Power supply detection method, device, equipment and medium based on sparse hybrid expert model
By employing a power supply detection method based on a sparse hybrid expert model, and training the power supply detection model using multiple expert networks, the problem of detection bias in UPS power supply detection is solved, and higher accuracy in predicting power supply health status and remaining service life is achieved.
Patent Information
- Application Number
- CN202611241573.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, deep learning models for UPS power supplies have difficulty identifying the internal degradation physical characteristics, battery capacity scale, and attenuation rate differences of different power supplies, resulting in discrepancies between the detection results and reality. Furthermore, a single model cannot adapt to the multi-stage degradation process of the power supply, leading to low detection accuracy.
A sparse hybrid expert model is adopted. By acquiring historical operating data and capacity degradation data of the power supply, a multi-dimensional input sequence is constructed. Multiple expert networks are used to train the model and generate a sparse hybrid expert model, which is used to determine the remaining lifetime and health status of the power supply.
It effectively avoids the detection bias of a single deep learning model, can adapt to the multi-stage degradation process of the power supply, and improves the detection accuracy and precision.
Smart Images

Figure CN123069264A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of power supply detection, and in particular to a power supply detection method, apparatus, equipment, and medium based on a sparse hybrid expert model. Background Technology
[0002] Switches, communication nodes, and related control equipment in power distribution communication networks typically require uninterruptible power supplies (UPS) to ensure continuous operation of communication links and dispatch control services under conditions of abnormal mains power, short-term power outages, or power fluctuations. The lithium-ion batteries in UPS power supplies gradually age under conditions such as long-term float charging, periodic charging and discharging, temperature changes, and load fluctuations, resulting in a degradation in their capacity, internal resistance, and discharge capability over time.
[0003] Therefore, it is necessary to test the UPS power supply batteries to determine their State of Health (SOH) and Remaining Life (RUL) so that technicians can perform maintenance, replacement of spare parts, and risk warnings. One common testing method is to pre-collect the voltage, current, and temperature parameters of the power supply, use deep learning (such as CNN, RNN, or standard Transformer) on these parameters to obtain a power supply testing model, and then call the power supply testing model to perform testing to determine the power supply's State of Health (SOH) and Remaining Life (RUL).
[0004] However, the above method has the following technical problems: different power supplies have different internal degradation physical characteristics, battery capacity scale and decay rate. Various differences may cause distribution drift. Deep learning models trained only by voltage, current and temperature parameters are difficult to identify various differences, resulting in deviations between the detection results and the actual situation. Moreover, a single deep learning model cannot adapt to the multi-stage degradation process of the power supply, further increasing the detection bias. Summary of the Invention
[0005] This invention provides a power supply detection method, apparatus, device, and medium based on a sparse hybrid expert model, which can solve one or more technical problems existing in the prior art.
[0006] A first aspect of this invention provides a power source detection method based on a sparse hybrid expert model, the method comprising: Acquire historical operating data and capacity degradation data of the power supply separately; The incremental capacity characteristics are determined based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process; A multidimensional input sequence is constructed using the incremental capacity features and the historical operating data, wherein the multidimensional input sequence includes fused input features, historical input sequences, and predicted output sequences; The preset expert architecture is trained based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks. The sparse hybrid expert model is invoked to determine the power supply's detection parameters based on the power supply's real-time operating data. The detection parameters include: remaining lifetime value and health status percentage value.
[0007] This invention uses a model containing multiple expert networks for power supply detection, which avoids the problem that a single deep learning model is unable to identify the internal degradation physical characteristics, battery capacity scale, and decay rate differences of different power supplies, resulting in deviations between the detection results and reality. Moreover, multiple expert networks can also adapt to the multi-stage degradation process of the power supply, further improving the accuracy of detection.
[0008] In conjunction with the first aspect, in one implementation, determining the incremental capacity characteristics based on the historical operating data and the capacity degradation data includes: Constant current charging data is obtained from the historical operating data, wherein the constant current charging data is the voltage sequence and capacity sequence of the power supply during the constant current charging stage; An interpolated capacity curve is constructed using the constant current charging data, and the interpolated capacity curve is converted into a fixed-dimensional incremental capacity feature vector. The interpolated capacity curve is a curve obtained by constructing a curve using the constant current charging data and performing interpolation processing. The incremental capacity feature vector is subjected to Gaussian smoothing filtering to obtain a filtered feature vector, and the filtered feature vector is divided into training parameter samples according to the capacity degradation data to obtain incremental capacity features.
[0009] In conjunction with the first aspect, in one implementation, the construction of a multidimensional input sequence using the incremental capacity feature and the historical operating data includes: Operational scalar feature data is extracted from the historical operation data, wherein the operational scalar feature data includes: average voltage, average current, average temperature, and cumulative cycle count; The scalar feature data of operation and the incremental capacity feature are concatenated to obtain the fused input feature; The fused input features are stacked and concatenated with feature vectors from several historical time steps in chronological order to obtain a historical input sequence; Starting with the fused input features, the output target of the model is constructed based on the preset prediction step size to obtain the predicted output sequence.
[0010] In conjunction with the first aspect, in one implementation, the step of training a preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model includes: The multidimensional input sequence is subjected to reversible instance normalization to obtain normalized temporal features; The model is trained on the preset expert architecture using the training parameter samples of the normalized temporal features and the incremental capacity features to obtain a sparse hybrid expert model.
[0011] In conjunction with the first aspect, in one implementation, the reversible instance normalization process performed on the multidimensional input sequence to obtain normalized temporal features includes: The time series parameters are calculated using the historical input sequence, wherein the time series parameters include: time series mean and time series standard deviation; The time-series parameters are used to normalize the fused input features and the predicted output sequence to obtain normalized fused features and normalized predicted sequences, respectively. The normalized fusion features and the normalized prediction sequence are respectively subjected to reversible instance normalization processing to obtain normalized time-series features.
[0012] In conjunction with the first aspect, in one implementation, the preset expert architecture includes: a multi-resolution prediction head, multiple different expert networks, and a router; The router is used to determine the selection probability of multiple different expert networks based on the hidden state, which is generated based on the normalized time series features. The multi-resolution prediction head includes a short-term prediction branch and a long-term trend prediction branch.
[0013] In conjunction with the first aspect, in one implementation, after the step of calling the sparse hybrid expert model to determine the power supply's detection parameters based on the power supply's real-time operating data, the method further includes: If the percentage value of the health status is less than a preset health threshold or the remaining lifespan value is less than a preset power maintenance value, an early warning message will be generated. or; Multi-level early warning processing is performed based on the percentage value of health status and the value of remaining lifespan.
[0014] A second aspect of this invention provides a power supply detection device based on a sparse hybrid expert model, the device comprising: The acquisition module is used to acquire historical operating data and capacity degradation data of the power supply, respectively. The determination module is used to determine the incremental capacity characteristics based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process; A construction module is used to construct a multidimensional input sequence using the incremental capacity features and the historical operating data, wherein the multidimensional input sequence includes fused input features, historical input sequences, and predicted output sequences; The training module is used to train the preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks. The detection module is used to call the sparse hybrid expert model to determine the detection parameters of the power supply based on the real-time operating data of the power supply. The detection parameters include: remaining lifetime value and health status percentage value.
[0015] Compared to existing technologies, the power supply detection method, apparatus, device, and medium provided in this invention, based on a sparse hybrid expert model, offers the following advantages: This invention can acquire historical operating data and capacity degradation data of the power supply separately; determine incremental capacity features based on the historical operating data and capacity degradation data; construct a multi-dimensional input sequence using the incremental capacity features and historical operating data; train a preset expert architecture based on the multi-dimensional input sequence and incremental capacity features to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks; and call the sparse hybrid expert model to determine the remaining lifetime value and health status percentage value of the power supply based on the real-time operating data of the power supply. This invention uses a model containing multiple expert networks for power supply detection, which avoids the problem of deviations between detection results and reality caused by a single deep learning model's inability to identify the internal degradation physical characteristics, battery capacity scale, and decay rate differences of different power supplies. Furthermore, multiple expert networks can adapt to the multi-stage degradation process of the power supply, further improving the detection accuracy. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating a power supply detection method based on a sparse hybrid expert model according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a power supply detection device based on a sparse hybrid expert model provided in an embodiment of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Switches, communication nodes, and related control equipment in power distribution communication networks typically require uninterruptible power supplies (UPS) to ensure continuous operation of communication links and dispatch control services under conditions of abnormal mains power, short-term power outages, or power fluctuations. The lithium-ion batteries in UPS power supplies gradually age under conditions such as long-term float charging, periodic charging and discharging, temperature changes, and load fluctuations, resulting in a degradation in their capacity, internal resistance, and discharge capability over time.
[0019] Therefore, it is necessary to test the UPS power supply batteries to determine their State of Health (SOH) and Remaining Life (RUL) so that technicians can perform maintenance, replacement of spare parts, and risk warnings. One common testing method is to pre-collect the voltage, current, and temperature parameters of the power supply, use deep learning (such as CNN, RNN, or standard Transformer) on these parameters to obtain a power supply testing model, and then call the power supply testing model to perform testing to determine the power supply's State of Health (SOH) and Remaining Life (RUL).
[0020] However, the above method has the following technical problems: different power sources have different internal degradation physical characteristics, battery capacity scale and decay rate. Various differences may cause distribution drift. Deep learning models trained only by voltage, current and temperature parameters are difficult to identify various differences, resulting in deviations between the detection results and the actual situation. Moreover, a single deep learning model cannot adapt to the multi-stage degradation process of the power source (such as early linear decay, mid-term capacity regeneration and late-term rapid decay), which further increases the detection bias.
[0021] To address the aforementioned issues, the following specific embodiments will provide a detailed description and explanation of a power supply detection method, apparatus, device, and medium based on a sparse hybrid expert model provided in this application.
[0022] To address the technical problems of biased detection and low detection accuracy in existing technologies that rely on a single deep learning model, this paper refers to... Figure 1 The diagram shows a schematic flowchart of a power supply detection method based on a sparse hybrid expert model according to an embodiment of the present invention.
[0023] As an example, the power source detection method based on a sparse hybrid expert model may include: S11. Acquire the historical operating data and capacity degradation data of the power supply respectively.
[0024] In one embodiment, historical operating data and capacity degradation data of the UPS power supply battery can be collected.
[0025] Specifically, data such as voltage, current, temperature, cycle count, battery capacity, and corresponding timestamps are collected from the UPS power supply battery during the charging and discharging process. This data can originate from an actual UPS power supply monitoring system, publicly available battery degradation datasets, or laboratory battery testing platforms.
[0026] In one embodiment, the capacity degradation data is a time-series measured dataset that characterizes the capacity decay of lithium batteries during charge-discharge cycles and float charging aging. It is the core foundational data for extracting IC curves and generating SOH / RUL real labels.
[0027] Specifically, the capacity degradation data includes: the true value of the base capacity time series, the full life cycle aging time series markers, and the basic data of the corresponding charging segments. Among them, the true value of the base capacity time series includes: the measured usable full-charge capacity Q of the battery under each charge-discharge cycle / verification test. t The rated capacity of the matching battery is Qᵣ. ate d, used to calculate the true value of SOH (e.g., SOH t =Q t / Qᵣ ate (d×100%). The full lifecycle aging timeline includes: cumulative charge / discharge cycles c. t The critical capacity / critical state of harmlessness (SOH) threshold at which the battery reaches end-of-life (EOL); the complete lifetime timeline from the start of data collection to battery failure (for labeling with the actual RUL tag). Supporting basic charging segment data (for calculating the IC incremental capacity curve) includes: the time-series voltage V sampled during the constant current charging phase. k , capacity Q k The sampling point sequence, through ΔQ / ΔV, calculates the IC (V) incremental capacity curve, which is a derivative basic material of capacity degradation data.
[0028] The distinction between capacity degradation data and historical operational data is as follows: Historical operating data: Real-time operating condition scalars (instantaneous voltage, current, temperature, real-time cycle count); Capacity degradation data: Measured full-charge capacity time-series data obtained from periodic charge-discharge tests. This data is specifically used to quantify the long-term battery degradation trend, generate ground truth labels for model training, and calculate the incremental capacity IC physical characteristics. It serves as a quantitative benchmark for battery aging. Capacity degradation data is either labeled or ground truth data; it does not provide raw time-series sampling data for constant current charging.
[0029] When used for model training, publicly available battery datasets can serve as the training data source in the embodiments. In practical applications, data collected online by the UPS power supply of the power distribution communication network switch can be used as model input.
[0030] S12. Determine the incremental capacity characteristics based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process.
[0031] In one embodiment, incremental capacity characteristics can be determined based on historical operating data and capacity degradation data, wherein the incremental capacity characteristics are the changes in the incremental capacity of the power supply during the charging process.
[0032] Since historical operating data and capacity degradation data are collected from the power monitoring system, data during the charging process can be found in this data, and the characteristics of incremental capacity changes can be extracted to obtain incremental capacity characteristics.
[0033] In an optional embodiment, determining the incremental capacity characteristics based on the historical operating data and the capacity degradation data may include the following sub-steps: S121. Obtain constant current charging data from the historical operating data, wherein the constant current charging data is the voltage sequence and capacity sequence of the power supply during the constant current charging stage.
[0034] S122. Construct an interpolated capacity curve using the constant current charging data, and convert the interpolated capacity curve into a fixed-dimensional incremental capacity feature vector, wherein the interpolated capacity curve is a curve obtained by constructing a curve using the constant current charging data and performing interpolation processing.
[0035] S123. Perform Gaussian smoothing filtering on the incremental capacity feature vector to obtain a filtered feature vector, and divide the filtered feature vector into training parameter samples according to the capacity degradation data to obtain incremental capacity features.
[0036] In one embodiment, data from the constant current charging phase of the battery can be selected from historical operating data to obtain the voltage and capacity sequences during the charging process. Incremental capacity curves are calculated based on the capacity and voltage increments. Then, the incremental capacity curves corresponding to different charge-discharge cycles are interpolated to a unified voltage grid so that the incremental capacity characteristics under different cycles have the same dimension. The interpolated incremental capacity curves are smoothed and filtered to reduce the impact of sampling quantization errors and sensor noise on the curve peaks. Finally, the processed incremental capacity curves are correlated with the capacity, SOH, or lifetime labels of the corresponding cycles to form training samples, validation samples, and test samples.
[0037] Specifically, data from the constant current charging phase of the battery is prioritized for incremental capacity analysis. The current during constant current charging is relatively stable, and the relationship between voltage and capacity can effectively reflect the internal aging state of the battery. The incremental capacity curve is calculated based on the capacity and voltage increments, as shown in the following formula: ; In the above formula, This represents the incremental capacity value corresponding to voltage V. , These represent the measured capacity and measured voltage at the k-th sampling point during the constant current charging phase of the battery. , These represent the capacity difference and voltage difference between two adjacent sampling points, respectively.
[0038] When the UPS power supply switches from float charging to charging (such as after a power grid anomaly recovery or after a periodic discharge test), local data of this constant current charging stage is extracted (e.g., a charging segment with voltage between 3.0V and 4.0V) to obtain the voltage and capacity sequences during the local charging process; even if the sequence is an incomplete local segment, it can still be used for subsequent feature extraction.
[0039] Because the charging voltage ranges of different batteries or different cycles may vary, the original incremental capacity curve is interpolated to a uniform voltage grid to ensure consistent model input dimensions. For example, interpolation can be performed at fixed voltage intervals within a preset voltage range to obtain an interpolated capacity curve. Then, the interpolated capacity curve can be converted into a fixed-dimensional incremental capacity feature vector.
[0040] In actual UPS operation in power distribution and communication networks, deep charge and discharge from 0% to 100% is extremely rare. However, UPS systems face two common operating conditions: first, shallow discharge and subsequent recovery charging caused by short-term power grid fluctuations or outages; and second, battery verification discharge tests performed periodically (e.g., quarterly) by the BMS system and subsequent recharging. These processes generate partial or fragmented constant current charging data.
[0041] The method of this invention is designed to address this pain point by incorporating an "interpolation to a uniform voltage grid" step. This step does not require a complete charging curve, but rather extracts a local constant current charging segment (e.g., a specific voltage window from 3.0V to 4.0V), aligns it, and thereby extracts the local incremental capacity characteristics.
[0042] Specifically, the incremental capacity feature vector with fixed dimensions can be represented by the following formula: .
[0043] In the above formula, It was set up to unify dimensions. Equally spaced voltage grid nodes.
[0044] Subsequently, the incremental capacity feature vector can be Gaussian smoothed to reduce the impact of sampling quantization errors, measurement noise, and outliers on the curve peak.
[0045] The filtered feature vector can be represented by the following formula: ; In the above formula, The standard deviation is A one-dimensional Gaussian smoothing filter function is used to filter out quantization noise.
[0046] The peak position, peak amplitude, and peak shape changes in the above curves can reflect degradation phenomena such as loss of active materials, lithium inventory loss, and impedance changes within the battery. Therefore, incorporating these curves as physical degradation characteristics into the model can improve the interpretability of model predictions and its generalization ability across batteries.
[0047] After extracting the incremental capacity curve, the incremental capacity features corresponding to each cycle or time window are mapped to capacity labels, SOH labels, or lifetime labels, and training samples, validation samples, and test samples are divided in chronological order.
[0048] This invention introduces incremental capacity curves as physical degradation characteristics into the UPS power supply life prediction process. Compared with methods that only use raw data such as voltage, current and temperature, it can more effectively characterize the internal aging state of the battery and improve the interpretability of SOH estimation and RUL prediction.
[0049] S13. Construct a multidimensional input sequence using the incremental capacity features and the historical operating data, wherein the multidimensional input sequence includes fused input features, historical input sequences, and predicted output sequences.
[0050] In one embodiment, voltage, current, temperature, and cycle count for each cycle or sampling period can be extracted from historical operating data, and capacity estimates can be extracted from capacity degradation data. A multidimensional input sequence is constructed using these parameters and a filtered feature vector of incremental capacity features. This multidimensional input sequence includes fused input features, historical input sequences, and a predicted output sequence.
[0051] In one embodiment, constructing a multidimensional input sequence using the incremental capacity feature and the historical operating data may include the following sub-steps: S131. Extract operational scalar feature data from the historical operational data, wherein the operational scalar feature data includes: average voltage, average current, average temperature, and cumulative cycle count.
[0052] S132. The scalar feature data of operation and the incremental capacity feature are spliced together to obtain the fused input feature.
[0053] S133. The fused input features are stacked and concatenated with several historical time step feature vectors in chronological order to obtain a historical input sequence.
[0054] S134. Starting from the fused input features, construct the output target of the model based on the preset prediction step size to obtain the predicted output sequence.
[0055] In one operating mode, voltage, current, temperature, and cumulative cycle count for each cycle or sampling period can be extracted from historical operating data, along with a capacity estimate. The capacity estimate can be calculated by the BMS based on the time-series integration of real-time voltage and current, and used as input features for the model. Capacity degradation data (labeled ground truth) verifies the true capacity and true state of equilibrium (SOH) obtained from discharge tests; it serves as sample labels and is used for supervised learning during training, not for input feature construction. For each cycle or sampling window, operating state features such as voltage, current, temperature, cycle count, and capacity estimate are concatenated and fused with the incremental capacity curve features.
[0056] Specifically, the above-mentioned operating status characteristics are spliced with the incremental capacity curve characteristics to obtain fused characteristics that include electrochemical degradation information and external operating information, thus obtaining fused input characteristics.
[0057] Arrange the fused input features and historical fused input features in chronological order to obtain a historical input sequence of length L; Next, the prediction step size H can be set according to the prediction requirements, so that the model can output the capacity or SOH prediction sequence for the next H time steps based on the historical input sequence of length L.
[0058] Specifically, in the t-th sampling window, the average voltage within that window is extracted. Average current Average temperature and cumulative number of cycles The local IC curve corresponding to the sampling window is interpolated to M fixed voltage nodes (e.g., one node every 20mV, truncating from 3.5V to 4V, then M=25), resulting in a high-dimensional vector of dimension M. The above four scalars are directly concatenated with the M-dimensional IC vector along the feature dimension to obtain... Following the chronological order, the feature vectors of the current time t and its L-1 past historical times are taken and stacked along the time step dimension to form a shape of size . The two-dimensional input matrix. The concatenated matrix. Perform subsequent operations such as reversible instance normalization.
[0059] In one specific implementation, scalar operating state features such as voltage, current, and temperature are concatenated with the incremental capacity curve vector after interpolation and smoothing to form multidimensional input features. Arranged in chronological order, a historical input sequence is obtained. Based on this historical input sequence, the model outputs a future capacity or SOH prediction sequence, the mathematical definition of which is provided below.
[0060] Specifically, the fusion of input features can be represented by the following formula: ; Specifically, the historical input sequence can be represented by the following formula: ; Specifically, the predicted output sequence can be represented by the following formula: ; For the t-th time step or the t-th loop, the fused input features, historical input sequence, and predicted output sequence can be defined as shown in the above equation.
[0061] Where L represents the length of the historical sequence and H represents the prediction step size. This is the fused feature input vector at time step t. The features include: , , These represent the extracted voltage, current, and temperature values at the t-th time step, respectively. This represents the feature vector of the incremental capacity curve after preprocessing, uniform dimension interpolation, and Gaussian smoothing at step t. This represents the cumulative number of charge-discharge cycles of the battery corresponding to the t-th time step. The total dimension of the multidimensional input sequence formed by concatenating the above features. It is an input matrix consisting of L consecutive historical time step vectors. L represents the time window length of the input historical sequence. It is the sequence of target values predicted by the model for the next H time steps. It is the predicted value at the i-th time step.
[0062] S14. Train the preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks.
[0063] In one embodiment, a sparse hybrid expert model can be obtained by training a preset expert architecture based on training samples of multidimensional input sequences and incremental capacity features, wherein the preset expert architecture is composed of multiple expert networks.
[0064] By conducting comprehensive testing through multiple expert networks, we can avoid the problem that a single deep learning model is unable to identify the internal degradation physical characteristics, battery capacity scale, and decay rate differences of different power sources, which leads to deviations between the test results and reality. Furthermore, comprehensive testing through multiple expert models can also adapt to the multi-stage degradation process of power sources (such as early linear decay, mid-term capacity regeneration, and late-stage rapid decay), further improving the accuracy of the testing.
[0065] In one embodiment, training a model on a preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model may include the following sub-steps: S141. Perform reversible instance normalization on the multidimensional input sequence to obtain normalized temporal features.
[0066] S142. The model is trained on the preset expert architecture using the training parameter samples of the normalized temporal features and the incremental capacity features to obtain a sparse hybrid expert model.
[0067] In one embodiment, the multidimensional input sequence is subjected to reversible instance normalization to obtain normalized temporal features. Then, the training parameter samples of the normalized temporal features and incremental capacity features are used to train a model on a preset expert architecture to obtain a sparse hybrid expert model.
[0068] The real labels used in model training are divided into three categories: capacity labels, SOH labels, and remaining useful life (RUL / lifetime) labels.
[0069] The three types of labels can be obtained from the measured data of the UPS local monitoring system (in actual engineering deployment scenarios). Specifically, the measured data of the UPS local monitoring system can come from the actual UPS power supply monitoring system, or from publicly available battery degradation datasets or laboratory battery testing platforms. When used for model training, publicly available battery datasets can be used as the training data source in the examples; in practical applications, the data collected online by the UPS power supply of the power distribution communication network switch can be used as the model input.
[0070] The UPS regularly performs verification charge and discharge tests to measure the full charge capacity of the battery and calculate the true SOH (State of Health) based on the rated capacity; it also records the corresponding number of cycles and failure time to obtain the true RUL (Rated Life Label).
[0071] The three types of labels can also be obtained from the laboratory battery charge and discharge test platform (model offline training dataset). The lithium battery can be subjected to a complete aging cycle in the laboratory, and the measured capacity and SOH of each cycle can be recorded until the battery reaches the EOL life end threshold. The entire set of time-series measured data can be directly used as the true value of the label.
[0072] Alternatively, the three types of labels can also be obtained from publicly available battery degradation datasets (general model training materials). Specifically, publicly available industry lithium battery degradation datasets (including complete capacity and SOH time series ground truth values) can be used for model pre-training and generalization validation.
[0073] The processed incremental capacity curve can then be mapped to the capacity, SOH, or lifetime labels of the corresponding cycles to form training samples, validation samples, and test samples.
[0074] In one embodiment, the preset expert architecture includes: a multi-resolution prediction head, multiple different expert networks, and a router; The router is used to determine the selection probability of multiple different expert networks based on the hidden state, which is generated based on the normalized time series features. The multi-resolution prediction head includes a short-term prediction branch and a long-term trend prediction branch.
[0075] In one embodiment, the sparse hybrid expert model processes physical feature input, RevIN reversible instance normalization, input embedding, imaginary part enhanced rotational position encoding, unnormalized Derf module, sparse hybrid expert module and multi-resolution prediction head in that order, and outputs SOH estimation results and RUL prediction results.
[0076] The input embedding module is used to map the fused multidimensional input sequence to the hidden feature space. The imaginary part enhanced rotational position encoding module is used to obtain the real part attention score and the imaginary part attention score in the attention calculation, and to fuse the two in a weighted manner, thereby enhancing the model's ability to model local charging and discharging changes and long-term capacity degradation trends. The unnormalized Derf module is used to replace or partially replace the normalization operation in the traditional Transformer. It suppresses anomalous noise through soft saturation nonlinear mapping based on the error function, while retaining amplitude information related to the degree of battery degradation. The sparse hybrid expert module includes a router and multiple expert networks. The router selects the Top-K expert networks to participate in the computation based on the input hidden state, so that different expert networks can adapt to different aging stages or different degradation modes. The multi-resolution prediction head, including a short-term prediction branch and a long-term trend prediction branch, is used to output capacity or SOH prediction results at multiple time scales and integrate short-term fluctuations and long-term degradation trends to obtain the final prediction sequence.
[0077] Furthermore, the sparse hybrid expert module comprises N expert networks and a router. For the input hidden state, the router outputs the weights corresponding to each expert network, and selects the K experts with the highest weights to participate in the computation of the current sample or the current time step. The model output is obtained by weighted summation of the outputs of the selected expert networks according to the routing weights. By activating only some expert networks, the model can improve expressive power while reducing online inference computation, and enable different expert networks to learn different degradation characteristics such as linear degradation, slow decay, capacity regeneration, and rapid capacity decline at the end of their lifetime.
[0078] In one operation mode, the normalized temporal features after the reversible instance normalization process in step S141 are input into the embedding module and mapped to the hidden dimension of the model.
[0079] In one implementation, the input embedding module can employ a SwiGLU gating structure, as shown in the following equation: ; This represents the initial hidden layer state features that are input to the model and output from the embedding layer. , , , , Both are learnable weight matrices and bias vectors in the input embedding module. It is a smooth nonlinear activation function, and its calculation formula is: . This represents the Hadamard product, which is the element-wise multiplication of a matrix or vector.
[0080] For the high-dimensional physical degradation features contained in the incremental capacity curve, the input embedding module can unify them with operating state features such as voltage, current, and temperature into the same hidden representation space. Among these, the symbols... This indicates element-wise multiplication.
[0081] For the imaginary part-enhanced rotational position encoding module, battery life prediction depends not only on short-term charge-discharge fluctuations but also on long-term degradation history. Traditional rotational position encoding typically utilizes the real part information in the complex attention score, while underutilizing the imaginary part information. This embodiment introduces imaginary part-enhanced rotational position encoding into the attention calculation, enabling the model to simultaneously obtain local correlation information and global phase distance information.
[0082] Specifically, during the attention calculation process, after rotating and encoding the query vector and key vector, the real part attention score and the imaginary part attention score can be calculated separately. Then, the real part attention score and the imaginary part attention score are weighted and fused before being input into the Softmax function to obtain the attention weights. The real part attention score is used to characterize the local charging and discharging relationship between adjacent or short-distance time steps, while the imaginary part attention score is used to enhance the model's ability to perceive degradation trends over longer time spans. η is a preset weight or a learnable weight.
[0083] In one embodiment, the real part attention score can be calculated as follows: ; In one embodiment, the imaginary part attention score can be calculated as follows: ; The attention weights can be calculated as follows: ; In the above formula, and Let be the real and imaginary parts of the complex attention fraction generated by time step t on time step s, respectively. and These are the query vector and key vector in the complex field, respectively, with superscripts. It indicates complex conjugate. The nth rotation angle frequency is preset for encoding the rotation position. and These are the query vector and key vector in the real number field, respectively. This represents a rotation matrix with a relative position distance of st. Indicates phase offset as The rotation matrix is used to extract the imaginary part information. d is the feature dimension of the attention head, used as a scaling factor to prevent the gradient from vanishing due to excessively large dot product values. It is the fusion weight coefficient that controls the intensity of the imaginary attention branch.
[0084] For the unnormalized Derf module, while LayerNorm or RMSNorm in traditional Transformers are beneficial for stable training, the amplitude changes of capacity, voltage curves, and incremental capacity curves may themselves contain important degradation information in battery lifetime prediction tasks. Over-normalization may weaken this amplitude information. Therefore, this embodiment uses a Derf module based on an error function to replace or partially replace the traditional normalization operation.
[0085] The unnormalized Derf module performs a nonlinear mapping of the hidden state based on an error function. Specifically, the nonlinear mapping can be expressed as follows: ; The error function and its derivative can be expressed as follows: ; .
[0086] This represents the hidden state vector input to the Transformer layer. and These are the learnable channel scaling and bias vectors from the Derf module, used to perform affine transformations. It is a Gaussian error function that can perform soft saturation truncation on maxima outliers. and These are the learnable scaling and offset parameters within the error function.
[0087] The unnormalized Derf module suppresses extreme noise and high-frequency perturbations through soft saturation while preserving capacity degradation amplitude information. Therefore, the model can reduce its dependence on traditional normalization layers and maintain good training stability and prediction robustness.
[0088] For sparse hybrid expert modules, UPS power batteries exhibit different degradation modes at different stages of their lifespan. For example, early on, they may show relatively slow linear degradation; in the middle stage, capacity fluctuations or capacity regeneration may occur; and in the later stage, rapid capacity decline may occur. A single feedforward network cannot adequately adapt to all degradation modes simultaneously.
[0089] This embodiment uses a sparse hybrid expert module to replace the feedforward network in the standard Transformer. This module includes multiple expert networks and a router. The router outputs the expert selection probability based on the current hidden state, as shown in the following formula: ; In the above formula, This is represented as the gating probability vector of each expert network output by the router. It is the learnable gating weight matrix of the router.
[0090] Next, the K expert networks with the highest weights can be selected to participate in the calculation. The outputs of the selected expert networks are weighted and summed according to their routing weights to obtain the output of this module, as shown in the following formula: ; In the above formula, This represents the final weighted output feature of the sparse hybrid expert module. This represents the set of expert network indexes that rank the top K in terms of gating probability. This represents the probability of the routing weight assigned to the i-th selected expert. This represents the calculation output of the i-th expert feedforward network for the input state.
[0091] Through the sparse routing mechanism described above, different expert networks can adaptively learn the features of different aging stages or degradation modes during training. For example, some experts are better suited to modeling slow linear decay stages, some are better suited to modeling stages with rapid capacity decline, and some are better suited to handling stages with capacity regeneration or measurement noise. Since only a portion of the expert networks are activated during each inference, this module can control the computational load of online inference while improving model capacity.
[0092] To ensure both short-term prediction accuracy and long-term trend stability, this embodiment employs a multi-resolution prediction head. The multi-resolution prediction head includes a short-term prediction branch and a long-term trend prediction branch, with different branches corresponding to different prediction scales. During training, the multi-resolution prediction loss can be calculated to enable the model to simultaneously learn short-term fluctuations and long-term degradation trends. Specifically, the multi-resolution prediction loss can be expressed as follows: ; In the above formula, This represents the main task prediction loss generated by the multi-resolution prediction head. M is the number of time scales set for multi-resolution prediction (e.g., evaluating 12, 24, 48, and 96 steps respectively). This represents the loss weight coefficient at the m-th resolution scale. , Let represent the true label sequence and the model prediction sequence at the m-th resolution, respectively.
[0093] The above model structure can be used to train UPS power state estimation and remaining service life prediction models.
[0094] Specifically, the sparse hybrid expert Transformer model is trained using training samples. Training objectives may include capacity prediction error, SOH prediction error, and multi-resolution prediction error. To prevent unbalanced expert utilization in the sparse hybrid expert module, an expert load balancing loss can be added, as shown in the following equation: The sample or token allocation ratio and average routing probability corresponding to the i-th expert are represented by the variables in the above formula.
[0095] The total training loss can be expressed as follows: ; In the above formula, This represents the load balancing auxiliary loss in sparse expert networks, used to prevent some experts from being overused and causing expert crashes. It is a constant representing the weight of the auxiliary loss term in the total loss. N is the total number of expert networks included in the model. This represents the statistical proportion of tokens assigned to the i-th expert within the training batch. Let L be the statistical probability assigned by the router to the i-th expert within the training batch. Let L denote the total loss function optimized by joint backpropagation of the model.
[0096] After model training is completed, model parameters are obtained for UPS power state estimation and remaining service life prediction. The model can be deployed on an operation and maintenance server, edge computing device, or UPS power monitoring system to receive online data and output prediction results.
[0097] This invention uses a non-normalized Derf module to replace or partially replace the traditional normalization operation. While suppressing abnormal noise and high-frequency disturbances in the sensor, it retains the battery capacity degradation amplitude information and avoids the weakening of the degradation degree characterization by the traditional normalization operation.
[0098] Moreover, this invention employs imaginary part enhanced rotational position encoding, which simultaneously models local charge-discharge changes and long-period degradation trends in attention calculation, thus improving the RUL extrapolation capability under long prediction step sizes.
[0099] Furthermore, this invention employs a sparse hybrid expert module, dynamically selecting some expert networks to participate in the calculation based on the current degradation state of the battery. This allows different expert networks to learn the degradation patterns of different aging stages, thereby improving the model's expressive power while maintaining online inference efficiency. Moreover, by using a multi-resolution prediction head to simultaneously constrain short-term and long-term trend predictions, the model can account for both short-term capacity fluctuations and overall degradation trends, making it suitable for online status estimation and maintenance early warning of UPS power supplies in power distribution and communication network switches.
[0100] As an example, the reversible instance normalization process performed on the multidimensional input sequence to obtain normalized temporal features may include the following sub-steps: S1411. Calculate time series parameters using the historical input sequence, wherein the time series parameters include: time series mean and time series standard deviation.
[0101] S1412. The time-series parameters are used to normalize the fused input features and the predicted output sequence to obtain normalized fused features and normalized predicted sequences, respectively.
[0102] S1413. Perform reversible instance normalization processing on the normalized fusion features and the normalized prediction sequence respectively to obtain normalized time series features.
[0103] In one embodiment, the mean and standard deviation of the input sequence for each battery sample can be calculated separately for each sample and for each feature dimension. The input sequence is then normalized using the mean and standard deviation, enabling the model to learn relative degradation trends rather than relying solely on the absolute capacity scale. After the model outputs the prediction result, the mean and standard deviation of the corresponding sample and corresponding feature dimension are used to perform inverse normalization at the output, restoring the prediction result to the physical quantity scale. Through the above reversible instance normalization process, the model's generalization ability is improved when facing UPS batteries of different models, different initial capacities, and different degradation rates.
[0104] Because UPS power supply batteries may come from different manufacturers, batches, or operating environments, their initial capacity, voltage range, and degradation rate vary. If data from different individuals is directly input into the model, the model may over-rely on absolute capacity metrics and ignore the actual degradation trend. Therefore, this embodiment employs Reversible Instance Normalization (RevIN) to independently normalize the input sequence for each battery sample.
[0105] Specifically, the mean and standard deviation are calculated for each sample sequence by sample and by feature dimension. In one operation mode, time-series parameters can be calculated using historical input sequences, wherein the time-series parameters include: time-series mean and time-series standard deviation.
[0106] Specifically, the time series mean can be expressed as follows: ; Specifically, the time series standard deviation can be expressed as follows: ; Then, the fused input features are normalized, as shown in the following formula: ; After the model completes the prediction, the prediction result is inversely normalized to the original physical scale at the output, as shown in the following formula: The overall RevIN process can be represented by the following formula: ; In the above formula, and represents the time mean and time standard deviation of the input historical sequence on the d-th feature dimension, respectively. This represents the original feature value of the d-th dimension at step t. This represents the relative eigenvalues after normalization. This represents the normalized result value of the model's predicted output. This indicates the predicted value restored to the true physical scale after the inverse normalization operation. This represents the normalization operation function for forward reversible instances. This represents the main body of the sparse hybrid expert Transformer deep network model.
[0107] This invention employs reversible instance normalization to address capacity scale differences and distribution drift among different UPS batteries, enabling the model to better adapt to battery data of different models, initial capacities, and usage environments.
[0108] S15. Call the sparse hybrid expert model to determine the power supply's detection parameters based on the power supply's real-time operating data, wherein the detection parameters include: remaining lifetime value and health status percentage value.
[0109] In one embodiment, real-time operating data of the UPS power supply battery under test can be collected online, and the real-time operating data can be used for feature extraction and sequence construction in accordance with steps S12 and S13. The constructed input sequence is then input into a trained sparse hybrid expert model to obtain a future capacity or SOH prediction sequence. The state of health (SOH) is calculated based on the ratio of predicted capacity to rated capacity or initial capacity, resulting in a percentage value of the state of health. The point at which the SOH or capacity first falls below the preset end-of-life threshold is determined as the predicted end-of-life point. The difference between the predicted end-of-life point and the current point in time or the current number of cycles is used as the RUL prediction result.
[0110] In one specific implementation, the health status percentage can be calculated based on the ratio of the current capacity to the rated capacity or the initial capacity, as shown in the following formula: ; When the predicted SOH is lower than the preset lifetime termination threshold, the corresponding cycle number or time point is determined as the predicted lifetime termination point, as shown in the following formula: .
[0111] The difference between the current cycle number or the current time point and the predicted lifetime end point is the RUL prediction result, from which the remaining lifetime value is obtained, as shown in the following formula: .
[0112] In the above formula, It represents the current percentage of battery health at time step t. It is the maximum available capacity predicted by the model or actually measured at time step t. It is the rated nominal capacity of the battery when it leaves the factory. It is the number of cycles or the time point at which the battery is predicted to reach the end-of-life. It represents the health status during the c-th cycle of the predicted trajectory. It is a preset end-of-life threshold (often set at 80% in industry). It is the remaining service life, which is the predicted number of cycles that can continue to operate. This is the current starting loop number when executing this prediction task.
[0113] In one implementation, the end-of-life threshold can be set to 80%, or it can be adjusted according to the UPS power supply operation and maintenance specifications or the importance level of the equipment; the above 80% is only an example of threshold setting and does not constitute the only limitation on the end-of-life threshold.
[0114] After the step of calling the sparse hybrid expert model to determine the power supply's detection parameters based on the power supply's real-time operating data, the method may further include the following sub-steps: S16. When the percentage value of the health status is less than the preset health threshold or the remaining lifespan value is less than the preset power maintenance value, an early warning message is generated. or; S17. Perform multi-level early warning processing based on the percentage value of health status and the value of remaining lifespan.
[0115] In one embodiment, a preset health threshold and a preset power maintenance value can be set for the UPS power battery; when the predicted SOH is lower than the preset health threshold, or the predicted RUL is less than the preset power maintenance value, maintenance warning information is generated; corresponding operation and maintenance suggestions can also be output according to different warning levels, including continuing monitoring, shortening the inspection cycle, arranging battery retesting, planning to replace the battery, or replacing the battery immediately.
[0116] Specifically, warning thresholds are set based on the importance, maintenance cycle, and safety margin of the UPS power supply in the power distribution communication network switch. When the model-predicted State of Health (SOH) is lower than the warning threshold, or the predicted Range of Limiting Usage (RUL) is less than the preset maintenance lead time, the system generates a warning message. The warning message may include the current SOH, predicted RUL, recommended maintenance time, recommended retest time, and recommended replacement level.
[0117] In one implementation, the system can be configured with multiple warning levels. For example, when the Range of Limits (RUL) is long but the State of Health (SOH) is showing a downward trend, the system outputs a recommendation to continue monitoring or shorten the inspection cycle; when the RUL is close to the maintenance lead time, the system outputs a recommendation to plan replacement; when the SOH is below the end-of-life threshold or the RUL is extremely short, the system outputs a risk warning to immediately replace or deactivate the system.
[0118] In this embodiment, the present invention provides a power supply detection method based on a sparse hybrid expert model. Its advantages are as follows: the present invention can acquire historical operating data and capacity degradation data of the power supply separately; determine incremental capacity features based on the historical operating data and capacity degradation data; construct a multi-dimensional input sequence using the incremental capacity features and historical operating data; train a preset expert architecture based on the multi-dimensional input sequence and incremental capacity features to obtain a sparse hybrid expert model, wherein the preset expert architecture is composed of multiple expert networks; and call the sparse hybrid expert model to determine the remaining lifetime value and health status percentage value of the power supply based on the real-time operating data of the power supply. The present invention uses a model containing multiple expert networks for power supply detection, which avoids the problem that a single deep learning model is unable to identify the internal degradation physical characteristics, battery capacity scale, and decay rate differences of different power supplies, leading to deviations between the detection results and reality. Furthermore, multiple expert network models can adapt to the multi-stage degradation process of the power supply, further improving the detection accuracy.
[0119] This invention also provides a power supply detection device based on a sparse hybrid expert model, see [link to relevant documentation]. Figure 2 The diagram shows a schematic of the structure of a power detection device based on a sparse hybrid expert model according to an embodiment of the present invention.
[0120] As an example, the power supply detection device based on the sparse hybrid expert model may include: The acquisition module 201 is used to acquire the historical operating data and capacity degradation data of the power supply, respectively. The determining module 202 is used to determine the incremental capacity characteristics based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process; The construction module 203 is used to construct a multidimensional input sequence using the incremental capacity features and the historical operation data, wherein the multidimensional input sequence includes fused input features, historical input sequences and predicted output sequences; Training module 204 is used to train a model on a preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks. The detection module 205 is used to call the sparse hybrid expert model to determine the detection parameters of the power supply based on the real-time operating data of the power supply. The detection parameters include: remaining lifetime value and health status percentage value.
[0121] Optionally, determining the incremental capacity characteristics based on the historical operating data and the capacity degradation data includes: Constant current charging data is obtained from the historical operating data, wherein the constant current charging data is the voltage sequence and capacity sequence of the power supply during the constant current charging stage; An interpolated capacity curve is constructed using the constant current charging data, and the interpolated capacity curve is converted into a fixed-dimensional incremental capacity feature vector. The interpolated capacity curve is a curve obtained by constructing a curve using the constant current charging data and performing interpolation processing. The incremental capacity feature vector is subjected to Gaussian smoothing filtering to obtain a filtered feature vector, and the filtered feature vector is divided into training parameter samples according to the capacity degradation data to obtain incremental capacity features.
[0122] Optionally, constructing a multidimensional input sequence using the incremental capacity feature and the historical operating data includes: Operational scalar feature data is extracted from the historical operation data, wherein the operational scalar feature data includes: average voltage, average current, average temperature, and cumulative cycle count; The scalar feature data of operation and the incremental capacity feature are concatenated to obtain the fused input feature; The fused input features are stacked and concatenated with feature vectors from several historical time steps in chronological order to obtain a historical input sequence; Starting with the fused input features, the output target of the model is constructed based on the preset prediction step size to obtain the predicted output sequence.
[0123] Optionally, the step of training a model on a preset expert architecture based on the multidimensional input sequence and the incremental capacity features to obtain a sparse hybrid expert model includes: The multidimensional input sequence is subjected to reversible instance normalization to obtain normalized temporal features; The model is trained on the preset expert architecture using the training parameter samples of the normalized temporal features and the incremental capacity features to obtain a sparse hybrid expert model.
[0124] Optionally, the reversible instance normalization process performed on the multidimensional input sequence to obtain normalized temporal features includes: The time series parameters are calculated using the historical input sequence, wherein the time series parameters include: time series mean and time series standard deviation; The time-series parameters are used to normalize the fused input features and the predicted output sequence to obtain normalized fused features and normalized predicted sequences, respectively. The normalized fusion features and the normalized prediction sequence are respectively subjected to reversible instance normalization processing to obtain normalized time-series features.
[0125] Optionally, the preset expert architecture includes: a multi-resolution prediction head, multiple different expert networks, and a router; The router is used to determine the selection probability of multiple different expert networks based on the hidden state, which is generated based on the normalized time series features. The multi-resolution prediction head includes a short-term prediction branch and a long-term trend prediction branch.
[0126] Optionally, the device further includes: The early warning information module is used to generate early warning information when the health status percentage value is less than a preset health threshold or the remaining lifespan value is less than a preset power maintenance value after the step of calling the sparse hybrid expert model to determine the power supply detection parameters based on the real-time operating data of the power supply. or; The early warning processing module is used to perform multi-level early warning processing based on the health status percentage value and the remaining lifetime value after the step of calling the sparse hybrid expert model to determine the detection parameters of the power supply based on the real-time operating data of the power supply.
[0127] Those skilled in the art will understand that, for ease of description and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0128] Furthermore, embodiments of this application also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the power detection method based on a sparse hybrid expert model as described in the above embodiments.
[0129] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer-executable program for causing a computer to execute the power supply detection method based on a sparse hybrid expert model as described in the above embodiments.
[0130] In the description of the embodiments of the present invention, it should be noted that the terms "above," "below," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. When an element such as a layer, region, or substrate is referred to as being "above" or "on top of" another element, it may be directly on the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly on" or "above" another element, there is no intermediate element. It should also be understood that when an element is referred to as being "below" or "under" another element, it may be directly below or under the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly below" or "under" another element, there is no intermediate element. Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0131] Those skilled in the art will understand that embodiments of this application may also include computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0132] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), devices, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0135] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A power source detection method based on a sparse hybrid expert model, characterized in that, The method includes: Acquire historical operating data and capacity degradation data of the power supply separately; The incremental capacity characteristics are determined based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process; A multidimensional input sequence is constructed using the incremental capacity feature and the historical operating data, wherein the multidimensional input sequence includes fused input features, historical input sequence, and predicted output sequence; The preset expert architecture is trained based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks. The sparse hybrid expert model is invoked to determine the power supply's detection parameters based on the power supply's real-time operating data. The detection parameters include: remaining lifetime value and health status percentage value.
2. The power supply detection method based on a sparse hybrid expert model according to claim 1, characterized in that, The step of determining incremental capacity characteristics based on the historical operating data and the capacity degradation data includes: Constant current charging data is obtained from the historical operating data, wherein the constant current charging data is the voltage sequence and capacity sequence of the power supply during the constant current charging stage; An interpolated capacity curve is constructed using the constant current charging data, and the interpolated capacity curve is converted into a fixed-dimensional incremental capacity feature vector. The interpolated capacity curve is a curve obtained by constructing a curve using the constant current charging data and performing interpolation processing. The incremental capacity feature vector is subjected to Gaussian smoothing filtering to obtain a filtered feature vector, and the filtered feature vector is divided into training parameter samples according to the capacity degradation data to obtain incremental capacity features.
3. The power supply detection method based on a sparse hybrid expert model according to claim 1, characterized in that, The construction of a multidimensional input sequence using the incremental capacity feature and the historical operating data includes: Operational scalar feature data is extracted from the historical operation data, wherein the operational scalar feature data includes: average voltage, average current, average temperature, and cumulative cycle count; The scalar feature data of operation and the incremental capacity feature are concatenated to obtain the fused input feature; The fused input features are stacked and concatenated with feature vectors from several historical time steps in chronological order to obtain a historical input sequence; Starting with the fused input features, the output target of the model is constructed based on the preset prediction step size to obtain the predicted output sequence.
4. The power supply detection method based on a sparse hybrid expert model according to claim 1, characterized in that, The step of training a sparse hybrid expert model based on the multidimensional input sequence and the incremental capacity features to obtain a preset expert architecture includes: The multidimensional input sequence is subjected to reversible instance normalization to obtain normalized temporal features; The model is trained on the preset expert architecture using the training parameter samples of the normalized temporal features and the incremental capacity features to obtain a sparse hybrid expert model.
5. The power supply detection method based on a sparse hybrid expert model according to claim 4, characterized in that, The reversible instance normalization process performed on the multidimensional input sequence to obtain normalized temporal features includes: The time series parameters are calculated using the historical input sequence, wherein the time series parameters include: time series mean and time series standard deviation; The time-series parameters are used to normalize the fused input features and the predicted output sequence to obtain normalized fused features and normalized predicted sequences, respectively. The normalized fusion features and the normalized prediction sequence are respectively subjected to reversible instance normalization processing to obtain normalized time-series features.
6. The power supply detection method based on a sparse hybrid expert model according to claim 4, characterized in that, The preset expert architecture includes: a multi-resolution prediction head, multiple different expert networks, and a router; The router is used to determine the selection probability of multiple different expert networks based on the hidden state, which is generated based on the normalized time series features. The multi-resolution prediction head includes a short-term prediction branch and a long-term trend prediction branch.
7. The power supply detection method based on a sparse hybrid expert model according to claim 6, characterized in that, After the step of calling the sparse hybrid expert model to determine the power supply's detection parameters based on the power supply's real-time operating data, the method further includes: If the percentage value of the health status is less than a preset health threshold or the remaining lifespan value is less than a preset power maintenance value, an early warning message will be generated. or; Multi-level early warning processing is performed based on the percentage value of health status and the value of remaining lifespan.
8. A power supply detection device based on a sparse hybrid expert model, characterized in that, The device includes: The acquisition module is used to acquire historical operating data and capacity degradation data of the power supply, respectively. The determination module is used to determine the incremental capacity characteristics based on the historical operating data and the capacity degradation data, wherein the incremental capacity characteristics are the incremental capacity characteristics of the power supply during the charging process; A construction module is used to construct a multidimensional input sequence using the incremental capacity features and the historical operating data, wherein the multidimensional input sequence includes fused input features, historical input sequences, and predicted output sequences; The training module is used to train the preset expert architecture based on the multidimensional input sequence and the incremental capacity feature to obtain a sparse hybrid expert model, wherein the preset expert architecture includes multiple expert networks. The detection module is used to call the sparse hybrid expert model to determine the detection parameters of the power supply based on the real-time operating data of the power supply. The detection parameters include: remaining lifetime value and health status percentage value.
9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the power supply detection method based on a sparse hybrid expert model as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the power supply detection method based on a sparse hybrid expert model as described in any one of claims 1-7.
