Lithium-ion battery pack connection fault diagnosis method combined with model and deep learning

Abstract

This invention belongs to the field of energy system fault diagnosis technology and discloses a lithium-ion battery pack connection fault diagnosis method that integrates model and deep learning. The method includes the following steps: collecting and processing voltage and current data of each battery cell; training a CAE-LSTM model to regress the connection fault magnitude of the battery cell; constructing a dimensionality-reduced equivalent circuit model of the battery cell to estimate its internal resistance, comparing it with the average internal resistance of the battery cell to determine the connection fault, thus achieving the first layer of fault diagnosis; fusing the voltage and current data before and after the current jump with the online voltage and current data, and inputting this data into the trained CAE-LSTM model to regress the fault magnitude of the current battery cell; and comparing the fault magnitude with a preset threshold to determine the second layer of fault diagnosis. This invention employs the above-mentioned connection fault diagnosis method, balancing the complexity and accuracy of the algorithm, and ensuring robustness in the presence of measurement noise.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for aircraft energy systems, and in particular to a method for diagnosing faults in lithium-ion battery pack connections that integrates modeling and deep learning. Background Technology

[0002] In recent years, with the increasing strategic value of near-space, research on near-space vehicles has developed rapidly. As a high-altitude pseudo-satellite platform, near-space vehicles have broad application prospects in many fields such as communication relay, reconnaissance and surveillance, surveying and mapping, and weather forecasting. To fully leverage their unique advantages, these vehicles need to possess strong long-duration endurance while maintaining a predetermined cruising altitude.

[0003] In this context, the energy management system of a near-space vehicle can be viewed as a DC microgrid structure composed of multiple photovoltaic energy storage nodes. Each photovoltaic energy storage node includes photovoltaic modules, a maximum power point tracking control module, a DC-DC converter, a lithium-ion battery pack, and a load. Among them, the lithium-ion battery pack, as the core energy storage unit, not only provides stable power support for the normal operation of the vehicle, but is also a key component for achieving continuous day and night flight and long-endurance operation capabilities.

[0004] Energy storage systems typically consist of numerous lithium-ion battery cells connected in series and parallel. The safety of lithium-ion batteries remains a persistent challenge, as battery failures can severely impact mission efficiency and limit long-endurance flight capabilities. During operation, vibration and aging of connectors can lead to loosening between battery electrodes and connecting components, resulting in battery pack connection failures. When connection failures occur in the battery system, the increased contact resistance not only leads to energy loss but also further amplifies inconsistencies caused by increased heat generation, ultimately potentially causing thermal runaway.

[0005] Existing battery fault diagnosis methods have several limitations. Data-driven methods typically rely on strong assumptions about fault modes and require large amounts of high-quality data, making them sensitive to measurement noise and difficult to assess fault severity. Model-based methods, while interpretable and computationally efficient, are inevitably affected by modeling errors and measurement noise, potentially leading to false positives or false negatives in practical applications. Therefore, there is an urgent need for a method that balances algorithmic complexity and accuracy while ensuring robustness against measurement noise in lithium-ion battery pack fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to provide a lithium-ion battery pack connection fault diagnosis method that integrates model and deep learning, balancing the complexity and accuracy of the algorithm, and ensuring robustness in the presence of measurement noise.

[0007] To achieve the above objectives, this invention provides a lithium-ion battery pack connection fault diagnosis method that integrates modeling and deep learning, comprising the following steps: Step S1, Data Acquisition and Model Training Stage: Collect and process the voltage and current data of each battery cell in the lithium-ion battery pack. The processed data is used to train the CAE-LSTM model to regress the connection fault magnitude of the battery cell. Step S2, First-level online fault diagnosis stage: Based on the equivalent circuit model of a battery cell, according to the load current change amplitude, a reduced-dimensional equivalent circuit model of the battery cell is constructed to estimate the internal resistance of the battery cell. The estimated internal resistance of the battery cell is compared with the average internal resistance of the battery cell to determine the connection fault of the battery cell and realize the first-level fault diagnosis. Step S3, Second-layer online fault diagnosis stage: The voltage and current data before and after the current jump are fused with the online voltage and current data. The fused and normalized data is input into the trained CAE-LSTM model to regress the fault amplitude of the current battery cell. The fault amplitude is compared with a preset threshold to determine whether a connection fault has occurred, thus realizing the second-layer fault diagnosis.

[0008] Preferably, in step S1, the data acquisition and model training phase, the specific process is as follows: Step S11: Based on the lithium-ion battery pack fault diagnosis experimental platform, inject faults into individual battery cells according to the preset fault range, and collect the current and voltage of each corresponding individual battery cell. Step S12: Divide the collected current and voltage data to form a dataset for training and validating the CAE-LSTM model; Step S13: Normalize the battery operation data using a sliding window, and train a CAE-LSTM model based on the processed data to regress the connection fault magnitude of the battery cell. Within the window, the voltage of each battery cell at each moment is normalized by subtracting the average voltage of each battery cell from the voltage of each battery cell at each moment. The normalized voltage and current are used as inputs to the CAE-LSTM model for battery pack connection fault diagnosis.

[0009] Preferably, the input data for the CAE-LSTM model is as follows: ; ; ; in, and It is the moment of each battery cell Normalized voltage and current; and It is the moment of each battery cell Normalized voltage and current; and It is the moment of each battery cell Normalized voltage and current; It is the width of the window; This is the input data for the CAE-LSTM model; It is an index of the individual battery cells; It represents the total number of individual battery cells.

[0010] Preferably, in the CAE-LSTM model, the convolutional autoencoder (CAE) consists of an encoder and a decoder; the encoder compresses information by encoding the input data into a low-dimensional representation of the latent vectors in the hidden layer. The encoder consists of two-dimensional convolutional layers, activation layers, and convergence layers; the two-dimensional convolutional layer is represented as: ; in, For multi-channel output of a two-dimensional convolutional layer; Input data; and For channel Weight matrix and bias on; This represents the transpose convolution operation; It is the ReLU activation function; For channel indexing; Total number of channels; The decoder consists of deconvolutional layers and activation layers; the deconvolutional layer is represented as: ; in, This represents the output feature map of the transposed convolutional layer. Indicates the th from the previous convolutional layer One input feature map; Indicates corresponding to the first Transposed convolution kernels for each channel; This indicates the bias term.

[0011] Preferably, the input sequence is processed by CAE to obtain a denoised time series, and the time series is used as the input of the Long Short-Term Memory (LSTM) cells stacked after CAE; the LSTM integrates the fault information at each time step and regresses the wiring fault magnitude. An LSTM unit consists of an input layer, a storage layer, and an output layer; the equations for an LSTM unit are defined as follows: ; in, This represents the input at the current moment; Represents the short-term hidden state at the current moment (output); This represents the current long-term storage state; long-term storage state and short-term hidden state Together they form the storage cells in LSTM, storing the state for a long time. Combine the current input with historical information from all previous time points. Used as the short-term hidden state and output of the LSTM unit; This indicates the long-term storage state at the previous moment; Represents the short-term hidden state at the previous moment (output); Indicates the candidate cell state; The sigmoid activation function is represented by tanh; the hyperbolic tangent function is represented by tanh. Long-term storage status and short-term hidden state Controlled by three gating structures: forget gate Input gate and output gate ; , , and These represent the inputs corresponding to the current time. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and These represent the short-term hidden states corresponding to the previous time step. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and The bias vectors representing the states of the forget gate, input gate, output gate, and candidate cells; One-dimensional convolutional layers and fully connected layers are used to regress the magnitude of connection faults.

[0012] Preferably, in step S2, the first-level online fault diagnosis stage, the specific process is as follows: Step S21: Based on the equivalent circuit model of a battery cell, construct a dimension-reduced equivalent circuit model of the cell according to the load current change amplitude between the current sampling time and the previous sampling time, record the voltage and current data before and after the current jump, and estimate the internal resistance of the battery cell. Step S22: Compare the estimated internal resistance of the battery cell with the average internal resistance of the battery cell. When the estimated internal resistance exceeds the preset threshold, it is determined that the battery cell has a connection fault, realizing the first layer of fault diagnosis and triggering the second layer of fault diagnosis.

[0013] Preferably, according to Kirchhoff's laws, the equivalent circuit model of a single battery cell is as follows: ; in, This is the open-circuit voltage; For ohmic internal resistance, Equivalent fault resistor; , and These are polarization resistance, polarization capacitance, and polarization voltage, respectively. This is the terminal voltage; This is the current flowing through the battery; For time; Battery voltage consists of three parts, including the state of charge. The open-circuit voltage affected by the ohmic internal resistance and the battery voltage drop caused by the equivalent fault resistance, as well as the polarization voltage generated by the polarization resistance and polarization capacitor; connection faults are identified by the abnormal internal resistance of the battery.

[0014] Preferably, a single-unit dimension-reduced equivalent circuit model is established, and the specific process is as follows: When the change in load current between the current moment and the previous moment exceeds a given threshold, the ratio of the voltage difference to the current difference between the two moments is used as the estimated battery internal resistance, and the voltage and current before and after the current jump are recorded. When the change in load current at the current moment is less than a given threshold, the voltage value of the dimensionality-reduced model is obtained by removing the average dynamic response from the battery cell voltage, and the ratio of the voltage difference to the current difference in the dimensionality-reduced model before and after the current jump is used as the estimated battery internal resistance. Let the change in load current The equivalent circuit model of a single unit with reduced dimensions is as follows: ; ; ; ; ; in, and These represent the time intervals of the i-th battery cell. The current and voltage, ; and These represent the current and voltage of a single battery cell before the current mutation occurred; and These represent the current and voltage of a single battery cell after the mutation occurs; The threshold value for current change; This represents the estimated internal resistance of each individual battery cell; The formula for calculating the average internal resistance of a single cell is as follows: ; when Greater than the given fault diagnosis internal resistance threshold At that time, the first-level fault diagnosis determines that a connection fault has occurred in battery cell i, and triggers the second-level fault diagnosis to achieve accurate fault amplitude estimation.

[0015] Preferably, in step S3, the second-level online fault diagnosis stage, the specific process is as follows: Step S31: After a fault is detected in the first-level online fault diagnosis stage, the corresponding voltage and current data before and after the current jump collected and recorded in the first-level online fault diagnosis stage are fused with the online operating voltage and current data obtained through the sliding window mechanism to construct a time series dataset containing current jump characteristics. Step S32: Normalize the fused multidimensional voltage and current data to eliminate the impact of dimensional differences and amplitude fluctuations on the model regression accuracy. Step S33: Input the normalized data into the pre-trained CAE-LSTM model to calculate the fault magnitude of the battery cell at the current time. Perform regression estimation; Step S34: Estimate the fault amplitude output by the CAE-LSTM model. With fault diagnosis internal resistance threshold Comparison, when exceeding When a fault is detected in unit i, output is generated. This enables the diagnosis and identification of faults in the second-layer battery cells.

[0016] Therefore, the present invention employs a lithium-ion battery pack connection fault diagnosis method that integrates the aforementioned model with deep learning, and the beneficial effects are as follows: (1) This invention integrates the dimension-reduced equivalent circuit model and the deep learning model. The model-driven method enables rapid estimation of the internal resistance of a single battery cell and anomaly localization. After an anomaly is detected, a data-driven method is introduced for fine diagnosis, thus realizing hierarchical and phased fault diagnosis.

[0017] (2) The present invention effectively reduces computational complexity by adopting a dimension-reduced equivalent circuit model; at the same time, the CAE-LSTM-based deep learning model is used for fault amplitude estimation, which improves diagnostic accuracy and ensures both efficiency and diagnostic performance.

[0018] (3) The CAE-LSTM model used in this invention can effectively extract and filter the measurement noise in voltage, current and other signals, thereby reducing the false alarm rate and the missed alarm rate and improving the reliability of fault diagnosis results.

[0019] (4) The present invention can maintain high diagnostic accuracy even in the presence of measurement noise and inconsistent battery parameters, and is applicable to actual battery pack operating conditions.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a lithium-ion battery pack connection fault diagnosis method that integrates modeling and deep learning according to the present invention. Figure 2 This is a schematic diagram of the LSTM network unit structure used for time series data processing in this invention; Figure 3 This is a schematic diagram of the CAE-LSTM network structure used in the second-layer fault diagnosis of this invention; Figure 4 This is a schematic diagram of the equivalent circuit model of a single battery cell in this invention; Figure 5 This is a schematic diagram of the battery cell connection structure in the lithium-ion battery pack studied in this invention. Figure 6 This is a schematic diagram of the experimental platform structure used in the embodiments of the present invention; Figure 7 This is a schematic diagram of the first-level fault diagnosis results when the algorithm is executed recursively in an embodiment of the present invention; wherein, (a) is the algorithm result of the first-level online fault diagnosis under 3 mV / 3 mA Gaussian noise; (b) is the algorithm result of the first-level online fault diagnosis under 6 mV / 6 mA Gaussian noise; Figure 8 The present invention provides a schematic diagram of the fault amplitude estimation results in the second-layer fault diagnosis when the algorithm is executed in a recursive manner; wherein, (a) is the algorithm result of the second-layer online fault diagnosis under 3 mV / 3 mA Gaussian noise; (b) is the algorithm result of the second-layer online fault diagnosis under 6 mV / 6 mA Gaussian noise. Figure 9The diagram shows the experimental results of the comparison method in the embodiment of the present invention; wherein, (a) is the result of a comparison algorithm under 3 mV / 3 mA Gaussian noise; (b) is the result of a first comparison algorithm under 6 mV / 6 mA Gaussian noise; (c) is the result of another comparison algorithm under 3 mV / 3 mA Gaussian noise; and (d) is the result of another comparison algorithm under 6 mV / 6 mA Gaussian noise. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] like Figure 1 As shown, a method for diagnosing connection faults in lithium-ion battery packs that integrates modeling and deep learning includes the following steps: Step S1, Data Acquisition and Model Training Stage: Collect and process the voltage and current data of each battery cell in the lithium-ion battery pack. The processed data is used to train the CAE-LSTM model to regress the connection fault magnitude of the battery cell. Step S2, First-level online fault diagnosis stage: Based on the equivalent circuit model of a battery cell, according to the load current change amplitude, a reduced-dimensional equivalent circuit model of the battery cell is constructed to estimate the internal resistance of the battery cell. The estimated internal resistance of the battery cell is compared with the average internal resistance of the battery cell to determine the connection fault of the battery cell and realize the first-level fault diagnosis. Step S3, Second-layer online fault diagnosis stage: The voltage and current data before and after the current jump are fused with the online voltage and current data. The fused and normalized data is input into the trained CAE-LSTM model to regress the fault amplitude of the current battery cell. The fault amplitude is compared with a preset threshold to determine whether a connection fault has occurred, thus realizing the second-layer fault diagnosis.

[0024] Example 1 Step S1, Data Acquisition and Model Training Stage: Collect and process the voltage and current data of each battery cell in the lithium-ion battery pack. The processed data is used to train the CAE-LSTM model to regress the connection fault magnitude of the battery cell.

[0025] Step S11: Based on the lithium-ion battery pack fault diagnosis experimental platform, inject faults into individual battery cells according to the preset fault range, and collect the current and voltage of each corresponding individual battery cell. Step S12: Divide the collected current and voltage data to form a dataset for training and validation of the convolutional autoencoder and long short-term memory network model (CAE-LSTM model); Step S13: A sliding window is used to normalize the battery operation data. Based on the processed data, a CAE-LSTM model is trained to regress the connection fault magnitude of individual battery cells. Specifically, within the window, the voltage of each battery cell at each moment is normalized by subtracting the average voltage of all battery cells from the voltage at each moment. The normalized voltage and current are used as inputs to the CAE-LSTM model for accurate battery pack connection fault diagnosis.

[0026] The input data for the CAE-LSTM model is shown below: ; ; ; in, and It is the moment of each battery cell Normalized voltage and current; and It is the moment of each battery cell Normalized voltage and current; and It is the moment of each battery cell Normalized voltage and current; It is the width of the window; This is the input data for the CAE-LSTM model; It is an index of the individual battery cells; It represents the total number of individual battery cells.

[0027] In the CAE-LSTM model, the convolutional autoencoder (CAE) consists of an encoder and a decoder; the encoder compresses information by encoding the input data into a low-dimensional representation of the latent vectors in the hidden layer.

[0028] The encoder consists of two-dimensional convolutional layers, activation layers, and convergence layers; the two-dimensional convolutional layer is represented as: ; in, For multi-channel output of a two-dimensional convolutional layer; Input data; and For channel Weight matrix and bias on; This represents the transpose convolution operation; It is the ReLU activation function; For channel indexing; This represents the total number of channels.

[0029] The decoder consists of deconvolutional layers and activation layers; the deconvolutional layer is represented as: ; in, This represents the output feature map of the transposed convolutional layer. Indicates the th from the previous convolutional layer One input feature map; Indicates corresponding to the first Transposed convolution kernels for each channel; This indicates the bias term.

[0030] Since battery pack measurement data often contains noise and bias, the input sequence is processed using CAE to obtain a denoised time series. The new time series is then used as input to a Long Short-Term Memory (LSTM) array stacked after the CAE. The LSTM synthesizes fault information at each time point and regresses the magnitude of wiring faults.

[0031] An LSTM unit consists of an input layer, a storage layer, and an output layer; the structure of an LSTM unit is as follows: Figure 2 As shown. The LSTM element equations are defined as follows: ; in, This represents the input at the current moment; Represents the short-term hidden state at the current moment (output); This represents the current long-term storage state; long-term storage state and short-term hidden state Together they form the storage cells in LSTM, storing the state for a long time. Combine the current input with historical information from all previous time points. Used as the short-term hidden state and output of the LSTM unit; This indicates the long-term storage state at the previous moment; Represents the short-term hidden state at the previous moment (output); Indicates the candidate cell state; The sigmoid activation function is represented by tanh; the hyperbolic tangent function is represented by tanh. Long-term storage status and short-term hidden state Controlled by three gating structures: forget gate Input gate and output gate ; , , and These represent the inputs corresponding to the current time. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and These represent the short-term hidden states corresponding to the previous time step. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and The bias vector represents the state of the forget gate, input gate, output gate, and candidate cells.

[0032] One-dimensional convolutional layers and fully connected layers are used to regress the magnitude of connection faults. The overall network structure is as follows: Figure 3 As shown.

[0033] Step S2, First-level online fault diagnosis stage: Based on the equivalent circuit model of a battery cell, according to the load current change amplitude, a reduced-dimensional equivalent circuit model of the battery cell is constructed to estimate the internal resistance of the battery cell. The estimated internal resistance of the battery cell is compared with the average internal resistance of the battery cell to determine the connection fault of the battery cell and realize the first-level fault diagnosis.

[0034] Step S21: Based on the equivalent circuit model of a battery cell, construct a dimension-reduced equivalent circuit model of the cell according to the load current change amplitude between the current sampling time and the previous sampling time, record the voltage and current data before and after the current jump, and estimate the internal resistance of the battery cell. Step S22: Compare the estimated internal resistance of the battery cell with the average internal resistance of the battery cell. When the estimated internal resistance exceeds the preset threshold, it is determined that the battery cell has a connection fault, realizing the first layer of fault diagnosis and triggering the second layer of fault diagnosis.

[0035] The first-order equivalent circuit model of a single battery cell is as follows: Figure 4 As shown, It is open-circuit voltage. Indicates the internal resistance of the ohm. The fault resistor is caused by a wiring fault. , and These are polarization resistance, polarization capacitance, and polarization voltage. It is the terminal voltage. It is the current flowing through the battery; it is positive when charging and negative when discharging.

[0036] According to Kirchhoff's laws, the equivalent circuit model of a single battery cell is shown below: ; in, For time.

[0037] Battery voltage consists of three parts, including the state of charge (SOC). The open-circuit voltage affected by the ohmic internal resistance and the equivalent fault resistance, the battery voltage drop caused by the ohmic internal resistance and the equivalent fault resistance, and the polarization voltage generated by the polarization resistance and polarization capacitor.

[0038] When the input current changes suddenly, the battery voltage undergoes a dynamic response. This is because the connecting resistor is connected in series in the circuit. Figure 5 As shown, this also has a similar effect on the battery's internal resistance. Therefore, it is feasible to identify abnormal internal resistance of the battery as a reflection of a connection fault.

[0039] The specific steps for establishing a reduced-dimensional equivalent circuit model for a single unit are as follows: When the current changes abruptly, the voltage change in the battery is primarily caused by the battery's internal resistance. Since the dynamic response of the battery has a limited impact on the voltage over the sampling time interval, the effect of polarization voltage on the internal resistance estimation can be ignored. If the current changes slowly or remains constant, the voltage change during this process mainly comes from the change in polarization voltage. Because the batteries in a series battery pack exhibit similar dynamic responses, the average dynamic response of the batteries is used, rather than the dynamic response of each individual battery.

[0040] Based on the magnitude of the load current change, make the following judgment: When the change in load current between the current moment and the previous moment exceeds a given threshold, the ratio of the voltage difference to the current difference between the two moments is used as the estimated battery internal resistance, and the voltage and current before and after the current jump are recorded.

[0041] When the change in load current at the current moment is less than a given threshold, the voltage value of the dimensionality-reduced model is obtained by removing the average dynamic response from the battery cell voltage, and the ratio of the voltage difference to the current difference in the dimensionality-reduced model before and after the current jump is used as the estimated battery internal resistance.

[0042] Let the change in load current The equivalent circuit model of a single unit with reduced dimensions is as follows: ; ; ; ; ; in, and These represent the time intervals of the i-th battery cell. The current and voltage, ; and These represent the current and voltage of a single battery cell before the current mutation occurred; and These represent the current and voltage of a single battery cell after the mutation occurs; The threshold value for current change; This represents the estimated internal resistance of each individual battery cell.

[0043] The formula for calculating the average internal resistance of a single cell is as follows: ; when Greater than the given fault diagnosis internal resistance threshold At that time, the first-level fault diagnosis determines that a connection fault has occurred in battery cell i, and triggers the second-level fault diagnosis to achieve accurate fault amplitude estimation.

[0044] Step S3, Second-layer online fault diagnosis stage: The voltage and current data before and after the current jump are fused with the online voltage and current data. The fused and normalized data is input into the trained CAE-LSTM model to regress the fault amplitude of the current battery cell. The fault amplitude is compared with a preset threshold to determine whether a connection fault has occurred, thus realizing the second-layer fault diagnosis.

[0045] Step S31: After the fault is detected in step S2, the corresponding voltage and current data before and after the current jump collected and recorded in step S2 are fused with the online operating voltage and current data obtained through the sliding window mechanism to construct a time series dataset containing current jump characteristics. Step S32: Normalize the fused multidimensional voltage and current data to eliminate the impact of dimensional differences and amplitude fluctuations on the model regression accuracy. Step S33: Input the normalized data into the pre-trained CAE-LSTM model to calculate the fault magnitude of the battery cell at the current time. Perform regression estimation; Step S34: Estimate the fault amplitude output by the CAE-LSTM model. With fault diagnosis internal resistance threshold Comparison, when exceeding When a fault is detected in unit i, output is generated. This enables the diagnosis and identification of faults in the second-layer battery cells.

[0046] Example 2 This embodiment presents a lithium-ion battery pack connection fault diagnosis method that integrates modeling and deep learning, including the following steps: Step S1: In this embodiment, the lithium-ion battery pack fault diagnosis experimental platform includes a monitor, a temperature control device, and battery testing equipment, used to monitor the battery pack's operating status and collect data, such as... Figure 6 As shown.

[0047] During aircraft operation, mechanical vibration and aging of battery pack connection devices can lead to connection failures, typically manifested as abnormal changes in connection resistance. Therefore, during battery pack operation, connection failures are simulated by introducing additional resistors at preset locations, and the voltage and current data of each battery cell are measured and recorded during operation.

[0048] In this embodiment, a battery pack consisting of four lithium-ion battery cells connected in series is selected as the research object, and the battery pack is tested under constant temperature conditions.

[0049] First, the battery pack is charged with constant current and constant voltage according to the preset rate; then, after being left to stand, it is discharged under dynamic stress test (DST) conditions to simulate the operating state under actual application scenarios.

[0050] Under different initial voltage conditions, additional resistors of varying amplitudes were introduced into individual battery cells to construct various connection fault conditions, and corresponding battery operation data were collected at a sampling frequency of 1 Hz. The collected data was preprocessed and divided into model training and validation datasets. Part of the data was used to train the CAE-LSTM model, while the remaining data was used to validate the model's diagnostic performance.

[0051] Overall network structure as follows Figure 3 As shown, the input is a voltage and current sequence, and the output is the fault amplitude. First, the acquired voltage and current sequence is normalized within a window with a width of 32 sampling points. Next, it is input into the CAE module to obtain a denoised sequence, which is then input into the LSTM module. Finally, the fault amplitude is obtained through a one-dimensional convolutional layer and a fully connected layer. The CAE consists of an encoder and a decoder. The structure of the LSTM unit is shown below. Figure 2 As shown.

[0052] Step S2, based on the equivalent circuit model of a single battery cell, as follows: Figure 3 As shown, the battery voltage consists of open-circuit voltage, polarization voltage, and the voltage drop caused by ohmic internal resistance and equivalent fault resistance. Since the connection resistance is connected in series in the circuit, its changes affect the terminal voltage of the individual battery cells. Therefore, it is feasible to identify abnormal internal resistance of the battery as a reflection of a connection fault. Because the connected battery cells share a common current input, they exhibit similar polarization voltage changes. Considering that the open-circuit voltage of a battery cell is affected by its state of charge, its change process is relatively slow. Therefore, a reduced-dimensional equivalent circuit model of the individual cell is constructed to estimate its internal resistance.

[0053] The number of floating-point operations (FLOPs) required is used as a metric for evaluating algorithm complexity. Comparing this invention with the Recursive Least Squares (RLS) method for internal resistance estimation, the internal resistance estimation of the single-unit reduced-dimensional equivalent circuit model in this invention requires an average of 2.25 to 5.5 FLOPs per unit per step, while the RLS method requires up to 10⁸ FLOPs. Furthermore, the performance of this estimation method avoids the influence of window size and battery operating conditions on estimation accuracy in the RLS method.

[0054] In this embodiment, when the algorithm is executed recursively, the first-level fault diagnosis result is as follows: Figure 7 As shown. Gaussian noise with standard deviations of 3 mV / 3 mA and 6 mV / 6 mA was added to the voltage and current of the battery in use to simulate sensor measurement errors. To verify the robustness of the algorithm under conditions of initial SoC inconsistency between cells, the initial voltages of cells 1 and 2 were 4.05 V, and the initial voltages of cells 3 and 4 were 3.9 V. A 50 mΩ connection fault was injected into cell 2 at 500 s, and the fault diagnosis threshold was set to 30 mΩ. A connection fault was determined to exist in the battery when the estimated internal resistance of the battery exceeded the average value of the threshold. The results are as follows. Figure 7 (a) and Figure 7 As shown in (b), the estimated internal resistance of monomer 2 increases significantly at 500 s and subsequently exceeds the preset threshold. This demonstrates that the method of the present invention can achieve timely detection of connection faults.

[0055] Step S3: The voltage and current data recorded in Step S2 before and after the current jump are fused with the online voltage and current data obtained through sliding window processing; the fused data is normalized, and the normalized data is input into the CAE-LSTM model trained in Step S1 for fault amplitude estimation. The result is as follows: Figure 8 As shown. Figure 8 (a) and Figure 8 As shown in (b), the output fault amplitude of the abnormal cell 2 detected by the first-layer fault diagnosis increases significantly after 500s, exceeding the preset 30 mΩ threshold, and eventually stabilizes at about 50 mΩ. This demonstrates that the method of the present invention can effectively identify battery cell connection faults under online operating conditions and reasonably estimate the fault amplitude.

[0056] The proposed method was compared and validated against two existing fault diagnosis methods. To ensure fairness, all methods were trained, validated, and tested using the same dataset. One comparison method involved estimating the voltage of individual battery cells based on a combination of convolutional neural networks and long short-term memory networks, and then combining this with changes in the correlation coefficients between cells to determine faults. The other comparison method involved constructing multiple correlation coefficients to characterize the synchronization characteristics of individual cell voltages, converting the correlation sequences into image features using Gram angle field mapping, and then classifying the fault states using a convolutional neural network and an adaptive weighting mechanism.

[0057] Figure 9 The diagnostic results of each comparison method under the conditions of step S1 are presented. For example... Figure 9 As shown in (a), at approximately 500 s, the correlation coefficients between some individual components show a significant decrease, indicating an anomaly in the system. However, when the fault persists, the correlation coefficients subsequently recover to a relatively stable state, making it difficult for this method to reliably identify persistent connection faults. Meanwhile, as... Figure 9 As shown in (b), under fault-free operating conditions, the correlation coefficient is still affected by measurement noise and fluctuates significantly, which has an adverse effect on diagnostic stability.

[0058] like Figure 9 (c) and Figure 9 As shown in (d), the other comparison method has multiple missed detections after the fault occurs, and the missed detections are relatively concentrated in time. Its diagnostic results are easily affected by changes in load current. The fault detection rate and false positive rate on the validation set are 3.10% and 0.93%, respectively.

[0059] Furthermore, the diagnostic effectiveness of both of the aforementioned comparative methods is highly dependent on the sliding window length, thus limiting their applicability in practical scenarios. In contrast, the method of this invention exhibits more stable diagnostic performance and higher reliability in fault identification under the same test conditions.

[0060] Therefore, this invention employs a lithium-ion battery pack connection fault diagnosis method that integrates a model and deep learning, combining a dimensionality-reduced equivalent circuit model with a deep learning model. Model-driven methods enable rapid estimation of the internal resistance of individual battery cells and anomaly localization. After anomaly detection, a data-driven approach is introduced for refined diagnosis, achieving layered and phased fault diagnosis. This invention effectively reduces computational complexity by using a dimensionality-reduced equivalent circuit model. Simultaneously, the CAE-LSTM-based deep learning model is used for fault amplitude estimation, improving diagnostic accuracy and balancing efficiency with diagnostic performance. The CAE-LSTM model used in this invention can effectively extract and filter measurement noise from signals such as voltage and current, thereby reducing false alarm and false negative rates and improving the reliability of fault diagnosis results. This invention maintains high diagnostic accuracy even in the presence of measurement noise and inconsistent battery parameters, making it suitable for actual battery pack operating conditions.

[0061] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing connection faults of a lithium-ion battery pack by combining a model with deep learning, characterized by, Includes the following steps: Step S1, Data Acquisition and Model Training Stage: Collect and process the voltage and current data of each battery cell in the lithium-ion battery pack. The processed data is used to train the CAE-LSTM model to regress the connection fault magnitude of the battery cell. Step S2, First-level online fault diagnosis stage: Based on the equivalent circuit model of a battery cell, according to the load current change amplitude, a reduced-dimensional equivalent circuit model of the battery cell is constructed to estimate the internal resistance of the battery cell. The estimated internal resistance of the battery cell is compared with the average internal resistance of the battery cell to determine the connection fault of the battery cell and realize the first-level fault diagnosis. Step S3, Second-layer online fault diagnosis stage: The voltage and current data before and after the current jump are fused with the online voltage and current data. The fused and normalized data is input into the trained CAE-LSTM model to regress the fault amplitude of the current battery cell. The fault amplitude is compared with a preset threshold to determine whether a connection fault has occurred, thus realizing the second-layer fault diagnosis. 2.The model and deep learning integrated lithium-ion battery pack connection fault diagnosis method according to claim 1, characterized in that, In step S1, the data acquisition and model training phase, the specific process is as follows: Step S11: Based on the lithium-ion battery pack fault diagnosis experimental platform, inject faults into individual battery cells according to the preset fault range, and collect the current and voltage of each corresponding individual battery cell. Step S12: Divide the collected current and voltage data to form a dataset for training and validating the CAE-LSTM model; Step S13: Normalize the battery operation data using a sliding window, and train a CAE-LSTM model based on the processed data to regress the connection fault magnitude of the battery cell. Within the window, the voltage of each battery cell at each moment is normalized by subtracting the average voltage of each battery cell from the voltage of each battery cell at each moment. The normalized voltage and current are used as inputs to the CAE-LSTM model for battery pack connection fault diagnosis. 3.The model and deep learning integrated lithium-ion battery pack connection fault diagnosis method according to claim 2, characterized in that, The input data for the CAE-LSTM model is shown below: ； ； ； wherein, and are the normalized voltage and current of each battery cell at time ; and are the normalized voltage and current of each battery cell at time ; and are the normalized voltage and current of each battery cell at time ; is the width of the window; is the input data of the CAE-LSTM model; is the index of the battery cell; is the total number of battery cells. 4.The model and deep learning integrated lithium-ion battery pack connection fault diagnosis method according to claim 3, characterized in that, In the CAE-LSTM model, the convolutional autoencoder (CAE) consists of an encoder and a decoder; the encoder compresses information by encoding the input data into a low-dimensional representation of the latent vectors in the hidden layer. The encoder consists of two-dimensional convolutional layers, activation layers, and convergence layers; the two-dimensional convolutional layer is represented as: ； wherein, is a multi-channel output of a two-dimensional convolutional layer; is input data; and is a weight matrix and bias on a channel ; denotes a transpose convolution operation; is a ReLU activation function; is a channel index; is a total number of channels; The decoder consists of deconvolutional layers and activation layers; the deconvolutional layer is represented as: ； in, This represents the output feature map of the transposed convolutional layer. Indicates the th from the previous convolutional layer One input feature map; Indicates corresponding to the first Transposed convolution kernels for each channel; This indicates the bias term.

5. The lithium-ion battery pack connection fault diagnosis method based on the fusion of model and deep learning according to claim 4, characterized in that, The input sequence is processed by CAE to obtain a denoised time series, which is then used as the input to a Long Short-Term Memory (LSTM) stacked after CAE. The LSTM integrates the fault information at each time step and regresses the magnitude of the wiring fault. An LSTM unit consists of an input layer, a storage layer, and an output layer; the equations for an LSTM unit are defined as follows: ； in, This represents the input at the current moment; Represents the short-term hidden state at the current moment (output); This represents the current long-term storage state; long-term storage state and short-term hidden state Together they form the storage cells in LSTM, storing the state for a long time. Combine the current input with historical information from all previous time points. Used as the short-term hidden state and output of the LSTM unit; This indicates the long-term storage state at the previous moment; Represents the short-term hidden state at the previous moment (output); Indicates the candidate cell state; The sigmoid activation function is represented by tanh; the hyperbolic tangent function is represented by tanh. Long-term storage status and short-term hidden state Controlled by three gating structures: forget gate Input gate and output gate ; , , and These represent the inputs corresponding to the current time. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and These represent the short-term hidden states corresponding to the previous time step. The weight matrix of the forget gate, input gate, output gate, and candidate cell state; , , and The bias vectors representing the states of the forget gate, input gate, output gate, and candidate cells; One-dimensional convolutional layers and fully connected layers are used to regress the magnitude of connection faults.

6. The lithium-ion battery pack connection fault diagnosis method based on the fusion of model and deep learning according to claim 5, characterized in that, In step S2, the first-level online fault diagnosis stage, the specific process is as follows: Step S21: Based on the equivalent circuit model of a battery cell, construct a dimension-reduced equivalent circuit model of the cell according to the load current change amplitude between the current sampling time and the previous sampling time, record the voltage and current data before and after the current jump, and estimate the internal resistance of the battery cell. Step S22: Compare the estimated internal resistance of the battery cell with the average internal resistance of the battery cell. When the estimated internal resistance exceeds the preset threshold, it is determined that the battery cell has a connection fault, realizing the first layer of fault diagnosis and triggering the second layer of fault diagnosis.

7. The lithium-ion battery pack connection fault diagnosis method based on the fusion of model and deep learning according to claim 6, characterized in that, According to Kirchhoff's laws, the equivalent circuit model of a single battery cell is shown below: ； in, This is the open-circuit voltage; For ohmic internal resistance, Equivalent fault resistor; , and These are polarization resistance, polarization capacitance, and polarization voltage, respectively. This is the terminal voltage; This is the current flowing through the battery; For time; Battery voltage consists of three parts, including the state of charge. The open-circuit voltage affected by the ohmic internal resistance and the battery voltage drop caused by the equivalent fault resistance, as well as the polarization voltage generated by the polarization resistance and polarization capacitor; connection faults are identified by the abnormal internal resistance of the battery.

8. The lithium-ion battery pack connection fault diagnosis method based on the fusion of model and deep learning according to claim 7, characterized in that, The specific process for establishing a dimension-reduced equivalent circuit model for a single unit is as follows: When the change in load current between the current moment and the previous moment exceeds a given threshold, the ratio of the voltage difference to the current difference between the two moments is used as the estimated battery internal resistance, and the voltage and current before and after the current jump are recorded. When the change in load current at the current moment is less than a given threshold, the voltage value of the dimensionality-reduced model is obtained by removing the average dynamic response from the battery cell voltage, and the ratio of the voltage difference to the current difference in the dimensionality-reduced model before and after the current jump is used as the estimated battery internal resistance. Let the change in load current The equivalent circuit model of a single unit with reduced dimensions is as follows: ； ； ； ； ； in, and These represent the time intervals of the i-th battery cell. The current and voltage, ; and These represent the current and voltage of a single battery cell before the current mutation occurred; and These represent the current and voltage of a single battery cell after the mutation occurs; The threshold value for current change; This represents the estimated internal resistance of each individual battery cell; The formula for calculating the average internal resistance of a single cell is as follows: ； when Greater than the given fault diagnosis internal resistance threshold At that time, the first-level fault diagnosis determines that a connection fault has occurred in battery cell i, and triggers the second-level fault diagnosis to achieve accurate fault amplitude estimation.

9. The lithium-ion battery pack connection fault diagnosis method based on the fusion of model and deep learning according to claim 8, characterized in that, In step S3, the second-level online fault diagnosis stage, the specific process is as follows: Step S31: After a fault is detected in the first-level online fault diagnosis stage, the corresponding voltage and current data before and after the current jump collected and recorded in the first-level online fault diagnosis stage are fused with the online operating voltage and current data obtained through the sliding window mechanism to construct a time series dataset containing current jump characteristics. Step S32: Normalize the fused multidimensional voltage and current data to eliminate the impact of dimensional differences and amplitude fluctuations on the model regression accuracy. Step S33: Input the normalized data into the pre-trained CAE-LSTM model to calculate the fault magnitude of the battery cell at the current time. Perform regression estimation; Step S34: Estimate the fault amplitude output by the CAE-LSTM model. With fault diagnosis internal resistance threshold Comparison, when exceeding When a fault is detected in unit i, output is generated. This enables the diagnosis and identification of faults in the second-layer battery cells.

Images