Aviation lubricating oil access management method, device and equipment and storage medium

Abstract

The invention discloses an aviation lubricating oil access management method, device and equipment and a storage medium, and relates to the technical field of image processing. The two long-standing core problems that the manual operation efficiency is low and the material account consistency is difficult to guarantee in the aviation lubricating oil management field are solved. A traditional management mode depends on manual verification and paper recording, and has the defects of response delay, error proneness and incapability of real-time monitoring, however, according to the scheme, through cooperation of a weighing sensor of the lubricating oil supply cabinet and a computer vision algorithm, automatic tracking of the whole lubricating oil storing and taking process is achieved. And the risk of property account inconsistency caused by human negligence in traditional management is avoided. In addition, a face recognition authorization mechanism ensures strict management and control of the operation authority, and the compliance requirement of the aviation maintenance field for high-danger article management is met from the technical level.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus, device and storage medium for the storage and management of aviation lubricating oil. Background Technology

[0002] The management of aviation lubricating oil has always been a major challenge in the aviation industry due to its special characteristics, high risks, and strict application requirements. The management of aviation lubricating oil by aviation maintenance departments is directly related to flight safety, maintenance efficiency, and environmental compliance.

[0003] Currently, the management of aviation lubricating oil is primarily based on a designated storage and manual management system. This involves storing the lubricating oil on shelves or in cabinets in a designated warehouse. When maintenance personnel need to use the oil, they first contact the management personnel, then go to the designated warehouse together to verify the name, part number, expiration date, and quantity of the lubricating oil to be borrowed. A borrowing registration is then completed. Upon return, the management personnel must verify the consistency between the inventory and the records and sign for confirmation. For aviation maintenance departments with complete manuals and strict management, dedicated personnel will be assigned to manage aviation lubricating oil in the same way.

[0004] However, in order to save costs, most maintenance departments often cannot assign dedicated personnel to specific positions. Instead, the management is usually handled part-time by on-duty aircraft parts personnel, tool room personnel, or aircraft maintenance personnel. Manual management has high time and labor costs and is prone to errors. In particular, part-time management is more likely to result in unclear division of labor, overlapping positions, and absenteeism, leading to low efficiency in the management of aviation lubricating oil. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, equipment, and storage medium for the storage and management of aviation lubricating oil, so as to solve the problem of low management efficiency of aviation lubricating oil in the prior art.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A method for storing and managing aviation lubricating oil, wherein the method is applied to several lubricating oil tanks of the same specifications placed in a lubricating oil supply cabinet, the lubricating oil supply cabinet having a weighing function, and all lubricating oil tanks being in at least one state of being packed in cartons or in bulk, the method comprising:

[0008] Step S1: In response to the request from an external worker to open the lubricating oil supply cabinet, the FaceNet deep convolutional neural network is used to extract facial features, and authorization information matching the facial features is retrieved from an external database.

[0009] Step S2: If the authorization information exists, the lubricating oil supply cabinet is opened. After the external staff completes the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the whole package / bulk package combination.

[0010] Step S3: Send the optimal solution for extraction to the external staff. After the external staff completes the extraction operation, the bottle mouth features of the lubricating oil can are identified by OpenCV Hough circle detection, the corner points of the packaged carton are identified by YOLO algorithm, and the three-dimensional geometric features of the opened carton are identified by Mask R-CNN instance segmentation model to obtain the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons.

[0011] Step S4: Construct a pre-trained Siamese neural network contrastive model that includes visual and weight branches and shares convolutional kernel parameters;

[0012] Step S5: Input the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons into the visual branch in the form of feature vectors, and input the weighing data into the weight branch in the form of time-series data;

[0013] Step S6: The cosine similarity score between the visual branch and the weight branch is output through the twin neural network comparison model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated.

[0014] As a further improvement to this application, step S6 involves outputting the cosine similarity score between the visual branch and the weight branch through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated. Following this, the method further includes:

[0015] Step S10: The number of returned empty cans, the number of packaged cartons, the number of bulk cartons, the weighing data, and the cosine similarity score are uploaded to the consortium blockchain through different dedicated channels.

[0016] Step S20: Associate all dedicated channels with the composite key DataType_NodeID_Timestamp, and construct the access port of the consortium blockchain by periodically replacing consensus nodes based on the RBFT consensus mechanism;

[0017] Step S30: Create smart contract code on the blockchain using Solidity to manage the access port;

[0018] Step S40: Deploy the zk-SNARK protocol on the blockchain and encrypt the smart contract code based on zero-knowledge proof privacy protection, so that the smart contract code is only open to authorized identities of external regulatory authorities;

[0019] Step S50: The smart contract code is called through a preset Python script, and the newly added number of empty cans returned, the number of fully packed cartons, the number of loose cartons, the weighing data, and the cosine similarity score are uploaded to the blockchain based on the authorized identity.

[0020] As a further improvement to this application, step S6 involves outputting the cosine similarity score between the visual branch and the weight branch through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated. Following this, the method further includes:

[0021] Step S100: Extract several historical oil consumption data based on the lubricating oil supply cabinet. Each historical maintenance data includes the model code of each maintenance, the amount of lubricating oil used, the maintenance duration, and seasonal parameters.

[0022] Step S200: The machine model code, lubricating oil consumption, maintenance duration, and seasonal parameters for each maintenance are aggregated according to the matching maintenance event ID using a time series alignment algorithm to obtain a training dataset containing timestamps.

[0023] Step S300: Construct a pre-trained bidirectional LSTM network architecture, wherein the loss function of the bidirectional LSTM network architecture adopts Huber loss, the optimizer is Nadam, and the initial learning rate is set to 0.001.

[0024] Step S400: Input the training dataset into the bidirectional LSTM network architecture, and output a prediction dataset based on a preset prediction step size through the bidirectional LSTM network architecture. The prediction dataset includes predicted oil consumption data.

[0025] Step S500: Extract real-time oil consumption data based on the lubricating oil supply cabinet, and obtain the difference between the real-time oil consumption data and the predicted oil consumption data;

[0026] Step S600: When the difference exceeds the preset deviation threshold, a lubricating oil purchase request is generated and sent to the external monitoring terminal.

[0027] As a further improvement to this application, step S1, in response to an external worker's request to open the lubricating oil supply cabinet, involves extracting facial features using a FaceNet deep convolutional neural network and retrieving authorization information matching the facial features from an external database, including:

[0028] Step S11: In response to the request from an external worker to open the lubricating oil supply cabinet, acquire the facial image of the external worker using an external imaging device;

[0029] Step S12: Extract high-dimensional facial feature vectors from the facial image using the FaceNet deep convolutional neural network;

[0030] Step S13: Perform L2 normalization on the high-dimensional face feature vector to obtain a standardized feature vector;

[0031] Step S14: Obtain the image confidence of the standardized feature vector using the MTCNN face detection algorithm;

[0032] Step S15: When the image confidence exceeds a preset confidence threshold, the standardized feature vector is transmitted to the external database.

[0033] Step S16: Calculate the similarity between the standardized feature vector and all pre-stored facial records in the external database;

[0034] Step S17: When the similarity is greater than a preset similarity threshold, retrieve the authorization information that matches the facial features from an external database.

[0035] As a further improvement to this application, in step S2, if the authorization information exists, the lubricating oil supply cabinet is opened. After the external staff completes the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the complete / bulk combination, including:

[0036] Step S21: If the authorization information exists, then open the lubricating oil supply cabinet;

[0037] Step S22: After the external staff completes the storage operation, the weighing function collects weight data for a preset continuous duration at a preset collection frequency.

[0038] Step S23: Extract the weight change rate features of all weight data using the first-order difference method to obtain a three-dimensional time-series feature matrix including the original weight, the amount of change, and the acceleration of change.

[0039] Step S24: Obtain the historical access data of the lubricating oil supply cabinet based on several historical moments and divide it into a sample set and a verification set;

[0040] Step S25: Construct the gradient boosting decision tree model based on the XGBoost framework, and define the objective variable of the gradient boosting decision tree model as taking the optimal solution, wherein taking the optimal solution is to maximize the number of fully packaged cartons and minimize the number of loose cartons;

[0041] Step S26: Input the sample set into the gradient boosting decision tree model and select the best splitting feature for each decision tree using a greedy algorithm. The best splitting feature of the current decision tree includes the score of the current decision tree based on each complete / disassembled combination.

[0042] Step S27: Obtain the F1-score performance index of each best splitting feature based on the validation set, and define that the weight of each decision tree is proportional to its respective F1-score performance index;

[0043] Step S28: Calculate the weighted total score of all decision trees based on each package / bulk combination, according to the scores and weights of all decision trees.

[0044] Step S29: Define the assembled / bulk combination with the highest weighted total score as the optimal solution to be extracted.

[0045] As a further improvement to this application, step S4 involves constructing a pre-trained Siamese neural network contrastive model that includes a visual branch and a weight branch, sharing convolutional kernel parameters, including:

[0046] Step S41: Construct a symmetric neural network structure that includes a visual branch and a weight branch, with both branches using ResNet-18 as the basic architecture.

[0047] Step S42: Initialize the weight values ​​of the two branches using the Xavier method to share the convolution kernel parameters of the two branches;

[0048] Step S43: Obtain visual features of empty lubricating oil cans, unopened lubricating oil cans, fully packaged cartons, and bulk cartons as anchor point samples;

[0049] Step S44: Weighing data from normal operation is used as positive samples, and weighing data from abnormal operation is used as negative samples.

[0050] Step S45: The anchor sample, the positive sample, and the negative sample are pre-trained using the triplet loss function to enable the symmetric neural network structure to form compact intra-class distances and dispersed inter-class distances in the feature space, thereby obtaining the Siamese neural network comparison model.

[0051] As a further improvement to this application, step S6 involves outputting the cosine similarity score between the visual branch and the weight branch through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated, including:

[0052] Step S61: Perform L2 normalization on the visual feature vector output by the visual branch and the time-series weighing feature vector output by the weight branch to obtain the standardized visual feature vector and the standardized time-series weighing feature vector in sequence.

[0053] Step S62: The standardized visual feature vector and the standardized temporal weighing feature vector are mapped to the feature space through the Siamese neural network contrast model, and the cosine similarity score is calculated using equation (1):

[0054] (1);

[0055] in, The cosine similarity score is given. The standardized visual feature vector, The standardized time-series weighing feature vector;

[0056] Step S63: When the cosine similarity score is lower than the preset score threshold, the access anomaly signal is generated.

[0057] To achieve the above objectives, this application also provides the following technical solutions:

[0058] An access management device for aviation lubricating oil, the access management device being applied to the access management method described above, the access management device comprising:

[0059] The face matching module is used to respond to the request from external staff to open the lubricating oil supply cabinet. It uses FaceNet deep convolutional neural network to extract facial features and retrieves authorization information that matches the facial features from an external database.

[0060] The optimal solution extraction module is used to open the lubricating oil supply cabinet if the authorization information exists, and after the external staff completes the storage operation, it obtains the weighing data through the weighing function and processes the weighing data based on the gradient boosting decision tree model to output the optimal extraction solution for the whole / bulk combination.

[0061] The image recognition module is used to send the optimal solution for retrieval to the external staff. After the external staff completes the retrieval operation, the module uses OpenCV Hough circle detection to identify the bottle mouth features of the lubricating oil can, YOLO algorithm to identify the corner points of the packaged carton, and Mask R-CNN instance segmentation model to identify the three-dimensional geometric features of the opened carton, thereby obtaining the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons.

[0062] The contrastive model definition module is used to construct a pre-trained Siamese neural network contrastive model that includes a visual branch and a weight branch and shares convolutional kernel parameters.

[0063] The comparison model input module is used to input the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons into the visual branch in the form of feature vectors, and to input the weighing data into the weight branch in the form of time-series data.

[0064] The comparison model output module is used to output the cosine similarity score between the visual branch and the weight branch through the Siamese neural network comparison model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated.

[0065] To achieve the above objectives, this application also provides the following technical solutions:

[0066] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the access management method described above.

[0067] To achieve the above objectives, this application also provides the following technical solutions:

[0068] A computer-readable storage medium storing program instructions that, when executed by a processor, enable the access management method described above.

[0069] Effective effects:

[0070] This application responds to an external worker's request to open the lubricating oil supply cabinet by employing a FaceNet deep convolutional neural network to extract facial features and retrieving authorization information matching the facial features from an external database. If authorization information exists, the lubricating oil supply cabinet is opened. After the external worker completes the storage operation, weighing data is obtained through a weighing function and processed using a gradient boosting decision tree model to obtain the optimal solution for retrieving the combined packaged and bulk oil. The optimal solution is sent to the external worker. After the external worker completes the retrieval operation, the bottle mouth features of the lubricating oil can are identified using OpenCV Hough circle detection, corner points of the packaged cardboard box are identified using the YOLO algorithm, and the mask... The R-CNN instance segmentation model identifies the 3D geometric features of opened cardboard boxes, obtaining the number of returned empty cans, the number of fully packed boxes, and the number of loose boxes. A pre-trained Siamese neural network contrastive model is constructed, including a visual branch and a weight branch, sharing convolutional kernel parameters. The number of returned empty cans, fully packed boxes, and loose boxes are input into the visual branch as feature vectors, and the weighing data is input into the weight branch as time-series data. The Siamese neural network contrastive model outputs the cosine similarity score between the visual branch and the weight branch. When the cosine similarity score is lower than a preset threshold, an access anomaly signal is generated.

[0071] This application addresses two core issues in aviation lubricant management—low ​​efficiency of manual operations and difficulty in ensuring consistency between inventory and records—through the integration of a multimodal intelligent sensing and decision-making system. Traditional management methods rely on manual verification and paper records, which suffer from delays, errors, and a lack of real-time monitoring. This solution, however, achieves automated tracking of the entire lubricant storage and retrieval process through the collaboration of weighing sensors in the lubricant supply cabinet and computer vision algorithms (FaceNet, YOLO, Mask R-CNN). The core technological breakthrough lies in dynamic optimal solution calculation, which analyzes weighing data using a gradient boosting decision tree model to intelligently recommend the combination of bulk / packaged lubricant retrieval, reducing manual decision-making time. Multimodal cross-validation utilizes a Siamese neural network to compare visual recognition results with weight change data; when the cosine similarity falls below a threshold, an anomaly alarm is triggered, fundamentally eliminating the risk of discrepancies between inventory and records caused by human error in traditional management. Furthermore, a facial recognition authorization mechanism ensures strict control over operating permissions, technically fulfilling the compliance requirements for managing high-risk materials in the aviation maintenance field. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the steps of one embodiment of the aviation lubricating oil storage and management method of this application;

[0073] Figure 2 This is a schematic diagram of the functional modules of one embodiment of the aviation lubricating oil storage and retrieval management device of this application;

[0074] Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0075] Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0076] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0077] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0078] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0079] like Figure 1 As shown, this embodiment provides an example of a storage and management method for aviation lubricating oil. In this embodiment, the storage and management method is applied to several lubricating oil tanks of the same specifications placed in a lubricating oil supply cabinet. The lubricating oil supply cabinet has a weighing function, and all lubricating oil tanks are in at least one state of being packed in cartons or in bulk.

[0080] Preferably, the common packaging specifications for lubricating oil cans are 946ml / can or 946ml / barrel, suitable for aircraft turbine engines; the large packaging specifications include 15KG / barrel (such as aviation hydraulic oil for ground use) and 3kg / barrel (such as aviation grease).

[0081] Preferably, the packaging method is typically 24 x 1 quart / box (24 qt / ctn) or 5 gallons / drum (5 gal / pail).

[0082] Preferably, this embodiment is mainly used for the unified management of lubricating oil tanks of the same model. For lubricating oil tanks of multiple specifications, it is recommended to use one set of storage and retrieval management methods and storage and retrieval management devices provided in this embodiment for each specification.

[0083] Specifically, the access management method includes the following steps:

[0084] Step S1: In response to the request from external staff to open the lubricating oil supply cabinet, the FaceNet deep convolutional neural network is used to extract facial features, and authorization information matching the facial features is retrieved from an external database.

[0085] Preferably, facial recognition relies on an external camera. Specifically, when an external worker triggers an opening request for the oil supply cabinet via a physical button or electronic interface, the system immediately initiates the facial recognition process. The request signal is transmitted to the control center via an IoT module, triggering the camera module to start.

[0086] The next step is to use a high-resolution infrared camera to capture facial images of staff to ensure clear imaging even in low-light conditions.

[0087] Preferably, image standardization can be achieved by converting to RGB format, adjusting to a 224×224 pixel resolution, and normalizing pixel values ​​to the range of [-1,1].

[0088] The next step is to load a pre-trained FaceNet model, such as the InceptionResNetV1 architecture, into the pre-processed image. The face features are represented by a 128-dimensional embedding vector, which is invariant to translation, rotation, and illumination.

[0089] The next step is to calculate the similarity between the collected 128-dimensional feature vector and the registered features in the authorization database.

[0090] Preferably, liveness detection technology, such as blink detection, can be used to prevent photo deception; in order to prevent misjudgment caused by long-term appearance changes, the database features need to be updated regularly to cope with changes in people's appearance.

[0091] Step S2: If authorization information exists, the lubricating oil supply cabinet is opened. After the external staff completes the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the complete / bulk combination.

[0092] Preferably, after FaceNet verification is successful, an opening command is sent to the lubricating oil supply cabinet via the CAN bus protocol. After the cabinet is opened, the weighing sensor preheating program of the weighing function is started, and the input parameters of the gradient boosting decision tree (GBDT) model are initialized at the same time.

[0093] Preferably, in the weighing data acquisition and preprocessing stage, real-time weight data is acquired through a high-precision weighing sensor and the raw data is filtered by a sliding window, for example, with a window size of 5 and a threshold of 0.2 kg to eliminate mechanical vibration interference. Then, the weight change ΔW = W_after - W_before is calculated.

[0094] Preferably, GBDT feature engineering can construct a four-dimensional feature vector including the current total weight (reflecting the inventory base), weight change (reflecting the scale of operation), time characteristics (working days / holidays), and historical operation frequency (sliding window statistics), and perform one-hot encoding on the packaged / bulk status, for example, packaged = 1, 0; bulk = 0, 1.

[0095] Preferably, the GBDT model is implemented using the XGBoost architecture, with 100 decision trees and a maximum depth of max_depth=6. The next step is to convert the weighing data into feature vectors and input them into the GBDT model. The GBDT model outputs three predicted values: the number of whole-pack suggestions (discrete value), the number of bulk suggestions (continuous value), and the operation priority score (0-1).

[0096] Preferably, the greedy algorithm described below can be used, or the NSGA-II multi-objective optimization algorithm can be used to generate the Pareto optimal solution strategy. In step S3, the optimal solution is sent to the external staff. After the external staff completes the retrieval operation, the bottle mouth features of the lubricating oil can are identified by OpenCV Hough circle detection, the corner points of the packaged carton are identified by YOLO algorithm, and the three-dimensional geometric features of the opened carton are identified by MaskR-CNN instance segmentation model to obtain the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons.

[0097] Preferably, after the external staff completes the removal operation, the multi-angle camera group with at least three viewing angles pre-installed in the lubricating oil supply cabinet is triggered to capture images simultaneously. The next step is to preprocess the images from each viewing angle: convert them to HSV color space, apply adaptive threshold binarization, perform morphological closing operations to eliminate noise, and detect circular features. It is worth noting that the detection parameters of the source features need to be adjusted according to the actual bottle opening size.

[0098] Next, a YOLO model pre-trained for cardboard box corner detection is loaded, and the number of corners is counted for each detected cardboard box. A complete cardboard box should have four visible corners. Next, a pre-trained Mask R-CNN model is fine-tuned using the COCO dataset, and the 3D geometric features of the cardboard box are calculated based on the instance segmentation results: surface area change rate, vertex displacement, and opening angle. Finally, spatial association rules are established, and the final count is performed to achieve result fusion and statistical analysis. Specifically, spatial association rules are established by matching the detection results of the same physical object from different perspectives using 3D coordinates, and the attribution relationship between the bottle opening and the cardboard box is determined using a distance threshold method.

[0099] Step S4: Construct a pre-trained Siamese neural network contrast model that includes a visual branch and a weight branch and shares convolutional kernel parameters.

[0100] Preferably, the model architecture of the Siamese neural network comparison model can employ a shared convolutional feature extractor consisting of 3 layers of convolution + ReLU + BN. Specifically, the visual branch is followed by a 3-layer fully connected network to process the bottle / carton quantity features; the weight branch is followed by an LSTM layer to process time-series weighing data. For the visual branch, the 3D features (empty cans / packaged / bulk quantities) can be flattened before processing; for the weight branch, time-series modeling can be performed on the shared features.

[0101] Next, we will use Triplet Loss as the contrastive loss function and define the sampling strategy for anchor samples, positive samples, and negative samples through the following key code.

[0102] In step S5, the number of returned empty cans, the number of fully packed cartons, and the number of loose cartons are input into the visual branch in the form of feature vectors, and the weighing data is input into the weight branch in the form of time-series data.

[0103] Preferably, the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons obtained in step S3 can be converted into a unique thermal encoding vector. The dimension of the vector can be set according to the maximum possible number. For example, assuming the maximum number is 10, then each category corresponds to a 10-dimensional vector.

[0104] The next step is to perform z-score normalization on the time-series data acquired by the weighing sensor, converting the data into a sliding window sequence with a fixed window length, such as 10.

[0105] The next step is to convert the one-hot encoded vector into a tensor format suitable for the Siamese neural network contrast model and input it into the visual branch, and to convert the time series data into a three-dimensional tensor of sample number × time step × feature number and input it into the weight branch.

[0106] The next step is to calculate the cosine similarity between the outputs of the two branches and compare it with a preset threshold, such as 0.7. If the similarity is lower than the threshold, an anomaly alarm is triggered.

[0107] Step S6: The cosine similarity score between the visual branch and the weight branch is output through the Siamese neural network comparison model. When the cosine similarity score is lower than the preset score threshold, an access anomaly signal is generated.

[0108] Further, in step S6, the cosine similarity score between the visual branch and the weight branch is output through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated. Afterwards, the following steps are also included:

[0109] In step S10, the number of returned empty cans, the number of fully packed cartons, the number of loose cartons, the weighing data, and the cosine similarity score are uploaded to the consortium blockchain through different dedicated channels.

[0110] Preferably, the number of dedicated channels in this embodiment corresponds to five types of data: the number of returned empty cans, the number of packaged cartons, the number of loose cartons, weighing data, and cosine similarity. Then, the channel ID of each dedicated channel can be set as channel_empty (number of returned empty cans), channel_boxed (number of packaged cartons), channel_loose (number of loose cartons), channel_weight (weighing data), and channel_score (cosine similarity score).

[0111] The next step is to define the initial configuration for each channel in configtx.yaml, specifying the participating organizations (such as Org1, Org2) and anchor node strategies.

[0112] The next step is to use the configtxgen tool to generate the genesis block and channel configuration transactions for each channel, and to create an independent chaincode package for each channel, which contains smart contract functions for data uploading.

[0113] The next step is to implement a dedicated upload interface for each type of data, use the channel context of the Fabric SDK, and then achieve data association through cross-channel queries, using QueryByTxID to obtain the transaction hash.

[0114] Step S20: Associate all dedicated channels with the composite key DataType_NodeID_Timestamp, and build access ports for the consortium blockchain by periodically replacing consensus nodes based on the RBFT consensus mechanism.

[0115] Preferably, a composite key generation function can be defined in the chaincode to combine DataType (data type), NodeID (node ​​identifier), and Timestamp (time stamp) into a unique key, and the RBFT consensus node group can be configured in configtx.yaml to specify the initial node list and rotation period.

[0116] The next step is to implement the chaincode function for periodic node replacement, manage consensus node members through smart contracts, and implement the RBFT protocol extension in the orderer node to handle node join / leave events.

[0117] Finally, a cross-channel query function is implemented to verify data consistency using composite keys.

[0118] Step S30: Create a smart contract code monitoring access port on the blockchain using Solidity.

[0119] Preferably, creating a smart contract code monitoring access port requires first creating the monitoring contract infrastructure, which includes state variables, event definitions, and access control modifiers.

[0120] The next step is to add core port monitoring logic, including request verification, permission approval, and data integrity checks.

[0121] Step S40: Deploy the zk-SNARK protocol on the blockchain, and use zero-knowledge proofs to protect the privacy of encrypted smart contract code, so that the smart contract code is only open to authorized identities of external regulatory authorities.

[0122] Preferably, a zero-knowledge proof verification interface can be added to interact with the zk-SNARK protocol.

[0123] Finally, the chaincode is deployed to the network using the Fabric CLI tool.

[0124] In step S50, the smart contract code is called through a preset Python script, and the data of the newly added returned empty cans, the number of packaged cartons, the number of bulk cartons, the weighing data, and the cosine similarity score are uploaded to the blockchain based on the authorized identity.

[0125] Preferably, step S50 adopts the process of Fabric SDK initialization → data preparation and formatting → zk-SNARK authorization verification → smart contract call → on-chain result verification to achieve secure data on-chain.

[0126] Further, in step S6, the cosine similarity score between the visual branch and the weight branch is output through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated. Afterwards, the following steps are also included:

[0127] Step S100: Extract several historical oil usage data based on the lubricating oil supply cabinet. Each historical maintenance data includes the model code of each maintenance, the amount of lubricating oil used, the maintenance duration, and seasonal parameters.

[0128] In step S200, the machine model code, lubricating oil consumption, maintenance duration, and seasonal parameters for each maintenance are aggregated according to the matching maintenance event ID using a time series alignment algorithm to obtain a training dataset containing timestamps.

[0129] Preferably, the time series alignment algorithm can remove records with more than 30% missing values, ultimately forming a complete training dataset containing timestamps.

[0130] Preferably, feature transformation can be performed on the cleaned data, converting the model code into a 16-dimensional dense vector through an embedding layer, standardizing the maintenance duration using Z-score, and converting the seasonal parameters into sine / cosine periodic features. Training samples are generated using a sliding window technique, with each window containing a 30-day sequence of maintenance events and a window step size of 7 days, forming a three-dimensional tensor (number of samples × time step × number of features).

[0131] Step S300: Construct a pre-trained bidirectional LSTM network architecture. The loss function of the bidirectional LSTM network architecture is Huber loss, the optimizer is Nadam, and the initial learning rate is set to 0.001.

[0132] Step S400: Input the training dataset into the bidirectional LSTM network architecture, and output the prediction dataset based on the preset prediction step size through the bidirectional LSTM network architecture. The prediction dataset includes predicted oil consumption data.

[0133] Preferably, the model prediction results can be input into the Bayesian optimization module, using the prediction error distribution of the past 12 months as the prior probability, and combined with constraints such as current inventory levels, procurement cycles, and flight schedules to calculate the optimal inventory warning threshold. The threshold update follows a reinforcement learning strategy; when the prediction accuracy exceeds a certain percentage (e.g., 95%) for five consecutive times, the threshold adjustment range is automatically widened to reduce frequent alarms.

[0134] Step S500: Extract real-time oil consumption data based on the lubricating oil supply cabinet, and obtain the difference between the real-time oil consumption data and the predicted oil consumption data.

[0135] Step S600: When the difference exceeds the preset deviation threshold, generate a lubricating oil purchase request and send it to the external monitoring terminal.

[0136] Preferably, when the amount of a certain type of lubricating oil consumed exceeds the predicted value + 3σ (σ is the historical error standard deviation) for three consecutive times, i.e., a preset deviation threshold, the isolated forest anomaly detection algorithm is triggered. After verifying that the anomaly is real, the system generates a purchase recommendation form containing the recommended quantity and urgency level based on the supplier's delivery cycle and the Economic Order Quantity (EOQ) model.

[0137] Preferably, incremental model training can be performed monthly. Newly generated operational data is denoised using an adversarial autoencoder (AAE) and then fine-tuned with a learning rate of 10%. Simultaneously, edge node data from multiple airports is aggregated through a federated learning framework to update global model parameters, enabling the prediction model to continuously adapt to changes in consumption patterns brought about by new aircraft maintenance modes.

[0138] Further, in step S1, in response to an external worker's request to open the lubricating oil supply cabinet, facial features are extracted using a FaceNet deep convolutional neural network, and authorization information matching the facial features is retrieved from an external database. This specifically includes the following steps:

[0139] Step S11: In response to the request from the external staff to open the lubricating oil supply cabinet, acquire the facial image of the external staff through an external imaging device.

[0140] Preferably, the lubricating oil supply cabinet can be equipped with an industrial-grade camera (1920×1080 resolution, 30fps) to continuously monitor the cabinet door area. When a person is detected approaching (triggered by an infrared distance sensor), the camera automatically focuses and captures a facial image. The system's built-in FaceNet model (pre-trained on the CASIA-WebFace dataset) runs in real time, extracting a 512-dimensional facial feature vector.

[0141] Step S12: Extract high-dimensional facial feature vectors from facial images using the FaceNet deep convolutional neural network.

[0142] Step S13: Perform L2 normalization on the high-dimensional face feature vector to obtain a standardized feature vector.

[0143] Preferably,

[0144] Step S14: Obtain the image confidence of the standardized feature vector using the MTCNN face detection algorithm.

[0145] Step S15: When the image confidence exceeds the preset confidence threshold, the standardized feature vector is transmitted to an external database.

[0146] Preferably, the confidence threshold can be set to 0.7. When occlusion (such as a mask or goggles) or low clarity is detected, i.e. when the confidence is less than 0.7, an abnormal warning is triggered.

[0147] Step S16: Calculate the similarity between the standardized feature vector and all pre-stored facial records in the external database.

[0148] Preferably, the standardized feature vector is transmitted to an external authorization database via an encrypted channel (AES-256). This database is stored using a graph database (Neo4j), and each node contains attributes such as personnel ID, facial feature hash value, permission level (e.g., repairman / administrator), and effective time window. During a query, the cosine similarity between the feature vector and all records in the database is calculated, and authorized records with a similarity > 0.8 are returned.

[0149] Step S17: When the similarity is greater than the preset similarity threshold, retrieve the authorization information that matches the facial features from the external database.

[0150] Preferably, the preset similarity threshold can be set to 0.8.

[0151] Further, in step S2, if authorization information exists, the lubricating oil supply cabinet is opened. After the external personnel complete the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the complete / bulk combination, including:

[0152] Step S21: If authorization information exists, the lubricating oil supply cabinet is opened.

[0153] Step S22: After the external staff completes the storage operation, the weight data for a preset continuous duration is collected through the weighing function at a preset collection frequency.

[0154] Step S23: Extract the weight change rate features of all weight data using the first-order difference method to obtain a three-dimensional time-series feature matrix including the original weight, the amount of change, and the acceleration of change.

[0155] Preferably, after external personnel complete the storage operation, the weighing sensor of the lubricating oil supply cabinet can collect weight data continuously for 30 seconds at a frequency of 10Hz. Mechanical vibration noise is eliminated by sliding window mean filtering, and the weight change rate feature is extracted using the first-order difference method to form a three-dimensional time-series feature matrix containing the original weight, the change, and the acceleration of the change.

[0156] Step S24: Obtain historical access data of the lubricating oil supply cabinet based on several historical moments and divide it into a sample set and a validation set.

[0157] Preferably, the training set and the validation set can be divided in a 7:3 ratio.

[0158] Step S25: Construct a gradient boosting decision tree model based on the XGBoost framework, and define the objective variable of the gradient boosting decision tree model as extracting the optimal solution, which is to maximize the number of fully packaged cartons and minimize the number of loose cartons.

[0159] Preferably, the XGBoost framework is used to construct the model, with the target variable being the historically optimal retrieval scheme (a combination of the number of whole containers N_closed and the number of loose containers N_open). Parameters such as tree depth (max_depth=6) and learning rate (eta=0.1) are adjusted through Bayesian hyperparameter optimization, and early stopping is used to prevent overfitting.

[0160] Step S26: Input the sample set into the gradient boosting decision tree model and select the best splitting feature for each decision tree using a greedy algorithm. The best splitting feature of the current decision tree includes the score of the current decision tree based on each complete / disassembled combination.

[0161] Preferably, the weight of each tree is defined to be proportional to its F1 score on the validation set, and the weight distribution is normalized using the softmax function. An initial threshold (default 0.85) is determined based on ROC curve analysis, and the final score is determined by performing an argmax operation on the weighted voting results at the point of maximum Youden index (sensitivity ≥ 90%, specificity ≥ 80%). The model uses 100 CART trees with a maximum depth of 6 as mentioned above, and each tree selects the optimal splitting feature based on the information gain ratio using a greedy algorithm.

[0162] Preferably, in the model prediction stage, XGBoost generates the final predicted category through weighted voting of multiple decision trees. Typically, a classification threshold (such as 0.5) needs to be set to convert the probability output into a clear binary classification result.

[0163] The next step is to construct a confusion matrix to statistically analyze the prediction results: calculate true positives (TP, the number of samples correctly predicted as positive by the model), false positives (FP, the number of samples incorrectly predicted as positive by the model), true negatives (TN, the number of samples correctly predicted as negative by the model), and false negatives (FN, the number of samples incorrectly predicted as negative by the model).

[0164] The next step is to calculate the core metrics based on the confusion matrix: precision (P, i.e., TP / (TP+FP)) and recall (R, i.e., TP / (TP+FN)), and then obtain the F1-score (calculated as 2PR / (P+R)) by the harmonic average of the two.

[0165] Preferably, for multi-class scenarios, the F1-score can be calculated using either macro-averaging (taking the arithmetic mean of F1 scores for each class) or micro-averaging (calculating a comprehensive index based on global TP / FP / FN). In XGBoost, the F1-score evaluation function can be called through the scikit-learn interface.

[0166] Step S27: Obtain the F1-score performance index of each best splitting feature based on the validation set, and define that the weight of each decision tree is proportional to its respective F1-score performance index.

[0167] Preferably, the linearly normalized weights of the original F1-score can be used, i.e.:

[0168] The weight of a single tree = the F1 score of that tree / the sum of the F1 scores of all trees. For example, if the F1 scores of three trees are 0.7, 0.8 and 0.9 respectively, then the weights are 0.7 / 2.4≈0.29, 0.8 / 2.4≈0.33 and 0.9 / 2.4≈0.38.

[0169] Step S28: Calculate the weighted total score of all decision trees based on each package / bulk combination, according to the scores and weights of all decision trees.

[0170] Preferably, a common weighted average can be used.

[0171] Step S29: Define the complete / dispersed combination with the highest weighted total score as the optimal solution.

[0172] Preferably, since there is no situation where a single bottle of lubricating oil is not used up and is recycled, there is no situation where a half-bottle or one-third-bottle is opened and not used up. Therefore, the combination of full package / bulk packaging is actually a combination of the quantity used in the whole box and the quantity used in a single bottle. Thus, the greedy algorithm described above aims to use as much whole box of lubricating oil as possible and use as few single bottles of lubricating oil as possible.

[0173] Further, step S4 involves constructing a pre-trained Siamese neural network contrastive model that includes a visual branch and a weight branch, sharing convolutional kernel parameters. This specifically includes the following steps:

[0174] Step S41: Construct a symmetric neural network structure that includes a visual branch and a weight branch, with both branches using ResNet-18 as the basic architecture.

[0175] Preferably, the visual branch input is a 128×128×3 RGB image (from the bottle mouth / carton detection result), and the weight branch input is 64-dimensional time-series weighing data (normalized by a sliding window). The two branches are kept consistent at the feature extraction level by sharing convolution kernel parameters.

[0176] Step S42: Initialize the weight values ​​of the two branches using the Xavier method to share the convolution kernel parameters of the two branches.

[0177] Step S43: Obtain the visual features of empty lubricating oil cans, unopened lubricating oil cans, fully packaged cartons, and bulk cartons as anchor point samples.

[0178] Step S44: Weighing data from normal operation is used as positive samples, and weighing data from abnormal operation is used as negative samples.

[0179] Step S45: Pre-train the anchor samples, positive samples, and negative samples using the triplet loss function to enable the symmetric neural network structure to form compact intra-class distances and dispersed inter-class distances in the feature space, thus obtaining the Siamese neural network comparison model.

[0180] Preferably, the anchor sample is a standard feature for visual recognition (such as the bottle mouth feature of 8 empty cans), the positive sample is the matched weight sensor data (8 × empty can baseline weight ± 0.3% error), and the negative sample is the weight data corresponding to abnormal operation. Through a dynamic hard sample mining strategy, the most difficult negative sample to distinguish is selected in each round of training, so that the model forms a compact intra-class distance and a dispersed inter-class distance in the feature space.

[0181] Preferably, a contrastive loss can be introduced in the later stage of pre-training to force the output features of the visual branch and the weight branch to satisfy the cosine similarity of matching sample pairs > 0.9 and the cosine similarity of unmatched sample pairs < 0.3. The Adam optimizer (learning rate 0.0001) is used for 200 rounds of iterative training, and the early stopping method is used to monitor the validation set loss.

[0182] Preferably, after training, a Bayesian optimization module is connected to model the relationship between similarity score and false alarm rate using Gaussian process regression. The Bayesian optimization module continuously receives actual operating data and automatically adjusts the judgment threshold (initial value 0.85, update step size ±0.01) to ensure stable performance of the system under extreme scenarios such as metal reflection and rapid operation.

[0183] Further, in step S6, the cosine similarity score between the visual branch and the weight branch is output through the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated, specifically including the following steps:

[0184] Step S61: Perform L2 normalization on the visual feature vector output by the visual branch and the temporal weighing feature vector output by the weight branch to obtain the standardized visual feature vector and the standardized temporal weighing feature vector in turn.

[0185] Preferably, the visual branch features include the geometric features of the metal ring at the bottle opening and the position of the corner of the carton, while the weight branch features include time-series statistics such as the rate of weight change and the amplitude of fluctuation.

[0186] Step S62: The standardized visual feature vector and the standardized temporal weighing feature vector are mapped to the feature space through the Siamese neural network contrast model, and the cosine similarity score is calculated using equation (1):

[0187] (1);

[0188] in, Scoring the cosine similarity. To standardize visual feature vectors, This is a standardized time-series weighing feature vector.

[0189] Preferably, the score range is between [-1, 1], where 1 represents a perfect match and -1 represents a complete conflict.

[0190] Step S63: When the cosine similarity score is lower than the preset score threshold, an access anomaly signal is generated.

[0191] Preferably, the preset score threshold can be set to 0.85 and adjusted in real time through the Bayesian optimization module.

[0192] Preferably, after each abnormal event is triggered, relevant data (visual features, weight data, judgment results) can be added to the training set, the Siamese neural network model can be fine-tuned through online learning, and the model weights can be updated using the FTRL-Proximal algorithm, making subsequent detections more sensitive to similar anomalies.

[0193] In this embodiment, in response to an external worker's request to open the lubricating oil supply cabinet, a FaceNet deep convolutional neural network is used to extract facial features, and authorization information matching the facial features is retrieved from an external database. If authorization information exists, the lubricating oil supply cabinet is opened. After the external worker completes the storage operation, weighing data is obtained through the weighing function and processed based on a gradient boosting decision tree model to obtain the optimal solution for retrieving the packaged / bulk combination. The optimal solution is sent to the external worker. After the external worker completes the retrieval operation, the bottle mouth features of the lubricating oil can are identified using OpenCV Hough circle detection, corner points of the packaged cardboard box are identified using the YOLO algorithm, and the mask... The R-CNN instance segmentation model identifies the 3D geometric features of opened cardboard boxes, obtaining the number of returned empty cans, the number of fully packed boxes, and the number of loose boxes. A pre-trained Siamese neural network contrastive model is constructed, including a visual branch and a weight branch, sharing convolutional kernel parameters. The number of returned empty cans, fully packed boxes, and loose boxes are input into the visual branch as feature vectors, and the weighing data is input into the weight branch as time-series data. The Siamese neural network contrastive model outputs the cosine similarity score between the visual branch and the weight branch. When the cosine similarity score is lower than a preset threshold, an access anomaly signal is generated.

[0194] This embodiment addresses two core issues in aviation lubricant management—low ​​efficiency of manual operations and difficulty in ensuring consistency between inventory and records—by integrating a multimodal intelligent sensing and decision-making system. Traditional management methods rely on manual verification and paper records, which suffer from delays, errors, and a lack of real-time monitoring. This solution, however, achieves automated tracking of the entire lubricant storage and retrieval process through the collaboration of weighing sensors in the lubricant supply cabinet and computer vision algorithms (FaceNet, YOLO, Mask R-CNN). The core technological breakthrough of this embodiment lies in dynamic optimal solution calculation. By analyzing weighing data using a gradient boosting decision tree model, it intelligently recommends the appropriate combination of bulk / packaged lubricant, reducing manual decision-making time. Multimodal cross-validation utilizes a Siamese neural network to compare visual recognition results with weight change data. When the cosine similarity falls below a threshold, an anomaly alarm is triggered, fundamentally eliminating the risk of discrepancies between inventory and records caused by human error in traditional management. Furthermore, a facial recognition authorization mechanism ensures strict control over operating permissions, technically fulfilling the compliance requirements for managing high-risk materials in the aviation maintenance field.

[0195] like Figure 2 As shown, this embodiment provides an example of an aviation lubricating oil access management device. In this embodiment, the access management device is applied to the access management method as described in the above embodiment.

[0196] Specifically, the access management device includes a face matching module 1, an optimal solution calculation module 2, an image recognition module 3, a comparison model definition module 4, a comparison model input module 5, and a comparison model output module 6, which are electrically or signalally connected in sequence.

[0197] The system comprises the following modules: Face matching module 1 responds to external staff's request to open the lubricating oil supply cabinet, extracts facial features using a FaceNet deep convolutional neural network, and retrieves authorization information matching the facial features from an external database; Optimal solution calculation module 2 opens the lubricating oil supply cabinet if authorization information exists, and after the external staff completes the storage operation, acquires weighing data through a weighing function and processes the weighing data based on a gradient boosting decision tree model, outputting the optimal solution for retrieving both packaged and bulk oil; Image recognition module 3 sends the optimal solution to the external staff, and after the external staff completes the retrieval operation, identifies the bottle mouth features of the lubricating oil can using OpenCV Hough circle detection, identifies corner points of the packaged cardboard box using the YOLO algorithm, and identifies the mask... The R-CNN instance segmentation model identifies the 3D geometric features of opened cardboard boxes, sequentially obtaining the number of returned empty cans, the number of fully packed cardboard boxes, and the number of loose cardboard boxes. The comparison model definition module 4 is used to construct a pre-trained Siamese neural network comparison model that includes a visual branch and a weight branch, sharing convolutional kernel parameters. The comparison model input module 5 is used to input the number of returned empty cans, the number of fully packed cardboard boxes, and the number of loose cardboard boxes into the visual branch in the form of feature vectors, and input the weighing data into the weight branch in the form of time-series data. The comparison model output module 6 is used to output the cosine similarity score between the visual branch and the weight branch through the Siamese neural network comparison model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated.

[0198] Furthermore, the access management device also includes a data uplink module, a consortium blockchain access port construction module, a consortium blockchain access port monitoring module, a zk-SNARK protocol deployment module, and an additional data uplink module, which are electrically or signalally connected in sequence; the data uplink module is electrically or signalally connected to the comparison model output module 6.

[0199] The data upload module is used to upload the number of returned empty cans, the number of fully packed cartons, the number of loose cartons, weighing data, and cosine similarity scores to the consortium blockchain through different dedicated channels. The consortium blockchain access port construction module is used to associate all dedicated channels with the composite key DataType_NodeID_Timestamp and to build the consortium blockchain access port by periodically replacing consensus nodes based on the RBFT consensus mechanism. The consortium blockchain access port supervision module is used to create smart contract code on the blockchain through Solidity to supervise the access port. The zk-SNARK protocol deployment module is used to deploy the zk-SNARK protocol on the blockchain and encrypt the smart contract code based on zero-knowledge proof privacy protection so that the smart contract code is only open to authorized identities of external supervisors. The additional data upload module is used to call the smart contract code through a preset Python script and upload the newly added data such as the number of returned empty cans, the number of fully packed cartons, the number of loose cartons, weighing data, and cosine similarity scores to the blockchain based on the authorized identity.

[0200] Furthermore, the access management device also includes a historical oil consumption data extraction module, a training dataset aggregation module, a bidirectional LSTM network architecture construction module, a prediction dataset acquisition module, an oil consumption data difference acquisition module, and a lubricating oil purchase request generation and sending module, which are electrically or signalally connected in sequence; the historical oil consumption data extraction module and the comparison model output module are electrically or signalally connected.

[0201] The system comprises several modules: a historical oil consumption data extraction module for extracting historical oil consumption data from the lubricating oil supply cabinet, where each historical maintenance data includes the machine model code, lubricating oil consumption, maintenance duration, and seasonal parameters for each maintenance; a training dataset aggregation module for aggregating the machine model code, lubricating oil consumption, maintenance duration, and seasonal parameters for each maintenance according to the matching maintenance event ID using a time series alignment algorithm, resulting in a training dataset containing timestamps; a bidirectional LSTM network architecture construction module for constructing a pre-trained bidirectional LSTM network architecture, using Huber loss as the loss function, Nadam as the optimizer, and an initial learning rate of 0.001; a prediction dataset acquisition module for inputting the training dataset into the bidirectional LSTM network architecture, outputting a prediction dataset based on a preset prediction step size, which includes predicted oil consumption data; an oil consumption data difference acquisition module for extracting real-time oil consumption data from the lubricating oil supply cabinet and obtaining the difference between the real-time oil consumption data and the predicted oil consumption data; and a lubricating oil purchase request generation and sending module for generating and sending lubricating oil purchase requests to an external monitoring terminal when the difference exceeds a preset deviation threshold.

[0202] Furthermore, the face matching module 1 specifically includes a first face matching unit, a second face matching unit, a third face matching unit, a fourth face matching unit, a fifth face matching unit, a sixth face matching unit, and a seventh face matching unit that are electrically or signal-connected in sequence; the seventh face matching unit is electrically or signal-connected to the optimal solution extraction calculation module 2.

[0203] The system comprises the following components: a first face matching unit, which responds to an external worker's request to open the lubricating oil supply cabinet by acquiring the worker's facial image using an external imaging device; a second face matching unit, which extracts a high-dimensional facial feature vector from the facial image using a FaceNet deep convolutional neural network; a third face matching unit, which performs L2 normalization on the high-dimensional facial feature vector to obtain a standardized feature vector; a fourth face matching unit, which obtains the image confidence score of the standardized feature vector using the MTCNN face detection algorithm; a fifth face matching unit, which transmits the standardized feature vector to an external database when the image confidence score exceeds a preset confidence threshold; a sixth face matching unit, which calculates the similarity between the standardized feature vector and all pre-stored facial records in the external database; and a seventh face matching unit, which retrieves authorization information matching the facial features from the external database when the similarity score is greater than a preset similarity threshold.

[0204] Furthermore, the optimal solution extraction calculation module 2 specifically includes a first optimal solution extraction calculation unit, a second optimal solution extraction calculation unit, a third optimal solution extraction calculation unit, a fourth optimal solution extraction calculation unit, a fifth optimal solution extraction calculation unit, a sixth optimal solution extraction calculation unit, a seventh optimal solution extraction calculation unit, an eighth optimal solution extraction calculation unit, and a ninth optimal solution extraction calculation unit that are electrically or signal-connected in sequence; the first optimal solution extraction calculation unit is electrically or signal-connected to the seventh face matching unit, and the ninth optimal solution extraction calculation unit is electrically or signal-connected to the image recognition module 3.

[0205] The system comprises five sub-units: a first optimal solution retrieval unit, used to open the lubricating oil supply cabinet if authorized information exists; a second optimal solution retrieval unit, used to collect weight data for a preset continuous duration via weighing at a preset frequency after external personnel complete the storage operation; a third optimal solution retrieval unit, used to extract the weight change rate features of all weight data using the first-order difference method, obtaining a three-dimensional time-series feature matrix including the original weight, change amount, and change acceleration; a fourth optimal solution retrieval unit, used to acquire historical access data of the lubricating oil supply cabinet based on several historical moments and divide it into a sample set and a validation set; and a fifth optimal solution retrieval unit, used to construct a gradient boosting decision tree model based on the XGBoost framework, defining the objective variable of the gradient boosting decision tree model as retrieving the optimal solution. To maximize the number of fully packed cartons and minimize the number of loose cartons; the sixth optimal solution calculation unit is used to input the sample set into the gradient boosting decision tree model and select the best splitting feature for each decision tree using a greedy algorithm. The best splitting feature of the current decision tree includes the score of the current decision tree based on each fully packed / loose combination; the seventh optimal solution calculation unit is used to obtain the F1-score performance index of each optimal splitting feature based on the validation set, and defines the weight of each decision tree as proportional to its respective F1-score performance index; the eighth optimal solution calculation unit is used to calculate the weighted total score of all decision trees based on each fully packed / loose combination according to the scores and weights of all decision trees; the ninth optimal solution calculation unit is used to define the fully packed / loose combination with the highest weighted total score as the optimal solution.

[0206] Furthermore, the comparison model definition module 4 specifically includes a first comparison model definition unit, a second comparison model definition unit, a third comparison model definition unit, a fourth comparison model definition unit, and a fifth comparison model definition unit that are electrically or signal-connected in sequence; the first comparison model definition unit is electrically or signal-connected to the image recognition module 3, and the fifth comparison model definition unit is electrically or signal-connected to the comparison model input module 5.

[0207] The first contrast model definition unit is used to construct a symmetric neural network structure containing a visual branch and a weight branch, both of which use ResNet-18 as the basic architecture. The second contrast model definition unit is used to initialize the weight values ​​of the two branches using the Xavier method to share the convolution kernel parameters of the two branches. The third contrast model definition unit is used to acquire the visual features of empty lubricating oil tanks, unopened lubricating oil tanks, packaged cartons, and bulk cartons as anchor samples. The fourth contrast model definition unit is used to use weighing data from normal operations as positive samples and weighing data from abnormal operations as negative samples. The fifth contrast model definition unit is used to pre-train the anchor samples, positive samples, and negative samples using a triplet loss function to make the symmetric neural network structure form compact intra-class distances and dispersed inter-class distances in the feature space, thus obtaining the Siamese neural network contrast model.

[0208] Furthermore, the comparison model output module 6 specifically includes a first comparison model output unit, a second comparison model output unit, and a third comparison model output unit that are electrically or signal-connected in sequence; the first comparison model output unit is electrically or signal-connected to the comparison model input module 5.

[0209] The first comparison model output unit is used to perform L2 normalization on the visual feature vector output by the visual branch and the time-series weighing feature vector output by the weight branch, respectively, to obtain the standardized visual feature vector and the standardized time-series weighing feature vector.

[0210] The output unit of the second contrast model is used to map the standardized visual feature vector and the standardized temporal weighing feature vector to the feature space through the Siamese neural network contrast model, and calculate the cosine similarity score using equation (1):

[0211] (1);

[0212] in, Scoring the cosine similarity. To standardize visual feature vectors, This is a standardized time-series weighing feature vector.

[0213] The third comparison model output unit is used to generate an access anomaly signal when the cosine similarity score is lower than a preset score threshold.

[0214] It should be noted that this embodiment is a functional module embodiment based on the above method embodiment. For the preferred, extended, limited, exemplified and principle explanation parts of this embodiment, please refer to the above embodiment. This embodiment will not repeat them.

[0215] In this embodiment, in response to an external worker's request to open the lubricating oil supply cabinet, a FaceNet deep convolutional neural network is used to extract facial features, and authorization information matching the facial features is retrieved from an external database. If authorization information exists, the lubricating oil supply cabinet is opened. After the external worker completes the storage operation, weighing data is obtained through the weighing function and processed based on a gradient boosting decision tree model to obtain the optimal solution for retrieving the packaged / bulk combination. The optimal solution is sent to the external worker. After the external worker completes the retrieval operation, the bottle mouth features of the lubricating oil can are identified using OpenCV Hough circle detection, corner points of the packaged cardboard box are identified using the YOLO algorithm, and the mask... The R-CNN instance segmentation model identifies the 3D geometric features of opened cardboard boxes, obtaining the number of returned empty cans, the number of fully packed boxes, and the number of loose boxes. A pre-trained Siamese neural network contrastive model is constructed, including a visual branch and a weight branch, sharing convolutional kernel parameters. The number of returned empty cans, fully packed boxes, and loose boxes are input into the visual branch as feature vectors, and the weighing data is input into the weight branch as time-series data. The Siamese neural network contrastive model outputs the cosine similarity score between the visual branch and the weight branch. When the cosine similarity score is lower than a preset threshold, an access anomaly signal is generated.

[0216] This embodiment addresses two core issues in aviation lubricant management—low ​​efficiency of manual operations and difficulty in ensuring consistency between inventory and records—by integrating a multimodal intelligent sensing and decision-making system. Traditional management methods rely on manual verification and paper records, which suffer from delays, errors, and a lack of real-time monitoring. This solution, however, achieves automated tracking of the entire lubricant storage and retrieval process through the collaboration of weighing sensors in the lubricant supply cabinet and computer vision algorithms (FaceNet, YOLO, Mask R-CNN). The core technological breakthrough of this embodiment lies in dynamic optimal solution calculation. By analyzing weighing data using a gradient boosting decision tree model, it intelligently recommends the appropriate combination of bulk / packaged lubricant, reducing manual decision-making time. Multimodal cross-validation utilizes a Siamese neural network to compare visual recognition results with weight change data. When the cosine similarity falls below a threshold, an anomaly alarm is triggered, fundamentally eliminating the risk of discrepancies between inventory and records caused by human error in traditional management. Furthermore, a facial recognition authorization mechanism ensures strict control over operating permissions, technically fulfilling the compliance requirements for managing high-risk materials in the aviation maintenance field.

[0217] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0218] The memory 72 stores program instructions for implementing the aviation lubricating oil access management method of any of the above embodiments.

[0219] The processor 71 is used to execute program instructions stored in the memory 72 for the management of access to aviation lubricating oil.

[0220] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0221] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 In this embodiment of the application, the storage medium 8 stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. 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, or terminal devices such as computers, servers, mobile phones, and tablets.

[0222] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, signal, or other forms.

[0223] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for storing and managing aviation lubricating oil, wherein the method is applied to several lubricating oil tanks of the same specifications placed in a lubricating oil supply cabinet, the lubricating oil supply cabinet having a weighing function, and all lubricating oil tanks being in at least one state of being packed in cartons or in bulk, characterized in that, The access management method includes: Step S1: In response to the request from an external worker to open the lubricating oil supply cabinet, the FaceNet deep convolutional neural network is used to extract facial features, and authorization information matching the facial features is retrieved from an external database. Step S2: If the authorization information exists, the lubricating oil supply cabinet is opened. After the external staff completes the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the whole package / bulk package combination. Step S3: Send the optimal solution for retrieval to the external staff. After the external staff completes the retrieval operation, the bottle mouth features of the lubricating oil can are identified by OpenCV Hough circle detection, the corner points of the packaged carton are identified by YOLO algorithm, and the three-dimensional geometric features of the opened carton are identified by MaskR-CNN instance segmentation model to obtain the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons. Step S4: Construct a pre-trained Siamese neural network contrastive model that includes visual and weight branches and shares convolutional kernel parameters; Step S5: Input the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons into the visual branch in the form of feature vectors, and input the weighing data into the weight branch in the form of time-series data; Step S6: The cosine similarity score between the visual branch and the weight branch is output through the twin neural network comparison model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated.

2. The access management method according to claim 1, characterized in that, Step S6: Output the cosine similarity score between the visual branch and the weight branch using the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, generate an access anomaly signal. Following this, the process further includes: Step S10: The number of returned empty cans, the number of packaged cartons, the number of bulk cartons, the weighing data, and the cosine similarity score are uploaded to the consortium blockchain through different dedicated channels. Step S20: Associate all dedicated channels with the composite key DataType_NodeID_Timestamp, and construct the access port of the consortium blockchain by periodically replacing consensus nodes based on the RBFT consensus mechanism; Step S30: Create smart contract code on the blockchain using Solidity to manage the access port; Step S40: Deploy the zk-SNARK protocol on the blockchain and encrypt the smart contract code based on zero-knowledge proof privacy protection, so that the smart contract code is only open to authorized identities of external regulatory authorities; Step S50: The smart contract code is called through a preset Python script, and the newly added number of empty cans returned, the number of fully packed cartons, the number of loose cartons, the weighing data, and the cosine similarity score are uploaded to the blockchain based on the authorized identity.

3. The access management method according to claim 1, characterized in that, Step S6: Output the cosine similarity score between the visual branch and the weight branch using the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, generate an access anomaly signal. Following this, the process further includes: Step S100: Extract several historical oil consumption data based on the lubricating oil supply cabinet. Each historical maintenance data includes the model code of each maintenance, the amount of lubricating oil used, the maintenance duration, and seasonal parameters. Step S200: The machine model code, lubricating oil consumption, maintenance duration, and seasonal parameters for each maintenance are aggregated according to the matching maintenance event ID using a time series alignment algorithm to obtain a training dataset containing timestamps. Step S300: Construct a pre-trained bidirectional LSTM network architecture, wherein the loss function of the bidirectional LSTM network architecture adopts Huber loss, the optimizer is Nadam, and the initial learning rate is set to 0.

001. Step S400: Input the training dataset into the bidirectional LSTM network architecture, and output a prediction dataset based on a preset prediction step size through the bidirectional LSTM network architecture. The prediction dataset includes predicted oil consumption data. Step S500: Extract real-time oil consumption data based on the lubricating oil supply cabinet, and obtain the difference between the real-time oil consumption data and the predicted oil consumption data; Step S600: When the difference exceeds the preset deviation threshold, a lubricating oil purchase request is generated and sent to the external monitoring terminal.

4. The access management method according to claim 1, characterized in that, Step S1, in response to an external worker's request to open the lubricating oil supply cabinet, uses a FaceNet deep convolutional neural network to extract facial features and retrieves authorization information matching the facial features from an external database, including: Step S11: In response to the request from an external worker to open the lubricating oil supply cabinet, acquire the facial image of the external worker using an external imaging device; Step S12: Extract high-dimensional facial feature vectors from the facial image using the FaceNet deep convolutional neural network; Step S13: Perform L2 normalization on the high-dimensional face feature vector to obtain a standardized feature vector; Step S14: Obtain the image confidence of the standardized feature vector using the MTCNN face detection algorithm; Step S15: When the image confidence exceeds a preset confidence threshold, the standardized feature vector is transmitted to the external database. Step S16: Calculate the similarity between the standardized feature vector and all pre-stored facial records in the external database; Step S17: When the similarity is greater than a preset similarity threshold, retrieve the authorization information that matches the facial features from an external database.

5. The access management method according to claim 1, characterized in that, Step S2: If the authorization information exists, the lubricating oil supply cabinet is opened. After the external staff completes the storage operation, the weighing data is obtained through the weighing function and processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the complete / bulk combination, including: Step S21: If the authorization information exists, then open the lubricating oil supply cabinet; Step S22: After the external staff completes the storage operation, the weighing function collects weight data for a preset continuous duration at a preset collection frequency. Step S23: Extract the weight change rate features of all weight data using the first-order difference method to obtain a three-dimensional time-series feature matrix including the original weight, the amount of change, and the acceleration of change. Step S24: Obtain the historical access data of the lubricating oil supply cabinet based on several historical moments and divide it into a sample set and a verification set; Step S25: Construct the gradient boosting decision tree model based on the XGBoost framework, and define the objective variable of the gradient boosting decision tree model as taking the optimal solution, wherein taking the optimal solution is to maximize the number of fully packaged cartons and minimize the number of loose cartons; Step S26: Input the sample set into the gradient boosting decision tree model and select the best splitting feature for each decision tree using a greedy algorithm. The best splitting feature of the current decision tree includes the score of the current decision tree based on each complete / disassembled combination. Step S27: Obtain the F1-score performance index of each best splitting feature based on the validation set, and define that the weight of each decision tree is proportional to its respective F1-score performance index; Step S28: Calculate the weighted total score of all decision trees based on each package / bulk combination, according to the scores and weights of all decision trees. Step S29: Define the assembled / bulk combination with the highest weighted total score as the optimal solution to be extracted.

6. The access management method according to claim 1, characterized in that, Step S4, construct a pre-trained Siamese neural network contrastive model that includes a visual branch and a weight branch, sharing convolutional kernel parameters, including: Step S41: Construct a symmetric neural network structure that includes a visual branch and a weight branch, with both branches using ResNet-18 as the basic architecture. Step S42: Initialize the weight values ​​of the two branches using the Xavier method to share the convolution kernel parameters of the two branches; Step S43: Obtain visual features of empty lubricating oil cans, unopened lubricating oil cans, fully packaged cartons, and bulk cartons as anchor point samples; Step S44: Weighing data from normal operation is used as positive samples, and weighing data from abnormal operation is used as negative samples. Step S45: The anchor sample, the positive sample, and the negative sample are pre-trained using the triplet loss function to enable the symmetric neural network structure to form compact intra-class distances and dispersed inter-class distances in the feature space, thereby obtaining the Siamese neural network comparison model.

7. The access management method according to claim 5, characterized in that, Step S6: Output the cosine similarity score between the visual branch and the weight branch using the Siamese neural network contrast model. When the cosine similarity score is lower than a preset score threshold, generate an access anomaly signal, including: Step S61: Perform L2 normalization on the visual feature vector output by the visual branch and the time-series weighing feature vector output by the weight branch to obtain the standardized visual feature vector and the standardized time-series weighing feature vector in sequence. Step S62: The standardized visual feature vector and the standardized temporal weighing feature vector are mapped to the feature space through the Siamese neural network contrast model, and the cosine similarity score is calculated using equation (1): （1）； in, The cosine similarity score is given. The standardized visual feature vector, The standardized time-series weighing feature vector; Step S63: When the cosine similarity score is lower than the preset score threshold, the access anomaly signal is generated.

8. An access management device for aviation lubricating oil, wherein the access management device is applied to the access management method as described in any one of claims 1 to 7, characterized in that, The access management device includes: The face matching module is used to respond to the request from external staff to open the lubricating oil supply cabinet. It uses FaceNet deep convolutional neural network to extract facial features and retrieves authorization information that matches the facial features from an external database. The optimal solution extraction module is used to open the lubricating oil supply cabinet if the authorization information exists, and after the external staff completes the storage operation, it obtains the weighing data through the weighing function and processes the weighing data based on the gradient boosting decision tree model to output the optimal extraction solution for the whole / bulk combination. The image recognition module is used to send the optimal solution for retrieval to the external staff. After the external staff completes the retrieval operation, the module uses OpenCV Hough circle detection to identify the bottle mouth features of the lubricating oil can, YOLO algorithm to identify the corner points of the packaged carton, and Mask R-CNN instance segmentation model to identify the three-dimensional geometric features of the opened carton, thereby obtaining the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons. The contrastive model definition module is used to construct a pre-trained Siamese neural network contrastive model that includes a visual branch and a weight branch and shares convolutional kernel parameters. The comparison model input module is used to input the number of returned empty cans, the number of packaged cartons, and the number of bulk cartons into the visual branch in the form of feature vectors, and to input the weighing data into the weight branch in the form of time-series data. The comparison model output module is used to output the cosine similarity score between the visual branch and the weight branch through the Siamese neural network comparison model. When the cosine similarity score is lower than a preset score threshold, an access anomaly signal is generated.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the access management method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the access management method as described in any one of claims 1 to 7.

Images