Forklift lithium battery voltage fault detection method and device with physical interpretability

Abstract

The invention belongs to the technical field of lithium battery fault detection, and discloses a forklift lithium battery voltage fault detection method and a forklift lithium battery voltage fault detection device with physical interpretability, which adopt a mode of combining a dynamic auto-encoder with contrast learning and a physical model, simplify the complexity of traditional pure physical modeling, avoid the problem that parameters are not easy to obtain, and improve the detection accuracy. And meanwhile, the detection robustness under a complex working condition is improved. The comparative learning enables the model to better distinguish the health states of different batteries by optimizing the separability of a low-dimensional vector space, and weakens the influence caused by individual differences to a certain extent. An empirical voltage model of the forklift lithium battery is embedded in a decoder module of the dynamic auto-encoder, so that a model prediction result not only conforms to data distribution, but also follows an actual physical mechanism, and the interpretability of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery fault detection, and particularly relates to a method and device for detecting voltage faults of forklift lithium batteries with physical interpretability. Background Art

[0002] Lithium batteries have become an ideal power source for electric forklifts due to their high energy density, long cycle life and other advantages, and are increasingly widely used in the forklift field. The operating environment of forklifts is complex and diverse, and they often need to work frequently in narrow spaces, high and low temperature environments, and dusty and humid conditions. Such special usage scenarios pose many challenges to forklift lithium batteries. Especially under realistic conditions such as high-frequency charge and discharge, complex thermal environments, and diverse user usage habits, the battery system faces a series of challenges such as poor heat dissipation performance, high risk of thermal runaway, fast life attenuation, and poor operating stability. Therefore, developing an intelligent fault prediction and anomaly identification method for forklift lithium batteries is of great significance for improving the performance of the battery system throughout its life cycle and realizing the safe and controllable operation of forklifts.

[0003] Currently, the existing forklift lithium battery fault detection technologies mainly include physical modeling-based methods and data-driven artificial intelligence methods. Physical modeling-based methods, by establishing thermal, electrical, and chemical kinetic models of the battery, deduce and estimate the internal state, and have strong physical interpretability, but face problems such as complex modeling, difficult parameter acquisition, and sensitivity to individual differences. Data-driven artificial intelligence methods, such as deep neural network methods, can mine feature information from large-scale operation data to achieve automatic prediction and judgment, and have good engineering adaptability, but often lack interpretability of model outputs and have insufficient stability when facing cross-condition migration. Summary of the Invention

[0004] Aiming at the defects in the prior art, the present invention provides a method and device for detecting voltage faults of forklift lithium batteries with physical interpretability, which improves the detection robustness, health state discrimination ability and interpretability under complex working conditions.

[0005] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0006] In the first aspect: A method for detecting voltage faults of forklift lithium batteries with physical interpretability is provided, including:

[0007] Collect historical multi-dimensional time series data during the actual operation of forklift lithium batteries, and generate multiple sample sequences after standardization processing and sliding time window processing;

[0008] Add labels to each sample sequence, and construct positive sample pairs and negative sample pairs;

[0009] Divide the sample sequence into an input sequence and an output sequence, input the input sequence and the output sequence into the encoder module, obtain the low-dimensional feature representation output by the encoder module, and calculate the contrast loss function based on the low-dimensional feature representations in the positive sample pairs and negative sample pairs;

[0010] Input the low-dimensional feature representation of the sample sequence and the input sequence into the decoder module, obtain the reconstructed output sequence output by the decoder module, and construct a reconstruction loss function based on the reconstructed output sequence;

[0011] Calculate the empirical voltage value based on the empirical voltage model, and calculate the physical consistency loss function according to the empirical voltage value and the reconstructed voltage value in the reconstructed output sequence;

[0012] Update the encoder module and the decoder module by integrating the contrast loss function, the reconstruction loss function, and the physical consistency loss function until the training ends, and use the encoder module and the decoder module after the training ends to obtain the reconstructed output sequences of all sample sequences during the training, and determine the voltage fault threshold;

[0013] Compare the voltage fault threshold with the voltage value of the forklift lithium battery collected in real time to obtain the voltage fault detection result of the forklift lithium battery.

[0014] The following also provides several optional methods, which are not additional limitations to the above overall solution, but are only further supplements or optimizations. Without technical or logical contradictions, each optional method can be combined with the above overall solution alone, or multiple optional methods can be combined with each other.

[0015] Preferably, each data point in the historical multi-dimensional time series data includes a voltage value, a current value, a temperature, and a state of charge, and each data point in the input sequence includes a current value and a state of charge, and each data point in the output sequence includes a voltage value and a temperature.

[0016] Preferably, the empirical voltage model is expressed as follows:

[0017]

[0018] Where, is the empirical voltage value corresponding to the th sample sequence at the th moment, is the state of charge of the th sample sequence at the th moment, is the current value of the th sample sequence at the th moment, is the temperature of the th sample sequence at the th moment, , , , , , and are empirical coefficients.

[0019] Preferably, calculating the physical consistency loss function based on the empirical voltage value and the reconstructed voltage value in the reconstructed output sequence includes:

[0020] Calculating the time derivative of the reconstructed voltage value to obtain the change rate of the reconstructed voltage value;

[0021] Calculating the time derivative of the empirical voltage value to obtain the change rate of the empirical voltage value;

[0022] Calculating the physical consistency loss function based on the difference between the change rate of the reconstructed voltage value and the change rate of the empirical voltage value.

[0023] Preferably, calculating the physical consistency loss function based on the difference between the change rate of the reconstructed voltage value and the change rate of the empirical voltage value includes:

[0024]

[0025] Wherein, is the physical consistency loss function, is the physical residual corresponding to the th sample sequence at the th moment, is the total number of sample sequences, is the start time of the th sample sequence, is the end time of the th sample sequence, is the length of the sliding time window, is the reconstructed voltage value corresponding to the th sample sequence at the th moment, is the empirical voltage value corresponding to the th sample sequence at the th moment, is the change rate of the reconstructed voltage value , is the change rate of the empirical voltage value , is the L2 norm.

[0026] Preferably, determining the voltage fault threshold includes:

[0027] Take the reconstructed voltage values in all the reconstructed output sequences, calculate the percentile values of all the reconstructed voltage values, and use the percentile values as the voltage fault thresholds.

[0028] Preferably, compare the voltage fault threshold with the voltage value of the forklift lithium battery collected in real time to obtain the voltage fault detection result of the forklift lithium battery, including:

[0029] If the voltage value of the forklift lithium battery collected in real time is greater than the voltage fault threshold, it is determined that the forklift lithium battery has a voltage fault; otherwise, the forklift lithium battery has no voltage fault.

[0030] Second aspect: A forklift lithium battery voltage fault detection device with physical interpretability, including a processor and a memory storing a number of computer instructions. When the computer instructions are executed by the processor, the steps of the forklift lithium battery voltage fault detection method with physical interpretability are implemented.

[0031] The forklift lithium battery voltage fault detection method and device with physical interpretability provided by the present invention have the following beneficial effects compared with the prior art:

[0032] 1. By adopting the method of combining dynamic autoencoders with contrastive learning and physical models, the complexity of traditional pure physical modeling is simplified, the problem of difficult parameter acquisition is avoided, and the detection robustness under complex working conditions is improved. Contrastive learning optimizes the separability of the low-dimensional vector space, enabling the model to better distinguish the health states of different batteries and weakening the influence brought by individual differences to a certain extent.

[0033] 2. Embed the empirical voltage model of the forklift lithium battery in the decoder module of the dynamic autoencoder, so that the model prediction results not only conform to the data distribution but also follow the actual physical mechanism, improving the interpretability of the model;

[0034] 3. Jointly optimize the reconstruction error, state discrimination loss, and physical residuals to achieve the deep integration of data-driven and knowledge-driven. Description of the Drawings

[0035] Figure 1 It is a flowchart of the forklift lithium battery voltage fault detection method with physical interpretability of the present invention. Detailed Embodiments

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention.

[0038] Example 1: As Figure 1 shown, this example proposes a method for detecting voltage faults in forklift lithium batteries with physical interpretability, including the following steps:

[0039] Step S1, data input and preprocessing. Collect multi-dimensional time series data of forklift lithium batteries during actual operation, including voltage values , current values , temperature and state of charge , and construct the input sequence , , where

[0040] (1)

[0041] Among them, is the normalized sample, , , , are the mean values of voltage, current, temperature and state of charge respectively, , , , are the corresponding standard deviations. The normalized data will be processed by a sliding time window and sliced into sample sequences on each time window: , where , is the total number of sample sequences, is the sliding time window length, is the th sample sequence.

[0042] Step S2, construction of positive and negative sample pairs. Calibrate the category of sample pairs through the health labels (normal or faulty) of the samples. Specifically, if two samples belong to the same health state, they form a positive sample pair; if two samples belong to different health states, they form a negative sample pair. The construction form of the sample pairs is as follows:

[0043] (2)

[0044] Among them, is a sample pair, is the th sample sequence, represents a positive sample pair, represents a negative sample pair.

[0045] Step S3: Generation of sample dynamic mapping relationship and construction of contrast loss function. First, the standardized samples are divided into two parts: input and output:

[0046] (3)

[0047] (4)

[0048] Among them, is the th input sequence of the sample sequence, is the th output sequence of the sample sequence, is the th current value at time in the input sequence of the sample sequence, is the th state of charge at time in the input sequence of the sample sequence, is the th voltage value at time in the output sequence of the sample sequence, is the th temperature at time in the output sequence of the sample sequence.

[0049] Secondly, the dynamic mapping relationship obtained by the input sequence and the output sequence passing through a parameterized compression network can be expressed as:

[0050] (5)

[0051] Among them, represents the dynamic characterization of the sample sequence in the latent space, that is, the low-dimensional feature representation in the low-dimensional vector space, represents the learnable parameters of the encoder module, represents the encoder module. The encoder module consists of multiple convolutional units, non-linear activation functions, and compression structures, and is used to extract and compress the feature information of the time series signal layer by layer. In this embodiment, a dynamic autoencoder is used to encode the input-output joint sequence to obtain a low-dimensional feature representation of the battery dynamic behavior, thereby transforming the high-dimensional anomaly detection problem into a structure modeling problem in the low-dimensional vector space.

[0052] This embodiment introduces a contrastive learning mechanism to optimize the distribution of low-dimensional feature representations, enabling the model to distinguish healthy states. Specifically, for any pair of samples , the features obtained after passing through the compression network represent the low-dimensional feature representations corresponding to the sample sequence . The Euclidean distance is used to measure the feature similarity. The Euclidean distance between the low-dimensional feature representation and the low-dimensional feature representation is:

[0053] (6)

[0054] where represents and 's Euclidean distance (two-norm). For positive sample pairs, the contrastive training objective is to maximize similarity, and the positive sample pair loss is:

[0055] (7)

[0056] For negative sample pairs, to prevent non-convergence during training, the Euclidean distance is restricted not to exceed 1, and the loss is:

[0057] (8)

[0058] Therefore, the final contrastive loss function is:

[0059] (9)

[0060] Step S4, reconstruct the decoder features of the forklift lithium battery voltage model. Feed the low-dimensional feature representation together with the input into the decoder module to restore the original output sequence:

[0061] (10)

[0062] where is the reconstructed output sequence output by the encoder module, including the reconstructed voltage value and the reconstructed temperature , is the decoder module, are its parameters.

[0063] Further extract the reconstructed voltage time series from the reconstruction result, . Next, calculate the time derivative of the reconstructed voltage value to obtain the change rate of the reconstructed voltage value. In this embodiment, the time derivative of the reconstructed voltage is calculated based on an automatic differentiation tool (such as the PyTorch tool) to obtain its change rate:

[0064] (11)

[0065] where is the automatic differentiation function. At the same time, aiming at the problem of insufficient physical interpretability in the existing forklift lithium battery fault detection using a dynamic autoencoder, an empirical voltage model of the forklift lithium battery is introduced to construct a known physical evolution function of the forklift lithium battery. This function is used to simulate the theoretical change rate of the voltage under the influence of variables (current value, state of charge, temperature). Its form is derived from the empirical model and is defined as:

[0066] (12)

[0067] where is the empirical voltage value corresponding to the th sample sequence at the th moment, is the state of charge of the th sample sequence at the th moment, is the current value of the th sample sequence at the th moment, is the temperature of the th sample sequence at the th moment, , , , , , and are empirical coefficients.

[0068] Furthermore, the time derivative is also calculated for the empirical voltage value calculated by the empirical voltage model. Define the physical residual as the difference between the derivative of the reconstructed voltage value and the derivative of the empirical voltage value. Based on the physical residuals at all time steps, define a physical consistency loss function, and optimize the decoder parameters through the physical consistency loss function to make the reconstruction result consistent with the physical law reflected by the empirical voltage model, thereby realizing the physical constraint of the empirical voltage model on the decoder output and used to punish the reconstruction behavior that does not conform to the physical law:

[0069] (13)

[0070] where is the physical consistency loss function, is the th sample sequence's physical residual at the th moment, is the total number of sample sequences, is the start time of the th sample sequence, is the end time of the th sample sequence, is the length of the sliding time window, is the reconstructed voltage value corresponding to the th sample sequence at the th moment, is the empirical voltage value corresponding to the th sample sequence at the th moment, is the change rate of the reconstructed voltage value , is the change rate of the empirical voltage value , is the L2 norm, used to measure the sum of squared differences.

[0071] Step S5, Model training and unified loss optimization. To optimize the reconstruction accuracy, physical rationality, and feature discrimination ability simultaneously, in this embodiment, the loss functions are unified and integrated as the objective function for model training.

[0072] First, define the reconstruction loss function to measure the difference between the reconstructed output and the true output:

[0073] (14)

[0074] where, is the reconstructed output sequence corresponding to the th sample sequence at the th moment, the th sample sequence's true output sequence at the th moment. This loss encourages the model to more accurately restore the dynamic behavior of the system.

[0075] Secondly, introduce the contrast loss function to improve the separability of the encoded features, and then combine it with the physical consistency loss function to ensure that the reconstruction result conforms to the known physical laws. Finally, the total model loss function is uniformly expressed as:

[0076] (15)

[0077] where, They are the weight factors for the three types of losses respectively, which are set according to the importance of specific tasks. All model parameters (including ) are jointly updated through backpropagation. The model training uses an optimizer based on gradient descent (such as Adam) to iteratively optimize this joint loss function on the training set until the loss converges or reaches the set number of iterations. Ensure that the model can have the ability to discriminate states while maintaining prediction accuracy and conform to the physical characteristics of the battery, and finally achieve accurate modeling and fault identification of the operating state of the forklift lithium battery.

[0078] After the model training is completed, use the trained encoder module to process the multi-dimensional time series data of the forklift lithium battery collected during training, and reconstruct the reconstructed voltage value of the sample through the decoder module. Statistically analyze the reconstructed voltage value and calculate its percentile value, and set this percentile value as the voltage fault threshold.

[0079] Step S6: In the real-time fault detection of the forklift lithium battery, compare the actually measured voltage value obtained by real-time monitoring with this voltage fault threshold. If the actually measured voltage value exceeds the voltage fault threshold, it is determined that the forklift lithium battery may have a fault; otherwise, there is no fault.

[0080] Embodiment 2: This embodiment provides a forklift lithium battery voltage fault detection device with physical interpretability, including a processor and a memory storing several computer instructions. When the computer instructions are executed by the processor, the steps of the forklift lithium battery voltage fault detection method with physical interpretability are implemented.

[0081] For the specific limitations of the forklift lithium battery voltage fault detection device with physical interpretability, reference can be made to the limitations of the forklift lithium battery voltage fault detection method with physical interpretability in the above text, and details will not be elaborated here.

[0082] The memory and the processor are directly or indirectly electrically connected to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, and the processor realizes the method of the present invention by running the computer program stored in the memory.

[0083] Among them, the memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electric Erasable Programmable Read-Only Memory (EEPROM), etc.

[0084] The processor may be an integrated circuit chip with data processing capabilities. The above-mentioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0085] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0086] The above-described embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A method for detecting voltage faults of a forklift lithium battery with physical interpretability, characterized in that, The physically interpretable forklift lithium battery voltage fault detection method includes: Collect historical multi-dimensional time series data during the actual operation of the forklift lithium battery, and generate multiple sample sequences after standardization processing and sliding time window processing; Add labels to each sample sequence and construct positive sample pairs and negative sample pairs; Divide the sample sequences into input sequences and output sequences, input the input sequences and output sequences into the encoder module, obtain the low-dimensional feature representations output by the encoder module, and calculate the contrast loss function based on the low-dimensional feature representations in the positive sample pairs and negative sample pairs; Input the low-dimensional feature representations of the sample sequences and the input sequences into the decoder module, obtain the reconstructed output sequences output by the decoder module, and construct a reconstruction loss function based on the reconstructed output sequences; Calculate the empirical voltage value based on the empirical voltage model, and calculate the physical consistency loss function according to the empirical voltage value and the reconstructed voltage value in the reconstructed output sequence; Update the encoder module and the decoder module comprehensively based on the contrast loss function, the reconstruction loss function, and the physical consistency loss function until the training ends, and use the encoder module and the decoder module after the training ends to obtain the reconstructed output sequences of all sample sequences during training, and determine the voltage fault threshold; Compare the voltage fault threshold with the voltage value of the forklift lithium battery collected in real time to obtain the forklift lithium battery voltage fault detection result.

2. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 1, characterized in that Each data point in the historical multi-dimensional time series data includes a voltage value, a current value, a temperature, and a state of charge, and each data point in the input sequence includes a current value and a state of charge, and each data point in the output sequence includes a voltage value and a temperature.

3. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 1, characterized in that, The empirical voltage model is expressed as follows: ； Among them, is the empirical voltage value corresponding to the -th sample sequence at the -th moment, is the state of charge of the -th sample sequence at the -th moment, is the current value of the -th sample sequence at the -th moment, is the temperature of the -th sample sequence at the -th moment, , , , , , and are empirical coefficients.

4. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 1, characterized in that The calculation of the physical consistency loss function according to the empirical voltage value and the reconstructed voltage value in the reconstructed output sequence includes: Perform a time derivative calculation on the reconstructed voltage value to obtain the change rate of the reconstructed voltage value; Perform a time derivative calculation on the empirical voltage value to obtain the change rate of the empirical voltage value; Calculate the physical consistency loss function according to the difference between the change rate of the reconstructed voltage value and the change rate of the empirical voltage value.

5. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 4, wherein The calculation of the physical consistency loss function according to the difference between the change rate of the reconstructed voltage value and the change rate of the empirical voltage value includes: ； Among them, is the physical consistency loss function, is the physical residual corresponding to the -th sample sequence at the -th moment, is the total number of sample sequences, is the start time of the -th sample sequence, is the end time of the -th sample sequence, is the length of the sliding time window, is the reconstructed voltage value corresponding to the -th sample sequence at the -th moment, is the empirical voltage value corresponding to the -th sample sequence at the -th moment, is the change rate of the reconstructed voltage value , is the change rate of the empirical voltage value , is the L2 norm.

6. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 1, characterized in that The determination of the voltage fault threshold includes: Take the reconstructed voltage values in all the reconstructed output sequences, calculate the percentile value of all the reconstructed voltage values, and use the percentile value as the voltage fault threshold.

7. The method for detecting voltage faults of a forklift lithium battery with physical interpretability according to claim 1, characterized in that, The comparison of the voltage fault threshold with the voltage value of the forklift lithium battery collected in real time to obtain the forklift lithium battery voltage fault detection result includes: If the voltage value of the forklift lithium battery collected in real time is greater than the voltage fault threshold, it is determined that the forklift lithium battery has a voltage fault; otherwise, the forklift lithium battery does not have a voltage fault.

8. A forklift lithium battery voltage fault detection device with physical interpretability, comprising a processor and a memory storing a number of computer instructions, characterized in that, When the computer instructions are executed by the processor, the steps of the physically interpretable forklift lithium battery voltage fault detection method described in any one of claims 1 to 7 are implemented.

Images