An intelligent monitoring method and system for an intelligent turnover cabinet and a storage medium

Abstract

This application relates to the field of intelligent monitoring technology, and discloses an intelligent monitoring method, system, and storage medium for intelligent turnover cabinets. The method includes: obtaining storage location status data through sensor cross-verification; edge nodes receiving data and storing it in a cache when communication is interrupted to obtain a distributed dataset; combining biometrics to determine permissions and extracting temporal relationships to obtain behavioral feature sequences; obtaining spatial distribution data through multi-frequency phase difference measurement and triangulation; and extracting temporal features and inputting them into a long short-term memory network to predict risky assets. This application solves the problems of low monitoring reliability, weak system fault tolerance, lack of proactive early warning, coarse positioning granularity, and insufficient predictive ability in existing technologies, realizing an upgrade of the intelligent turnover cabinet monitoring system from passive response to proactive prediction.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology, and in particular to an intelligent monitoring method, system and storage medium for an intelligent turnover cabinet. Background Technology

[0002] Intelligent storage lockers, as crucial equipment for electricity metering asset management, typically employ sensors to monitor storage location status and record asset entry and exit information. Current intelligent storage lockers primarily rely on single-type sensors, such as photoelectric sensors or barcode scanners, to identify assets. A centralized data acquisition architecture uploads monitoring data to a management platform, and alerts are issued for overdue assets based on preset rules. These systems play a fundamental role in the daily management of electricity metering assets.

[0003] However, existing technologies have the following shortcomings: single sensors are prone to misjudgment due to environmental interference or equipment aging; when a photoelectric sensor detects that an asset is in place but the barcode scanner cannot read it, it is impossible to determine whether it is a sensor malfunction or an asset anomaly; centralized architecture relies on a master control node, and once the master control node fails or the network is interrupted, the entire monitoring system will stop working and data will be lost; existing systems can only passively record after assets enter and leave the warehouse, and cannot intercept unauthorized access attempts in advance, nor can they identify violations such as operators not picking up goods according to instructions; for assets stored in boxes, existing technologies can only identify box-level information and cannot perceive changes in the position of individual meters within the box; asset management relies on human experience and lacks the ability to predict the risk of asset verification period expiration in advance, and problems are often only discovered after the assets have expired.

[0004] Further analysis revealed several issues: First, the lack of a sensor cross-verification mechanism prevented the system from automatically switching to a backup identification mode when a sensor malfunctioned, leading to decreased data reliability. Second, the absence of a distributed fault-tolerance mechanism prevented effective caching and synchronization of monitoring data generated during network anomalies, causing data integrity problems. Third, the lack of proactive behavior prediction methods meant the system could only issue alerts after violations occurred, failing to prevent them beforehand. Fourth, the lack of precise asset location technology within the storage compartments prevented managers from monitoring single-table layout changes during full-scale storage. Fifth, the lack of an intelligent prediction model integrating business rules prevented the system from proactively identifying overdue inventory risks and optimizing inventory turnover. These problems collectively hampered the reliability, real-time performance, and intelligence level of the intelligent turnover cabinet monitoring system. Summary of the Invention

[0005] This application provides an intelligent monitoring method, system, and storage medium for intelligent turnover cabinets. By employing technologies such as dual-mode sensor mutual verification, distributed edge computing, proactive behavior monitoring, RFID phase difference positioning, and LSTM prediction with fused periodic constraints, it solves the problems of low monitoring reliability, weak system fault tolerance, lack of proactive early warning, coarse positioning granularity, and insufficient prediction capability in existing technologies, thereby upgrading the intelligent turnover cabinet monitoring system from passive response to proactive prediction.

[0006] Firstly, this application provides an intelligent monitoring method for an intelligent turnover cabinet, the intelligent monitoring method for the intelligent turnover cabinet comprising: Step S1: Acquire the photoelectric signal and identification signal of the storage location, and perform sensor cross-verification based on the fluctuation characteristics of the photoelectric signal and the validity of the identification signal to obtain the storage location status data; Step S2: The edge node receives the storage space status data and triggers reporting based on the status change. When a communication interruption is detected, the election mechanism is started and the storage space status data is stored in the cache to obtain a distributed dataset. Step S3: Obtain the target's location information and combine it with biometric distance to determine permissions. When permission verification fails, extract the temporal relationship of state changes from the distributed dataset to obtain a behavioral feature sequence. Step S4: Measure the tag phase difference at multiple frequency points using the antenna array and unwrap it to obtain the distance. Calculate the spatial location based on triangulation and map it to the grid coordinate system. Associate the behavioral feature sequence to obtain spatial distribution data. Step S5: Extract the verification period attribute and circulation history from the spatial distribution data to construct a time series feature vector, input it into the long short-term memory network with fused period constraints to output the remaining period and circulation probability, screen risky assets, and obtain early warning information.

[0007] Secondly, this application provides an intelligent monitoring system for an intelligent turnover cabinet, the intelligent monitoring system for the intelligent turnover cabinet comprising: The acquisition module is used to acquire photoelectric signals and identification signals of the storage location, and to perform sensor cross-verification based on the fluctuation characteristics of the photoelectric signals and the validity of the identification signals to obtain storage location status data. The receiving module is used to receive the storage space status data from the edge node and trigger reporting based on the status change. When a communication interruption is detected, an election mechanism is started and the storage space status data is stored in the cache to obtain a distributed dataset. The extraction module is used to obtain the target's location information and combine it with biometric distance to determine permissions. When the permission verification fails, it extracts the temporal relationship of state changes from the distributed dataset to obtain a behavioral feature sequence. The association module is used to measure the tag phase difference at multiple frequency points through the antenna array and unwrap it to obtain the distance, calculate the spatial position based on triangulation and map it to the grid coordinate system, and associate the behavioral feature sequence to obtain spatial distribution data; The input module is used to extract the verification period attribute and circulation history from the spatial distribution data to construct a time-series feature vector, input the long short-term memory network with fused period constraints to output the remaining period and circulation probability, screen risky assets, and obtain early warning information.

[0008] Thirdly, an intelligent monitoring device for an intelligent turnover cabinet is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the intelligent monitoring device for the intelligent turnover cabinet to execute the aforementioned intelligent monitoring method for the intelligent turnover cabinet.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned intelligent monitoring method for intelligent turnover cabinets.

[0010] The technical solution provided in this application performs sensor mutual verification based on the fluctuation characteristics of photoelectric signals and the validity of identification signals. When photoelectric detection is stable but the identification signal is invalid, it automatically switches to radio frequency identification (RFID), solving the problem of easy misjudgment by a single sensor in the prior art. This mutual verification mechanism determines the sensor's working status by calculating the voltage fluctuation coefficient of the photoelectric signal and determines the anomaly type based on the scanning timeout flag. When the barcode scanning module is determined to be faulty, RFID backup identification is immediately activated, avoiding monitoring interruptions caused by sensor failures and ensuring high reliability of storage location status data acquisition. The traditional master-slave polling mode is replaced by an event-driven mechanism that triggers reporting based on state changes at edge nodes. When a communication interruption is detected, an election mechanism is initiated to select a temporary master node and store the data in the cache, solving the single-point failure problem in the prior art where a failure of the central node leads to the paralysis of the entire cabinet. This distributed architecture exchanges identifiers between nodes and selects the node with the smallest value as the temporary master node. The temporary master node takes over the data aggregation function and writes the storage location status data into non-volatile memory in timestamp order. By compressing and encoding to reduce storage space, a distributed dataset containing offline time period records is formed, ensuring the integrity and recoverability of monitoring data during network anomalies. By obtaining the target's location information and combining it with biometric distance to determine permissions, when permission verification fails, the temporal relationship of status changes is extracted from the distributed dataset, solving the problem that existing technologies cannot proactively prevent unauthorized operations. This proactive monitoring mechanism uses microwave radar to pre-detect approaching targets and trigger facial recognition, calculates the Euclidean distance between the feature vector and the pre-stored feature library, and determines unauthorized access and locks the cabinet door when the minimum distance exceeds the threshold. At the same time, the temporal relationship of storage location status changes is extracted from the distributed dataset, identifying abnormal operation patterns where the change timestamp of non-indication storage locations is earlier than that of indication storage locations, recording abnormal operations to form a behavioral feature sequence, realizing the transformation from post-event alarm to pre-event warning.

[0011] This RFID phase difference positioning method measures the tag phase difference at multiple frequency points using an antenna array and unwraps the tag to obtain the distance. Based on triangulation, the spatial location is calculated and mapped to a grid coordinate system. This solves the problem in existing technologies where whole-box storage can only identify the box level but cannot locate individual meters. The method measures the phase value of the reflected signal at multiple discrete frequency points, calculates the phase difference between adjacent frequency points, and calculates the distance difference based on the phase difference and frequency point interval. The distances from each antenna to the tag are accumulated. The least squares method is used to iteratively solve the distance equations to obtain the tag's three-dimensional spatial coordinates. The drawer storage space is divided into grid cells, and the grid index number is calculated based on the tag coordinates. Correlation of behavioral feature sequences yields spatial distribution data containing tag codes, grid locations, and operation records. This refines asset tracking granularity from the box level to the individual meter level, achieving precise monitoring of asset locations within the box. Furthermore, by extracting the verification cycle attribute and circulation history from the spatial distribution data, a time-based system is constructed. This intelligent prediction method, by inputting a long short-term memory network with fusion periodic constraints into an ordered feature vector, outputs the remaining period and turnover probability. It addresses the problem of existing technologies lacking predictive capabilities and only providing post-event warnings. This method calculates the proportion of time spent in storage by statistically analyzing the number of outbound and inbound trips within a preset time period. It constructs a multi-dimensional time-series feature vector by combining current storage status, date cycle, and other features. The long short-term memory network processes these time-series features through gating mechanisms such as forget gates, input gates, and output gates. A verification period constraint term is added to the loss function, penalizing cases where the predicted remaining period exceeds the verification period standard value. This ensures the model output conforms to electricity metering business rules. Assets with predicted remaining periods less than the warning threshold and turnover probabilities less than the stagnation threshold are selected as risk assets. Warning information containing asset identification and location information is generated, enabling early warning and priority outbound strategies. This transforms asset management from experience-driven to data-intelligent driven. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an embodiment of the intelligent monitoring method for intelligent turnover cabinets in this application. Figure 2 This is a schematic diagram illustrating the prediction of the remaining asset verification period and risk screening in the embodiments of this application; Figure 3 This is a schematic diagram of one embodiment of the intelligent monitoring system for the intelligent turnover cabinet in this application. Figure 4 This is a schematic block diagram of the intelligent monitoring device of the intelligent turnover cabinet in this embodiment of the invention. Detailed Implementation

[0014] This application provides an intelligent monitoring method, system, and storage medium for an intelligent turnover cabinet. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent monitoring method for intelligent turnover cabinets in this application includes: Step S1: Acquire the photoelectric signal and identification signal of the storage location, and perform sensor cross-verification based on the fluctuation characteristics of the photoelectric signal and the validity of the identification signal to obtain the storage location status data; Step S2: The edge node receives the storage space status data and triggers reporting based on the status change. When a communication interruption is detected, the election mechanism is started and the storage space status data is stored in the cache to obtain a distributed dataset. Step S3: Obtain the target's location information and combine it with biometric distance to determine permissions. When permission verification fails, extract the temporal relationship of state changes from the distributed dataset to obtain the behavioral feature sequence. Step S4: Measure the tag phase difference at multiple frequency points using the antenna array and unwrap it to obtain the distance. Calculate the spatial location based on triangulation and map it to the grid coordinate system. Associate the behavioral feature sequence to obtain spatial distribution data. Step S5: Extract the verification period attribute and circulation history from the spatial distribution data to construct a time series feature vector, input it into the long short-term memory network with fused period constraints to output the remaining period and circulation probability, screen risky assets, and obtain early warning information.

[0016] It is understood that the executing entity of this application can be the intelligent monitoring system of the intelligent turnover cabinet, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as the executing entity for illustration.

[0017] Specifically, in the intelligent monitoring method of the intelligent turnover cabinet, step S1 collects infrared obstruction signals within the storage space using a photoelectric sensor. When a measuring instrument is placed in the storage space, the infrared light path is blocked, causing a change in the output voltage of the photoelectric sensor. The system continuously samples this voltage signal to obtain a voltage sampling sequence. From this sequence, the maximum voltage value, minimum voltage value, and average voltage value are extracted. The voltage fluctuation coefficient is calculated as the difference between the maximum and minimum voltage values ​​divided by the average voltage value. This fluctuation coefficient reflects the stability of the photoelectric signal. Simultaneously, the barcode scanning module optically scans the barcode on the surface of the asset within the storage space. If a valid ASCII character sequence is not decoded within a preset timeout period, a scan failure flag is recorded. When the voltage fluctuation coefficient is less than a stability threshold, it indicates that the photoelectric detection is stable. If a scanning failure indicator is present, the scanning module is deemed faulty. In this case, the system automatically switches to radio frequency identification (RFID) mode and sends a read command to the RFID reader. The reader transmits radio frequency signals through a circularly polarized antenna. The electronic tag affixed to the asset receives the radio frequency energy and generates backscatter modulation. The reader receives and demodulates the reflected signal to obtain the electronic tag code stored in the tag. The system binds the obtained electronic tag code or barcode scanning result with the storage location number. Combined with the current on-site detection status of the photoelectric sensor, it forms storage location status data containing three fields: storage location number, asset code, and on-site status. This mutual verification mechanism determines the sensor's working status and selects the most reliable identification method by mutually verifying the stability of the photoelectric detection and the validity of the identification signal.

[0018] Edge nodes receive storage location status data via an RS485 fieldbus. The microcontroller within the edge node parses the data packet and checks the presence status change flag field. This flag field is a single-bit flag; when it changes from 0 to 1, it indicates a change in storage location status. The edge node immediately encapsulates the storage location status data into a data message conforming to the communication protocol. After adding source address, destination address, and message type control information to the message header, it triggers a report to the central node via the CAN bus. Simultaneously, the edge node starts a timer to send heartbeat packets to the central node at fixed intervals. The heartbeat packet is a short message with a specific format containing a node identifier and a timestamp. After sending the heartbeat packet, the edge node starts a response waiting timer. If a response packet is received from the central node before the timer expires, communication is considered normal, and the heartbeat counter is reset. If no response packet is received for a preset number of consecutive times, the heartbeat counter increments. When the counter reaches a preset threshold, communication with the central node is considered interrupted. The edge node records the current system timestamp as the communication interruption start timestamp and enters offline working mode. In offline mode, each edge node... Nodes exchange their node identifiers via the I2C bus. The node identifier is a pre-configured unique numerical code. After receiving the identifiers from other nodes, each node compares their numerical values ​​and selects the node with the smallest identifier value as the temporary master node. This temporary master node takes over the data aggregation function of the original central node and is responsible for receiving the storage position status data reported by other edge nodes. The edge nodes write the storage position status data into the circular buffer of the non-volatile Flash memory in ascending order of the timestamp field. The circular buffer adopts a circular queue structure and maintains a head pointer and a tail pointer. When the tail pointer catches up with the head pointer, the earliest record is overwritten. Before writing to memory, the storage position status data is compressed and encoded using LZ77. The compression process involves searching for a historical sequence that matches the current byte sequence to be encoded within a sliding window. If a match is found, a tuple of offset distance and matching length is recorded. If no match is found, the original byte is recorded. The compressed data is appended to the currently active Flash data block, forming a distributed dataset containing the communication interruption start timestamp, the total number of records during offline period, and the compressed data.

[0019] The microwave radar sensor transmits a linear frequency modulated continuous wave signal with a center frequency of 24 GHz. The frequency of this signal increases linearly from the starting frequency to the ending frequency within the scanning period. After receiving the signal reflected back from the target human body, the radar mixes it with the transmitted signal. The intermediate frequency (IF) signal after mixing is proportional to the target distance. A fast Fourier transform is performed on the IF signal to convert the time-domain signal to the frequency-domain signal. The frequency value corresponding to the peak value is extracted from the spectrum. The target distance is calculated based on this frequency value, which is equal to the frequency value multiplied by the speed of light and then divided by the frequency modulation slope. Simultaneously, the target velocity is extracted from the Doppler shift of the spectrum, which is equal to the Doppler shift multiplied by the speed of light and then divided by twice the carrier frequency. The system is designed... The system defines the distance and velocity ranges within a detection area, filters targets whose distance and velocity fall within these ranges as valid targets, and obtains target location information. Simultaneously, the camera captures a video stream and extracts the current video frame. A multi-task cascaded convolutional neural network is applied to this video frame for face detection. This network consists of three cascaded stages: candidate window generation, window refinement, and output of the face region. A 128-dimensional feature vector is extracted from the detected face region, with each dimension of the feature vector representing the output value of the last fully connected layer of the convolutional neural network. A pre-stored feature library of authorized users, containing feature vectors from multiple authorized users, is also retrieved from a database. The system calculates the Euclidean distance between the currently extracted feature vector and the feature vector of each user in the feature library. The Euclidean distance is calculated by taking the square root of the sum of the squares of the differences in the corresponding dimensions. The minimum value is selected from all calculated Euclidean distances. If this minimum Euclidean distance exceeds a preset distance threshold, it indicates that the current face does not match any authorized users, and the permission verification fails. At this time, the system reads the storage space status data records within a preset time window from the distributed dataset generated in step S2. The time window is obtained by subtracting the window duration from the current timestamp to obtain the starting timestamp. The system extracts the storage space number, the status change identifier, and the timestamp field from each record, and then... The timestamp field values ​​are sorted from smallest to largest to obtain the state change time sequence. This time sequence records the order in which the state changes of each storage location change. The system analyzes this time sequence to identify abnormal operation patterns. Specifically, the system extracts a list of storage location numbers from the outbound instructions that should have their indicator lights lit as the indicated storage locations. It then iterates through each record in the state change time sequence. If a storage location number of a certain record is not in the indicated storage location list and the timestamp of that record is less than the timestamp of any indicated storage location record, it is determined to be an abnormal operation, i.e., the operator took assets from a non-indicated storage location first. The storage location number, timestamp, and operation type corresponding to this abnormal operation are recorded to obtain a behavioral feature sequence.

[0020] The RFID reader sequentially activates each antenna in the antenna array via control signals. When each antenna is activated, the reader transmits RFID signals at multiple discrete frequency points, such as 920.5MHz, 921.5MHz, 922.5MHz, and 923.5MHz. After receiving the RFID signal, the RFID tag on the asset modulates the stored data into the reflected signal via backscattering. The reader receives this reflected signal and extracts the in-phase and quadrature components through quadrature demodulation. It calculates the phase value of the reflected signal based on the in-phase and quadrature components, obtaining the phase value by dividing the in-phase component by the quadrature component using the arctangent function. The reader records the phase value corresponding to each frequency point and calculates the phase difference between adjacent frequency points, which is equal to the phase difference of the next frequency point. Subtracting the previous frequency point's phase value from the position value, since the phase change during radio frequency signal propagation is proportional to the propagation distance, the distance difference is calculated based on the phase difference and frequency point interval. The distance difference equals the phase difference multiplied by the speed of light and then divided by the product of the frequency point interval and 4π. The distance differences calculated at each frequency point are summed to obtain the distance from the antenna to the tag. Due to the 2π periodic ambiguity of the phase value, phase unwrapping is achieved to eliminate ambiguity by measuring and calculating the distance difference at multiple frequency points. The reader sequentially activates the four antennas in the antenna array and repeats the above process to obtain the distance values ​​from the four antennas to the tag. The known spatial coordinates of each antenna in the antenna array are obtained. These coordinates are the three-dimensional positions of the antennas in the drawer storage coordinate system. The distance between two points in space is equal to the square root of the sum of the squares of their coordinate differences. To determine the relationship between the antenna coordinates and the tag's coordinates, a system of four equations is established. Each equation describes the distance relationship between an antenna coordinate and the tag's coordinates to be determined. An error function is constructed as the sum of squares of the differences between the measured distances of each antenna and the theoretical distances calculated based on the current coordinate guesses. Partial derivatives of the error function are calculated with respect to the horizontal, vertical, and lateral coordinates. These partial derivatives reflect the rate of change of the error function in each coordinate direction. The three partial derivatives form a three-dimensional gradient vector. The initial coordinate guesses are set to the geometric center coordinates of the drawer storage location. The coordinate values ​​are updated along the negative direction of the gradient vector. The update formula is: new coordinates equal old coordinates minus the learning rate multiplied by the gradient component in the corresponding direction. The learning rate is a preset positive number that controls the step size of each iteration. This process is repeated. In the gradient update process, the error function value is recalculated after each iteration. Iteration stops when the absolute value of the difference between the error function value of the current iteration and the error function value of the previous iteration is less than the convergence threshold. The three coordinate components that finally converge are used as the three-dimensional spatial coordinates of the label. The system pre-divides the drawer storage space into cubic grid cells according to preset dimensions. The dimensions of the grid cells in the horizontal, vertical, and height directions are set respectively, and a grid coordinate system is established. The grid index is obtained by dividing the three-dimensional spatial coordinates of the label by the grid size in each direction and rounding down. The grid indices in the three directions are combined according to a specific formula to calculate a unique index number. The index number is equal to the horizontal index multiplied by the product of the number of horizontal and vertical grid cells and the number of vertical grid cells.In addition to the horizontal and vertical indexes multiplied by the number of vertical grid cells, and the vertical height index, this index number is data-bound with the label code and stored in the association table. Records matching the current drawer location number are extracted from the behavioral feature sequence generated in step S3. The operation timestamp and operation type fields are read from the matching records and associated with the aforementioned index number. Two fields, timestamp and operation type, are added to the association table, forming spatial distribution data containing the label code, grid location index number, operation timestamp, and operation type.

[0021] The system reads the tag code field of each record from the spatial distribution data table. Based on the tag code, it queries the asset database to obtain the corresponding asset's verification date and asset type. The verification date is the timestamp of the last verification completed for that asset. The asset type is an enumerated value identifying categories such as single-phase meters, three-phase meters, and low-voltage transformers. The system pre-configures the standard verification cycle values ​​corresponding to each asset type. For example, the standard verification cycle value for single-phase and three-phase meters is 8 years (2920 days), and the standard verification cycle value for low-voltage transformers is 5 years (1825 days). The system reads the current system timestamp and calculates the difference between the current timestamp and the verification date timestamp. This difference is in seconds. Dividing this difference by 86400 converts it to days to obtain the used cycle. The standard verification cycle value is then subtracted from the used cycle. The remaining verification period is obtained. Operation records for each asset are read from the spatial distribution data table. Records with the operation type "outbound" are filtered, and the number of outbound operations within a preset time period (e.g., the past 180 days) is counted to obtain the outbound count. Records with the operation type "inbound" are filtered, and the time difference between adjacent inbound and outbound operations is calculated and summed to obtain the total inbound duration. The total inbound duration is divided by the total length of the time period to obtain the in-stock duration percentage. A time-series feature vector is constructed, containing multiple dimensions: the first dimension is the current in-stock status (1 if the asset is currently in the cabinet, 0 otherwise); the second dimension is the number of outbound operations; the third dimension is the in-stock duration percentage; the fourth dimension is the code for the current day of the week; and the fifth dimension is the code for the current day of the month. The sequential feature vector is input into the Long Short-Term Memory (LSTM) network model, which consists of an input layer, a first LSTM layer, a second LSTM layer, a fully connected layer, and an output layer. The input layer receives the temporal feature vector. The first LSTM layer contains multiple LSTM units, each maintaining a hidden state vector and a cell state vector. A forget gate controls how many old cell states are retained, an input gate controls how many new inputs are received, and an output gate controls how many hidden states are output. The output of the first LSTM layer serves as the input to the second LSTM layer, which also processes sequence information through a gating mechanism. The hidden state at the last time step of the second LSTM layer is input to the fully connected layer, which undergoes a linear transformation using a weight matrix and a bias vector before activation. The model's output layer contains two neurons: the first neuron outputs the predicted remaining cycle, and the second neuron outputs the future transition probability. During model training, a loss function is defined, consisting of two parts: the first part is the mean squared error, calculated as the sum of the squares of the difference between the predicted and actual remaining cycles and the squares of the difference between the predicted and actual transition probabilities; the second part is a test cycle constraint term, which is the square of the excess if the predicted remaining cycle exceeds the test cycle standard value, otherwise it is zero. The total loss function is the mean squared error plus the constraint term multiplied by the weight coefficients. The network weights are updated using backpropagation and a gradient descent optimizer. After training, the model is used for inference, with forward propagation performed on the input temporal feature vector.The input layer passes the feature vector to the first LSTM layer. At each time step, the first LSTM layer calculates the hidden state. The second LSTM layer receives the output from the first LSTM layer and continues processing. A fully connected layer performs a linear transformation and non-linear activation on the hidden state at the last time step of the second LSTM layer. The two neurons in the output layer output the predicted remaining cycle and future turnover probability values, respectively. The system sets a warning cycle threshold (e.g., 30 days) and a stagnation probability threshold (e.g., 0.2). It iterates through all assets, filtering those with a predicted remaining cycle less than the warning cycle threshold and a future turnover probability less than the stagnation probability threshold. These assets face the risk of an impending expiration of their inspection period but a low probability of being released in the near future. The system reads the corresponding tag codes and grid location index numbers from the spatial distribution data table to generate warning information. This warning information includes the asset identifier (tag code), location information (grid location index number), predicted remaining cycle value, and future turnover probability value. The system reports the warning information to the management platform via a network interface for managers to view and take appropriate measures.

[0022] In one specific embodiment, step S1 includes: Infrared occlusion signals of the storage location are collected by photoelectric sensors. The infrared occlusion signals are continuously sampled to obtain a voltage sampling sequence. The maximum voltage value, minimum voltage value, and average voltage value are calculated based on the voltage sampling sequence. The voltage fluctuation coefficient is obtained by the ratio of the difference between the maximum voltage value, minimum voltage value, and average voltage value to the average value. The barcode of the asset in the storage location is scanned by the barcode scanning module. When the scan times out and no valid character code is returned, the scan failure mark is recorded. The sensor abnormality type is determined by comparing the scan failure mark with the threshold of the voltage fluctuation coefficient. When the voltage fluctuation coefficient is less than the stability threshold and there is a scan failure indicator, a tag reading command is sent to the radio frequency reader. The reflected signal of the asset electronic tag is received through the radio frequency antenna and demodulated to obtain the electronic tag code. By binding the electronic tag encoding or barcode scanning results with the storage location number and combining them with the on-site detection status of photoelectric sensors, storage location status data containing the storage location number, asset code, and on-site status can be obtained.

[0023] Specifically, the infrared emitting tube inside the photoelectric sensor continuously emits an infrared beam, and the receiving tube receives the beam and outputs a voltage signal. When a single-phase meter or three-phase meter or other measuring instrument is placed in the storage space, the asset blocks the infrared light path, causing the output voltage of the receiving tube to drop. The analog-to-digital converter continuously samples this voltage at a frequency of 1000 times per second, collecting 100 voltage values ​​within a 100-millisecond time window to form a voltage sampling sequence. The maximum value is recorded as the maximum voltage value, and the minimum value is recorded as the minimum voltage value. The average voltage value is obtained by summing all the sampled values ​​and dividing by the number of sampling points. The difference between the maximum and minimum voltage values ​​is calculated and divided by the average value. The voltage value yields the voltage fluctuation coefficient, which reflects the degree of voltage signal jitter. A small fluctuation coefficient indicates a stable signal, while a large fluctuation coefficient indicates an unstable signal. The image sensor of the barcode scanning module optically images the barcodes on the surface of the assets in the storage space. The decoding chip binarizes the image, extracts the width information of the black and white stripes, and decodes it into an ASCII character sequence according to the barcode standard rules. After the scanning module starts, if the decoding chip outputs a complete character sequence within a 200-millisecond timeout, the scan is successful. If no output is made within the timeout or the output character verification error occurs, the scan failure flag is set to 1. The judgment logic is to check whether the voltage fluctuation coefficient is less than the stability threshold. Simultaneously, the scan failure flag is checked for a value of 1. When the voltage fluctuation coefficient is less than the stability threshold and the scan failure flag is 1, the scanning module is deemed faulty while the photoelectric sensor is normal. At this time, the main controller sends a hexadecimal command code to the RFID reader's serial port to initiate RFID reading. The reader drives the circularly polarized antenna to radiate a 920.5MHz radio frequency signal. The RFID electronic tag antenna on the asset surface receives the radio frequency signal and generates an induced current to power the chip. The chip modulates the stored 96-bit EPC code into the reflected signal by changing the antenna impedance to achieve backscatter modulation. After receiving the reflected signal, the reader performs downconversion and quadrature demodulation, and then baseband processing. The device performs Manchester decoding on the demodulated signal to recover the digital bit stream and extracts the EPC code to obtain the electronic tag code. The main controller maintains a mapping table between the storage location number and the asset code. When the barcode is successfully scanned, it is used as the asset code. When the RFID is successfully read, it is used as the asset code. The association between the storage location number and the asset code is established in the mapping table. The current output level of the photoelectric sensor is read. A low level indicates that the light path is blocked and the asset is determined to be in place. A high level indicates that the light path is unobstructed and the asset is determined to be in place. A data structure containing three fields, namely storage location number, asset code and in-place status, is constructed to form the storage location status data.

[0024] The edge node microcontroller connects to the fieldbus via an RS485 transceiver, configured to operate at a baud rate of 115200, with 8 data bits, 1 stop bit, and no parity. After the storage location connection board detects a status change, it sends the storage location status data to the bus via the Modbus RTU protocol. The edge node, acting as the master, polls and reads bit 1 of the slave station register address 0x0003 to obtain the storage location status change indicator. If bit 1 is 1, it indicates a storage location status change. It reads the storage location number from register 0x0001, the asset barcode from 0x0004 to 0x000E, and the storage location status from bit 0 of 0x0003. This data is assembled into a CAN extended frame. The 29-bit identifier stores the edge node ID in the high 3 bits, the storage location number in the next 5 bits, and the message type in the next 5 bits. The 8-byte data field stores the storage location number sequentially. Asset coding and on-premises status are recorded. The CAN controller's transmit interface sends data frames to the CAN bus. The edge node maintenance software timer triggers a heartbeat packet every 10 seconds. The heartbeat packet is a standard CAN frame with the node ID as the identifier and a timestamp and liveness flag (0xAA) in the data field. After sending the heartbeat packet, a 5-second timeout timer is started to wait for a response packet from the central node. The highest bit of the response packet identifier is set to 1 to indicate the response type. The maintenance heartbeat counter is initially 0, incremented by 1 with each transmission, and reset to zero upon receiving a response. When the counter reaches 3, a communication interrupt is triggered after three consecutive unresponsive heartbeats. The data is then read. The clock obtains the Unix timestamp record as the start time of the communication interruption. Each edge node exchanges node identifiers via the I2C bus. After collecting the identifiers of other nodes, it compares the values ​​and selects the node with the smallest value. This node switches to temporary master node mode to receive the storage status data reported by other edge nodes. The storage status data is inserted into the circular buffer of the Flash memory in ascending order of timestamp. The circular buffer maintains a head pointer pointing to the earliest record and a tail pointer pointing to the next writable position. Before writing, LZ77 compression is performed on the data. The compression algorithm maintains a 32KB sliding window to match the current byte sequence with the historical sequence within the window. The hash value of the current sequence is calculated, and the historical position with the same hash value is searched in the hash table. The maximum matching length and offset distance of the records are compared to see if the bytes are identical. If the matching length is greater than or equal to 3 bytes, a triple containing the offset distance, matching length, and the next byte is output. If the matching length is less than 3 bytes, the current byte is directly output. The compressed data is appended to the currently active block of Flash. When the block is full (4KB), a new block is allocated to continue writing. The Flash block index table records the starting address, the number of valid records, and the timestamp range of each block, forming a distributed dataset.

[0025] The microwave radar sensor generates a linear frequency modulated (LFM) signal with a period of 60 microseconds, ranging from 77 GHz to 81 GHz. The transmitting antenna radiates the signal reflected from the human target. The receiving antenna receives the reflected signal and mixes it with the transmitted signal to output an intermediate frequency (IF) signal. The IF signal frequency is proportional to the target distance. After sampling by the analog-to-digital converter (ADC), the digital signal processor performs a 2048-point fast Fourier transform to obtain the frequency domain spectrum. Spectral lines with amplitudes greater than the detection threshold are identified. The target distance is calculated from the spectral line frequency by multiplying the spectral line frequency by the speed of light, multiplying by the frequency modulation period, and dividing by twice the frequency modulation bandwidth. The distance is extracted from the changes in the spectrum over multiple periods. The Doppler frequency shift calculation speed is equal to the Doppler frequency shift multiplied by the speed of light divided by twice the carrier frequency. Targets with a distance between 0.3 meters and 1.5 meters and a speed between -0.5 meters per second and +0.5 meters per second are selected as valid targets to obtain target location information. The camera outputs a video stream at a rate of 30 frames per second to extract the current RGB image data. This data is input into a multi-task cascaded convolutional neural network. The first stage generates candidate windows, the second stage refines the windows, and the third stage outputs the coordinates of the face region. After cropping the face region to a fixed size, it is input into a face recognition network to extract a 128-dimensional feature vector. The authorized user feature library is read from the database. Each record is traversed, and the feature vector is extracted. The Euclidean distance between the feature vector and the current feature vector is calculated. The distance is the square root of the sum of the squares of the differences in the corresponding dimensions. The minimum value among all distances is recorded. If the minimum distance exceeds the distance threshold, the authorization verification is deemed to have failed. Data blocks whose timestamp range overlaps with the current time window are searched from the Flash block index table of the distributed dataset. The starting timestamp of the time window is the current timestamp minus the window duration. The data blocks are read and LZ77 decompression is performed. During decompression, when triples are encountered, the corresponding length is copied from historical data according to the offset distance. The byte sequence is appended with the next byte. Single bytes are output directly. The original storage location status data is decompressed. The storage location number in-situ status and timestamp fields are extracted and sorted by timestamp from smallest to largest using quick sort. The sorted sequence is the time sequence relationship of status changes. The list of storage location numbers that should light up the green light is parsed from the outbound instruction. Each record in the time sequence relationship is traversed to check if the storage location number is in the indication list. If it does not exist, it is a non-indication storage location. The timestamp is compared. If it is smaller than the timestamp of any indication storage location, it is determined to be an abnormal operation. The storage location number, timestamp and operation type of the abnormal operation are recorded to form a behavior feature sequence.

[0026] The RF reader / writer writes control registers to the RF front-end chip via the SPI interface to set the antenna port transmit power and operating frequency. Setting the antenna port selection field to 1 activates the first antenna, and the transmit power register is set to 28dBm corresponding to the code value. The operating frequency register is sequentially written with four frequency words from 920.5MHz to 923.5MHz. The phase-locked loop generates a local oscillator signal based on the frequency words, which is then amplified by the transmission link and radiated through the antenna. The tag receives the RF signal, and the rectifier circuit converts it to DC voltage for power supply. The tag chip reads the EPC code according to the EPC... The Gen2 protocol uses Manchester encoding. The reflected signal is modulated by switching the antenna impedance. The reader receives the reflected signal, down-converts it to an intermediate frequency (IF), and quadrature demodulates it to output in-phase and quadrature components. The analog-to-digital converter samples the signal and calculates the phase value, which is equal to the arctangent function divided by the quadrature component and the in-phase component. The current frequency phase value is recorded, and the measurement is repeated at the next frequency. After measuring four frequencies, the phase difference between adjacent frequencies is calculated as the difference between the previous and subsequent phases. The phase difference is less than -π plus 2π and greater than positive π minus 2π, completing the unwrapping. The distance difference is calculated based on the phase difference, which is equal to the phase difference multiplied by the speed of light divided by the frequency interval multiplied by 4π. The frequency interval is 1MHz, and the speed of light is 3 x 10⁸ meters per second. The three distance differences are summed to obtain the total distance from the antenna to the tag. The four antennas are switched sequentially for repeated measurements to obtain four distance values. The spatial coordinates of the four antennas are read from the configuration file as the four corners of the drawer. The three components of the tag's coordinates are set as unknowns. Four equations are established based on the spatial distance formula, and an error function is constructed. The gradient vector is obtained by subtracting the sum of squared differences between the left and right sides of the four equations and taking the partial derivative of the error function with respect to the three coordinate components. The initial coordinate guess is the geometric center of the drawer, i.e., the length, width, and height are each divided by 2. The coordinates are updated along the negative gradient direction. The new coordinates are equal to the old coordinates minus the learning rate multiplied by the gradient component. The updated error function value is calculated, and the absolute value of the difference between the old and new errors is compared. The iteration stops when the difference is less than the convergence threshold. After convergence, the three coordinate components are the three-dimensional spatial coordinates of the label. The number of grids in each direction is calculated based on the drawer size and grid size. Each component of the label coordinates is divided by the grid size and rounded down to obtain three grid indices. The index number is equal to the horizontal index multiplied by the product of the vertical grid number and the height grid number, plus the vertical index multiplied by the height grid number and the height index. A record is inserted into the database to establish a label code and bind it to the grid index number. The operation timestamp and operation type are extracted from the record matching the storage location number from the behavioral feature sequence. The database table is updated to add timestamp and operation type fields to form a spatial distribution data table.

[0027] The spatial distribution data table is used to query the tag code. Based on the code, the asset database is used to query the inspection date and asset type. The inspection date is a Unix timestamp, and the asset type is an enumerated integer. The configuration table is used to query the standard value of the inspection period in days based on the asset type. The current system timestamp is read, and the inspection date timestamp is subtracted to get the time difference. This time difference is divided by 86400 to convert it to days, which gives the used period. The standard value of the inspection period is subtracted from the used period to get the remaining inspection period. Records with the operation type of "outbound" are filtered from the spatial distribution data table, and the number of outbound records in the past 180 days is counted to get the outbound frequency. Records with the operation type of "inbound" are filtered, grouped by asset code, and then sorted by time. The system uses a time-stamp sorting method to find adjacent inbound and outbound records. It calculates the outbound timestamp minus the inbound timestamp to obtain the duration of a single inbound transaction. The total inbound duration is then summed, and divided by 180 days and seconds to obtain the percentage of inbound duration. A five-dimensional temporal feature vector is constructed: the first dimension represents the current inbound status (1 if present, 0 if absent); the second dimension represents the number of outbound transactions; the third dimension represents the percentage of inbound duration; the fourth dimension represents the day of the week (1 to 7); and the fifth dimension represents the day of the month. This temporal feature vector is input into a Long Short-Term Memory (LSTM) network model. The input layer is passed to the first LSTM layer, which contains 64 units maintaining 64 dimensions of hidden states and cell states. The forgetting gate uses a sigmoid function to calculate the forgetting coefficient and the old cell state. Selective forgetting is achieved through state multiplication. The input gate controls the reception of new inputs, and the output gate controls the output of the hidden state. The output of the first LSTM layer serves as the input to the second LSTM layer, which contains 32 units. The hidden state of the second LSTM layer is input at the last time step. The fully connected layer is linearly transformed by the weight matrix and bias vector, and then activated by the activation function. The output layer has two neurons; the first output predicts the remaining cycle time, and the second outputs the future transition probability. The loss function during training includes mean squared error and a cycle time constraint term. The mean squared error is the square of the difference between the predicted and actual remaining cycle time plus the square of the difference between the predicted and actual transition probabilities. The constraint term is the prediction... When the remaining period exceeds the standard value, the excess portion is squared; otherwise, it is zero. The total loss is the mean squared error plus the constraint term multiplied by the weight coefficient. The network weights are updated through backpropagation and gradient descent. During inference, the input feature vector is used to perform forward propagation and output the predicted remaining period and circulation probability values. A warning period threshold and a stagnation probability threshold are set to filter assets whose predicted remaining period is less than the warning threshold and whose circulation probability is less than the stagnation threshold as risk assets. The label code and grid location index number of the risk assets are read from the spatial distribution data table to generate warning information containing asset identification location information, predicted remaining period, and circulation probability, which is reported to the management platform through the network interface.

[0028] In one specific embodiment, step S2 includes: Edge nodes receive storage location status data via fieldbus, detect the presence status change flag in the storage location status data, and when the presence status change flag is detected, encapsulate the storage location status data into a data packet and trigger reporting. Edge nodes detect the communication status with the central node by periodically sending heartbeat packets and listening for response packets. When no response packet is received for a preset number of consecutive times, the communication is considered to be interrupted, and the start timestamp of the communication interruption is recorded. When a communication interruption is detected, edge nodes compare priorities by exchanging node identifiers and select the edge node with the smallest identifier value as the temporary master node, which then takes over the data aggregation function. The storage location status data is written into a circular buffer of non-volatile memory in timestamp order. The storage location status data is then compressed and encoded before storage to obtain a distributed dataset containing offline time period records.

[0029] Specifically, the transceiver of the edge node microcontroller is connected to the fieldbus, and the serial port is configured in standard communication mode. After the storage location connection board detects the placement of a single-phase meter or three-phase meter in the storage location, it sends the storage location status data to the bus via the communication protocol. The edge node, acting as the master station, reads the status change flag bit from the slave station register. This flag bit is a single-bit flag; when the value is set, it indicates a change in the storage location status. The edge node simultaneously reads the storage location number, asset barcode, and location status from the corresponding register and assembles these fields into a data frame. The identifier of the data frame stores node information in segments: the high-order segment stores the edge node identifier to distinguish different nodes, the middle segment stores the storage location number corresponding to the single-phase meter position, three-phase meter position, and transformer position, and the low-order segment stores the message type distinction. Status changes, anomaly alarms, and heartbeat packets are processed. The data field sequentially arranges the storage location number and asset barcode or tag code. The edge node calls the controller's transmit register to write frame data and triggers transmission. The controller broadcasts the data frame to the central node and other edge nodes via differential signals on the bus. The edge node's internal maintenance software timer is implemented based on the system tick timer, configured for periodic triggering. Each time the timer expires, it triggers an interrupt service routine to construct a heartbeat packet. The heartbeat packet format is a standard frame, with the edge node identifier filled in. The data field stores the current timestamp and a fixed liveness flag byte. After the heartbeat packet is sent, a timeout timer is started and the node enters a waiting state. Upon receiving the heartbeat packet, the central node immediately constructs a response packet. The highest bit of the identifier in the response packet is... The remaining bits are copied from the heartbeat identifier. The edge node listens for all frames on the bus during the receive interrupt. When a frame is received where the highest bit of the identifier is set and the remaining bits match the node's, it is considered an acknowledgment packet. The edge node maintains a heartbeat counter variable stored in memory, initially zero. The counter increments after each heartbeat packet is sent and resets after each acknowledgment packet is received. When the counter value reaches a preset threshold, it indicates that no acknowledgment has been received from the central node for several consecutive periods. The edge node determines that the communication link with the central node has been interrupted, reads the real-time clock chip to obtain time information through the bus interface, converts the time information into a timestamp format, and records this timestamp as the communication interruption start time. This timestamp is stored in non-volatile memory to prevent loss due to power failure. Each edge node detects... Upon detecting a communication interruption, communication resumes via an inter-node bus, which includes clock and data lines. Each edge node is configured with a unique slave address. A node, acting as the master, sends read requests to other nodes sequentially. Each slave responds by returning its node identifier value. The master collects all node identifiers and compares their values. The comparison logic involves iterating through all identifiers and finding the one with the smallest value. The node with the smallest value is elected as the temporary master. The temporary master broadcasts the election result message to other nodes via the bus, with the message identifier set to broadcast type. Upon receiving the election result, other nodes switch their state to slave mode. The temporary master then begins listening to the bus to receive storage position status data reported by other edge nodes and caches it in its local memory.Edge nodes insert the received storage status data into the circular buffer of non-volatile memory according to the numerical order of the timestamp field. The circular buffer allocates contiguous address space in memory, maintains a head pointer variable pointing to the starting address of the earliest record, and a tail pointer variable pointing to the address of the next writable record. Before inserting a new record, it checks whether the tail pointer plus the record length exceeds the end of the buffer. If it does, it wraps back to the starting address of the buffer. If the tail pointer catches up with the head pointer after wrapping back, the head pointer moves forward by one record length to overwrite the oldest record. Before writing, a lossless compression algorithm is performed on the storage status data. The compression algorithm maintains a sliding window to store the most recently processed data. The current byte sequence to be encoded is matched against the historical sequences within the sliding window. The matching process involves calculating the hash value of the first few bytes of the current sequence. The hash value is calculated by summing the byte values ​​by position and then taking the modulo. The hash table is then searched for the historical sequence position corresponding to the same hash value. Starting from that historical position, the current sequence is compared byte by byte. The maximum number of consecutively matched bytes is recorded as the matching length and the offset distance of the historical position relative to the current position. If the matching length reaches or exceeds the minimum matching threshold, a valid match is considered found. The output triple contains the offset distance, the matching length, and the next character after the matching ends. If the matching length is less than the threshold, no replacement is performed, and the original value of the current byte is directly output. The compression process proceeds byte by byte until the entire data is processed. The length of the compressed data is usually reduced. The compressed data is appended to the currently active block of memory. Memory is organized by block. When the remaining space of the current block is insufficient, the next free block is allocated. The block index table records the starting address, number of used bytes, earliest record timestamp, and latest record timestamp of each block. The index table itself is stored in the front area of ​​memory and occupies a fixed amount of space. The block index table is used to locate all storage location status data records during offline periods to form a distributed dataset. This dataset includes the communication interruption start timestamp. The data consists of fields, a total number of offline records, and compressed data for each block. Once the network is restored, the central node queries the missing data from the offline period based on the start timestamp. Edge nodes read the block index table to find all data blocks for the corresponding time period, and sequentially read the compressed data of each block to decompress and restore the original data records. The decompression process involves reading the compressed data stream; when a triple is encountered, a specified length of historical byte sequence is copied from the already decompressed output buffer based on the offset distance, and the next byte is appended to the output buffer. When a single byte is encountered, it is directly written to the output buffer. The decompressed original data is then transmitted to the central node via the network interface to complete data synchronization.

[0030] For example, a power supply station's intelligent turnover cabinet is configured with multiple edge nodes to manage single-phase meter storage locations, three-phase meter storage locations, and transformer storage locations respectively. During normal operation, when the first edge node detects that a single-phase meter has been removed from a certain storage location, the status change flag bit in the slave station register changes from low to high, triggering a status change. The first edge node reads the storage location number and asset barcode, then assembles a data frame. The high-order segment of the identifier identifies the first node, the middle segment identifies the corresponding storage location, and the data field stores the complete storage location status data. The first edge node sends this frame to the central node via the bus. After receiving it, the central node updates its database and returns a response. Simultaneously, the first edge node periodically sends heartbeat packets. The heartbeat packet identifier is the node identifier, and the data field contains the current timestamp and a fixed flag. The central node... Upon receiving a heartbeat packet, the first edge node immediately sends a response packet. Upon receiving the response, the first edge node resets its heartbeat counter. Suppose that at some point, the central node's main control board malfunctions and stops responding. If the first edge node's heartbeat packet times out without a response, the counter increments. After multiple consecutive heartbeat packet timeouts, the counter reaches the threshold. The first edge node determines that communication has been interrupted and reads the real-time clock to obtain the timestamp, recording the interruption start time. The first edge node reads identifiers from other nodes via the inter-node bus, compares all node identifiers, and determines the node with the smallest value. This node becomes the temporary master node and broadcasts the election result via the bus. Other nodes switch to slave node mode. Subsequently, the second edge node detects the entry of a three-phase meter into a storage location. The second node then updates the storage location status. Status data is sent via the bus. The ephemeral master node receives and buffers the data. It then inserts the data into a circular buffer in memory according to its timestamp. The buffer's tail pointer currently points to a certain address. Before writing, the data is compressed. Assuming the first few bytes contain fixed-format fields such as storage location number and asset type, they are designed to perfectly match the same fields previously recorded in the sliding window. When the matching length reaches a threshold, the offset distance is the position of the previous record. The output triple contains the offset value, matching length, and the next byte. Subsequent bytes with asset barcodes that differ from historical records are output byte-by-byte. After compression, the data length is reduced. The compressed data is written to the memory address, and the tail pointer is updated. The block index table records the increase in the number of bytes used in the block and retains the earliest timestamp. The latest timestamp is updated, and data is continuously cached during offline periods until the storage utilization rate approaches the limit. When the central node recovers, the temporary master node detects the heartbeat response recovery. The temporary master node queries the block index table to find all data blocks from the start timestamp to the current time, reads each block of compressed data in sequence and decompresses it. When reading a triple during decompression, it offsets the output buffer forward by a specified number of bytes, copies the specified number of bytes to the current output position, and then appends the next byte. After decompression is completed, the original data is uploaded to the central node in batches via the network. The central node inserts these data into the database according to the timestamp to complete the records during the offline period. After the temporary master node completes data synchronization, it clears the offline data in the storage cache to release storage space and switches back to normal node mode.

[0031] In one specific embodiment, step S3 includes: The microwave radar sensor transmits a linear frequency modulated continuous wave signal and receives the reflected signal. The Fourier transform of the reflected signal is performed to extract the target distance and velocity. Valid targets within a preset area are then selected to obtain the target location information. The system acquires video frames and performs face detection to obtain face regions. It then extracts feature vectors from these face regions and calculates the Euclidean distance between the feature vectors and feature vectors in a pre-stored feature library. If the minimum Euclidean distance exceeds a distance threshold, the system determines that the permission verification has failed. When the permission verification fails, the storage location status data within the time window is read from the distributed dataset, the status change identifier and timestamp of each storage location are extracted and sorted by timestamp to obtain the time sequence relationship of status changes. Based on the temporal relationship of state changes, identify abnormal operations where the change timestamp of non-indication storage locations is earlier than that of indication storage locations. Record the storage location number and timestamp of the abnormal operation to obtain the behavioral feature sequence.

[0032] Specifically, the radio frequency front-end of the microwave radar sensor generates a frequency-modulated signal that varies linearly between a start frequency and a stop frequency. The transmitting antenna radiates this signal into the space in front of the cabinet door. When a human target is located in the detection area, some energy is reflected. The receiving antenna captures the reflected signal and mixes it with the transmitted signal in a mixer to generate an intermediate frequency (IF) signal. The frequency of the IF signal is proportional to the target distance. After the analog-to-digital converter samples the IF signal, the digital signal processor performs a fast Fourier transform to convert the time-domain waveform into a frequency-domain spectrum. The basic principle of the Fourier transform is to decompose a complex signal into a superposition of sine waves of different frequencies. The processor traverses the transformed spectrum data to find the spectral lines whose amplitude exceeds the detection threshold, and calculates the target distance from the frequency value corresponding to the spectral line. The distance is calculated using the frequency value. The process multiplies the target's speed by the speed of light, then by the frequency modulation period, and finally divides by twice the frequency modulation bandwidth. Simultaneously, it extracts the Doppler shift from the spectral changes of multiple consecutive frequency modulation periods. The Doppler shift reflects the radial velocity of the target relative to the radar; the velocity is calculated by multiplying the Doppler shift by the speed of light and then dividing by twice the carrier frequency. The processor filters targets based on preset detection ranges for distance and velocity, retaining only targets whose distance and velocity are within the set ranges as valid targets. Target position information is obtained from the distance and angle information of the valid targets. The camera image sensor outputs a video stream at a fixed frame rate. Color image data of the current frame is extracted from the video stream and input into a multi-task cascaded convolutional neural network for face detection. This network consists of three cascaded stages; the first stage is a shallow layer... A convolutional network quickly scans the entire image to generate a large number of candidate windows that may contain faces. The second-stage network refines these candidate windows, eliminating obvious non-face windows and retaining those with higher confidence. The third-stage network precisely locates the retained windows, outputting the coordinates and size of the face region. The face region is then cropped from the image based on its coordinates and adjusted to a fixed size. This is input into a face recognition convolutional neural network to extract feature vectors. This network maps the face image into a high-dimensional feature vector through multiple convolutional pooling and fully connected layers. Each dimension of the feature vector represents the activation value of the corresponding neuron in the final fully connected layer of the network. The network also reads the authorized user feature library from a database table. Each record in this table contains a user identifier field and a feature vector field. The segment is stored as a floating-point array. It iterates through all records in the table to extract the feature vector of each user. The Euclidean distance between the extracted feature vector and the user's feature vector is calculated. The Euclidean distance is calculated as the square root of the sum of the squares of the differences in corresponding dimensions. The minimum value among all calculated distances is recorded. It is then determined whether this minimum Euclidean distance exceeds a preset distance threshold. If it does, it indicates that the current face does not match any authorized users, and the permission verification fails. When permission verification fails, the data block whose timestamp range overlaps with the specified time window is searched from the storage block index table of the distributed dataset. The start timestamp of the time window is the current timestamp minus the window duration, and the end timestamp is the current timestamp. Compressed data from the data block that matches the time range is read.The decompression algorithm is executed to restore the original storage location status data records. The decompression process involves reading the compressed data stream; when a matching triplet is encountered, a byte sequence of a specified length is copied from the already decompressed historical data based on the offset value, and the next byte is appended to the output. When a single byte is encountered, it is output directly. Decompression yields multiple original storage location status data records. From each record, the storage location number field, status change identifier field, and timestamp field are extracted. The status change identifier is a Boolean value indicating whether a status change has occurred in the storage location. The extracted records are sorted in ascending order of the timestamp value. A quicksort algorithm is used, selecting the middle record as the pivot, placing records with timestamps less than the pivot on the left and records with timestamps greater than the pivot on the right, recursively sorting. The sorted record sequence represents the time-series relationship of status changes for each storage location. The sequence of changes is determined by parsing the outbound instruction message to extract a list of storage location numbers whose indicator lights should be illuminated. This list contains storage locations for which the system instructs the operator to retrieve goods. Each record in the status change sequence is iterated to check if its storage location number exists in the instruction storage location list. If the storage location number is not in the list, the storage location is considered a non-instruction storage location. The timestamp of the non-instruction storage location record is compared with the timestamps of all instruction storage location records. If the timestamp is less than the timestamp of any instruction storage location record, it is considered an abnormal operation, meaning the operator retrieved assets from a non-instruction storage location instead of following the system instruction. The storage location number, timestamp, and operation type corresponding to this abnormal operation are recorded. The operation type is read from the records to distinguish between inbound, outbound, and transfer types. All abnormal operation records are arranged chronologically to form a behavioral characteristic sequence.

[0033] For example, when a warehouse manager at a power supply station is handling outbound operations, a microwave radar sensor detects a person approaching the cabinet door. The radar emits a frequency-modulated signal and receives the reflected signal from the person. The mixed intermediate frequency signal is sampled and Fourier transformed to obtain the spectrum. The target distance is calculated from the frequency corresponding to the peak of the spectrum, which is within a certain range in front of the cabinet door. Simultaneously, the Doppler frequency shift is extracted from the spectrum changes over multiple periods to calculate the target velocity, which is within the speed range of approaching the cabinet door. The radar selects targets whose distance and velocity both match the preset detection area as valid targets, thus obtaining the target location information. A camera simultaneously captures video frames, which are input into a multi-task cascaded convolutional neural network. The first stage of the network quickly scans the image. The process involves generating several candidate windows, refining them in the second stage of the network, and retaining high-confidence windows. The third stage network precisely locates and outputs the coordinates of the face region. Based on these coordinates, the face region is cropped from the image, resized, and then input into the face recognition network. The network's convolutional and fully connected layers extract the high-dimensional feature vector of the face. The feature library of authorized users is read from the database. The Euclidean distance between the feature vector of each authorized user and the current feature vector is calculated. The Euclidean distance is the square root of the sum of squared differences in the corresponding dimensions. The minimum value among all distances is recorded. If the minimum distance exceeds a preset threshold, it indicates that the face does not belong to any authorized user, and the authorization verification fails. At this point, the distribution... The dataset queries the storage location status data within the most recent time window. The start timestamp of the time window is the current time minus the window duration. Data blocks with overlapping timestamp ranges are located from the block index table. The compressed data of these blocks is read and decompressed. During decompression, when encountering triples, the byte sequence is copied from historical data based on the offset value; single bytes are output directly. After obtaining the original records, the storage location number, status change identifier, and timestamp of each record are extracted. These are then sorted in ascending order of timestamp using a quicksort algorithm. The sorted data yields the temporal relationship of status changes. Assuming the retrieval instruction requests the assets of three single-phase table storage locations with specific storage location numbers, these three storage location numbers constitute the indicator storage location list. The table iterates through the records in the time sequence to check the storage location number. If a storage location number of a certain record is not in the instruction list, it is a non-instruction storage location. If the timestamp of this record is compared and found to be earlier than the timestamps of all instruction storage location records, the operation is determined to be an abnormal operation, that is, the operator took the asset from the non-instruction storage location first. The storage location number, timestamp and operation type of the abnormal operation are recorded as outbound. All abnormal operation records are arranged in chronological order to form a behavioral feature sequence. This sequence records the operator's illegal picking behavior. Subsequently, this sequence is associated with the corresponding grid position in the spatial distribution data. The management platform generates anomaly alarms based on the behavioral feature sequence and records them in the operator's behavior profile.

[0034] In one specific embodiment, step S4 includes: The RF reader sequentially activates each antenna in the antenna array to transmit RF signals at multiple discrete frequency points, receives the reflected signals from the asset tags, extracts the phase value corresponding to each frequency point, calculates the phase difference between adjacent frequency points, calculates the distance difference based on the phase difference and frequency point interval, and accumulates them to obtain the distance from the antenna to the tag. Based on the known spatial coordinates of each antenna in the antenna array and the distance from the antenna to the tag, a set of distance equations is established. The least squares method is used to iteratively solve the set of distance equations to obtain the three-dimensional spatial coordinates of the tag. Divide the drawer storage space into grid cells of preset size and establish a grid coordinate system. Calculate the index number of the grid cell to which the label belongs based on the three-dimensional spatial coordinates of the label, and bind the index number to the label code. Extract the operation timestamp and operation type corresponding to the current storage location from the behavioral feature sequence, and associate the operation timestamp and operation type with the index number to obtain spatial distribution data containing tag code, grid location and operation record.

[0035] Specifically, the microcontroller inside the RFID reader writes configuration parameters for antenna port selection, transmit power, and operating frequency to the RFID front-end chip via a serial interface. The value of the antenna port selection register determines which antenna in the antenna array is activated. The reader sequentially sets this register to the values ​​corresponding to the first, second, third, and fourth antennas. The transmit power register is set to the encoded value corresponding to the preset power. The operating frequency register is sequentially written with frequency words corresponding to multiple discrete frequency points. The phase-locked loop circuit of the RFID front-end chip generates a local oscillator signal of the corresponding frequency based on the frequency words. The transmit link amplifies the local oscillator signal and radiates the RFID signal through the currently activated antenna. The UHF tag affixed to the asset inside the drawer receives the RFID signal and then... The rectifier circuit of the chip converts radio frequency energy into DC voltage to power the chip. The tag chip reads the electronic tag code in the memory and modulates the code into a digital bit stream according to the communication protocol. Backscatter modulation is achieved by switching the antenna impedance between high and low impedance states. The impedance switching causes a change in the antenna reflection coefficient, thereby modulating the amplitude and phase of the reflected signal. After the reader antenna receives the reflected signal, the radio frequency front-end chip down-converts it to an intermediate frequency. The quadrature demodulator outputs two baseband signals, one in phase and one quadrature. The analog-to-digital converter samples the two signals, and the digital signal processor calculates the phase value of the reflected signal. The phase value is calculated by taking the arctangent function of the ratio of the quadrature component to the in phase component. After recording the phase value corresponding to the current frequency point, the circuit switches to the next frequency. The above measurement process is repeated for discrete frequency points. After completing the phase measurement for all frequency points, the phase value corresponding to each frequency point is obtained. The phase difference between adjacent frequency points is calculated. The phase difference is equal to the phase value of the next frequency point minus the phase value of the previous frequency point. Since the phase value cycles between negative π and positive π, when the phase difference is less than negative π, 2π is added for unwrapping; when the phase difference is greater than positive π, 2π is subtracted for unwrapping. The unwrapped phase difference reflects the actual change in phase of the radio frequency signal propagating between two frequency points. The distance difference is calculated based on the phase difference. The distance difference is calculated by multiplying the phase difference by the speed of light, dividing by the frequency point interval, and then dividing by 4π. The frequency point interval is the frequency difference between two adjacent frequency points, and the speed of light is a known constant. The distance differences calculated between all adjacent frequency points are summed to obtain the distance from the antenna to the tag. The reader sequentially switches to the second, third, and fourth antennas, repeating the multi-frequency phase measurement and distance calculation process to obtain the distance values ​​from each of the four antennas in the antenna array to the tag. The known spatial coordinates of the four antennas are read from the configuration file; these coordinates were measured and recorded during installation. The four antennas are arranged at the four corners of a rectangle within the drawer storage location. Assuming the tag's unknown spatial coordinates include three unknowns: a horizontal component, a horizontal component, and a vertical component, a system of distance equations is established based on the distance formula between two points in space. Each antenna corresponds to one equation. The left side of the equation represents the sum of the squares of the differences between the antenna coordinates and the tag coordinates in the three directions, and the right side represents the square of the measured distance from the corresponding antenna. Four equations are established for the four antennas, forming a system of equations.The least squares method is used to iteratively solve this system of nonlinear equations. The basic idea of ​​the least squares method is to find the solution that minimizes the sum of squared errors of all equations. The error function is constructed as the sum of squared differences between the left and right sides of the four equations. Partial derivatives of the error function with respect to the three coordinate components are calculated. The partial derivatives reflect the trend of the error function in each coordinate direction. The three partial derivatives form a three-dimensional gradient vector pointing in the direction of the fastest growth of the error function. The initial coordinate guess is set as the geometric center coordinates of the drawer storage position, which is half of the length, width and height of the drawer. The error function value corresponding to this initial guess is calculated. The coordinate guess is updated along the negative direction of the gradient vector. The update formula is that the new coordinates equal the old coordinates. The coordinates are subtracted from the preset learning rate multiplied by the gradient component in the corresponding direction. The learning rate controls the step size of each iteration. The gradient update process is repeated. After each iteration, the error function value is recalculated. Iteration stops when the absolute value of the difference between the error function value of the current iteration and the error function value of the previous iteration is less than the preset convergence threshold. The three coordinate components at convergence are the three-dimensional spatial coordinates of the label. The number of grids in each direction is calculated based on the length, width, and height of the drawer and the preset grid cell size. The number of grids is equal to the drawer size divided by the grid size and rounded up. A grid coordinate system is established with the lower left corner of the drawer as the origin. Each component of the label's three-dimensional spatial coordinates is divided by the grid cell size in the corresponding direction. The results are rounded down to obtain three grid indices, each indicating the grid cell in which the tag is located in each direction. These three grid indices are combined to calculate a unique index number. The index number is calculated as follows: horizontal index multiplied by the product of the total number of horizontal and vertical grid cells and the total number of vertical grid cells, plus the horizontal index multiplied by the total number of vertical grid cells and the vertical index. A record is inserted into the database table to establish a link between the tag code field and the index number field. The tag code field stores the electronic tag code, and the index number field stores the calculated index number. Records matching the current drawer storage location number are then searched from the behavioral feature sequence. The feature sequence contains records of abnormal operations by operators. From the matched records, the operation timestamp and operation type fields are extracted. The operation timestamp records the time the operation occurred, and the operation type distinguishes different operations such as inbound, outbound, and transfer. The database table is updated by adding the operation timestamp and operation type fields to existing records. The extracted operation timestamp and operation type are written to the corresponding fields and associated with the index number. After data association, each record in the database table contains four fields: tag code, grid location index number, operation timestamp, and operation type. This table represents the spatial distribution data, which records the precise grid location and related operation history of each asset tag.

[0036] For example, a drawer storage compartment contains a turnover box with several single-phase meters inside. An RF reader activates the first antenna (located in the front left corner of the drawer) using its antenna port register. The operating frequency register sequentially writes the frequency words corresponding to four discrete frequency points. A phase-locked loop generates an RF signal at the corresponding frequency, which is radiated through the first antenna. A tag on one of the single-phase meters in the turnover box receives the RF signal and performs backscatter modulation. The reader receives the reflected signal, demodulates it, and outputs in-phase and quadrature components. The phase value of the first frequency point is calculated using the arctangent function, divided by the in-phase component. The measurement is repeated at the second frequency point to obtain the phase value of the second frequency point. The first and... The phase difference at the second frequency point is equal to the phase at the second frequency point minus the phase at the first frequency point. If the phase difference is negative (less than negative π), add 2π for unwrapping. Calculate the distance difference based on the unwrapped phase difference and frequency interval: the distance difference is equal to the phase difference multiplied by the speed of light, divided by the frequency interval, and then divided by 4π. Calculate the distance differences between the second and third frequency points, and between the third and fourth frequency points sequentially. Summing these three distance differences gives the total distance from the first antenna to the tag. The reader switches the antenna port register to activate the second antenna, located at the front right corner of the drawer. Repeat the multi-frequency measurement and distance calculation to obtain the distance from the second antenna to the tag. Activate the third antenna, located at the rear left corner of the drawer, and the fourth antenna, located at the rear right corner of the drawer, to complete the measurement. The distances from the four antennas to the tag are obtained. The spatial coordinates of the four antennas are read from the configuration file. The three components of the tag coordinates are set as unknowns. Four equations are established based on the principle that the sum of the squares of the differences between each antenna coordinate and the tag coordinate equals the square of the corresponding distance. The error function is constructed as the sum of the squares of the differences on both sides of the four equations. The partial derivatives of the error function with respect to the three coordinate components are used to obtain the gradient vector. The initial coordinate guesses are set as the center coordinates of the drawer (half the length, half the width, and half the height). The coordinates are updated along the negative gradient direction. The new coordinates are equal to the old coordinates minus the learning rate multiplied by the gradient component. This iterative update continues until the change in the error function value is less than the convergence threshold. After convergence... Obtain the three-dimensional spatial coordinates of the label. Calculate the number of grids based on the drawer's length, width, height, and grid size. Divide each component of the label coordinates by the grid size and round down to obtain three grid indices. Combine the grid indices to calculate the index number, which is equal to the horizontal index multiplied by the vertical grid number multiplied by the height grid number plus the vertical index multiplied by the height grid number plus the height index. Insert a record into the database, binding the label code to the index number. Search for records matching the drawer location from the behavioral feature sequence, extract the operation timestamp and operation type, and update the database record by adding these two fields, forming a spatial distribution data record containing the label code, grid location index number, operation timestamp, and operation type.

[0037] In one specific embodiment, a set of distance equations is established based on the known spatial coordinates of each antenna in the antenna array and the distance from the antenna to the tag. The least squares method is used to iteratively solve the set of distance equations to obtain the three-dimensional spatial coordinates of the tag, including: Obtain the known spatial coordinates of each antenna in the antenna array. Based on the fact that the sum of the squares of the differences between the spatial coordinates to be determined of the tag and the known spatial coordinates of each antenna is equal to the square of the distance from the corresponding antenna to the tag, establish a set of distance equations containing multiple antennas. The error function is constructed as the sum of squares of the differences between the measured distance and the theoretical calculated distance of each antenna. The partial derivatives of the error function with respect to the horizontal coordinate, the horizontal longitudinal coordinate, and the vertical height coordinate are obtained to obtain the three-dimensional gradient vector. The initial coordinates are set as the geometric center coordinates of the drawer storage position. Each coordinate component is updated along the negative direction of the three-dimensional gradient vector. Each update is the product of the preset learning rate and the corresponding gradient component. The update is repeated iteratively until the change in the error function is less than the convergence threshold. The three converged coordinate components are used as the three-dimensional spatial coordinates of the label.

[0038] Specifically, the microcontroller of the RF reader configures the control register of the RF front-end chip through a serial interface, sequentially setting the antenna port selection, transmit power, and operating frequency parameters. The antenna port selection register determines which antenna in the antenna array is activated, and the reader activates each antenna in sequence. The operating frequency register writes frequency words for multiple discrete frequency points. The phase-locked loop of the RF front-end chip generates the corresponding local oscillator signal based on the frequency words. After amplification by the transmit link, the RF signal is radiated through the activated antenna. The UHF tag on the asset in the drawer receives the RF signal, and the rectifier circuit converts the RF energy into DC voltage for power supply. The tag chip reads the electronic tag code in the memory and modulates it into a digital bit stream according to the communication protocol. The antenna impedance is switched between high-impedance and low-impedance states. Backscatter modulation is achieved through inter-frequency switching. Impedance switching causes a change in the antenna reflection coefficient, thus modulating the reflected signal. After the reader receives the reflected signal, the RF front-end chip down-converts it to an intermediate frequency. The quadrature demodulator outputs two signals: an in-phase component and a quadrature component. After sampling by the analog-to-digital converter, the digital signal processor calculates the phase value. The phase value is the arctangent of the quadrature component divided by the in-phase component. After recording the phase value of the current frequency point, the measurement is repeated at the next frequency point. After completing the measurement of all frequency points, the phase difference between adjacent frequency points is calculated. The phase difference is equal to the phase of the next frequency point minus the phase of the previous frequency point. Since the phase value cycles between negative π and positive π, when the phase difference is less than negative π, 2π is added for unwrapping; when the phase difference is greater than positive π, 2π is subtracted for unwrapping. The unwrapped phase difference reflects the propagation phase of the RF signal. The actual changes are calculated based on the phase difference. The distance difference is calculated by multiplying the phase difference by the speed of light, dividing by the frequency interval, and then dividing by 4π. The frequency interval is the frequency difference between adjacent frequency points. The total distance from the antenna to the tag is obtained by summing the distance differences of all adjacent frequency points. The reader switches to other antennas sequentially to repeat the multi-frequency measurement and distance calculation to obtain the distance value from each antenna to the tag. The spatial coordinates of each antenna are read from the configuration file. The antennas are arranged in the drawer storage space with the four corners of a rectangle. Assume that the spatial coordinates of the tag to be calculated contain three components: horizontal, vertical, and vertical. A system of equations is established based on the distance formula between two points in space. Each antenna corresponds to one equation. The left side of the equation is the sum of the squares of the differences between the antenna coordinates and the tag coordinates in the three directions, and the right side of the equation is the sum of the squares of the measured distances of the corresponding antennas. The equations are structured into a system and solved iteratively using the least squares method. The error function is constructed as the sum of squares of the differences between the left and right sides of each equation. Partial derivatives of the error function with respect to the three coordinate components are calculated. These partial derivatives reflect the changing trend of the error function in each coordinate direction. The three partial derivatives form the gradient vector. The initial coordinate guesses are set as the coordinates of the geometric center of the drawer. The coordinates are updated along the negative gradient direction using the formula: new coordinates equal old coordinates minus the learning rate multiplied by the gradient component. This iteration is repeated until the change in the error function value is less than the convergence threshold. The coordinate components at convergence are the 3D spatial coordinates of the label. The number of grids in each direction is calculated based on the drawer size and grid cell size. The number of grids equals the drawer size divided by the grid size and rounded up. A grid coordinate system is established with the lower left corner of the drawer as the origin.Each component of the label's spatial coordinates is divided by the corresponding grid size and rounded down to obtain three grid indices. These grid indices indicate the grid cell number in each direction where the label is located. A unique index number is calculated by combining these three grid indices: the horizontal index multiplied by the product of the total number of vertical and height grid cells, plus the vertical index multiplied by the total height grid cells, plus the height index. A record is inserted into the database to establish a link between the label code and the index number. Records matching the current drawer's storage location number are searched from the behavioral feature sequence. The operation timestamp and operation type are extracted, and the database record is updated with these two fields. After the linking is complete, each record in the database table contains the label code, grid location index number, operation timestamp, and operation type. This table represents spatial distribution data.

[0039] For example, a drawer storage compartment contains a tote box with several single-phase meters. The reader activates the first antenna located at the front left corner of the drawer. The operating frequency register sequentially writes frequency words for multiple discrete frequency points. The phase-locked loop generates corresponding frequency radio frequency signals, which are radiated through the antenna. A tag of a single-phase meter inside the tote box receives the signal and performs backscatter modulation. The reader receives the reflected signal, demodulates it, and outputs in-phase and quadrature components. The phase value of the first frequency point is calculated as the arctangent function divided by the quadrature function. Switching to the second frequency point, the phase value of the second frequency point is measured, and the phase difference between the first and second frequency points is calculated. The distance difference is equal to the phase difference of the second frequency point minus the phase difference of the first frequency point. If the phase difference is negative (less than negative π), add 2π for unwrapping. Calculate the distance difference based on the unwrapped phase difference and frequency interval: equal to the phase difference multiplied by the speed of light, divided by the frequency interval, and then divided by 4π. Calculate the distance differences of other adjacent frequency points sequentially, and sum them to obtain the total distance from the first antenna to the tag. Switch and activate the second antenna located at the front right corner of the drawer. Repeat the measurement to obtain the distance from the second antenna to the tag. Activate the third antenna located at the rear left corner and the fourth antenna located at the rear right corner sequentially to obtain the distances from each of the four antennas to the tag. (From the configuration...) The file reads the spatial coordinates of four antennas. Assuming the three components of the tag coordinates are unknowns, four equations are established based on the sum of the squares of the differences between each antenna coordinate and the tag coordinate, which equals the square of the corresponding distance. An error function is constructed as the sum of the squares of the differences on both sides of these four equations. The gradient vector is obtained by taking the partial derivatives of the error function with respect to the three coordinate components. The initial coordinates are set as the center coordinates of the drawer. The coordinates are updated along the negative gradient direction by subtracting the learning rate from the old coordinates and multiplying by the gradient component. This process is iterated until the change in the error function value is less than the convergence threshold, yielding the tag's three-dimensional spatial coordinates. The number of grids is calculated based on the drawer size and grid size. Each component of the tag coordinates is divided by the grid size and rounded down to obtain the grid index. The combined index number is calculated as the horizontal index multiplied by the vertical grid number multiplied by the height grid number plus the vertical index multiplied by the height grid number plus the height index. A record is inserted into the database, binding the tag code to the index number. Records matching the drawer are searched from the behavioral feature sequence to extract the operation timestamp and operation type. These two fields are added to the updated database record, forming a spatial distribution data record containing the tag code, grid position index number, operation timestamp, and operation type.

[0040] In one specific embodiment, step S5 includes: Extract the tag code of each asset from the spatial distribution data, query the verification date and asset type corresponding to each tag code, determine the corresponding verification cycle standard value according to the asset type, calculate the difference between the current timestamp and the verification date to obtain the used cycle, and subtract the used cycle from the verification cycle standard value to obtain the remaining verification cycle. Operation records for each asset are extracted from spatial distribution data. The number of outbound transactions and the duration of inbound transactions within a preset time period are counted. The ratio of the total inbound duration to the total length of the time period is calculated to obtain the inbound duration ratio. A time-series feature vector containing time and frequency dimensions is constructed by combining the number of outbound transactions and the inbound duration ratio. The temporal feature vector is input into the long short-term memory network model. A test period constraint term is added to the loss function of the long short-term memory network model. The test period constraint term is a penalty term for the predicted remaining period exceeding the test period standard value. The predicted remaining period and future turnover probability are obtained through forward propagation. Assets with a predicted remaining period less than the warning period threshold and a future turnover probability less than the stagnation probability threshold are selected as risk assets. The tag codes and grid locations of the risk assets are extracted to generate warning information containing asset identification and location information.

[0041] Specifically, the system queries the label code field of all records in the spatial distribution data database table. Based on the label code, it queries the corresponding verification date field and asset type field in the asset database. The verification date field is stored as a timestamp to record the date the asset was last verified. The asset type field is stored as an enumeration value to distinguish categories such as single-phase meters, three-phase meters, low-voltage transformers, and concentrators. The system queries the configuration table for the corresponding verification cycle standard value based on the asset type. The verification cycle standard value is stored in days. The system reads the current timestamp, calculates the time difference in seconds by subtracting the verification date timestamp from the current timestamp, divides the time difference by 86400 to convert it to days to obtain the used cycle. The verification cycle standard value is then subtracted from the used cycle. The remaining verification period is obtained using the cycle. The operation record field for each asset is read from the spatial distribution data table. Records with the operation type field set to "outbound" are filtered, and the number of these records within a preset time period is counted to obtain the outbound count. Records with the operation type field set to "inbound" are grouped by asset code. Each group is sorted by operation timestamp, and adjacent inbound and outbound records are found. The outbound timestamp is subtracted from the inbound timestamp to obtain the single inbound duration. All single inbound durations are summed to obtain the total inbound duration. The total inbound duration is divided by the preset time period length to obtain the inbound duration percentage. A time-series feature vector is constructed, containing multiple dimensions. The first dimension is the current inbound status. The spatial distribution data table is queried to determine if the current asset has a record. If it does, then... The first dimension is 1, otherwise 0. The second dimension is the number of times the item was taken out of the warehouse. The third dimension is the percentage of time the item was in the warehouse. The fourth dimension is the day of the week for the current date, converted to an integer code from 1 to 7. The fifth dimension is the day of the month for the current date, converted to an integer code from 1 to 31. The temporal feature vector is used as the input to the Long Short-Term Memory (LSTM) network model. The LSM network model includes an input layer, a first LSTM layer, a second LSTM layer, a fully connected layer, and an output layer. The input layer receives the temporal feature vector and passes it to the first LSTM layer. The first LSTM layer contains multiple LSTM units, each maintaining a hidden state vector and a cell state vector. The forget gate controls how many old cell states are retained. The forget gate uses a sigmoid activation function. The forgetting coefficient is calculated and selectively forgotten by multiplying it element-wise with the old cell state. An input gate controls the amount of new input received, also using a sigmoid activation function. An output gate controls the amount of hidden states output. The output of the first LSTM layer serves as the input to the second LSTM layer, which also processes sequence information using a gating mechanism. The hidden state of the second LSTM layer at the last time step is input to a fully connected layer. This fully connected layer undergoes a linear transformation using the weight matrix and bias vector, followed by an activation function. The output layer contains two neurons: the first neuron outputs the predicted remaining cycle time, and the second neuron outputs the future transition probability. During model training, a loss function is defined, comprising two parts.The first part calculates the mean squared error (MSE) by summing the squares of the difference between the predicted and actual remaining periods and the squares of the difference between the predicted and actual turnover probabilities. The second part is the verification period constraint term. This constraint term is calculated as the square of the excess period if the predicted remaining period exceeds the verification period standard value; otherwise, it is zero. The total loss function is the MSE plus the verification period constraint term multiplied by the weight coefficient. The network weights are updated using backpropagation and a gradient descent optimizer. The trained model is then used for inference. Forward propagation is performed on the input temporal feature vector. The input layer passes the feature vector to the first LSTM layer. The first LSTM layer calculates the hidden state at each time step. The second LSTM layer receives the first... The output of the LSTM layer continues to be processed. The fully connected layer performs a linear transformation and non-linear activation on the hidden state of the last time step of the second LSTM layer. The two neurons in the output layer output the predicted remaining cycle and future turnover probability values, respectively. Warning cycle thresholds and stagnation probability thresholds are set. All assets are traversed to filter those with a predicted remaining cycle less than the warning cycle threshold and a future turnover probability less than the stagnation probability threshold. These assets face the risk of an impending expiration of their inspection period but a low probability of being released in the near future. The tag codes and grid location index numbers corresponding to these risky assets are read from the spatial distribution data table to generate warning information containing the asset identifier (tag code), location information (grid location index number), predicted remaining cycle value, and future turnover probability value.

[0042] For example, a power supply station's warehouse stores several single-phase meters and three-phase meters. The spatial distribution data table is used to query all tag codes. Based on the tag code of a particular single-phase meter, the asset database is searched to find its calibration date is a specific timestamp, and its asset type is single-phase meter. The configuration table shows the standard calibration period for single-phase meters is 2920 days. The current system timestamp is read, and the time difference is calculated by subtracting the calibration date timestamp from the current timestamp. This time difference is divided by 86400 to convert to days, yielding the used period. Subtracting the used period from the standard calibration period gives the remaining calibration period for the single-phase meter. The operation records for this single-phase meter are then read from the spatial distribution data table and filtered... For records with the operation type "outbound," count the number of outbound records over the past 180 days to obtain the outbound frequency. For records with the operation type "inbound," sort by timestamp, find adjacent inbound and outbound records, and calculate the single inbound duration by subtracting the inbound timestamp from the outbound timestamp for each pair. Summate all single inbound durations to obtain the total inbound duration. Divide the total inbound duration by 180 days to convert to seconds to obtain the inbound duration percentage. Construct a time-series feature vector with the first dimension being 1 (current inbound), the second dimension being the outbound frequency, the third dimension being the inbound duration percentage, the fourth dimension being the current day of the week, and the fifth dimension being the current day of the month. Input the length of the time-series feature vector. In the time-phase memory network model, the input layer is passed to the first LSTM layer. The LSTM units in the first LSTM layer retain the old cell state through the forgetting gate. The forgetting gate calculates the sigmoid activation function and outputs the forgetting coefficient. The forgetting coefficient is multiplied element-wise with the old cell state. The input gate controls the reception of new input, and the output gate controls the output of the hidden state. The output of the first LSTM layer is passed to the second LSTM layer for further processing. Finally, the hidden state at the last time step is input to the fully connected layer, which performs linear transformation and activation. The first neuron in the output layer outputs the predicted remaining period of the single phase table, and the second neuron outputs the future flow. The system sets a warning period threshold and a stagnation probability threshold. It determines whether the predicted remaining period of the single-phase meter is less than the warning threshold and whether the circulation probability is less than the stagnation threshold. If both conditions are met, the single-phase meter is marked as a risk asset. The system reads the tag code and grid location index number of the single-phase meter from the spatial distribution data table and generates a warning message containing the tag code, grid location index number, predicted remaining period, and circulation probability of the single-phase meter. After receiving the warning message, the management platform sets the storage location indicator light of the single-phase meter to purple to indicate priority for outbound delivery. When there is a subsequent outbound task, the storage location of the single-phase meter is given priority to light up a green light to guide the operator to pick up the goods.

[0043] Figure 2 This is a schematic diagram illustrating the prediction of the remaining asset verification period and risk screening in the embodiments of this application; Figure 2This figure illustrates the prediction results of the remaining verification period for 30 power metering assets and the risk asset screening process using the Long Short-Term Memory (LSTM) network model in this embodiment. The horizontal axis represents asset numbers (1-30), and the vertical axis represents the remaining verification period (in days). Solid lines connecting hollow dots represent the remaining verification period values ​​predicted by the LSTM network with fused period constraints for each asset. The horizontal dashed line represents the warning period threshold (set to 30 days). As can be seen from the figure, the predicted remaining periods for assets numbered 1-8 are all below the 30-day warning threshold. These assets will be marked as high-risk assets by the system, and warning information will be generated, prompting management personnel to prioritize verification or removal from inventory. Assets numbered 9-20 are in the medium-risk range. Assets numbered 21-30 have sufficient remaining verification periods and are considered low-risk. This figure visually demonstrates the proactive prediction and risk screening functions implemented by the LSTM model in this application, verifying the technological upgrade effect from passive response to proactive warning.

[0044] The intelligent monitoring method for the intelligent turnover cabinet in the embodiments of this application has been described above. The intelligent monitoring system for the intelligent turnover cabinet in the embodiments of this application is described below. Please refer to [link / reference]. Figure 3 One embodiment of the intelligent monitoring system for the intelligent turnover cabinet in this application includes: The acquisition module is used to acquire photoelectric signals and identification signals of the storage location, and to perform sensor cross-verification based on the fluctuation characteristics of the photoelectric signals and the validity of the identification signals to obtain storage location status data. The receiving module is used to receive the storage space status data from the edge node and trigger reporting based on the status change. When a communication interruption is detected, an election mechanism is started and the storage space status data is stored in the cache to obtain a distributed dataset. The extraction module is used to obtain the target's location information and combine it with biometric distance to determine permissions. When the permission verification fails, it extracts the temporal relationship of state changes from the distributed dataset to obtain a behavioral feature sequence. The association module is used to measure the tag phase difference at multiple frequency points through the antenna array and unwrap it to obtain the distance, calculate the spatial position based on triangulation and map it to the grid coordinate system, and associate the behavioral feature sequence to obtain spatial distribution data; The input module is used to extract the verification period attribute and circulation history from the spatial distribution data to construct a time-series feature vector, input the long short-term memory network with fused period constraints to output the remaining period and circulation probability, screen risky assets, and obtain early warning information.

[0045] above Figure 3 The intelligent monitoring system of the intelligent turnover cabinet in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The intelligent monitoring device of the intelligent turnover cabinet in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0046] Reference Figure 4 This invention also provides an intelligent monitoring device for an intelligent turnover cabinet. This intelligent monitoring device can be a server, and its internal structure can be as follows: Figure 4 As shown, the intelligent monitoring device of this intelligent turnover cabinet includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the intelligent monitoring device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the intelligent monitoring device stores the data corresponding to this embodiment. The network interface of the intelligent monitoring device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0047] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the intelligent monitoring device of the intelligent turnover cabinet to which the present invention is applied.

[0048] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the intelligent monitoring method of the intelligent turnover cabinet.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an intelligent monitoring device (which may be a personal computer, server, or network device, etc.) of an intelligent turnover cabinet to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligently monitoring an intelligent turnover cabinet, characterized in that, The method comprises: Step S1: Obtain the photoelectric signal and identification signal of the storage location, and perform sensor mutual verification according to the fluctuation characteristics of the photoelectric signal and the validity of the identification signal to obtain storage location state data; Step S2: The edge node receives the storage location state data and triggers reporting based on state changes, starts the election mechanism when communication interruption is detected, and stores the storage location state data to the cache to obtain a distributed data set; Step S3: Obtain the location information of the target and combine the biological feature distance to determine the authority, and extract the state change time sequence from the distributed data set when the authority verification fails to obtain a behavior feature sequence; Step S4: Measure the phase difference of the tag at multiple frequency points through an antenna array and get the distance by unwrapping, calculate the spatial position based on triangular positioning and map to the grid coordinate system, associate the behavior feature sequence, and obtain spatial distribution data; Step S5: Extract the test cycle attribute and flow history from the spatial distribution data to construct a time sequence feature vector, input the long short-term memory network fused with cycle constraints to output the remaining cycle and flow probability, and screen risk assets to obtain early warning information.

2. The method of claim 1, wherein, The step S1 comprises: An infrared blocking signal of the storage location is collected by an optoelectronic sensor, a voltage sampling sequence is obtained by continuously sampling the infrared blocking signal, the maximum voltage value, the minimum voltage value and the average voltage value are calculated based on the voltage sampling sequence, and the voltage fluctuation coefficient is obtained according to the ratio of the difference value of the maximum voltage value, the minimum voltage value and the average voltage value to the average value; A bar code scanning module scans the bar code of the asset in the storage location, and records a scanning failure identifier when scanning times out without returning a valid character code, and determines the sensor abnormal type according to the threshold comparison result of the scanning failure identifier and the voltage fluctuation coefficient; When the voltage fluctuation coefficient is less than a stable threshold value and the scanning failure identifier exists, a tag reading instruction is sent to a radio frequency reader, a reflection signal of an asset electronic tag is received through a radio frequency antenna, and an electronic tag code is obtained by demodulation; The electronic tag code or bar code scanning result is bound with the storage location number, and the in-place detection state of the optoelectronic sensor is combined to obtain storage location state data containing the storage location number, asset code and in-place state.

3. The method of claim 1, wherein, The step S2 comprises: The edge node receives the storage location state data through a field bus, detects the in-place state change identifier in the storage location state data, encapsulates the storage location state data as a data packet and triggers reporting when the in-place state change identifier is detected; The edge node detects the communication state with the center node by sending a heartbeat packet at a fixed time and listening to the response packet, and determines that the communication is interrupted when no response packet is received for a continuous preset number of times, and records the starting time stamp of the communication interruption; When the communication interruption is determined, the edge nodes compare the priorities by exchanging node identifiers, and select the edge node with the smallest identifier value as the temporary master node, and the temporary master node takes over the data aggregation function; The storage location state data is written into the ring buffer area of the non-volatile memory in time stamp order, and the storage location state data is stored after compression encoding to obtain a distributed data set containing offline time period records.

4. The method of claim 1, wherein, Step S3 includes: The system transmits a linear frequency modulated continuous wave signal using a microwave radar sensor and receives the reflected signal. The reflected signal is then subjected to a Fourier transform to extract the target distance and velocity. Valid targets within a preset area are then selected to obtain the target location information. The system acquires video frames and performs face detection to obtain face regions. It then extracts feature vectors from these face regions and calculates the Euclidean distance between the feature vectors and feature vectors in a pre-stored feature library. If the minimum Euclidean distance exceeds a distance threshold, the system determines that the permission verification has failed. When the permission verification fails, the storage location status data within the time window is read from the distributed dataset, the status change identifier and timestamp of each storage location are extracted and sorted by timestamp to obtain the temporal relationship of status change. Based on the temporal relationship of the state changes, identify abnormal operations where the change timestamp of the non-indication storage location is earlier than the change timestamp of the indication storage location, record the storage location number and timestamp of the abnormal operation, and obtain the behavioral feature sequence.

5. The method of claim 1, wherein, Step S4 includes: The RF reader sequentially activates each antenna in the antenna array to transmit RF signals at multiple discrete frequency points, receives the reflected signals from the asset tag, extracts the phase value corresponding to each frequency point, calculates the phase difference between adjacent frequency points, calculates the distance difference based on the phase difference and frequency point interval, and accumulates them to obtain the distance from the antenna to the tag. Based on the known spatial coordinates of each antenna in the antenna array and the distance from the antenna to the tag, a set of distance equations is established. The least squares method is used to iteratively solve the set of distance equations to obtain the three-dimensional spatial coordinates of the tag. The drawer storage space is divided into grid cells of a preset size and a grid coordinate system is established. The index number of the grid cell to which the label belongs is calculated based on the three-dimensional spatial coordinates of the label, and the index number is bound to the label code. Extract the operation timestamp and operation type corresponding to the current storage location from the behavioral feature sequence, and associate the operation timestamp and operation type with the index number to obtain spatial distribution data containing tag code, grid position and operation record.

6. The method of claim 5, wherein, The process involves establishing a set of distance equations based on the known spatial coordinates of each antenna in the antenna array and the distance from the antenna to the tag, and then iteratively solving the set of distance equations using the least squares method to obtain the three-dimensional spatial coordinates of the tag, including: Obtain the known spatial coordinates of each antenna in the antenna array. Based on the fact that the sum of the squares of the differences between the spatial coordinates to be determined of the tag and the known spatial coordinates of each antenna is equal to the square of the distance from the corresponding antenna to the tag, establish a set of distance equations containing multiple antennas. An error function is constructed as the sum of squares of the differences between the measured distance and the theoretical calculated distance of each antenna. The partial derivatives of the error function with respect to the horizontal coordinate, the horizontal longitudinal coordinate, and the vertical height coordinate are calculated to obtain the three-dimensional gradient vector. The initial coordinates are set as the geometric center coordinates of the drawer storage position. Each coordinate component is updated along the negative direction of the three-dimensional gradient vector. Each update is the product of the preset learning rate and the corresponding gradient component. The update is repeated iteratively until the change in the error function is less than the convergence threshold. The three converged coordinate components are used as the three-dimensional spatial coordinates of the label.

7. The method of claim 1, wherein, Step S5 includes: Extract the tag code of each asset from the spatial distribution data, query the verification date and asset type corresponding to each tag code, determine the corresponding verification cycle standard value according to the asset type, calculate the difference between the current timestamp and the verification date to obtain the used cycle, and subtract the used cycle from the verification cycle standard value to obtain the remaining verification cycle. Operation records of each asset are extracted from the spatial distribution data. The number of outbound transactions and the duration of inbound transactions within a preset time period are counted. The ratio of the total inbound duration to the total length of the time period is calculated to obtain the inbound duration ratio. A time-series feature vector containing time and frequency dimensions is constructed by combining the number of outbound transactions and the inbound duration ratio. The time-series feature vector is input into the Long Short-Term Memory (LSTM) network model. A test period constraint term is added to the loss function of the LTM network model. The test period constraint term is a penalty term for the predicted remaining period exceeding the test period standard value. The predicted remaining period and future turnover probability are obtained through forward propagation. Assets whose predicted remaining period is less than the warning period threshold and whose future turnover probability is less than the stagnation probability threshold are selected as risk assets. The tag codes and grid positions of the risk assets are extracted to generate warning information containing asset identification and location information.

8. An intelligent monitoring system for an intelligent turnover cabinet, characterized in that, The intelligent monitoring system for the intelligent turnover cabinet as described in any one of claims 1-7 comprises: The acquisition module is used to acquire photoelectric signals and identification signals of the storage location, and to perform sensor cross-verification based on the fluctuation characteristics of the photoelectric signals and the validity of the identification signals to obtain storage location status data. The receiving module is used to receive the storage space status data from the edge node and trigger reporting based on the status change. When a communication interruption is detected, an election mechanism is started and the storage space status data is stored in the cache to obtain a distributed dataset. The extraction module is used to obtain the target's location information and combine it with biometric distance to determine permissions. When the permission verification fails, it extracts the temporal relationship of state changes from the distributed dataset to obtain a behavioral feature sequence. The association module is used to measure the tag phase difference at multiple frequency points through the antenna array and unwrap it to obtain the distance, calculate the spatial position based on triangulation and map it to the grid coordinate system, and associate the behavioral feature sequence to obtain spatial distribution data; The input module is used to extract the verification period attribute and circulation history from the spatial distribution data to construct a time-series feature vector, input the long short-term memory network with fused period constraints to output the remaining period and circulation probability, screen risky assets, and obtain early warning information.

9. A smart monitoring device of a smart turnover cabinet, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the intelligent monitoring method of the intelligent turnover cabinet according to any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is run by the processor, it causes the processor to execute the intelligent monitoring method for the intelligent turnover cabinet as described in any one of claims 1 to 7.

Images