A management method and system of an intelligent battery replacement cabinet

Abstract

The application relates to a management method and system of an intelligent battery replacement cabinet, solves the problem that it is difficult to adapt to complex scenes and diversified needs of outdoor battery replacement of electric vehicles, and restricts the safety performance and large-scale development of battery replacement services, the method comprises the following steps: responding to an electrical signal of a storage battery inserted into a warehouse body, completing adaptability authentication through an electrical signal detection component, and generating a connection health score in combination with a battery connector physical contact state and an electrical signal conduction quality; when the adaptation passes and the score is not lower than a preset abnormal threshold, triggering face authentication, comparing an authorized face database to obtain an identity authentication result; based on the collaborative decision of the two, generating a battery replacement authorization, auxiliary authentication or connection abnormal instruction, recording full-process data and synchronizing to a cloud management platform after corresponding operations are performed. The application has the following effects: guaranteeing the core safety of battery replacement connection, and balancing safety and energy supplement efficiency through battery and identity authentication collaboration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent battery replacement management and control of electric vehicles, and in particular to a management method and system of an intelligent battery replacement cabinet. BACKGROUND

[0002] With the popularity of electric vehicles (including electric bicycles, electric motorcycles, etc.), the battery replacement mode has become the mainstream energy supplement choice in the electric vehicle travel field due to its high efficiency, no need for users to charge themselves, and other advantages. The intelligent battery replacement cabinet, as the core terminal equipment of the electric vehicle battery replacement service, undertakes key tasks such as electric vehicle battery adaptation verification, connection state monitoring, user identity verification, and abnormal processing. The scientificity and reliability of its management scheme directly affect the energy supplement experience, travel safety of electric vehicle users, and the efficiency of the large-scale promotion of the battery replacement service. The industry has higher requirements for the full-process fine management and dynamic adaptation capability of the electric vehicle intelligent battery replacement cabinet.

[0003] In related technologies, the electric vehicle intelligent battery replacement cabinet has realized basic battery replacement process closed-loop management: the adaptation judgment is completed by detecting the basic electrical parameters of the electric vehicle battery, the battery connection state is monitored relying on simple sensing data, the electric vehicle user identity verification is completed using a traditional authentication method, and the battery replacement operation or abnormal alarm is triggered based on fixed rules, which basically meets the electric vehicle battery replacement needs in simple scenarios.

[0004] However, the existing management scheme still has obvious limitations: the monitoring and adaptation judgment of the battery state are single in dimension and lack precision, the user authentication and battery state management lack effective linkage, the decision logic is fixed, and it is difficult to adapt to the complex scenarios and diversified needs of electric vehicle outdoor battery replacement, which restricts the safety and efficiency of the battery replacement service and the large-scale development. SUMMARY

[0005] In order to protect the core safety of the battery replacement connection, balance safety and energy supplement efficiency through battery and identity authentication, the present application provides a management method and system of an intelligent battery replacement cabinet.

[0006] In a first aspect, the present application provides a management method of an intelligent battery replacement cabinet, which adopts the following technical scheme:

[0007] A management method of an intelligent battery replacement cabinet, comprising:

[0008] In response to an electrical signal generated by the insertion of the storage battery into the cabinet body, performing an adaptability authentication of the storage battery through an electrical signal detection component, and generating a battery connection health score based on the physical contact state of the battery connector and the electrical signal conduction quality;

[0009] When the battery compatibility certification is passed and the battery connection health score is not lower than the preset abnormal threshold, the face authentication process is triggered. The operator's face image is captured by the monitoring camera of the battery swapping cabinet and compared with the pre-stored authorized face database to obtain the user identity authentication result.

[0010] The system makes collaborative decisions based on battery connection health score and user authentication results: if user authentication is successful, a battery swap authorization instruction is generated; if user authentication fails but the battery connection health score is not lower than the first threshold, an auxiliary authentication instruction is generated; if the battery connection health score is lower than the first threshold, a connection abnormality instruction is generated.

[0011] The system executes corresponding operations based on the instructions generated by the decision: if it is a battery swap authorization instruction, it controls the battery swapping cabinet to perform a battery swapping operation; if it is an auxiliary authentication instruction, it guides the user to perform secondary authentication through the human-machine interface; if it is a connection error instruction, it triggers the audible and visual alarm component and displays battery connection guidance information on the human-machine interface.

[0012] After the operation is completed, the entire battery swap process data will be automatically recorded and synchronized to the cloud management platform.

[0013] By adopting the above technical solutions, and through battery compatibility certification and quantitative assessment of connection health, a solid safety barrier for battery swapping is built from the source; by combining facial recognition and dual-factor collaborative decision-making, a precise balance between safety control and energy replenishment efficiency is achieved; and the entire process data is synchronized to the cloud, providing support for the iteration of management strategies, thereby fundamentally improving the safety, reliability, and scenario adaptability of electric vehicle battery swapping services.

[0014] Secondly, this application provides a management system for an intelligent battery swapping cabinet, which adopts the following technical solution:

[0015] A management system for an intelligent battery swapping cabinet includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is loaded and executed by the processor, it implements the management method for the intelligent battery swapping cabinet as described in the first aspect. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a management method for an intelligent battery swapping cabinet according to an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of another embodiment of the present application, which shows the process of generating a battery connection health score based on the physical contact state and electrical signal conduction quality of the battery connector. Detailed Implementation

[0018] The present application will be further described in detail below with reference to the accompanying drawings.

[0019] ReferenceFigure 1 The present application discloses a management method for an intelligent battery swapping cabinet, comprising:

[0020] In step S100, in response to the electrical signal generated when the battery is inserted into the battery swapping cabinet, the battery compatibility is certified by the electrical signal detection component, and a battery connection health score is generated based on the physical contact state of the battery connector and the electrical signal conduction quality.

[0021] Among them, battery compatibility certification: A testing component identifies and verifies the batteries inserted into the battery swapping cabinet to ensure they meet the cabinet's specifications and can be safely and effectively charged and discharged. Battery connection health rating: Based on the physical contact state and electrical signal conduction quality of the battery connectors, a series of testing and evaluation methods are used to quantitatively score the reliability and stability of the battery-to-battery swapping cabinet connection, determining whether the connection is normal.

[0022] Necessary process description:

[0023] When a battery is inserted into the battery swapping cabinet, the system uses electrical signal detection components (including voltage sensors, current sensors, and contact resistance detection modules) to collect real-time physical contact parameters (such as contact resistance, insertion / removal force, and connection stability) and electrical signal conduction parameters (such as voltage drop, current stability, and CAN communication quality) of the battery connector. Simultaneously, temperature and humidity sensors acquire the temperature and humidity parameters of the battery swapping cabinet's operating environment and establish an environmental compensation model based on these parameters to dynamically correct the collected data, eliminating interference from environmental factors such as high temperature and high humidity on the detection results. For example, for the hard-plug connection method used in the battery swapping cabinet, the system will focus on monitoring the insertion / removal force and changes in contact resistance to ensure reliable connection between the battery and the charging contacts inside the cabinet.

[0024] After filtering and preprocessing, the corrected data is used to extract key features, including contact resistance change rate, voltage drop fluctuation amplitude, and CAN communication bit error rate. Entropy weighting is used to calculate the weights of each feature: by statistically analyzing the frequency distribution of features across different value ranges, the entropy value of each feature is calculated to assess its discriminative power, thereby determining the weight ratios of contact resistance change rate, voltage drop fluctuation amplitude, and communication bit error rate. After weighted fusion, a comprehensive evaluation index is formed, which, combined with a pre-set health scoring model (which can be trained based on machine learning algorithms such as support vector machines) and environmental compensation factors, ultimately generates a battery connection health score.

[0025] Step S200: When the battery compatibility certification is passed and the battery connection health score is not lower than the preset abnormal threshold, the face authentication process is triggered. The operator's face image is captured by the monitoring camera of the battery swapping cabinet and compared with the pre-stored authorized face database to obtain the user identity authentication result.

[0026] The facial recognition process involves several steps: After the battery is physically connected, the built-in monitoring camera in the battery swapping cabinet captures the operator's facial image. This image is then compared using a facial recognition algorithm with a pre-stored authorized facial database to verify the operator's identity. The pre-stored authorized facial database, located locally in the battery swapping cabinet or on a cloud management platform, contains facial feature vector data of registered users for identity verification. The database allows users to input their own data and temporarily authorize others to use it, meeting the needs of relatives and friends to swap batteries for them.

[0027] The process of obtaining user authentication results can be found in steps S210 to S270, and will not be elaborated here.

[0028] Step S300: Make a collaborative decision based on the battery connection health score and the user authentication result: If the user authentication is successful, generate a battery swap authorization instruction; if the user authentication fails but the battery connection health score is not lower than the first threshold, generate an auxiliary authentication instruction; if the battery connection health score is lower than the first threshold, generate a connection abnormality instruction.

[0029] Collaborative decision-making refers to a decision-making mechanism whereby the system integrates information from two dimensions—battery connection health score and user authentication result—and judges according to preset logical rules to generate differentiated operation instructions. The first threshold is a dynamic critical value used to distinguish whether the battery connection status is normal. A value below this threshold indicates a serious abnormality in the battery connection, requiring immediate intervention; a value not below this threshold but below the abnormal threshold allows for a second authentication attempt.

[0030] Necessary process description:

[0031] The system makes collaborative decisions based on the user authentication result output in step S200 and the battery connection health score generated in step S100. The decision engine first reads the health score value and compares it with a preset first threshold in real time. This first threshold is not a fixed value, but a dynamic value output by adjusting the model based on the real-time environmental parameters (such as temperature and humidity) of the battery swapping cabinet input threshold according to preset dynamic threshold rules. For example, the threshold requirement can be appropriately reduced in high temperature and high humidity environments to adapt to the objective influence of the environment on contact resistance.

[0032] If user authentication is successful, the system directly generates a battery swapping authorization command. This command will trigger the battery swapping cabinet to perform the battery replacement operation, including core actions such as unlocking the compartment door, switching battery status, and recording the swapping current. If user authentication fails but the battery connection health score is not lower than the first threshold, it indicates that the battery physical connection is good but the identity verification has failed. In this case, the system generates an auxiliary authentication command and guides the user to perform secondary authentication through the 13.3-inch human-computer interaction interface. For example, it may prompt the user to adjust their face position, clean the camera, or select an alternative authentication method (such as a temporary authorization code). The authentication process will be called again after the user tries again.

[0033] If the battery connection health score falls below the first threshold, the system will determine it as a connection error regardless of the user's authentication result and immediately generate a connection error command. This command will trigger the audible and visual alarm components (including a speaker beep and an LED warning light flashing) and display battery connection guidance information on the human-machine interface, instructing the user to check whether the battery is properly inserted, whether the connector is clean, or whether there is any physical damage, ensuring that connection problems are identified and resolved in a timely manner.

[0034] In step S400, the corresponding operation is executed according to the instruction generated by the decision: if it is a battery swap authorization instruction, the battery swap cabinet is controlled to perform a battery swap operation; if it is an auxiliary authentication instruction, the user is guided to perform secondary authentication through the human-machine interface; if it is a connection abnormality instruction, the audible and visual alarm component is triggered, and battery connection guidance information is displayed on the human-machine interface.

[0035] The system provides several instruction packages: Battery Swapping Authorization Instruction: This is the final authorization instruction generated by the system based on a comprehensive assessment of the battery connection health score and user authentication results, instructing the battery swapping station to perform a physical battery replacement. Auxiliary Authentication Instruction: This is a guiding instruction generated when user authentication fails but the battery connection is good, used to assist the user in completing secondary authentication through human-computer interaction. Connection Anomaly Instruction: This is a fault instruction generated when the battery connection health score falls below a first threshold, triggering an alarm mechanism and initiating a connection repair guidance process.

[0036] The necessary process is described below:

[0037] The system executes differentiated operations based on the instruction type generated in step S300. Upon receiving a battery swap authorization instruction, the battery swapping cabinet controller immediately executes the battery swapping operation process: first, it sends an unlock signal to the target compartment, opens the compartment door, and activates the battery rail system; then, it interacts with the battery BMS via CAN communication, switching the status of the battery to be charged to charging mode while simultaneously releasing the fully charged battery; after the swap is completed, the system updates the compartment status information, records the swapping current, and simultaneously deducts the user's account balance or card usage. The entire process is monitored in real time, with the progress displayed on the 13.3-inch screen as a "Replacing, please wait" status message.

[0038] If a secondary authentication command is received, the system guides the user through a human-machine interface to perform secondary authentication. The screen displays specific guidance content, such as "Please adjust your face position," "Please remove any obstructions," or "Please use a temporary authorization code," while a synchronized voice prompt plays through the speaker. The system provides an alternative authentication method entry point; the user can choose to scan a QR code for verification or enter a dynamic password. While the user is retrying authentication, the battery swapping cabinet maintains the current battery locked state to ensure operational safety.

[0039] If a connection error command is received, the system immediately triggers the audible and visual alarm components: the speaker emits a continuous buzzing warning sound (frequency 2Hz), the LED strip on the door switches to a flashing red state, and a warning pop-up appears on the screen. The human-machine interface simultaneously displays battery connection guidance information, including a 3D animation demonstrating the correct insertion posture, connector cleaning methods, and steps for troubleshooting physical damage. For hard-plug connections, a special prompt states, "Please push firmly to the bottom until you hear a click." After the user reinserts and removes the battery according to the guidance, the system automatically initiates a connection status reassessment mechanism, re-collects parameters, and generates a health score. If the score meets the requirements, the alarm is automatically deactivated, and the battery swapping process resumes.

[0040] Step S500: After the operation is completed, the entire battery swap process data will be automatically recorded and synchronized to the cloud management platform.

[0041] Among them, the full-process data of battery swapping covers all operation information from battery insertion to replacement completion, including structured datasets such as battery parameters (model, health score, BMS status), user authentication information (identity ID, authentication result, failure reason), operation logs (timestamp, compartment number, instruction type), environmental parameters (temperature and humidity) and equipment operating status (charging power, communication quality).

[0042] Cloud Management Platform: A remote management system built on the RUOYI framework (SpringCloud) microservice architecture. It communicates with the battery swapping cabinet in real time via 4G network or WiFi to realize functions such as data storage, equipment monitoring, fault early warning and operation analysis.

[0043] The necessary process is described below:

[0044] Upon completion of the operation, the system immediately initiates the data recording process. First, it collects all data from each functional module for this battery swap: battery model, health score, and BMS communication logs from the battery management module; user ID, facial recognition similarity value, and authentication result from the identity authentication module; operation time, compartment number, and instruction execution status from the compartment control board; and temperature and humidity data from environmental sensors. All data is packaged in a preset format (such as a JSON structure) and a data verification code (CRC32) is attached to ensure integrity. Simultaneously, the device's local timestamp is recorded as the operation time reference.

[0045] The packaged data is temporarily stored in the local storage of the battery swapping cabinet (such as embedded Flash or SD card), forming an operation log queue. The system attempts to establish a connection with the cloud management platform through the built-in network module (4G or WiFi) and uses the MQTT protocol for data push. If the network is available, the data is encrypted with AES and uploaded in real time through a TLS secure channel. The cloud receives the data, verifies the checksum, and parses it into the database. If the network is abnormal (such as weak signal or no network), the data is temporarily stored in the local queue and retransmitted in batches in chronological order after the network is restored. A breakpoint resume mechanism is supported to avoid data loss. After synchronization is complete, the local data is cyclically overwritten according to a preset retention strategy (such as the most recent 1000 records) to prevent storage overflow.

[0046] Based on the physical contact state and electrical signal conduction quality of the battery connector, a battery connection health score is generated, including:

[0047] Step S110: Real-time acquisition of physical contact parameters and electrical signal conduction parameters of the battery connector, and data preprocessing. The physical contact parameters include contact resistance, insertion and removal force, and connection stability. The electrical signal conduction parameters include voltage drop, current stability, and communication quality.

[0048] Physical contact parameters refer to quantitative indicators reflecting the physical contact quality between the battery connector and the charging interface of the battery swapping cabinet, including contact resistance (measuring the conductivity of the electrical contact point), insertion and extraction force (mechanical characteristics during battery insertion), and connection stability (continuous reliability of the contact state). Electrical signal conduction parameters refer to indicators characterizing the electrical connection and communication quality between the battery and the battery swapping cabinet, including voltage drop (potential difference when current passes through the contact point), current stability (fluctuations in charging current), and communication quality (bit error rate and response delay of CAN bus data transmission).

[0049] The necessary process is described below:

[0050] The system acquires physical contact and electrical signal conduction parameters in real time through an electrical signal detection component. For physical contact parameters, at the moment the battery is inserted into the compartment, a four-wire micro-resistance measurement circuit detects the contact resistance value (measurement current 10mA, accuracy ±0.1mΩ), while a pressure sensor built into the compartment guide rail records the insertion and removal force curve (sampling frequency 100Hz). During the 3-second stabilization period after battery insertion, the system continuously monitors the contact resistance fluctuation amplitude to assess connection stability. For electrical signal conduction parameters, a high-precision voltage sensor measures the voltage drop between the battery's positive and negative terminals and the cabinet contact points (resolution 1mV), a Hall current sensor monitors the current stability at the moment of charging start-up (acquisition time 500ms, standard deviation calculated), and a CAN communication module calculates the bit error rate of interaction with the battery BMS (statistical window 100 frames of data).

[0051] The collected raw data first underwent preprocessing: a moving average filter (window length 10 points) was used to remove high-frequency noise introduced by mechanical vibration, Z-score standardization was used to eliminate the influence of different physical dimensions, and obvious outliers were removed based on the 3σ criterion. The preprocessed data formed a standardized parameter set, providing high-quality input for subsequent environmental compensation and feature extraction.

[0052] Step S120: Based on the environmental parameters obtained by the temperature sensor and humidity sensor, an environmental compensation model is established, and the preprocessed physical contact parameters and electrical signal parameters are dynamically corrected. The compensated data is output, and the environmental compensation factor is calculated based on the environmental parameters and the preset influence coefficient matrix built into the environmental compensation model.

[0053] Among them, environmental parameters refer to the temperature and humidity data of the battery swapping cabinet's operating environment, which are acquired in real time through temperature and humidity sensors. This data is used to quantitatively assess the interference of environmental factors on battery connection detection. The environmental compensation model is a dynamic correction algorithm based on historical data. It uses real-time environmental parameters as input variables to perform nonlinear corrections on physical contact parameters and electrical signal parameters, eliminating the influence of temperature and humidity on measurements such as contact resistance and voltage drop, and outputting compensated standardized data. The preset influence coefficient matrix is ​​a parameter mapping table pre-set in the environmental compensation model, defining the compensation coefficients for various physical quantities (contact resistance, voltage drop, etc.) within different temperature and humidity ranges, used for rapid calculation of correction amounts.

[0054] The necessary process is described below:

[0055] The system collects environmental parameters in real time using temperature and humidity sensors deployed inside the battery swapping cabinet. The temperature sensor uses a digital probe, covering the operating temperature range of 0°C to 50°C, and is installed on the side wall of the cabinet to avoid interference from battery heating. The humidity sensor detects humidity levels from 5% to 75% for non-condensing operation and is placed near the cabinet's ventilation openings to obtain accurate environmental data. Both sensors sample synchronously at a frequency of 1Hz, and the data is digitally filtered to form an environmental parameter vector.

[0056] Based on the collected environmental parameters, the system invokes a preset environmental compensation model to dynamically correct the preprocessed physical contact parameters and electrical signal parameters. This model employs a multiple linear regression algorithm framework with a built-in preset influence coefficient matrix. This matrix is ​​stored in a two-dimensional lookup table, with row coordinates representing temperature ranges (in 5°C increments) and column coordinates representing humidity ranges (in 10% increments). Each cell stores four compensation coefficients: contact resistance temperature coefficient, voltage drop humidity coefficient, current stability temperature coefficient, and communication quality humidity coefficient. For example, when the ambient temperature is 35°C and the humidity is 60%, the lookup matrix yields a contact resistance temperature coefficient of 1.03 and a voltage drop humidity coefficient of 1.08. The correction process calculates the compensation amount in real time: compensated contact resistance = original contact resistance × temperature coefficient, compensated voltage drop = original voltage drop × humidity coefficient, thereby eliminating the baseline drift of the measured values ​​caused by environmental factors and outputting a standardized set of corrected parameters.

[0057] While performing dynamic correction, the system calculates an environmental compensation factor based on environmental parameters and a preset influence coefficient matrix. This factor is calculated using a weighted arithmetic mean method, with the formula: Environmental Compensation Factor = Σ(Compensation Coefficient_i × Weight_i), where i takes values ​​from 1 to 4, corresponding to the four compensation coefficients. The weight allocation is preset based on the degree of influence of each parameter on the final health score: contact resistance temperature coefficient weight 0.4, voltage drop humidity coefficient weight 0.3, current stability temperature coefficient weight 0.2, and communication quality humidity coefficient weight 0.1. Taking the above 35℃, 60% humidity environment as an example, if the four coefficients found from the matrix are 1.03, 1.08, 1.02, and 1.05 respectively, then the environmental compensation factor = 1.03×0.4 + 1.08×0.3 + 1.02×0.2 + 1.05×0.1 = 0.412 + 0.324 + 0.204 + 0.105 = 1.045. In step S140, this factor directly affects the output of the health score model, enabling dynamic adjustment of the scoring benchmark. For example, if the factor is calculated to be 0.95 under high temperature and high humidity conditions, the final health score needs to be multiplied by 0.95 to eliminate environmental interference and ensure consistency in equipment evaluation standards across different seasons and regions.

[0058] Step S130: Extract key features from the compensated data: contact resistance change rate, voltage drop fluctuation amplitude, and communication error rate. Use the entropy weight method to determine the weight of each feature and perform weighted fusion to form a comprehensive evaluation index.

[0059] Among them, the contact resistance change rate refers to the degree of change in contact resistance over time after battery insertion. It quantifies the stability of the contact state by calculating the percentage change in resistance value within adjacent sampling periods. Voltage drop fluctuation amplitude reflects the stability of voltage drop when current flows through the battery connector. It assesses connection quality by the difference between the maximum and minimum voltage drop values ​​within a statistical time window. Communication bit error rate characterizes the reliability of data transmission via CAN communication between the battery swapping cabinet and the battery BMS. It is calculated by statistically analyzing the proportion of erroneous frames to the total number of transmitted frames. Entropy weighting method is a mathematical method that objectively determines the weight of each indicator based on the entropy value of feature information. The smaller the entropy value, the stronger the distinguishing ability of the feature, and the higher the corresponding weight.

[0060] The specific process of determining the weights of each feature and weighting and integrating them using the entropy weight method to form a comprehensive evaluation index can be found in steps S131 to S135, and will not be elaborated here.

[0061] Step S140: Based on the preset health score model, the battery connection health score is generated by taking the comprehensive evaluation index as the core input and combining environmental compensation factors.

[0062] The necessary procedures are as follows:

[0063] The system receives the comprehensive evaluation index output in step S130 and the environmental compensation factor calculated in step S120, and inputs them together into a preset health score model. This model uses a piecewise linear mapping function as its basic architecture: the numerical range of the comprehensive evaluation index [0,1] is divided into several sub-intervals, each corresponding to a preset health score range. For example, when the comprehensive evaluation index is in the 0-0.3 range, it is mapped to a health score of 90-100; the 0.3-0.6 range is mapped to 70-90; and the 0.6-1.0 range is mapped to 50-70. The mapping process uses linear interpolation to calculate the initial score, ensuring the continuity and precision of the score output.

[0064] After the initial score is generated, the system introduces an environmental compensation factor for dynamic correction. The correction formula is: Final health score = Initial score × Environmental compensation factor. This factor typically ranges from 0.9 to 1.1. When environmental conditions are harsh (e.g., high temperature and humidity), the factor is less than 1.0, and the score is appropriately lowered to eliminate environmental interference; under ideal environmental conditions, the factor is close to 1.0, and the scoring benchmark remains unchanged. For example, if the comprehensive evaluation index for a certain test is 0.25, and the initial score obtained through model mapping is 95, while the environmental compensation factor is 0.95, then the final health score is 95 × 0.95 = 90.25. The rounded integer score is passed to step S200 as a basis for collaborative decision-making, and simultaneously cached in local memory for data synchronization in step S500. This mechanism effectively ensures the consistency of scoring standards across different regions and seasons, avoiding misjudgments caused by environmental factors.

[0065] The entropy weight method is used to determine the weights of each feature and then weighted and fused to form a comprehensive evaluation index, including:

[0066] Step S131: Extract three key features from the compensated data: contact resistance change rate, voltage drop fluctuation amplitude, and communication error rate. Statistically analyze the distribution frequency of each feature in different value ranges according to sample batches to form a feature value distribution table.

[0067] Among them, the contact resistance change rate reflects the fluctuation of contact resistance over time after battery insertion. It measures connection stability by calculating the percentage change in resistance value between adjacent sampling periods; a smaller value indicates more stable contact. Voltage drop fluctuation amplitude characterizes the stability of voltage drop when current flows through the battery connector. It quantifies connection quality by statistically analyzing the difference between the maximum and minimum voltage drop values ​​within a fixed time window. Communication bit error rate characterizes the reliability of data transmission in CAN communication between the battery swapping cabinet and the battery BMS. It is calculated by statistically analyzing the proportion of erroneous frames to the total number of transmitted frames. Feature value distribution table: A structured statistical table is generated by statistically analyzing the sample distribution frequency of the three key features in different value ranges according to sample batches, providing basic data for subsequent entropy weight calculation.

[0068] The necessary procedures are as follows:

[0069] The system extracts three key features from the compensated data output in step S120. For the contact resistance change rate, the percentage change in resistance values ​​at adjacent sampling points within a 10-second stabilization period after battery insertion is calculated and averaged by absolute value. For example, if the contact resistance sequence for a given test is [1.20mΩ, 1.22mΩ, 1.25mΩ], the change rate is [(1.22-1.20) / 1.20+(1.25-1.22) / 1.22] / 2=2.05%. For the voltage drop fluctuation amplitude, voltage drop data within 500ms after charging starts is collected, and the difference between the maximum and minimum values ​​is calculated. For example, if the voltage drop fluctuates within the range of [48mV, 52mV, 51mV, 49mV] within 10ms, the amplitude is 52-48=4mV. For the communication error rate, the number of erroneous frames is counted within a 100-frame data window after establishing communication with the battery BMS. For example, if 3 erroneous frames are detected, the error rate is 3%. The three features are normalized to form feature sample values, which are then used as statistical inputs.

[0070] The system calculates the frequency distribution of three features in different value ranges according to preset sample batches (each batch contains the most recent 100 battery swap operation records). First, the value intervals are divided: the contact resistance change rate is divided into four intervals: [0-2%], (2-5%], (5-10%], and (10-15%]; the voltage drop fluctuation amplitude is divided into four intervals: [0-3mV], (3-5mV], (5-8mV], and (8-15mV]; and the communication bit error rate is divided into four intervals: [0-1%], (1-3%], (3-5%], and (5-10%). Then, all samples within the batch are iterated, and the number of samples for each feature falling into each interval is counted, forming a feature value distribution table. For example, statistics for a certain batch show that the contact resistance change rate is distributed 45 times in the (2-5%) interval, the voltage drop fluctuation amplitude is distributed 30 times in the (3-5mV) interval, and the communication bit error rate is distributed 60 times in the (1-3%) interval. This distribution table is stored in matrix form, providing structured data input for the entropy calculation in step S132.

[0071] Step S132: Based on the preset entropy weight calculation rules, using the feature value distribution table as input, calculate the normalized proportion of each feature in different samples, and output the entropy value of each feature through the entropy value algorithm.

[0072] The entropy weighting calculation rule is based on an objective weighting method determined by information entropy theory. It assesses the information value of a feature by calculating its dispersion within the sample. A smaller entropy value indicates a stronger ability to distinguish samples, and therefore a higher weight should be given in the comprehensive evaluation. Normalized proportion refers to the ratio of the number of samples containing a particular feature within a specific value range to the total number of samples containing that feature. It is used to eliminate the influence of differences in sample size among different features, making each feature comparable. Entropy value is a measure of feature uncertainty calculated using the entropy algorithm, ranging from 0 to 1. A larger entropy value indicates less information provided by the feature and a lower contribution to the evaluation results.

[0073] The necessary process is described below:

[0074] The system reads the feature value distribution table generated in step S131 and counts the total number of samples for each feature and the number of samples in each interval. When calculating the normalized percentage, the sample frequency of each interval is divided by the total number of samples for that feature to obtain the probability distribution of each interval. For example, if the total number of samples for the contact resistance change rate in a certain batch is 100, with 20 samples in the (0-2%) interval, 45 samples in the (2-5%) interval, 30 samples in the (5-10%) interval, and 5 samples in the (10-15%) interval, the corresponding normalized percentages are 0.20, 0.45, 0.30, and 0.05, respectively.

[0075] The information entropy of each feature is calculated based on the normalized proportion, using the standard entropy algorithm formula: Entropy = -Σ(p_ij×ln(p_ij)) / ln(n), where p_ij represents the normalized proportion of the i-th feature in the j-th interval, and n is the number of intervals. Taking the aforementioned contact resistance change rate as an example, its entropy value is calculated as: -[0.20×ln(0.20)+0.45×ln(0.45)+0.30×ln(0.30)+0.05×ln(0.05)] / ln(4)≈0.823. Similarly, the voltage drop fluctuation amplitude entropy value is calculated to be 0.756, and the communication bit error rate entropy value is 0.892. The calculation results form a feature entropy vector, which is used for subsequent steps of discrimination ranking and weight determination. When the number of samples in a certain interval is 0, its normalized proportion is specified to take a minimum value ε (such as 0.0001) to avoid logarithmic calculation errors.

[0076] Step S133: Based on the entropy values ​​of each feature, directly evaluate its discrimination and form a discrimination ranking.

[0077] Discrimination ranking refers to the process of sorting features from highest to lowest based on their entropy values, according to their ability to distinguish differences between samples. A smaller entropy value indicates a greater degree of dispersion of the feature within the samples, providing more information and thus a stronger discriminative ability.

[0078] The necessary process is described below:

[0079] The system receives the entropy values ​​of each feature calculated in step S132 and directly evaluates the feature discrimination by comparing the entropy values. The evaluation rule is clear: the smaller the entropy value, the stronger the feature's ability to distinguish the battery connection state, and the higher its importance in the evaluation system. For example, if a batch has a calculated entropy value of 0.823 for contact resistance change rate, 0.756 for voltage drop fluctuation amplitude, and 0.892 for communication error rate, the system will sort them in ascending order of entropy value to form a discrimination ranking: voltage drop fluctuation amplitude (entropy value 0.756) > contact resistance change rate (entropy value 0.823) > communication error rate (entropy value 0.892). This ranking result is directly used to guide the weight allocation in step S134, ensuring that the feature with the highest discrimination receives the maximum weight, thereby improving the accuracy of the comprehensive evaluation index.

[0080] Step S134: Based on the discrimination ranking results, determine the weight ratios of contact resistance change rate, voltage drop fluctuation amplitude, and communication error rate through a preset weight normalization rule.

[0081] The weight normalization rule is a mathematical rule that calculates and standardizes weights based on feature discrimination to ensure that the sum of the weights of each feature is 1, forming a comparable weight allocation scheme. The weight ratio refers to the normalized weight value of the three features—contact resistance change rate, voltage drop fluctuation amplitude, and communication bit error rate—in the comprehensive evaluation index, directly determining the degree of influence of each feature on the final score.

[0082] The necessary process is described below:

[0083] The system reads the discrimination ranking results generated in step S133 and assigns weights according to the principle that the smaller the entropy value, the greater the weight. Specifically, the difference coefficient for each feature is calculated first: Difference coefficient = 1 - Entropy value. For example, if the entropy value of voltage drop fluctuation amplitude is 0.756, its difference coefficient is 0.244; the entropy value of contact resistance change rate is 0.823, and its difference coefficient is 0.177; the entropy value of communication bit error rate is 0.892, and its difference coefficient is 0.108.

[0084] Subsequently, the system calculates the weight ratio of each feature using a weight normalization rule: Weight ratio = Difference coefficient_i / Σ (Difference coefficient). Taking the above data as an example, the sum of the difference coefficients of the three features is 0.244 + 0.177 + 0.108 = 0.529. Therefore, the weight of voltage drop fluctuation amplitude is approximately 0.46, the weight of contact resistance change rate is approximately 0.177 / 0.529, and the weight of communication error rate is approximately 0.20. This weight ratio ensures that the voltage drop fluctuation amplitude, which has the highest distinguishing power, dominates the comprehensive evaluation, while also taking into account the contributions of other features.

[0085] After the calculation is completed, the system outputs a structured weight vector [0.33, 0.46, 0.20], which is directly passed to step S135 for weighted fusion. This weight vector is synchronously cached in local memory for reuse by subsequent samples in the same batch, avoiding redundant calculations and improving processing efficiency.

[0086] Step S135: According to the preset weighted fusion rules, the compensated feature values ​​are integrated with the determined feature weights to generate a comprehensive evaluation index.

[0087] The weighted fusion rule refers to a mathematical calculation criterion that uses the weight ratio determined in step S134 as coefficients to linearly weight and sum the three normalized feature values, ensuring that different features are integrated into a single evaluation index according to their degree of impact on connection health. The comprehensive evaluation index is a standardized value generated by weighting and fusing three features—contact resistance change rate, voltage drop fluctuation amplitude, and communication bit error rate—to intuitively reflect the overall health status of the battery connection. The value range is typically from 0 to 1.

[0088] The necessary process is described below:

[0089] The system reads the weight vector output in step S134 (e.g., contact resistance change rate weight 0.33, voltage drop fluctuation amplitude weight 0.46, and communication error rate weight 0.20) and the compensated feature values ​​output in step S120. The three feature values ​​are first normalized: the contact resistance change rate is normalized by 15% of its maximum value, the voltage drop fluctuation amplitude by 15mV, and the communication error rate by 10%, mapping each feature value to the [0,1] interval. The normalization formula is: Normalized value = Original value / Feature upper limit value, ensuring comparability of features with different dimensions.

[0090] After normalization, the system performs integrated calculations according to weighted fusion rules: Comprehensive evaluation index = Σ(Normalized feature value_i × Weight_i). For example, in a certain test, the normalized contact resistance change rate is 0.15, the voltage drop fluctuation amplitude is 0.30, and the communication bit error rate is 0.25, then the comprehensive evaluation index = 0.15 × 0.33 + 0.30 × 0.46 + 0.25 × 0.20 = 0.0495 + 0.138 + 0.05 = 0.2375. The closer this index value is to 0, the more ideal the connection status; the closer it is to 1, the more serious the connection problem.

[0091] The calculated comprehensive evaluation index is transmitted in real time to step S140 as the core input of the health scoring model, and is also cached in local storage for historical data analysis and model optimization in step S500. To ensure computational robustness, the system sets anomaly boundary protection for normalized feature values: if a feature value exceeds the normal upper limit (e.g., contact resistance change rate > 20%), it is automatically truncated to 1.0 and an anomaly flag is triggered to avoid extreme data affecting the overall evaluation accuracy.

[0092] The operator's facial image is captured by the monitoring camera of the battery swapping cabinet and compared with a pre-stored authorized facial database to obtain the user's identity authentication result, including:

[0093] Step S210: The operator's face image is captured in real time by the monitoring camera built into the battery swapping cabinet, and the ambient light intensity, camera pitch angle and yaw angle parameters are recorded simultaneously during the acquisition.

[0094] Among them, the monitoring camera is an image acquisition device integrated above or at the top of the screen in front of the battery swapping cabinet. It is used to capture the operator's facial image in real time, and usually has 1080P resolution, wide dynamic range and infrared night vision function, adapting to complex indoor and outdoor lighting environments. Pitch angle and yaw angle are the deflection angle parameters of the camera relative to its standard installation position. The pitch angle refers to the tilt angle of the camera in the vertical direction, and the yaw angle refers to the rotation angle in the horizontal direction, used to record the geometric attitude during image acquisition.

[0095] The necessary process is described below:

[0096] Once the system triggers the facial recognition process, the monitoring camera built into the battery swapping cabinet immediately begins image acquisition. The camera captures the operator's facial video stream in real time at a rate of 30 frames per second and automatically selects the single frame with the highest clarity as the image to be processed.

[0097] The ambient light intensity during data acquisition is recorded synchronously, measured by the camera's built-in photosensor or a separate illuminance sensor, with a range of 0-20000 Lux and a resolution of 1 Lux. The data is stored bound to the image frame. This light intensity value is subsequently used for dynamic threshold adjustment in step S250. For example, in low-light environments (illuminance <100 Lux), the face matching threshold can be appropriately reduced to improve the pass rate.

[0098] The camera's pitch and yaw angle parameters are acquired in real time via a built-in MEMS angle sensor, which is installed inside the camera module. The sensor has a measurement accuracy of ±1° and a sampling frequency of 10Hz. When the user's height difference causes the face to deviate from the camera center, the system records the current angle deviation value (e.g., pitch angle +15°, yaw angle -8°). This data is used in the angle alignment correction step S230 to ensure that the face image can be geometrically transformed to a standard frontal pose. All acquired parameters (light intensity, angle values) and image data are packaged into a structured data packet with timestamps as the association key and transmitted to the image preprocessing module via the internal bus, providing complete metadata support for subsequent image quality screening and environmental adaptive correction.

[0099] Step S220: Qualified images are selected based on preset image quality rules. The image edge sharpness is calculated using the Laplacian variance algorithm based on a preset sharpness score threshold. If the score is lower than the preset threshold, the image is judged as blurry. The face region is detected by a semantic segmentation model based on a preset occlusion area ratio threshold. If the occlusion area exceeds the preset threshold, the image is judged as over-occluded. For blurry or over-occluded images, a re-sampling mechanism is triggered to adjust the camera to a preset angle range and then re-sample.

[0100] The image quality rules refer to pre-defined multi-dimensional image qualification criteria, including a sharpness scoring threshold (usually set to a Laplacian variance of 100-150) and an occlusion area percentage threshold (usually set to 30%-40%), used to quantitatively evaluate whether the captured image meets the requirements for face recognition. The Laplacian variance algorithm calculates the variance of pixel values ​​after convolution with the Laplacian operator to quantitatively evaluate the edge sharpness of the image; a larger variance value indicates a sharper image. The semantic segmentation model uses a lightweight convolutional neural network (such as the U-Net architecture based on MobileNetV2) to detect face regions at the pixel level, accurately identify facial contours and the location of facial features, and provide a segmentation mask for calculating the occlusion area. The re-acquisition mechanism is an automatically triggered re-acquisition process when an image is judged to be blurry or excessively occluded, re-acquiring the image by adjusting the camera's pitch angle (±20° range) and yaw angle (±30° range).

[0101] The necessary process is described below:

[0102] The system performs dual quality checks on the face image captured in step S210. First, blurriness is determined: the Laplacian variance algorithm is used to calculate image sharpness. Specifically, a 3×3 Laplacian kernel convolution operation is performed on the grayscale image, and the variance of the resulting matrix is ​​used as the sharpness score. If the score is lower than a preset threshold of 120, the image is determined to be blurry. For example, if the image edges are blurred due to user hand tremors, the calculated variance may only be 80, immediately triggering a re-capture.

[0103] Simultaneously, occlusion detection is performed: a pixel-level mask for the face region is generated using a semantic segmentation model, and the total area of ​​the mask is compared with the area of ​​the occluded region (such as facial pixels covered by masks, hats, hair, etc.). If the occluded area exceeds a preset threshold of 35%, it is considered excessive occlusion. For example, when a user is wearing a KN95 mask, if the model detects that the area of ​​the mouth and nose is occluded by 60%, the system will determine it as unacceptable.

[0104] When an image is deemed blurry or excessively occluded, the system immediately triggers a re-sampling mechanism. The speaker issues a voice prompt, "Please remain still" or "Please remove obstructions," while simultaneously controlling the camera drive motor to adjust the angle: the pitch angle scans vertically in 5° increments within ±20°, and the yaw angle scans horizontally in 8° increments within ±30°, searching for the optimal shooting angle. A maximum of three re-sampling attempts are made, with a one-second interval between each attempt. If the image is still unsatisfactory after three attempts, specific guidance information is displayed on the 13.3-inch screen, such as "Please move closer to the camera 30-50cm" or "Please adjust to a well-lit position," guiding the user to cooperate until a satisfactory image is obtained or the maximum number of attempts is reached, after which the system transitions to the error handling process.

[0105] Step S230: Perform preset environmental adaptive correction on qualified images: eliminate backlight / low light interference through illumination compensation rules, align the face to the frontal orientation based on angle alignment rules, fill the temporary occlusion area through occlusion repair rules, and generate a preprocessed face image.

[0106] The system includes several key components: Environmental Adaptive Correction: A set of algorithms for standardizing acquired images under complex outdoor lighting conditions and varying user postures at battery swapping stations. This includes three core steps: illumination compensation, angle alignment, and occlusion repair, ensuring the generation of standard frontal face images that meet recognition requirements. Illumination Compensation Rules: Image enhancement algorithms based on histogram equalization and Retinex theory are used to eliminate underexposure caused by backlighting or overexposure caused by strong light, improving overall image contrast and detail visibility. Angle Alignment Rules: Through facial landmark detection and affine transformation matrix calculation, non-frontal face images are rotated and corrected to a standard frontal orientation, eliminating the impact of pitch and yaw angle deviations on feature extraction. Occlusion Repair Rules: For minor occlusions (such as temporary mask wear or glare from glasses), a generative adversarial network (lightweight Pix2Pix model) is used for pixel-level repair, filling in missing facial areas and restoring a complete facial structure.

[0107] The necessary process is described below:

[0108] After receiving the qualified images selected in step S220, the system performs three environmental adaptive corrections in sequence.

[0109] First, illumination compensation rules are applied: the image grayscale histogram is analyzed. If backlighting (average background brightness > 150 Lux and face area < 50 Lux) or low light (overall average brightness < 30 Lux) is detected, the Adaptive Histogram Equalization (CLAHE) algorithm is activated. The image is divided into 8×8 blocks, and histogram equalization is performed independently on each block with a contrast limit threshold of 3.0 to avoid noise amplification. Simultaneously, Retinex multi-scale enhancement is performed, extracting illumination components at Gaussian filter scales of 15, 80, and 250 channels and weighted fusion to eliminate uneven illumination. For example, in a backlit afternoon scene, the original image's face area brightness is only 40 Lux; after compensation, it is increased to the 90-110 Lux range, making facial details clearly visible.

[0110] Next, angle alignment rules are executed: 68 facial landmarks are located using a lightweight facial landmark detection model (PFLD), and the pitch and yaw angles of the current facial pose are calculated. If the angle deviation exceeds the threshold (pitch angle > 15° or yaw angle > 20°), an affine transformation matrix is ​​constructed, and a bilinear interpolation algorithm is used to rotate and correct the image to a standard frontal view. For example, when the user is short, resulting in a camera pitch angle of +18°, the system detects the deviation of the angle between the nose tip and the line connecting the eyes, calculates the rotation matrix, and rotates the entire image counterclockwise by 18° to align the facial pose to a standard frontal view.

[0111] Finally, the occlusion repair rules are implemented: a semantic segmentation model is used to detect occluded areas. If the occlusion area is less than 30% and concentrated in non-critical areas (such as the chin and forehead), a lightweight generative adversarial network (a Pix2Pix model based on MobileNetV2) is used to fill in the occluded areas with pixels. The model input is an occluded facial image and a segmentation mask, and the output is the repaired complete facial image. For example, if a user's temporary mask covers the mouth and nose area, the model learns from a large amount of unoccluded facial data distribution to generate filling content consistent with skin tone and texture, so that the repaired image meets the feature extraction requirements. After triple correction, a preprocessed face image is generated, maintaining a resolution of 1080P, a facial area ratio of >25%, a pose deviation of <5°, and a brightness uniformity of >80%, which is directly passed to step S240 for feature extraction.

[0112] Step S240: Using the preprocessed face image as input, a preset lightweight deep learning model is used to extract multi-dimensional features: the geometric information of facial contours and facial features is obtained through the structural feature extraction module, and the skin details and local texture information are obtained through the texture feature extraction module. After splicing, the features are normalized to generate a fixed-length face feature vector.

[0113] The model comprises two main modules: **Structural Feature Extraction Module:** This module focuses on geometric information, using keypoint and edge detection algorithms to extract structured information such as facial contours, feature positions, and facial proportions, forming feature vectors representing the geometric attributes of the face. **Texture Feature Extraction Module:** This module focuses on detailed information, using depthwise separable convolution and attention mechanisms to extract microscopic features such as skin texture, pore distribution, and local brightness variations, forming feature vectors representing the detailed attributes of the face. **Face Feature Vector:** This is a normalized, fixed-length numerical array (typically 128-dimensional or 256-dimensional) used to uniquely identify a face and serves as the core data carrier for subsequent comparison and authentication.

[0114] The necessary process is described below:

[0115] The system receives a standard face image (uniform size of 112×112 pixels, pose deviation <5°, brightness uniformity >80%) preprocessed in step S230 as input, and calls a preset lightweight deep learning model for multi-dimensional feature extraction. This model adopts a dual-branch parallel architecture, with a total parameter size controlled within 3.2MB and inference time <50ms, meeting the real-time processing requirements of the battery swapping cabinet. During model loading, the pre-trained weight file is read from local encrypted storage, and forward propagation calculations are performed based on the OpenCV and NCNN frameworks, ensuring efficient operation on the embedded platform.

[0116] The structural feature extraction module first locates 68 key coordinate points on the face using a facial landmark detection network (PFLD-68 model), including contour lines, eyebrows, eyes, nose, and mouth. Based on these key points, geometric relationship vectors are calculated, such as the ratio of eye spacing to face width, the vertical distance from the tip of the nose to the corners of the eyes, and the angle between the line connecting the corners of the mouth and the horizontal line, totaling 12 geometric parameters. Simultaneously, the Canny edge detection algorithm is used to extract facial contour edge maps, and these edge maps are encoded into 16-dimensional structural feature sub-vectors through a small convolutional network (3 layers of 3×3 convolutions). Finally, the key point geometric vectors and the edge encoding vectors are concatenated to form a 28-dimensional structural feature representation, used to capture the rigid geometric properties of the face.

[0117] The texture feature extraction module employs a lightweight convolutional neural network (MobileFaceNet architecture) to perform deep encoding on the entire face region. The network first extracts multi-scale feature maps through five layers of depthwise separable convolutions, followed by a channel attention mechanism (SE module) after each convolution to enhance the response of key texture regions. Subsequently, the deep feature maps are unfolded into one-dimensional vectors and compressed into 100-dimensional texture feature sub-vectors through fully connected layers. This vector encodes microscopic details such as skin pore distribution, fine line direction, local pigmentation, and light and shadow transitions, playing a crucial role in facial identification. To prevent overfitting, the network employs a Dropout strategy (dropout rate of 0.3) and data augmentation during training.

[0118] The feature vectors output by the two modules (28-dimensional structural features + 100-dimensional texture features) are concatenated along the channel dimension to form a 128-dimensional original feature vector. The system then performs L2 normalization: normalized vector = original vector / ||original vector||2, mapping the feature vector onto a unit hypersphere to make the vector magnitude 1, eliminating the influence of differences in brightness and contrast between images on the feature amplitude. The normalized 128-dimensional feature vector is the final output, which is invariant to translation, rotation, and scaling, and can be directly used for similarity comparison calculation in step S250. The feature extraction process is completed locally, without transmitting the original image to the cloud, ensuring the security of user privacy data.

[0119] Step S250: Execute the preset dynamic comparison rules: calculate the similarity between the feature vector and the pre-stored authorized face database, dynamically adjust the matching threshold in combination with the collection environment parameters, and output the authentication result.

[0120] The dynamic comparison rule refers to a comparison strategy that adaptively adjusts the matching threshold based on real-time collected environmental parameters (light intensity, camera angle) to maintain the accuracy and pass rate of face recognition in complex environments. Similarity calculation: Calculates the cosine similarity between the current operator's facial feature vector and each feature vector in the pre-stored authorized face database, with a value ranging from -1 to 1. A higher value indicates a higher degree of identity matching. Matching threshold: The key threshold for determining whether identity authentication passes, typically set between 0.70 and 0.85, dynamically fluctuating according to the severity of the environment. Authentication result: Includes a binary judgment result of "pass" or "fail," as well as metadata such as similarity score, matching user ID, and environmental parameters.

[0121] The necessary process is described below:

[0122] The system reads the 128-dimensional facial feature vector generated in step S240 and calculates its similarity with the feature vectors in the pre-stored authorized facial database. Each authorized user in the database corresponds to one feature record. The system uses a cosine similarity algorithm: similarity = (vector A · vector B) / (||vector A||2 × ||vector B||2). To improve comparison efficiency, the system first uses the product quantization index (PQIndex) to quickly filter the database, reducing the candidate set to the top-10 most similar users. Then, it performs precise cosine similarity calculation on these 10 candidates, keeping the total comparison time within 100ms. Simultaneously with the similarity calculation, the system obtains environmental parameters from step S210, including illumination intensity, camera pitch angle, and yaw angle. The dynamic comparison rules adjust the matching threshold in real time based on these parameters: when the light intensity is below 50 Lux (weak light) or above 15000 Lux (strong light), the threshold is lowered by 0.05 to improve the pass rate; when the absolute value of the pitch angle exceeds 15° or the yaw angle exceeds 20°, the threshold is lowered by 0.03 to compensate for attitude deviation. For example, in a low-light environment at night (30 Lux), the base threshold of 0.75 is adjusted to 0.70; in a backlight environment in the afternoon (18000 Lux and pitch angle +18°), the threshold is adjusted to 0.72. The system selects the highest similarity value among all candidate users and compares it with the adjusted dynamic threshold. If the highest similarity is ≥ the dynamic threshold (e.g., 0.70), the authentication is successful, and the output includes the matched user ID, similarity score (e.g., 0.82), and a snapshot of environmental parameters; if the highest similarity is < the dynamic threshold, the authentication fails, and the output is marked with the reason for failure (e.g., "insufficient similarity" or "poor environment"). The authentication result is encapsulated in the form of a structure and passed to step S260 for database update or exception feedback. At the same time, it is cached in local storage for subsequent statistical analysis.

[0123] Step S260: After successful authentication, the optimized feature vector is automatically added to the database according to the preset database update strategy; if authentication fails, the reason for the exception is pushed to the management platform through the preset feedback mechanism.

[0124] Database Update Strategy: This refers to the preset rules by which the system automatically supplements or replaces the extracted feature vectors in the authorized face database after successful authentication. This includes incremental updates, overwrite updates, and quality assessment mechanisms to ensure continuous database optimization. Feedback Mechanism: After authentication failure, the system uses a standardized process to push the cause of the anomaly to the cloud management platform via structured messages, supporting MQTT protocol transmission. The messages include diagnostic information such as failure type, environmental parameters, and device ID. Optimized Feature Vectors: The 128-dimensional facial feature vectors extracted during this authentication process, after quality assessment (e.g., sharpness and pose deviation exceeding the database entry standards), are used to supplement or replace historical features in the database, enabling dynamic updates to the user's facial model.

[0125] The necessary process is described below:

[0126] When step S250 outputs a successful authentication result, the system initiates a database update strategy. First, the extracted feature vectors are evaluated for quality: the cosine similarity between the extracted vector and existing feature vectors in the database is calculated. If the similarity is between 0.75 and 0.85, and the environmental parameters collected this time are superior (e.g., illumination > 100 Lux, pose deviation < 5°), it is considered a high-quality new sample. The system will then perform incremental updates according to preset rules: a new feature record is added to the database for the user ID, labeled with the collection timestamp and environmental tag; if the number of feature records for that user has reached its limit (e.g., 10), the oldest record with the lowest quality score is automatically replaced to maintain a manageable database size. The update operation is completed in local encrypted storage, and the newly added vectors are synchronized to the cloud management platform via a 4G network to achieve data consistency across multiple devices.

[0127] When authentication fails, the system generates a structured exception message through a preset feedback mechanism. The message is encapsulated in JSON format and includes the failure reason code (01-insufficient similarity, 02-poor image quality, 03-face not registered), the highest similarity value collected this time, environmental parameters (light intensity, camera angle), device number, and timestamp.

[0128] For example, a failure due to a similarity score of 0.62 being below the threshold of 0.70 would result in a message content of {"error_code":"01","max_similarity":0.62,"light_level":35,"device_id":"BSS12-001","timestamp":"2025-11-25T14:30:00"}. The message, after being encrypted with AES, is pushed to the cloud management platform in real-time via the MQTT protocol. Upon receiving the message, the platform parses it and stores it in the exception log database. Simultaneously, it triggers an alarm notification (such as an SMS or app push) to the operations and maintenance personnel, prompting them to pay attention to authentication anomalies on specific devices.

[0129] After the database update and feedback push operations are completed, the system encapsulates the authentication result (including success / failure flags, similarity, environmental parameters, and operation time) into a standardized record and writes it to the local operation log queue, awaiting full data synchronization in step S500. All operations involving user biometrics are completed in a TEE (Trusted Execution Environment) to ensure the security and privacy protection of feature vectors during transmission and storage.

[0130] Step S270: Output the user authentication result.

[0131] If the battery connection health score is lower than the first threshold, a connection error command will be generated, including:

[0132] Step S310: Based on the preset dynamic threshold rules, input the real-time environmental parameters of the battery swapping cabinet into the threshold adjustment model and output the dynamic first threshold.

[0133] The dynamic threshold rule refers to a preset rule that automatically adjusts the first threshold based on the real-time environmental conditions (temperature and humidity) of the battery swapping cabinet. This rule is used to eliminate the benchmark drift of connection health determination caused by environmental factors, ensuring the accuracy of anomaly determination under different operating conditions. The threshold adjustment model is a parameterized calculation model established based on historical data analysis. It maps environmental parameters to threshold compensation amounts, typically using linear weighting or piecewise functions, outputting a dynamic threshold negatively correlated with the severity of the environment. The dynamic first threshold is the real-time judgment threshold after environmental compensation. Compared to the fixed threshold, it is appropriately relaxed in harsh environments such as high temperature and high humidity to avoid misjudgments caused by environmental interference and improve system robustness.

[0134] The necessary process is described below:

[0135] The system reads environmental parameters in real time from temperature and humidity sensors deployed inside the battery swapping cabinet. The temperature sensor monitors from 0°C to 50°C, and the humidity sensor monitors from 5% to 75% non-condensing humidity. Both sensors synchronously collect current temperature and humidity data at a frequency of 1Hz. The collected data is smoothed by a moving average filter (window length of 3 points) to form a stable environmental parameter vector (T, H), where T is the temperature value and H is the humidity value.

[0136] Based on preset dynamic threshold rules, the system inputs environmental parameter vectors into the threshold adjustment model. This model employs a piecewise linear function architecture and incorporates the influence coefficients for each environmental interval. For example, the preset rule is: Dynamic First Threshold = Baseline Threshold × (1 - α × (T - 25) / 50 - β × (H - 50) / 100), where the baseline threshold is 70 points, α is the temperature influence coefficient (0.15), and β is the humidity influence coefficient (0.10). When the ambient temperature is 35℃ and the humidity is 60%, the Dynamic First Threshold = 70 × (1 - 0.15 × (35 - 25) / 50 - 0.10 × (60 - 50) / 100) = 70 × (1 - 0.03 - 0.01) = 67.2 points.

[0137] If environmental parameters exceed the normal operating range (e.g., temperature > 50℃ or humidity > 75%), the model automatically triggers the upper limit protection mechanism, locking the dynamic first threshold to the minimum allowable value of 60 points to prevent excessive relaxation of the judgment criteria. The calculated dynamic first threshold is output in floating-point form and passed to step S320 for real-time comparison with the battery connection health score, ensuring that reasonable anomaly detection sensitivity can be maintained even in harsh environments.

[0138] Step S320: Compare the battery connection health score with a dynamic first threshold.

[0139] The necessary process is described below:

[0140] The system reads two floating-point variables from memory: the battery connection health score and the dynamic first threshold. The health score is a percentage value with one decimal place (e.g., 82.5 points), and the dynamic first threshold is a judgment benchmark after environmental compensation (e.g., 67.2 points). The comparison logic uses a simple numerical comparison: if the health score < the dynamic first threshold, the "abnormal" flag (Boolean value False) is output; otherwise, the "normal" flag (Boolean value True) is output.

[0141] The comparison operation is completed within 1ms by the MCU on the main control board of the battery swapping cabinet. The result is encapsulated in the form of a structure, including a comparison result flag, a health score value, a dynamic first threshold, and a timestamp. This structure is then passed to the instruction generation module in step S330 via an internal message queue. If the score is below the threshold, a "trigger anomaly" flag is added to the structure for subsequent triggering of audible and visual alarms and interface guidance. If the score is not below the threshold, a "continue process" flag is added to allow the authentication process to proceed normally. This comparison operation ensures that each step of the authentication has clear quantitative basis, avoiding subjective judgment errors.

[0142] Step S330: If the battery connection health score is lower than the dynamic first threshold, a connection abnormality command is generated.

[0143] The necessary process is described below:

[0144] The system monitors the comparison result flag bit output in step S320 in real time. When an "abnormal" flag is received (i.e., the health score is lower than the dynamic first threshold), the instruction generation module immediately executes the hard-coded conditional branch logic and instantiates a connection abnormal instruction object. This instruction object contains the instruction code (CMD_CONNECTION_FAULT), priority (HIGH), trigger timestamp, associated warehouse number, and current health score value, and is synchronously distributed to the audible and visual alarm module, human-machine interaction module, and log recording module via the CAN bus or internal message queue.

[0145] Upon generation of the command, the system immediately terminates the current battery swapping process, preventing subsequent facial recognition or battery replacement operations, ensuring a forced interruption of the process in an insecure state. Simultaneously, the command activates a 30-second anomaly timer. During the timer's validity period, all user operation requests will be intercepted and a message will be displayed: "Please handle connection error." If the user does not respond within 30 seconds, the system automatically pushes a timeout alarm to the cloud management platform. This mechanism achieves closed-loop control from health status detection to abnormal response, ensuring that connection problems are addressed promptly and explicitly.

[0146] If the command indicates a connection error, an audible and visual alarm will be triggered, and battery connection guidance information will be displayed on the human-machine interface, including:

[0147] Step S410: In response to the connection error command, risk levels are divided according to a preset difference range based on the numerical difference between the battery connection health score and the dynamic first threshold.

[0148] Risk level: This refers to the degree of danger categorized based on the difference between the battery connection health score and the dynamic first threshold. It is typically divided into three levels: low risk, medium risk, and high risk, used to differentiate the intensity of the audible and visual alarms and the complexity of the triggering process. Difference range: A preset numerical range threshold that maps continuous differences to discrete risk levels. For example, a difference ≤ 5 indicates low risk, 5-15 indicates medium risk, and > 15 indicates high risk.

[0149] The necessary process is described below:

[0150] After receiving the connection error command generated in step S330, the system immediately reads the current battery connection health score (e.g., 58.3 points) and the dynamic first threshold output in step S310 (e.g., 67.2 points), and calculates the difference between the two: difference = dynamic first threshold - health score.

[0151] In the example above, the difference is 67.2 - 58.3 = 8.9 points.

[0152] When classifying risk levels based on preset difference ranges, the system uses piecewise function logic: if the difference is ≤ 5 points, it is judged as low risk (risk level code 1), indicating a minor connection problem that may only require the user to unplug and replug the connector; if the difference is between 5 and 15 points, it is judged as medium risk (risk level code 2), indicating a moderate connection problem, requiring the user to check the cleanliness of the connector and the insertion force; if the difference is > 15 points, it is judged as high risk (risk level code 3), indicating a serious connection abnormality, possibly due to physical damage or BMS communication failure, requiring an immediate alarm and prompting the user to replace the battery or contact maintenance. For example, a difference of 8.9 points falls within the 5-15 point range, and the system judges it as medium risk level 2.

[0153] The risk level assessment result is output in structure form, including the level code, difference value, health score, and threshold. It is transmitted via an internal message queue to step S420 to trigger the corresponding audio-visual mode, and simultaneously to step S430 to invoke the matching guidance knowledge base content. This mechanism ensures the quantification and consistency of risk assessment, avoiding subjective judgment bias.

[0154] Step S420: Trigger the corresponding audible and visual alarm mode based on the determined risk level.

[0155] The audible and visual alarm mode includes a multi-level warning scheme preset according to the risk level. This scheme includes configurable parameter combinations such as speaker frequency (e.g., 2Hz / 5Hz / 10Hz), LED strip color (yellow / orange / red), and flashing duty cycle (e.g., 30% / 50% / 70%), used to visually convey the severity of the abnormality to the operator. The risk level mapping table, pre-configured in the battery swapping cabinet firmware, directly maps the risk level codes (1-3) output in step S410 to specific audible and visual control command codes (e.g., MODE_LOW, MODE_MID, MODE_HIGH) for rapid response.

[0156] The necessary process is described below:

[0157] After receiving the risk level code output in step S410, the system immediately queries the risk level mapping table to obtain the corresponding audible and visual alarm mode parameters. For low risk (code 1), the prompt mode is triggered: the speaker emits a single short beep (frequency 2Hz, lasting 0.5 seconds), the LED light strip on the top of the cabinet displays a solid yellow light, and the warning area on the screen displays a yellow warning icon, prompting the user "Minor connection abnormality, please check".

[0158] For medium risk (code 2), the warning mode is triggered: the speaker emits a continuous medium-frequency beep (frequency 5Hz, duty cycle 50%, lasting 10 seconds), the LED strip flashes orange (cycle 1 second, 700ms on and 300ms off), and the screen warning area is highlighted in orange to display "Moderate connection failure, please reinsert the battery".

[0159] For high-risk (code 3), emergency mode is triggered: the speaker emits a high-frequency urgent alarm (frequency 10Hz, duty cycle 70%, lasting for 30 seconds or until the user responds), the LED strip flashes red quickly (cycle 0.5 seconds, on for 350ms and off for 150ms), the screen warning area flashes red and displays "Severe connection abnormality, do not operate, contact maintenance", and at the same time, an emergency alarm is sent to the management platform via the 4G network.

[0160] The sound and light control commands directly drive the speaker driver circuit and LED controller via the GPIO interface, with a response delay of <100ms. If the user does not respond within 10 seconds, the system automatically repeats the current mode, repeating a maximum of 3 times before entering a forced standby state to prevent continuous disturbance.

[0161] Step S430: Based on the risk level and numerical difference, call the preset guidance knowledge base and perform quantification and clarification of the main contribution parameters: According to the preset contribution parameter judgment rules, clarify the main contribution parameters that lead to low scores.

[0162] The system includes a pre-defined knowledge base: a structured database stored in the local memory of the battery swapping cabinet, containing fault mode mapping tables for various connection parameters (contact resistance, voltage drop, communication quality), a diagnostic text library, and user profile matching rules, used to quickly locate the root cause of problems and generate targeted guidance content. The system also quantifies and clarifies key contributing parameters: based on a weighted calculation of feature weights and feature value deviations, it quantifies the contribution of each parameter to low scores, accurately identifying the core parameters causing connection anomalies. Finally, the system uses a pre-defined judgment logic, sorting parameters by contribution score from highest to lowest. If the highest-scoring parameter exceeds the second-highest score by more than 20% or has an absolute value >30 points, it is determined to be a "key contributing parameter."

[0163] The necessary process is described below:

[0164] The system receives the risk level, health score and dynamic first threshold value difference (e.g. 8.9 points) output in step S410, and reads the compensated feature values ​​(contact resistance change rate, voltage drop fluctuation amplitude, communication error rate) output in step S120 and the feature weight vector determined in step S134.

[0165] First, the contribution score of each parameter is calculated using the weighted deviation formula: Contribution = Normalized eigenvalue × Feature weight × 100. The normalized eigenvalue is calculated based on the percentile of the parameter among all historical samples. For example, if the voltage drop fluctuation amplitude in a certain test is 8mV, which is in the 75th percentile of historical samples (i.e., worse than the 75th percentile of historical values), then its normalized eigenvalue is 0.75; if the weight of the parameter is 0.46, then the contribution score = 0.75 × 0.46 × 100 = 34.5 points. The contact resistance change rate and communication bit error rate are calculated in the same way.

[0166] After obtaining the three contribution scores, the system executes the judgment rule: the scores are sorted from highest to lowest. If the score of the first place exceeds the score of the second place by more than 20% (e.g., 34.5 vs 25.0, a difference of 38%) or the absolute value of the first place is greater than 30 points, then the parameter is judged as a "major contributing parameter". For example, if the calculation results are ranked as follows: voltage drop fluctuation amplitude 34.5 points, contact resistance change rate 25.0 points, and communication error rate 18.2 points, then the voltage drop fluctuation amplitude is judged as a major contributing parameter.

[0167] The risk level and numerical difference are used to adjust the judgment sensitivity: at a high risk level, the judgment threshold is tightened to 15% (making it easier to determine the main contributing parameters). When the difference is >10 points, at least one main contributing parameter is forcibly judged to ensure that serious problems can be accurately located. The final output is structured data containing the name of the main contributing parameter, the contribution score, and the associated fault mode code, which is passed to step S440 to retrieve targeted guidance content.

[0168] Step S440: Based on the preset mapping relationship of the main contribution parameters, retrieve the corresponding guidance text from the preset guidance knowledge base; at the same time, adjust the guidance format according to the preset user profile tags.

[0169] The system includes: a preset mapping relationship: a parameter lookup table stored in the index layer of the guidance knowledge base, which directly maps the main contributing parameters determined in step S430 (such as "voltage drop fluctuation amplitude") to the corresponding fault mode code (such as FAULT_VOLT_DROP) and guidance text ID (such as GUIDE_003), enabling rapid and accurate location. A preset guidance knowledge base: a structured knowledge base deployed locally on the battery swapping cabinet in the form of an SQLite database, containing fault phenomenon descriptions, step-by-step troubleshooting guidance, links to illustrated tutorials, and user profile adaptation rules, supporting offline access and incremental updates in the cloud. User profile tags: behavioral feature tags automatically generated based on users' historical operation data, including dimensions such as age group (elderly / middle-aged / young), proficiency (novice / expert), and fault frequency (low-frequency / high-frequency), used to dynamically adjust the presentation format of the guidance content.

[0170] The necessary process is described as follows: The system queries the knowledge base based on the main contribution parameters to obtain the corresponding guidance text for the fault. At the same time, the system reads the current user's profile tags (such as elderly user, novice user, or experienced user) and automatically adjusts the guidance format: for elderly users, the font size is increased and voice broadcast is enabled; for novice users, detailed explanations and animated demonstrations are added; and for experienced users, the core instructions are simplified.

[0171] The revised guidance content is presented through a human-computer interaction interface. The warning area displays the risk level and alarm status, while the guidance area displays adapted text, animation links, and other information, forming a layered and differentiated user guidance system.

[0172] In step S450, the processing status is synchronously presented through the regional collaborative display module of the human-computer interaction interface. The warning area dynamically displays the current risk level and alarm mode, while the guidance area adaptively displays differentiated guidance content that matches the user profile.

[0173] The segmented collaborative display module divides the 13.3-inch human-machine interface of the battery swapping cabinet into multiple functional areas, including a warning area and a guidance area. This display control unit supports independent rendering and synchronous updates for each area, ensuring clear information presentation and highlighting key points. The warning area, a fixed area at the top of the interface, dynamically displays the current risk level and audible / visual alarm mode using color (yellow / orange / red) and flashing status, providing clear and immediate abnormal warnings. The guidance area, the main area of ​​the interface, adaptively displays differentiated guidance content based on user profiles, supporting text, animation, and voice formats for precise fault guidance.

[0174] The necessary process is described below:

[0175] After receiving the JSON-formatted display data packet generated in step S440, the system drives the touchscreen to perform zoned rendering. The warning zone immediately updates its background color (e.g., orange for medium risk), synchronously displays the risk level text ("Moderate Connection Problem") and icon (flashing orange LED), and refreshes the alarm mode status in real time (e.g., "Buzzer 5Hz") to ensure that the warning information is consistent with the audible and visual alarm hardware status.

[0176] The guidance area adapts its content based on user profile tags: in elderly user mode, the font size is increased to 24 points, key steps are displayed with high contrast in orange background and white text, and a voice broadcast button is automatically triggered; in novice user mode, each guide is followed by detailed explanatory text and an embedded link to a 3D animation demonstration; in experienced user mode, only fault codes and core instructions are displayed, while redundant explanations are hidden. The content in both areas is updated synchronously using a unified timestamp, and the interface automatically displays a "Reassessing" status after the user completes the guided operation, achieving a seamless transition in the process.

[0177] Step S460: After the user guidance is executed, the connection status re-evaluation mechanism is started: the battery connection parameters are re-collected and a new health score is generated. The effectiveness of the guidance is verified based on the preset validity judgment rule. The judgment rule is: if the new score reaches more than the preset percentage of the dynamic first threshold, the guidance is deemed effective; otherwise, the guidance is deemed invalid.

[0178] The connection status reassessment mechanism involves automatically re-collecting battery connection parameters and calculating a health score after the user completes the operation according to the instructions. This is a closed-loop detection process to verify whether the user's operation effectively resolved the connection problem. A preset validity judgment rule is used to determine whether the user guidance was successful; it is typically set at 85%-90% of a dynamic first threshold. If the new score reaches this standard, the guidance is considered effective.

[0179] The necessary process is described below:

[0180] The system re-collects the physical contact parameters and electrical conduction parameters of the battery connector through the electrical signal detection component, and repeats steps S110 to S140 to generate a new health score. The new score is compared with the dynamic first threshold output in step S310, and a ratio is calculated: Ratio = New Score / Dynamic First Threshold. If the ratio is ≥ a preset percentage (e.g., 85%), the guidance is deemed effective, and the system resumes the normal battery swapping process; if the ratio is < 85%, the guidance is deemed ineffective, the abnormal state is maintained, and the alarm level is escalated.

[0181] For example, if the dynamic first threshold is 67.2 points, and the user's new score is 61.5 points after re-inserting the battery, the proportion is 91.5% (≥85%), the guidance is deemed effective, the audible and visual alarm is automatically deactivated, and the user is allowed to enter the face authentication process; if the new score is 55 points, the proportion is 81.8% (<85%), the guidance is deemed invalid, a high-risk alarm is triggered, and a message is displayed: "Please replace the battery compartment or contact maintenance."

[0182] Step S470: Execute a preset hierarchical processing strategy based on the verification results.

[0183] The pre-defined hierarchical processing strategy based on the verification results includes:

[0184] Step S471: Using the verification result of the connection state re-evaluation mechanism as input, execute the preset hierarchical processing strategy according to different scenarios.

[0185] Step S472: When the guidance is determined to be effective, a preset weight optimization mechanism is executed to dynamically adjust the weight value of the guidance scheme in the knowledge base based on the verification results.

[0186] For details on the specific process, please refer to steps S4721 to S4725, which will not be elaborated here.

[0187] Step S473: When the guidance is determined to be invalid, a preset multi-objective decision-making mechanism is activated. With the resolution rate, processing time and user satisfaction as optimization objectives, a mapping relationship between risk level and optimal display level is established, and the guidance level is upgraded according to the preset upgrade path.

[0188] The multi-objective decision-making mechanism is a decision-making algorithm that comprehensively considers multiple optimization objectives (resolution rate, processing time, user satisfaction), selecting the optimal processing solution by weighing the priorities of different objectives. Optimal display level: Based on the risk level and optimization objectives, the most suitable display level for the current situation is determined, such as upgrading from basic guidance to advanced guidance. Predefined upgrade path: Predefined guidance level upgrade rules that gradually provide more detailed guidance content or switch to a higher priority processing solution based on the number of ineffective guidance attempts or the increase in risk level.

[0189] When step S460 determines that the guidance is invalid, the system activates a preset multi-objective decision-making mechanism, the specific process of which is as follows:

[0190] 1. Establish mapping relationship:

[0191] The system establishes a mapping relationship between the current risk level (low, medium, high) and the optimal display level based on the current risk level and historical data. The mapping relationship is as follows:

[0192] Low risk: Basic guidance (resolution rate first, processing time second).

[0193] Medium risk: Intermediate guidance (emphasis on both resolution rate and user satisfaction).

[0194] High risk: Advanced guidance (priority on user satisfaction, secondary on resolution rate).

[0195] Evaluation and optimization objectives:

[0196] The system evaluates the performance of the current guidance plan in three dimensions: resolution rate, processing time, and user satisfaction. Define the following optimization objective weights: resolution rate weight w1; processing time weight w2; user satisfaction weight w3;

[0197] Weight allocation is adjusted according to the risk level:

[0198] Low risk: w1 = 0.6, w2 = 0.3, w3 = 0.1.

[0199] Medium risk: w1 = 0.4, w2 = 0.3, w3 = 0.3.

[0200] High risk: w1 = 0.3, w2 = 0.3, w3 = 0.4.

[0201] Calculate the comprehensive score S:

[0202] S = w1 × resolution rate + w2 × processing time + w3 × user satisfaction.

[0203] Among them, the resolution rate, processing time, and user satisfaction are all normalized to the interval [0, 1].

[0204] Select the optimal solution system:

[0205] Select the optimal display level according to the comprehensive score S. Preset the comprehensive score threshold T (for example, T = 0.7). If the score S of the current guidance plan is < T, trigger the upgrade path and raise the guidance level to a higher level.

[0206] 4. Execute the upgrade path system:

[0207] Adjust the guidance content according to the preset upgrade path.

[0208] The upgrade path is as follows:

[0209] Basic guidance → Intermediate guidance: Add detailed troubleshooting steps and provide video tutorial links.

[0210] Intermediate guidance → Advanced guidance: Enable the remote assistance function and provide expert contact information.

[0211] 5. Update the guidance content system: Regenerate the guidance content according to the new display level and display it to the user through the human-computer interaction interface. At the same time, the system records the execution situation of the upgrade path for subsequent optimization and analysis.

[0212] Step S474: Execute the preset closed-loop optimization process, and dynamically adjust the knowledge base strategy based on the comparison of historical data and real-time verification results through a preset threshold mechanism.

[0213] When executing the pre-defined closed-loop optimization process, the system first integrates historical operation and maintenance data with real-time verification results, and combines environmental parameters and user behavior data to construct a multi-dimensional feature dataset. For example, the dataset may include historical health scores, ambient temperature, humidity, user operating habits, etc. This data will be used to analyze the correlation between the effectiveness of the guidance program and the environment and user behavior.

[0214] Next, the system employs time series analysis algorithms (such as the ARIMA model) to establish a dynamic baseline model of key performance indicators (such as resolution rate, processing time, and user satisfaction) as they change with environmental factors. This model can predict the expected performance of the guided solution under different environmental conditions. For example, the model predicts that the resolution rate of a certain guided solution may decrease by 5% under high-temperature conditions.

[0215] Then, a predictive model for the effectiveness of the guidance program is constructed using the random forest algorithm. This model predicts the effectiveness of the guidance program based on a multidimensional feature dataset. The prediction results are compared with real-time validation results to evaluate the accuracy of the model and the actual effectiveness of the guidance program. For example, if the random forest model predicts a 90% resolution rate for a guidance program among novice users, while the actual resolution rate is 85%, this indicates that the prediction model may have overestimated the effectiveness of the program.

[0216] A dual-threshold decision-making mechanism is implemented: when the actual effectiveness of a solution is consistently higher than the dynamic baseline and exceeds a preset upper threshold (e.g., 10%), its weight is increased; when it is lower than the dynamic baseline and falls below a preset lower threshold (e.g., 5%), a downgrade process is initiated. For example, if a pilot solution exceeds the preset upper threshold in three consecutive iterations, the system will automatically increase the weight of that solution.

[0217] Based on the comparison results, the guidance knowledge base is incrementally updated, and the effectiveness of the strategy is verified through A / B testing. For example, the updated guidance scheme is A / B tested to compare the performance of the new and old schemes in similar environments, ensuring that the new scheme is at least as effective as the old scheme.

[0218] Finally, a monitoring mechanism is established to track key metrics. When these metrics exceed preset tolerances, an optimization rollback is triggered, and threshold parameters are recalibrated. (For example, if user satisfaction with a particular onboarding strategy continues to decline, the system will automatically reduce the weight of that strategy and reassess relevant parameters.) This closed-loop optimization process ensures that the knowledge base strategy is always adapted to the latest operational data and user feedback, achieving continuous performance improvement.

[0219] Step S475: Through a preset distributed learning framework, the optimization experience of each terminal is aggregated, the global knowledge base is updated, and synchronized to each battery swapping cabinet terminal.

[0220] Among them, the distributed learning framework is a machine learning architecture based on multi-terminal collaboration. It achieves dynamic updates and synchronization of the global knowledge base by aggregating the optimization experience of each terminal (battery swapping cabinet).

[0221] Global Knowledge Base: A central knowledge base stored on a cloud server, containing optimization experience, fault mode mapping, and guidance schemes for all battery swapping cabinet terminals, used for unified management and distribution of optimization strategies.

[0222] Terminal optimization experience: The improvement strategies and guidance scheme update records generated by each battery swapping terminal after executing the optimization process locally are used to update the global knowledge base.

[0223] The specific process is as follows: The system collects optimization experience from each battery swapping cabinet terminal through a pre-set distributed learning framework, including improvement strategies and guidance scheme update records. This data is uploaded to a global knowledge base in the cloud. The cloud server integrates and analyzes the collected experience, updating the fault mode mapping and guidance scheme in the global knowledge base. After the update is completed, the cloud synchronizes the changes to all battery swapping cabinet terminals via the network, ensuring that each terminal can apply the latest optimization strategies, improving the overall system performance and user experience.

[0224] When the guidance is deemed effective, a preset weight optimization mechanism is executed to dynamically adjust the weight value of the guidance scheme in the knowledge base based on the verification results, including:

[0225] Step S4721: Taking the verification result of the connection state re-evaluation mechanism as input, when the boot is determined to be effective, extract the successful application data packet of the current boot scheme, which includes risk level, main contribution parameters, environmental characteristics and user profile tags.

[0226] Successful application data package: When the bootstrapping is successful, the system extracts a set of key data, including risk level, main contribution parameters, environmental characteristics and user profile tags, which are used as input for the subsequent weight optimization model.

[0227] The general process is as follows: When the guidance is deemed valid, the system extracts a set of key information to form a successful application data package. This information includes: risk level (e.g., low, medium, high risk), main contributing parameters (e.g., voltage drop fluctuation amplitude, contact resistance change rate), environmental characteristics (e.g., temperature, humidity), and user profile tags (e.g., elderly user, novice user). This data is obtained through real-time monitoring and recording during the guidance process. It serves as input to the subsequent weight optimization model, used to construct the state space and action space. This process utilizes data encapsulation technology to integrate different types of data into a structured data package, ensuring the integrity and usability of the information and providing accurate input for subsequent reinforcement learning algorithms, thereby achieving dynamic optimization of the guidance scheme weights.

[0228] Step S4722: Construct a weight optimization model based on a preset reinforcement learning framework. Construct a state space with risk level, main contribution parameters and environmental features, and use the guidance scheme as the action space. Iterate and update the Q-value table of each guidance scheme through the Bellman equation of the Q-learning algorithm.

[0229] The detailed process is as follows:

[0230] 1. State Space Construction: The system combines the risk level (low, medium, high), key contributing parameters (such as voltage drop fluctuation amplitude and contact resistance change rate), and environmental characteristics (temperature, humidity) into a state vector. For example, the state vector can be represented as:

[0231] Status = (Risk Level, Main Contributing Parameters, Temperature, Humidity);

[0232] Specifically, the numerical value is: Status = (Medium risk, voltage drop fluctuation amplitude, 30°C, 60%);

[0233] 2. Action Space Definition: The system treats all possible bootstrapping schemes as options in the action space. Each bootstrapping scheme corresponds to a specific method for resolving connectivity issues, for example:

[0234] Action 1: Reinsert the battery;

[0235] Action 2: Clean the battery connector;

[0236] Action 3: Replace the battery compartment;

[0237] 3. Q-value table initialization: The system initializes the Q-value table, which is a two-dimensional array where rows correspond to the state space and columns correspond to the action space. The initial Q-value can be set to 0 or a small random value.

[0238] 4. Q-value table update: The Q-value table is iteratively updated using the Bellman equation through the Q-learning algorithm. After each successful bootstrapping, the corresponding Q-value is updated based on the current state and the bootstrapping strategy adopted. The specific update formula is as follows: .

[0239] Where Q(s,a) is the Q-value of taking action a in the current state s. α is the learning rate (0 < α ≤ 1), controlling the degree to which new information covers old information. r is the reward signal, a positive reward (e.g., +1) for successful guidance and a negative reward (e.g., -1) for failure. γ is the discount factor (0 ≤ γ < 1), controlling the degree of discounting future rewards. s′ is the new state after executing action a. It is the maximum Q value of all possible actions under the new state s′.

[0240] 5. Weight Adjustment: Calculate new weight values ​​for each bootstrapping scheme based on the updated Q-value table. A higher weight value indicates a more effective bootstrapping scheme in the current state.

[0241] Step S4723: Based on the updated Q-value table, calculate the new weight values ​​of each guidance scheme using a preset weight mapping function. Among them, the guidance scheme corresponding to the user profile tag with high matching degree receives a higher weight increase.

[0242] The specific process is as follows: First, the system analyzes the Q-value table to identify the performance of each guidance program under different states. The Q-value table contains the expected return of each guidance program in a specific state, providing a quantitative basis for weight adjustment. Next, the system evaluates the matching degree between user profile tags and guidance programs. This step identifies which guidance programs are most suitable for specific user groups, such as novice users or elderly users, by analyzing user behavior data and preferences. Then, the system uses a preset weight mapping function to calculate a new weight value for each guidance program. This function comprehensively considers the magnitude of the Q-value and the matching degree of the user profile, ensuring that guidance programs that perform well in practical applications and are highly matched with user needs receive higher weights. The design of the weight mapping function may involve complex algorithms, such as neural networks, that can process and integrate data from different dimensions.

[0243] Step S4724: Establish a weight adjustment traceability mechanism to record the decision path and effect data of each weight optimization. When the cumulative number of effective times of the same guidance scheme reaches a preset threshold within a preset period, its priority in the knowledge base will be automatically increased by one level.

[0244] The implementation steps of the weight adjustment traceability mechanism are as follows:

[0245] 1. Record Decision Path: The system records the decision path in detail for each weight optimization. This includes which guidance schemes were selected for weight adjustment, the magnitude of the adjustment, the Q-value changes used as a basis, and relevant user profile tags.

[0246] 2. Performance data collection: At the same time, the system collects performance data after weight adjustment, such as the success rate of the guidance plan in subsequent applications, user satisfaction feedback, processing time and other key performance indicators.

[0247] 3. Cumulative Valid Times: For each onboarding solution, the system tracks its cumulative valid times over a preset period (e.g., a quarter or a year). This involves analyzing user feedback after each onboarding session to confirm whether the onboarding truly resolved the user's problem.

[0248] 4. Preset Threshold Setting: The system sets a preset threshold for the number of effective attempts for each bootstrapping scheme. This threshold is based on historical data analysis and business needs, reflecting the minimum number of successful cases required for a bootstrapping scheme to be considered effective and worthy of increased weight.

[0249] 5. Automatic Priority Boosting: When a guidance solution's cumulative effective usage reaches or exceeds a preset threshold within a preset period, the system automatically upgrades the solution's priority in the knowledge base by one level. This indicates that the solution performs exceptionally well in solving a specific problem and should be used more frequently in future guidance sessions.

[0250] Step S4725: The stability of the weights of each guidance scheme is monitored by a preset weight convergence detection algorithm. When the weight fluctuation coefficient is detected to be lower than the preset threshold, it is determined that the weight optimization has reached the convergence state, a weight optimization completion report is generated and synchronized to the knowledge base version management system.

[0251] Among them, the weight convergence detection algorithm is an algorithm used to monitor and evaluate the stability of the weights in the guidance scheme. It determines whether the optimization process has reached a stable state by analyzing the trend of weight changes.

[0252] Specifically, the system employs a weight convergence detection algorithm, which quantifies weight fluctuation by calculating the percentage difference between two consecutive weight updates. Specifically, the algorithm calculates the weight fluctuation coefficient using the following formula:

[0253] Weight fluctuation coefficient = (new weight - old weight) ÷ old weight × 100%. This fluctuation coefficient reflects the magnitude of weight adjustment. If the value is lower than the preset threshold (e.g., 5%), the weight is considered to be stable enough, indicating that the optimization process has converged.

[0254] To ensure the stability of the weights, the system generates a weight optimization completion report, which details each iteration, weight changes, user feedback, and evaluation of the guiding effect during the optimization process. This data provides a transparent and traceable record for weight optimization.

[0255] Subsequently, the system utilizes the synchronization mechanism within the distributed learning framework to push the updated weight information to the cloud-based knowledge base version management system. This synchronization process employs incremental update technology, transmitting only the changed data to reduce network bandwidth consumption and improve synchronization efficiency.

[0256] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A management method of an intelligent battery replacement cabinet, characterized in that, The application comprises the following steps: In response to the electrical signal generated by the battery inserted into the battery swap cabinet, the adaptability of the battery is authenticated by the electrical signal detection component, and the battery connection health score is generated based on the physical contact state of the battery connector and the electrical signal conduction quality; When the battery adaptability authentication is passed and the battery connection health score is not lower than the preset abnormal threshold, the face authentication process is triggered, the operator's face image is captured by the monitoring camera of the battery swap cabinet, and the face image is compared with the pre-stored authorized face database to obtain the user identity authentication result; Based on the battery connection health score and the user identity authentication result, the following decisions are made: if the user identity authentication is successful, a battery swap authorization instruction is generated; If the user identity authentication fails but the battery connection health score is not lower than the first threshold, an auxiliary authentication instruction is generated; If the battery connection health score is lower than the first threshold, a connection abnormal instruction is generated; According to the instructions generated by the decision, the corresponding operation is performed: if it is a battery swap authorization instruction, the battery swap cabinet is controlled to perform the battery replacement operation; if it is an auxiliary authentication instruction, the user is guided to perform secondary authentication through the human-computer interaction interface; if it is a connection abnormal instruction, the sound and light alarm component is triggered, and the battery connection guidance information is displayed on the human-computer interaction interface; After the operation is completed, the whole process data of this battery swap is automatically recorded and synchronized to the cloud management platform; Based on the physical contact state of the battery connector and the electrical signal conduction quality, the battery connection health score is generated, which comprises the following steps: Real-time collection of physical contact parameters and electrical signal conduction parameters of the battery connector, and data preprocessing, wherein the physical contact parameters include contact resistance, plug-in force, connection stability, and the electrical signal conduction parameters include voltage drop, current stability and communication quality; Based on the temperature sensor and the humidity sensor, the environmental parameters are obtained, the environmental compensation model is established, the pre-processed physical contact parameters and electrical signal conduction parameters are dynamically corrected, and the compensated data is output. At the same time, based on the environmental parameters and the preset influence coefficient matrix in the environmental compensation model, the environmental compensation factor is calculated; From the compensated data, the key features of contact resistance change rate, voltage drop fluctuation amplitude and communication error rate are extracted, the entropy weight method is used to determine the weight of each feature and weighted fusion is performed to form a comprehensive evaluation index; Based on the preset health score model, the comprehensive evaluation index is taken as the core input, and the environmental compensation factor is used to generate the battery connection health score; The entropy weight method is used to determine the weight of each feature and weighted fusion is performed to form a comprehensive evaluation index, which comprises the following steps: From the compensated data, the key features of contact resistance change rate, voltage drop fluctuation amplitude and communication error rate are extracted, the distribution frequency of each feature in different value intervals is counted according to sample batches, and a feature value distribution table is formed; Based on the preset entropy weight calculation rule, the normalized proportion of each feature in different samples is calculated based on the feature value distribution table, and the entropy value of each feature is output through the entropy algorithm; Based on the entropy value of each feature, its discriminability is directly evaluated and a discriminability ranking is formed; According to the discriminability ranking result, the weight proportion of contact resistance change rate, voltage drop fluctuation amplitude and communication error rate is determined through the preset weight normalization rule; The compensated feature data is integrated according to a preset weighted fusion rule to generate a comprehensive evaluation index. 2.The management method of the intelligent battery replacement cabinet according to claim 1, characterized in that, The face image of the operator is captured by the monitoring camera of the battery swap cabinet, and is compared with the pre-stored authorized face database to obtain a user identity authentication result, including: The face image of the operator is captured in real time by the monitoring camera built in the battery swap cabinet, and the environmental light intensity, camera pitch angle and yaw angle parameters at the time of collection are recorded synchronously; Based on a preset image quality rule, qualified images are screened, a preset definition score threshold is used to calculate the image edge sharpness by a Laplacian variance algorithm, and a score lower than the preset threshold is judged as a blurred image, a preset occlusion area ratio threshold is used to detect the face region by a semantic segmentation model, and an occlusion area exceeding the preset threshold is judged as an excessive occlusion image, and the blurred or excessive occlusion image triggers a re-collection mechanism to adjust the camera to a preset angle range and then re-collects; A preset environmental self-adaptive correction is performed on the qualified image: the backlight / weak light interference is eliminated by a light compensation rule, the face is aligned to a front face orientation based on an angle alignment rule, and the temporary occlusion area is filled by an occlusion repair rule to generate a pre-processed face image; The pre-processed face image is input, and a preset lightweight deep learning model is used to extract multi-dimensional features: geometric information of facial contours and feature positions is obtained by a structure feature extraction module, and skin details and local texture information are obtained by a texture feature extraction module, and a fixed-length face feature vector is generated after normalization processing; A preset dynamic comparison rule is executed: the feature vector is compared with the pre-stored authorized face database for similarity calculation, the matching threshold is dynamically adjusted combined with the collection environment parameters, and an authentication result is output; After authentication succeeds, the optimized feature vector of this time is automatically supplemented into the database according to a preset database updating strategy; when authentication fails, the abnormal reason is pushed to the management platform through a preset feedback mechanism; An output user identity authentication result is output. 3.The management method of the intelligent battery replacement cabinet according to claim 1, characterized in that, If the battery connection health score is lower than the first threshold, a connection abnormality instruction is generated, including: Based on a preset dynamic threshold rule, the real-time environmental parameters of the battery swap cabinet are input into a threshold adjustment model to output a dynamic first threshold; The battery connection health score is compared with the dynamic first threshold; If the battery connection health score is lower than the dynamic first threshold, a connection abnormality instruction is generated. 4.The management method of the intelligent battery replacement cabinet according to claim 3, characterized in that, If it is a connection abnormality instruction, a sound and light alarm component is triggered, and battery connection guidance information is displayed on a human-computer interaction interface, including: In response to the connection abnormality instruction, based on the numerical difference between the battery connection health score and the dynamic first threshold, the risk level is divided according to a preset difference interval; According to the determined risk level, the corresponding sound and light alarm mode is triggered; Based on the risk level and the numerical difference, the preset guidance knowledge base is called and the main contribution parameter is quantified and clarified: according to the preset contribution parameter judgment rule, the main contribution parameter causing the low score is clarified; According to the preset mapping relationship of the main contribution parameter, the corresponding guidance text is called from the preset guidance knowledge base; at the same time, the guidance form is adjusted according to the preset user portrait label. The processing state is synchronously presented by a sub-region cooperative display module of the human-computer interaction interface, wherein the current risk level and the alarm mode are dynamically displayed in the early warning area, and the difference content matched with the user portrait is adaptively displayed in the guide area. After the user guidance is executed, a connection state reevaluation mechanism is started: the battery connection parameters are reacquired, a new connection health score is generated, the guidance effectiveness is verified based on a preset effectiveness determination rule, and the determination rule is: if the new score reaches a preset percentage of the dynamic first threshold, the guidance is determined to be effective; otherwise, the guidance is determined to be invalid. A preset hierarchical processing strategy is executed based on the verification result. 5.The management method of the intelligent battery replacement cabinet according to claim 4, characterized in that, The preset hierarchical processing strategy based on the verification result comprises: Taking the verification result of the connection state reevaluation mechanism as input, the preset hierarchical processing strategy is executed according to different situations; When the guidance is determined to be effective, a preset weight optimization mechanism is executed, and the weight value of the guidance scheme in the knowledge base is dynamically adjusted based on the verification result; When the guidance is determined to be invalid, a preset multi-objective decision mechanism is started to solve the optimization target of the rate, processing time and user satisfaction, a mapping relationship between the risk level and the optimal display level is established, and the guidance level is improved according to a preset upgrade path; A preset closed-loop optimization process is executed, the knowledge base strategy is dynamically adjusted through a preset threshold mechanism based on the comparison between the historical data and the real-time verification result; Through a preset distributed learning framework, the optimization experience of each terminal is aggregated, the global knowledge base is updated and synchronized to each battery swap cabinet terminal. 6.The management method of the intelligent battery replacement cabinet according to claim 5, characterized in that, When the guidance is determined to be effective, the preset weight optimization mechanism is executed, and the weight value of the guidance scheme in the knowledge base is dynamically adjusted based on the verification result, which comprises: Taking the verification result of the connection state reevaluation mechanism as input, when the guidance is determined to be effective, the success application data packet of the current guidance scheme is extracted, including the risk level, the main contribution parameter, the environmental feature and the user portrait label; A weight optimization model is built based on a preset reinforcement learning framework, a state space is built with the risk level, the main contribution parameter and the environmental feature, an action space is built with the guidance scheme, and the Q value table of each guidance scheme is iteratively updated through the Bellman equation of the Q-learning algorithm; According to the updated Q value table, a preset weight mapping function is used to calculate the new weight value of each guidance scheme, wherein the guidance scheme corresponding to the user portrait label with high matching degree obtains a higher weight improvement amplitude; A weight adjustment tracing mechanism is established to record the decision path and effect data of each weight optimization, and when the cumulative effective number of the same guidance scheme reaches a preset threshold within a preset period, the priority of the guidance scheme in the knowledge base is automatically improved by one level; Through a preset weight convergence detection algorithm, the stability of the weight of each guidance scheme is monitored, when the weight fluctuation coefficient is detected to be lower than a preset threshold, it is determined that the weight optimization reaches a convergence state, a weight optimization completion report is generated and synchronized to a knowledge base version management system.

7. A management system of an intelligent battery replacement cabinet, characterized in that, The program can be loaded and executed by the processor to implement the management method of the intelligent battery swap cabinet according to any one of claims 1 to 6.

Images