A battery replacement cabinet safety control method based on intelligent internet of things

Abstract

The application discloses a battery replacement cabinet safety control method based on an intelligent Internet of Things, relates to the technical field of thermal safety monitoring and control of a battery replacement cabinet, and utilizes a multivariate time sequence anomaly detection model combined with a long short-term memory autoencoder and an attention mechanism to extract deep features from voltage, current and environmental temperature data, effectively identify internal heat accumulation initial signals that are difficult to capture by external sensors, overcome the limitations of traditional fixed threshold monitoring, dynamically learn normal charging and discharging modes, sensitively capture subtle abnormal changes, thereby advance the early warning time, and reduce the risk of missed reports caused by surface temperature lag; a dynamic threshold mechanism is based on quantile numbers of a target false alarm rate and working condition parameters and is adaptively adjusted, so that the system can flexibly cope with complex factors such as different environmental temperatures and battery aging degrees, avoids the false alarm problem caused by a unified threshold, and improves the early warning accuracy.

Description

Technical Field

[0001] This invention relates to the field of thermal safety monitoring and control technology for battery swapping cabinets, and in particular to a safety control method for battery swapping cabinets based on the Internet of Things. Background Technology

[0002] Battery swapping cabinets are widely used in urban mobility and instant logistics. Due to their high-density storage, multi-path parallel charging, and frequent turnover, batteries are more prone to thermal risk accumulation in confined spaces. Existing safety controls are mostly based on external measurements such as surface temperature, terminal voltage, and current, supplemented by threshold alarms and multi-sensor fusion. Some systems use anomaly detection or prediction models based on historical data, while others use equivalent thermal models, charge-discharge curves, and impedance parameters to indirectly infer the internal state.

[0003] However, in the early stage of local heat source formation caused by internal defects, the surface temperature changes little and is greatly affected by the environment and mutual thermal coupling, making it difficult for surface sensors to capture in time; fixed thresholds are difficult to adapt to different SOH and environmental conditions, easily resulting in false alarms / missed alarms; model-based methods are highly dependent on parameters and calibration, and are not adaptable to changes in ventilation / load / stacking methods; increasing sensor types or density can improve coverage, but it also increases cost, deployment and maintenance complexity.

[0004] Therefore, achieving early online identification and location of internal heat accumulation in individual batteries within the cabinet without significantly increasing hardware costs, and maintaining low detection latency and false alarm rate under different battery aging and environmental conditions, remains a challenge and requirement for current technology. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a safety control method for battery swapping cabinets based on the Internet of Things, which solves the problems of lagging and insufficient adaptability in the early detection of internal heat accumulation in existing battery swapping cabinet safety control methods, resulting in untimely warnings and false alarms or missed alarms.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a safety control method for a battery swapping cabinet based on a smart Internet of Things, comprising:

[0009] Step S1: Real-time acquisition of battery terminal voltage, charging and discharging current and ambient temperature inside the battery swapping cabinet at the edge computing node to obtain synchronized multivariate time-series data;

[0010] Step S2 involves constructing and normalizing features for the time-series data. The features include at least one of the following: statistical features, time derivative features, and time-frequency features.

[0011] Step S3: Input the processed features into a pre-trained multivariate temporal anomaly detection model. The anomaly detection model is an architecture that integrates a long short-term memory autoencoder and an attention mechanism, and outputs the corresponding anomaly score.

[0012] Step S4: Determine the anomaly score based on the dynamic threshold to obtain the risk level;

[0013] Step S5: When the risk level reaches the preset threshold, a safety warning is issued, and the charging power is adjusted and / or interrupted according to the handling strategy. At the same time, the event is reported to the cloud monitoring platform through IoT communication.

[0014] As a preferred embodiment of the smart IoT-based battery swapping cabinet safety control method of the present invention, the training data of the anomaly detection model includes historical normal charge and discharge cycle data under multiple cabinets, multiple environments and multiple rates, and a sliding time window is used for modeling.

[0015] As a preferred embodiment of the smart Internet of Things-based safety control method for battery swapping cabinets described in this invention, the dynamic threshold is determined based on the quantile of the target false alarm rate and is adaptively adjusted according to the ambient temperature inside the cabinet and the battery health status, wherein the battery health status is estimated through capacity retention rate and / or equivalent internal resistance trend.

[0016] The dynamic threshold is determined based on the quantiles of the target false positive rate as follows:

[0017] During normal operation periods without alarms, a reference window for anomaly scores is maintained on a rolling basis to estimate the high-quantile threshold base corresponding to the target false alarm rate; therefore, at time... calculate Threshold basis is obtained from quantiles:

[0018] ,

[0019] in, Indicates time quantile threshold bases, Indicates the target false alarm rate. This represents the threshold independent variable used to determine the supremacy. This represents the set of reference sample indices used for estimation. Represents a set Sample size Indicates an indicator function, Indicates that the index is Abnormal scores, Indicates the index at the current time. Indicates the sample index;

[0020] Mapping the effects of temperature and health status on risk to a monotonically scaled threshold basis yields a dynamic threshold:

[0021] ,

[0022] in, Indicates time The dynamic threshold, Indicates the quantile threshold basis. and These represent the temperature correction factor and the health correction factor, respectively. Indicates the temperature offset. Indicates the ambient temperature inside the cabinet. Indicates reference temperature. Indicators representing health status;

[0023] In the formula:

[0024] ,

[0025] in, These represent the weights of the capacity channel and the internal resistance channel, respectively. and These represent the current available capacity and the initial capacity, respectively. and These represent the current equivalent internal resistance and the initial equivalent internal resistance, respectively.

[0026] right Lightweight smoothing is employed with a minimum update interval set, allowing for faster updates during load abrupt changes or temperature jumps, and samples from abnormal regions are not re-fed to the reference window;

[0027] Verify the actual false alarm rate using historical normal data. The fit is checked periodically during operation.

[0028] As a preferred embodiment of the smart IoT-based battery swapping cabinet safety control method described in this invention, the anomaly score is a weighted fusion result of reconstruction error and prediction residual, and the weights are dynamically allocated by an attention mechanism for different time steps and different feature channels.

[0029] As a preferred embodiment of the smart Internet of Things-based safety control method for battery swapping cabinets described in this invention, the edge computing node is configured with a thermal observer to estimate the internal temperature and temperature rise rate of the battery based on voltage, current and ambient temperature inside the cabinet. The estimation results are used to adaptively adjust the dynamic threshold and determine the risk level.

[0030] The state update method of the thermal observer is as follows:

[0031] The internal temperature estimate and heating trend are updated using discrete energy balance; the heating power is first constructed from electrical and environmental quantities.

[0032] ,

[0033] in, Indicates time Heating power, Indicates the Joule heat coefficient, Represents the equivalent coefficients related to polarization / entropy change. Indicates time The charging and discharging current, Indicates time The equivalent internal resistance, Indicates time The terminal voltage, Indicates the index at the current time;

[0034] Forward Euler prediction under the single heat capacity-thermal resistance model:

[0035] ,

[0036] in, Indicates based on Time information Predicted internal temperature values, Indicates time Internal temperature estimate Indicates the discrete time step. Indicates equivalent heat capacity, Indicates the equivalent thermal conductivity. Indicates the ambient temperature inside the cabinet;

[0037] The prediction bias is corrected using the energy balance residual, and the heating rate is obtained:

[0038] ,

[0039] in, Indicates time Energy balance residual, This represents the estimated internal temperature at the previous moment;

[0040] Perform a correction based on the residual and provide the heating rate:

[0041] ,

[0042] in, Indicates the corrected Estimated internal temperature at any given time This represents the residual correction gain of the temperature channel. Indicates time Estimation of the internal temperature rise rate;

[0043] Introduce bias states and iterate:

[0044] ,

[0045] in, express Estimation of heat generation bias at time point. Indicates time The heat generation bias estimation, Indicates the bias forgetting factor. This represents the residual correction gain of the bias channel;

[0046] The system employs segmented operation during air cooling startup, load transition, and rest periods. ,limit The physical upper and lower bounds and for Perform a light smoothing process;

[0047] Initialization with the steady-state phase of the environment and When current or voltage sensing is abnormal, the correction channel is frozen, and only the model prediction branch is retained.

[0048] As a preferred embodiment of the safety control method for a battery swapping cabinet based on the intelligent Internet of Things described in this invention, the time-frequency features include the fundamental and harmonic amplitudes, phases, and their proportional features obtained by performing a fast Fourier transform on the current waveform during the charging phase. The time-frequency features and statistical features are jointly input into the anomaly detection model.

[0049] As a preferred embodiment of the smart Internet of Things-based battery swapping cabinet safety control method described in this invention, the handling strategy is a graded strategy, which includes at least: derated charging, forced air cooling, isolation of suspicious battery trays and interruption of charging, and automatic reset according to preset recovery conditions after the risk is eliminated.

[0050] As a preferred embodiment of the smart IoT-based battery swapping cabinet security control method described in this invention, the cloud monitoring platform is used to receive edge events, store logs and model versions, issue control commands to the battery swapping cabinet and perform remote updates, and the edge computing node enables local security rules for independent judgment when the network is interrupted.

[0051] Secondly, the present invention provides a safety control system for a battery swapping cabinet based on the intelligent Internet of Things, comprising,

[0052] The data acquisition module is used to synchronously collect and time-mark battery terminal voltage, charging and discharging current and ambient temperature inside the battery swapping cabinet at the edge computing node, forming multivariate time-series data;

[0053] The feature processing module is used to perform preprocessing and normalization on the time series data, construct statistical features, time derivative features and time-frequency features based on the sliding time window, and perform fast Fourier transform on the current waveform during the charging stage to obtain the fundamental and harmonic related features, and output the feature sequence.

[0054] The anomaly detection module is used to run a multivariate temporal anomaly detection model that integrates a long short-term memory autoencoder and an attention mechanism, and to reason about the feature sequence to obtain an anomaly score that is a weighted fusion of reconstruction error and prediction residual.

[0055] The dynamic threshold determination module is used to set an initial threshold based on the quantile of the target false alarm rate, and adaptively adjust the threshold in combination with the ambient temperature inside the cabinet and the battery health status. The battery health status is estimated by the capacity retention rate and / or equivalent internal resistance trend, and the abnormal scores are compared accordingly to give a risk level.

[0056] The thermal monitoring module is used to estimate the internal temperature and temperature rise rate of the battery based on voltage, current and ambient temperature inside the cabinet, and provides the estimation results to the dynamic threshold determination module for threshold adjustment and enhanced risk assessment.

[0057] The disposal execution module is used to implement a graded disposal strategy according to the risk level, including at least: reducing charging power, starting forced air cooling, isolating the suspected battery tray and interrupting charging, and automatically resetting according to preset recovery conditions;

[0058] The IoT communication module is used to report alarm events, logs and model version information to the cloud monitoring platform, receive control commands and model / policy updates issued by the platform, and realize edge-cloud collaboration.

[0059] The edge management and fault tolerance module is used to enable local security rules to independently complete risk assessment and handling when the network is interrupted, and to synchronize data and policies after the network is restored.

[0060] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the safety control method for a battery swapping cabinet based on the smart Internet of Things as described in the first aspect of the present invention.

[0061] The beneficial effects of this invention are as follows: This invention significantly improves the early detection capability of internal heat accumulation and the overall safety performance of the battery through a smart Internet of Things-based battery swapping cabinet safety control method.

[0062] This invention utilizes a multivariate temporal anomaly detection model combined with a long short-term memory autoencoder and attention mechanism to extract deep features from voltage, current, and ambient temperature data, effectively identifying initial signals of internal thermal accumulation that are difficult for external sensors to capture. This data-driven approach overcomes the limitations of traditional fixed threshold monitoring. By dynamically learning normal charging and discharging patterns, it achieves sensitive capture of subtle anomalies, thereby advancing the warning timing and reducing the risk of missed detections due to surface temperature lag. The dynamic threshold mechanism of this invention is based on the quantile of the target false alarm rate and adaptive adjustment of operating parameters, enabling the system to flexibly cope with complex factors such as different ambient temperatures and battery aging levels, avoiding false alarms caused by a uniform threshold and improving warning accuracy. The introduction of a thermal observer further enhances internal state inference. By estimating the battery's internal temperature and temperature rise rate through an energy balance model, it provides supplementary evidence for risk level determination, especially in fast charging or high-temperature scenarios, enabling more accurate assessment of thermal runaway risk. In addition, the edge-cloud collaborative architecture ensures a balance between real-time response and remote management. Edge computing nodes process local data and execute rapid control, while the cloud platform is responsible for model updates and log analysis, improving the system's reliability and scalability. The tiered response strategy automatically adjusts charging power or activates isolation measures based on the risk level, achieving closed-loop management from early warning to control and reducing the possibility of the accident spreading.

[0063] This invention not only enhances the adaptability of the battery swapping cabinet under varying operating conditions, but also reduces the dependence on external hardware through algorithm optimization, thereby lowering operation and maintenance costs and providing sustainable assurance for the safe and stable operation of the electric vehicle battery swapping network. Attached Figure Description

[0064] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0065] Figure 1 This is a flowchart illustrating the safety control method for a battery swapping cabinet based on the Internet of Things in this embodiment.

[0066] Figure 2 This is a schematic diagram of the framework of the smart IoT-based battery swapping cabinet safety control system in the embodiment. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0069] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0070] This application proposes a safety control method for battery swapping cabinets based on the intelligent Internet of Things, combined with... Figure 1 As shown, the method includes:

[0071] Step S1: Real-time acquisition of battery terminal voltage, charging and discharging current, and ambient temperature inside the battery swapping cabinet at the edge computing node to obtain synchronized multivariate time-series data. In this embodiment, synchronized multivariate time-series data refers to voltage, current, and ambient temperature channel data with a unified timestamp under the same time reference. The time reference can be obtained by the local high-precision clock of the edge node or unified time synchronization by the site. The default sampling period is 100 milliseconds, which can be adjusted in the range of 50 milliseconds to 1000 milliseconds according to the site noise and computing power. The time alignment error is preferably no more than 50 milliseconds to ensure the comparability of cross-channel features. The range and resolution of voltage and current are set according to the rated parameters of the monitored battery and charging module. The ambient temperature inside the cabinet is set at one location each in the ventilation channel and the high thermal coupling area, or combined into an equivalent quantity. Optionally, when short-term communication congestion causes missing channel data, the most recent valid value is maintained for no more than 2 seconds, and the window is marked as not participating in modeling when the duration exceeds this time to avoid false triggering of judgment.

[0072] Step S2 involves feature construction and normalization of the time-series data. Features include at least one of the following: statistical features, time derivative features, and time-frequency features. Specifically, statistical features include mean, variance, extreme values, and root mean square (RMS). Time derivative features utilize first-order difference and short window slope. Time-frequency features extract dimensionally consistent indicators such as energy and harmonic proportion from the windowed data. The default feature window length is 20 to 60 seconds, the sliding step is 1 to 5 seconds, and the window overlap is 50% to balance timeliness and robustness. Normalization employs centering and scaling calculated per window. The scaling parameter is initialized during the stable phase after each cabinet starts and updated at a fixed period. Optionally, in scenarios with limited computing power, only four core features—mean, variance, first-order difference, and harmonic energy proportion—are retained to maintain consistency between the input and subsequent model interfaces. When the percentage of valid samples within a window is less than 80%, the current window feature is discarded and retried after one step.

[0073] Step S3 involves inputting the processed features into a pre-trained multivariate temporal anomaly detection model. The anomaly detection model is an architecture integrating a long short-term memory autoencoder and an attention mechanism, and outputs the corresponding anomaly score. For example, the length of the input sequence is consistent with the window length in step S2. The default hidden state size is set to a medium size to adapt to edge computing power, and the latent representation dimension is kept consistent with that in the training phase to avoid distribution drift. The single latency of online inference should be controlled within 200 milliseconds. The model parameters use a stable version that has passed offline verification and are atomically switched during low-load periods. When the model file verification fails or the version is inconsistent, inference is paused and the calculation of simplified anomaly scores based solely on statistical features is rolled back until the model is available again.

[0074] Step S4: Determine the risk level by judging the abnormal score based on the dynamic threshold. Similarly, the risk level is mapped according to the graded threshold from low to high, which by default includes three levels: prompt, warning and emergency. Each level should meet the minimum continuous judgment time to suppress jitter, which is recommended to be 3 to 10 seconds and adjusted according to the site load. In order to avoid frequent crossing of levels, uplink and downlink asymmetric hysteresis is adopted, and the exit condition is more relaxed than the entry condition by a fixed ratio. When the dynamic threshold has not been updated or is in a frozen state, the previous effective threshold is maintained for no more than 60 seconds to ensure continuity.

[0075] Step S5: When the risk level reaches the preset threshold, a safety warning is issued, and the charging power is adjusted and / or interrupted according to the handling strategy. At the same time, the event is reported to the cloud monitoring platform via IoT communication. Optionally, the event report shall include at least a timestamp, cabinet identifier, tray identifier, risk level, summary of the abnormal score in the most recent window, and summary of the handling action. A single event shall be retried no less than 3 times and exponential backoff shall be used to alleviate congestion. When the reporting channel is unavailable, the event and key summary shall be cached locally for no less than 72 hours and retransmitted sequentially after the channel is restored. At the same time, it shall be ensured that the handling action is based on local judgment and does not depend on the reporting result.

[0076] In one embodiment, the training data for the anomaly detection model includes historical normal charge-discharge cycle data under multiple cabinets, environments, and rates, and a sliding time window is used for modeling to improve the generalization ability to different operating conditions. Furthermore, the training samples are selected only from normal sections confirmed by manual or rule-based filtering, excluding periods with over-temperature, over-current, communication anomalies, and maintenance operations. The sample distribution should cover common seasons and ventilation conditions. The default window length is consistent with the online data, and the sliding step size is 1 to 5 seconds to increase the sample size while controlling the dependence of adjacent samples. Stratified sampling is used across cabinets and across sites to reduce bias. The model selection uses site-specific cross-validation to evaluate the generalization ability, and the configuration with the lowest false alarm rate under the target false alarm rate constraint is selected first.

[0077] In one embodiment, the dynamic threshold is determined based on the quantile of the target false alarm rate and is adaptively adjusted according to the ambient temperature inside the cabinet and the battery health status, wherein the battery health status is estimated by capacity retention and / or equivalent internal resistance trends.

[0078] The dynamic threshold is determined based on the quantiles of the target false positive rate as follows:

[0079] During normal operation periods without alarms, a reference window for anomaly scores is maintained on a rolling basis to estimate the high-quantile threshold base corresponding to the target false alarm rate; therefore, at time... calculate Threshold basis is obtained from quantiles:

[0080] ,

[0081] in, Indicates time quantile threshold bases, Indicates the target false alarm rate. This represents the threshold independent variable used to determine the supremacy. This represents the set of reference sample indices used for estimation (containing only samples judged to be normal). Represents a set Sample size Indicates an indicator function, Indicates that the index is Abnormal scores, Indicates the index at the current time. Indicates the sample index;

[0082] The impact of temperature and health status on risk is mapped to a monotonically scaled threshold, resulting in a dynamic threshold for online assessment.

[0083] ,

[0084] in, Indicates time The dynamic threshold, Indicates the quantile threshold basis. and These represent the temperature correction factor and the health correction factor, respectively. Indicates the temperature offset. Indicates the ambient temperature inside the cabinet. Indicates reference temperature. Indicators representing health status (range of values) );

[0085] In the formula:

[0086] ,

[0087] in, These represent the weights of the capacity channel and the internal resistance channel, respectively. and These represent the current available capacity and the initial capacity, respectively. and These represent the current equivalent internal resistance and the initial equivalent internal resistance, respectively.

[0088] right Lightweight smoothing is employed with a minimum update interval set, allowing for faster updates during load abrupt changes or temperature jumps, and samples from abnormal regions are not re-fed to the reference window;

[0089] Verify the actual false alarm rate using historical normal data. The fit is checked periodically during operation, and minor adjustments are made within the allowable range of the project when necessary. and The possible values ​​of ;

[0090] Specifically, the threshold determination is broken down into two layers: statistical quantile and operating condition correction, forming an online threshold that aligns with the target false alarm rate while being sensitive to environmental and degradation conditions. The first layer relies on the windowed distribution of normal segments to set the high-quantile threshold base, maintaining consistency in judgment criteria despite differences in fractional scales caused by different cabinets and scaling ratios, without imposing distribution assumptions, facilitating cross-site migration. The second layer uses an exponential mapping to scale the threshold base, where temperature offset and health degradation are linearly combined into the exponential term, causing the threshold to tighten as high temperature and degradation increase, and loosen as low temperature or new battery status increases, improving the operating condition adaptability of alarms. The health indicator is composed of a weighted average of capacity and internal resistance, covering capacity degradation on the one hand, and reflecting the amplification effect of impedance increase on thermal and safety risks on the other hand. The weights are constrained by a simplex shape, facilitating robust calibration on different datasets. Stabilization and update rhythm are used to suppress jitter and oversensitivity, and abnormal segments are removed to avoid contaminating the baseline. The closed-loop verification of offline and online operations keeps the target false alarm rate controllable during operation. In this embodiment, the commonly used range for the target false alarm rate parameter is 0.01% to 2%. The lower limit of the sample size of the reference window is recommended to be no less than 500 valid fractional samples. The reference temperature is based on a normal temperature of 25 degrees Celsius and can be set according to the long-term average of the site. The temperature correction coefficient and the health correction coefficient are tuned within a small range, usually calibrated through playback tests of historical normal and slightly abnormal periods, so that the threshold tightening during high temperature or health deterioration is controlled between several percentage points and more than ten percentage points. Under the constraint that the sum of the capacity weight and internal resistance weight of the health indicator is one, the commonly used range is 30% to 70% of each. The specific proportion is determined according to the relative contribution of capacity decay and internal resistance increase to risk in the sample. The threshold smoothing update cycle is recommended to be 1 to 5 seconds, and the minimum update interval is consistent with the site load. The reference window is not re-fed in abnormal sections to maintain the purity of the quantile threshold base.

[0091] In one embodiment, the outlier score is a weighted fusion result of the reconstruction error and the prediction residual. The weights are dynamically allocated by the attention mechanism for different time steps and different feature channels. Specifically, to ensure the stability of the fusion, the outputs of the reconstruction branch and the prediction branch are first scaled on their respective normal segment statistics, and then fused according to the dynamic weights. By default, the baseline weights of the two branches each account for half, and the fluctuation range of the weights is limited to a preset range to avoid extreme bias. Optionally, when either branch shows obvious drift or missing data in the current window, the weight of the other branch is temporarily increased to no less than 70%, and after the drift disappears, it is smoothly restored to the baseline ratio in an exponential manner.

[0092] In one embodiment, the edge computing node is configured with a thermal observer to estimate the internal temperature and temperature rise rate of the battery based on voltage, current and ambient temperature inside the cabinet. The estimation results are used for adaptive adjustment of dynamic thresholds and determination of risk level.

[0093] The thermal observer's state update method is as follows:

[0094] The internal temperature estimate and heating trend are updated using discrete energy balance; the heating power is first constructed from electrical and environmental quantities.

[0095] ,

[0096] in, Indicates time Heating power, Indicates the Joule heat coefficient, Represents the equivalent coefficients related to polarization / entropy change. Indicates time The charging and discharging current, Indicates time The equivalent internal resistance (from health estimates). Indicates time The terminal voltage, Indicates the index at the current time;

[0097] Forward Euler prediction under the single heat capacity-thermal resistance model:

[0098] ,

[0099] in, Indicates based on Time information Predicted internal temperature values, Indicates time Internal temperature estimate Indicates the discrete time step. Indicates equivalent heat capacity, Indicates the equivalent thermal conductivity. Indicates the ambient temperature inside the cabinet;

[0100] The prediction bias is corrected using the energy balance residual, and the heating rate is obtained:

[0101] ,

[0102] in, Indicates time Energy balance residual, This represents the estimated internal temperature at the previous moment;

[0103] Perform a correction based on the residual and provide the heating rate:

[0104] ,

[0105] in, Indicates the corrected Estimated internal temperature at any given time This represents the residual correction gain of the temperature channel. Indicates time Estimation of the internal temperature rise rate;

[0106] To absorb slowly varying mismatches, a bias state is introduced and recursively calculated:

[0107] ,

[0108] in, express Estimation of heat generation bias at time point. Indicates time The heat generation bias estimation, Indicates the bias forgetting factor (range) ), This represents the residual correction gain of the bias channel;

[0109] The system employs segmented operation during air cooling startup, load transition, and rest periods. ,limit The physical upper and lower bounds and for Perform a light smoothing process;

[0110] Initialization with the steady-state phase of the environment and When current or voltage sensing is abnormal, the correction channel is frozen, and only the model prediction branch is retained;

[0111] Specifically, this update mechanism uses a simple first-order thermal network as its framework, mapping measurable voltage, current, and ambient temperature to heating power. It then advances prior predictions of internal temperature through discrete energy balance. Since internal temperature cannot be directly measured, a pseudo-observation residual based on energy conservation is introduced to provide a lightweight correction to the prediction, while also providing the heating rate. This facilitates the use of a two-dimensional quantization combining temperature level and heating rate in risk classification. A bias state is used to absorb model order mismatch, installation differences, and changes in heat conduction paths, while a forgetting factor ensures gradual bias changes. Gain segmentation and boundary management reduce jitter and drift risks during operating condition switching, and a bypass strategy under abnormal operating conditions maintains estimation continuity. This observer relies only on quantities already collected in the scheme and existing health estimates, resulting in a compact structure. This facilitates implementation at edge nodes with low computing power, synergizing with the aforementioned dynamic thresholds and risk assessments. For example, equivalent heat capacity and equivalent thermal conductivity can be obtained through offline pulse conditions or bench steady-state experiments and written during initial site deployment, with common orders of magnitude being hundreds to thousands of joules per Kelvin and a few tenths to several watts per Kelvin, respectively. The discrete time step is consistent with the sampling period of step S1, and the temperature correction gain and bias correction gain are tuned within a small range to balance response and steady-state error. The forgetting factor is taken as a value close to one to reflect the slow variation assumption. The calculation results of the internal temperature are limited to a reasonable range near the ambient temperature, and the heating rate is lightly smoothed. When the voltage or current channel is missing for more than one window, only the predicted branch is retained and the correction channel is frozen until the channel is restored.

[0112] In one embodiment, the time-frequency features include the fundamental and harmonic amplitudes, phases, and their proportional characteristics obtained by performing a fast Fourier transform on the current waveform during the charging phase. The time-frequency features and statistical features are jointly input into the anomaly detection model. Similarly, the sampling rate of the time-frequency analysis is consistent with step S1. The default spectral resolution meets the resolution requirements of the power frequency and several harmonics by using an appropriate number of points and window length. The harmonic order preferably covers the second to fifteenth order. The proportional characteristics are mainly the ratio of harmonic energy to the fundamental frequency. When computing power or storage is limited, only a few robust features such as the proportion of low-order harmonics and total harmonic energy can be retained to reduce the computational load, while keeping the feature interface unchanged. When the current waveform contains obvious DC bias or short-term spikes, detrending and limiting processing are performed first to reduce the interference of spectral leakage.

[0113] In one embodiment, the handling strategy is a tiered strategy, including at least: derating charging, forced air cooling, isolating the suspected battery tray, and interrupting charging, and automatically resetting according to preset recovery conditions after the risk is eliminated; optionally, the default power step of derating charging is a number of equal parts of the nominal power, and after a single step of derating, an evaluation period of not less than 30 seconds is maintained before deciding whether to further drate or reset; forced air cooling can be triggered in the warning mode, and the isolation of the tray and interruption of charging are only triggered in the emergency mode and the handling time is recorded; to avoid oscillation, the reset condition is set with hysteresis and minimum observation time, and after the reset, it is prohibited to upgrade again for a short period of time.

[0114] In one embodiment, the cloud monitoring platform is used to receive edge events, store logs and model versions, issue control commands to the battery swapping cabinet and perform remote updates. When the network is interrupted, the edge computing node enables local security rules to make independent judgments. Furthermore, the data synchronization cycle between the edge and the cloud does not exceed 60 seconds by default under normal network conditions. Model and policy updates are completed atomically during off-peak hours and verification information is sent back after success. When the network quality is lower than a set threshold, the reporting frequency is automatically reduced and emergency events are sent first. Ordinary events are delayed until the network is restored and then batch retransmitted. During this period, the edge continuously executes local rules to ensure that safe handling does not depend on cloud reachability.

[0115] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements a smart IoT-based battery swapping cabinet safety control method as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] This application also proposes a safety control system for battery swapping cabinets based on the Internet of Things, combined with... Figure 2 As shown, it includes:

[0117] The data acquisition module is used to synchronously collect and time-mark battery terminal voltage, charging and discharging current and ambient temperature inside the battery swapping cabinet at the edge computing node, forming multivariate time-series data;

[0118] The feature processing module is used to perform preprocessing and normalization on time series data, construct statistical features, time derivative features and time-frequency features based on the sliding time window, and perform fast Fourier transform on the current waveform during the charging stage to obtain the fundamental and harmonic related features, and output the feature sequence.

[0119] The anomaly detection module is used to run a multivariate temporal anomaly detection model that integrates a long short-term memory autoencoder and an attention mechanism, and to reason about the feature sequence to obtain an anomaly score that is a weighted fusion of reconstruction error and prediction residual.

[0120] The dynamic threshold determination module is used to set an initial threshold based on the quantile of the target false alarm rate, and adaptively adjust the threshold in combination with the ambient temperature inside the cabinet and the battery health status. The battery health status is estimated by the capacity retention rate and / or equivalent internal resistance trend, and the abnormal scores are compared accordingly to give a risk level.

[0121] The thermal monitoring module is used to estimate the internal temperature and temperature rise rate of the battery based on voltage, current and ambient temperature inside the cabinet, and provides the estimation results to the dynamic threshold determination module for threshold adjustment and enhanced risk assessment.

[0122] The disposal execution module is used to implement graded disposal strategies according to the risk level, including at least: reducing charging power, starting forced air cooling, isolating the suspected battery tray and interrupting charging, and automatically resetting according to preset recovery conditions;

[0123] The IoT communication module is used to report alarm events, logs and model version information to the cloud monitoring platform, receive control commands and model / policy updates issued by the platform, and realize edge-cloud collaboration.

[0124] The edge management and fault tolerance module is used to enable local security rules to independently complete risk assessment and handling when the network is interrupted, and to synchronize data and policies after the network is restored. In this embodiment, each module runs collaboratively on the processor through program instructions, and the data flow corresponds one-to-one with steps S1 to S5, with the minimum refresh period being consistent with the sampling period. When resources are insufficient, priority is given to ensuring the real-time performance of the three core links of data acquisition, dynamic threshold assessment and handling execution, while secondary tasks are suspended until resources are restored to ensure the continuous availability of security functions.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0126] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A smart Internet of Things-based battery swap cabinet safety control method, characterized in that, The method comprises the following steps: Step S1, collecting battery terminal voltage, charging and discharging current and cabinet internal environment temperature in real time at the edge computing node of the battery swap cabinet to obtain synchronous multivariate time series data; Step S2, performing feature construction and normalization processing on the time series data, and the features include at least one of statistical features, time derivative features and time-frequency features; Step S3, inputting the processed features into a pre-trained multivariate time series anomaly detection model, the anomaly detection model is an architecture integrating long and short-term memory autoencoder and attention mechanism, and outputting corresponding anomaly scores; Step S4, determining the anomaly scores according to a dynamic threshold to obtain a risk level; Step S5, when the risk level reaches a preset threshold, issuing a safety warning, adjusting and / or interrupting the charging power according to a disposal strategy, and reporting the event to a cloud monitoring platform through Internet of Things communication; The dynamic threshold is determined based on quantiles of a target false positive rate, and is adaptively adjusted according to the cabinet internal environment temperature and the battery health state, wherein the battery health state is estimated by capacity retention rate and / or equivalent internal resistance trend; The dynamic threshold determined based on quantiles of a target false positive rate is as follows: In the normal operation interval without triggering the alarm, the reference window of the rolling maintenance abnormal score is used to estimate the high quantile threshold base to correspond to the target false positive rate; for this, at time Calculate Quantile threshold base: ， in, Indicates time quantile threshold bases, Indicates the target false alarm rate. This represents the threshold independent variable used to determine the supremacy. This represents the set of reference sample indices used for estimation. Represents a set Sample size Indicates an indicator function, Indicates that the index is Abnormal scores, Indicates the index at the current time. Indicates the sample index; Map the influence of temperature and health state on risk to the monotonic scaling of threshold base to obtain the dynamic threshold: ， wherein, represents a dynamic threshold at a time point, represents a quantile threshold base, respectively represent a temperature correction coefficient and a health correction coefficient, represents a temperature offset, represents an in-cabinet ambient temperature, represents a reference temperature, represents a health status indicator;​​ In the formula: ， wherein, respectively represent the weight of the capacity channel and the internal resistance channel, and respectively represent the current available capacity and the initial capacity, and respectively represent the current equivalent internal resistance and the initial equivalent internal resistance; To Light smoothing is used with a minimum update interval, and the update is allowed to be accelerated when there is a load mutation or temperature jump. The abnormal section sample is not backfilled to the reference window. Verify the actual false alarm rate using historical normal data. The fit is checked periodically during operation; The edge computing node is configured with a thermal observer for estimating the battery internal temperature and temperature rise rate according to the voltage, current and cabinet internal environment temperature, and the estimation results are used for adaptive adjustment of the dynamic threshold and determination of the risk level; The state updating method of the thermal observer is as follows: Update the internal temperature estimation and temperature rise trend by discrete energy balance; first construct the heat generation power from electrical and environmental quantities: ， wherein, represents the heating power at time represents the Joule heat coefficient, represents the equivalent coefficient related to polarization / entropy change, represents the charging / discharging current at time represents the equivalent internal resistance at time represents the terminal voltage at time represents the current time index;​​​​ Make forward Euler prediction under the single thermal capacity-thermal resistance model: ， wherein, represents the internal temperature prediction value of the cabinet based on represents the internal temperature prediction value of the cabinet based on represents the internal temperature estimation value of the cabinet at the time represents the internal temperature estimation value of the cabinet at the time represents a discrete time step, represents an equivalent heat capacity, represents an equivalent thermal conductivity, represents the cabinet internal ambient temperature;​ Correct the prediction deviation by energy balance residual error, and obtain the temperature rise rate: ， wherein, represents the energy balance residual at the time instant represents the internal temperature estimate at the previous time instant;​ Make a correction based on the residual error and give the temperature rise rate: ， wherein, represents the corrected time internal temperature estimate, represents the residual correction gain for the temperature channel, represents the internal temperature ramp rate estimate at the time t. Introduce the bias state and recursively update it: ， wherein, represents a heating bias estimate at time instant represents a heating bias estimate at time instant a heating bias estimate at time instant represents a bias forgetting factor, represents a bias channel residual correction gain; In the wind cooling opening, load transition and static interval adopts the subsection , limit The physical upper and lower boundaries and make light smoothing ; Initialize with environmental steady state segment With Freeze the correction path when current or voltage sensing is abnormal, leaving only the model prediction leg.

2. The battery swap cabinet safety control method based on the smart Internet of Things according to claim 1, characterized in that, The training data of the anomaly detection model includes historical normal charging and discharging cycle data under multiple cabinets, multiple environments and multiple rates, and a sliding time window is used for modeling. 3.The battery swap cabinet security control method based on smart Internet of Things according to claim 1, characterized in that, The anomaly score is a weighted fusion result of reconstruction error and prediction residual error, and the weight is dynamically allocated to different time steps and different feature channels by the attention mechanism. 4.The battery swap cabinet security control method based on smart Internet of Things according to claim 1, characterized in that, The time-frequency features include fundamental wave and harmonic amplitude, phase and proportion features obtained by performing fast Fourier transform on the charging stage current waveform, and the time-frequency features and statistical features are jointly input into the anomaly detection model. 5.The battery swap cabinet security control method based on smart Internet of Things according to claim 1, characterized in that, The disposal strategy is a hierarchical strategy, which at least includes de-rating charging, forced air cooling, isolating suspicious battery trays and interrupting charging, and automatically resetting according to preset recovery conditions after the risk is resolved.

6. The battery swap cabinet safety control method based on the smart Internet of Things according to claim 1, characterized in that, The cloud monitoring platform is used for receiving edge events, storing logs and model versions, issuing control instructions to the battery swap cabinet and performing remote updates, and the edge computing node enables independent determination by using local safety rules when the network is interrupted.

7. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by a processor to implement the steps of the battery swap cabinet safety control method based on intelligent Internet of Things according to any one of claims 1-6.

Images