A neural network-based forklift battery SOC prediction method and device

Abstract

The application relates to a forklift battery SOC prediction method based on a neural network, which comprises the following steps: acquiring a voltage signal at a forklift battery end and forklift working condition state data; obtaining a multi-dimensional feature vector based on the voltage signal and the forklift working condition state data, wherein the multi-dimensional feature vector comprises an original voltage value, a voltage change rate, a working condition state code and a historical reference voltage; performing inference on the multi-dimensional feature vector based on a lightweight neural network model to obtain an initial SOC prediction value; and correcting the initial SOC prediction value based on a static open-circuit voltage label and a charging event label to obtain a target SOC value. The application can better combine the actual working conditions of the forklift to estimate the SOC of the forklift lead-acid battery, and improve the estimation accuracy.

Description

Technical Field

[0001] This application relates to the technical field of battery power prediction, and in particular to a method and device for predicting the state of charge (SOC) of a forklift battery based on a neural network. Background Technology

[0002] In forklift lead-acid battery applications, accurately estimating the battery's remaining state of charge (SOC) is crucial for vehicle energy management and efficient operation. As the primary power source for forklifts, the accuracy of SOC estimation directly impacts forklift operation planning, charging management, and lifespan.

[0003] Existing methods for estimating the state of charge (SOC) of lead-acid batteries for forklifts often fail to adequately incorporate the actual operating conditions of the forklift when using voltage signals for estimation, and lack effective correction mechanisms, resulting in insufficient estimation accuracy and failing to meet the need for accurate SOC estimation of batteries under different operating conditions for forklifts. Summary of the Invention

[0004] This application provides a method and apparatus for predicting the state of charge (SOC) of a forklift battery based on a neural network, which aims to better estimate the SOC of the lead-acid battery of the forklift by combining the actual working conditions of the forklift, thereby improving the accuracy of the estimation.

[0005] To address the aforementioned technical problems, this application provides the following technical solutions:

[0006] A neural network-based method for predicting the state of charge (SOC) of a forklift battery, comprising:

[0007] Acquire the voltage signal from the forklift battery and the forklift's operating status data;

[0008] Based on the voltage signal and the forklift operating condition data, a multi-dimensional feature vector is obtained, which includes the original voltage value, voltage change rate, operating condition code and historical reference voltage.

[0009] The initial SOC prediction value is obtained by reasoning about the multidimensional feature vector based on a lightweight neural network model.

[0010] Based on the static open-circuit voltage tag and the charging event tag, the initial SOC prediction value is corrected to obtain the target SOC value.

[0011] Accordingly, embodiments of this application also provide a forklift battery SOC prediction device based on a neural network, comprising:

[0012] The voltage acquisition and operating condition detection module is used to acquire the voltage signal at the forklift battery terminal and the forklift operating condition data.

[0013] The feature vector acquisition module is used to obtain a multi-dimensional feature vector based on the voltage signal and the forklift operating condition data. The multi-dimensional feature vector includes the original voltage value, voltage change rate, operating condition code, and historical reference voltage.

[0014] The edge SOC prediction module is used to infer the multidimensional feature vector based on a lightweight neural network model to obtain an initial SOC prediction value.

[0015] The correction module is used to correct the initial SOC prediction value based on the static open-circuit voltage tag and the charging event tag to obtain the target SOC value.

[0016] The beneficial effects of this application are as follows:

[0017] 1. This application obtains the voltage signal at the forklift battery terminal and the forklift operating condition data, and combines the two to obtain a multi-dimensional feature vector containing the original voltage value, voltage change rate, operating condition code and historical reference voltage. This can more comprehensively reflect the battery state under different operating conditions, improve the quality of the basic data for SOC prediction, and thus help improve the prediction accuracy.

[0018] 2. The initial SOC prediction value is obtained by reasoning from the multi-dimensional feature vector based on the lightweight neural network model. The lightweight feature makes it adaptable to the limited computing resources of the forklift terminal, reducing the requirements for the terminal hardware performance. At the same time, the ability of the neural network to process complex nonlinear relationships is utilized to improve the rationality of the initial prediction.

[0019] 3. The initial SOC prediction value is corrected based on the static open-circuit voltage tag and the charging event tag. The two tags provide reliable references from static and charging scenarios respectively, which can effectively correct the deviation in the initial prediction, further improve the accuracy of the target SOC value, and reduce the error impact of a single prediction method. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a forklift battery SOC prediction method based on a neural network, provided in an embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the overall structure of a forklift battery SOC prediction device based on a neural network, provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the fields of industrial production and logistics transportation, forklifts are important handling equipment, and the battery life of forklifts directly affects operational efficiency and scheduling. Accurately grasping the remaining state of charge (SOC) of the battery is a key prerequisite for ensuring the continuous and stable operation of forklifts.

[0025] Currently, forklift battery SOC prediction relies heavily on traditional power calculation methods or simple sensor data feedback. These methods typically estimate based on a single or few parameters such as battery voltage and current. Some technologies combine basic battery models to deduce the state of charge. The overall solution is relatively simplified in structural design and is difficult to adapt to complex actual working conditions.

[0026] However, existing technologies have significant limitations: on the one hand, forklifts frequently switch between operating conditions during operation, such as heavy-load lifting, high-speed driving, and idling, resulting in significant differences in battery discharge characteristics. Traditional methods struggle to dynamically adapt to these changes in operating conditions, leading to large deviations in SOC prediction. On the other hand, battery voltage signals are susceptible to interference from factors such as temperature, aging, and charging / discharging history. Simple parameter combinations or basic model derivations cannot effectively eliminate the errors caused by these interferences. Especially in low-battery or high-load scenarios, prediction accuracy drops significantly, making it difficult to meet the accurate SOC requirements of refined forklift management.

[0027] Therefore, in order to solve the above problems, refer to Figure 1 This application provides a forklift battery SOC prediction method based on neural networks, including:

[0028] 101. Obtain the voltage signal from the forklift battery and the forklift operating status data.

[0029] This includes: real-time voltage analog signals collected by battery sensors, analog-to-digital conversion processing of the real-time voltage analog signals, and generation of raw voltage values.

[0030] Based on the change of the original voltage value within a preset time window, the original voltage value is subjected to time difference processing to generate the voltage change rate and historical reference voltage, where the historical reference voltage is the initial voltage value within the time window.

[0031] Based on the operating status signals (such as stationary, driving, lifting, and charging) output by the forklift controller, the generated voltage change rate and historical reference voltage are correlated and encoded (the operating status signals are converted into digital codes) to obtain the voltage signal at the forklift battery terminal and the forklift operating status data. The voltage signal includes the original voltage value, voltage change rate, and historical reference voltage, and the operating status data is the operating status code.

[0032] Specifically, battery sensors (such as voltage sensors) continuously collect real-time analog voltage signals between the positive and negative terminals of the forklift's lead-acid battery. These signals are continuously changing electrical signals and cannot be directly processed by digital systems. Therefore, the forklift's onboard terminal uses a built-in analog-to-digital converter (ADC) to perform analog-to-digital conversion on the real-time voltage signals. Following a preset sampling frequency (e.g., sampling once every 100 milliseconds), the analog signals are converted into discrete digital quantities. Each digital quantity corresponds to a specific voltage value, generating the raw voltage value, which directly reflects the battery's voltage state at a given moment.

[0033] Then, based on the generated raw voltage values, a preset time window (e.g., the past minute) is set, and all raw voltage value data within this time window are extracted. The difference between the first and last raw voltage values ​​within this time window is calculated, and then divided by the duration of the time window (e.g., 60 seconds) to obtain the voltage change per unit time, i.e., the voltage change rate, which reflects the voltage rise and fall trend during this period. Simultaneously, the first raw voltage value within this time window is designated as the historical reference voltage, serving as the benchmark reference value for subsequent voltage changes.

[0034] Next, the forklift controller outputs real-time operating status signals reflecting the forklift's operational status (such as stationary, moving, lifting, and charging status signals transmitted via the CAN bus). Based on these signals, the generated voltage change rate and historical reference voltage are correlated and encoded: different operating statuses are converted into corresponding digital codes (e.g., stationary corresponds to 00, moving corresponds to 01, lifting corresponds to 10, and charging corresponds to 11), and this code is associated and stored with the voltage change rate, historical reference voltage, and the generated original voltage value to form a complete data record. Through this processing, the final voltage signal obtained at the forklift battery terminal includes the original voltage value, voltage change rate, and historical reference voltage, while the forklift operating status data is presented in the form of operating status codes.

[0035] For example, the analog voltage signal collected by the battery sensor at a certain moment is 12.6V. After analog-to-digital conversion, the original voltage value of 12.6V is generated. A preset time window is set to the past minute. Within this window, the first original voltage value is 12.7V, and the last original voltage value is 12.5V. The voltage change rate is calculated to be (12.5-12.7) / 60≈-0.0033V / second. Simultaneously, 12.7V is designated as the historical reference voltage. At this time, the operating status signal output by the forklift controller is the driving status signal. This signal is converted to code 01 and associated with the original voltage value of 12.6V, the voltage change rate of -0.0033V / second, and the historical reference voltage of 12.7V for storage. This completes the acquisition of the forklift battery voltage signal and operating status data.

[0036] The technical solution provided in this application has the following technical advantages:

[0037] 1. By calculating the voltage change rate and determining the historical reference voltage through a preset time window, the voltage data not only includes the real-time status, but also reflects the dynamic change trend and benchmark reference, providing richer information dimensions for subsequent feature analysis and enhancing the utilization value of the data.

[0038] 2. The operating condition signal is converted into an encoding and associated with voltage-related data, realizing the organic integration of voltage information and operating condition information. This avoids the limitation that single voltage data cannot reflect the battery characteristics under different operating conditions, enabling subsequent processing to analyze the battery status in conjunction with specific operating conditions and reducing the interference of operating condition differences on the final result.

[0039] 3. By generating and integrating the original voltage value, voltage change rate, historical reference voltage, and operating condition code in stages, a more comprehensive and systematic basic data is formed. Compared with the scattered and single data processing method in traditional schemes, it provides higher quality input for subsequent SOC prediction, which helps to improve the accuracy and stability of prediction.

[0040] 102. Based on voltage signals and forklift operating condition data, a multidimensional feature vector is obtained. The multidimensional feature vector includes the original voltage value, voltage change rate, operating condition code, and historical reference voltage.

[0041] This includes: calculating the difference between the original voltage value and the historical reference voltage in the voltage signal to generate voltage difference features.

[0042] Based on the operating condition code in the forklift operating condition data, the generated voltage difference feature and the voltage change rate in the voltage signal are correlated and mapped to generate operating condition correlation features.

[0043] Based on the multidimensional feature integration rules, the generated working condition correlation features are aggregated with the original voltage value, historical reference voltage, and working condition code in the forklift working condition data to obtain a multidimensional feature vector.

[0044] Specifically, the process extracts the original voltage value and the historical reference voltage from the voltage signal. The original voltage value is the current battery voltage reading, and the historical reference voltage is the initial voltage value within a preset time window (such as the first voltage value within the past minute). Based on these two values, a voltage difference feature is generated by calculating the difference between them (original voltage value minus historical reference voltage). This feature intuitively reflects the degree of deviation of the current voltage from the historical benchmark. For example, a positive voltage difference indicates that the current voltage is higher than the historical reference value, and vice versa, providing a basis for subsequent analysis based on operating conditions.

[0045] Then, the operating condition code (e.g., 00 for stationary, 01 for traveling, 10 for lifting, and 11 for charging) is obtained from the forklift operating condition data. This code represents the current operating load state of the forklift. Based on this code, the generated voltage difference feature and the voltage change rate in the voltage signal are correlated and mapped. That is, according to the influence weight of the voltage feature under different operating conditions, the correspondence between the voltage difference feature, the voltage change rate, and the operating condition code is established. For example, under the traveling condition (code 01), the voltage change rate has a more significant impact on the power consumption, which enhances its correlation with the voltage difference feature; under the stationary condition (code 00), the original state of the voltage difference feature is emphasized more. Through this mapping, operating condition correlation features that integrate operating condition information are generated.

[0046] Multidimensional feature integration rules refer to the structured combination of features from different sources according to a preset format, ensuring that each feature dimension is clear and recognizable by the model. Based on this rule, the generated operating condition-related features are aggregated with the original voltage value, historical reference voltage, and operating condition code from the forklift operating condition data. This means arranging these four types of features in a fixed order to form a vector structure containing multidimensional information. For example, combining them in the order of "original voltage value - voltage change rate - operating condition code - historical reference voltage - operating condition-related features" ultimately yields a multidimensional feature vector containing all necessary information.

[0047] For example, suppose the original voltage value in the voltage signal is 12.5V, the historical reference voltage is 12.8V, and the voltage change rate is -0.05V / minute; the operating condition code in the forklift operating condition data is driving (01). Calculate the difference between 12.5V and 12.8V to obtain a voltage difference feature of -0.3V. Then, based on the driving code 01, associate the voltage difference feature -0.3V with the voltage change rate -0.05V / minute. Considering that the load is larger during driving and the impact of the voltage change rate is more prominent, generate an operating condition association feature (e.g., the fused feature value is -0.32). Next, according to the integration rules, aggregate the original voltage value 12.5V, the voltage change rate -0.05V / minute, the operating condition code 01, the historical reference voltage 12.8V, and the operating condition association feature -0.32 to form a multi-dimensional feature vector [12.5, -0.05, 01, 12.8, -0.32].

[0048] The technical solution provided in this application has the following technical advantages:

[0049] 1. Traditional solutions often simply concatenate raw voltage and operating condition data as input. This application, however, generates voltage difference features by calculating the difference between the raw voltage value and the historical reference voltage. This highlights the dynamic changes in voltage relative to the historical benchmark, avoiding the problem that static values ​​in the raw data fail to reflect trends, and providing more targeted feature dimensions for subsequent analysis.

[0050] 2. By mapping voltage difference features and voltage change rate based on operating condition coding, the limitation of separating voltage features and operating condition information in traditional schemes is overcome. By dynamically adjusting the correlation weight between the two according to different operating conditions (such as driving, stationary, etc.), the features can better fit the battery characteristics under specific operating conditions, reducing the interference of operating condition differences on the effectiveness of features.

[0051] 3. Based on the multi-dimensional feature integration rules, various features are aggregated. Compared with the disordered data piling in traditional solutions, this structured integration ensures the standardization and integrity of feature vectors, enabling lightweight neural network models to extract key information more efficiently, improving the model's utilization efficiency of input data, and providing higher quality feature input for subsequent inference.

[0052] Furthermore, in one embodiment, before predicting the SOC value, the model needs to be trained to enable inference. Therefore, before inferring the multi-dimensional feature vector based on the lightweight neural network model to obtain the initial SOC prediction value, the following steps are also included:

[0053] The historical operating data of the forklift lead-acid battery was preprocessed to obtain the training dataset;

[0054] Based on the training dataset, deep neural networks or temporal networks are trained to obtain cloud-based teacher models;

[0055] Based on the soft output of the cloud-based teacher model, the initial lightweight neural network model is trained by knowledge distillation to obtain a preliminary lightweight model;

[0056] Based on a preset pruning threshold, the preliminary lightweight model is pruned to obtain the pruned model.

[0057] The pruned model is quantized to obtain a trained lightweight neural network model.

[0058] Specifically, historical operating data of forklift lead-acid batteries under different working conditions is collected. This data includes battery terminal voltage signals, forklift operating status (such as stationary, driving, lifting, charging, etc.), and SOC tags generated through methods such as static open-circuit voltage and charging events. The collected data is cleaned to filter out abnormal noise data caused by sensor malfunctions or external interference. At the same time, the data is precisely aligned with the corresponding SOC tags in time to ensure that each data point is matched with an accurate tag. After this preprocessing, a dataset that can be used for model training is formed.

[0059] Then, using the training dataset obtained above, deep neural networks or temporal networks (such as LSTM, GRU, etc.) are trained in the cloud. During training, the backpropagation algorithm is used, and the error between the model's output SOC estimate and the actual label value is measured through a loss function. The network parameters are continuously adjusted to reduce the error. At the same time, methods such as cross-validation are used to evaluate the model's generalization ability on different forklift data. The model performance is optimized by adjusting hyperparameters such as network depth and width, ultimately obtaining a cloud-based teacher model with high accuracy.

[0060] Next, based on the cloud-based teacher model obtained above, it is used to infer from a large number of samples, yielding soft outputs containing SOC predictions and related uncertainty estimates. These soft outputs are then used as teacher signals to guide the training of an initial lightweight neural network model (such as a lightweight feedforward network or a shallow temporal network). During training, by minimizing the difference between the outputs of the initial lightweight model and the teacher model, the smaller model learns the knowledge contained in the teacher model, thereby significantly reducing the number of parameters while maintaining high accuracy, resulting in a preliminary lightweight model.

[0061] For the obtained preliminary lightweight model, based on a preset pruning threshold, a pruning algorithm is used to remove redundant connections and neurons that contribute little to the SOC estimation. This can be achieved by resetting weights below the threshold to zero based on their absolute values, or by using structured pruning to delete certain neurons across an entire layer. After pruning, the model is retrained (fine-tuned) to recover the accuracy lost due to pruning, resulting in the final pruned model.

[0062] Finally, the pruned model is quantized, using fixed-point quantization to convert the model's weights and inference calculations from floating-point numbers to low-bit-width representations (e.g., 8-bit integers). During quantization, quantization-aware adjustments are made, employing either symmetric or asymmetric quantization schemes, and minimum / maximum scale calibration is used to ensure the SOC output accuracy meets requirements. After quantization, the model's memory footprint is significantly reduced, and inference speed is significantly improved, ultimately resulting in a lightweight neural network model that can be efficiently run on a terminal.

[0063] Furthermore, in some embodiments, the initial lightweight neural network model originates from a cloud-based model. A high-precision complex model is first trained in the cloud, and then, through knowledge distillation, the initial lightweight neural network model is fitted to the output distribution of the complex cloud model, absorbing its inherent knowledge. This significantly reduces the number of parameters while maintaining high accuracy, thus adapting to the resource constraints of the forklift terminal.

[0064] The technical solution provided in this application has the following technical advantages:

[0065] 1. Traditional solutions struggle to adequately handle the complex nonlinear characteristics and diverse operating conditions of forklift lead-acid batteries. This application, however, preprocesses historical operating data, integrating a large number of samples covering various operating conditions and battery aging levels, providing rich data for model training. The cloud-based teacher model trained on this data can deeply mine data patterns, and knowledge distillation allows the lightweight model to learn the knowledge of complex models. Combined with pruning and quantization, key information is preserved, enabling the model to accurately estimate SOC in various scenarios, overcoming the limitation of traditional methods to single or few features.

[0066] 2. Traditional complex models struggle to run efficiently on resource-constrained edge devices (such as forklift T-Boxes). This application addresses this by obtaining a preliminary lightweight model through knowledge distillation, followed by pruning to remove redundant parameters and quantizing and compressing data representation, significantly reducing model size and computational load. This enables the model to perform rapid inference on forklift terminals, meeting real-time requirements while reducing hardware resource consumption, thus resolving the mismatch between traditional models and edge device resources.

[0067] 3. Traditional solutions suffer from either overly simple but inefficient models, or overly complex models with poor generalization capabilities. In this application, the cloud-based teacher model employs a complex structure to ensure performance, while knowledge distillation transfers its generalization capabilities to a lightweight model. Pruning and quantization compress the model while maintaining good performance through optimization. This results in an edge-end model that possesses sufficient accuracy and can adapt to diverse working conditions of different forklifts, exhibiting superior generalization capabilities compared to traditional single, fixed models.

[0068] 4. Traditional models are difficult to update and upgrade once deployed. However, this application is based on historical data for training, and new cloud-based teacher models can be continuously trained by accumulating new data. After adjustments, a new lightweight model can be obtained and deployed to the edge. This mechanism enables the system to continuously adapt to conditions such as battery aging and changes in operating conditions, achieving dynamic optimization of the model and maintaining estimation accuracy over a longer period compared to traditional static models.

[0069] 103. Based on a lightweight neural network model, reason about the multidimensional feature vectors to obtain the initial SOC prediction value.

[0070] This includes: scaling the original voltage value and historical reference voltage in the multidimensional feature vector to generate voltage reference features;

[0071] Based on voltage reference features and voltage change rate and operating condition coding in multi-dimensional feature vectors, the three are correlated and fused to generate comprehensive operating condition voltage features.

[0072] Based on a lightweight neural network model, pattern matching inference is performed on the comprehensive operating condition voltage characteristics to obtain the initial SOC prediction value.

[0073] Specifically, the original voltage value and historical reference voltage are extracted from the multi-dimensional feature vector. The original voltage value is the current instantaneous voltage reading of the battery, while the historical reference voltage is the voltage value at a certain stable point in the past (such as the voltage at the end of the last resting period). Based on these two values, a scaling process is performed: the ratio of the original voltage value to the historical reference voltage is calculated and mapped to a preset standard range (such as 0-1) to generate a voltage benchmark feature. This feature eliminates the influence of initial voltage differences between different batteries, unifies the voltage benchmark, and makes the voltage characteristics of different batteries or the same battery at different times comparable, laying the foundation for subsequent fusion of other features.

[0074] Then, the generated voltage reference feature is correlated and fused with the voltage change rate (voltage change per unit time) and operating condition code (e.g., 00 for stationary, 01 for driving, etc.) in the multi-dimensional feature vector. Specifically, the voltage change rate and operating condition code are first converted to the same standard range as the voltage reference feature, and then the three are combined into a comprehensive value using a preset fusion rule (e.g., weighted summation based on 40% voltage reference feature, 30% voltage change rate, and 30% operating condition code) to generate a comprehensive operating condition voltage feature. This feature integrates voltage reference, dynamic changes, and operating condition information, and can more comprehensively reflect the battery's charge-related characteristics in the current state.

[0075] The lightweight neural network model pre-stores a large number of patterns corresponding to comprehensive operating condition voltage features and SOC values ​​(obtained through training with historical data). Based on this model, pattern matching inference is performed on the generated comprehensive operating condition voltage features. That is, the model searches for the historical pattern most similar to the current comprehensive operating condition voltage features, uses the SOC value corresponding to the historical pattern as a reference, and combines it with the fine-tuning rules learned by the model (such as making small adjustments based on the degree of feature difference) to finally obtain the initial SOC prediction value. This process utilizes the model's memory and matching ability for feature patterns to quickly achieve the mapping from comprehensive features to SOC values.

[0076] For example, assuming the original voltage value in the multidimensional feature vector is 12.4V and the historical reference voltage is 12.7V, the ratio of the two is 0.976. After mapping to the 0-1 interval, a voltage baseline feature of 0.98 is generated. The voltage change rate is -0.02V / minute, which is 0.3 after mapping; the operating condition code is 01 (driving), which is 0.5 after mapping. Calculated according to the fusion rule: 0.98×40%+0.3×30%+0.5×30%=0.392+0.09+0.15=0.632, generating a comprehensive operating condition voltage feature of 0.632. This feature is input into a lightweight neural network model. The model matches the most similar historical pattern, corresponding to a SOC value of 68%, and then fine-tuning is performed.

[0077] The overall operating condition voltage characteristic is 0.632, while the most similar historical pattern in the model corresponds to an overall operating condition voltage characteristic of 0.63, with a corresponding SOC value of 68%. The difference between the two is 0.002, which is relatively small. According to the preset fine-tuning rules, when the difference is in the range of 0.001-0.003, the adjustment range for the SOC value is ±1%. Since the current overall operating condition voltage characteristic of 0.632 is slightly larger than the historical pattern's 0.63, and considering the pattern learned by the model (a slightly larger overall operating condition voltage characteristic corresponds to a slightly lower SOC value), 1% is subtracted from 68% to obtain an initial predicted SOC value of 67%.

[0078] The technical solution provided in this application has the following technical advantages:

[0079] 1. Traditional solutions often use the original voltage value directly for processing, ignoring the reference differences between different batteries or the same battery at different times. However, this application generates voltage reference characteristics by scaling the original voltage value with the historical reference voltage, which can eliminate the reference deviation caused by individual differences and time factors, making the voltage characteristics comparable under a unified standard, and providing a more reliable basis for subsequent fusion.

[0080] 2. Traditional solutions often process features such as voltage change rate and operating condition separately, making it difficult to reflect the correlation between features. This application integrates voltage reference features with voltage change rate and operating condition coding, and integrates multi-dimensional information through unified interval conversion and weighting rules, so that the generated comprehensive operating condition voltage feature can more comprehensively reflect the dynamic state of the battery under specific operating conditions, avoiding the limitations of a single feature.

[0081] 3. Traditional solutions rely on simple linear mappings or fixed rules during inference, which are difficult to cope with complex battery characteristics. This application uses a lightweight neural network model for pattern matching inference, which captures the complex relationship between features and SOC through a large number of historical patterns stored in the model, and handles subtle differences by fine-tuning the rules. While ensuring the efficient operation of the terminal, it improves the accuracy of prediction and overcomes the problem of poor adaptability of traditional models.

[0082] Furthermore, in some embodiments, inference is performed on the multi-dimensional feature vectors based on a lightweight neural network model to obtain an initial SOC prediction value, including:

[0083] Based on the difference between the original voltage value and the historical reference voltage, the voltage change rate is dynamically weighted and fused to obtain the voltage dynamic trend characteristics.

[0084] Based on the load conditions indicated by the operating condition code, the voltage dynamic trend characteristics are subjected to situational adaptation correction to obtain situational voltage characteristics.

[0085] The initial SOC prediction value is obtained by performing nonlinear mapping processing on the contextualized voltage characteristics based on a lightweight neural network model.

[0086] Specifically, the voltage change rate is dynamically weighted and fused based on the difference between the original voltage value and the historical reference voltage. The original voltage value is the current instantaneous voltage of the battery, while the historical reference voltage includes the highest open-circuit voltage since the last full charge and the most recent open-circuit voltage measured during rest. The difference between these two values ​​reflects the relative magnitude of voltage change. The voltage change rate reflects the dynamic trend of voltage over time. By combining the above differences and weighting them, a higher weight is given to the voltage change rate when the difference is large, and a lower weight is given when the difference is small. This generates a voltage dynamic trend feature that more accurately reflects the dynamic trend of battery capacity.

[0087] Then, based on the forklift load scenario indicated by the operating condition code, the generated voltage dynamic trend features are adjusted for contextual adaptation. Operating condition codes include states such as stationary, moving, lifting, and charging. The battery load varies under different states; for example, the load is higher during moving and lifting, resulting in a larger voltage drop component, while the voltage is closer to the true open-circuit voltage when stationary. To address these differences, the voltage dynamic trend features are adjusted. For instance, the trend features are appropriately amplified during moving to compensate for the load voltage drop, while the original trend is retained during stationary states, thus generating contextualized voltage features that fit the current operating condition.

[0088] Next, a lightweight neural network model is used to perform nonlinear mapping processing on the generated contextualized voltage features. The lightweight neural network model has learned the complex nonlinear relationship between contextualized voltage features and State of Charge (SOC) through training. After inputting this feature into the model, the model performs multi-level information extraction and integration through internal neuron connections and activation function operations, ultimately outputting an initial SOC prediction value reflecting the remaining battery capacity.

[0089] For example, assuming the initial voltage is 12.5V and the most recent open-circuit voltage measured at rest in the historical reference voltage is 12.8V, the difference is 12.5 - 12.8 = -0.3V. The voltage change rate is obtained by calculating the voltage change over the past minute. If the voltage drops from 12.6V to 12.5V in the past minute, the voltage change rate is (12.5 - 12.6) / 1 = -0.1V / minute. Since the absolute value of the difference is relatively large at 0.3V, the voltage change rate is assigned a weight of 0.8, resulting in a voltage dynamic trend characteristic of -0.1 × 0.8 = -0.08V / minute.

[0090] Because the operating condition is driving (with a large load), the voltage will contain a significant voltage drop component. Therefore, the voltage dynamic trend characteristics need to be amplified and corrected. Assuming a correction coefficient of 1.2, the contextualized voltage characteristic is -0.08 × 1.2 = -0.096 V / min. This contextualized voltage characteristic is input into a lightweight neural network model. The model performs calculations based on the learned nonlinear relationship, combining the voltage dynamic trend reflected by this characteristic with the operating condition information, and finally outputs an initial SOC prediction value of 65%.

[0091] Of these, 65% of the initial SOC predictions are outputs from the lightweight neural network model after performing nonlinear mapping on the contextualized voltage features. In the example, the voltage dynamic trend feature obtained after the above process is -0.08V / min, and after contextual adaptation correction, the contextualized voltage feature is -0.096V / min. This feature is input into the lightweight neural network model, which, based on the correlation between voltage dynamic trends, operating conditions, and SOC learned during training, outputs the final SOC prediction value through internal computational mechanisms (such as the connection weights between neurons and activation function processing). This result reflects the model's comprehensive analysis of the battery charge information contained in the input features, and its specific value is determined by the mapping rules learned during model training.

[0092] The technical solution provided in this application has the following technical advantages:

[0093] 1. Traditional solutions often directly use the voltage change rate or simply combine it with the original voltage value for SOC prediction, ignoring the impact of the voltage change relative to historical benchmarks on trend judgment. This application, however, dynamically weights and fuses the voltage change rate based on the difference between the original voltage value and the historical reference voltage. By adjusting the weight of the voltage change rate according to the magnitude of the difference (e.g., assigning higher weight to larger differences), it can more accurately capture the true dynamic trend of battery capacity, avoiding the misleading effect of a single voltage change rate under different benchmarks and improving the effectiveness of trend features.

[0094] 2. Traditional solutions often fail to differentiate between operating conditions when processing voltage characteristics, leading to interpretation discrepancies in the same voltage trend under different load scenarios. This application performs scenario-adaptive correction on the voltage dynamic trend characteristics based on the load scenario (such as driving, stationary, etc.) indicated by the operating condition state code. It makes targeted adjustments to the voltage characteristics under different loads (such as the greater impact of load voltage drop during driving), making the characteristics more consistent with the battery state under actual operating conditions and reducing the interference of operating condition differences on prediction.

[0095] 3. Traditional solutions often use linear or simple nonlinear models to handle voltage characteristics, making it difficult to characterize the complex relationship between voltage and SOC. This application utilizes a lightweight neural network model to perform nonlinear mapping on contextualized voltage characteristics. This model can learn complex nonlinear relationships under the influence of multiple factors through neural networks, and its lightweight nature adapts to terminal resource constraints. Compared with traditional models, it improves terminal operating efficiency while maintaining prediction accuracy, thus balancing accuracy and real-time performance.

[0096] Furthermore, in some embodiments, a lightweight neural network model is used to perform nonlinear mapping processing on the contextualized voltage characteristics to obtain an initial SOC prediction value, including:

[0097] Based on contextualized voltage features and preset feature mapping rules, the contextualized voltage features are processed by multi-dimensional dynamic weight allocation to generate dynamic weighted features.

[0098] Based on dynamic weighted features and a battery capacity decay model, nonlinear correlation adaptation processing is performed on the dynamic weighted features to generate capacity-related features.

[0099] Based on the power consumption correlation features and the SOC interval mapping matrix, the power consumption correlation features are processed by interval probability distribution transformation to obtain the initial SOC prediction value.

[0100] Specifically, the preset feature mapping rule is based on the correlation between contextualized voltage features and SOC summarized from a large amount of historical operating data of forklift lead-acid batteries, covering the influence weight of voltage trends on power capacity under different operating conditions. Based on this rule, the contextualized voltage features are processed by multi-dimensional dynamic weight allocation: First, the sub-features such as voltage change amplitude and change rate contained in the contextualized voltage features are analyzed. Then, according to the current forklift operating condition (e.g., the weight of voltage drop amplitude is emphasized when driving, and the weight of steady-state voltage value is emphasized when stationary), a weight value that is dynamically adjusted in real time is assigned to each sub-feature. Finally, each sub-feature is multiplied by its corresponding weight and then summed to generate a dynamically weighted feature that can highlight key power capacity information.

[0101] Then, the battery capacity decay model is constructed based on the chemical characteristics of lead-acid batteries, reflecting the nonlinear relationship between the capacity decay rate and voltage characteristics at different SOC stages. For example, at low SOC, the voltage is more sensitive to changes in capacity. Based on this model, nonlinear correlation adaptation processing is performed on the dynamically weighted features: the dynamically weighted features are input into the model, and through the model's built-in nonlinear transformation function (such as simulating the curve relationship between battery internal resistance and SOC), the feature values ​​are matched with the capacity decay law, correcting the linear deviation caused by battery aging and temperature effects, so that the feature values ​​form a more accurate nonlinear correlation with the actual capacity state, and finally generating capacity correlation features.

[0102] The battery capacity decay model is constructed based on the chemical characteristics of lead-acid batteries, combined with a large amount of cyclic charge-discharge experimental data, to quantify the nonlinear relationship between the capacity decay rate and voltage characteristics at different SOC stages. The specific steps are as follows:

[0103] Data Acquisition: Conduct full life cycle tests on multiple groups of lead-acid batteries for forklifts of the same type, and record the voltage changes, capacity decay, and corresponding cycle counts in different SOC ranges (such as 100%-90%, 90%-80%...10%-0%).

[0104] Feature extraction: For each SOC range, calculate the voltage sensitivity to changes in power (i.e., the voltage change corresponding to a unit change in power). For example, at low SOC (10%-20%), the voltage drops by 0.05V for every 1% decrease in power; at high SOC (90%-100%), the voltage drops by only 0.01V for every 1% decrease in power.

[0105] Model fitting: The relationship between sensitivity and SOC is fitted by a nonlinear function (such as a quadratic curve or an exponential curve), while incorporating influencing factors such as temperature and number of cycles (e.g., voltage sensitivity in the low SOC range increases by 10% for every 5°C decrease in temperature) to form dynamically adjusted model parameters.

[0106] For example, when constructing a capacity decay model for a certain model of forklift lead-acid battery, experimental data showed that: in the SOC range of 80%-100%, the voltage sensitivity was 0.01V / % (a 1% decrease in capacity corresponds to a 0.01V decrease in voltage); in the SOC range of 50%-80%, the voltage sensitivity was 0.02V / %; in the SOC range of 20%-50%, the voltage sensitivity was 0.03V / %; and in the SOC range of 0%-20%, the voltage sensitivity was 0.05V / %. The model fitted these data into curves and set the following parameters: when the battery cycle count exceeds 500, the sensitivity in each range increases by 20%; and when the temperature is below 10℃, the sensitivity in the 0%-20% range increases by an additional 15%.

[0107] The SOC interval mapping matrix is ​​established through training with big data in the cloud. The matrix stores the correspondence between electricity-related feature values ​​and the probability distribution of SOC intervals (e.g., 1%~100%), with each feature value corresponding to a probability value in a different SOC interval. Based on this matrix, the electricity-related features are processed through interval probability distribution transformation: the electricity-related feature values ​​are matched with the feature intervals in the matrix to extract the corresponding SOC interval probability distribution. Then, through a probability fusion algorithm (e.g., selecting the midpoint value of the interval with the highest probability, or weighted calculation of the expected value of the probability distribution), the probability distribution is converted into a specific SOC value, thus obtaining the initial SOC prediction value.

[0108] For example, the contextualized voltage feature includes two sub-features: voltage change amplitude (currently 0.3V) and voltage change rate (currently -0.02V / minute, the negative sign indicates voltage drop). The preset feature mapping rules stipulate that under driving conditions (current conditions), the voltage change amplitude has a weight of 60%, and the voltage change rate has a weight of 40%; the opposite is true under stationary conditions.

[0109] Calculate the dynamic weighted characteristics: voltage change amplitude contribution = 0.3V × 60% = 0.18V, voltage change rate contribution = (-0.02V / min) × 40% = -0.008V / min, dynamic weighted characteristics = 0.18V + (-0.008V / min) = 0.172V.

[0110] The dynamically weighted feature generated above is 0.172V (corresponding to approximately 60%-70% SOC). Applying the battery capacity degradation model, the voltage sensitivity in this range is 0.02V / %, and the current battery cycle count is 600 cycles (after 500 cycles, the sensitivity increases by 20%, i.e., 0.024V / %). Nonlinear transformation: Using the model's built-in function, the dynamically weighted feature 0.172V is matched with the sensitivity 0.024V / %, calculating the corresponding capacity degradation correlation value: 0.172V ÷ 0.024V / % ≈ 7.17%, meaning this voltage feature corresponds to a capacity degradation of 7.17%, generating a capacity correlation feature of 7.17.

[0111] In the SOC interval mapping matrix, the SOC interval probability distribution corresponding to the power-related feature 7.17 is: 65%-70% (probability 60%), 60%-65% (probability 30%), and 70%-75% (probability 10%).

[0112] The expected value is calculated using a weighted average: the midpoint of the 65%-70% interval is 67.5%, and the contribution value is 67.5% × 60% = 40.5%; the midpoint of the 60%-65% interval is 62.5%, and the contribution value is 62.5% × 30% = 18.75%; the midpoint of the 70%-75% interval is 72.5%, and the contribution value is 72.5% × 10% = 7.25%; the initial SOC prediction value is 40.5% + 18.75% + 7.25% = 66.5% (rounded to 67%).

[0113] The technical solution provided in this application has the following technical advantages:

[0114] 1. Traditional solutions often use fixed weights to process voltage features, which is difficult to adapt to battery characteristics under different operating conditions. However, this application performs multi-dimensional dynamic weight allocation on contextualized voltage features based on preset feature mapping rules. The weights of sub-features such as voltage change amplitude and change rate are adjusted in real time according to the current operating conditions (such as driving or standing still). This can highlight key power information, avoid feature distortion caused by fixed weights, and improve the feature's ability to represent the power status.

[0115] 2. Traditional solutions often correlate voltage characteristics with charge through linear relationships, which cannot reflect the nonlinear effects of factors such as battery aging and temperature. This application combines a battery charge decay model to perform nonlinear correlation adaptation on dynamically weighted characteristics. By simulating the nonlinear conversion function of chemical characteristics such as changes in battery internal resistance, the linear deviation is corrected, making the correlation between characteristics and actual charge state more accurate and enhancing the effectiveness of characteristics.

[0116] 3. Traditional schemes often directly map feature values ​​to a single SOC value, ignoring the uncertainty of prediction. This application performs interval probability distribution transformation based on the SOC interval mapping matrix, and converts feature values ​​into specific SOC values ​​through probability fusion algorithm. This can reflect the possibility of different SOC intervals, reduce the error risk of single value prediction, and further improve the reliability of the initial SOC prediction value.

[0117] 104. Based on the static open-circuit voltage tag and the charging event tag, the initial SOC prediction value is corrected to obtain the target SOC value.

[0118] This includes: performing linear compensation processing on the initial SOC prediction value based on the absolute difference between the static open-circuit voltage tag and the initial SOC prediction value to generate an open-circuit compensation value;

[0119] Based on the deviation between the charging amount data in the charging event tag and the open circuit compensation value, the open circuit compensation value is proportionally corrected to generate a charging correction value.

[0120] Based on the preset fusion rules, the charging correction value is subjected to range constraint processing to obtain the target SOC value.

[0121] Specifically, the system retrieves the stationary open-circuit voltage tag (i.e., the SOC reference value calculated from the open-circuit voltage after the battery has been stationary for a long time; this value is unaffected by load and closely approximates the actual battery capacity) and the initial SOC prediction value (the preliminary result output by the lightweight neural network model). The absolute difference between the two is calculated, and linear compensation is performed based on this difference. A compensation coefficient is set (e.g., the compensation amount increases by 0.8% for every 1% increase in the difference). The absolute difference is multiplied by the compensation coefficient to obtain the compensation amount, which is then added to the initial SOC prediction value (if the initial value is lower than the stationary tag, the compensation amount is positive; if the initial value is higher than the stationary tag, the compensation amount is negative) to generate the open-circuit compensation value. This linear relationship directly corrects the deviation between the initial prediction and the stationary reference value, leveraging the high reliability of the stationary tag to lay the foundation for correction.

[0122] Then, the charging amount data (i.e., the SOC increment corresponding to the actual amount of electricity charged into the battery during a charging process, such as 50% if charging from 30% to 80%) is extracted from the charging event tags. The deviation between this charging amount data and the theoretical increment of the generated open-circuit compensation value within the corresponding time period is calculated (for example, if the open-circuit compensation value shows that the SOC should increase by 45% during this time period, but the charging amount data is 50%, the deviation is 5%). Based on this deviation, the open-circuit compensation value is proportionally corrected. A correction ratio is set (e.g., 20% of the deviation), the deviation is multiplied by the correction ratio to obtain the correction amount, and then the correction amount is added to the open-circuit compensation value to generate the charging correction value. Further calibration is performed by observing the actual changes in electricity during the charging process to compensate for the lack of timeliness of static tags (e.g., after multiple charge and discharge cycles).

[0123] The core of the preset fusion rule is the physical boundary constraint of the SOC value (i.e., the battery capacity cannot be lower than 0% or higher than 100%). Based on this rule, the generated charging correction value is subjected to range constraint processing: if the charging correction value is within the range of 0%-100%, it is directly used as the target SOC value; if it exceeds the upper limit (e.g., 105%), the constraint is 100%; if it is lower than the lower limit (e.g., -3%), the constraint is 0%. This ensures that the final result conforms to the actual physical characteristics of the battery capacity and avoids unreasonable values ​​caused by calculation errors.

[0124] For example, assuming the static open-circuit voltage label is 70%, the initial predicted SOC is 65%, the absolute difference between the two is 5%, and the compensation coefficient is set to 0.8%, then the compensation amount = 5% × 0.8 = 4%, and the open-circuit compensation value = 65% + 4% = 69%. Next, the charging event label shows that the battery charged from 60% to 85% within the corresponding time period, with a charging amount of 25%; while the open-circuit compensation value shows that the SOC increased from 60% to 69% during the same time period, with a theoretical increase of 9%, resulting in a deviation of 25% - 9% = 16% from the charging amount data. With the correction ratio set to 20%, the correction amount = 16% × 20% = 3.2%, and the charging correction value = 69% + 3.2% = 72.2%. 72.2% is within the 0%-100% range and requires no constraint; the final target SOC value is 72.2%.

[0125] Among them, the static open-circuit voltage tag is obtained by converting the open-circuit voltage-SOC curve of the lead-acid battery after being static for a preset time, and the charging event tag is calculated by the charging end time, the charging ampere-hours, and the rated capacity of the lead-acid battery.

[0126] Specifically, when the system detects that the forklift is stationary and the battery current is nearly zero via the motion sensor at the forklift terminal, it starts timing and continuously monitors the battery terminal voltage. After the resting time reaches a preset duration (usually several hours to meet the requirement of stable open-circuit voltage for lead-acid batteries), the steady-state voltage at this point is recorded as the open-circuit voltage (OCV). Since there is a stable correlation between the open-circuit voltage and SOC of a lead-acid battery, the system calls a pre-calibrated OCV-SOC curve to convert the recorded open-circuit voltage value into the corresponding SOC value, which serves as the resting open-circuit voltage label.

[0127] The forklift terminal monitors battery voltage changes and charging signals in real time. When it detects that the forklift has connected to the charging pile and entered the charging state, it records the charging start time and initial voltage. During the charging process, it continuously tracks changes in charging voltage and current. When it detects that the charging voltage has entered the constant voltage stage and the current is approaching zero, or when it receives a charging completion signal from the charging pile, it determines that charging has ended and marks the SOC at this time as 100%. At the same time, if the charging pile provides charging ampere-hours information, the system combines the rated capacity of the lead-acid battery and calculates the proportion of the charged amount to the rated capacity to deduce the SOC value at the start of charging (i.e., 100% minus the proportion of charged amount), thus obtaining the SOC tag at the start of charging.

[0128] The technical solution provided in this application has the following technical advantages:

[0129] 1. Traditional methods lack clear quantitative compensation basis for calibrating the initial SOC prediction value. However, this application performs linear compensation based on the absolute difference between the static open-circuit voltage tag and the initial SOC prediction value. By setting a fixed compensation coefficient, the difference is directly related to the compensation amount, making the compensation process quantifiable and traceable. This avoids the arbitrariness of subjective experience adjustment in traditional calibration and improves the standardization of calibration.

[0130] 2. Traditional methods often directly overwrite the predicted value with the charging amount when using charging data for calibration, ignoring the incremental deviation analysis between the predicted value and the charging amount. This application calculates the theoretical incremental deviation between the charging amount data and the open-circuit compensation value and corrects it proportionally. This retains the basis of static calibration while combining it with the actual charging increment for fine-tuning, avoiding the dominant influence of measurement errors from a single charging data point on the results and enhancing the robustness of the calibration.

[0131] 3. Traditional methods do not impose strict physical boundary constraints on the correction results, leading to unreasonable values ​​exceeding the 0%-100% range. This application imposes range constraints on the charging correction value based on preset fusion rules, ensuring that the final target SOC value conforms to the physical characteristics of the battery capacity, avoiding invalid values ​​caused by accumulated calculation errors, and guaranteeing the rationality and practicality of the results.

[0132] Furthermore, in some embodiments, the initial SOC prediction value is corrected based on the resting open-circuit voltage tag and the charging event tag to obtain the target SOC value, including:

[0133] Based on the deviation between the static open-circuit voltage tag and the initial SOC prediction value, the initial SOC prediction value is dynamically proportionally adjusted to generate an open-circuit correction intermediate value.

[0134] Based on the temporal correlation between the full charge mark time in the charging event tag and the corresponding initial SOC prediction value, the open circuit correction intermediate value is processed by timing compensation to generate the charging correction intermediate value.

[0135] Based on the credibility weights of the static open-circuit voltage tag and the charging event tag, the intermediate values ​​of charging correction are weighted and fused to obtain the target SOC value.

[0136] Specifically, the deviation between the stationary open-circuit voltage tag and the initial SOC prediction value is first calculated. This deviation is determined by the ratio of the absolute difference between the two to the stationary open-circuit voltage tag value, reflecting the degree of deviation between the initial prediction value and the actual SOC reference value under stationary conditions. Based on this deviation, the initial SOC prediction value is dynamically adjusted proportionally: if the deviation is small (e.g., less than 5%), a smaller adjustment ratio (e.g., 10%) is used to correct the initial value; if the deviation is large (e.g., greater than 10%), a larger adjustment ratio (e.g., 30%) is used to make the adjusted value closer to the SOC level indicated by the stationary open-circuit voltage tag, ultimately generating an open-circuit correction intermediate value.

[0137] Then, the full-charge marker time (i.e., the time point when SOC = 100%) is extracted from the charging event tags, and the temporal correlation between this time and the corresponding initial SOC prediction value is analyzed. For example, the time interval between the two and the forklift's operating conditions during this period (such as the cumulative duration of driving and lifting) are calculated. Based on this temporal correlation, time-series compensation processing is performed on the open-circuit correction intermediate value: if the time interval between the full-charge marker time and the initial prediction value is short and the forklift's operating load is stable, it indicates that the initial prediction value is less affected by time-series fluctuations, and a small compensation amount is applied; if the interval is long or the load fluctuation is large, then according to the power decay pattern under similar time sequences in historical data, the corresponding compensation amount is applied to generate the charging correction intermediate value.

[0138] Next, confidence weights are pre-assigned to the open-circuit voltage tag and the charging event tag based on the resting time and the stability of the charging process. For example, the weight of the tag obtained from long-term resting (e.g., more than 8 hours) is set to 0.7, the weight of the charging event tag is set to 0.3, and the weight of the tag from short-term resting is appropriately reduced. Based on these weights, the intermediate charging correction value is weighted and fused: the intermediate open-circuit correction value is multiplied by the resting tag weight, and the intermediate charging correction value is multiplied by the charging tag weight. Then, the two products are added together to obtain the fused value. If the fused result exceeds the reasonable range of SOC (0%~100%), truncation is performed to finally obtain the target SOC value.

[0139] The truncation process refers to adjusting the value obtained through weighted fusion to a reasonable range (0%~100%) when it exceeds the acceptable range of SOC. For example, if the fusion result is 105%, it is truncated to 100%; if the fusion result is -3%, it is truncated to 0%, to ensure that the final target SOC value conforms to the actual physical range of battery capacity.

[0140] For example, assuming the static open-circuit voltage label (true SOC reference value) is 80%, and the initial predicted SOC value is 75%. Calculate the deviation: the absolute difference is |75%-80%|=5%, the deviation = 5%÷80%=6.25% (between 5% and 10%). Dynamic adjustment: according to the rules, a deviation of 6.25% corresponds to an adjustment ratio of 20%, the adjustment amount = (80%-75%)×20%=1%, therefore the median open-circuit correction value = 75%+1%=76%.

[0141] The full charge marker in the charging event tags is extracted as "10:00" (SOC=100%), and the initial SOC prediction corresponds to "14:00", with a time interval of 4 hours. During this period, the forklift was driven for a total of 2 hours and lifted for a total of 1 hour (with significant load fluctuations). Historical data shows that under similar timeframes, the battery capacity decreases by approximately 3% per hour, with a total decrease of approximately 12% (4 hours × 3%). Compensation processing: Due to the long interval and large load fluctuations, a compensation of 12% is applied, resulting in a charging correction median of 76% + 12% = 88%.

[0142] The known open-circuit voltage tag originates from a 10-hour period of inactivity (long-term inactivity), with a confidence weight of 0.7; the charging event tag indicates a stable charging process, with a weight of 0.3. Weighted fusion: Open-circuit correction median (76%) × Inactivity weight (0.7) = 53.2%; Charging correction median (88%) × Charging weight (0.3) = 26.4%; Fusion result = 53.2% + 26.4% = 79.6% (within the 0%-100% range, no truncation required). The final target SOC value is 79.6%.

[0143] The technical solution provided in this application has the following technical advantages:

[0144] 1. Traditional methods adjust the initial SOC prediction value using a fixed percentage, ignoring the impact of deviation on the adjustment range. This application dynamically determines the adjustment percentage based on the deviation, making minor adjustments for small deviations and larger adjustments for large deviations, avoiding over-adjustment or under-adjustment, making the calibration more closely match the actual state, and improving the accuracy of open-circuit calibration.

[0145] 2. Traditional solutions rarely consider the impact of time on power. This application performs time-series compensation based on the time correlation between the full-charge mark and the initial predicted value. It applies compensation based on the power decay law of the forklift's operating conditions, which can correct the prediction deviation caused by the passage of time and load changes, and reduce the interference of time-series fluctuations on the results.

[0146] 3. Traditional solutions assign equal weights to calibration tags in different scenarios. This application allocates credibility weights based on resting time and charging stability (e.g., tags with longer resting time have higher weights), and integrates the calibration results of the two types of tags through weighted fusion, avoiding the dominant influence of single tag errors and further improving the reliability of the target SOC value.

[0147] Furthermore, in some embodiments, after correcting the initial SOC prediction value based on the resting open-circuit voltage tag and the charging event tag to obtain the target SOC value, the method further includes:

[0148] The historical operating data of the forklift lead-acid battery and the generated SOC tags are processed to obtain the training set and the validation set;

[0149] Based on the training set and validation set, deep neural networks or temporal networks are trained to obtain new cloud models;

[0150] Based on pre-set test data, the estimation error of the new cloud-based high-performance model and the currently deployed model are compared and evaluated to obtain the model evaluation results;

[0151] Based on the model evaluation results, if the new cloud-based high-performance model is superior, it will be identified as the model version to be released, and a model to be released will be generated.

[0152] The model to be released is published to the forklift terminal and the old model is replaced to achieve the update of the lightweight neural network model.

[0153] Specifically, when processing the historical operating data and generated SOC tags of forklift lead-acid batteries, it is first necessary to integrate multi-source data, including historical operating data such as voltage signals, current changes, and running time of different forklifts under various working conditions (such as driving, lifting, stationary, and charging), as well as SOC tags (i.e., the actual battery reference value at the corresponding moment) generated through methods such as stationary open-circuit voltage calibration and full-charge marking. Next, data cleaning is performed to remove outliers (such as voltage jumps caused by sensor malfunctions), fill in missing values ​​(such as filling short-term missing values ​​with the average of adjacent moments), and align the data with the SOC tags in chronological order. Then, stratified sampling is used to divide the training and validation sets, such as a 7:3 ratio, to ensure consistent sample distribution across different working conditions and SOC intervals in both sets (e.g., the training set contains 60% driving condition samples, and the validation set maintains the same proportion), avoiding poor model generalization ability due to uneven sample distribution. The final training set is used for model parameter learning, and the validation set is used for performance monitoring during the training process.

[0154] Then, based on the obtained training and validation sets, a deep neural network (such as LSTM) or a time-series network (such as TCN) suitable for processing time-series data is selected for training. During training, historical operating data in the training set (such as voltage change rate and operating condition codes over the past 5 minutes) is used as input, and the corresponding SOC label is used as output. The network parameters (such as neuron weights and bias values) are continuously adjusted through the backpropagation algorithm to gradually reduce the error between the model's output SOC prediction value and the label. At the same time, after a certain number of training rounds (such as 10 rounds), the model performance is evaluated using the validation set (such as calculating the mean absolute error). If the validation set error continues to rise, training is stopped (to avoid overfitting). Finally, the optimal model parameters are saved to obtain a new cloud model. This model can deeply mine the power change patterns in historical data and capture the battery characteristics under complex forklift operating conditions.

[0155] The pre-set test data consists of historical data independent of the training and validation sets, covering a period of time, such as typical operating scenarios of different forklifts within the past three months (e.g., frequent start-stop, long periods of inactivity, alternating fast and slow charging), and accurately labeled with SOC tags. Based on this test data, the input features (e.g., voltage signals, operating condition data) of the new cloud model and the currently deployed model are input into the models respectively to obtain the SOC estimates of both, and then the estimation error (the absolute difference between the predicted value and the label) for each sample is calculated. The two sets of errors are compared and evaluated by statistically analyzing the mean, maximum value, and 90th percentile of the errors: if the new model has a lower average error and a more concentrated error distribution (e.g., the average error decreases from 3.5% to 2.8%, and the maximum error decreases from 10% to 7%), the model evaluation result is "the new model is better"; otherwise, "the current model is better".

[0156] Then, based on the model evaluation results above, clear judgment criteria are set. If the average estimation error of the new cloud model on the test data is at least 15% lower than that of the currently deployed model, and the maximum error is reduced by more than 10%, it is judged as "superior". When this condition is met, the new model is marked as a release version, and the model's parameter configuration (such as the number of network layers, input feature dimensions), training logs (such as the best validation set error), and adapted forklift battery models and other metadata are recorded, generating a release model package containing the model file and metadata. If the new model does not meet the standard, the release is abandoned, and training continues to be optimized based on more historical data to avoid a decrease in terminal prediction accuracy due to insufficient model updates.

[0157] The model package to be published is transmitted to the T-Box device of the forklift terminal via a vehicle-to-everything (V2X) communication protocol (such as MQTT). During transmission, a breakpoint resumption mechanism is used to ensure file integrity (e.g., resuming transmission from the breakpoint when reconnecting after an interruption). After receiving the model package, the terminal first verifies the file integrity (e.g., checking hash values) and compatibility (e.g., confirming that the model input / output format matches the terminal hardware). After successful verification, the old model file is automatically replaced during non-working hours of the forklift (e.g., when the forklift is shut down at night), and information such as the replacement time and the old model version is recorded. After the replacement is completed, the terminal starts the new model for the first inference test (e.g., predicting SOC using current voltage data). If the test is normal, the update is completed; if an anomaly occurs, it automatically rolls back to the old model to ensure the continuous and stable operation of the terminal.

[0158] For example, suppose a forklift fleet's historical operating data contains records of 1,000 forklifts over 6 months, and the SOC tag is generated using the static open-circuit voltage and the fully charged marking.

[0159] After cleaning, 500,000 valid samples were obtained and divided into a training set (350,000 samples) and a validation set (150,000 samples) in a 7:3 ratio. Both sets included samples from driving (40%), lifting (25%), stationary (20%), and charging (15%) operating conditions. An LSTM network was selected, and the voltage change rate and operating condition codes of the past 10 minutes were used as inputs. After 100 training rounds, the average error of the validation set stabilized at 2.9%, and the model at this point was saved as the new cloud model. The test data contained 50,000 samples. The average error of the currently deployed model was 3.6%, while the average error of the new model was 2.7%, and the maximum error decreased from 9.5% to 6.8%. The evaluation result was "the new model is better". Because the average error of the new model decreased by 25% (>15%) and the maximum error decreased by 28% (>10%), it was marked as a release version, and a package containing the LSTM model file and a list of compatible battery models was generated. The model package was sent to each forklift T-Box via MQTT, and the old model was automatically replaced at night. After the replacement, the terminal test showed that the SOC prediction was normal, and the update was completed.

[0160] The technical solution provided in this application has the following technical advantages:

[0161] 1. Traditional solutions often use raw historical data directly to train models. However, this application first aggregates and preprocesses the historical operating data of forklift lead-acid batteries and the generated SOC tags to generate training and validation sets. This process can integrate multi-source data, remove noise and redundant information, improve data quality, provide a more reliable foundation for model training, and help improve the training effect of the model.

[0162] 2. Traditional solutions use only a single type of network or lack a systematic training and optimization process. This application trains and optimizes deep neural networks or time-series networks based on training and validation sets to generate new cloud models. Deep neural networks are good at capturing complex features, while time-series networks are suitable for processing time series data. By combining the advantages of both and through optimization, the model can learn the battery operation rules more accurately and improve the model accuracy.

[0163] 3. Traditional solutions lack a rigorous evaluation process when updating models. This application compares the estimation error between the new cloud model and the currently deployed model using test data and generates evaluation results. This ensures that only the better new model is selected as the release version, avoiding the risk of updates due to poor model performance and improving the reliability of model updates.

[0164] 4. Traditional model updates rely on manual operation or local upgrades, which are inefficient and costly. This application uses the vehicle network to distribute the model to be released to the forklift terminal and replace the old model, thereby realizing the automatic update of the lightweight neural network model. This reduces manual intervention, improves the efficiency of model updates, and can quickly apply the cloud optimization results to the terminal, ensuring that the terminal model is always in a better state.

[0165] Furthermore, refer to Figure 2 This application also provides a forklift battery SOC prediction device based on a neural network, characterized in that it includes:

[0166] Voltage acquisition and operating condition detection module 1 is used to acquire the voltage signal at the forklift battery terminal and the forklift operating condition data;

[0167] Feature vector acquisition module 2 is used to obtain multi-dimensional feature vectors based on voltage signals and forklift operating condition data. The multi-dimensional feature vectors include the original voltage value, voltage change rate, operating condition code and historical reference voltage.

[0168] Edge SOC prediction module 3 is used to infer multi-dimensional feature vectors based on a lightweight neural network model to obtain initial SOC prediction values;

[0169] The correction module 4 is used to correct the initial SOC prediction value based on the static open-circuit voltage tag and the charging event tag to obtain the target SOC value.

[0170] The voltage acquisition and operating condition detection module 1 acquires the voltage signal at the forklift battery terminal and the forklift operating condition status data; the feature vector acquisition module 2 obtains a multi-dimensional feature vector based on the voltage signal and forklift operating condition status data acquired by the voltage acquisition and operating condition detection module 1. The multi-dimensional feature vector includes the original voltage value, voltage change rate, operating condition code, and historical reference voltage; the edge SOC prediction module 3 infers the multi-dimensional feature vector obtained by the feature vector acquisition module 2 based on a lightweight neural network model to obtain an initial SOC prediction value; the correction module 4 corrects the initial SOC prediction value obtained by the edge SOC prediction module 3 based on the static open circuit voltage tag and the charging event tag to obtain the target SOC value.

[0171] Furthermore, in some embodiments, a forklift battery SOC prediction device based on a neural network further includes:

[0172] Model training module 5 is used to process the historical operating data of the forklift lead-acid battery and the generated SOC tags to obtain training and validation sets; based on the training and validation sets, deep neural networks or time-series networks are trained to obtain new cloud models.

[0173] Model evaluation decision module 6 is used to compare and evaluate the estimation error of the new cloud-based high-performance model and the currently deployed model based on preset test data, and obtain the model evaluation result; based on the model evaluation result, if the new cloud-based high-performance model is better, it is determined as the model version to be released, and the model to be released is generated.

[0174] Model distribution and update module 7 is used to publish the model to be published to the forklift terminal and replace the old model, so as to realize the update of the lightweight neural network model.

[0175] After receiving the data uploaded by the forklift in the cloud, the model training module 5 processes the historical operating data of the forklift's lead-acid battery and the generated SOC tags to obtain a training set and a validation set. Based on the training set and validation set, a deep neural network or a temporal network is trained to obtain a new cloud model. The model evaluation and decision module 6 compares and evaluates the estimation error between the new cloud model obtained by the model training module 5 and the currently deployed model based on preset test data to obtain a model evaluation result. Based on the model evaluation result, if the new cloud model is better, it is determined as the model version to be released, and a model to be released is generated. The model distribution and update module 7 publishes the model to be released generated by the model evaluation and decision module 6 to the forklift terminal (i.e., the edge terminal) and replaces the old model to realize the update of the lightweight neural network model.

[0176] It should be understood that this application is not limited to the processes and structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting the state of charge (SOC) of a forklift battery based on a neural network, characterized in that, The forklift battery is a lead-acid battery for forklifts, and the method includes: Acquire the voltage signal from the forklift battery and the forklift's operating status data; Based on the voltage signal and the forklift operating condition data, a multi-dimensional feature vector is obtained, which includes the original voltage value, voltage change rate, operating condition code and historical reference voltage. The initial SOC prediction value is obtained by reasoning about the multidimensional feature vector based on a lightweight neural network model. Based on the static open-circuit voltage tag and the charging event tag, the initial SOC prediction value is corrected to obtain the target SOC value; Specifically, the initial SOC prediction value is obtained by reasoning about the multidimensional feature vector based on a lightweight neural network model, including: dynamically weighting and fusing the voltage change rate based on the difference between the original voltage value and the historical reference voltage to obtain voltage dynamic trend features; performing context adaptation correction on the voltage dynamic trend features based on the load situation indicated by the operating condition code to obtain contextualized voltage features; and performing nonlinear mapping processing on the contextualized voltage features based on the lightweight neural network model to obtain the initial SOC prediction value. The method of performing nonlinear mapping processing on contextualized voltage features based on a lightweight neural network model to obtain an initial SOC prediction value includes: performing multi-dimensional dynamic weight allocation processing on contextualized voltage features based on contextualized voltage features and preset feature mapping rules to generate dynamically weighted features; performing nonlinear correlation adaptation processing on the dynamically weighted features based on the dynamically weighted features and the battery capacity decay model to generate capacity-related features; and performing interval probability distribution transformation processing on the capacity-related features based on the capacity-related features and the SOC interval mapping matrix to obtain the initial SOC prediction value.

2. The forklift battery SOC prediction method according to claim 1, characterized in that, Before inferring the multidimensional feature vector based on the lightweight neural network model to obtain the initial SOC prediction value, the method further includes: The historical operating data of the forklift lead-acid battery was preprocessed to obtain the training dataset; Based on the training dataset, a deep neural network or a temporal network is trained to obtain a cloud-based teacher model; Based on the soft output of the cloud-based teacher model, the initial lightweight neural network model is trained by knowledge distillation to obtain a preliminary lightweight model. Based on a preset pruning threshold, the preliminary lightweight model is subjected to network pruning to obtain the pruned model. The pruned model is then quantized to obtain a trained lightweight neural network model.

3. The forklift battery SOC prediction method according to claim 1, characterized in that, The static open-circuit voltage tag is obtained by converting the open-circuit voltage-SOC curve of the lead-acid battery after it has been static for a preset time. The charging event tag is calculated by the charging end time, the charging ampere-hours, and the rated capacity of the lead-acid battery.

4. The forklift battery SOC prediction method according to claim 1, characterized in that, After correcting the initial SOC prediction value based on the static open-circuit voltage tag and the charging event tag to obtain the target SOC value, the method further includes: The historical operating data of the forklift lead-acid battery and the generated SOC tags are processed to obtain the training set and the validation set; Based on the training set and the validation set, a deep neural network or a temporal network is trained to obtain a new cloud model; Based on preset test data, the estimation error of the new cloud-based high-performance model is compared and evaluated with that of the currently deployed model to obtain the model evaluation result; Based on the model evaluation results, if the new cloud-based high-performance model is superior, it will be identified as the model version to be released, and a model to be released will be generated. The model to be published is published to the forklift terminal and the old model is replaced to update the lightweight neural network model.

5. The forklift battery SOC prediction method according to claim 1, characterized in that, The step of correcting the initial SOC prediction value based on the static open-circuit voltage tag and the charging event tag to obtain the target SOC value includes: Based on the deviation between the static open-circuit voltage tag and the initial SOC prediction value, the initial SOC prediction value is dynamically proportionally adjusted to generate an open-circuit correction intermediate value. Based on the temporal correlation between the full charge mark time in the charging event tag and the corresponding initial SOC prediction value, the open circuit correction intermediate value is subjected to time-series compensation processing to generate the charging correction intermediate value. Based on the credibility weights of the static open-circuit voltage tag and the charging event tag, the intermediate value of the charging correction is weighted and fused to obtain the target SOC value.

6. A forklift battery SOC prediction device based on neural networks, characterized in that, The forklift battery is a lead-acid battery for forklifts, and the device includes: The voltage acquisition and operating condition detection module is used to acquire the voltage signal at the forklift battery terminal and the forklift operating condition data. The feature vector acquisition module is used to obtain a multi-dimensional feature vector based on the voltage signal and the forklift operating condition data. The multi-dimensional feature vector includes the original voltage value, voltage change rate, operating condition code, and historical reference voltage. The edge SOC prediction module is used to infer the multidimensional feature vector based on a lightweight neural network model to obtain an initial SOC prediction value. The correction module is used to correct the initial SOC prediction value based on the static open-circuit voltage tag and the charging event tag to obtain the target SOC value; Specifically, the initial SOC prediction value is obtained by reasoning about the multidimensional feature vector based on a lightweight neural network model, including: dynamically weighting and fusing the voltage change rate based on the difference between the original voltage value and the historical reference voltage to obtain voltage dynamic trend features; performing context adaptation correction on the voltage dynamic trend features based on the load situation indicated by the operating condition code to obtain contextualized voltage features; and performing nonlinear mapping processing on the contextualized voltage features based on the lightweight neural network model to obtain the initial SOC prediction value. The method of performing nonlinear mapping processing on contextualized voltage features based on a lightweight neural network model to obtain an initial SOC prediction value includes: performing multi-dimensional dynamic weight allocation processing on contextualized voltage features based on contextualized voltage features and preset feature mapping rules to generate dynamically weighted features; performing nonlinear correlation adaptation processing on the dynamically weighted features based on the dynamically weighted features and the battery capacity decay model to generate capacity-related features; and performing interval probability distribution transformation processing on the capacity-related features based on the capacity-related features and the SOC interval mapping matrix to obtain the initial SOC prediction value.

7. The forklift battery SOC prediction device according to claim 6, characterized in that, Also includes: The model training module is used to process the historical operating data of the forklift lead-acid battery and the generated SOC tags to obtain a training set and a validation set; based on the training set and the validation set, a deep neural network or a time series network is trained to obtain a new cloud model. The model evaluation decision module is used to compare and evaluate the estimation error of the new cloud-based high-performance model and the currently deployed model based on preset test data, and obtain the model evaluation result. Based on the model evaluation results, if the new cloud-based high-performance model is superior, it will be identified as the model version to be released, and a model to be released will be generated. The model distribution and update module is used to publish the model to be published to the forklift terminal and replace the old model, so as to realize the update of the lightweight neural network model.

Images