A method and system for sorting retired power battery mobile robots and a storage medium

Abstract

The application discloses a kind of retired power battery mobile robot sorting method, system and storage medium;The method comprises: obtaining the voltage and current time series of retired power battery and binding with battery number;The time series is preprocessed to form fixed-length input segment;Artificial statistical feature vector is constructed;Feature extraction is carried out on input segment using one-dimensional convolutional neural network, and deep dynamic response feature vector is output;Deep dynamic response feature vector and artificial statistical feature vector are fused to generate fusion feature vector;XGBoost model is used to make hierarchical decision on fusion feature vector, and output sorting grade and sorting confidence;According to the mapping target sorting area of sorting grade and generating robot sorting task;Control mobile sorting robot to execute material taking, transfer and unloading action;The application realizes the precise sorting closed loop from data analysis, intelligent grading to automation execution.

Description

Technical Field

[0001] This invention relates to the field of recycling technology for retired power batteries, specifically to a mobile robot sorting method, system, and storage medium for retired power batteries. Background Technology

[0002] With the large-scale industrial application of new energy vehicles and electrochemical energy storage systems, power batteries are entering a concentrated retirement phase. Retired power batteries come from diverse sources and are of various models, exhibiting significant inconsistencies in their historical operating conditions, aging paths, and State of Health (SOH). Rapid and accurate sorting of retired batteries before secondary use or material recycling is crucial in determining their subsequent utilization value and safety. However, SOH assessment of retired batteries faces multiple uncertainties, including missing historical data, non-uniform degradation, and differences in chemical systems. Traditional uniform rules or simple threshold methods are ill-suited to this complex retirement scenario.

[0003] Current sorting technologies mainly rely on the following methods: manual sorting based on experience is inefficient, subjective, and inconsistent; single-threshold sorting based on capacity or internal resistance is intuitive, but full-capacity testing is time-consuming, and internal resistance testing cannot fully reflect the battery's dynamic load response capability, making it difficult to accurately distinguish batteries with similar performance but different aging mechanisms; traditional machine learning sorting methods have made some improvements, but often rely only on a few manually designed statistical features (such as average voltage and estimated internal resistance). These features have limited ability to represent deep dynamic information such as local voltage drop, recovery rate, and plateau decay contained in the voltage / current curve, resulting in limited model accuracy and insufficient generalization ability on new batches and new models of batteries. More importantly, most existing algorithms only output evaluation reports or sorting grades, and still require manual handling, classification, and warehousing operations. The algorithm decision-making and physical execution are severely disconnected, failing to form an efficient closed loop from intelligent judgment to automatic execution, which restricts the construction of large-scale, automated sorting production lines for retired batteries.

[0004] Therefore, how to fully extract deep characteristics that characterize battery health and consistency from limited voltage and current detection data, and how to directly and reliably transform the grading results into automated sorting actions to achieve an integrated and precise closed loop of data analysis, sorting decision-making, and robot execution, has become a pressing technical problem in this field. Constructing cost-effectiveness models and improving data quality and scenario coverage are crucial to supporting the rapid development of autonomous driving perception technology. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a sorting method, system and storage medium for retired power batteries using a mobile robot.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This application provides a method for sorting retired power batteries using a mobile robot, including the following steps: Obtain the voltage and current time series of retired power batteries and bind them with the battery's unique serial number; The voltage and current time series are preprocessed, including standardization and sliding window segmentation, to form input segments of fixed length. Construct an artificial statistical feature vector, which contains statistical features extracted from the voltage time series and the current time series; A one-dimensional convolutional neural network is used to extract features from the input segment, and a deep dynamic response feature vector is output. The deep dynamic response feature vector is fused with the artificial statistical feature vector to generate a fused feature vector; The extreme gradient boosting tree (XGBoost) model is used to make hierarchical decisions on the fused feature vectors, and the sorting level and sorting confidence of the battery are output. The target sorting area is determined based on the sorting level and sorting confidence mapping, and a robot sorting task is generated. The robot sorting task includes at least the battery number, picking point, target delivery point and action type. The robot sorting task is sent to the mobile sorting robot to control the mobile sorting robot to perform material picking, transfer and unloading actions, and put the battery into the target sorting area.

[0007] Preferably, the standardization process and sliding window segmentation satisfy the following formula: ; ; in, and The first Voltage and current values ​​at each sampling point , These are the mean values ​​of the voltage and current sequences, respectively. , These are the standard deviations of the voltage series and the current series, respectively. To prevent small constants with a denominator of zero; For the first A sliding window, For the preset window length, This is the sliding step size.

[0008] Preferably, the convolution operation for feature extraction in the one-dimensional convolutional neural network is represented as follows: ; In the formula, For the lth layer The convolutional channel at the _th convolutional channel in the _th ... Output at each time position Input data for the previous layer, For convolution kernel weights, For bias terms, The kernel length is 1. This represents the number of channels in the previous layer. For time location, The position number inside the convolution kernel. Enter the channel number.

[0009] Preferably, the XGBoost model is used to perform hierarchical decision-making on the fused feature vector, outputting the battery sorting level and sorting confidence, satisfying the following formula: ; ; ; in, For the first The battery sample belongs to the first The probability of each sorting grade. The total number of sorting grades, The original score output by the XGBoost model. The final sorting grade is determined by... The confidence level for sorting.

[0010] Preferably, it also includes a consistency optimization grouping step: for batteries belonging to the same sorting level, construct an engineering feature vector. ,in This is a capacity estimate. This is the estimated internal resistance value. This is an estimate of health status. For short-term energy, The slope of the voltage change. The current fluctuation coefficient is used to calculate the weighted distance between batteries of the same class. ,in For the first The weights of each feature The number of features; clustering or matching is performed based on the weighted distance, such that the grouping satisfies the objective function. ,in Indicates within-group differences. Indicates the robot's execution time. Indicates the number of retests. , , These are the weighting coefficients.

[0011] Preferably, when multiple mobile sorting robots exist, the scheduling cost is calculated based on the current position, remaining battery power, and task queue length of each robot, and the optimal robot is selected to perform the robot sorting task, satisfying the following formula: ; In the formula, Number the robot to be scheduled. This is the robot's current position. Here are the coordinates of the material collection point. The robot has remaining battery power. This is the length of the robot's task queue. , , , These are preset weighting coefficients.

[0012] Preferably, the mobile sorting robot is a differential wheel mobile chassis, and its motion control satisfies the following formula: ; ; in, For the robot's linear velocity, Angular velocity, and These are the linear velocities of the right and left wheels, respectively. This refers to the wheel track.

[0013] Preferably, the method further includes a model closed-loop update step: when the sorting confidence level is lower than a preset threshold, the corresponding battery is sent to the retesting area for retesting; the retesting results are recorded, and the XGBoost model is updated using the overall loss function, wherein the overall loss function is: ; in, For graded loss, To correct the loss for the retested samples, For consistency grouping constraint loss, and These are the weighting coefficients.

[0014] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.

[0015] A mobile robot sorting system for retired power batteries includes: The data input and preprocessing module is used to acquire the voltage and current time series of retired power batteries, bind them with the battery's unique number, and perform standardization and sliding window segmentation to form input segments of fixed length. The artificial statistical feature construction module is used to extract artificial statistical feature vectors from the voltage time series and current time series. A one-dimensional convolutional neural network feature extraction module is used to extract deep dynamic response feature vectors from the input segment using a one-dimensional convolutional neural network; The feature fusion module is used to fuse the deep dynamic response feature vector with the manually statistical feature vector to generate a fused feature vector; The XGBoost hierarchical decision module is used to classify the fused feature vector using the extreme gradient boosting tree XGBoost model, and output the battery sorting level and sorting confidence. The target area mapping and task generation module is used to determine the target sorting area according to the sorting level and sorting confidence mapping, and generate a robot sorting task that includes at least the battery number, picking point, target delivery point and action type. A mobile sorting robot is used to receive the sorting task from the robot and autonomously perform material picking, transfer and unloading actions to place the batteries into the target sorting area.

[0016] Compared with the prior art, this application has the following beneficial effects: This invention proposes a mobile robot sorting method, system, and storage medium for retired power batteries. By fusing deep dynamic response features automatically extracted by 1D-CNN with manually generated statistical features with clear physical meaning, and using an XGBoost model for hierarchical decision-making, it overcomes the shortcomings of traditional methods in representing implicit degradation information such as local morphology and temporal evolution of voltage / current curves, while maintaining the interpretability and stability of the model, thus significantly improving the sorting accuracy and generalization ability. On this basis, by directly mapping the sorting level and confidence level output by the model to robot-executable tasks including material picking points, target placement points, and action types, and driving the mobile robot to complete autonomous transfer and unloading, it truly establishes a complete closed loop from data input, intelligent grading to automated execution. This effectively solves the problems of disconnect between algorithm decision-making and physical execution and high dependence on manual labor in existing technologies, and greatly improves the automation, efficiency, and traceability of retired power battery sorting. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the precision sorting system for retired power batteries of the present invention.

[0018] Figure 2This is a diagram of the one-dimensional CNN temporal feature extraction structure of the present invention.

[0019] Figure 3 This is a logic diagram of the CNN-XGBoost precise sorting algorithm of the present invention.

[0020] Figure 4 A closed-loop diagram for task generation and control of the mobile sorting robot.

[0021] Figure 5 Connection diagram for a mobile robot sorting system for retired power batteries.

[0022] Figure 6 This is a block diagram of the robot control system and circuit principle.

[0023] Figure 7 This is a state machine diagram of the mobile sorting robot of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Furthermore, in this invention, an element referred to as fixed to or disposed on another element may be directly disposed on the other element, or there may be an intermediate element. When an element is considered to be connected to another element, it may be directly connected to the other element, or there may be an intermediate element present simultaneously. The terms vertical, horizontal, left, right, and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0026] This invention proposes a sorting system and method for retired power batteries using a mobile robot based on a CNN-XGBoost fusion algorithm. The overall technical concept is as follows: given the obtained voltage and current detection data, a one-dimensional convolutional neural network is used to automatically extract deep dynamic response features such as voltage drop, recovery, plateau decay, and current fluctuation from the voltage and current time series. The deep features are then fused with manually statistical features using an extreme gradient boosting tree (XGBoost) to output the sorting level, confidence level, and target sorting area. The algorithm results are then directly mapped to the target area, path task, loading method, and unloading action that the mobile robot can execute. Finally, the mobile sorting robot completes the automatic transfer and unloading.

[0027] Example 1 This embodiment is combined with the appendix Figure 1 -Appendix Figure 7This application provides a detailed description of its technical solution: A mobile robot sorting method for retired power batteries is provided. The execution entity of this method can be a host computer, industrial control computer, edge computing terminal, robot scheduling server, or an embedded controller integrated into the mobile sorting robot. The method input consists of voltage and current data already obtained and bound to the battery number, and the method output consists of sorting level, sorting confidence level, target sorting area, and robot execution task. This invention does not focus on how the voltage and current data are collected at the front end as the core protection point, but rather uses the acquired voltage and current time series and the statistical characteristics derived from them as the input starting point, focusing on protecting the back-end accurate sorting closed loop of data analysis, sorting decision, and robot execution. The method specifically includes the following steps S1-S8: Step S1. Obtain the voltage time series and current time series of the retired power battery and bind them with the battery's unique number.

[0028] The system first establishes a sample record for each retired power battery, assigning each battery a unique number. Batch number Voltage time series Current time series Together with optional model, temperature, or historical information Mi, a traceable sample to be sorted is formed, which can be represented as... The purpose of this step is to ensure that the sorting level output by the subsequent algorithm can be accurately traced back to the specific battery and can be further transformed into a robot task. The system does not limit the specific acquisition method of voltage and current data, but only requires that each battery sample has a unique number and can be associated with voltage sequence, current sequence, batch information and optional model information.

[0029] Step S2. Preprocess the voltage time series and current time series. The preprocessing includes standardization and sliding window segmentation to form input segments of fixed length.

[0030] The purpose of this step is to organize the battery data that has been tested into a unified sample that can be directly identified by the algorithm; preprocessing includes two aspects: standardization and sliding window segmentation.

[0031] Standardization is used to eliminate dimensional differences caused by different voltage platforms, capacity specifications, and test scales, allowing subsequent models to focus more on the curve's variation rather than being influenced by absolute numerical scales. Standardization satisfies the following equation: ; in, and The first Voltage and current values ​​at each sampling point , These are the mean values ​​of the voltage and current sequences, respectively. , These are the standard deviations of the voltage series and the current series, respectively. To prevent small constants with a denominator of zero.

[0032] After standardization, the system uses a sliding window to segment time series of different lengths into fixed-length input segments to meet the input dimension requirements of subsequent one-dimensional convolutional neural networks, satisfying: ;

[0033] in, For the first A sliding window, For the preset window length, This is the sliding step size.

[0034] Step S3. Construct an artificial statistical feature vector, which contains statistical features extracted from the voltage time series and the current time series.

[0035] The purpose of this step is to extract a set of statistical features with engineering interpretation significance from the voltage and current sequences. Although the subsequent one-dimensional convolutional neural network can automatically learn deep curve features, deep features themselves are not easy to interpret. Therefore, this invention also retains manually generated statistical features to improve the stability and interpretability of XGBoost hierarchical decision-making.

[0036] Artificial statistical features primarily reflect the battery's voltage decay rate, current stability, short-time energy output capability, equivalent internal resistance level, and voltage fluctuation range during short-term testing. The constructed artificial statistical feature vector... It can be represented as: ; in, The slope of the voltage change reflects the battery's ability to maintain its platform. This is the current fluctuation coefficient, used to identify test stability; Characterized by short-time energy output; This is an estimate of the equivalent internal resistance; This refers to the voltage range; and These are the average voltage and average current, respectively.

[0037] One possible way to calculate the above features is as follows: ; ; ; It should be noted that the above formula only provides one feasible feature construction method. In practical applications, capacity estimation, SOH estimation, recovery voltage features, or temperature rise features can be added according to existing data conditions. The core of this invention is that artificial statistical features and deep CNN features are used together for accurate sorting, rather than limiting it to a single feature.

[0038] Step S4. Use a one-dimensional convolutional neural network to extract features from the input segment and output a deep dynamic response feature vector.

[0039] This step is the deep feature extraction stage in the precise sorting algorithm; retired power batteries often contain subtle aging information in their short-time voltage and current curves, such as voltage drop at the moment of loading, voltage recovery speed, voltage plateau decay, current disturbance, and local abrupt changes in the curve; relying solely on manual statistical features may not be sufficient to fully describe these local differences, therefore, this invention uses a one-dimensional convolutional neural network to automatically extract features from the time series.

[0040] Unlike two-dimensional image convolution, one-dimensional convolution kernels slide along the time axis, acting like a local dynamic feature scanner. When the kernel passes through regions of voltage drop, recovery, plateau decay, or current fluctuation, it automatically extracts the response pattern of that local segment. By setting convolution kernels of different lengths, the model can simultaneously focus on short-term abrupt changes, medium-scale variations, and longer-term trends, thus forming a more complete characterization of the battery's dynamic response.

[0041] The convolution operation of a one-dimensional convolutional neural network is specifically represented as follows: ;

[0042] In the formula, For the first Layer The convolutional channel at the _th convolutional channel in the _th ... Output at each time position; Input data for the previous layer; These are the kernel weights; For bias terms; The kernel length is [length]. This represents the number of channels in the previous layer. For time location; Number the positions inside the convolution kernel; The input channel number is used; this formula describes how the convolutional kernel slides across the time series and extracts local features.

[0043] After convolutional layers, the output undergoes batch normalization and ReLU non-linear activation. Then, global average pooling is used to compress the local information extracted from different time points into a deep dynamic response feature vector of uniform length. , can be represented as ,in This represents the total number of convolutional channels; ;

[0044] In the formula, Indicates batch normalization. Represents a non-linear activation function. This is the output of the c-th convolutional channel after global average pooling. This is the deep dynamic response feature vector that is the final output of 1D-CNN. Through global pooling, the model can compress local information extracted from different time positions into a feature vector of uniform length, which is convenient for subsequent fusion with manually statistical features.

[0045] Step S5. Fuse the deep dynamic response feature vector with the artificial statistical feature vector to generate a fused feature vector.

[0046] The purpose of this step is to merge the deep curve features automatically extracted by the one-dimensional convolutional neural network with the manually calculated statistical features to form a unified fused feature vector for XGBoost grading. The reason for this design is that CNNs are good at discovering hidden patterns in curves, but their interpretability is weak; manually calculated statistical features have clear physical meaning, but they are difficult to cover all nonlinear changes; after the two are merged, both good discriminative ability and engineering interpretability can be obtained at the same time.

[0047] Specifically, the system will use the deep dynamic response feature vector of each battery sample Artificial statistical feature vectors And optional model or batch information Perform a concatenation operation to form a fused feature vector. Furthermore, the fusion features are standardized to reduce dimensional differences between different batches and specifications of batteries, satisfying the following formula: ;

[0048] ;

[0049] In the formula, To fuse feature vectors, For the standardized fusion features, and These are the mean and standard deviation of the fusion features, respectively; the output of this step is... It will be used directly as input to the subsequent XGBoost classifier.

[0050] Step S6. Use the Extreme Gradient Boosting Tree (XGBoost) model to perform hierarchical decision-making on the fused feature vector, and output the sorting level and sorting confidence of the battery.

[0051] This step is the core decision-making process that transforms the fused features into specific sorting levels. XGBoost consists of multiple decision trees, each of which classifies and judges the features from different perspectives. Finally, the results of multiple trees are integrated to output the probability that the battery belongs to each sorting level. Compared with single threshold sorting, XGBoost can handle nonlinear relationships between multiple features. Compared with completely black-box deep networks, XGBoost can also explain the sorting criteria through feature importance analysis.

[0052] The ensemble output of the XGBoost model can be represented as: ;

[0053] In the formula, For model integration output, For the number of decision trees, For the k-th tree, This is a set of decision tree functions; the formula indicates that XGBoost does not rely on a single tree for judgment, but rather multiple trees work together to complete hierarchical decisions.

[0054] Subsequently, the model output is subjected to probability transformation, and the sorting level and confidence level are determined based on the maximum probability, satisfying the following formula: ;

[0055] ;

[0056] ;

[0057] in, For the i-th battery sample, it belongs to the i-th battery sample. The probability of each sorting grade. The total number of sorting grades, The raw score output by the XGBoost model. The final sorting grade is determined by... The confidence level for sorting.

[0058] The confidence threshold-based traffic splitting mechanism is a key inventive technical feature of this invention. If the confidence threshold... If the confidence level is higher than the preset threshold, the system will directly proceed to the next step to generate a robot sorting task; if the confidence level is lower than the preset threshold, the system will send the battery to the retest buffer area or the manual confirmation area instead of directly issuing a sorting instruction. This hierarchical, confidence level judgment and diversion mechanism effectively avoids the flow of low-confidence erroneous sorting into the subsequent automatic execution stage, and improves the reliability and robustness of the entire sorting system.

[0059] This invention recommends setting five sorting results: A, B, C, D, and R. The typical characteristics and uses of each level are as follows: Class A has high capacity, low internal resistance, and stable curves, suitable for high-value recycling; Class B has medium capacity and relatively stable state, suitable for ordinary recycling; Class C has discrete parameters and requires further grouping, suitable for secondary grouping or equalization processing; Class D has low capacity and high internal resistance, suitable for dismantling and recycling; Class R has insufficient confidence or abnormal data, requiring retesting or manual confirmation.

[0060] In addition, the XGBoost model can output feature importance ranking, so that the sorting results are not only black box level outputs, but can also be interpreted as hierarchical conclusions caused by certain dynamic features, further enhancing the engineering interpretability of the system.

[0061] Step S7. Determine the target sorting area based on the sorting level and sorting confidence mapping, and generate a robot sorting task. The robot sorting task includes at least the battery number, picking point, target delivery point, and action type.

[0062] The purpose of this step is to transform the algorithm's output levels into on-site executable sorting areas and robot tasks; the target area mapping satisfies the following formula: ;

[0063] In the formula, For the first The target sorting area corresponding to each battery; The function is used to map sorting grade, confidence level, and batch information to physical sorting bins or sorting areas.

[0064] The typical characteristics, algorithm judgment criteria, robot target areas, and subsequent uses of each level are shown in the table below:

[0065] The core function of this step is to transform the algorithm's judgment result into a target position that the robot can execute.

[0066] When generating a robot sorting task, the system packages the target area with information such as battery number, picking point, delivery point, loading method, and priority to form a structured executable instruction, which can be represented as: ;

[0067] In the formula, For the first The task of robotically sorting individual batteries; Each battery has a unique serial number. The coordinates of the material picking point; Coordinates of the target delivery point; Sorting regions for the target; For loading method; This determines the task priority.

[0068] Consistency optimization grouping is an important technical measure for improving sorting quality in this invention; after XGBoost grade judgment, this step further performs consistency optimization grouping on batteries of the same grade; firstly, an engineering feature vector for consistency grouping is constructed. ,in This is a capacity estimate. This is the estimated internal resistance value. This is an estimate of health status. For short-term energy, The slope of the voltage change. The current fluctuation coefficient is used; then the weighted distance between batteries of the same class is calculated: ;

[0069] in Let m be the weight of the m-th feature. This represents the total number of feature dimensions; clustering or matching is performed based on this weighted distance, ensuring that the grouping results satisfy a multi-objective optimization function: ;

[0070] in Indicates within-group differences. Indicates the robot's execution time. Indicates the number of retests. , , The weighting coefficients are used to optimize the grouping results. This optimization function comprehensively balances intra-group consistency, robot transportation efficiency, and retesting costs, ensuring that the grouping results not only pursue algorithmic accuracy but also take into account the overall efficiency of system operation.

[0071] In multi-robot scenarios, the system can calculate the scheduling cost based on each robot's current position, remaining battery power, and task queue length, and select the optimal robot to execute the task, satisfying the following formula: ;

[0072] In the formula, The robot to be scheduled is assigned a number; This is the robot's current position. The coordinates of the material picking point; The robot's remaining battery power; denoted as the length of the robot task queue; a, b, c, and d are preset weight coefficients; if the system has only one robot, the scheduling formula can be simplified to a fixed allocation.

[0073] Step S8. Send the robot sorting task to the mobile sorting robot to control the mobile sorting robot to perform material picking, transfer and unloading actions, and place the battery into the target sorting area.

[0074] This step explains how the robot completes the actual sorting action according to the task instructions. After receiving the task, the robot first moves from its current position to the picking point to load the battery or battery tray; then, it plans a path according to the target area coordinates and moves to the corresponding sorting area; finally, it performs the unloading action and sends a completion signal back to the host computer.

[0075] During the experimental phase, the robot chassis can be a two-wheel differential chassis, a four-wheel differential chassis, or a Mecanum wheel chassis. Navigation methods can include QR code positioning, magnetic strip navigation, visual positioning, or simple fixed routes; for industrial applications, this can be expanded to AGV / AMR, laser SLAM, UWB positioning, or multi-sensor fusion navigation. Regardless of the navigation method used, the core principle is that the sorting area output by the algorithm can be converted into position, path, and motion sequence by the robot control system.

[0076] Path planning can be used The algorithm, whose cost function is: ,in The actual cost from the starting point to node n. For the node The estimated cost to reach the target point; for the motion control of the differential moving chassis, satisfying... , ,in For the robot's linear velocity, Angular velocity, and The linear velocities of the right and left wheels are respectively. The wheelbase is used to determine the speed difference between the left and right wheels; by controlling the speed difference between the left and right wheels, the robot can move straight, turn, and rotate in place.

[0077] To ensure the reliability of the robot's sorting actions, the system uses a state machine for control. Each state must wait for sensor confirmation or timeout judgment before entering the next state, avoiding problems such as moving before the battery is loaded, unloading before it is in place, or the target area being incorrect. The state machine includes states such as standby, receiving task, moving to the picking point, picking and docking, loading confirmation, moving to the target sorting area, unloading confirmation, result feedback, and returning to standby. When any state experiences a timeout, obstacle avoidance trigger, emergency stop trigger, communication interruption, or loading failure, the robot enters an abnormal handling state and marks the current task as abnormal and pending.

[0078] Task completion confirmation requires all of the following conditions to be met: ;

[0079] in This indicates that the loading confirmation signal is valid. This indicates that the arrival confirmation signal is valid. This indicates that the uninstallation confirmation signal is valid. This indicates that the database update is complete. Only when all of the above conditions are met will the system determine that the battery sorting task is complete, thereby avoiding the mistaken conclusion that the task is over if the robot is not loaded, not in place, or not unloaded.

[0080] Model closed-loop update is another important technical feature of this invention. After the robot completes unloading, it returns the execution result to the host computer. The system records the battery number, sorting level, target area, robot number, execution time, anomaly flag, and retest result. If the retest result deviates significantly from the model output, the sample is marked as an incremental training sample. The XGBoost model is updated using an overall loss function that includes the retest sample correction loss and consistency constraint loss, satisfying the following conditions: ;

[0081] in For graded loss, To correct the loss for the retested samples, For consistency grouping constraint loss, and The weighting coefficients are used to continuously evolve the model during use, thereby improving sorting accuracy and adaptability to new scenarios.

[0082] Through the above steps S1 to S8, the system can complete a complete closed-loop process from voltage and current data input to CNN-XGBoost accurate classification, and then to automatic execution by the mobile sorting robot and result feedback update.

[0083] Example 2 To implement the method described in Embodiment 1, this embodiment provides a mobile robot sorting system for retired power batteries. This system is not simply a software algorithm, nor is it a single transport vehicle. Instead, it integrates algorithm decision-making, robot scheduling, mobile carrying, circuit control, sorting area, and data traceability into an executable closed loop.

[0084] The virtual functional module structure of this system mainly consists of a data input and number binding module, a data preprocessing module, a 1D-CNN feature extraction module, a manual statistical feature construction module, an XGBoost hierarchical decision module, a consistency optimization grouping module, a robot task scheduling module, and a robot execution feedback module.

[0085] Each module is connected sequentially according to the logic of data input, data processing, feature extraction, intelligent grading, group optimization, robot execution, and result feedback. The input end of the data input and number binding module is connected to an external detection device or database to receive the obtained voltage and current data and bind it to the battery number. Its output end is connected to the data preprocessing module. The data preprocessing module performs outlier handling, resampling, standardization, and sliding window reconstruction on the input data. Its output is connected to the 1D-CNN feature extraction module and the manual statistical feature construction module, respectively. The 1D-CNN feature extraction module automatically extracts deep features of dynamic response, and the manual statistical feature construction module calculates statistical features with engineering interpretability. After the two types of features are input into the feature fusion module, they are passed to the XGBoost grading decision module, which outputs the grade probability, sorting grade, and feature importance of the batteries to be sorted. The consistency optimization grouping module performs difference reduction and combination optimization on batteries within the same grade. The robot task scheduling module generates picking, transportation, and unloading tasks according to the target area. The robot execution feedback module receives sensor feedback signals to perform loading, placement, unloading, and abnormal status confirmation, thus forming a closed-loop sorting process.

[0086] The hardware structure of this system's physical robot is centered around a mobile sorting robot. It can be used for experimental verification or expanded into an industrial-grade AGV or AMR robot. The experimental mobile sorting robot mainly consists of a 24V power supply system, a mobile chassis, a main control unit, a navigation and positioning unit, a carrying mechanism, a drive unit, safety sensors, and multi-level sorting areas. The mobile chassis can be a four-wheel differential car chassis; the main control unit can be a Raspberry Pi or STM32 series microcontroller; the navigation and positioning unit can use QR codes, magnetic strips, line following, or AprilTag methods; the carrying mechanism can be an electric lifting pallet or a small roller platform; the drive unit can use a DC geared motor, encoder, and H-bridge drive circuit; safety sensors include photoelectric sensors, limit switches, ultrasonic sensors, and emergency stop buttons; the sorting areas are set as different levels of material boxes or ground areas such as A, B, C, D, and R. In industrial applications, the above structure can be further upgraded to an AGV / AMR chassis, a PLC or industrial robot controller, laser SLAM or vision SLAM navigation, an automatic roller docking platform, a servo drive system, safety radar, and an automatic hopper.

[0087] The robot control system adopts a layered structure consisting of a host computer algorithm layer, a main control execution layer, a drive power layer, and a sensor feedback layer. The host computer is responsible for CNN-XGBoost model inference and task allocation; the main control execution layer is responsible for the state machine and action timing; the drive power layer is responsible for the chassis motor, lifting motor, and roller motor; and the sensor feedback layer is responsible for positioning, loading, limit switching, obstacle avoidance, and completion confirmation. The power supply uses a 24V lithium battery pack, which is stepped down to 12V and 5V via a DC-DC module to supply the sensors, main control board, and communication module, respectively. The main power input is equipped with a fuse, emergency stop switch, reverse connection protection, and overcurrent protection. Communication between the host computer and the main controller can be achieved via USB serial port, RS485, CAN, Wi-Fi, Ethernet, or MQTT. The chassis motor drive circuit adopts an H-bridge drive structure. The main controller outputs PWM signals and direction signals to control the speed of the left and right wheels, and the encoder feeds back the actual speed to form a speed closed loop.

[0088] To ensure the reliability of the robot's sorting actions, the system uses a state machine for control. The state machine includes states such as standby, receiving task, moving to the picking point, picking and docking, loading confirmation, moving to the target sorting area, unloading confirmation, result transmission, and returning to standby. Each state must wait for sensor confirmation or timeout judgment before entering the next state. When any state experiences timeout, obstacle avoidance trigger, emergency stop trigger, communication interruption, or loading failure, the robot enters an abnormal handling state and marks the current task as abnormal and pending.

[0089] Example 3 The scope of protection of this invention is not limited to the specific embodiments described above; without departing from the core inventive concept of this invention, the following alternative solutions can achieve the same or similar technical objectives.

[0090] At the algorithm model level, 1D-CNN can be replaced by deep temporal feature extraction networks such as LSTM and Transformer-Encoder; XGBoost can be replaced by ensemble tree models such as LightGBM and Random Forest; at the feature input level, in addition to voltage and current time series and manually statistical features, auxiliary features such as temperature, SOC estimates, and BMS historical data can be added; at the consistency grouping level, the weighted distance combined with clustering method can be replaced by Gaussian mixture model, DBSCAN, spectral clustering, or genetic algorithm. At the robot morphology level, differential mobile robots can be upgraded to industrial AGVs, AMRs, Mecanum wheel robots, or track-mounted transport robots; at the navigation level, QR code / magnetic strip navigation can be upgraded to laser SLAM, visual SLAM, or UWB positioning; at the main control system level, STM32 can be replaced by Raspberry Pi, Jetson, industrial PC, or ROS controller; at the communication level, serial ports can be replaced by CAN, EtherCAT, ModbusTCP, or MQTT.

[0091] All the above alternative solutions retain the core logic of this invention: deep temporal feature extraction, integrated model hierarchical decision-making, confidence gating and diversion, consistency optimization grouping, and closed-loop execution of mobile robot tasks. Therefore, they are still equivalent or extended implementations of this invention.

[0092] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0093] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for sorting decommissioned power batteries using a mobile robot, characterized in that, Includes the following steps: Obtain the voltage and current time series of retired power batteries and bind them with the battery's unique serial number; The voltage and current time series are preprocessed, including standardization and sliding window segmentation, to form input segments of fixed length. Construct an artificial statistical feature vector, which contains statistical features extracted from the voltage time series and the current time series; A one-dimensional convolutional neural network is used to extract features from the input segment, and a deep dynamic response feature vector is output. The deep dynamic response feature vector is fused with the artificial statistical feature vector to generate a fused feature vector; The extreme gradient boosting tree (XGBoost) model is used to make hierarchical decisions on the fused feature vectors, and the sorting level and sorting confidence of the battery are output. The target sorting area is determined based on the sorting level and sorting confidence mapping, and a robot sorting task is generated. The robot sorting task includes at least the battery number, picking point, target delivery point and action type. The robot sorting task is sent to the mobile sorting robot to control the mobile sorting robot to perform material picking, transfer and unloading actions, and put the battery into the target sorting area.

2. The method according to claim 1, characterized in that, The standardization process and sliding window segmentation satisfy the following equation: ； ； in, and The first Voltage and current values ​​at each sampling point , These are the mean values ​​of the voltage and current sequences, respectively. , These are the standard deviations of the voltage series and the current series, respectively. To prevent small constants with a denominator of zero; For the first A sliding window, For the preset window length, This is the sliding step size.

3. The method according to claim 2, characterized in that, The convolution operation for feature extraction in the one-dimensional convolutional neural network is represented as follows: ； In the formula, This represents the output of the c-th convolutional channel in the l-th layer at the j-th time position. Input data for the previous layer, For convolution kernel weights, For bias terms, The kernel length is 1. This represents the number of channels in the previous layer. For time location, The position number inside the convolution kernel. Enter the channel number.

4. The method according to claim 3, characterized in that, The XGBoost model, an extreme gradient boosting tree model, is used to perform hierarchical decision-making on the fused feature vectors, outputting the battery sorting level and sorting confidence, satisfying the following formula: ； ； ； in, Let be the probability that the i-th battery sample belongs to the k-th sorting level, and C be the total number of sorting levels. The original score output by the XGBoost model. The final sorting grade is determined by... The confidence level for sorting.

5. The method according to claim 4, characterized in that, It also includes a consistency optimization grouping step: for batteries belonging to the same sorting level, construct engineering feature vectors. ,in This is a capacity estimate. This is the estimated internal resistance value. This is an estimate of health status. For short-term energy, The slope of the voltage change. The current fluctuation coefficient is used to calculate the weighted distance between batteries of the same class. ,in For the first The weights of each feature The number of features; clustering or matching is performed based on the weighted distance, such that the grouping satisfies the objective function. ,in Indicates within-group differences. Indicates the robot's execution time. Indicates the number of retests. , , These are the weighting coefficients.

6. The method according to claim 5, characterized in that, When multiple mobile sorting robots exist, the scheduling cost is calculated based on each robot's current position, remaining battery power, and task queue length. The optimal robot is then selected to perform the sorting task, satisfying the following formula: ； In the formula, Number the robot to be scheduled. This is the robot's current position. Here are the coordinates of the material collection point. The robot has remaining battery power. This is the length of the robot's task queue. , , , These are preset weighting coefficients.

7. The method according to claim 6, characterized in that, The mobile sorting robot is a differential wheel mobile chassis, and its motion control satisfies the following formula: ； ； in, For the robot's linear velocity, Angular velocity, and These are the linear velocities of the right and left wheels, respectively. This refers to the wheel track.

8. The method according to claim 7, characterized in that, It also includes a model closed-loop update step: when the sorting confidence level is lower than a preset threshold, the corresponding battery is sent to the retesting area for retesting; the retesting results are recorded, and the XGBoost model is updated using the overall loss function, which is: ； in, For graded loss, To correct the loss for the retested samples, For consistency grouping constraint loss, and These are the weighting coefficients.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 8.

10. A mobile robot sorting system for retired power batteries, characterized in that, include: The data input and preprocessing module is used to acquire the voltage and current time series of retired power batteries, bind them with the battery's unique number, and perform standardization and sliding window segmentation to form input segments of fixed length. The artificial statistical feature construction module is used to extract artificial statistical feature vectors from the voltage time series and current time series. A one-dimensional convolutional neural network feature extraction module is used to extract deep dynamic response feature vectors from the input segment using a one-dimensional convolutional neural network; The feature fusion module is used to fuse the deep dynamic response feature vector with the manually statistical feature vector to generate a fused feature vector; The XGBoost hierarchical decision module is used to classify the fused feature vector using the extreme gradient boosting tree XGBoost model, and output the battery sorting level and sorting confidence. The target area mapping and task generation module is used to determine the target sorting area according to the sorting level and sorting confidence mapping, and generate a robot sorting task that includes at least the battery number, picking point, target delivery point and action type. A mobile sorting robot is used to receive the sorting task from the robot and autonomously perform material picking, transfer and unloading actions to place the batteries into the target sorting area.

Images