Lithium ion battery internal short circuit fault detection method based on time-frequency fusion

Abstract

The invention discloses a lithium ion battery internal short circuit fault detection method based on time-frequency fusion, and the method comprises the steps: firstly constructing a fractional order equivalent circuit model fused with an electrochemical aging mechanism, and simulating the paths of conductivity reduction, active material loss and lithium inventory reduction in combination with an aging empirical formula; joint modeling of the aging process and the random internal short circuit is achieved, and multi-cycle voltage and current and electrochemical impedance spectroscopy data are obtained; secondly, extracting time domain features by using a time domain attention enhanced long-short-term memory network, and analyzing frequency domain information by using a multi-scale frequency sensing convolutional neural network; and then weighting and screening two types of modal features through a dynamic gating fusion module, introducing a cross attention mechanism to establish dependency mapping between time-frequency domain features, and finally outputting an internal short circuit fault detection result. The method can solve the problems that in the lithium ion battery aging process, due to lithium dendrite growth, the internal short circuit early fault is high in concealment, and single time-frequency characteristics are difficult to detect, and early high-precision recognition of the internal short circuit fault can be achieved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery safety monitoring technology, and in particular discloses a method for detecting internal short-circuit faults based on time-frequency fusion. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have become core components of energy storage systems and electric vehicles. Their operational safety is directly related to critical infrastructure and personal safety. Internal short circuits, as the most insidious and destructive fault mode, are mainly caused by lithium dendrite growth, separator aging, or mechanical stress. They can lead to localized thermal runaway and trigger chain reactions such as fires and explosions. Early internal short circuit fault signals are weak and easily masked by operating noise. Traditional detection methods based on voltage difference, temperature threshold, or equivalent circuit parameters lack sensitivity and are difficult to effectively detect before the thermal runaway window (usually <60 seconds).

[0003] Existing internal short-circuit detection technologies are mainly divided into three categories: methods based on consistency differences identify abnormal cells by voltage / temperature differences between batteries within the module, but battery aging causes parameter drift, which significantly reduces the signal-to-noise ratio and results in a high false alarm rate; methods based on model analysis use equivalent circuit parameters to identify short-circuit resistance, but integer-order models are difficult to accurately describe the frequency domain response characteristics of the battery, and the coupling between the aging process and fault characteristics can interfere with the accuracy of parameter estimation; and data-driven methods use machine learning to directly mine fault modes from operational data, but single-modal features (only time domain or only frequency domain) are difficult to capture early weak faults, and the limitation of relying on massive labeled samples leads to insufficient generalization ability.

[0004] The core bottlenecks currently faced by existing internal short-circuit detection technologies are: the separation of time and frequency characteristics, with time-domain methods (such as voltage platform monitoring) being sluggish in responding to early micro-short circuits, and frequency-domain methods (such as EIS phase angle analysis), although sensitive to interface polarization, relying on dedicated equipment and having low sampling frequencies, failing to meet real-time monitoring requirements; aging-fault coupling interference, with mechanisms such as conductivity loss and active material degradation caused by battery aging masking impedance changes caused by short circuits, making it difficult for traditional methods to decouple these two types of effects; and poor adaptability across operating conditions, with single-feature detection thresholds easily failing under complex operating conditions such as dynamic loads and temperature fluctuations.

[0005] Therefore, in order to achieve high-precision early detection of internal short circuits in high-noise and multi-aging scenarios, it is urgent to break through the technical bottlenecks of time-frequency feature fusion and aging-fault decoupling, and establish a multimodal diagnostic system that combines physical interpretability and intelligent discrimination capabilities. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the shortcomings mentioned in the background technology, this invention discloses a time-frequency fusion-based method for detecting internal short-circuit faults in lithium-ion batteries. This method aims to solve the technical problems of insufficient sensitivity for early internal short-circuit detection, fragmented time-frequency features, and aging-fault coupling interference in existing technologies under strong noise and multi-aging scenarios. The method achieves high-precision early identification of internal short-circuit faults through fractional-order equivalent circuit modeling and neural network fusion architecture.

[0008] (II) Technical Solution

[0009] This invention discloses a method for detecting internal short-circuit faults in lithium-ion batteries based on time-frequency fusion, comprising the following steps:

[0010] Step S100: Data preparation. Time-series data during the charging and discharging process of lithium-ion batteries are collected in real time using voltage and current sensors. Electrochemical impedance spectroscopy (EIS) of batteries under different states of charge is measured using an electrochemical workstation.

[0011] Step S200: Establish a fractional-order equivalent circuit model that integrates conductivity loss, active material loss, and lithium inventory loss; fit the empirical function of model parameter evolution based on a public dataset to achieve dynamic simulation of the aging process.

[0012] Step S300: Randomly inject internal short-circuit faults of different degrees into the model to generate time-domain and frequency-domain data with multiple cycles and aging stages;

[0013] Step S400: Establish a temporal attention-enhanced long short-term memory network to parse the temporal dependencies of voltage and current sequences, and extract features such as charge / discharge plateau shift and enhanced voltage fluctuations; use the temporal attention-enhanced long short-term memory network to parse multi-cycle time-series data of voltage, current, and state of charge. This network extracts cross-cycle temporal dependency features through a bidirectional long short-term memory layer. Its forward propagation process is dynamically controlled by a gating mechanism, and the forget gate dynamically adjusts the historical memory retention rate according to the voltage deviation. After optimization, a temporal feature vector F is finally generated. t ;

[0014] Step S500: Establish a multi-scale frequency-aware convolutional neural network to analyze EIS and capture the frequency domain response of impedance collapse and phase angle anomalies in the mid-to-low frequency range. The multi-scale frequency-aware convolutional neural network includes a feature preprocessing layer, a multi-scale convolutional block, a frequency channel attention module, and a feature compression layer. The input is the electrochemical impedance spectrum corresponding to a single charge-discharge cycle, and the output is the frequency domain feature vector F. f ;

[0015] Step S600: Use a dynamic gating fusion module to process the time-domain feature vector F t and frequency domain eigenvector F fWeighted selection is performed, and then a time-frequency feature dependency mapping is established through a cross-attention mechanism to generate a fused feature vector, which is then input into the classification discriminator to complete the internal short-circuit fault detection.

[0016] Furthermore, steps S200-S600 of the present invention also include details such as the modeling and calculation methods of formulas (1) to (35). To avoid redundancy, please refer to the following text for details.

[0017] Another aspect of the present invention relates to a short-circuit fault detection system for lithium-ion batteries based on time-frequency fusion, comprising:

[0018] Measurement module: Used for data preparation, it acquires time-series data of lithium-ion battery charging and discharging process in real time through voltage and current sensors, and measures the electrochemical impedance spectrum of the battery under different states of charge through an electrochemical workstation;

[0019] Model building module: used to build a fractional-order equivalent circuit model that integrates conductivity loss, active material loss and lithium inventory loss, and fits the empirical function of model parameter evolution based on public dataset to realize dynamic simulation of the aging process;

[0020] Fault injection module: used to randomly inject internal short-circuit faults of different degrees into the model, generating time-domain and frequency-domain data with multiple cycles and aging stages;

[0021] The first network module is used to establish a temporal attention-enhanced long short-term memory network to analyze the temporal dependencies of voltage and current sequences and extract features such as charge / discharge plateau shift and voltage fluctuation enhancement.

[0022] The second network module is used to establish a multi-scale frequency-aware convolutional neural network to analyze EIS and capture frequency domain responses such as impedance collapse and phase angle anomalies in the mid-to-low frequency range.

[0023] Fault detection module: The dynamic gating fusion module performs weighted selection of time-domain and frequency-domain features, and then establishes a time-frequency feature dependency mapping through a cross-attention mechanism to generate a fused feature vector, which is then input into the classification discriminator to complete the internal short-circuit fault detection.

[0024] Furthermore, this invention also discloses a lithium-ion battery internal short-circuit fault detection system based on time-frequency fusion, comprising:

[0025] At least one processor; and at least one memory communicatively connected to said processor, wherein:

[0026] The memory stores program instructions that can be executed by the processor, which can call the program instructions to execute the lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion as described above.

[0027] (III) Beneficial Effects

[0028] (1) This invention provides a method for detecting internal short-circuit faults in lithium-ion batteries based on time-frequency feature fusion. By establishing a fractional-order equivalent circuit model containing constant-phase elements and injecting multi-physics field coupled faults, a dynamic characteristic dataset covering the entire life cycle is constructed. The time-domain dynamic evolution features are extracted using a time-attention enhanced long short-term memory network, and the impedance spectrum frequency domain response is analyzed by a multi-scale frequency-aware convolutional neural network. A dynamic gating fusion mechanism is designed to realize adaptive weighting of time-frequency features, and a feature cross-mapping network is constructed to complete feature dimension alignment. Finally, internal short-circuit fault detection is achieved based on the fused features.

[0029] (2) The lithium-ion battery internal short-circuit fault detection method based on time-frequency feature fusion provided by this invention first combines fractional calculus theory with fault injection technology to construct an equivalent battery model containing constant-phase elements and generate multi-condition aging-fault coupling samples. The core technological breakthroughs are: proposing a dual-channel deep feature extraction architecture, strengthening the dynamic process feature capture capability through a time-domain attention mechanism, and utilizing frequency-domain multi-scale convolution to mine impedance phase response laws; designing a dynamic gated fusion device to decouple the time-frequency feature coupling relationship, constructing a feature cross-mapping network to eliminate modal dimension differences; and adopting an aging sample decoupling training strategy to separate aging interference and fault features. Simulation results comparing with other algorithms show that, compared with traditional single-domain analysis methods, this invention significantly improves the early internal short-circuit detection sensitivity under complex operating conditions, effectively extends the thermal runaway handling window period, and significantly reduces the false alarm rate, providing reliable technical support for battery safety detection. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below:

[0031] Figure 1 This is a flowchart illustrating an embodiment of the lithium-ion battery internal short-circuit fault detection method based on time-frequency feature fusion provided by the present invention.

[0032] Figure 2 This is a diagram of the third-order fractional-order equivalent circuit structure in this invention;

[0033] Figure 3 This is a structural diagram of the electro-aging coupling model in this invention;

[0034] Figure 4 This is a comparison of the internal short-circuit detection results of the time-frequency algorithm proposed in this invention with those of traditional neural networks LSTM and CNN.

[0035] Figure 5 This is a functional block diagram of an embodiment of the lithium-ion battery internal short-circuit fault detection method based on time-frequency feature fusion provided by the present invention.

[0036] Explanation of icon numbers:

[0037] 10. Measurement module; 20. Model building module; 30. Fault injection module; 40. First network module; 50. Second network module; 60. Fault detection module. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0039] like Figure 1 As shown, in order to address the technical problems of insufficient sensitivity for early internal short circuit detection, fragmented time-frequency characteristics, and aging-fault coupling interference in existing technologies under strong noise and multi-aging scenarios, this invention designs a lithium-ion battery internal short circuit fault detection method based on time-frequency fusion. The lithium-ion battery internal short circuit fault detection method based on time-frequency fusion specifically includes the following steps:

[0040] Step S100: Data preparation. Time-series data during the charging and discharging process of the lithium-ion battery are collected in real time using voltage and current sensors. Electrochemical impedance spectroscopy of the battery under different states of charge is measured using an electrochemical workstation. Hybrid Pulse Power Characteristic (HPPC) test is performed on a polymer 912995 type soft-pack lithium-ion battery. The terminal voltage across the battery is measured using a voltage sensor, and the current of the battery is measured using a current sensor.

[0041] Step S200: Establish a fractional-order equivalent circuit model that integrates conductivity loss, active material loss, and lithium inventory loss. Figure 2 As shown, an empirical function for the evolution of model parameters is fitted based on a publicly available dataset to achieve dynamic simulation of the aging process;

[0042] exist Figure 2 In the third-order fractional RC equivalent circuit shown, the first parallel branch consisting of the fractional inductor L, the ohmic internal resistance R0, the resistor R1 and the constant-phase element CPE1, the second parallel branch consisting of the resistor R2 and the constant-phase element CPE2, and the third parallel branch consisting of the resistor R3 and the constant-phase element CPE3 are connected in series.

[0043] Step S200 specifically includes the following steps:

[0044] Step S210: Construct a fractional-order equivalent circuit model by integrating three aging mechanisms: conductivity loss, active material loss, and lithium inventory loss, as shown below. Figure 3 As shown.

[0045] Model parameters With the number of loops The dynamic evolution is based on fitting and transfer to a public dataset, and is ultimately achieved through the following set of formulas:

[0046] The conductivity loss branch simulates the increase in ohmic resistance caused by current collector corrosion and conductive agent shedding:

[0047] (1)

[0048] Increased polarization resistance and decreased double-layer capacitance due to electrode pulverization and desorption in the active material loss branch simulation:

[0049]

[0050] The lithium inventory loss branch characterizes the exponential increase in diffusion impedance and the decay of interfacial capacitance caused by the consumption of active lithium in side reactions:

[0051]

[0052] in, For the random disturbance term, the coefficients are... to , to , Determined by fitting a publicly available aging dataset, and satisfying capacity decay constraints:

[0053]

[0054] In formulas (1) to (4), C(N) and C0 are the discharge capacity of the battery in the Nth cycle and the initial discharge capacity of the battery, respectively. β and β are the capacity decay rate coefficient and the cyclic aging acceleration factor, respectively. , , , These are the resistance values ​​of R0 to R3 after the Nth cycle. , , , These are the resistance values ​​of R0 to R3 before the battery discharges. , These are the parameter values ​​for CPE2 and CPE3 after the Nth iteration, respectively.

[0055] Step S220: The state-space equations of the model are derived based on Caputo fractional differential theory; the state variables of the model are defined as follows: ,in These correspond to the polarization voltages of the three parallel branches, Given a charged state, with the subscript k representing the current time; the discrete form of the state-space equation is expressed as:

[0056]

[0057] in, For current input, For the model parameter set, For process noise; Newton binomial coefficient matrix From fractional order The decision is made if its elements satisfy:

[0058]

[0059] State transition function The expression is:

[0060]

[0061] Measurement function The expression y k for:

[0062]

[0063] Where the subscript m is an integer ranging from 1 to 3, It is an open-circuit voltage function. For Coulomb efficiency, For rated capacity, To measure noise, The sampling time interval, This is the Gamma function.

[0064] Step S300: Randomly inject internal short-circuit faults of different degrees into the model to generate time-domain and frequency-domain data with multiple cycles and aging stages.

[0065] Step S300 specifically includes the following steps:

[0066] Step S310: The internal short-circuit fault injection mechanism is achieved by connecting a variable resistor in parallel across the fractional-order equivalent circuit model. The total impedance function under fault conditions is expressed as follows:

[0067]

[0068] in, The impedance function of the battery after the Nth cycle without an internal short circuit is given, and the short-circuit resistance is given. The value retrieval strategy is defined as follows: the normal state is set to Simulating an open circuit state, the fault states are distributed uniformly. Random sampling was used to cover scenarios ranging from soft short circuits to hard short circuits; fault injection timing and aging stages were implemented using a grouping strategy, with the early injection group (10 battery groups) being selected based on the number of cycles. Interval-based random fault triggering, mid-term injection group (50 battery groups) in Interval random triggering, late injection group (20 battery groups) in Random triggering within a range, with an additional fault-free control group (20 battery groups) maintained throughout. ;

[0069] Step S320: The multi-source data generation process includes the collaborative generation of time-domain and frequency-domain data: During constant current discharge, the terminal voltage response is solved iteratively through fractional-order discrete state-space equations, where the polarization voltage component... The calculation formula is:

[0070]

[0071] The terminal voltage output is:

[0072]

[0073] Where L is the fractional inductance value, for The polarization current component under the current, This represents the state of charge at the current sampling time n.

[0074] With a sampling frequency of 1Hz, 3600 time-domain data points are generated in a single discharge cycle, forming a time-domain data matrix. Each row contains voltage, current, and SOC values, forming a matrix:

[0075]

[0076] Among them, superscript Indicates the first One charge-discharge cycle.

[0077] At the end of each cycle, the electrochemical impedance spectroscopy is calculated based on the current parameters, and the frequency points are... ( Values ​​are taken at ten frequency points and distributed on a logarithmic scale, and expressed through the impedance function:

[0078]

[0079] Generate frequency domain data matrix Each row contains the real and imaginary parts of the impedance:

[0080]

[0081] The data augmentation process introduces operating noise disturbances, and Gaussian white noise is added to the time-domain data. Add relative error to frequency domain data Sample labeling uses quadruples Record the cycle count, fault status flag, short-circuit resistance value, and most recent frequency index of the fault trigger point in sequence. , Let p be the frequency value at the p-th frequency point. The characteristic frequency corresponding to the fault trigger moment;

[0082] Step S330: The dynamic aging-fault coupling mechanism is implemented through a loop-driven process: First, the aging parameters are updated. When the number of iterations is satisfied Time settings Subsequently, a 1C constant current discharge is performed and time-domain data is recorded. The EIS spectrum at the current SOC is calculated to generate frequency-domain data. Finally, the cycle counter is incremented and the process is repeated. Key parameter settings include a total of 100 cycles to cover the entire lifespan aging period, a time-domain sampling frequency of 1Hz, EIS using 50 logarithmically distributed frequency points based on ten frequency points, and noise intensity set to... Simulates sensor error, with short-circuit resistance ranging from 5 to 500Ω;

[0083] At the physical mechanism level, internal short-circuit bypass current This leads to abnormal loss of discharge capacity. In the time domain, this manifests as an early attenuation of the voltage plateau; the frequency domain response distortion is quantized by the phase angle shift.

[0084]

[0085] In the characteristic frequency range Within Hz As a fault-sensitive area; the final generated multi-period dataset structure is as follows: .

[0086] Step S400: Establish a temporal attention-enhanced long short-term memory network to parse the temporal dependencies of voltage and current sequences, and extract features such as charge / discharge plateau shift and voltage fluctuation enhancement; use the temporal attention-enhanced long short-term memory network to parse multi-cycle time-series data of voltage, current, and state of charge, with its input being a time-domain tensor composed of discharge cycles. ,in For batch size, For single-cycle timing, the feature dimension includes voltage. Current State of charge and timestamp This network extracts cross-period temporal dependency features through a bidirectional long short-term memory layer. Its forward propagation process is dynamically regulated by a gating mechanism, including a forgetting gate. Dynamically adjust the historical memory retention rate based on voltage deviation: when detected... Lowering the forget gate activation value at certain times can suppress the memory of historical plateau voltages. This is the predicted voltage value at the current moment. For voltage standard deviation, input gate This strengthens the current abnormal response characteristics, and the gating calculation follows:

[0087]

[0088] in, For Sigmoid activation functions, including W f W i and W o weight matrix With including b f b i and b o bias matrix As learnable parameters, residual connections are achieved through... To achieve information fusion between levels, For layer normalization function, Let l represent the hidden state of the l-th layer at time t, where the superscript l is the index of the neural network layer.

[0089] Forward and backward hidden states , After splicing, a 128-dimensional feature tensor is formed. Subsequently, a temporal attention mechanism is used to weight and aggregate the features of the key discharge platform region. The temporal attention mechanism module focuses on the hidden state sequence with weights and calculates the attention score.

[0090] (17)

[0091] The attention weight distribution is determined by both the voltage statistics and the second derivative, and is calculated as follows:

[0092]

[0093] in Let T be the average plateau voltage and T be the total length of the time series. The attention weight vector is the transpose. Let be the attention weight at time t. This design, based on the voltage standard deviation, focuses the network on the mid-to-late stages of the charge state. The abnormal fluctuation area.

[0094] Finally, feature representation is optimized through residual connections and layer normalization:

[0095]

[0096] in, For output features, The original temporal features output by the temporal attention module. These are the learnable weight parameters in the residual connection layer. This is the residual mapping function.

[0097] A periodic Dropout strategy is employed to randomly mask neurons with a 30% probability, generating a 128-dimensional time-domain feature vector that is then fed into the time-frequency fusion module.

[0098] For output features Offset of the charging / discharging platform Voltage fluctuation intensity and capacity decay rate The relationship is explained as follows:

[0099] Charge / discharge platform offset Defined as the discharge curve at Average voltage in the plateau region Reference values ​​for health status The difference:

[0100]

[0101] in, The number of sampling points in the platform area. For the first Point voltage characterizes the energy loss caused by an internal short circuit.

[0102] Voltage fluctuation intensity is quantified by the standard deviation of voltage in the plateau region:

[0103]

[0104] in, For sequence length, The average voltage in the plateau region. Increase the reflection of local current disturbances caused by short circuits.

[0105] The capacity decay rate is defined based on the capacity decay ratio:

[0106]

[0107] in, For the initial capacity, The current loop capacity is determined by... Sequence integral estimation.

[0108] Step S500: Establish a multi-scale frequency-aware convolutional neural network to analyze EIS and capture frequency domain responses such as impedance collapse and phase angle anomalies in the mid-to-low frequency range. The multi-scale frequency-aware convolutional neural network includes a feature preprocessing layer, a multi-scale convolutional block, a frequency channel attention module, and a feature compression layer. The input is the electrochemical impedance spectrum corresponding to a single charge-discharge cycle, including the real part. and the virtual part Data, time-domain tensor ,in For batch size, The data is standardized and Fourier smoothing filters are applied to suppress high-frequency noise, representing the number of frequency points.

[0109]

[0110] in, For the mean and standard deviation of the data, It is a Gaussian low-pass kernel. Determined through grid search;

[0111] The feature preprocessing layer passes through convolution Increase feature dimensions:

[0112]

[0113] This operation enhances the linear separability of frequency domain features;

[0114] The multi-scale convolutional blocks perform convolution in parallel. , , Feature extraction at three scales:

[0115]

[0116] The output features are concatenated along the channel dimension. ,in Convolution captures local impedance fluctuations. Convolutional recognition of mid-range relaxation patterns Convolution extracts the global frequency domain trend; ReLU is a linear rectified function, and BN is a batch normalization function.

[0117] The frequency channel attention module To perform frequency-domain sensitive weighting, first generate a global spectrum descriptor:

[0118]

[0119] Calculate channel weights by sharing a fully connected network:

[0120]

[0121] in It is the ReLU activation function. For the Sigmoid function, This indicates a splicing operation, where b1 and b2 are the bias terms. , (reduction rate) The final weighted feature is:

[0122]

[0123] This mechanism strengthens Contributes to the mid-to-low frequency range, accurately capturing impedance collapse and phase angle anomalies;

[0124] The feature compression layer passes through Convolutional dimensionality reduction and global average pooling The subsequent processing yields the frequency domain feature vector F. f vector encoding :

[0125]

[0126] The vector encoding features low-frequency impedance collapse, phase angle shift, and relaxation time distribution distortion, which are used for subsequent time-frequency fusion.

[0127] Step S600: A dynamic gating fusion module is used to weight and select the time-domain feature vector and the frequency-domain feature vector. Then, a time-frequency feature dependency mapping is established through a cross-attention mechanism to generate a fused feature vector, which is then input into the classification discriminator to complete the internal short-circuit fault detection. The time-domain feature vector... Generated by a temporal attention-enhanced long short-term memory network, frequency domain feature vector Generated by a multi-scale frequency-aware convolutional neural network;

[0128] The dynamic gating fusion module performs adaptive weighted selection:

[0129]

[0130] Wherein, the time-domain feature vector F t Specifically, the output feature in formula (19) Frequency domain eigenvector F f Specifically, the vector encoding in formula (29) , It is the ReLU activation function. For the Sigmoid function, the weight matrix , , , Bias term b t1 b t2 b f3 and b f4 Generate gate weights G as learnable parameters. t and G f Post-execution feature weighting:

[0131]

[0132] This module dynamically adjusts the weight allocation according to the degree of battery aging, strengthening the contribution of frequency domain features in the early stage of aging and strengthening the contribution of time domain features in the late stage of aging.

[0133] The cross-attention mechanism establishes a time-frequency feature dependency mapping, which weights the time-domain features. as query vector Weighted frequency domain features Zero-fill up to 128 dimensions: Set key-value pairs ; Calculate multi-head attention output:

[0134]

[0135] in Let Q be the scaling factor, K be the query vector, K be the key vector, and V be the value vector, respectively. The output should be... Encoding time-frequency features with nonlinear interaction;

[0136] The feature fusion layer concatenates the attention output with the original temporal features to obtain... :

[0137]

[0138] The compressed fused feature vector is obtained by reducing the dimensionality to 192 through a fully connected layer. :

[0139]

[0140] The classification discriminator receives Output the internal short-circuit probability P isc :

[0141]

[0142] in W c Let b be the classification layer weight matrix. h For the hidden layer bias term, b cls For the classification layer bias term, The activation function is ReLU, the Dropout probability is 0.5, and the final binary classification threshold is 0.5.

[0143] Figure 4 The comparison results of the detection performance of the present invention and the comparative method are presented. Through this example, the detection rate of the present invention for internal short-circuit faults is significantly better than that of traditional Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs). This indicates that the present invention not only improves the detection sensitivity but also reduces the possibility of false alarms.

[0144] like Figure 5 As shown, Figure 5This is a functional block diagram of an embodiment of the lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion provided by the present invention. In this embodiment, the lithium-ion battery internal short-circuit fault detection system based on time-frequency fusion provided by the present invention includes a measurement module 10, a model building module 20, a fault injection module 30, a first network module 40, a second network module 50, and a fault detection module 60. The measurement module 10 is used for data preparation, acquiring time-series data of the lithium-ion battery charging and discharging process in real time through voltage and current sensors, and measuring the electrochemical impedance spectrum of the battery under different states of charge using an electrochemical workstation. The model building module 20 is used to establish a fractional-order equivalent circuit model that integrates conductivity loss, active material loss, and lithium inventory loss, and fitting the model parameter evolution based on a publicly available dataset. The system employs a verification function to dynamically simulate the aging process; a fault injection module 30 randomly injects internal short-circuit faults of varying degrees into the model, generating time-domain and frequency-domain data across multiple cycles and aging stages; a first network module 40 uses a temporal attention-enhanced long short-term memory network to analyze the temporal dependencies of voltage and current sequences, extracting features such as charge / discharge plateau shift and voltage fluctuation enhancement; a second network module 50 uses a multi-scale frequency-aware convolutional neural network to analyze electrochemical impedance spectroscopy, capturing mid-to-low frequency impedance collapse and phase angle anomalies in the frequency domain response; and a fault detection module 60 uses a dynamic gating fusion module to weight and select time-domain and frequency-domain features, establishes a time-frequency feature dependency mapping through a cross-attention mechanism, generates a fused feature vector, and inputs it into a classification discriminator to complete internal short-circuit fault detection.

[0145] The specific functions of the measurement module 10, model building module 20, fault injection module 30, first network module 40, second network module 50 and fault detection module 60 are as shown in the above method, and will not be repeated here.

[0146] This embodiment discloses a time-frequency fusion-based method and system for detecting internal short-circuit faults in lithium-ion batteries, applicable to the safety monitoring of lithium-ion battery energy storage systems. Methods based on single time-domain or frequency-domain features are currently the main means of internal short-circuit detection. However, during battery aging, early internal short-circuit fault signals caused by lithium dendrite growth are weak and highly concealed, easily masked by operating noise. Simultaneously, aging-induced conductivity loss, active material loss, and lithium inventory loss mechanisms can couple and interfere with fault characteristics, leading to insufficient detection sensitivity and high false alarm rates in traditional methods. Furthermore, the features of time-domain and frequency-domain methods are fragmented, making it difficult to collaboratively capture the complete characterization of early faults. To address these issues, this invention first constructs a fractional-order equivalent circuit model integrating electrochemical aging mechanisms. This model, combined with empirical aging formulas, simulates the paths of conductivity decline, active material loss, and lithium inventory reduction, achieving joint modeling of the aging process and random internal short circuits. Multi-cycle voltage-current and electrochemical impedance spectroscopy data are then obtained. Secondly, a time-domain attention-enhanced long short-term memory network is used to extract time-domain features, and a multi-scale frequency-aware convolutional neural network is employed to analyze frequency-domain information. Subsequently, a dynamic gating fusion module is used to weightedly filter two types of modal features, and a cross-attention mechanism is introduced to establish a dependency mapping between time- and frequency-domain features. Finally, the internal short-circuit fault detection result is output. Simulation results comparing this method with other algorithms demonstrate that the proposed method significantly improves the internal short-circuit detection sensitivity under strong noise and multiple aging scenarios, effectively extends the thermal runaway handling window, and significantly reduces the false alarm rate, providing reliable technical support for battery safety testing.

[0147] In the embodiments provided by this invention, the fault detection method and its corresponding steps can be implemented in the form of a software program, which can be stored in a computer-readable storage medium. The aforementioned software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 of the technical features; and these modifications 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 the present invention.

Claims

1. A lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion, characterized by, The method comprises the following steps: Step S100: data preparation, collecting time series data in the charging and discharging process of the lithium ion battery in real time through a voltage sensor and a current sensor, and measuring the electrochemical impedance spectrum EIS of the battery at different states of charge through an electrochemical workstation; Step S200: establishing a fractional order equivalent circuit model fusing the loss of conductivity, the loss of active material and the loss of lithium inventory, fitting an experience function of model parameter evolution based on a public data set, and realizing dynamic simulation of the aging process; Step S300: injecting different degrees of internal short circuit faults into the model to generate time domain data and frequency domain data in multiple cycles and multiple aging stages; Step S400: establishing a time series attention enhanced long short-term memory network to analyze the time series dependence of the voltage and current sequence, and extracting the features of the charging and discharging platform offset and the enhanced voltage fluctuation; The multi-period time sequence data of voltage, current and state of charge is parsed using a time sequence attention enhanced long short-term memory network, the network extracts cross-period time sequence dependent features through a bidirectional long short-term memory layer, a forward propagation process thereof is dynamically regulated by a gating mechanism, a forgetting gate dynamically adjusts a history memory retention rate according to a voltage deviation, and finally generates a time domain feature vector F after optimization t ; Step S500: Establishing a multi-scale frequency-aware convolutional neural network to analyze EIS, capturing the frequency domain response of the impedance collapse and phase angle anomaly in the low frequency; the multi-scale frequency-aware convolutional neural network comprises a feature preprocessing layer, a multi-scale convolutional block, a frequency channel attention module and a feature compression layer, the input is the electrochemical impedance spectrum corresponding to a single charge-discharge cycle, and the output is a frequency domain feature vector F f ; Step S600: using a dynamic gating fusion module to perform weighted selection on the time domain feature vector F t and the frequency domain feature vector F f , and then establishing a time-frequency feature dependency mapping through a cross-attention mechanism to generate a fusion feature vector input into a classification discriminator to complete internal short circuit fault detection.

2. The method for detecting internal short circuit fault in Li-ion battery based on time-frequency fusion according to claim 1, characterized in that, Step S200 specifically comprises the following steps: Step S210: constructing a fractional order equivalent circuit model by fusing the three aging mechanisms of the loss of conductivity, the loss of active material and the loss of lithium inventory; The fractional order equivalent circuit model is a third-order fractional order RC equivalent circuit, in which a first parallel branch composed of a fractional order inductor L, an ohmic internal resistance R0 and a constant phase element CPE1, a second parallel branch composed of a resistance R2 and a constant phase element CPE2, and a third parallel branch composed of a resistance R3 and a constant phase element CPE3 are connected in series; Model parameters Dynamic evolution with cycle number Fitting and transfer according to public data sets, finally through the following formula group: The loss of conductivity branch simulates the growth of ohmic internal resistance caused by the corrosion of current collector and the shedding of conductive agent: （1） The active material loss branch simulates the increase of polarization resistance caused by electrode pulverization and the attenuation of double-layer capacitance: The lithium inventory loss branch represents the exponential growth of diffusion impedance and the attenuation of interface capacitance caused by the consumption of active lithium by side reactions: wherein, is a random perturbation term, each coefficient to , to , determined by fitting to a published aging dataset and satisfying a capacity fade constraint: In the formulas (1) to (4), C(N) and Co are the discharge capacity of the battery at the Nth cycle and the initial discharge capacity of the battery, respectively, and β are the capacity fade rate coefficient and the cycle aging acceleration factor, respectively, , , , are the resistance values of R0 to R3 after the Nth cycle, respectively, , , , are the resistance values of R0 to R3 before the discharge of the battery, respectively, , are the parameter values of CPE2 and CPE3 after the Nth cycle, respectively; Step S220: the state space equation of the model is derived based on Caputo fractional differential theory; the state variable of the model is defined as wherein corresponding to the polarization voltage of the three parallel branches respectively, is the state of charge, and subscript k is the current time; the discrete form of the state space equation is represented as: wherein is the current input, is the set of model parameters, is the process noise; Newton binomial coefficient matrix is determined by the fractional order whose elements satisfy: State transition function The expression for the state transition function is measuring function the expression y k is: wherein subscript m is an integer having a value of 1 to 3, is an open circuit voltage function, is a coulombic efficiency, is a rated capacity, is a measurement noise, is a sampling time interval, is a Gamma function. 3.The lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion according to claim 2, characterized in that, Step S300 specifically comprises the following steps: Step S310: The internal short circuit fault injection mechanism injects the internal short circuit fault by connecting a variable resistor in parallel across the fractional order equivalent circuit model The implementation, whose fault state total impedance function is represented as: wherein, Z(N) is the impedance function of the battery after the Nth cycle, short circuit resistance The value strategy of Z(N) is defined as: normal state is set to The simulation of the open circuit state, the fault state is randomly sampled from the uniform distribution to cover the soft short circuit to the hard short circuit scene; Step S320: The multi-source data generation process includes the cooperative generation of time domain and frequency domain data: In the constant current discharge process, the terminal voltage response is solved by iterative solution of the fractional order discrete state space equation, in which the calculation formula of the polarization voltage component is: The terminal voltage output is: wherein L is a fractional inductance value, is a polarization current component under SoC(n) is the state of charge at the current sampling instant n. The sampling frequency is set to 1 Hz, and 3600 groups of time domain data points are generated in a single discharge cycle to form a time domain data matrix Each row contains voltage, current and SOC values to form a matrix: wherein the superscript denotes the th charge-discharge cycle; The electrochemical impedance spectrum is calculated based on the current parameters after each cycle, and the frequency points , , are taken as values and distributed on a logarithmic scale, and the impedance function is passed through Generating a frequency domain data matrix Each row contains the real and imaginary parts of the impedance: The data enhancement process introduces working condition noise disturbance, time domain data adds Gaussian white noise , frequency domain data adds relative error , sample annotation adopts quadruple Record the cycle number, fault state flag, short circuit resistance value and fault trigger point nearest frequency index in turn , is the frequency value of the pth frequency point, is the characteristic frequency corresponding to the fault trigger time; Step S330: Dynamic aging-failure coupling mechanism is realized by a loop driven process: first, update the aging parameters , when the number of cycles meets , set , then perform 1C rate constant current discharge and record time domain data, calculate EIS spectrum at current SOC to generate frequency domain data, finally, increment the cycle counter and repeat the process; Key parameter settings include total cycle number 100 covering the full life cycle aging, time domain sampling frequency 1 Hz, EIS according to ten frequency points take 50 logarithmic distribution frequency points, noise intensity is set to Simulate sensor error, short-circuit resistance range covers 5-500Ω; Physical mechanism, internal short circuit bypass current Abnormal loss of discharge capacity In the time domain, the voltage platform decays in advance; in the frequency domain, the response distortion is quantified by the phase angle: Within the characteristic frequency interval Hz, the following holds as a fault sensitive region; the resulting multi-cycle data set structure is .

4. The lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion according to any one of claims 1-3, characterized in that, Step S400 specifically comprises: The application discloses a method for analyzing multi-period time series data of voltage, current and state of charge by using a long short-term memory network enhanced by time attention, wherein the input is a time domain tensor formed by discharge cycles , wherein is a batch size, is a single cycle time series length, and the feature dimension includes voltage , current , state of charge and time stamp , the network extracts cross-period time series dependent features through a bidirectional long short-term memory layer, and the forward propagation process is dynamically regulated by a gating mechanism, wherein the forget gate is dynamically adjusted according to the voltage deviation to reduce the activation value of the forget gate to inhibit the memory of the historical platform voltage when is detected, is the voltage prediction value at the current moment, is the voltage standard deviation, and the input gate then strengthens the current abnormal response feature, and the gating calculation follows: wherein, is a sigmoid activation function, including W f , W i and W o weight matrices and bias matrices f , b i and b o are learnable parameters, and residual connections are implemented through to fuse information across layers, is a layer normalization function, is the hidden state of the l-th layer at time t, and the superscript l is the index of the neural network layer.​ forward and backward hidden states , after splicing form 128-dimensional feature tensor , then use the timing attention mechanism to weight and aggregate the key discharge platform area features, the timing attention mechanism module focuses on the hidden state sequence, calculates the attention score: （ 17） The attention weight distribution is determined by the voltage statistical characteristics and the second derivative, and the calculation formula is: wherein is the average platform voltage, T is the total length of the time series, is the transpose of the attention weight vector, is the attention weight at time t, is the voltage standard deviation, this design makes the network focus on the abnormal fluctuation area of the middle and later state of charge interval ​ Finally, the feature expression is optimized through residual connection and layer normalization: wherein, is an output feature, is an original temporal feature output by the temporal attention module, is a learnable weight parameter in the residual connection layer, is a residual mapping function; And a periodic Dropout strategy is adopted to randomly shield neurons with a probability of 30%, generating a 128-dimensional time domain feature vector to be delivered to the time-frequency fusion module.

5. The lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion according to claim 4, characterized in that, Step S400 further comprises: For output characteristics With the charge and discharge platform offset , voltage fluctuation intensity And the relationship between the capacity attenuation rate Explained as follows: Charge and discharge plateau shift defined as the difference between the average voltage of the discharge curve in the plateau region and the average voltage of the charge curve in the plateau region Plateau region average voltage from the state of health reference value the difference wherein, is the number of sampling points for the platform region, is the voltage at the first point voltage, characterizing the energy loss due to internal short circuit; The voltage fluctuation intensity is quantified by the standard deviation of the voltage in the platform region: wherein, is the sequence length, is the average voltage of the platform region, increases the local current disturbance caused by the short circuit; The capacity decay rate is defined based on the capacity decay rate: wherein, is the initial capacity, is the current cycle capacity, by sequence integral estimation. 6.The lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion according to claim 5, characterized in that, Step S500 specifically comprises: The input of the multi-scale frequency-aware convolutional neural network is the electrochemical impedance spectrum corresponding to a single charge-discharge cycle, including real part and imaginary part Data, time domain tensor , wherein is the batch size, is the number of frequency points, the data is standardized and Fourier smoothing filter is applied to suppress high frequency noise: wherein, is the data mean and standard deviation, is a Gaussian low-pass kernel, determined by grid search; The feature preprocessing layer is configured to convolution lifting feature dimension: This operation enhances the linear separability of the frequency domain features; The multi-scale convolution block performs convolution in parallel , , Three scales of feature extraction: The output features are spliced along the channel dimension as wherein The convolution captures local impedance fluctuations, The convolution identifies mid-range relaxation patterns, The convolution extracts global frequency domain trends, ReLU is a linear rectifier function, and BN is a batch normalization function. The frequency channel attention module is used for sensitive weighting in frequency domain, and a frequency spectrum global descriptor is first generated: The channel weight is calculated through a shared fully connected network: where is the ReLU activation function, is the Sigmoid function, denotes the concatenation operation, and b1and b2are bias terms, , (reduction rate ), and the final weighted feature is: This mechanism strengthens The middle and low frequency region contributes to accurately capture impedance collapse and phase angle anomaly; The feature compression layer is configured to compress the feature vector F Convolutional dimension reduction and global average pooling The processing obtains subsequent frequency domain feature vectors F f Vector encoding : This vector encodes the features of low-frequency impedance collapse, phase angle offset and relaxation time distribution distortion, which are used for subsequent time-frequency fusion.

7. The lithium-ion battery internal short-circuit fault detection method based on time-frequency fusion according to claim 6, characterized in that, Step S600 specifically comprises: The fusion of the time domain feature and the frequency domain feature is realized through a dynamic gating fusion module and a cross attention mechanism; the time domain feature vector is generated by a time sequence attention enhanced long short-term memory network, and the frequency domain feature vector is generated by a multi-scale frequency perception convolutional neural network; The dynamic gating fusion module performs adaptive weighted selection: wherein the time-domain feature vector F t Specifically, the output feature in formula (19) , the frequency-domain feature vector F f Specifically, the vector encoding in formula (29) , is a ReLU activation function, is a Sigmoid function, the weight matrix , , , , the bias term b t1 , b t2 , b f3 , and b f4 are learnable parameters, the generation gating weight G t , and G f perform feature weighting: This module dynamically adjusts the weight distribution according to the aging degree of the battery, strengthens the contribution of frequency domain features in the early stage of aging, and strengthens the contribution of time domain features in the late stage of aging; The cross-attention mechanism establishes a time-frequency feature dependency mapping, and weights the time-domain features as query vectors ; weight the frequency-domain features zero-padded to 128 dimensions: ; set key-value pairs ; calculate multi-head attention output: wherein is a scaling factor, Q, K and V are the query vector, key vector and value vector respectively, and output encoding the time-frequency features non-linearly The feature fusion layer splices attention output and original time domain features to obtain : dimensioned to 192 dimensions by a fully connected layer to obtain compressed fused feature vectors : The classification discriminator receives and outputs an internal short circuit probability P isc : where , W c is the classification layer weight matrix, b h is the hidden layer bias term, b cls is the classification layer bias term, is the ReLU activation function, the Dropout probability is 0.5, and the final binary classification threshold is 0.

5.

8. A time-frequency fusion based short circuit fault detection system for lithium ion battery, characterized in that, It comprises: A measurement module (10) is used for data preparation, collecting time series data in the charging and discharging process of the lithium ion battery in real time through a voltage sensor and a current sensor, and measuring the electrochemical impedance spectrum EIS of the battery at different states of charge through an electrochemical workstation; The model establishing module (20) is configured to establish a fractional-order equivalent circuit model fusing conductive loss, active material loss and lithium inventory loss, fit an evolution experience function of model parameters based on a public data set, and realize dynamic simulation of an aging process. The fault injection module (30) is configured to randomly inject internal short circuit faults of different degrees in the model, and generate time domain data and frequency domain data of multiple cycles and multiple aging stages. The first network module (40) is used for establishing a time sequence attention enhanced long short-term memory network to analyze the time sequence dependency of the voltage and current sequence, and extracting the features of the charge and discharge platform offset and voltage fluctuation enhancement; the time sequence attention enhanced long short-term memory network is used for analyzing the multi-period time sequence data of the voltage, current and state of charge; the network extracts the cross-period time sequence dependency features through a bidirectional long short-term memory layer, and the forward propagation process is dynamically regulated by a gating mechanism; the forget gate dynamically adjusts the history memory retention rate according to the voltage deviation, and finally generates the time domain feature vector F after optimization t ​ The second network module (50) is used for establishing a multi-scale frequency-aware convolutional neural network to analyze EIS, and capturing the frequency domain response of the impedance collapse and the phase angle anomaly in the middle and low frequencies. The multi-scale frequency-aware convolutional neural network comprises a feature preprocessing layer, a multi-scale convolutional block, a frequency channel attention module and a feature compression layer, the input is an electrochemical impedance spectrum corresponding to a single charge-discharge cycle, and the output is a frequency domain feature vector F f . The fault detection module (60) adopts a dynamic gate fusion module to perform weighted selection on the time domain feature vector F t and the frequency domain feature vector F f , and then establishes a time-frequency feature dependency mapping through a cross-attention mechanism to generate a fusion feature vector input into a classification discriminator to complete internal short circuit fault detection.

9. A time-frequency fusion based short circuit fault detection system for lithium ion battery, characterized in that, The method comprises the following steps: at least one processor; and at least one memory connected in communication with the processor, wherein: the memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the lithium ion battery internal short circuit fault detection method based on time-frequency fusion according to any one of claims 1 to 7.

Images