Method, device and equipment for managing access to aviation lubricants and storage medium

Abstract

The application discloses an aviation lubricating oil storage and access management method and device, equipment and a storage medium, relates to the technical field of image processing, and solves the two core problems of low artificial operation efficiency and difficulty in guaranteeing the consistency of physical and accounting records in the field of aviation lubricating oil management through the integration of a multi-modal intelligent sensing and decision system. The traditional management mode relies on manual verification and paper records, and has defects such as response delay, error-prone and inability to monitor in real time. Through the cooperation of the weighing sensor of the lubricating oil supply cabinet and the computer vision algorithm, the present application realizes the automatic tracking of the whole process of lubricating oil storage and access. The risk of inconsistent physical and accounting records caused by human negligence in traditional management is eliminated. In addition, the face recognition authorization mechanism ensures strict control of operation permission, and realizes the compliance requirement of aviation maintenance field for high-risk product management from the technical aspect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an aviation oil storage and access management method, device and equipment and a storage medium. BACKGROUND

[0002] The management of aviation oil has always been an important challenge in the aviation field due to its particularity, high risk and strict application scene requirements. The management of aviation oil by the aviation maintenance department is directly related to flight safety, maintenance efficiency and environmental compliance.

[0003] Currently, the management of aviation oil is basically a fixed-point storage and manual management scheme, that is, aviation oil is stored in shelves or cabinets in designated warehouses. Maintenance personnel first contact the management personnel when using the aviation oil, go to the designated warehouse together, verify the name, part number, expiration date and quantity of the aviation oil to be borrowed, and then complete the borrowing registration. When returning, the management personnel needs to review and sign the consistency of the aviation oil account. For aviation maintenance departments with complete manual specifications and strict management, a dedicated person is arranged to manage the aviation oil as described above.

[0004] However, most maintenance departments cannot arrange a dedicated person to save costs, and most of them are managed by on-duty aviation material personnel, tool room personnel or aircraft maintenance personnel on a part-time basis. The time cost and labor cost of manual management are high and prone to errors, especially the part-time management method is more likely to cause unclear division of labor, overlapping positions and lack of positions, resulting in low management efficiency of aviation oil. SUMMARY

[0005] The main purpose of the present application is to provide an aviation oil storage and access management method, device, equipment and storage medium to solve the problem of low management efficiency of aviation oil in the prior art.

[0006] In order to achieve the above purpose, the present application provides the following technical solutions:

[0007] An aviation oil storage and access management method, the storage and access management method is applied to a plurality of same-specification oil tanks placed in an oil supply cabinet, the oil supply cabinet has a weighing function, and all the oil tanks are in at least one of a carton whole or bulk state, the storage and access management method comprising:

[0008] Step S1, in response to an oil supply cabinet opening request of an external worker, extracting a face feature using a FaceNet deep convolutional neural network, and calling authorization information matching the face feature from an external database;

[0009] Step S2, if the authorization information exists, open the oil supply cabinet, based on the external staff complete the storage operation, obtain the weighing data through the weighing function and process the weighing data based on the gradient boosting decision tree model to obtain the optimal solution of taking out the combination of bulk / solid;

[0010] Step S3, send the optimal solution of taking out to the external staff, after the external staff complete the taking out operation, through OpenCV Hough circle detection to identify the bottle opening features of the oil tank, Yolo algorithm to identify the corner points of the solid paper box, Mask R-CNN instance segmentation model to identify the three-dimensional geometric features of the uncapped paper box, to obtain the number of returned empty cans, the number of solid paper boxes and the number of bulk paper boxes;

[0011] Step S4, construct a pre-trained twin neural network contrast model including a visual branch and a weight branch, which shares convolution kernel parameters;

[0012] Step S5, input the returned empty can quantity, the solid paper box quantity and the bulk paper box quantity into the visual branch in the form of feature vectors respectively, and input the weighing data into the weight branch in the form of time series data;

[0013] Step S6, output the cosine similarity score of the visual branch and the weight branch through the twin neural network contrast model, and generate an access exception signal when the cosine similarity score is lower than a preset score threshold.

[0014] As a further improvement of the present application, step S6, output the cosine similarity score of the visual branch and the weight branch through the twin neural network contrast model, and generate an access exception signal when the cosine similarity score is lower than a preset score threshold, and then further comprising:

[0015] Step S10, upload the returned empty can quantity, the solid paper box quantity, the bulk paper box quantity, the weighing data and the cosine similarity score to the alliance chain through different special channels;

[0016] Step S20, associate all special channels through composite key DataType_NodeID_Timestamp, and build access ports of the alliance chain through periodic replacement of consensus nodes based on RBFT consensus mechanism;

[0017] Step S30, create smart contract code to supervise the access port through Solidity in the blockchain;

[0018] Step S40, deploying a zk-SNARK protocol on the blockchain, and encrypting the smart contract code based on zero-knowledge proof privacy protection, so that the smart contract code is only open to the authorized identity of the external supervision end;

[0019] Step S50, calling the smart contract code through a preset Python script, and based on the authorized identity, uploading data of the newly added returned empty can quantity, full carton quantity, bulk carton quantity, weighing data, and cosine similarity score to the chain.

[0020] As a further improvement of the present application, step S6, the twin neural network comparison model outputs the cosine similarity score of the visual branch and the weight branch, and generates an access exception signal when the cosine similarity score is lower than a preset score threshold. Then, it further includes:

[0021] Step S100, extracting a plurality of historical oil consumption data based on the oil supply cabinet, each historical maintenance data including aircraft model code, oil consumption, maintenance duration, and season parameter;

[0022] Step S200, aggregating the aircraft model code, oil consumption, maintenance duration, and season parameter according to the matched maintenance event ID through a time series alignment algorithm, to obtain a training data set containing a timestamp;

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

[0024] Step S400, inputting the training data set into the bidirectional LSTM network architecture, and outputting a prediction data set based on a preset prediction step through the bidirectional LSTM network architecture, the prediction data set including predicted oil consumption data;

[0025] Step S500, extracting real-time oil consumption data based on the oil supply cabinet, and obtaining a difference value between the real-time oil consumption data and the predicted oil consumption data;

[0026] Step S600, when the difference value exceeds a preset deviation threshold, generating an oil purchase request and sending it to the external monitoring end.

[0027] As a further improvement of the present application, step S1, in response to an oil supply cabinet opening request of an external staff, a FaceNet deep convolutional neural network is used to extract facial features, and authorized information matching the facial features is called 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, inputting the sample set into the gradient boosting decision tree model and selecting the best split feature of each decision tree by a greedy algorithm, the best split feature of the current decision tree including the score of the current decision tree based on each whole / sparse combination;

[0042] Step S27, obtaining the F1-score performance indicator of each best split feature based on the validation set respectively, and defining the weight of each decision tree to be proportional to the respective F1-score performance indicator;

[0043] Step S28, calculating the weighted total score of all decision trees based on each whole / sparse combination according to the scores and weights of all decision trees;

[0044] Step S29, defining the whole / sparse combination with the highest weighted total score as the optimal solution.

[0045] As a further improvement of the present application, step S4, a pre-trained twin neural network contrast model including a visual branch and a weight branch sharing convolution kernel parameters is constructed, comprising:

[0046] Step S41, a symmetric neural network structure containing a visual branch and a weight branch is constructed, both branches using ResNet-18 as the basic architecture;

[0047] Step S42, the weight values of the two branches are initialized by the Xavier method to share the convolution kernel parameters of the two branches;

[0048] Step S43, the visual features of empty oil tanks, unopened oil tanks, whole paper boxes, and bulk paper boxes are obtained as anchor point samples;

[0049] Step S44, the weighing data of normal operation as positive samples and the weighing data of abnormal operation as negative samples;

[0050] Step S45, pre-training the anchor point samples, the positive samples, and the negative samples by a triplet loss function to make the symmetric neural network structure form a compact intra-class distance and a dispersed inter-class distance in the feature space, obtaining the twin neural network contrast model.

[0051] As a further improvement of the present application, step S6, outputting the cosine similarity score of the visual branch and the weight branch by the twin neural network contrast model, and generating an access abnormal signal when the cosine similarity score is lower than a preset score threshold, comprising:

[0052] Step S61, the visual feature vector output by the visual branch and the time-series weighing feature vector output by the weight branch are respectively subjected to L2 normalization, and a standardized visual feature vector and a standardized time-series weighing feature vector are sequentially obtained;

[0053] Step S62, the standardized visual feature vector and the standardized time-series weighing feature vector are mapped to the feature space by the twin neural network comparison model, and the cosine similarity score is calculated by formula (1):

[0054] (1);

[0055] wherein, is the cosine similarity score, is the standardized visual feature vector, is the standardized time-series weighing feature vector;

[0056] Step S63, when the cosine similarity score is lower than the preset score threshold, the access abnormality signal is generated.

[0057] In order to achieve the above purpose, the present application also provides the following technical scheme:

[0058] An access management device of aviation lubricating oil, the access management device is applied to the access management method as described above, and the access management device comprises:

[0059] A face matching module is configured to extract face features by using a FaceNet deep convolutional neural network in response to a lubricating oil supply cabinet opening request of an external worker, and to call authorized information matched with the face features from an external database;

[0060] An optimal solution taking-out calculation module is configured to open the lubricating oil supply cabinet if the authorized information exists, to obtain weighing data by using the weighing function after the external worker completes a storage operation, to process the weighing data based on a gradient boosting decision tree model, and to output an optimal solution of taking out a combination of bulk and packaged.

[0061] An image recognition module is configured to send the optimal solution of taking out to the external worker, to recognize a bottle opening feature of a lubricating oil tank by using OpenCV Hough circle detection, to recognize a packaged carton corner point by using Yolo algorithm, and to recognize a three-dimensional geometric feature of an uncapped carton by using a Mask R-CNN instance segmentation model after the external worker completes a taking-out operation, and to sequentially obtain a number of returned empty cans, a number of packaged cartons, and a number of bulk cartons.

[0062] A comparison model definition module is configured to construct a pre-trained twin neural network comparison model comprising a visual branch and a weight branch and sharing convolution kernel parameters.

[0063] a comparison model input module, configured to input the returned empty can quantity, the whole carton quantity, and the bulk carton quantity in the form of feature vectors into the visual branch respectively, and input the weighing data in the form of time series data into the weight branch;

[0064] a comparison model output module, configured to output, by the twin neural network comparison model, a cosine similarity score of the visual branch and the weight branch, and generate an access exception signal when the cosine similarity score is lower than a preset score threshold.

[0065] To achieve the above object, the present 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; the processor implements the access management method as described above when executing the program instructions stored in the memory.

[0067] To achieve the above object, the present application also provides the following technical solutions:

[0068] A computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are executed by the processor to implement the access management method as described above.

[0069] Effective effect:

[0070] In response to the oil supply cabinet opening request of the external staff, the FaceNet deep convolutional neural network is used to extract the face features, and the authorized information matched with the face features is called from the external database; if there is authorized information, the oil supply cabinet is opened, based on the storage operation completed by the external staff, the weighing data is obtained through the weighing function, and the weighing data is processed based on the gradient boosting decision tree model to obtain the whole / bulk combined optimal solution for taking out; the optimal solution for taking out is sent to the external staff, after the external staff completes the taking out operation, the bottle opening features of the oil tank are recognized through the OpenCV Hough circle detection, the Yolo algorithm is used to recognize the whole carton corner points, and the Mask R-CNN instance segmentation model is used to recognize the three-dimensional geometric features of the opened carton, to obtain the returned empty can quantity, the whole carton quantity, and the bulk carton quantity; a pre-trained twin neural network comparison model including a visual branch and a weight branch is constructed, which shares convolution kernel parameters; the returned empty can quantity, the whole carton quantity, and the bulk carton quantity are input into the visual branch in the form of feature vectors respectively, and the weighing data is input into the weight branch in the form of time series data; the cosine similarity score of the visual branch and the weight branch is output by the twin neural network comparison model, and an access exception signal is generated when the cosine similarity score is lower than a preset score threshold.

[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", "third", etc. in the present application are only for descriptive purpose and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "plurality" is 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 the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0078] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive of other embodiments. It is expressly understood that the embodiments described herein can be combined with other embodiments.

[0079] As shown in Figure 1 The present embodiment provides an embodiment of an access management method for aviation lubricating oil, which is applied to a plurality of lubricating oil tanks of the same specification placed in a lubricating oil supply cabinet with weighing function, and all the lubricating oil tanks are in at least one of a carton-packed or bulk-packed state.

[0080] Preferably, the specification of the lubricating oil tank is commonly 946ml / tank or 946ml / pail for small packages, which is suitable for aviation turbine engines; the large package specification includes 15KG / pail (such as ground aviation hydraulic oil) and 3kg / pail (such as aviation lubricating grease).

[0081] Preferably, the packaging mode is usually 24x1 quart / box (24qt / ctn), 5 gallons / pail (5gal / pail).

[0082] Preferably, the present embodiment is mainly used for unified management of lubricating oil tanks of the same model, and for lubricating oil tanks of multiple specifications, it is suggested to use a set of access management methods and access management devices provided in the present embodiment for one 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, load the pre-trained YOLO model for the corner points of the carton, and count the number of corner points for each detected carton. A complete carton should have 4 visible corner points. Next, fine-tune the pre-trained Mask R-CNN model through the COCO dataset, and calculate the stereo geometric features of the carton through the instance segmentation result: surface area change rate, vertex displacement, and opening angle. Finally, the result fusion and quantity statistics are realized by establishing spatial correlation rules and final quantity calculation. The spatial correlation rules are established by matching the 3D coordinates of the detection results of the same physical object at different angles, and the attribution relationship between the bottle mouth and the carton is determined by the distance threshold method.

[0099] Step S4, construct a pre-trained twin neural network contrast model including a visual branch and a weight branch, which share convolution kernel parameters.

[0100] Preferably, the model architecture design of the twin neural network contrast model can adopt a 3-layer convolution + ReLU + BN shared convolution feature extractor. The visual branch shares features followed by a 3-layer fully connected network to process bottle mouth / carton quantity features; the weight branch shares features followed by an LSTM layer to process time-series weighing data. Specifically, for the visual branch, the 3D features (empty cans / whole cartons / bulk cartons) can be flattened and processed; for the weight branch, the shared features can be time-series modeled.

[0101] Next, use Triplet Loss as the contrast loss function, and define the sampling strategy of anchor samples (anchor), positive samples (positive), and negative samples (negative) through the following key code.

[0102] Step S5, input the returned empty can quantity, whole carton quantity, and bulk carton quantity in the form of feature vectors into the visual branch, and input the weighing data in the form of time-series data into the weight branch.

[0103] Preferably, the returned empty can quantity, whole carton quantity, and bulk carton quantity obtained in step S3 can be converted into one-hot encoding vectors. The dimension of the vector can be set according to the maximum possible quantity, for example, assuming the maximum quantity is 10, then each class corresponds to a 10-dimensional vector.

[0104] Next, the time-series data obtained by the weighing sensor can be standardized by z-score normalization, and the data can be converted into a sliding window sequence with a fixed window length, for example, 10.

[0105] Next, convert the one-hot encoding vector into a tensor format suitable for the twin neural network contrast model and input it into the visual branch, and convert the time-series data into a three-dimensional tensor of sample number x time step x feature number and input it into the weight branch.

[0106] Next, the cosine similarity of the two branch outputs is calculated and compared with a preset threshold, for example 0.7, and if it is lower than the threshold, an abnormal alarm is triggered.

[0107] Step S6, the cosine similarity score of the visual branch and the weight branch output by the twin neural network comparison model is compared, and when the cosine similarity score is lower than the preset score threshold, an access exception signal is generated.

[0108] Further, in step S6, the cosine similarity score of the visual branch and the weight branch output by the twin neural network comparison model is compared, and when the cosine similarity score is lower than the preset score threshold, an access exception signal is generated, and then the following steps are further included:

[0109] Step S10, the returned empty can quantity, boxed carton quantity, loose carton quantity, weighing data, and cosine similarity score are uploaded to the consortium chain through different special channels.

[0110] Preferably, the number of special channels of the embodiment corresponds to the five types of data of returned empty can quantity, boxed carton quantity, loose carton quantity, weighing data, and cosine similarity, and the channel ID of each special channel can be set as returned empty can quantity channel_empty, boxed carton quantity channel_boxed, loose carton quantity channel_loose, weighing data channel_weight, and cosine similarity score channel_score.

[0111] Next, the initial configuration of each channel can be defined in configtx.yaml, and the participating organizations (such as Org1 and Org2) and anchor node strategies are specified.

[0112] Next, use the configtxgen tool to generate the genesis block and channel configuration transaction of each channel, and create an independent chain code package (chaincode) for each channel, which includes the smart contract function for data upload.

[0113] Next, implement a special upload interface for each type of data, and use the channel context of the Fabric SDK to implement data association through cross-channel queries, and use QueryByTxID to obtain the transaction hash.

[0114] Step S20, all special channels are associated through the composite key DataType_NodeID_Timestamp, and the access port of the consortium chain is constructed based on the RBFT consensus mechanism through periodic replacement of consensus nodes.

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

[0116] Next, the chaincode function for periodic node replacement is implemented to manage consensus node members through smart contracts, and the RBFT protocol extension is implemented in the orderer node to handle node join / exit events.

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

[0118] Step S30, create smart contract code supervision access port on blockchain through Solidity.

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

[0120] Next, add the core logic of port supervision, including request verification, permission approval and data integrity check.

[0121] Step S40, deploy zk-SNARK protocol on blockchain to encrypt smart contract code based on zero-knowledge proof privacy protection, so that the smart contract code is only open to authorized identities of external supervision ends.

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

[0123] Finally, deploy the chaincode to the network through the Fabric CLI tool.

[0124] Step S50, call the smart contract code through the preset Python script, and based on the authorized identity, chain the newly added empty can quantity, full carton quantity, bulk carton quantity, weighing data and cosine similarity score.

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

[0126] Further, in step S6, the twin neural network comparison model outputs the cosine similarity score of the visual branch and the weight branch. When the cosine similarity score is lower than the preset score threshold, an access exception signal is generated. Then, the following steps are included:

[0127] Step S100, extract a plurality of historical oil consumption data based on the oil supply tank, each historical maintenance data including the model code of each maintenance, the oil consumption, the maintenance duration, and the seasonal parameter.

[0128] Step S200, aggregate the model code of each maintenance, the oil consumption, the maintenance duration, and the seasonal parameter according to the matched maintenance event ID through a time series alignment algorithm, to obtain a training data set containing a timestamp.

[0129] Preferably, the time series alignment algorithm can eliminate records with missing values exceeding 30%, and finally form a complete training data set containing a timestamp.

[0130] Preferably, feature transformation can be performed on the cleaned data, the model code is converted into a 16-dimensional dense vector through an embedding layer, the maintenance duration is standardized by Z-score, and the seasonal parameter is converted into a sine / cosine periodic feature. A sliding window technique is used to generate training samples, each window contains a sequence of maintenance events for 30 consecutive days, the window step is 7 days, and a three-dimensional tensor (number of samples x time step x number of features) is formed.

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

[0132] Step S400, input the training data set into the bidirectional LSTM network architecture, and output a prediction data set based on a preset prediction step through the bidirectional LSTM network architecture, the prediction data set including predicted oil consumption data.

[0133] Preferably, the model prediction result can be input into a Bayesian optimization module, taking the prediction error distribution of the past 12 months as the prior probability, combining the current inventory level, the procurement cycle, the flight plan and other constraint conditions, to calculate the optimal inventory warning threshold. The threshold value is updated according to a reinforcement learning strategy, when the prediction accuracy rate exceeds a certain percentage (for example, 95%) for 5 consecutive times, the threshold value adjustment range is automatically relaxed to reduce frequent alarms.

[0134] Step S500, extract real-time oil consumption data based on the oil supply tank, and obtain the difference between the real-time oil consumption data and the predicted oil consumption data.

[0135] Step S600, when the difference exceeds a preset deviation threshold, generate an oil purchase request and send it to an external monitoring end.

[0136] Preferably, the Isolation Forest anomaly detection algorithm is triggered when the usage of a certain type of oil exceeds the predicted value + 3σ (σ is the historical error standard deviation) for 3 consecutive times, which is the preset deviation threshold. After verifying the authenticity of the anomaly, the system generates a purchase suggestion sheet containing the recommended quantity and emergency level based on the supplier's delivery cycle and the Economic Order Quantity (EOQ) model.

[0137] Preferably, the model incremental training can be performed every month. After the newly generated operation data is denoised by the Adversarial Autoencoder (AAE), it is fine-tuned at a learning rate of 10%. At the same time, the federated learning framework aggregates the edge node data of multiple airports to update the global model parameters, so that the prediction model continues to adapt to the changes in consumption patterns brought about by new aircraft maintenance modes.

[0138] Further, in step S1, in response to the oil supply cabinet opening request of the external staff, the FaceNet deep convolutional neural network is used to extract the face features, and the authorized information matched with the face features is called from the external database, including the following steps:

[0139] Step S11, in response to the oil supply cabinet opening request of the external staff, the face image of the external staff is obtained through the external shooting device.

[0140] Preferably, the oil supply cabinet can be equipped with an industrial-grade camera (resolution 1920x1080, frame rate 30fps) to continuously monitor the cabinet door area. When detecting personnel approaching (triggered by an infrared distance sensor), the camera automatically focuses and captures the face image. The FaceNet model (pre-trained on the CASIA-WebFace dataset) built-in the system runs in real time to extract a 512-dimensional face feature vector.

[0141] Step S12, the high-dimensional face feature vector of the face image is extracted through the FaceNet deep convolutional neural network.

[0142] Step S13, the high-dimensional face feature vector is L2 normalized to obtain a standardized feature vector.

[0143] Preferably,

[0144] Step S14, the image confidence of the standardized feature vector is obtained through the MTCNN face detection algorithm.

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

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

[0147] Step S16, the similarity between the standardized feature vector and all pre-stored face records in the external database is calculated.

[0148] Preferably, the standardized feature vector is transmitted to the external authorized database through an encrypted channel (AES-256). The database uses a graph database (Neo4j) for storage, and the nodes contain personnel ID, face feature hash value, permission level (such as maintenance personnel / administrator), valid time window, etc. When querying, the cosine similarity between the feature vector and all records in the library is calculated, and the authorized records with similarity>0.8 are returned.

[0149] Step S17, when the similarity is greater than the preset similarity threshold, the authorized information matched with the face feature is called from the external database.

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

[0151] Further, in step S2, if there is authorized information, the lubricating oil supply cabinet is opened, and after the external worker completes the storage operation, the weighing data is obtained through the weighing function and the weighing data is processed based on the gradient boosting decision tree model to obtain the optimal solution for taking out the combination of bulk and bulk, including:

[0152] Step S21, if there is authorized information, the lubricating oil supply cabinet is opened.

[0153] Step S22, after the external worker completes the storage operation, the weight data of a preset continuous time length is collected at a preset collection frequency through the weighing function.

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

[0155] Preferably, when the external worker completes the storage operation, the weighing sensor of the lubricating oil supply cabinet can collect weight data for 30 seconds at a frequency of 10Hz. Mechanical vibration noise is eliminated by sliding window mean filtering, and then the weight change rate features are extracted by first-order difference method to form a three-dimensional time sequence feature matrix including original weight, change amount and change acceleration.

[0156] Step S24, the historical access data of the lubricating oil supply cabinet based on a plurality of historical time points are obtained and divided into a sample set and a validation set.

[0157] Preferably, the training set and the validation set can be divided according to 7:3.

[0158] Step S25, a gradient boosting decision tree model is built based on the XGBoost framework, and the target variable of the gradient boosting decision tree model is defined as taking out the optimal solution, and the optimal solution is to maximize the number of full cartons and minimize the number of bulk cartons.

[0159] Preferably, the model is built using the XGBoost framework, and the target variable is the historical optimal taking solution (combination of N_closed and N_open). The tree depth (max_depth=6), learning rate (eta=0.1) and other parameters are adjusted through Bayesian hyperparameter optimization, and early stopping method is used to prevent overfitting.

[0160] Step S26, input the sample set into the gradient boosting decision tree model and select the best split feature of each decision tree through the greedy algorithm, and the best split feature of the current decision tree includes the score of the current decision tree based on each full / bulk combination.

[0161] Preferably, the weight of each tree is defined to be proportional to its F1-score (F1 value) on the validation set, and the weight distribution is normalized by the softmax function; the initial threshold (default 0.85) is determined according to the ROC curve analysis, and the argmax operation is performed on the weighted voting result to determine the final score corresponding to the point where the Youden index is maximum (sensitivity ≥ 90%, specificity ≥ 80%). The model uses 100 CART trees with a maximum depth of 6 mentioned above, and each tree selects the best split feature based on information gain ratio through the greedy algorithm.

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

[0163] Next, a confusion matrix is constructed to count the prediction results: 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] Next, based on the confusion matrix, the core indicators are calculated: precision (P, i.e. TP / (TP+FP)) and recall (R, i.e. TP / (TP+FN)), and then the harmonic average of the two is obtained to get the F1-score (the calculation formula is 2PR / (P+R)).

[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, the weight values of the two branches are initialized by Xavier method to share the convolution kernel parameters of the two branches.

[0177] Step S43, the visual features of empty oil tanks, unopened oil tanks, whole cartons, and bulk cartons are obtained as anchor point samples.

[0178] Step S44, the weighing data of normal operation is taken as a positive sample, and the weighing data of abnormal operation is taken as a negative sample.

[0179] Step S45, the anchor point sample, the positive sample, and the negative sample are pre-trained by a triplet loss function, so that the symmetric neural network structure forms a compact intra-class distance and a dispersed inter-class distance in the feature space, and a Siamese neural network contrast model is obtained.

[0180] Preferably, the anchor point sample is a standard feature of visual recognition (such as the bottle mouth features of 8 empty tanks), the positive sample is the matching weight sensor data (8x empty tank reference 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 to distinguish negative samples are selected in each training round, so that the model forms a compact intra-class distance and a dispersed inter-class distance in the feature space.

[0181] Preferably, a contrast loss can be introduced in the later stage of pre-training, which forces the output features of the visual branch and the weight branch to satisfy the cosine similarity of the matching sample pair >0.9, the cosine similarity of the unmatched sample pair <0.3, and 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, and the relationship between the similarity score and the false alarm rate is modeled by Gaussian process regression. The Bayesian optimization module continuously receives actual running data, automatically adjusts the decision threshold (initial value 0.85, update step ±0.01), and makes the system maintain stable performance in extreme scenarios such as metal reflection and fast operation.

[0183] Further, step S6, the cosine similarity score of the visual branch and the weight branch is output by the Siamese neural network contrast model, and when the cosine similarity score is lower than a preset score threshold, an access abnormal signal is generated, which specifically includes the following steps:

[0184] Step S61, the visual feature vector output by the visual branch and the time series weighing feature vector output by the weight branch are respectively subjected to L2 normalization, and standardized visual feature vectors and standardized time series weighing feature vectors are sequentially obtained.

[0185] Preferably, the visual branch features include geometric features of the bottle mouth metal ring, paper box corner point positions, etc., and the weight branch features include weight change rate, fluctuation amplitude, etc. time series statistics.

[0186] In step S62, the standardized visual feature vector and the standardized time series weight feature vector are mapped to the feature space by the twin neural network comparison model, and the cosine similarity score is calculated by formula (1):

[0187] (1);

[0188] wherein, is the cosine similarity score, is the standardized visual feature vector, is the standardized time series weight feature vector.

[0189] Preferably, the score range is between [-1, 1], 1 indicates complete match, and -1 indicates complete conflict.

[0190] In step S63, when the cosine similarity score is lower than the preset score threshold, an access abnormal signal is generated.

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

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

[0193] The embodiment extracts face features by using a FaceNet deep convolutional neural network in response to an external worker's lubricating oil supply cabinet opening request, and calls authorized information matching the face features from an external database; if there is authorized information, the lubricating oil supply cabinet is opened, and based on the external worker completing the storage operation, weighing data is obtained through a weighing function and the weighing data is processed based on a gradient boosting decision tree model to obtain a take-out optimal solution of the combination of bulk packaging and non-bulk packaging; the take-out optimal solution is sent to the external worker, and after the external worker completes the take-out operation, the bottle opening features of the lubricating oil tank are recognized by OpenCV Hough circle detection, the corner points of the bulk packaging carton are recognized by Yolo algorithm, and the three-dimensional geometric features of the uncapped carton are recognized by Mask R-CNN instance segmentation model to obtain the number of returned empty cans, the number of bulk packaging cartons, and the number of non-bulk packaging cartons; a pre-trained twin neural network comparison model is constructed, which includes a visual branch and a weight branch and shares convolution kernel parameters; the number of returned empty cans, the number of bulk packaging cartons, and the number of non-bulk packaging 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; the cosine similarity score of the visual branch and the weight branch is output by the twin neural network comparison model, and when the cosine similarity score is lower than a preset score threshold, an access abnormal signal is generated.

[0194] The embodiment solves the two core problems of low efficiency of manual operation and difficulty in guaranteeing the consistency of material and account in the field of aviation lubricating oil management through the integration of a multi-modal intelligent perception and decision system. The traditional management mode relies on manual verification and paper records, which has the defects of response delay, easy error and inability to monitor in real time, while the present scheme realizes the automatic tracking of the whole process of lubricating oil access through the cooperation of the weighing sensor of the lubricating oil supply cabinet and the computer vision algorithm (FaceNet, YOLO, Mask R-CNN). The breakthrough of the core technology of the embodiment lies in dynamic optimal solution calculation, which analyzes the weighing data by a gradient boosting decision tree model, intelligently recommends the take-out combination of bulk packaging and non-bulk packaging lubricating oil, and reduces the manual decision time; multi-modal cross-validation uses a twin neural network to compare the visual recognition results and the weight change data, and triggers an abnormal alarm when the cosine similarity is lower than the threshold, fundamentally eliminating the risk of inconsistency between material and account caused by human negligence in traditional management. In addition, the face recognition authorization mechanism ensures strict control of operation permissions, and realizes compliance requirements for the management of high-risk goods in the field of aviation maintenance from a technical point of view.

[0195] As shown in Figure 2 , the embodiment provides an embodiment of an access management device for aviation lubricating oil.

[0196] Specifically, the access management device comprises, in sequence, an electrically or signal connected face matching module 1, an optimal solution extraction 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.

[0197] The face matching module 1 is configured to extract a face feature by using a FaceNet deep convolutional neural network in response to an oil supply cabinet opening request of an external worker, and call authorized information matched with the face feature from an external database; the optimal solution extraction calculation module 2 is configured to open the oil supply cabinet if the authorized information exists, obtain weighing data by a weighing function after the external worker completes a storage operation, and process the weighing data based on a gradient boosting decision tree model to output an optimal solution of combination of bulk and packaged; the image recognition module 3 is configured to send the optimal solution to the external worker, and obtain a returned empty tank quantity, a packaged carton quantity, and a bulk carton quantity by sequentially recognizing a bottle opening feature of an oil tank by an OpenCV Hough circle detection, recognizing a packaged carton corner point by a Yolo algorithm, and recognizing a three-dimensional geometric feature of an uncapped carton by a Mask R-CNN instance segmentation model after the external worker completes a taking operation; the comparison model definition module 4 is configured to construct a pre-trained twin neural network comparison model including a visual branch and a weight branch and sharing convolution kernel parameters; the comparison model input module 5 is configured to input the returned empty tank quantity, the packaged carton quantity, and the bulk carton quantity in the form of a feature vector to the visual branch, and input the weighing data in the form of time series data to the weight branch; and the comparison model output module 6 is configured to output a cosine similarity score of the visual branch and the weight branch by the twin neural network comparison model, and generate an access exception signal when the cosine similarity score is lower than a preset score threshold.

[0198] Further, the access management device further comprises, in sequence, a data chaining module, an alliance chain access port construction module, an alliance chain access port supervision module, a zk-SNARK protocol deployment module, and an additional data chaining module; the data chaining module is electrically or signal connected with the comparison model output module 6.

[0199] The data chaining module is configured to upload the returned empty can quantity, the whole carton quantity, the bulk carton quantity, the weighing data, and the cosine similarity score to the alliance chain through different special channels; the alliance chain access port construction module is configured to associate all the special channels through the composite key DataType_NodeID_Timestamp, and construct the access port of the alliance chain through periodic replacement of consensus nodes based on the RBFT consensus mechanism; the alliance chain access port supervision module is configured to supervise the access port through the Solidity in the blockchain to create the smart contract code; the zk-SNARK protocol deployment module is configured to deploy the zk-SNARK protocol in the blockchain, and encrypt the smart contract code based on the zero-knowledge proof privacy protection, so that the smart contract code is only open to the authorized identity of the external supervision end; and the additional data chaining module is configured to call the smart contract code through the preset Python script, and perform data chaining on the newly added returned empty can quantity, whole carton quantity, bulk carton quantity, weighing data, and cosine similarity score based on the authorized identity.

[0200] Further, the access management device further comprises a historical oil consumption data extraction module, a training data aggregation module, a bidirectional LSTM network architecture construction module, a predicted data acquisition module, an oil consumption data difference acquisition module, and an oil purchase request generation and sending module connected in sequence; the historical oil consumption data extraction module is electrically or signal connected with the comparison model output module 6.

[0201] The historical oil consumption data extraction module is configured to extract a plurality of historical oil consumption data based on the oil supply cabinet, and each historical maintenance data includes an aircraft model code, an oil consumption amount, a maintenance duration, and a season parameter; the training data aggregation module is configured to aggregate the aircraft model code, the oil consumption amount, the maintenance duration, and the season parameter according to a matched maintenance event ID through a time series alignment algorithm, to obtain a training data set containing a timestamp; the bidirectional LSTM network architecture construction module is configured to construct a pre-trained bidirectional LSTM network architecture, wherein a loss function of the bidirectional LSTM network architecture adopts a Huber loss, an optimizer selects Nadam, and an initial learning rate is set to 0.001; the predicted data acquisition module is configured to input the training data set into the bidirectional LSTM network architecture, and output a predicted data set based on a preset prediction step through the bidirectional LSTM network architecture, wherein the predicted data set includes predicted oil consumption data; the oil consumption data difference acquisition module is configured to extract real-time oil consumption data based on the oil supply cabinet, and acquire a difference between the real-time oil consumption data and the predicted oil consumption data; and the oil purchase request generation and sending module is configured to generate an oil purchase request and send it to an external monitoring end when the difference exceeds a preset deviation threshold.

[0202] Further, the face matching module 1 specifically comprises 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 connected in sequence in electrical or signal connection; the seventh face matching unit is connected in electrical or signal connection with the optimal solution taking calculation module 2.

[0203] The first face matching unit is configured to obtain a face image of an external worker through an external shooting device in response to an oil supply cabinet opening request of the external worker; the second face matching unit is configured to extract a high-dimensional face feature vector of the face image through a FaceNet deep convolutional neural network; the third face matching unit is configured to perform L2 normalization on the high-dimensional face feature vector to obtain a standardized feature vector; the fourth face matching unit is configured to obtain an image confidence of the standardized feature vector through an MTCNN face detection algorithm; the fifth face matching unit is configured to transmit the standardized feature vector to an external database when the image confidence exceeds a preset confidence threshold; the sixth face matching unit is configured to calculate a similarity between the standardized feature vector and all pre-stored face records in the external database; and the seventh face matching unit is configured to call authorized information matched with the face feature from the external database when the similarity is greater than a preset similarity threshold.

[0204] Further, the optimal solution taking calculation module 2 specifically comprises a first optimal solution taking calculation unit, a second optimal solution taking calculation unit, a third optimal solution taking calculation unit, a fourth optimal solution taking calculation unit, a fifth optimal solution taking calculation unit, a sixth optimal solution taking calculation unit, a seventh optimal solution taking calculation unit, an eighth optimal solution taking calculation unit, and a ninth optimal solution taking calculation unit connected in sequence in electrical or signal connection; the first optimal solution taking calculation unit is connected in electrical or signal connection with the seventh face matching unit, and the ninth optimal solution taking calculation unit is connected in electrical or signal connection with the image recognition module 3.

[0205] The first optimal solution taking out calculation unit is configured to open the lubricating oil supply cabinet if the authorization information exists; the second optimal solution taking out calculation unit is configured to collect weight data of a preset continuous time length at a preset collection frequency through a weighing function based on the external staff completing the storage operation; the third optimal solution taking out calculation unit is configured to extract weight change rate characteristics of all weight data through a first-order difference method to obtain a three-dimensional time sequence characteristic matrix including original weight, change amount and change acceleration; the fourth optimal solution taking out calculation unit is configured to obtain historical access data of the lubricating oil supply cabinet based on a plurality of historical time points and divide the historical access data into a sample set and a verification set; the fifth optimal solution taking out calculation unit is configured to construct a gradient boosting decision tree model based on an XGBoost framework, and define the target variable of the gradient boosting decision tree model as the optimal solution taking out, the optimal solution taking out being to maximize the number of full-packed cartons and minimize the number of bulk cartons; the sixth optimal solution taking out calculation unit is configured to input the sample set into the gradient boosting decision tree model and select the best split feature of each decision tree through a greedy algorithm, the best split feature of the current decision tree including a score of the current decision tree based on each full-packed / bulk-packed combination; the seventh optimal solution taking out calculation unit is configured to obtain an F1-score performance index of each best split feature based on the verification set respectively, and define the weight of each decision tree to be proportional to the respective F1-score performance index; the eighth optimal solution taking out calculation unit is configured to calculate a weighted total score of all decision trees based on each full-packed / bulk-packed combination according to the scores and weights of all decision trees; and the ninth optimal solution taking out calculation unit is configured to define the full-packed / bulk-packed combination with the highest weighted total score as the optimal solution taking out.

[0206] Further, the comparative model definition module 4 specifically includes a first comparative model definition unit, a second comparative model definition unit, a third comparative model definition unit, a fourth comparative model definition unit and a fifth comparative model definition unit which are sequentially electrically or signal connected; the first comparative model definition unit is electrically or signal connected with the image recognition module 3, and the fifth comparative model definition unit is electrically or signal connected with the comparative model input module 5.

[0207] The first contrast model defining unit is configured to construct a symmetric neural network structure comprising a visual branch and a weight branch, both of which adopt ResNet-18 as a basic architecture; the second contrast model defining unit is configured to initialize the weight values of the two branches by the Xavier method to share the convolution kernel parameters of the two branches; the third contrast model defining unit is configured to obtain the visual features of empty lubricating oil tanks, unopened lubricating oil tanks, whole paper boxes, and bulk paper boxes as anchor point samples; the fourth contrast model defining unit is configured to take the weighing data of normal operation as positive samples and the weighing data of abnormal operation as negative samples; and the fifth contrast model defining unit is configured to pre-train the anchor point samples, the positive samples, and the negative samples by a triplet loss function, so that the symmetric neural network structure forms a compact intra-class distance and a dispersed inter-class distance in a feature space, and a twin neural network contrast model is obtained.

[0208] Further, the contrast model output module 6 specifically comprises a first contrast model output unit, a second contrast model output unit, and a third contrast model output unit which are sequentially electrically or signal connected; the first contrast model output unit is electrically or signal connected with the contrast model input module 5.

[0209] The first contrast model output unit is configured 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 sequentially obtain a standardized visual feature vector and a standardized time-series weighing feature vector.

[0210] The second contrast model output unit is configured to map the standardized visual feature vector and the standardized time-series weighing feature vector to a feature space by the twin neural network contrast model, and calculate a cosine similarity score by formula (1):

[0211] (1);

[0212] wherein, is the cosine similarity score, is the standardized visual feature vector, is the standardized time-series weighing feature vector.

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

[0214] It should be noted that the present embodiment is a functional module embodiment based on the above-mentioned method embodiment, and the preferred, expanded, limited, exemplified, and principle explained parts of the present embodiment can be seen in the above-mentioned embodiment, which will not be described herein again.

[0215] The embodiment extracts face features by using a FaceNet deep convolutional neural network in response to an external worker's lubricating oil supply cabinet opening request, and calls authorized information matching the face features from an external database; if there is authorized information, the lubricating oil supply cabinet is opened, and based on the external worker completing the storage operation, weighing data is obtained through a weighing function and the weighing data is processed based on a gradient boosting decision tree model to obtain a take-out optimal solution of the combination of bulk packaging and loose packaging; the take-out optimal solution is sent to the external worker, and after the external worker completes the take-out operation, the bottle opening features of the lubricating oil tank are recognized by OpenCV Hough circle detection, the bulk packaging carton corner points are recognized by Yolo algorithm, and the three-dimensional geometric features of the uncapped carton are recognized by Mask R-CNN instance segmentation model to obtain the number of returned empty cans, the number of bulk packaging cartons and the number of loose packaging cartons; a pre-trained twin neural network comparison model is constructed, which includes a visual branch and a weight branch and shares convolution kernel parameters; the number of returned empty cans, the number of bulk packaging cartons and the number of loose packaging 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; the cosine similarity score of the visual branch and the weight branch is output by the twin neural network comparison model, and when the cosine similarity score is lower than a preset score threshold, an access abnormal signal is generated.

[0216] The embodiment solves the two core problems of low efficiency of manual operation and difficulty in guaranteeing the consistency of material and account in the field of aviation lubricating oil management through the integration of a multi-modal intelligent perception and decision system. The traditional management mode relies on manual verification and paper records, which has the defects of response delay, easy error and inability to monitor in real time, while the present scheme realizes the automatic tracking of the whole process of lubricating oil access through the cooperation of the weighing sensor of the lubricating oil supply cabinet and the computer vision algorithm (FaceNet, YOLO, Mask R-CNN). The breakthrough of the core technology of the embodiment lies in dynamic optimal solution calculation, which analyzes the weighing data by a gradient boosting decision tree model, intelligently recommends the take-out combination of bulk packaging and loose packaging lubricating oil, and reduces the manual decision time; multi-modal cross verification uses a twin neural network to compare the visual recognition result and the weight change data, and triggers an abnormal alarm when the cosine similarity is lower than the threshold, which fundamentally eliminates the risk of inconsistency between material and account caused by human negligence in traditional management. In addition, the face recognition authorization mechanism ensures strict control of operation permission, and realizes the compliance requirement of aviation maintenance field for the management of high-risk goods from a technical aspect.

[0217] Figure 3 Fig. 1 is a structural schematic diagram of an electronic device according to an embodiment of the present application. As shown in Figure 3 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 method of managing access to the aviation lubricant.

[0219] The processor 71 is configured to execute the program instructions stored in the memory 72 to manage access to the aviation lubricant.

[0220] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 can be an integrated circuit chip having a processing capability. The processor 71 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application-Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0221] Further, Figure 4 FIG. 8 is a structural schematic diagram of the storage medium according to an embodiment of the present application. Figure 4 The storage medium 8 according to the embodiment of the present application stores program instructions 81 capable of implementing all the methods described above. The program instructions 81 can be stored in the storage medium in the form of a software product, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the embodiments of the present application. The storage medium described above includes a U disk, a mobile hard disk, a ROM (Read-Only Memory), a RAM (Random Access Memory), a magnetic disk or an optical disk, and various media capable of storing program codes, or a terminal device such as a computer, a server, a mobile phone, a tablet, etc.

[0222] In the several embodiments provided in the present application, it should be understood that the disclosed system, system and method can be implemented in other ways. For example, the system embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, another division manner can be adopted, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces. The coupling or communication connection can be electrical, mechanical, signal or other forms.

[0223] In addition, the various functional units in the embodiments of the present application can be integrated in one processing unit, or each can exist physically as a separate unit, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware, or in the form of a software functional unit. The above is only an implementation of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An access management method for aviation lubricating oil, said access management method being applied to a plurality of lubricating oil tanks of the same size placed in a lubricating oil supply tank, said lubricating oil supply tank having a weighing function, all the lubricating oil tanks being in at least one of a full load state by a carton or a bulk state, characterized in that, The access management method comprises: Step S1, in response to an oil supply cabinet opening request of an external worker, extracting a face feature by using a FaceNet deep convolutional neural network, and calling authorization information matched with the face feature from an external database; Step S2, if the authorization information exists, opening the oil supply cabinet, obtaining weighing data by the weighing function after the external worker completes the storage operation, and processing the weighing data based on a gradient boosting decision tree model to obtain a take-out optimal solution of the bulk / combination; Step S3, sending the take-out optimal solution to the external worker, and after the external worker completes the take-out operation, identifying the bottle opening features of the oil tank by OpenCV Hough circle detection, identifying the whole paper box corner points by Yolo algorithm, and identifying the three-dimensional geometric features of the uncapped paper box by MaskR-CNN instance segmentation model to obtain the number of returned empty cans, the number of whole paper boxes, and the number of bulk paper boxes; Step S4, constructing a pre-trained twin neural network contrast model including a visual branch and a weight branch, which share convolution kernel parameters; Step S5, inputting the number of returned empty cans, the number of whole paper boxes, and the number of bulk paper boxes in the form of feature vectors into the visual branch, and inputting the weighing data in the form of time series data into the weight branch; Step S6, outputting the cosine similarity score of the visual branch and the weight branch by the twin neural network contrast model, and generating an access abnormal signal when the cosine similarity score is lower than a preset score threshold; Step S4, constructing a pre-trained twin neural network contrast model including a visual branch and a weight branch, which share convolution kernel parameters, comprising: Step S41, constructing a symmetric neural network structure including a visual branch and a weight branch, both branches using ResNet-18 as the basic architecture; Step S42, initializing the weight values of the two branches by the Xavier method to share the convolution kernel parameters of the two branches; Step S43, obtaining the visual features of empty oil tanks, unopened oil tanks, whole paper boxes, and bulk paper boxes as anchor point samples; Step S44, taking the weighing data of normal operation as positive samples and the weighing data of abnormal operation as negative samples; Step S45, pre-training the anchor point samples, the positive samples, and the negative samples by a triplet loss function to make the symmetric neural network structure form a compact intra-class distance and a dispersed inter-class distance in the feature space, and obtaining the twin neural network contrast model.

2. The access management method according to claim 1, characterized by, Step S6, outputting the cosine similarity score of the visual branch and the weight branch by the twin neural network contrast model, and generating an access abnormal signal when the cosine similarity score is lower than a preset score threshold, and further comprising: Step S10, uploading the number of returned empty cans, the number of whole paper boxes, the number of bulk paper boxes, the weighing data, and the cosine similarity score to the alliance chain through different special channels. Step S20, all dedicated channels are associated through the composite key DataType_NodeID_Timestamp, and an access port of the consortium chain is constructed through periodic replacement consensus nodes based on an RBFT consensus mechanism; Step S30, the access port is supervised by creating a smart contract code in a blockchain through Solidity; Step S40, a zk-SNARK protocol is deployed in the blockchain, and the smart contract code is encrypted based on zero-knowledge proof privacy protection, so that the smart contract code is only open to the authorized identity of an external supervision end; Step S50, the smart contract code is called through a preset Python script, and the newly added returned empty can quantity, whole carton quantity, bulk carton quantity, weighing data, and cosine similarity score are data-chained based on the authorized identity.

3. The access management method according to claim 1, characterized by, Step S6, the cosine similarity score of the visual branch and the weight branch is output by comparing the twin neural network model, and when the cosine similarity score is lower than a preset score threshold, an access exception signal is generated, and then the method further comprises: Step S100, a plurality of historical oil consumption data are extracted based on the oil supply cabinet, and each historical maintenance data comprises an aircraft model code, an oil consumption amount, a maintenance duration, and a season parameter; Step S200, the aircraft model code, the oil consumption amount, the maintenance duration, and the season parameter of each maintenance are aggregated according to a matched maintenance event ID through a time series alignment algorithm, to obtain a training data set containing a timestamp; Step S300, a pre-trained bidirectional LSTM network architecture is constructed, a loss function of the bidirectional LSTM network architecture adopts Huber loss, an optimizer selects Nadam, and an initial learning rate is set to 0.001; Step S400, the training data set is input into the bidirectional LSTM network architecture, and a prediction data set based on a preset prediction step is output by the bidirectional LSTM network architecture, the prediction data set comprising predicted oil consumption data; Step S500, real-time oil consumption data are extracted based on the oil supply cabinet, and a difference between the real-time oil consumption data and the predicted oil consumption data is obtained; Step S600, when the difference exceeds a preset deviation threshold, an oil purchase request is generated and sent to an external monitoring end.

4. The access management method according to claim 1, characterized by, Step S1, in response to an oil supply cabinet opening request of an external worker, a face feature is extracted by using a FaceNet deep convolutional neural network, and authorized information matched with the face feature is called from an external database, comprising: Step S11, in response to an oil supply cabinet opening request of an external worker, a face image of the external worker is obtained through an external shooting device; Step S12, a high-dimensional face feature vector of the face image is extracted by using the FaceNet deep convolutional neural network; Step S13, the high-dimensional face feature vector is subjected to L2 normalization to obtain a standardized feature vector; Step S14, an image confidence of the standardized feature vector is obtained by using an 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 by, 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 5, wherein 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）； wherein, is the cosine similarity score, is the normalized visual feature vector, is the normalized temporal weighting feature vector; Step S63: When the cosine similarity score is lower than the preset score threshold, the access anomaly signal is generated.

7. An access management device for an aviation lubricant, the access management device being applied to the access management method according to any one of claims 1 to 6, characterized in that, The access management device comprises: A face matching module configured to, in response to an oil supply cabinet opening request from an external worker, extract a face feature using a FaceNet deep convolutional neural network and call authorized information matching the face feature from an external database; An optimal solution extraction module configured to, if the authorized information exists, open the oil supply cabinet, obtain weighing data through a weighing function after the external worker completes a deposit operation, and process the weighing data based on a gradient boosting decision tree model to output an optimal solution for extracting a combination of bulk and packaged goods; An image recognition module configured to send the optimal solution for extracting to the external worker, and after the external worker completes an extraction operation, recognize a bottle opening feature of an oil tank through OpenCV Hough circle detection, recognize a packaged carton corner point through Yolo algorithm, and recognize a three-dimensional geometric feature of an uncapped carton through a Mask R-CNN instance segmentation model to sequentially obtain a number of returned empty tanks, a number of packaged cartons, and a number of bulk cartons; A comparative model defining module configured to construct a pre-trained twin neural network comparative model including a visual branch and a weight branch that share convolution kernel parameters; A comparative model input module configured to input the number of returned empty tanks, the number of packaged cartons, and the number of bulk cartons in the form of feature vectors to the visual branch, and input the weighing data in the form of time series data to the weight branch; A comparative model output module configured to output a cosine similarity score of the visual branch and the weight branch through the twin neural network comparative model, and generate an access anomaly signal when the cosine similarity score is lower than a preset score threshold.

8. An electronic device, comprising: The computer readable storage medium stores program instructions that, when executed by a processor, implement the access management method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions that, when executed by a processor, implement the access management method of any one of claims 1 to 6.

Images