state estimation and outputting a loss state feature sequence. The first multidimensional loss feature slice is extracted sequentially from the multidimensional state feature sequence; The first multidimensional loss feature slice is input into the multi-task loss state representation network and transmitted to the shared feature extraction layer via the input processing layer. After extracting loss-related features in the shared feature extraction layer to obtain shared loss feature data, the shared loss feature data is transmitted to the multiple loss estimation task headers respectively. The loss state estimation based on the differentiated loss characterization target is performed in the multiple loss estimation task heads to obtain multiple loss estimation results; The multiple loss estimation results are correlated, verified and fused to output the first loss state feature frame.

5. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 4, characterized in that, The method involves correlating, verifying, and fusing the multiple loss estimation results to output a first loss state feature frame. Multiple sample state feature frames are generated based on the single-dimensional estimation results of multiple sets of samples. Based on the multiple sample state feature frames, state correlation analysis is performed on the single-dimensional estimation results of the multiple sets of samples to obtain multiple correlation relationships between multiple loss state characterization dimensions. Using the multiple loss state characterization dimensions as topology nodes, and constructing node connection relationships based on the multiple association relationships, a loss state association topology is formed. The association weights of the loss state association topology are adjusted using the single-dimensional estimation results of the multiple sets of samples and the state feature frames of multiple samples. The multiple loss estimation results are mapped at the node level to the loss state association topology, and consistency verification and fusion processing are performed based on the topology association relationship to output the first loss state feature frame.

6. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 1, characterized in that, Based on the historical state feature sequence and loss evolution feature sequence, operational patterns are mined to construct a loss evolution pattern model. The method includes: After performing time-related matching on the historical state feature sequence and the loss evolution feature sequence, state change correlation features are obtained by performing time-series change feature extraction. The historical state feature sequence is aggregated to obtain multiple state evolution window sequences of various sample running states; Based on the multiple state evolution window sequences, window association matching is performed on the state change association features to obtain multiple state association feature sequences; The loss evolution pattern under state constraints is analyzed on the multiple state-related feature sequences to obtain the repeated state change pattern and the repeated loss evolution pattern. Based on the repeated state change pattern and repeated loss evolution pattern, a correlation characterization of loss evolution law is constructed to obtain the loss evolution law model.

7. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 6, characterized in that, The method for performing pattern matching and judgment on the incremental multi-source state data includes: State features are extracted from the incremental multi-source state data to obtain incremental state association features; Match the incremental state association features with the loss evolution law association representation, and calculate the state similarity; When the state similarity meets the preset matching conditions, it is determined that the regularity triggering condition is met, and the incremental multi-source state data is input into the loss evolution law model for loss trend prediction.

8. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 2, characterized in that, The method further includes: During the loss evolution tracking process in continuous operation state in cyclic steps S1 to S4, historical operation data is accumulated to obtain the historical state feature sequence and the loss evolution feature sequence. Based on the time-series correlation analysis method, regular features are extracted from the historical state feature sequence and the loss evolution feature sequence to obtain the loss evolution correlation feature set; If the loss evolution correlation feature set meets the preset rule judgment conditions, then the operation rule is mined based on the historical state feature sequence and the loss evolution feature sequence to construct the loss evolution rule model.

9. The method for optimizing flat wire common-mode inductors based on high-frequency loss suppression as described in claim 8, characterized in that, After incrementally updating the historical state feature sequence and loss evolution feature sequence based on a preset update cycle, regular feature extraction is performed to obtain the update evolution correlation feature set. If the updated evolution correlation feature set satisfies the rule determination condition, then based on the operation rule mining of incremental data, the loss evolution rule correlation representation of the loss evolution rule model is updated. If the updated evolutionary correlation feature set does not meet the rule determination condition, the call to the loss evolution rule model is stopped, and loss suppression adjustment based on the real-time feedback control mechanism is used to reconstruct the loss evolution rule model under the re-accumulation of state feature sequence and loss feature sequence.

10. A flat wire common-mode inductor optimization system based on high-frequency loss suppression, characterized in that, For implementing the flat wire common-mode inductor optimization method based on high-frequency loss suppression as described in any one of claims 1 to 9, the system comprises: The state awareness module is used to synchronously perform multi-dimensional operation state awareness acquisition during the operation of the flat wire common mode inductor and generate online multi-source state data stream; The first suppression and regulation module is used to perform loss suppression and regulation on the online multi-source state data stream based on a real-time feedback control mechanism, so as to accumulate historical state feature sequences and loss evolution feature sequences. The pattern mining module is used to mine operational patterns based on the historical state feature sequence and the loss evolution feature sequence, and to construct a loss evolution pattern model. The second suppression and adjustment module is used to perform loss suppression and adjustment based on the incremental multi-source state data stream to obtain real-time loss state characteristic data. The loss trend prediction module is used to perform pattern matching and judgment on the incremental multi-source state data. If the incremental multi-source state data meets the preset pattern triggering condition, the incremental multi-source state data is input into the loss evolution law model to predict the loss trend and obtain the predicted loss state data. The optimization control module is used to perform collaborative decision processing on the real-time loss state characteristic data and the predicted loss state data to obtain optimization adjustment instructions, and to use the optimization adjustment instructions to perform loss suppression optimization control.

Images