Battery swapping device fault identification method and device

Abstract

The application provides a battery replacement equipment fault identification method and device, and relates to the technical field of artificial intelligence, real-time charging process data generated in the charging process of a charging bin of a battery replacement equipment is acquired, the total voltage of aliasing is decoupled according to physical effects, independent components contained therein are separated, a multi-channel decoupling matrix is constructed, each channel of the matrix respectively carries independent information with clear physical meaning, and a structured observation basis can be formed. The fault diagnosis model can perform independent change analysis and cross-channel correlation comparison on each channel of the matrix, the prediction process is completely based on the physical attribution of each channel, the diagnosis conclusion can be verified back through each channel, and the application has inherent physical interpretability. Any weak fault signal occurring in the charging process of the battery replacement equipment can be captured in time by the application, and early fault of the battery replacement equipment can be accurately identified.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for identifying faults in battery swapping equipment. Background Technology

[0002] With the rapid development of shared mobility and battery swapping models, electric bicycle battery swapping stations have been deployed extensively in cities. These stations address users' pain points of difficult and slow charging by providing centralized charging and rapid battery replacement services. However, due to frequent charging and discharging, plugging and unplugging of battery packs, and long-term operation in complex environments, battery swapping stations are prone to various types of malfunctions, such as poor contact, battery aging, overheating, and undervoltage. These malfunctions not only affect the user experience but may also lead to serious safety accidents such as fires. Therefore, real-time and accurate fault identification and early warning of the operating status of battery swapping equipment are crucial.

[0003] Currently, fault identification in battery swapping equipment typically involves directly extracting or classifying features from multi-source time-series data such as raw current, voltage, and temperature. However, the terminal voltage of a battery swapping device during charging is not determined by a single factor; its changes are influenced by multiple factors, including charging current, battery condition, and electrical connection status. Current fluctuations and operating condition switching during normal charging can also cause instantaneous voltage jumps. Directly analyzing the raw terminal voltage makes it difficult for diagnostic models to effectively distinguish voltage fluctuations caused by normal operating conditions from early fault characteristics such as poor contact, leading to difficulties in controlling false alarms and missed alarms, and insufficient accuracy in identifying early faults in battery swapping equipment. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for identifying faults in battery swapping equipment, which can accurately identify early-stage faults.

[0005] In a first aspect, embodiments of the present invention provide a method for fault identification of a battery swapping device. The method includes: in response to detecting that a battery pack has entered the charging compartment of the battery swapping device, acquiring real-time charging process data generated by the charging compartment during the charging process; the real-time charging process data includes at least current data, voltage data, and temperature data; decoupling the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data; constructing a multi-channel decoupling matrix based on the voltage components, current data, and temperature data; inputting the multi-channel decoupling matrix into a pre-constructed fault diagnosis model for the battery swapping device, and predicting faults in the battery swapping device based on the multi-channel decoupling matrix through the fault diagnosis model for the battery swapping device.

[0006] In conjunction with the first aspect, the present invention provides a first implementation of the first aspect, wherein the step of decoupling voltage data in real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data includes: determining the ohmic voltage drop component in the voltage data based on the ohmic internal resistance of the equivalent circuit model of the battery swapping device during the charging process; determining the polarization voltage drop component in the voltage data based on the polarization internal resistance and polarization capacitance of the equivalent circuit model of the battery swapping device during the charging process; and determining the open-circuit voltage component in the voltage data based on the state of charge data obtained by ampere-hour integration of current data during the charging process of the battery swapping device and the temperature data of the charging process.

[0007] In conjunction with the first aspect, this embodiment of the invention provides a second implementation of the first aspect, wherein the fault diagnosis model for the battery swapping equipment is constructed based on a dual-path time-frequency sensing neural network; the step of inputting a multi-channel decoupling matrix into the pre-constructed fault diagnosis model for the battery swapping equipment, and predicting faults of the battery swapping equipment based on the multi-channel decoupling matrix by the fault diagnosis model for the battery swapping equipment includes: inputting the data of each channel of the multi-channel decoupling matrix in parallel into the dual-path time-frequency sensing neural network; extracting features from the high-frequency feature representation and the low-frequency feature representation in the multi-channel decoupling matrix by the dual-path time-frequency sensing neural network to determine the fused feature representation contained in the multi-channel decoupling matrix; and performing fault waveform matching on the fused feature representation based on a pre-constructed fault waveform template to determine the fault category to which the current charging process of the battery swapping equipment belongs.

[0008] In conjunction with the first aspect, this embodiment of the invention provides a third implementation of the first aspect, wherein the step of extracting features from the high-frequency and low-frequency feature representations in the multi-channel decoupling matrix using a dual-path time-frequency sensing neural network to determine the fused feature representation contained in the multi-channel decoupling matrix includes: extracting trend features at different time spans in the multi-channel decoupling matrix and fusing the trend features step by step to determine the low-frequency feature representation in the multi-channel decoupling matrix; performing multi-branch parallel dilated convolution on the multi-channel decoupling matrix to determine the transient fault information of the multi-channel decoupling matrix at the corresponding time scale, and performing cross-branch attention aggregation on the transient fault information to determine the high-frequency feature representation in the multi-channel decoupling matrix; and performing adaptive weighted fusion of the high-frequency and low-frequency feature representations based on the operating condition information of the charging process of the battery swapping equipment to obtain the fused feature representation contained in the multi-channel decoupling matrix.

[0009] In conjunction with the first aspect, this embodiment of the invention provides a fourth implementation of the first aspect, wherein the step of adaptively weighting and fusing high-frequency feature representations and low-frequency feature representations based on the operating condition information of the charging process of the battery swapping equipment to obtain the fused feature representation contained in the multi-channel decoupling matrix includes: performing global average pooling and global max pooling on the high-frequency feature representations and low-frequency feature representations respectively to determine the high-frequency pooling feature vector corresponding to the high-frequency feature representation and the low-frequency pooling feature vector corresponding to the low-frequency feature representation; and performing weighted summation on the high-frequency pooling feature vector and the low-frequency pooling feature vector according to the adaptive fusion weights indicated by the operating condition information to obtain the fused feature representation contained in the multi-channel decoupling matrix.

[0010] In conjunction with the first aspect, this embodiment of the invention provides a fifth implementation of the first aspect, wherein the method further includes: calculating the category boundary margin of the battery swapping equipment fault diagnosis model based on the physical mechanism similarity between different fault categories of training samples; constructing a dynamically scaled classification loss based on the sample classification difficulty corresponding to the category boundary margin; constructing a total loss function of the battery swapping equipment fault diagnosis model based on the boundary margin loss and the dynamically scaled classification loss corresponding to the category boundary margin; and training the battery swapping equipment fault diagnosis model based on the total loss function.

[0011] In conjunction with the first aspect, this invention provides a sixth implementation of the first aspect, wherein the method further includes: determining a voltage residual sequence of the voltage data of the historical charging data based on multiple voltage components corresponding to the historical charging data of the battery swapping equipment; performing data preprocessing on the initial training samples corresponding to the historical charging data based on the voltage residual sequence to construct preprocessed training samples; determining the fault matching segment included in the preprocessed training samples using a preset fault waveform template matching; and performing waveform sharpening on the preprocessed training samples based on the local detail components in the fault matching segment to construct the final training samples.

[0012] Secondly, embodiments of the present invention also provide a fault identification device for battery swapping equipment. The device includes: a data acquisition module, configured to acquire real-time charging process data generated by the charging compartment during charging in response to monitoring the entry of a battery pack into the charging compartment of the battery swapping equipment; the real-time charging process data includes at least current data, voltage data, and temperature data; a data processing module, configured to decouple the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data; a data fusion module, configured to construct a multi-channel decoupling matrix based on the voltage components, current data, and temperature data; and an execution module, configured to input the multi-channel decoupling matrix into a pre-constructed fault diagnosis model for battery swapping equipment, and to predict faults in the battery swapping equipment based on the multi-channel decoupling matrix through the fault diagnosis model.

[0013] Thirdly, embodiments of the present invention also provide an electronic device, wherein the electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, and the processor executing the machine-executable instructions to implement the battery swapping equipment fault identification method of any of the above embodiments.

[0014] Fourthly, embodiments of the present invention also provide a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, which, when called and executed by a processor, cause the processor to implement the battery swapping equipment fault identification method of any of the above embodiments.

[0015] The embodiments of this invention bring the following beneficial effects: This invention provides a method and apparatus for fault identification in battery swapping equipment. By decoupling the total voltage of aliasing according to physical effects, the independent components contained therein are separated, allowing the electrical characterization of equipment circuit faults to focus on the physically related components, no longer affected by battery-side voltage fluctuations. Each channel in the constructed multi-channel decoupling matrix carries independent information with clearly defined physical meanings, forming a structured observation basis. This enables the fault diagnosis model to perform independent change analysis and cross-channel correlation comparison on each channel when performing fault prediction on this matrix. The prediction process is entirely based on the physical attribution of each channel, and the diagnostic conclusions can be verified channel by channel, possessing inherent physical interpretability. Any weak fault signal occurring during the charging process of the battery swapping equipment can be captured in a timely manner, thereby enabling accurate capture and differentiation of early faults.

[0016] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart of a method for identifying faults in a battery swapping device provided in an embodiment of the present invention; Figure 2This is a schematic diagram of the timing waveform of real-time charging process data provided in an embodiment of the present invention; Figure 3 A waveform diagram of a voltage component provided for an embodiment of the present invention; Figure 4 A schematic diagram of a fault prediction process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the ohmic internal resistance changing over time, provided as an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a fault identification device for battery swapping equipment provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] The raw signals collected during charging, such as total voltage, current, and temperature, are essentially products of the coupling between the battery's own characteristics and the state of the battery swapping equipment. When the connector contact resistance increases slightly or the heat dissipation duct begins to block, the resulting weak signal anomalies are easily overwhelmed by normal voltage fluctuations caused by changes in the battery's state of charge, aging, or ambient temperature. This "signal aliasing" makes it difficult for traditional methods to isolate pure equipment fault characteristics at the initial stage, leading to false alarms, missed alarms, and ambiguous localization. To solve this problem, this invention utilizes an equivalent circuit model of the battery to analyze the total voltage during charging into multiple physical components (such as ohmic voltage drop, polarization voltage drop, and open-circuit voltage). This allows the ohmic voltage drop component, directly affected by the battery swapping equipment's circuit resistance, to be completely isolated from the battery's own state fluctuations and displayed independently. At this time, the small resistance increase caused by connector spring fatigue or contact surface oxidation will appear as a pure step of the ohmic component, eliminating the need for ambiguous identification with changes in the battery's open-circuit voltage or polarization voltage, thus solving the problem of invisible equipment characteristics from the data source.

[0023] To facilitate understanding, a fault identification method for battery swapping equipment provided in an embodiment of the present invention will first be described, referring to... Figure 1 The method includes the following steps: Step S102: In response to detecting that the battery pack has entered the charging compartment of the battery swapping device, real-time charging process data generated by the charging compartment during the charging process is acquired.

[0024] In the daily operation of a battery swapping station, a single swapping system comprises several independently operating charging compartments. Each compartment contains a floating connector for docking with the battery pack, a charging power circuit, and heat dissipation components. When a depleted battery pack is pushed into the charging compartment and physically engages with the floating connector, the charging circuit is instantly activated. The system confirms this engagement event via a position sensor within the compartment or a communication handshake signal from the battery management system, and then initiates the data acquisition link. At this point, charging process data can be acquired in real time from Hall effect current sensors and voltage sampling circuits deployed within the charging compartment, as well as temperature sensor arrays distributed near the connector and inside the compartment. The data includes current data flowing through the charging compartment connector and power circuit, voltage data measured at both ends of the charging compartment connector, and temperature data from key temperature measurement points within the charging compartment.

[0025] Step S104: Decouple the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data.

[0026] Early faults in battery swapping equipment during charging often manifest as extremely weak signal changes at the electrical level. For example, an increase in connector contact resistance typically results in only millivolts of increase in the ohmic voltage drop across the circuit. However, the voltage fluctuations within the battery itself, caused by factors such as temperature and state of charge, naturally reach tens of millivolts in the total voltage collected from the charging compartment. Fault signals and normal battery fluctuations overlap in the total voltage, making fault identification based solely on this total voltage extremely difficult. This makes it challenging to isolate the abnormal components from the vastly different mixed signals, leading to frequent false alarms and missed detections of early faults. This invention, through mechanistic analysis of the electrical behavior of the charging circuit, recognizes that voltage changes from different physical sources belong to different voltage components within the total voltage. Faults in the device circuit, such as increased connector contact resistance, have their electrical characteristics implicit in the ohmic voltage drop component generated by current flowing through the physical impedance; battery polarization fluctuations and open-circuit voltage changes correspond to the electrochemical polarization component and the equilibrium potential component, respectively. Based on this, the embodiments of the present invention decompose the total voltage into independent components from the perspective of physical causes, so that the weak signals of equipment failure are naturally separated from the interference of battery-side fluctuations and appear independently in their own physical dimension.

[0027] Step S106: Construct a multi-channel decoupling matrix based on voltage component, current data, and temperature data.

[0028] The aforementioned voltage components physically separate the device circuit characteristics from the battery-side characteristics. This invention also constructs a multi-channel decoupling matrix based on these components to integrate them into a complete data structure that can be uniformly processed by the diagnostic model. In one implementation, the multi-channel decoupling matrix can be constructed by arranging each signal as an independent channel side-by-side. This matrix incorporates the resistive characteristics of the charging compartment connection circuit, battery polarization dynamics, battery equilibrium potential, charging excitation intensity, and thermal field distribution within the compartment into a single observation framework with clearly defined physical semantics between channels and no overlap between them. When the diagnostic model makes judgments based on this matrix, it no longer needs to deal with a single waveform entangled with information, but can directly address the independent changes of each channel and their cross-channel correlations.

[0029] Step S108: Input the multi-channel decoupling matrix into the pre-built fault diagnosis model of the battery swapping equipment, and use the fault diagnosis model of the battery swapping equipment to predict the fault of the battery swapping equipment based on the multi-channel decoupling matrix.

[0030] The aforementioned multi-channel decoupling matrix is ​​fed into a pre-built fault diagnosis model for battery swapping equipment. The model can then perform fault prediction based on the independent change patterns of each channel and the interrelationships across channels. In the offline phase, the model has learned from multi-channel decoupling matrix samples corresponding to various operating conditions of the battery swapping equipment, acquiring differentiated feature expressions for different faults in the equipment loop component channel, polarization component channel, equilibrium potential component channel, and current and temperature channels. In the online operation phase, the model can perform channel-by-channel analysis and cross-channel comparison of the input matrix: when the equipment loop component channel exhibits abnormal fluctuations independent of the polarization and equilibrium potential components, and this fluctuation is proportionally correlated with the current channel, the model can determine it as a charging compartment connection circuit fault; when the temperature channel exhibits abnormal changes while the equipment loop component channel remains stable, it points to an abnormality in the heat dissipation system; if the polarization component and temperature channel deviate in tandem, it may correspond to an abnormal state on the battery side. The judgment logic for these various fault types is determined based on the clear physical meaning of each column of the multi-channel decoupling matrix, and the diagnostic results can be verified channel-by-channel, making the fault identification of the battery swapping equipment physically interpretable.

[0031] In summary, the present invention provides a method and apparatus for fault identification in battery swapping equipment. By decoupling the total voltage of aliasing according to physical effects, it separates the independent components contained therein. This concentrates the electrical characterization of equipment circuit faults on the physically associated components, eliminating interference from battery-side voltage fluctuations. The constructed multi-channel decoupling matrix contains independent information with clearly defined physical meanings for each channel, forming a structured observation basis. This allows the fault diagnosis model to perform independent change analysis and cross-channel correlation comparison on each channel when predicting faults. The prediction process is entirely based on the physical attribution of each channel, and the diagnostic conclusions can be verified channel by channel, possessing inherent physical interpretability. For example, when an abnormal fluctuation proportional to the current occurs in a component channel of the equipment circuit, while the polarization component and equilibrium potential component channels remain stable, it is determined to be a fault in the charging compartment connection circuit. When the temperature channel is abnormal while the component channels of the equipment circuit are stable, it points to an abnormality in the heat dissipation system. Coordinated deviation between the polarization component and the temperature channel can distinguish abnormalities in the battery-side state. In summary, any weak fault signal that occurs during the charging process of the battery swapping equipment can be captured in a timely manner, thereby enabling accurate detection and differentiation of early faults.

[0032] Furthermore, based on the above embodiments, this embodiment of the invention also provides another method for identifying faults in battery swapping equipment. This method is used to further illustrate the above embodiments, and the steps are detailed below.

[0033] Step S202: In response to detecting that the battery pack has entered the charging compartment of the battery swapping device, real-time charging process data generated by the charging compartment during the charging process is acquired.

[0034] This step is the same as the steps in the above embodiment, and will not be repeated here. In one implementation, the data can be used to construct a training sample set for the above model. For the construction of the training sample set, the data acquisition process can rely on the intelligent battery swapping cabinets deployed at each battery swapping station. Through the high-precision sensors built into the battery swapping cabinets, key physical parameters of electric bicycle batteries during charging, discharging, and insertion / removal processes are collected in real time at a fixed frequency to obtain a monitoring sample set. For example, when a user inserts the battery pack into the charging compartment of the battery swapping cabinet, the data acquisition system is immediately activated, and at a sampling interval of 0.1 seconds, it synchronously records the current, voltage, and temperature data of the battery pack throughout the entire process of charging, standby, and battery insertion / removal switching. Each complete sampling window contains 4096 continuous sampling points, covering the complete operation cycle from battery access and charging process to possible fault events, until the user removes the battery, forming an independent original multi-source time-series data matrix. The three rows of this matrix correspond to the current sequence, voltage sequence, and temperature sequence at the same moment, respectively, which together constitute a sample unit for training or identification. To ensure data consistency over time, the system performs strict timestamp alignment on all sensor channels and preprocesses missing values ​​caused by communication jitter or momentary sensor failure using linear interpolation or forward padding to ensure that the current, voltage, and temperature data of each sample correspond one-to-one in the time dimension.

[0035] After data collection is completed, each sample can be precisely categorized by experienced equipment maintenance engineers or fault diagnosis experts. The categorization process can begin with initial classification based on alarm logs and fault codes automatically recorded by the battery swapping system. Further, experts can manually verify and confirm the data by combining the complete current, voltage, and temperature waveforms within the sampling window. For composite cases where multiple fault symptoms appear simultaneously within a sampling window, experts will categorize them into the most representative fault type based on the order of occurrence and the degree of their primary impact, or label them separately as composite fault samples. In one implementation, based on common fault physical mechanisms of battery swapping equipment, the labeled category set can include five categories, such as: 1. Normal samples, representing that all parameters of the equipment operate within a safe and stable range within the sampling window; 2. Poor contact faults, mainly manifested as non-periodic spikes, glitches, or jitters in the current or voltage waveform, usually related to loose connectors or contact oxidation; 3. Undervoltage / overvoltage faults, characterized by the voltage waveform continuously deviating from the standard threshold during charging or discharging; 4. Aging faults, reflected as long-term trend changes caused by battery capacity decay or increased internal resistance, such as shortened voltage plateau and prolonged charging time; 5. Overtemperature faults, manifested as a continuous abnormal rise in the temperature baseline or an excessively rapid rate of temperature rise. For each type of fault, experts can further label the specific time interval of the fault occurrence, providing accurate supervisory information for subsequent construction of fault waveform templates and training of fault identification models. Finally, all the labeled original multi-source time-series data matrices and their corresponding fault category labels can jointly form the original dataset used to train the fault identification model of battery swapping equipment. In one embodiment, Figure 2 This diagram illustrates the timing waveforms of real-time charging data, showing the raw current (blue), voltage (red), and temperature (green) timing waveforms of the battery pack during the charging and discharging process within the sampling window. The raw signals are coupled with multiple physical factors. The horizontal axis represents time (seconds), and the vertical axes represent current (amperes), voltage (volts), and temperature (degrees Celsius), respectively.

[0036] Furthermore, existing methods for constructing training samples for fault diagnosis models of battery swapping equipment typically use raw charging data or its simple statistical features directly as training input. However, in the early stages of faults, the signal deviations caused by equipment anomalies are extremely weak. Directly using raw data to construct training samples often dilutes the fault features within the samples due to normal operating condition fluctuations, making it difficult for the model to learn the key patterns that distinguish fault boundaries. This leads to training convergence difficulties and limited diagnostic sensitivity.

[0037] Building upon the above embodiments, the construction of training samples can be based on decoupling them from their physical domains, followed by performing local waveform enhancement and data reconstruction for transient contact anomalies and progressive aging faults, respectively. This allows the subsequent neural network to learn the contributions of current surges, contact impedance changes, electrochemical aging, and thermal accumulation to fault identification. In one implementation, a voltage residual sequence can be calculated for each decoupled voltage component, and this voltage residual sequence can be used to optimize the features of the original data. For example, redundant segments with stable and anomaly-free residuals can be removed, while effective intervals containing residual fluctuations can be retained, allowing the training samples to focus on time periods that may carry fault information.

[0038] The voltage components here are analogous to those in the voltage data of the above embodiment; the corresponding residual sequence can be determined based on the differences between multiple voltage components in the voltage data. That is, the original voltage sequence... The remaining portion after subtracting the open-circuit voltage component, ohmic voltage drop component, and polarization voltage drop component is used to evaluate measurement noise and unmodeled perturbations. If the residual within a certain window is significantly large, that window can be marked as noise for subsequent training as a reference for sample quality (e.g., directly removing that sample). For example, within a window containing 512 sampling points, the following voltage components of the voltage data can be estimated using the model: open-circuit voltage... Ohmic voltage drop and polarization voltage drop The voltage residual sequence can be expressed as: If you find the window The root mean square error (RMSE) is 0.05V, which is much higher than the historical normal window's 0.01V. This indicates the presence of physical phenomena within this window that are not described by the model, such as strong electromagnetic interference, sudden changes in sensor noise, or abnormal data acquisition. The system can mark this window as "low-quality data" or "abnormal noise." These can be directly removed during model training to prevent the model from learning non-fault-related noise features.

[0039] Furthermore, this embodiment of the invention also utilizes a preset fault waveform template to match the preprocessed samples, determines the fault matching segments contained in the samples, and then extracts the corresponding local detail components of the segments. This shrinks the fault information from the entire time period of the samples to a specific local location, providing the model with a clear label of the fault occurrence interval. In specific implementation, during the plugging and unplugging switching and relay operation processes of battery swapping equipment, short-duration spikes, dips, and abrupt slope changes are prone to occur. These fault segments account for a small proportion in the complete sampling window, and are easily weakened by convolutional averaging or downsampling operations when directly input into a deep network. This embodiment of the invention can first construct a fault waveform template using historical fault data, then use template matching to locate suspected fault segments, and finally perform local detail enhancement only on the matching segments to obtain a waveform sharpening data matrix. The specific steps are as follows: 1> Extract typical fault segments from historically labeled fault samples and construct a fault waveform template set, denoted as . .in, Indicates the number of fault waveform templates, with the following example values: ; Indicates the template index; each fault waveform template A standardized local waveform representing a typical fault can correspond to types such as current spikes, voltage sags, sudden changes in temperature rise slope, and relay jitter.

[0040] In practical implementation, historical fault samples are first segmented into local fault segments based on manual annotation or alarm logs. Then, length normalization, amplitude normalization, and time alignment are performed on each segment. Finally, a clustering method is used to obtain the center waveform of each segment, which serves as the corresponding fault waveform template. .

[0041] In one embodiment, for example, 200 fault segments labeled "relay jitter" are extracted from historical data. These segments vary in length, and the voltage waveforms exhibit a series of rapid, continuous spikes and dips. First, all segment lengths are normalized to 128 sampling points through interpolation. Then, the amplitude of each segment is normalized by maximum and minimum values. Finally, k-means clustering (k=1) is used to extract the central waveform of the cluster, which serves as the fault waveform template for "relay jitter". This template represents a typical, standardized relay bounce fault mode.

[0042] 2> Decoupling the data matrix from the physical domain Perform local window scanning on each channel and compare the local waveforms with each fault waveform template. Matching is performed one by one; a dynamic time warping distance is used during matching to account for differences caused by temporal scaling and amplitude offset. In the specific implementation, the local window length can be set to... , or Each sampling point, the window step size can be taken as... or Each sampling point, during local window scanning, represents the waveform (length) for each window. ) and each template (length ), calculate their dynamic time warping distance, which reflects the waveform of each window and each template. Similarity in shape, even if their time lengths or local scaling differ, is considered when the similarity exceeds the corresponding template threshold. At that time, it was considered that the window was related to the first One template was successfully matched, among which... Indicates the first The matching threshold for each template is determined by statistics from the training set.

[0043] In practical implementation, the matching threshold It can be determined through statistical analysis based on the training set data. For example, first, collect all data labeled as being similar to the template from the training set. Local fault segments corresponding to the fault type (e.g., "relay bounce"). ,for Each fault segment in Calculate its relationship with the template The dynamic time-warped distance yields a set of positive sample distance distributions. Then, a large number of normal segments that are not of this fault type can be randomly sampled from the training set. (Quantity can be) (equal to or more). For Each normal segment By calculating its relationship with the template The dynamic time-warped distance can be used to obtain a set of negative sample distance distributions. Furthermore, a target true positive rate (e.g., 95%) or false positive rate (e.g., 5%) can be set to... and Find an optimal threshold on the distribution. For example, you can choose The 95th percentile of the distribution is used as This ensures that 95% of positive samples can be correctly matched.

[0044] 3> Furthermore, in this embodiment of the invention, a matching mask is also generated based on the template matching result. To describe the first The template in the first Whether a matching segment's flag value is matched at time step i is determined. This matching mask can be used as a basis for judging the fault category. For example, based on the matching mask of a certain segment, its fault category can be determined. For example, when the i-th... When the value is within the matching window at a given time, it is taken as... Otherwise, the value is In one embodiment, for example, the matching mask is applied between times t=100 and t=200. A value of 1 indicates that the segment is marked as matching the first fault waveform template (such as a current spike template), and other times... A value of 0 indicates that the model considers local waveform features similar to the "current spike" template to exist within the time interval of 100-200 sampling points. If multiple adjacent windows successfully match the same template, the time intervals corresponding to these windows are merged into a continuous matching segment, and the corresponding values ​​within this continuous segment are... Set uniformly as To avoid excessive amplification caused by repetitive enhancement. In one embodiment, for example, the waveform is scanned using a window of length 64 and a step size of 8. Three consecutive windows, 20-83, 25-89, and 30-95, successfully match the "voltage dip" template. The system detects that these windows overlap and are consecutive in time, and therefore merges the smallest interval [20, 95] covered by these windows into a single consecutive matching segment. Then, for... All of them will have the corresponding matching mask Set to 1. Based on this, it is possible to ensure complete enhancement of the entire fault process (which may span multiple scan windows), rather than only performing discontinuous enhancements on the window edges that may introduce artifacts.

[0045] 4> Extract local detail components within the matching segment and perform waveform sharpening. Specifically, to ensure the stability of the enhancement process, the entire waveform is not directly amplified; instead, the data matrix is ​​first decoupled from the physical domain. Local detail components are extracted, and then only these detail components are weighted and enhanced. In one implementation, a short-window moving average is first applied to each channel to obtain a locally smoothed baseline. , Indicates the first The first channel in the The local smoothed baseline at each time point is used to characterize low-frequency trends without spikes or abrupt changes, where... This represents the channel index, with a value range of [value range missing]. to Then, the difference between the original signal and the local smoothed baseline is used as the local detail component, and sharpening intensity is applied to the segments marked by the matching mask. Enhancements will be made. Among them, This represents the waveform sharpening intensity coefficient, used to control the extent of detail magnification within the matching section. An example value is shown below. Based on this, the normal background section remains basically unchanged, while the mutation details similar to the fault template are amplified in a targeted manner.

[0046] In one embodiment, for example, for the c-th channel (such as the ohmic voltage drop channel) The original signal exhibits a weak voltage spike with an amplitude of only 0.02V within a matching interval from t=50 to t=70. First, a smoothed baseline is calculated using a moving average with a window length of 10. The baseline approximates a gentle curve in the interval from t=50 to t=70. Then, detail components are extracted. This detail component clearly shows a spike of approximately 0.02V in the range of t=50 to t=70. Finally, within the segment marked by the matching mask, the detail component is multiplied by the sharpening intensity factor. The enhanced waveform is obtained. Based on this, the original 0.02V peak was amplified to 0.026V, enhancing its significance, while the baseline... Remain unchanged. For non-matching segments, In other words, the waveform remains unchanged. Therefore, weak fault spikes are amplified in a targeted manner, while the background noise remains unchanged.

[0047] 5> Output waveform sharpening data matrix (Can be used as the final training sample). This represents multi-channel timing data after partial fault waveform enhancement, with a size of [size missing]. Its four channels are sequentially decoupled from the physical domain data matrix. To maintain consistency, this data matrix enhances transient fault characteristics such as short-term spikes, local dips, and slope anomalies while keeping the overall operating condition distribution unchanged, making it easier for subsequent networks to more stably identify contact anomalies and transient electrical faults.

[0048] It should be noted that the embodiments of the present invention adopt a combined processing method of "template positioning + local detail enhancement" instead of uniformly sharpening the entire time series. This is because the fault waveform of the battery swapping equipment usually only appears in a local time segment. If the entire sequence is uniformly enhanced, it is easy to amplify normal fluctuations and noise at the same time. The present invention first uses the fault waveform template to locate the suspicious segment, and then only enhances the local details in the matching segment. This can significantly improve the saliency of the fault features, while suppressing the amplification of background noise. It is especially suitable for fault modes with short duration and variable location, such as poor contact, relay jitter, and local thermal anomalies.

[0049] Step S204: Decouple the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data.

[0050] During the charging, discharging, and plugging / unplugging processes of electric bicycle battery swapping equipment, current, voltage, and temperature data are simultaneously affected by load changes, battery aging, connection status changes, and ambient temperature disturbances. The raw sampling results contain both fault information and a large amount of normal operating condition coupling information. If the raw multi-source time-series data is directly input into a classification model, the model may easily misidentify normal coupling fluctuations as fault features or ignore short-duration contact anomalies.

[0051] In one implementation, the battery pack in the battery swapping device can be characterized using a first-order resistor-capacitor equivalent circuit in fault identification scenarios. The terminal voltage can be decomposed into open-circuit voltage, ohmic voltage drop, polarization voltage drop, and measurement noise. Compared to directly using the original voltage waveform, this decomposition method can correspond connector loosening, contact oxidation, increased polarization, and temperature rise lag to different physical components, thereby reducing the learning difficulty of subsequent models. Each voltage component can be represented by the following steps: 1) Based on the ohmic internal resistance of the equivalent circuit model of the battery swapping equipment during the charging process, determine the ohmic voltage drop component in the voltage data.

[0052] The real-time charging process data collected by the battery swapping equipment can be characterized as a raw multi-source time-series data matrix. , Characterizes the raw multi-source time series data within a single sampling window, with a size of ,in This indicates the number of sampling points, with examples of possible values. ; The first line is the original current sequence. The second line is the original voltage sequence. The third line is the original temperature sequence. ,in, Represents the discrete time series point index, with a value range of 100. to Furthermore, the three channels are first aligned according to a unified timestamp, and linear interpolation or forward padding is performed on missing points (if missing values ​​exist) to ensure that the current, voltage, and temperature values ​​at the same time come from the same sampling time.

[0053] For the ohmic voltage drop component The estimated ohmic internal resistance can be used as a basis. Calculation, the calculation method can be characterized as .in, Indicates the first The instantaneous voltage drop caused by the ohmic internal resistance at a given moment is an important response quantity for faults such as connector loosening, contact oxidation, and unstable insertion / removal contact.

[0054] 2) Based on the polarization internal resistance and polarization capacitance of the equivalent circuit model of the battery swapping equipment during the charging process, determine the polarization voltage drop component in the voltage data.

[0055] For the polarization voltage drop component, the ohmic internal resistance can be estimated in real time using a sliding window state estimation method. and polarization resistance And recursively calculate the polarization voltage drop components. ,in, Indicates the first The internal resistance in ohms at any given moment is used to reflect the instantaneous conduction state inside the connector, wires, contacts, and battery. Indicates the first The polarization internal resistance at each moment is used to reflect the intensity of the battery polarization effect. Indicates the first The polarization voltage drop component at each moment is used to characterize the hysteresis voltage change caused by electrochemical polarization.

[0056] In the specific implementation, the original multi-source time-series data matrix is ​​processed segment by segment using a fixed-length sliding window. The window length can be set to... , or There are 10 sampling points. Within each sliding window, discrete state equations and observation equations are established using open-circuit voltage components, ohmic internal resistance, polarization internal resistance, and polarization voltage drop as state variables. Kalman filtering or extended Kalman filtering is then used for recursive estimation. and Polarization voltage drop component The recursive calculation method is expressed as follows:

[0057] in, Indicates the sampling interval, with examples of possible values. Second; This represents the polarization time constant, used to characterize the polarization response rate. Examples of its values ​​are as follows: ; Indicates the first The polarization voltage drop component at each moment; It is a natural constant. Preferably, It can be initialized to , and The initial value can be the calibration value or the state at the end of the previous window, thus ensuring the continuous and stable estimation process.

[0058] 3) Based on the state of charge data obtained by integrating the current data in ampere-hours during the charging process of the battery swapping equipment, and the temperature data during the charging process, determine the open-circuit voltage component in the voltage data.

[0059] Based on the original current sequence and the original temperature sequence Estimating open-circuit voltage components , Indicates the first The open-circuit voltage value corresponding to the intrinsic electromotive force of the battery at each moment is used to characterize the battery voltage baseline unaffected by instantaneous load voltage drop. Among them, the open-circuit voltage component... The estimation is based on the electrochemical principle that the open-circuit voltage of a battery in equilibrium is a deterministic function of its state of charge and temperature. First, real-time acquired current sequences are used... The state of charge at each time step is obtained recursively using the ampere-hour integral method. ,in, This indicates the battery's nominal capacity. Indicates the battery at time The state of charge reflects the ratio of the current remaining capacity to the nominal capacity. This represents the initial state of charge at the start of integration. Then, based on the "" calibrated beforehand through constant current intermittent titration or slow charge-discharge experiments... The "temperature-open-circuit voltage" three-dimensional characteristic surface mechanistically describes the relationship between the electrode equilibrium potential and the lithium-ion solid phase concentration and temperature. For any given time... ,Depend on and measured temperature The intrinsic electromotive force of the battery, which is unaffected by load voltage drop and polarization hysteresis, can be determined by referring to a table. This separates the battery's own state changes from the measured terminal voltage.

[0060] In one implementation, the state of charge (SOC) is first estimated using the ampere-hour integration method. This involves starting with the initial SOC at the sampling point and updating the SOC point-by-point using the current integral. The initial SOC can be obtained from a table of voltage values ​​under static conditions or from the equipment's factory calibration parameters. Then, using a three-dimensional lookup table of SOC-temperature-open-circuit voltage, bilinear interpolation can be used to obtain the open-circuit voltage component at each moment. In one implementation, the calibrated three-dimensional lookup table contains only a limited number of discrete grid points, while during the charging process... With continuous temperature changes, directly taking the nearest neighbor value will introduce a step error, masking the true temperature. A slow changing trend. Bilinear interpolation can be used to correlate temperature with... Linearly weighted neighboring grid values ​​in two dimensions yield a smooth and physically continuous result. The sequence. During sample construction, the slowly varying baseline extracted, correlated with the intrinsic state of the battery, is precisely this continuous, smooth sequence. The curve reflects the slow change of intrinsic electromotive force caused by increased discharge depth and temperature drift. After physical domain decoupling, this baseline is fully incorporated into the open-circuit voltage path and removed from the total voltage, allowing the ohmic voltage drop component characterizing the device circuit to be fully represented. No intrinsic battery changes remain, and the fault diagnosis model only needs to... The independent changes are identified, thus fundamentally avoiding misjudging slow, normal intrinsic electromotive force drift as equipment failures such as increased contact impedance.

[0061] In one embodiment, for example, when an electric bicycle is initially charging, the battery's state of charge (SOC) is 30%, at which point the intrinsic electromotive force (open-circuit voltage) is approximately 36V. When the charging station starts, a 10A current is applied instantaneously. Due to the ohmic internal resistance (e.g., 50mΩ), a 0.5V ohmic voltage drop is generated, causing the sampled terminal voltage to jump instantaneously from 36V to 36.5V. Without physical domain decoupling, the model would directly treat this 0.5V voltage jump as a fault characteristic. However, in reality, the open-circuit voltage component estimated through this step... It remains around 36V, thus distinguishing normal voltage fluctuations caused by load changes from faults such as loose connectors (the latter will cause changes in ohmic internal resistance, which in turn will cause different voltage fluctuation patterns), and avoiding misjudging normal charging start-up as a contact fault.

[0062] Step S206: Construct a multi-channel decoupling matrix based on voltage component, current data, and temperature data.

[0063] This multi-channel decoupling feature can be characterized as a physical domain decoupling data matrix. , This represents multi-channel timing data after physical domain decoupling. In this embodiment of the invention, the size is... . The first line is the original current sequence. The second line contains the ohmic voltage drop component. The third line is the polarization voltage drop component. The fourth line is the original temperature sequence. In summary, by decomposing the original voltage waveform into ohmic voltage drop and polarization voltage drop while retaining the original current and temperature, different channels can be respectively associated with current surges, contact impedance changes, electrochemical hysteresis, and thermal state accumulation. When the model performs fault determination on the multi-channel decoupling matrix, it no longer needs to automatically separate different physical factors within the same voltage channel, thus improving the stability and interpretability of fault identification.

[0064] It should also be noted that battery swapping equipment failures often occur simultaneously with sudden changes in connection impedance and increased polarization. These two factors are highly superimposed in the original voltage waveform. This invention does not directly perform statistical filtering on the original voltage sequence, but first performs physical decomposition based on the battery equivalent circuit of the battery swapping equipment, and then reassembles the decomposition results into a physical domain decoupled data matrix. By separating and then modeling, contact faults, aging faults and thermal faults can be mapped to different feature channels, thereby significantly reducing the interference of normal operating condition coupling fluctuations on model training and enhancing the model's separability for poor contact and aging faults.

[0065] In one embodiment, Figure 3The four channels after physical domain decoupling are shown: current, ohmic voltage drop, polarization voltage drop, and temperature, with time ranges equal to... Figure 2 Consistent. The original voltage is decomposed into ohmic voltage drop and polarization voltage drop using a first-order equivalent circuit model, separating contact faults (abrupt ohmic voltage drop) and electrochemical hysteresis (polarization voltage drop) into different channels. After decoupling, the physical meaning of each channel is clear, which is beneficial for subsequent models to learn different types of faults separately. The horizontal axis represents time (seconds), and the vertical axes represent current (A), ohmic voltage drop (V), polarization voltage drop (V), and temperature (°C), respectively.

[0066] Step S208: Input the data of each channel of the multi-channel decoupling matrix into the dual-path time-frequency sensing neural network in parallel. Extract the high-frequency feature representation and low-frequency feature representation in the multi-channel decoupling matrix through the dual-path time-frequency sensing neural network to determine the fused feature representation contained in the multi-channel decoupling matrix.

[0067] The fault identification model used is trained with optimal performance, and can be deployed on actual operating battery swapping equipment to perform online fault identification based on real-time collected data. The identification process begins at the battery swapping cabinet control unit, which continuously acquires real-time current, voltage, and temperature data of the battery pack during charging and discharging at a sampling interval of 0.1 seconds. Whenever the cumulative number of sampling points reaches the preset window length of 4096, the system assembles these data into a complete multi-source time-series data matrix as the current sample to be identified.

[0068] Among the faults in battery swapping equipment, there are both high-frequency transient faults such as poor contact and relay jitter, and low-frequency gradual faults such as capacity decay, heat dissipation degradation, and temperature baseline drift. Single-path neural networks often only favor one type of fault feature, leading to decreased stability in identifying the other type of fault. This invention constructs a dual-path time-frequency sensing neural network, with one path specifically extracting high-frequency transient fault features and the other path specifically extracting low-frequency gradual fault features. These two types of features are then dynamically fused together based on the current operating conditions of the battery swapping equipment to obtain the final fused feature vector used for classification. In specific implementation, the dual-path time-frequency sensing neural network receives the same multi-channel decoupling matrix at its front end, and then splits into a high-frequency transient path and a low-frequency gradual path: the high-frequency path consists of a pyramid composed of multiple parallel dilated convolutional branches, each branch using an exponentially increasing dilation rate to capture multi-scale transient details ranging from extremely short spikes to brief oscillations, and obtains a high-frequency feature map through cross-branch attention aggregation; the low-frequency path constructs a time-series feature pyramid through cascaded downsampling and generates a low-frequency feature map by fusing long-range trend information. Each of the two branches focuses on a type of physical time-frequency attribute. Ultimately, a dynamic fusion module based on operational condition awareness adaptively weights and concatenates the two types of features, achieving structured fault feature differentiation and coordination. Correspondingly, Figure 4 A schematic diagram of the fault prediction process is shown. (Refer to...) Figure 4 The corresponding process can be represented by the following steps: 1) Extract trend features of different time spans in the multi-channel decoupling matrix, and then pass and fuse the trend features step by step to determine the low-frequency feature representation in the multi-channel decoupling matrix.

[0069] Low-frequency gradual faults typically manifest as a slow shortening of the voltage plateau, a slow rise in the temperature baseline, and a gradual increase in polarization effects. These faults rely on trend changes over a relatively long time span. To simultaneously consider both local and global trends, this invention employs a temporal feature pyramid network to construct multi-layer time-scale representations and fuses them from top to bottom to form a low-frequency gradual feature map. Furthermore, to prevent low-frequency paths from being constrained by prior knowledge of high-frequency paths, this invention directly uses the input feature map... (Multi-channel decoupling matrix) is used as input, instead of high-frequency transient feature maps. As input, this ensures that low-frequency paths focus on extracting long-range trend information. The top-down fusion sequence representation starts from the highest layer (largest receptive field, lowest temporal resolution) and progressively propagates the fusion to shallower layers. Based on this, the global long-range trend semantic information extracted from deeper layers can be upsampled and reinjected into the shallow high-resolution feature map. The shallow layer gains contextual awareness over a larger time span while retaining its local detail localization capabilities, resulting in a final low-frequency gradient feature map. It can consistently express the complete trend of slow failures such as capacity decay and temperature accumulation.

[0070] The specific steps are as follows: a- Input feature map For input, construct Layered time sequence pyramid This represents the number of levels in the time-series pyramid, with examples of possible values. ;No. The layered temporal pyramid feature map is denoted as .in, This represents a hierarchical index, with a value range of [value range missing]. to In its specific implementation, the time-series pyramid has the following structure (using...). (Taking layers as an example): Layer 1: Input feature map After passing through a one-dimensional convolution with a kernel size of 3, a stride of 1, and "same" padding, the output feature map is obtained. The size is ,in, The number of feature channels in the time-series pyramid can be set to AND. Same; Level 2: Will By downsampling using a one-dimensional convolution with a stride of 2 (kernel size of 3), the time duration is halved, resulting in... The size is ; Level 3: Will By performing another one-dimensional convolution with a stride of 2, we obtain The size is ; Level 4: will By performing a third one-dimensional convolution with a stride of 2, we obtain The size is Each convolutional layer is followed by a batch normalization and ReLU activation function to extract trend features over the corresponding time span. Focus on local, detailed trends, and Then focus on the overall trend across the entire time window.

[0071] b - Perform top-down multi-scale fusion starting from the highest layer. Initially, the highest-layer temporal pyramid feature map can be used... As the top-level fusion feature, the fusion result of the previous layer is then upsampled to the time length of the next shallow layer using linear or bilinear interpolation, and then added or concatenated point by point with the temporal pyramid feature map of the next shallow layer before processing. Convolutional fusion then transmits long-term trend information from higher layers back to lower layers, allowing the lower layers to gain a larger temporal receptive field while maintaining detailed positioning capabilities.

[0072] c- Perform convolutional refinement on the fusion results at each level to suppress spurious changes introduced by upsampling. Specifically, after each fusion level is completed, a convolutional kernel with a size of [missing value] is applied. The one-dimensional convolution is smoothed and reshaped to further align the feature distributions from different levels. Based on this, the shallow features retain both high temporal resolution and can also incorporate the long-range trends of higher levels.

[0073] d- Use the lowest-level fusion result as a low-frequency gradient feature map .in, This represents the feature map of the final output of the low-frequency path. This feature map primarily characterizes gradual fault modes such as capacity decay, thermal degradation, temperature rise accumulation, and voltage plateau changes. The size is... ,in, This indicates the number of output channels for the low-frequency path. It should be noted that gradual faults are often not determined by a single instantaneous point, but by a slow trend over a long time span. This invention expands the time coverage through multi-level downsampling, and then brings the long-term trend back to the high-resolution layer through top-down fusion. This allows the model to see both the "long-term trend" and the "specific location of occurrence" at the same time, thereby significantly improving the ability to identify aging faults, thermal faults, and slow drift faults.

[0074] 2) Perform multi-branch parallel dilated convolution on the multi-channel decoupling matrix to determine the transient fault information of the multi-channel decoupling matrix at the corresponding time scale, and perform cross-branch attention aggregation on the transient fault information to determine the high-frequency feature representation in the multi-channel decoupling matrix; High-frequency transient faults typically manifest as short-term spikes, jitters, oscillations, and sharp drops, characterized by short timescales and large local amplitude variations. To simultaneously account for high-frequency anomalies of varying durations, this invention employs a multi-scale dilated convolutional pyramid to extract local abrupt features at different timescales in parallel, followed by adaptive aggregation using cross-branch attention. The specific steps are as follows: a- Input the data matrix (multi-channel coupling matrix) into the channel mapping layer to obtain the input feature map. . This represents the basic feature map entering the dual-path neural network, with a size of [size missing]. ,in, This represents the number of channels after mapping, with examples of possible values. or In its implementation, the channel mapping layer has the following structure: One-dimensional convolution: the input is... (4×L), using A 1×1 convolutional kernel with a stride of 1 and padding of 0, and an output size of [missing value]. ×L, this convolution operation maps the four physical channels linearly to a single value without changing the time length. Individual feature channels; batch normalization: for the convolutional... Each feature channel is batch normalized to accelerate convergence and stabilize training; Activation function: LeakyReLU activation function is used to introduce non-linearity, and its negative half-axis slope is set to 0.01 to retain some negative features.

[0075] b - Input feature map For input, build Several parallel dilated convolution branches are used to obtain the high-frequency branch feature maps of each branch. . Indicates the first The high-frequency feature map output by each dilated convolution branch has a size of [size missing]. This is used to characterize transient fault modes at the corresponding time scale. This indicates the number of high-frequency branches (i.e., parallel dilated convolution branches), with examples of possible values. ;No. Each branch contains two layers of one-dimensional dilated convolutions, with the first layer having a dilation rate of [value missing]. The expansion rate of the second layer is taken as ;when When, the branch focuses more on the extremely short transient peak; when As the size increases, the receptive field of the branch expands, making it more suitable for extracting oscillatory faults with slightly longer durations. In practical implementation, the size of each convolutional kernel can be set to... Using the same length filling method, the time length remains the same. Each convolutional layer is followed by batch normalization and LeakyReLU activation. Based on this, it is possible to cover a variety of fault timescales, from short-term spikes to longer high-frequency oscillations, without significantly increasing the number of parameters.

[0076] c - Feature maps of each high-frequency branch Perform cross-branch attention aggregation. Specifically, to generate branch weights, first process the high-frequency branch feature maps of each branch. Perform global max pooling along the time dimension to obtain the branch global feature vector. , Indicates the first The overall response strength of each branch over the full time window is used to reflect the branch's explanatory power for the current sample; then, Input a lightweight fully connected network, output the first... The attention scores of each branch are then normalized using Softmax to obtain the branch weights. , Indicates the first The aggregate weight of each branch, with a value range of [value missing]. arrive And satisfy the condition that the sum of the weights of all branches is In its implementation, the lightweight fully connected network has the following structure: Input: Global feature vector of each branch. Its dimensions are (Same as the number of channels in the input feature map). First fully connected layer: Mapping to a smaller dimension, for example (But not less than 8), to compress information, the activation function is ReLU activation function. Second fully connected layer: The output of the first fully connected layer is mapped to 1 neuron to obtain the unnormalized attention score of this branch.

[0077] d- Feature maps of each high-frequency branch according to branch weight Weighted summation yields the high-frequency transient feature map. . This represents the feature map of the final output of the high-frequency path. This feature map primarily preserves high-frequency fault modes such as poor contact, relay bounce, and transient current surges. Its size is... ,in, Indicates the number of output channels for the high-frequency path, set with... same.

[0078] 3) Based on the operating status information of the charging process of the battery swapping equipment, the high-frequency feature representation and the low-frequency feature representation are adaptively weighted and fused to obtain the fused feature representation contained in the multi-channel decoupling matrix.

[0079] The importance of high-frequency transient fault characteristics and low-frequency gradual fault characteristics varies under different operating conditions for battery swapping equipment. For example, high-frequency contact abnormalities are more pronounced under frequent plugging and unplugging and high-current switching conditions; while low-frequency heat accumulation and aging trends are more pronounced under continuous operation and high ambient temperature conditions.

[0080] This invention dynamically generates fusion weights based on the current operating conditions of the samples, enabling the model to adaptively focus on more relevant fault features for the current operating conditions. The specific steps are as follows: a- Perform global average pooling and global max pooling on the high-frequency feature representation and the low-frequency feature representation respectively to determine the high-frequency pooled feature vector corresponding to the high-frequency feature representation and the low-frequency pooled feature vector corresponding to the low-frequency feature representation.

[0081] High-frequency transient feature maps and low-frequency gradient feature map Perform global average pooling and global max pooling, and concatenate the results of the two pooling methods to obtain the high-frequency pooling feature vector. and low-frequency pooling feature vectors , This represents the overall statistical characteristics of high-frequency paths. It represents the overall statistical characteristics of low-frequency paths, thus simultaneously retaining two types of information: average response intensity and the location of the strongest response.

[0082] In practical implementation, for high-frequency transient feature maps (size Perform global average pooling and global max pooling along its time dimension (L), respectively. Global average pooling yields... For a dimensional vector, global max pooling also yields... Two vectors are then concatenated along the channel dimension to obtain... High-frequency pooling feature vectors of dimension Similarly, operations on low-frequency paths yield... Its dimensions are .

[0083] b- Based on the adaptive fusion weights indicated by the operating condition information, the high-frequency pooling feature vector and the low-frequency pooling feature vector are weighted and summed to obtain the fusion feature representation contained in the multi-channel decoupling matrix.

[0084] Among them, it can be obtained from the original multi-source time series data matrix Extract statistical features of operating conditions and construct a feature vector of operating conditions. . This represents the statistical description vector of the current sample's operating conditions, reflecting the equipment's operational status within the sampling window. In one implementation, the statistical features of the operating conditions include the mean current rate of change, the standard deviation of the current rate of change, the RMS current, the peak current, the mean temperature rate of change, the maximum temperature rate of change, and the voltage fluctuation intensity (which needs to be normalized to...). (interval), where the rate of change of current can be obtained from the difference between adjacent sampling points, and the rate of change of temperature can be obtained from the original temperature sequence. The effective value of the current can be calculated using the root mean square method after differential calculation.

[0085] In one implementation, the operating condition feature vector can be... Input the operating condition coding network to obtain the operating condition embedding vector. , This represents the working condition representation vector after nonlinear mapping, used in subsequent dynamic weight generation. In its specific implementation, the working condition coding network has the following structure: First fully connected layer: Input is the normalized working condition feature vector. The first layer fully connected network will Mapped to Dimension, among which, Indicates the first mapping dimension, for example The activation function used is ReLU. Dropout layer: A Dropout layer is added after the first layer to randomly drop a certain percentage (e.g., 0.2) of neurons to prevent overfitting. Second fully connected layer: Maps the output of the Dropout layer to... Dimensions. Among them, Indicates the second mapping dimension, for example, The ReLU activation function is used. Output: The output of the second fully connected layer is the job condition embedding vector. , dimension .

[0086] Furthermore, the high-frequency pooling feature vectors Low-frequency pooling feature vector and working condition embedding vector Channel-based concatenation, input to dynamic fusion network, output high-frequency fusion weights. and low-frequency fusion weights , This indicates the proportion of the high-frequency path's contribution to the final decision. This represents the proportion of the low-frequency path's contribution to the final decision; both values ​​are taken from... arrive Between, and satisfy In its implementation, the dynamic fusion network has the following structure: Input concatenation: High-frequency pooling feature vectors are concatenated. (dimension) ), low-frequency pooling feature vector (dimension) ) and working condition embedding vector (dimension) ) Segmented by channel to form a dimension The joint feature vector is input into a fully connected layer containing two output neurons, corresponding to the unnormalized scores of the high-frequency and low-frequency paths, respectively. Softmax activation: The two unnormalized scores are input into a softmax function to obtain the final fused weights. and .

[0087] In summary, high-frequency pooling of feature vectors and low-frequency pooling feature vectors Map each feature vector to the same dimension, then sum them according to their respective fusion weights to obtain the fused feature vector. . This represents the final fused feature vector used for fault classification. This vector adaptively adjusts the contribution ratio of high-frequency transient fault features to low-frequency gradual fault features based on the current operating conditions, enabling the model to focus on the most relevant fault information even under complex operating conditions. Its size is [size missing]. ,in, This represents the dimension of the fused features, with examples of possible values. In practice, dimension mapping is achieved by setting a fusion feature dimension, for example, 128, specifically using a fully connected layer. Will (dimension) ) mapped to dimensional vector Then use another fully connected layer Will (dimension) ) mapped to dimensional vector Then, the mapped vector and The final fused feature vector is obtained by weighted summation based on the weights generated by the dynamic fusion network. .Right now: .in, and It is a scalar, multiplied with a vector through a broadcast mechanism.

[0088] It should be noted that the saliency of fault characteristics of battery swapping equipment is strongly correlated with operating conditions. Fixed fusion strategies cannot take into account the significant differences in operating conditions such as frequent plugging and unplugging, high current load, continuous operation and high temperature environment. This invention introduces operating condition information into dual-path feature fusion instead of fixed equal weight fusion or static channel attention. Through operating condition perception dynamic fusion, the same model can automatically adjust the focus under different operating conditions, which significantly improves the generalization ability in complex operation scenarios.

[0089] Step S208: Based on the pre-built fault waveform template, perform fault waveform matching on the fused feature representation to determine the fault category to which the current charging process of the battery swapping equipment belongs.

[0090] In one implementation, the corresponding fused feature vector (fused feature representation) is fed into a classifier. Based on the classification surface learned during training, which has adaptive boundary margins for the categories, the classifier calculates the logical score of the sample belonging to each fault category, combining this with the matching results of the fault waveform template. The system applies a Softmax function to the logical score, converting it into a probability distribution, and uses the category with the highest probability as the final fault identification result for that sampling window. If the identification result is a fault, the system immediately reports the fault type, occurrence time, and key feature data (such as sudden changes in ohmic internal resistance, abnormal temperature rise rate, etc.) to the cloud-based operation and maintenance management platform, triggering corresponding alarms and handling procedures, such as notifying maintenance personnel to conduct on-site inspections or remotely restricting the continued use of the battery pack. If the identification result is normal, the system only records the summary information of this identification and continues processing the data for the next sampling window, thereby achieving continuous and real-time monitoring of the battery swapping equipment's operating status.

[0091] Furthermore, in the actual operating data of battery swapping equipment, the number of normal samples is usually far greater than the number of fault samples, and there are significant imbalances and differences in similarity between different fault categories. For example, poor contact and undervoltage faults may both manifest as voltage drops during certain periods, making them easily confused; overtemperature faults and aging faults may have overlapping characteristics in long-term trends. If ordinary cross-entropy loss is directly used for training, the model tends to favor the category with the larger number of samples and is insufficient in learning the boundaries between similar categories.

[0092] This invention introduces class adaptive boundary margin, dynamic scaling factor, and class weights during the training phase to enhance the learning ability for minority class faults and easily confused faults. Correspondingly, this can be represented by the following steps: 1) Calculate the category boundary margin of the fault diagnosis model for battery swapping equipment based on the physical mechanism similarity between different fault categories in the training samples.

[0093] Different fault categories exhibit varying degrees of physical similarity; the more similar the categories, the more deliberately the classification boundaries need to be widened. This invention first constructs category similarity based on prior fault physics and statistical analysis of training data. Then, it assigns different boundary margins to each fault category, enabling the model to proactively learn larger separation intervals for easily confused categories during training. The specific steps are as follows: 1> Fuse feature vectors Input the classifier to obtain the category logical output vector. , This represents the unnormalized discrimination score of the current sample for each fault category, with dimension 1. ,in, This represents the total number of fault categories. In its implementation, the classifier uses a "fully connected layer + batch normalization" structure, specifically: Fully connected layer: Input is the fused feature vector. (dimension) This fully connected layer will Mapping to the number of categories dimensional vector (i.e., the category logical output vector); Batch normalization (optional): Before the activation function, normalize the logical output... Batch normalization helps stabilize the training process, especially when there are many classes, and can prevent gradient vanishing or exploding.

[0094] 2> Construct a category prototype vector for each fault category and calculate the category similarity matrix.

[0095] Specifically, first, calculate the mean physical characteristics corresponding to each fault category on the training set, including ohmic resistance statistics, polarization resistance statistics, temperature rise slope, voltage plateau length, and current surge amplitude (which need to be normalized to...). The interval is used to form the category prototype vector for that category; then, the cosine similarity between any two category prototype vectors is calculated to obtain the category similarity matrix. In one embodiment, as an example, assume there are three fault categories: The prototype vector of poor contact is ; The prototype vector of undervoltage fault is ; The prototype vector of aging failure is .

[0096] Then, calculate the cosine similarity, defining... The cosine similarity function is expressed as: (Very high); (medium); (Lower).

[0097] So, for the category "poor contact", its set of neighboring categories... Includes the most similar (For example There are 10 categories, namely "undervoltage fault" (similarity 0.95).

[0098] Among them, for a certain category The nearest neighboring categories with the highest similarity can be selected from the category similarity matrix to form the nearest neighbor category set for that category. . Representation and Category A set of categories that are similar in physical mechanism or data distribution.

[0099] Furthermore, it can be based on the nearest category set. And class similarity, calculate class boundary margin for each class. . Indicates category The discriminant margin attached to the true class during training is used to actively widen the classification boundary between that class and similar classes. Its calculation method is as follows:

[0100] in, This represents the basic boundary margin, with examples of possible values. ; This represents the boundary margin adjustment coefficient, with examples of possible values. ; Indicates category With category The higher the similarity between categories, the easier it is for the category to be confused with surrounding categories, and the larger the corresponding boundary margin. Represents the set of neighboring categories The number of categories.

[0101] 2) Based on the sample classification difficulty corresponding to the class boundary margin, construct a dynamically scaled classification loss.

[0102] In long-tailed training sets, minority class samples are easily overwhelmed by majority class samples, and the classification difficulty varies among different samples. This invention reweights the loss from two perspectives: "class imbalance" and "sample difficulty imbalance," making the training process pay more attention to minority class samples and boundary samples. The specific steps are as follows: 1> Count the number of samples for each fault category in the training set and calculate the category weights. , Indicates category The training weights are used to balance the training bias caused by differences in the number of samples from different classes. In one implementation, the class weights are calculated as the reciprocal of the square root of the number of samples, i.e. .in, Indicates category The number of samples in the training set. Further, divide all class weights by the average, i.e. To avoid excessive overall losses.

[0103] 2> For each training sample, calculate the sample difficulty based on the true class probability adjusted for boundary margin. Specifically, for the ... Let there be a sample, and denote its true category as . The true class prediction probability after boundary margin adjustment is denoted as ,in The smaller the value, the more difficult the sample is to classify, and the more attention it should receive during training. The calculation method is expressed as follows:

[0104] in, Indicates the first The original logical output vector of each sample After adjusting the boundary margin, corresponding to the first Logical scores for each category. Indicates the first The original logical output vector of each sample After adjusting the boundary margin, corresponding to the first Logical scores for each category.

[0105] 3> Calculate the dynamic scaling factor based on the sample difficulty , Indicates the first The difficulty scaling factor for each sample is used to control the degree of focus on that sample in the loss, and is calculated as follows:

[0106] in, This represents the base scaling factor, with examples of possible values. ; This represents the adaptive adjustment coefficient, with examples of possible values. When the sample is more difficult to classify, The smaller, the corresponding The larger the value, the greater the contribution of that sample to the loss.

[0107] 4> Based on category weight and dynamic scaling factor Calculate the classification loss under dynamic scaling . This represents the classification loss used in long-tail data scenarios. Its core idea is to further increase the proportion of loss for minority class samples and hard-to-classify samples on top of ordinary cross-entropy. Specifically, for each sample, its true class log probability is multiplied by the class weight. and difficulty adjustment items Then, the average of all samples within a batch is calculated to obtain... The calculation method is expressed as follows:

[0108] in, To dynamically scale the classification loss, Indicates the first The true category of each sample The corresponding final category weight, This indicates the number of samples included in the training batch.

[0109] It should be noted that this invention does not simply increase the static weights of the minority class, but rather considers both class imbalance and sample difficulty, so that the model is neither dominated by majority class samples nor ignores those key samples that are most likely to be misjudged and located near the decision boundary, thereby improving the ability to identify rare and compound faults.

[0110] 3) Based on the boundary margin loss and dynamic scaling classification loss corresponding to the category boundary margin, construct the total loss function of the fault diagnosis model of the battery swapping equipment.

[0111] a- Regarding boundary margin loss: During the training phase, the true class score of each sample is subtracted from the corresponding class boundary margin. Then, calculate the class probability using the Softmax method, and calculate the boundary residual loss accordingly. . This represents a classification loss with adaptive class boundary constraints, which forces the model to learn clearer decision boundaries for easily confused classes. Based on this, the model no longer simply pursues a true class score greater than other classes, but requires the true class score to exceed other classes by a safe interval related to class similarity.

[0112] The true class score is the output vector from the class logic. Extracted from, if the true label of the current sample is a category Then the classifier output The first vector element This refers to the true class score of the sample. In this embodiment of the invention, the boundary margin loss... Its purpose is to proactively reduce the prediction scores of easily confused categories when calculating the classification loss, forcing the model to learn clearer decision boundaries for these categories. The specific calculation method is as follows: For each sample, the original logical output vector of the classifier is obtained. (dimension is) Then, based on the true category label of the sample. Construct an adjusted logical output vector Specifically, will Corresponding to the real category Subtract the boundary margin corresponding to the category from the element. Elements in other categories remain unchanged, that is ; in, Represents the adjusted logic output vector The Middle The value of each element, It is the original logic output vector The Middle The value of each element, This is the true class label of the sample. It is a category The boundary margin. Then, the adjusted logic output vector. Applying the Softmax function, we obtain the adjusted class probability distribution. , represented as ;in, This indicates that the adjusted sample belongs to the first... The probability of each category, since the score of the true category is subtracted. ,therefore It will be less than or equal to the original probability. Then, calculate the boundary margin loss. It is based on the adjusted probability distribution The purpose of calculating the cross-entropy loss is to ensure that the model, during training, not only reflects the raw score of the true class, but also... As high as possible, and more importantly, ensure it scores at least higher than all other categories. , is represented as: .

[0113] in, For boundary margin loss, This indicates that the adjusted sample belongs to its true category. The probability. It should be noted that the difficulty of distinguishing different fault categories is not uniform. Using the same classification interval will result in the boundaries of easily confused categories remaining blurred. This invention introduces a category adaptive boundary margin, which can impose stronger constraints on closely related categories such as poor contact, undervoltage, aging, and overtemperature, thereby significantly improving the separability between difficult-to-separate category pairs. In one embodiment, Figure 5 The graph shows the ohmic resistance as a function of time, estimated using a sliding window Kalman filter. In the faulty contact region (e.g., around 50 seconds), the ohmic resistance shows a momentary increase, consistent with the physical mechanisms of connector loosening or contact oxidation. The horizontal axis represents time (seconds), and the vertical axis represents ohmic resistance (ohms).

[0114] b-Regarding the total loss function: After completing the category boundary design and dynamic scaling design, all the aforementioned modules need to be trained uniformly so that the data preprocessing parameters, feature extraction parameters, dynamic fusion parameters, and classification parameters are optimized collaboratively under the same objective function. The specific steps are as follows: Boundary margin loss Dynamic scaling classification loss The regularization term is combined to form the total loss function. To ensure the integrity of the original design intent, the total loss function includes both boundary constraint terms and long-tail reweighting terms, expressed as:

[0115] in, This represents the boundary margin loss weight, with examples of possible values. ; This indicates dynamically scaling the classification loss weights; examples of possible values ​​are provided. ; This represents the L2 regularization coefficient, with examples of possible values. ; This represents the set of all learnable parameters in the model. Represents all learnable parameters of the model The square of the L2 norm.

[0116] 4) Train the fault diagnosis model of the battery swapping equipment based on the total loss function.

[0117] In one implementation, the model can be trained using a mini-batch approach, with each training batch inputting several original multi-source time-series data matrices. And its corresponding fault labels, batch size can be taken Each batch sequentially undergoes physical domain decoupling, waveform sharpening, high-frequency feature extraction, low-frequency feature extraction, condition-aware dynamic fusion, and classification output to obtain the category logical output vector z and the total loss function for each sample. The calculation results are used to determine the final classification result based on the category logical output vector z. After processing by the Softmax function, z is transformed into the probability distribution of each category, and the category with the highest probability is selected as the model's final classification result for the sample.

[0118] Furthermore, the backpropagation algorithm can be used to calculate the total loss function. The gradients of all learnable parameters are calculated, and the parameters are updated using the Adam optimizer. Preferably, the initial learning rate can be 10. -3 The system performs piecewise decay or cosine annealing decay based on the decrease in loss on the validation set. Furthermore, to prevent training instability, gradient pruning and early stopping strategies can be used. During training, the system continuously monitors changes in classification accuracy, recall, F1 score, and loss on the validation set. When the validation set metrics no longer improve within several consecutive rounds, training is stopped, and the optimal set of model parameters is saved as the final fault identification model.

[0119] In summary, this invention proposes a fault identification method for battery swapping equipment, which is innovative in the following aspects compared to existing technologies: 1. A physical domain decoupling method based on the first-order equivalent circuit model of a battery is proposed. The original voltage waveform is decomposed into open-circuit voltage, ohmic voltage drop and polarization voltage drop, so that fault information of different physical mechanisms such as contact impedance abrupt change and electrochemical polarization hysteresis are separated into independent channels, which reduces the difficulty for the model to learn fault features from the original coupled signal.

[0120] 2. An adaptive waveform sharpening mechanism is proposed. By constructing a fault waveform template library and combining it with dynamic time warping matching, the time segment in which transient faults occur is accurately located, and only the local detail components within the segment are enhanced, which highlights the characteristics of brief and weak faults while avoiding amplifying background noise.

[0121] 3. A dual-path time-frequency sensing neural network was constructed. High-frequency transient fault features were extracted using a multi-scale dilated convolutional pyramid, and low-frequency gradual fault features were extracted using a time-series feature pyramid network. Through a dynamic fusion module for operating condition sensing, the model can adaptively adjust the fusion weights of the two types of features according to real-time operating conditions such as current change rate and temperature change rate.

[0122] 4. A robust classification training strategy combining class adaptive boundary margin and dynamic scaling factor is proposed. Based on the physical mechanism of the fault, a larger classification margin is automatically assigned to easily confused classes. At the same time, the loss function is reweighted from two dimensions: class imbalance and sample classification difficulty, which improves the model's ability to identify minority class faults and boundary samples.

[0123] Furthermore, embodiments of the present invention also provide a fault identification device for battery swapping equipment, referring to... Figure 6The device includes: a data acquisition module 10, used to acquire real-time charging process data generated by the charging compartment during the charging process in response to the detection of a battery pack entering the charging compartment of the battery swapping equipment; the real-time charging process data includes at least current data, voltage data, and temperature data; a data processing module 20, used to decouple the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data; the voltage components include ohmic voltage drop components, polarization voltage drop components, and open circuit voltage components corresponding to the voltage data; a data fusion module 30, used to construct a multi-channel decoupling matrix based on the voltage components, current data, and temperature data; and an execution module 40, used to input the multi-channel decoupling matrix into a pre-constructed battery swapping equipment fault diagnosis model, and to predict the fault of the battery swapping equipment based on the multi-channel decoupling matrix through the battery swapping equipment fault diagnosis model.

[0124] The fault identification device for battery swapping equipment provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device and method embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0125] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figure 1 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figure 1 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 7 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figure 1 The method shown.

[0126] exist Figure 7In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7The diagram uses only a single double-headed arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figure 1 The method shown.

[0127] The computer program product of the battery swapping equipment fault identification method and device provided in this embodiment of the invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, which will not be repeated here. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of this invention, used to illustrate the technical solutions of this invention, and not to limit it. The scope of protection of this invention is not limited thereto. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for fault identification of battery swapping equipment, characterized in that, The method includes: In response to detecting that the battery pack has entered the charging compartment of the battery swapping device, real-time charging process data generated by the charging compartment during the charging process is acquired; the real-time charging process data includes at least current data, voltage data and temperature data. The voltage data in the real-time charging process data is decoupled according to physical effects to obtain multiple voltage components contained in the voltage data; Based on the voltage component, the current data, and the temperature data, a multi-channel decoupling matrix is ​​constructed; The multi-channel decoupling matrix is ​​input into a pre-built fault diagnosis model for battery swapping equipment, and the fault diagnosis model for battery swapping equipment is used to predict faults in the battery swapping equipment based on the multi-channel decoupling matrix.

2. The method according to claim 1, characterized in that, The step of decoupling the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data includes: Based on the ohmic internal resistance of the equivalent circuit model of the battery swapping equipment during the charging process, the ohmic voltage drop component in the voltage data is determined. Based on the polarization internal resistance and polarization capacitance of the equivalent circuit model of the battery swapping equipment during the charging process, the polarization voltage drop component in the voltage data is determined. Based on the state of charge data obtained by integrating the current data in ampere-hours during the charging process of the battery swapping device, and the temperature data during the charging process, the open-circuit voltage component in the voltage data is determined.

3. The method according to claim 1, characterized in that, The fault diagnosis model for battery swapping equipment is constructed based on a dual-path time-frequency sensing neural network; the steps of inputting the multi-channel decoupling matrix into the pre-constructed fault diagnosis model for battery swapping equipment, and using the fault diagnosis model to predict the faults of the battery swapping equipment based on the multi-channel decoupling matrix, include: The data of each channel of the multi-channel decoupling matrix are input in parallel into the dual-path time-frequency sensing neural network. The dual-path time-frequency sensing neural network extracts features from the high-frequency and low-frequency feature representations in the multi-channel decoupling matrix to determine the fused feature representation contained in the multi-channel decoupling matrix. Based on a pre-built fault waveform template, fault waveform matching is performed on the fused feature representation to determine the fault category to which the current charging process of the battery swapping equipment belongs.

4. The method according to claim 3, characterized in that, The step of extracting features from the high-frequency and low-frequency feature representations in the multi-channel decoupling matrix using the dual-path time-frequency sensing neural network to determine the fused feature representation contained in the multi-channel decoupling matrix includes: The trend features of different time spans in the multi-channel decoupling matrix are extracted and the trend features are passed and fused step by step to determine the low-frequency feature representation in the multi-channel decoupling matrix; Multi-branch parallel dilated convolution is performed on the multi-channel decoupling matrix to determine the transient fault information of the multi-channel decoupling matrix at the corresponding time scale, and cross-branch attention aggregation is performed on the transient fault information to determine the high-frequency feature representation in the multi-channel decoupling matrix; Based on the operating status information of the charging process of the battery swapping equipment, the high-frequency feature representation and the low-frequency feature representation are adaptively weighted and fused to obtain the fused feature representation contained in the multi-channel decoupling matrix.

5. The method according to claim 4, characterized in that, The step of adaptively weighting and fusing the high-frequency feature representation and the low-frequency feature representation based on the operating status information of the charging process of the battery swapping equipment to obtain the fused feature representation contained in the multi-channel decoupling matrix includes: Global average pooling and global max pooling are performed on the high-frequency feature representation and the low-frequency feature representation respectively to determine the high-frequency pooling feature vector corresponding to the high-frequency feature representation and the low-frequency pooling feature vector corresponding to the low-frequency feature representation; According to the adaptive fusion weights indicated by the operating condition information, the high-frequency pooling feature vector and the low-frequency pooling feature vector are weighted and summed to obtain the fusion feature representation contained in the multi-channel decoupling matrix.

6. The method according to claim 3, characterized in that, The method further includes: Based on the physical mechanism similarity between different fault categories in the training samples, the category boundary margin of the fault diagnosis model of the battery swapping equipment is calculated. Based on the sample classification difficulty corresponding to the category boundary margin, a dynamically scaled classification loss is constructed. Based on the boundary margin loss and dynamic scaling classification loss corresponding to the category boundary margin, the total loss function of the fault diagnosis model of the battery swapping equipment is constructed. The fault diagnosis model for the battery swapping equipment is trained based on the total loss function.

7. The method according to claim 3, characterized in that, The method further includes: Based on multiple voltage components corresponding to historical charging data of battery swapping equipment, determine the voltage residual sequence of voltage data of historical charging data. Based on the voltage residual sequence, the initial training samples corresponding to the historical charging data are preprocessed to construct preprocessed training samples. The fault matching segments contained in the preprocessed training samples are determined by using a preset fault waveform template matching method. Based on the local detail components in the fault matching section, the preprocessed training samples are waveform sharpened to construct the final training samples.

8. A fault identification device for battery swapping equipment, characterized in that, The device includes: The data acquisition module is used to acquire real-time charging process data generated by the charging compartment during the charging process in response to the detection that the battery pack has entered the charging compartment of the battery swapping equipment; the real-time charging process data includes at least current data, voltage data and temperature data. The data processing module is used to decouple the voltage data in the real-time charging process data according to physical effects to obtain multiple voltage components contained in the voltage data. The data fusion module is used to construct a multi-channel decoupling matrix based on the voltage component, the current data, and the temperature data; The execution module is used to input the multi-channel decoupling matrix into a pre-built fault diagnosis model for battery swapping equipment, and to predict the faults of the battery swapping equipment based on the multi-channel decoupling matrix through the fault diagnosis model for battery swapping equipment.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the battery swapping equipment fault identification method according to any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the fault identification method for battery swapping equipment as described in any one of claims 1 to 7.

Images