Tracking and tracing method and system based on RFID internet of things technology

Abstract

The application discloses a tracking and tracing method and system based on RFID Internet of Things technology, comprising: firstly, realizing full-link abnormal risk quantification through multi-source data fusion and dynamic evaluation. The method comprises: performing feature embedding processing on RFID full-link sensing data of a tracked article, and extracting a sensing feature vector; generating a fusion deep feature reflecting full-link correlation through deep feature fusion; evaluating the fusion deep feature based on multiple tracing links, and obtaining abnormal confidence of each link; analyzing the sensing feature vector to obtain an influence coefficient of each link on full-link abnormality; and dynamically aggregating the abnormal confidence and the influence coefficient, and calculating a full-link abnormal probability. The method solves the multi-source data time sequence alignment and feature fusion problem, realizes accurate positioning of each link abnormality and dynamic evaluation of full-link risk, is suitable for full-link tracking scenes of agricultural products, aquatic products and the like, and improves quality control efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a tracking and tracing method and system based on RFID Internet of Things technology. BACKGROUND

[0002] Traditional tracking and tracing methods mostly rely on manual recording or single sensor data, and have problems such as scattered data, chaotic time sequence, and fuzzy abnormal positioning. With the development of RFID Internet of Things technology, although automatic collection of full-link sensing data can be achieved, multi-source data is time sequence dislocated due to differences in collection frequency and device type, and traditional methods cannot efficiently fuse multi-source features, cannot quantify the influence weight of each link on full-link abnormality, and thus the abnormality evaluation result is not accurate enough to support dynamic quality control requirements. SUMMARY

[0003] The purpose of the present application is to provide a tracking and tracing method and system based on RFID Internet of Things technology.

[0004] In a first aspect, an embodiment of the present application provides a tracking and tracing method based on RFID Internet of Things technology, comprising:

[0005] performing feature embedding processing on RFID full-link sensing data of a tracking item to obtain a sensing feature vector of the RFID full-link sensing data;

[0006] performing deep feature fusion processing on the sensing feature vector to obtain a fused deep feature of the sensing feature vector;

[0007] determining a plurality of tracing links corresponding to the RFID full-link sensing data, performing link abnormality evaluation on the fused deep feature based on each tracing link to obtain a link abnormality confidence of each tracing link;

[0008] performing link influence degree analysis processing on the sensing feature vector to obtain a link influence coefficient of each tracing link;

[0009] determining a full-link abnormality probability of the tracking item based on the link abnormality confidence of each tracing link and the link influence coefficient of each tracing link.

[0010] In a second aspect, an embodiment of the present application provides a server system comprising a server, wherein the server is configured to execute the method of the first aspect.

[0011] Compared with the prior art, the application provides the beneficial effects including: adopting the application discloses a tracking and tracing method and system based on RFID Internet of Things technology, which comprises the following steps: firstly, realizing full-link abnormal risk quantification through multi-source data fusion and dynamic evaluation. The method comprises the following steps: performing feature embedding processing on RFID full-link sensing data of the tracked article, and extracting a sensing feature vector; generating a fusion deep feature reflecting full-link correlation through deep feature fusion; evaluating the fusion deep feature based on multiple traceability links to obtain abnormal confidence of each link; analyzing the sensing feature vector to obtain an influence coefficient of each link on the full-link abnormality; and dynamically aggregating the abnormal confidence and the influence coefficient to calculate a full-link abnormal probability. The method solves the problems of multi-source data time alignment and feature fusion, realizes accurate positioning of each link abnormality and dynamic evaluation of full-link risk, is suitable for full-link tracking scenes of agricultural products, aquatic products and the like, and improves quality control efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0013] Figure 1 A step flowchart of the tracking and tracing method based on RFID Internet of Things technology provided by the embodiments of the application is shown in the figure.

[0014] Figure 2 A structural schematic block diagram of the computer device provided by the embodiments of the application is shown in the figure. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the application more clear, the following will combine the drawings in the embodiments of the application to clearly and completely describe the technical solutions in the embodiments of the application. Obviously, the described embodiments are some of the embodiments of the application, not all the embodiments. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations.

[0016] The specific embodiments of the application will be described in detail below with reference to the drawings.

[0017] In order to solve the technical problems in the foregoing background art, Figure 1 A flowchart of the tracking and tracing method based on RFID Internet of Things technology provided by the embodiments of the present disclosure is shown in the figure, and the following will introduce the tracking and tracing method based on RFID Internet of Things technology in detail.

[0018] Step S101, feature embedding processing is performed on the RFID full-link perception data of the tracked article to obtain a perception feature vector of the RFID full-link perception data;

[0019] Step S102, deep feature fusion processing is performed on the perception feature vector to obtain a fusion deep feature of the perception feature vector;

[0020] Step S103, a plurality of traceability links corresponding to the RFID full-link perception data are determined, and the fusion deep feature is respectively evaluated for link abnormality based on each traceability link to obtain a link abnormality confidence of each traceability link;

[0021] Step S104, link influence degree analysis processing is performed on the perception feature vector to obtain a link influence coefficient of each traceability link;

[0022] Step S105, based on the link abnormality confidence of each traceability link and the link influence coefficient of each traceability link, a full-link abnormality probability of the tracked article is determined.

[0023] In the embodiment of the present application, an exemplary full-link tracking and tracing scenario of aquatic products (such as soft-shelled turtles) of a certain aquaculture enterprise is taken as an example, and the specific implementation process of the tracking and tracing method based on RFID Internet of Things technology is described in detail with the server as the execution subject. In the full-link process of aquatic products from cultivation to sales, RFID reading and writing devices and supporting sensors are deployed at each link (such as cultivation, transportation, warehousing, and sales) to collect perception information such as environmental data and operation records associated with aquatic products in real time, forming RFID full-link perception data. The server needs to analyze the data to evaluate the full-link abnormality risk. The specific execution steps are as follows:

[0024] First, the server needs to analyze the trajectory features of the RFID full-link sensing data and extract the sensing feature vector. Since the full-link sensing data usually comes from multiple sources (such as temperature and humidity sensors in breeding ponds, positioning and temperature control devices in transportation vehicles, and scanning records at sales terminals), there are time sequence misalignment (such as hourly collection in breeding stages and minute collection in transportation stages) and dimension differences (such as numerical temperature data and counting operation times). Therefore, the server calls a pre-trained traceability anomaly classification network, which includes a time sequence alignment submodule. The time sequence alignment submodule first identifies the device types and stage attributes corresponding to the multi-source sensing data (such as environmental sensors for breeding stages and vehicle-mounted devices for transportation stages), and divides the original data into multi-source time sequence segments (such as "water temperature time sequence segment" for breeding stages and "temperature and humidity time sequence segment" for transportation stages). Then, through the time sequence data input module, each segment is mapped to a pre-defined "time sequence-stage association space", which pre-stores the time feature range of each stage data (such as 0-180 days for breeding stages and 181-183 days for transportation stages). By matching the time stamp of the target segment with the "time sequence-stage feature section" in the association space (such as 181-183 days for transportation stages), the multi-source time sequence segments are aligned. The aligned segments are arranged into a continuous multi-source time sequence alignment sequence (a full-link data chain arranged in actual time order), and then a time sequence encoder (such as an LSTM network) is used for feature embedding mapping to convert the original sequence of length N into a low-dimensional, semantically rich sensing feature vector (e.g., a 128-dimensional vector, with each dimension representing "breeding water temperature fluctuation frequency", "transportation temperature and humidity exceeding time", etc.).

[0025] Next, the server performs deep feature fusion processing on the sensing feature vector to extract the fused deep feature. Although the sensing feature vector has been aligned, each dimension only reflects the local state of a single stage or device (such as "breeding water temperature" and "transportation emergency braking times"), and cannot directly express the associated influence of the full link. Therefore, the server calls the deep feature fusioner in the traceability anomaly classification network (such as a Transformer model or an attention mechanism network) to calculate the correlation weight of each feature dimension (e.g., "breeding water temperature continuous anomaly" has a higher correlation with "transportation temperature and humidity sudden change") and fuse the scattered feature dimensions into a fused deep feature that comprehensively reflects the full-link state (e.g., generating a chain feature that includes "breeding environment anomaly - transportation stress response - warehouse quality change").

[0026] Subsequently, the server needs to evaluate the abnormal confidence of each traceability link based on the fused deep features. The whole chain is usually divided into multiple traceability links such as breeding, transportation, storage, and sales. Each link may have specific abnormalities (such as low water temperature in the breeding link leading to weak water products, and excessive temperature and humidity in the transportation link accelerating corruption). Therefore, a link quality analysis network is integrated into the traceability anomaly classification network, which includes anomaly discriminators corresponding to each link (such as breeding discriminators and transportation discriminators). The server inputs the fused deep features into the anomaly discriminators of each link (pre-trained CNN or fully connected network), and the discriminators output the link abnormal confidence (probability value between 0 and 1) for the key features of the link (such as the breeding discriminator analyzing "water temperature" and "dissolved oxygen" sub-features, and the transportation discriminator analyzing "temperature and humidity" and "brake times" sub-features). For example, a breeding link confidence of 0.8 indicates an 80% probability of abnormality.

[0027] At the same time, the server needs to analyze the influence coefficient of each link on the whole chain. Different links have different contributions to the final result (such as breeding link abnormalities that may be weakened by timely control in subsequent transportation, and long-term storage temperature exceeding the standard in the sales link may directly lead to corruption). Therefore, the server calls the link influence degree analysis network (such as a regression model based on attention mechanism) in the traceability anomaly classification network to analyze the perception feature vector. This network learns from historical data (such as the full chain data of the past 1000 batches of water products) to calculate the correlation between each link abnormality and the final quality unqualified (i.e. link influence coefficient, such as a breeding link influence coefficient of 0.6 indicating that its abnormality accounts for 60% of the influence on the whole chain result).

[0028] Finally, the server integrates the abnormal confidence and influence coefficient of each link to calculate the whole chain abnormal probability. Specifically, through a dynamic aggregation method: multiply the abnormal confidence and influence coefficient of each link (such as breeding link 0.8 x 0.6 = 0.48, transportation link 0.3 x 0.3 = 0.09, and sales link 0.1 x 0.1 = 0.01), and then sum to get the whole chain abnormal probability (0.48 + 0.09 + 0.01 = 0.58, i.e. 58%). This probability directly reflects the risk of water product abnormalities in the whole chain, providing quality control basis for enterprises (such as focusing on checking the water temperature control equipment in the breeding link and strengthening the temperature and humidity monitoring in the transportation link).

[0029] In summary, this method collects full-chain perception data through RFID Internet of Things technology, combines deep learning network for time series alignment, feature fusion, link abnormality evaluation, and influence degree analysis, and finally realizes the precise quantification of water product whole chain abnormal risk, providing efficient and reliable technical support for agricultural product traceability.

[0030] In the embodiment of the present application, the feature embedding processing of the RFID full-link perception data of the tracked article is performed to obtain the perception feature vector of the RFID full-link perception data, which can be implemented through the following examples.

[0031] The multi-source time sequence alignment processing is performed on the RFID full-link perception data to obtain a multi-source time sequence alignment sequence of the RFID full-link perception data.

[0032] The time sequence feature embedding mapping is performed on the multi-source time sequence alignment sequence to obtain a perception feature vector of the multi-source time sequence alignment sequence.

[0033] In the embodiment of the present application, an example is taken from the whole-link tracking scene of soft-shelled turtles in a certain aquatic product enterprise. The server needs to perform feature embedding processing on the RFID sensing data of the breeding, transportation, and sales links. The specific execution process is as follows: first, the server performs multi-source time sequence alignment processing on the RFID whole-link sensing data. In the whole link of soft-shelled turtles from breeding to sales, the RFID sensing data is collected by multiple types of equipment: the underwater temperature and humidity sensor in the breeding link records the water temperature every hour (for example, 0-180 days, one data per hour), the vehicle-mounted RFID reader and the temperature and humidity meter in the transportation link collect the position and temperature every minute (for example, 181-183 days, one data per minute), and the supermarket RFID terminal in the sales link records the shelving time and storage temperature every day (for example, 184-185 days, one data per day). Due to the differences in collection frequency and equipment type, the original time stamps are misaligned (for example, the “water temperature on the 180th day at 20:00” in the breeding link and the “temperature and humidity on the 181st day at 08:00” in the transportation link are not continuous on the time axis), and the formats are mixed (numerical temperature values and text position descriptions). The server calls the time sequence alignment submodule in the pre-trained traceability anomaly classification network to first identify the “multi-source sensing data information” corresponding to each data: through data labels (such as “breeding-water temperature”, “transportation-temperature and humidity”, and “sales-storage temperature”), the data belonging to the link and the equipment type are determined, and the original data is segmented into independent multi-source time sequence segments (for example, the “water temperature time sequence segment” in the breeding link is [1st day 08:00, 25℃; 1st day 09:00, 26℃…], and the “temperature and humidity time sequence segment” in the transportation link is [181st day 08:00, 28℃ / 60%; 181st day 08:01, 28℃ / 62%…]). Then, the time sequence alignment submodule maps each segment to a pre-defined “time sequence-link correlation space”, which stores the time feature ranges of the data in each link (for example, the breeding link corresponds to 0-180 days, the transportation link corresponds to 181-183 days, and the sales link corresponds to 184-185 days). The server matches the time stamp of each time sequence segment with the “time sequence-link feature section” in the correlation space through the time sequence data input module (for example, the “181st day 08:00” in the transportation link is matched to the “181-183 day” section), adjusts the time stamp offset of the segment (for example, the “181st day 08:00” of the transportation link data is unified to “181st day 08:00” on the whole-link time axis), and fills in the missing time points (for example, when there is no data for a certain minute in the transportation link, the missing data is filled in by interpolation of the previous and subsequent data), and finally arranges the multi-source time sequence segments into a continuous “multi-source time sequence alignment sequence” (for example, [breeding-1st day 08:00-25℃, breeding-1st day 09:00-26℃… transportation-181st day 08:00-28℃ / 60%… sales-184th day 09:00-4℃]). After completing the multi-source time sequence alignment, the server performs time sequence feature embedding mapping on the sequence.The multi-source time sequence alignment sequence is arranged in chronological order, but the length can be as high as thousands or even tens of thousands (such as 180 days x 24 hours = 4320 pieces of data in the breeding link), and direct analysis is low in efficiency and redundant. The server calls the time sequence encoder (such as the LSTM network) in the time sequence alignment submodule, and inputs the sequence into the encoder. The encoder extracts time sequence features layer by layer through a recurrent neural network: the first layer learns local patterns at adjacent time points (such as "water temperature rises by 1℃ per hour"), the second layer captures trends across time periods (such as "water temperature is stable for 30 days before breeding, and drops by 5℃ on the 31st day"), and finally compresses the entire sequence into a 128-dimensional perception feature vector. Each dimension of the vector corresponds to a key feature (such as "standard deviation of breeding water temperature fluctuation", "proportion of time length of temperature and humidity exceeding standard during transportation", and "average storage temperature during sales"), which not only retains the time sequence information of the full-link data, but also removes redundant noise, providing efficient input for subsequent deep analysis. At this point, the server has completed the conversion from multi-source, misaligned RFID perception data to structured perception feature vectors, laying the foundation for full-link anomaly assessment.

[0034] In the embodiment of the application, the multi-source time sequence alignment processing of the RFID full-link perception data to obtain the multi-source time sequence alignment sequence of the RFID full-link perception data can be implemented by the following examples.

[0035] When the RFID full-link perception data of the tracked article is acquired, an origin abnormality classification network for classifying the tracked article according to the full-link origin is acquired; the origin abnormality classification network comprises a time sequence alignment submodule;

[0036] The multi-source perception data information corresponding to the RFID full-link perception data is determined through the time sequence alignment submodule, and the multi-source time sequence segment data of the RFID full-link perception data is extracted based on the multi-source perception data information;

[0037] The multi-source time sequence segment data is aligned by the time sequence alignment submodule to obtain aligned time sequence segment data of the multi-source time sequence segment data;

[0038] The aligned time sequence segment data of the multi-source time sequence segment data is arranged in full-link time sequence by the time sequence alignment submodule to obtain the multi-source time sequence alignment sequence of the RFID full-link perception data.

[0039] In the embodiment of the present application, an example is taken from the whole-link tracking scene of soft-shelled turtles in a certain aquatic product enterprise. The server processes the multi-source time alignment process of the RFID whole-link sensing data as follows: when the server receives the RFID whole-link sensing data of soft-shelled turtles (including environmental and operation data in the breeding, transportation, and sales links), it first calls the pre-trained “traceability anomaly classification network” from the local model library or the cloud. This network is a deep learning model pre-trained by the server for the whole-link traceability task of aquatic products, and its core function is to identify whole-link abnormal risks. Its internal “time alignment submodule” is specially used to solve the time disorder problem of multi-source sensing data. For example, this submodule has been trained through historical data (such as the whole-link data of the past 1000 batches of soft-shelled turtles), and is familiar with the time characteristic range and data format of the breeding (0-180 days), transportation (181-183 days), and sales (184-185 days) links. The server analyzes the meta-information (such as data labels, device IDs, and timestamps) of the RFID whole-link sensing data through the time alignment submodule, determines the “multi-source sensing data information”, i.e., the link and device type to which each piece of data belongs. For example: data with the label “breeding-temperature and humidity sensor” belongs to the breeding link and is collected by an underwater sensor, with a timestamp range of 0-180 days and a collection frequency of once an hour; data with the label “transportation-vehicle-mounted RFID” belongs to the transportation link and is collected by a vehicle-mounted RFID reader and a temperature and humidity meter, with a timestamp range of 181-183 days and a collection frequency of once a minute; data with the label “sales-terminal RFID” belongs to the sales link and is collected by a supermarket RFID terminal, with a timestamp range of 184-185 days and a collection frequency of once a day. Based on this information, the server divides the original whole-link data into independent “multi-source time segment data”: the temperature and humidity data in the breeding link is extracted as one time segment (such as [25℃ / 60% at 08:00 on the 1st day, 26℃ / 58% at 09:00 on the 1st day, … 24℃ / 62% at 20:00 on the 180th day]); the location and temperature and humidity data in the transportation link are extracted as another time segment (such as [East longitude 112° / North latitude 28°-28℃ / 60% at 08:00 on the 181st day, East longitude 112.1° / North latitude 28.1°-28℃ / 62% at 08:01 on the 181st day, …]); and the storage temperature data in the sales link is extracted as a third time segment (such as [4℃ at 09:00 on the 184th day, 3℃ at 09:00 on the 185th day]). The server aligns the above multi-source time segments through the “time data input module” of the time alignment submodule. The specific operation is as follows: match the time sequence-link association space: this space pre-stores the standard time characteristic sections of each link (such as 0-180 days for breeding, 181-183 days for transportation, and 184-185 days for sales).The server matches the timestamp of each segment with the corresponding section (e.g., "Day 181 08:00" of the transportation link matches the "181-183 days" section). Adjust the timestamp offset: if the timestamp of a segment is offset (e.g., the timestamp of a piece of data in the transportation link is incorrectly labeled as "Day 180 23:00"), the server corrects it to "Day 181 00:00" according to the section range, ensuring consistency with the link time range. Complete missing data: if a segment has missing data on the timeline (e.g., there is no data at 08:05 on Day 181 in the transportation link), the server fills it in through linear interpolation based on the two pieces of data before and after it (08:04-28℃ / 62%, 08:06-28℃ / 63%) to "08:05-28℃ / 62.5%". Finally, each segment is aligned into "aligned time series segment data" that is time-continuous and has no missing data (e.g., the cultivation segment ends at 24:00 on Day 180, the transportation segment starts at 00:00 on Day 181, and the sales segment starts at 00:00 on Day 184). The server uses the time series alignment submodule to concatenate the aligned multi-source time series segments in chronological order to form a "multi-source time series alignment sequence". For example: the last piece of data in the cultivation link is "Day 180 24:00-24℃ / 62%"; the first piece of data in the transportation link is "Day 181 00:00-112.2°E / 28.2°N-28℃ / 60%"; and the first piece of data in the sales link is "Day 184 00:00-4℃". After arranging these data in chronological order, a continuous data chain covering 0-185 days is formed (e.g., [cultivation-08:00 on Day 1-25℃ / 60%… cultivation-24:00 on Day 180-24℃ / 62%, transportation-00:00 on Day 181-28℃ / 60%… transportation-24:00 on Day 183-29℃ / 65%, sales-00:00 on Day 184-4℃…]). At this point, the server has completed the multi-source time series alignment processing, converting the originally scattered and misaligned RFID sensing data into a multi-source time series alignment sequence with clear structure and time continuity, providing a standardized input for subsequent feature embedding and anomaly analysis.

[0040] In the embodiment of the application, the time series alignment submodule includes a time series data input module; the multi-source time series segment data includes a plurality of multi-source time series segment data, which includes target multi-source time series segment data;

[0041] The time series segment alignment of the multi-source time series segment data by the time series alignment submodule can be implemented through the following examples.

[0042] determine a time-series-link correlation space of the target multi-source time-series segment data based on multi-source perception data information to which the target multi-source time-series segment data belongs; the time-series-link correlation space includes a plurality of time-series-link correlation values, each time-series-link correlation value having a time-series-link feature section;

[0043] determine the time-series-link feature section to which the target multi-source time-series segment data belongs through the time-series data input module, and determine the time-series-link correlation value of the time-series-link feature section to which the target multi-source time-series segment data belongs as the aligned time-series segment data of the target multi-source time-series segment data.

[0044] In the embodiments of the present application, an example is taken from the whole-link tracking scenario of soft-shelled turtles in a certain aquatic product enterprise. The server processes the time sequence alignment process of the "target multi-source time sequence fragment data" (i.e., the temperature and humidity and location data fragments of the transportation link) as follows: The server first analyzes the "multi-source perception data information" of the target multi-source time sequence fragment data, determines that the fragment belongs to the transportation link by the data label (such as "transportation-vehicle-mounted RFID-temperature and humidity"), device ID (such as "vehicle-mounted device 007"), and timestamp (such as "day 181 08:00-day 183 24:00"), and is collected by the vehicle-mounted sensor. The standard time range should be 181-183 days. Based on this, the server calls the pre-defined "time sequence-link association space" in the time sequence alignment submodule. The space is a structured database that stores "time sequence-link association values" of each link in the whole link. Each association value corresponds to a "time sequence-link characteristic section" (i.e., the time range and format requirements that the data of the link should cover). For example: the association value of the breeding link is "breeding-environment sensor", and the corresponding characteristic section is "0-180 days, time stamp accuracy hour level, data format [time, water temperature, dissolved oxygen]"; the association value of the transportation link is "transportation-vehicle-mounted RFID", and the corresponding characteristic section is "181-183 days, time stamp accuracy minute level, data format [time, longitude, latitude, temperature, humidity]"; the association value of the sales link is "sales-terminal RFID", and the corresponding characteristic section is "184-185 days, time stamp accuracy day level, data format [time, storage temperature]". The server matches the target multi-source time sequence fragment data (temperature and humidity and location data of the transportation link) with the characteristic section in the association space through the "time sequence data input module" of the time sequence alignment submodule. The specific operation is as follows: The server checks the timestamp of the target fragment: the timestamp of a record in the original data is "day 180 23:50" (belongs to the time range 0-180 days of the breeding link), but according to the multi-source perception data information (transportation link), the timestamp should belong to the 181-183 day section of the transportation link. At this time, the server determines that the timestamp is offset (may be caused by device clock error) and corrects it to "day 181 00:00" (the start time of the transportation link). There are time intervals that are not continuous in the target fragment (such as 08:00 on day 181 has data, 08:01 has no data, and 08:02 has data). The server fills in the missing points according to the minute-level accuracy requirement of the transportation link characteristic section by linear interpolation: the temperature of 08:00 is known to be 28℃, and the humidity is 60%; the temperature of 08:02 is 28℃, and the humidity is 62%; the temperature of 08:01 is 28℃ (no change), and the humidity is 61% (take the intermediate value). Some data formats in the target fragment are not standardized (such as the location information of a record is "east longitude 112 nearby", instead of the standard longitude and latitude values).The server invokes predefined format validation rules (the transport link feature segment requires "longitude and latitude to be numeric values ​​to one decimal place") to convert ambiguous descriptions into standard values ​​(e.g., "near 112° East longitude" is corrected to "112.0° East longitude"). After completing timestamp correction, missing values ​​completion, and format standardization, the server binds the target multi-source time series segment data to the transport link's time series-link association value ("Transportation-Onboard RFID") to generate "aligned time series segment data." For example, the corrected transport link data is: [Day 181 00:00 - 112.0° East longitude / 28.0° North latitude - 28°C / 60%; Day 181 00:01 - 112.1° East longitude / 28.1° North latitude - 28°C / 61%... Day 183 24:00 - 113.5° East longitude / 29.0° North latitude - 29°C / 65%]. The timestamp of this segment strictly falls within the characteristic segment of the transportation phase (days 181-183), the data format meets the requirements (minute-level accuracy, standard longitude and latitude, and temperature and humidity values), and the timeline is continuous and complete, seamlessly connecting with segments from other stages (such as days 0-180 of the breeding phase and days 184-185 of the sales phase). Through these steps, the server converts the original multi-source time series segment data of the transportation phase target, which originally had time misalignment and mixed formats, into aligned time series segment data that strictly matches the "Transportation-Onboard RFID" association value. This provides standardized input for subsequent full-link timing orchestration (concatenating the segments of each stage in time), ensuring the temporal continuity and format consistency of the data across the entire link.

[0045] In an embodiment of the present invention, the fused deep features are obtained by performing deep feature fusion processing on the perception feature vector by a deep feature fusion device in a timing alignment submodule; the timing alignment submodule belongs to a traceability anomaly classification network for performing full-link traceability anomaly classification on the tracked item;

[0046] The following embodiments are also provided.

[0047] Obtaining a first training trajectory set; wherein the first traceability instance data included in the first training trajectory set is traceability instance data without abnormal annotation;

[0048] Obtain multiple analysis tasks for the original timing alignment submodule;

[0049] Performing trajectory feature analysis on the first traceability instance data through the original time series alignment submodule to obtain time series fusion features corresponding to each analysis task, and determining multiple deviation information of the multiple analysis tasks based on the time series fusion features; one analysis task corresponds to one deviation information;

[0050] The first target bias information is determined based on the plurality of bias information, and a configuration parameter of the original time sequence alignment sub-module is weight optimized based on the first target bias information. The original time sequence alignment sub-module after weight optimization is determined as the time sequence alignment sub-module.

[0051] In the embodiment of the present application, the server extracts a "first training trajectory set" from the enterprise database, which contains the full-link RFID sensing data (such as water temperature, dissolved oxygen in the breeding link, temperature and humidity, location in the transportation link, storage temperature in the sales link) of 1000 batches of turtles in the past half year, but without labeled abnormal labels (because the frequency of abnormal events in the actual scene is low, and the cost of manual labeling is high). Each data only contains a timestamp, a device ID and an original sensing value (such as "breeding-1st day 08:00-25°C" "transportation-181st day 08:00-28°C / 60%"). The server sets two analysis tasks for the original time sequence alignment sub-module (initial model before training) to evaluate its feature extraction capability: time sequence reconstruction task: input the sensing data of the previous N time points in a link (such as water temperature data from the 1st to the 10th day in the breeding link), and require the model to output the predicted value at the N+1 time point (water temperature on the 11th day), to verify whether the model captures the time sequence rule; feature consistency task: the data collected by multiple source devices in the same link (such as water temperature sensor and dissolved oxygen sensor in the breeding link) should reflect the same environmental state, and the model is required to output feature vectors (water temperature feature, dissolved oxygen feature) with high similarity, to verify whether the model fuses the correlation of multi-source data. The server calls the original time sequence alignment sub-module (initial parameters are randomly initialized), analyzes the trajectory features of each data in the first training trajectory set, obtains the "time sequence fusion features" corresponding to each analysis task, and calculates the deviation: taking a data in the breeding link as an example: the original data is "1st day 08:00-25°C, 1st day 09:00-26°C…10th day 20:00-24°C" (a total of 240 time points). After the original sub-module extracts its sensing feature vector, it generates the time sequence fusion feature (including the water temperature change trend). The server requires the model to predict the water temperature at 21:00 on the 10th day (the actual value is 24°C), and the model outputs the predicted value as 22°C (because the initial parameters are not optimized, the rule of "water temperature stable at night" is not captured), and the mean square error ((24-22)²=4) is calculated as the deviation information of this task. Taking the water temperature sensor (data A) and the dissolved oxygen sensor (data B) in the breeding link as an example: the original sub-module extracts the feature vector V1 (such as [0.2, 0.5, 0.3]) of data A and the feature vector V2 (such as [0.1, 0.8, 0.1]) of data B, respectively. The server calculates the cosine similarity of the two vectors (0.2*0.1+0.5*0.8+0.3*0.1=0.45), because in reality water temperature and dissolved oxygen are positively correlated (increased water temperature may decrease dissolved oxygen), the ideal similarity should be higher (such as 0.7), so the deviation is 0.7-0.45=0.25.The server aggregates the deviation information for all analysis tasks (e.g., the average deviation for the time series reconstruction task is 3.2, and the average deviation for the feature consistency task is 0.3). It selects the time series reconstruction task with the largest deviation as the "first target deviation" (3.2), indicating that the original submodule has significant deficiencies in capturing temporal dependencies. Based on this first target deviation, the server uses a backpropagation algorithm to adjust the configuration parameters of the original submodule (e.g., the weight matrix of the LSTM layer and the query / key / value vectors of the attention mechanism). For example, the weight of the "forget gate" in the LSTM layer is increased to strengthen the model's memory of historical time points; the number of attention heads is adjusted to increase focus on key time points (e.g., nighttime water temperature). After multiple rounds of iterative training (e.g., 100 rounds), the average deviation for the time series reconstruction task decreases from 3.2 to 0.5 (significantly reducing the error between predicted and actual values), and the deviation for the feature consistency task also decreases to 0.1 (improving the similarity of feature vectors from multiple source devices). At this point, the server identifies the optimized original submodule as the final "time series alignment submodule," which can more accurately extract time series features from unlabeled data, providing reliable input for subsequent full-link anomaly assessment. Through an unsupervised training process, the server uses unlabeled historical data and optimizes the parameters of the timing alignment submodule through multi-task deviation analysis, enabling it to more accurately capture the timing patterns and multi-source correlations of full-link perception data, laying the model foundation for full-link tracking and traceability of aquatic products such as soft-shelled turtles.

[0052] In an embodiment of the present invention, the plurality of analysis tasks include a link interruption detection task;

[0053] The performing trajectory feature analysis on the first traceability instance data by the original time series alignment submodule to obtain time series fusion features corresponding to each analysis task, and determining multiple deviation information of the multiple analysis tasks based on the time series fusion features includes:

[0054] Performing multi-source timing alignment processing on the first traceability instance data by the original timing alignment submodule to obtain a first sample multi-source timing alignment sequence of the first traceability instance data; the first sample multi-source timing alignment sequence includes sample alignment timing segment data on multiple timing nodes;

[0055] Dynamically selecting a first timing node for injecting link noise from a plurality of timing nodes of the multi-source timing alignment sequence of the first sample;

[0056] Performing link noise injection processing on the sample alignment time series segment data at the first time series node in the first sample multi-source time series alignment sequence based on the noise link marker to obtain a first enhanced time series trajectory set;

[0057] The original timing alignment submodule is used for timing feature embedding mapping on the first enhanced timing trajectory set, to obtain a first enhanced feature vector of the first enhanced timing trajectory set, and the first enhanced feature vector is subjected to deep feature fusion processing to obtain a first timing fusion feature of the first enhanced feature vector. The first timing fusion feature belongs to the timing fusion feature;

[0058] The first timing fusion feature is subjected to noise link identification processing by a detector module corresponding to the link interruption detection task, to obtain the predicted alignment timing segment data of the noise link label in the first enhanced timing trajectory set;

[0059] Based on the sample alignment timing segment data on the first timing node in the first sample multi-source timing alignment sequence and the predicted alignment timing segment data, deviation information of the link interruption detection task is determined.

[0060] In the embodiments of the present application, an example is taken from the soft-shelled turtle full-link tracking system of a certain aquatic product enterprise. The server optimizes the process of the original time sequence alignment sub-module through the "link interruption detection task" as follows: the server first extracts a batch of data from the "first training trajectory set" (1000 batches of unannotated soft-shelled turtle full-link data) and performs multi-source time sequence alignment processing through the original time sequence alignment sub-module. The batch of data contains perception data of the breeding (0-180 days), transportation (181-183 days), and sales (184-185 days) links, and generates a "first sample multi-source time sequence alignment sequence" after alignment, which is a time-continuous full-link data chain containing thousands of "time sequence nodes" (each node corresponds to an aligned fragment data at a specific time point). For example, the "181st day 08:00" node data of the transportation link is "east longitude 112.0° / north latitude 28.0°-28°C / 60%", and the "181st day 08:01" node is "east longitude 112.1° / north latitude 28.1°-28°C / 61%". The server dynamically selects a "first time sequence node" (i.e., a time point simulating a link interruption) from the thousands of time sequence nodes of the alignment sequence for the link interruption detection task. The selection rule is based on data distribution: preferentially selecting a period with relatively stable data changes in the transportation link (such as 08:00-08:30 on the 181st day, during which the soft-shelled turtle is in transit and the temperature and humidity fluctuate little), to highlight the abnormality after noise injection. Finally, the "181st day 08:00" node is selected as the target. The server performs "link noise injection processing" on the original data of the "181st day 08:00" node ("east longitude 112.0° / north latitude 28.0°-28°C / 60%"): noise type: simulate a transportation link interruption (such as equipment failure), replace the temperature value of the node with an abnormal value (change from 28°C to 40°C), and delete the location information (set the east longitude / north latitude fields to null); noise marker: add a "noise link marker" (such as the label "transportation-181st day 08:00-noise") to the node, and record the original data ("28°C / 60%" "east longitude 112.0° / north latitude 28.0°") as a reference. After processing, a "first enhanced time sequence trajectory set" is generated, and only the "181st day 08:00" node in the original alignment sequence is contaminated, and the remaining nodes remain unchanged. The server calls the original time sequence alignment sub-module to perform feature processing on the enhanced trajectory set: time sequence feature embedding mapping: input the enhanced alignment sequence into the time sequence encoder (such as an LSTM network) of the sub-module, and output a "first enhanced feature vector" (containing the node information contaminated by noise) of 128 dimensions; deep feature fusion: fuse the enhanced feature vector through the deep feature fusioner (such as an attention network) of the sub-module, extract "first time sequence fusion features", and the features integrate the time sequence correlation of the full-link data, especially focusing on the transportation link disturbed by noise.The server calls the "detector module" (part of the original submodule, used to identify abnormal nodes) corresponding to the link interruption detection task, and inputs the first time sequence fusion feature into the module. The detector module locuses the "181st day 08:00" node as a noise link by analyzing the abnormal signals (such as sudden temperature rise, position missing) in the feature, and tries to predict the original alignment segment data (i.e. the correct value not contaminated by noise) of the node. For example, based on the temperature of the previous and next nodes (28℃ / 61% at 08:01, 28℃ / 59% at 07:59), the detector module predicts that the temperature at "the 181st day 08:00" should be 28℃, the humidity should be 60%, and the position should be "East 112.0° / North 28.0°". The server compares the prediction result of the detector module ("28℃ / 60%" "East 112.0° / North 28.0°") with the original data not contaminated by noise ("28℃ / 60%" "East 112.0° / North 28.0°"), and calculates the deviation: temperature deviation: the predicted value 28℃ is consistent with the original value 28℃, the deviation is 0; humidity deviation: the predicted value 60% is consistent with the original value 60%, the deviation is 0; position deviation: the predicted longitude and latitude are completely matched with the original value, the deviation is 0. If the original submodule is not optimized (such as random initial parameters), the detector module may predict the temperature to be 35℃ (without capturing the regularity of stable temperature during transportation), and the temperature deviation is (28-35)²=49. After training and optimization, the deviation gradually decreases, indicating that the recognition ability of the submodule for link interruption is improved. Through the link interruption detection task, the server simulates the abnormal interruption (such as equipment failure) of a certain link in the whole link, and optimizes the parameters of the original time sequence alignment submodule through the closed loop of noise injection-prediction-deviation calculation, so that it can more accurately identify and repair data interruption and improve the reliability of the whole link sensing data.

[0061] In the embodiments of the present application, the plurality of analysis tasks includes a link abnormality detection task;

[0062] The trajectory feature analysis of the first traceability instance data by the original time sequence alignment submodule obtains time sequence fusion features corresponding to each analysis task, and the plurality of deviation information of the plurality of analysis tasks is determined based on the time sequence fusion features, which can be implemented through the following examples.

[0063] The first sample multi-source time sequence alignment sequence of the first traceability instance data is obtained by performing multi-source time sequence alignment processing on the first traceability instance data by the original time sequence alignment submodule; the first sample multi-source time sequence alignment sequence contains sample alignment time sequence segment data on a plurality of time sequence nodes;

[0064] A second time sequence node to be injected with link disturbance is dynamically selected from the plurality of time sequence nodes of the first sample multi-source time sequence alignment sequence;

[0065] performing link disturbance injection processing on the sample alignment time segment data on the second time sequence node in the first sample multi-source time sequence alignment sequence based on a random link disturbance factor, to obtain a second enhanced time sequence track set; the random link disturbance factor is different from the sample alignment time segment data on the second time sequence node;

[0066] performing time sequence feature embedding mapping on the second enhanced time sequence track set through the original time sequence alignment sub-module to obtain a second enhanced feature vector of the second enhanced time sequence track set, and performing deep feature fusion processing on the second enhanced feature vector to obtain a second time sequence fusion feature of the second enhanced feature vector; the second time sequence fusion feature belongs to the time sequence fusion feature;

[0067] performing disturbance positioning analysis processing on the second time sequence fusion feature through a detector module corresponding to the link anomaly detection task, to obtain disturbance link prediction results of the plurality of time sequence nodes;

[0068] determining deviation information of the link anomaly detection task based on the disturbance link prediction results of the second time sequence node and the plurality of time sequence nodes.

[0069] In the embodiments of the present application, an example is taken from the whole link tracking system of a certain aquatic product enterprise. The server optimizes the process of the original time sequence alignment sub-module through the "link abnormality detection task" as follows: the server selects a batch of data from the "first training track set" (1000 batches of unannotated whole link data of soft-shelled turtles) and performs multi-source time sequence alignment processing through the original time sequence alignment sub-module. This batch of data covers the breeding (0-180 days), transportation (181-183 days), and sales (184-185 days) links, and generates a "first sample multi-source time sequence alignment sequence" after alignment, which is a time-continuous whole link data chain containing thousands of "time sequence nodes" (each node corresponds to the aligned fragment data of a specific time point). For example, the "181st day 09:00" node data of the transportation link is "East longitude 112.2° / North latitude 28.2°-28℃ / 65%" (stable temperature and humidity, which meets the normal range during transportation). The server dynamically selects a "second time sequence node" from the thousands of time sequence nodes in the alignment sequence for the link abnormality detection task, selects a stable period of data in the transportation link (such as the 181st day 09:00, during which the soft-shelled turtles do not experience sudden braking or temperature and humidity changes), and highlights the abnormality after disturbance. Finally, the "181st day 09:00" node is selected as the target. The server performs "link disturbance injection processing" on the original data of the "181st day 09:00" node ("East longitude 112.2° / North latitude 28.2°-28℃ / 65%"): disturbance factor: randomly generates a disturbance value different from the original data (such as increasing the temperature by 5℃ to 33℃ and reducing the humidity by 10% to 55%); disturbance marker: record the original data of the node ("28℃ / 65%" "East longitude 112.2° / North latitude 28.2°") as a reference, generate a "second enhanced time sequence track set", only the "181st day 09:00" node is disturbed, and the remaining nodes remain unchanged. The server calls the original time sequence alignment sub-module to perform feature processing on the second enhanced track set: time sequence feature embedding mapping: input the enhanced alignment sequence into the time sequence encoder (such as the LSTM network) of the sub-module, and output a 128-dimensional "second enhanced feature vector" (containing the disturbed node information); deep feature fusion: the enhanced feature vector is fused through the deep feature fusioner (such as the attention network) of the sub-module to extract "second time sequence fusion features", which comprehensively reflect the time sequence correlation of the whole link data and highlight the abnormal signals of the disturbed node (such as sudden temperature rise and sudden humidity drop). The server calls the "detector module" corresponding to the link abnormality detection task (part of the original sub-module, used to locate abnormal disturbances) and inputs the second time sequence fusion features into the module. The detector module analyzes the abnormal patterns in the features (such as a temperature deviation of 5℃ from the historical mean and a humidity deviation of 2 times the standard deviation), traverses all time sequence nodes, and outputs a "disturbance link prediction result", identifying that the "181st day 09:00" node is the disturbance link (probability 0.95), and the remaining nodes have no disturbance (probability less than 0.1).The server compares the prediction result of the detector module (the "181st day 09:00" is the disturbance node) with the second time sequence node (the "181st day 09:00") actually injected with the disturbance, calculates the deviation: positioning accuracy: the predicted node is completely consistent with the actual node, and the deviation is 0; if the original sub-module is not optimized (for example, the initial parameters are random), the detector module may misjudge as the "181st day 09:30" node (because the local feature of the temperature sudden change is not captured), and the positioning deviation is 30 minutes (the square of the time difference). After training and optimization, the deviation gradually decreases, indicating that the positioning ability of the sub-module to the link abnormality is improved. Through the link abnormality detection task, the server simulates the abnormal disturbance of a link in the whole link (such as temporary high temperature in the transportation process), and optimizes the parameters of the original time sequence alignment sub-module through the closed loop of disturbance injection-positioning-deviation calculation, so that the original time sequence alignment sub-module can more accurately identify and locate the data abnormality, and the abnormal detection ability of the whole link sensing data is improved.

[0070] In the embodiment of the application, the plurality of analysis tasks includes a link interference detection task.

[0071] The trajectory feature analysis of the first traceability instance data by the original time sequence alignment sub-module obtains time sequence fusion features corresponding to each analysis task, and the plurality of deviation information of the plurality of analysis tasks is determined based on the time sequence fusion features. The following example can be used for implementation.

[0072] The link interference feature is dynamically injected into the multi-source time sequence segment data of the first traceability instance data to obtain interference traceability instance data, and the multi-source time sequence alignment processing is performed on the interference traceability instance data to obtain a noise multi-source time sequence alignment sequence of the interference traceability instance data.

[0073] The multi-source time sequence alignment processing is performed on the first traceability instance data to obtain a first sample multi-source time sequence alignment sequence of the first traceability instance data.

[0074] The time sequence feature embedding mapping of the noise multi-source time sequence alignment sequence by the original time sequence alignment sub-module obtains a third enhanced feature vector of the noise multi-source time sequence alignment sequence, and the deep feature fusion processing is performed on the third enhanced feature vector to obtain a third time sequence fusion feature of the third enhanced feature vector. The third time sequence fusion feature belongs to the time sequence fusion feature.

[0075] The time sequence feature embedding mapping of the first sample multi-source time sequence alignment sequence by the original time sequence alignment sub-module obtains a fourth enhanced feature vector of the first sample multi-source time sequence alignment sequence, and the deep feature fusion processing is performed on the fourth enhanced feature vector to obtain a fourth time sequence fusion feature of the fourth enhanced feature vector. The fourth time sequence fusion feature belongs to the time sequence fusion feature.

[0076] determine bias information of the link interference detection task based on the third fusion feature and the fourth fusion feature.

[0077] In the embodiment of the present application, an example is taken from the soft-shelled turtle full-link tracking system of a certain aquatic product enterprise. The server optimizes the process of the original time sequence alignment sub-module through the "link interference detection task" as follows: the server selects a batch of original data (containing perception information of the breeding, transportation, and sales links) from the "first training track set" (1000 batches of unannotated soft-shelled turtle full-link data), and dynamically injects "link interference features" for the transportation link. The specific operation is as follows: interference link selection: select a stable data period in the transportation link (such as 10:00-10:30 on day 181, during which the soft-shelled turtles are in uniform transportation and the temperature and humidity fluctuate little); interference feature injection: for the "10:00 on day 181" node data in this period (the original value is "east longitude 112.3° / north latitude 28.3°-28℃ / 65%"), randomly generate interference features, modify the temperature value to an abnormal 35℃ (5-8℃ higher than the normal range), and add an erroneous position offset (east longitude 112.3° to 115.0°, simulating GPS signal interference), to generate "interference source instance data". The server respectively performs multi-source time sequence alignment processing on the interference source instance data and the original data: noise multi-source time sequence alignment sequence: after the interference data is processed by the time sequence alignment sub-module, a "noise multi-source time sequence alignment sequence" is generated, in which the "10:00 on day 181" node data is "east longitude 115.0° / north latitude 28.3°-35℃ / 65%" (the disturbed data), and the remaining nodes remain normal (such as "east longitude 112.3° / north latitude 28.3°-28℃ / 64%" on day 181 09:59 and "east longitude 112.3° / north latitude 28.3°-28℃ / 66%" on day 181 10:01); first sample multi-source time sequence alignment sequence: the original data generates a "first sample multi-source time sequence alignment sequence" after alignment, and the "10:00 on day 181" node data is the original correct value ("east longitude 112.3° / north latitude 28.3°-28℃ / 65%"), and the remaining nodes are consistent with the interference data. The server calls the original time sequence alignment sub-module to process the noise multi-source time sequence alignment sequence and the first sample multi-source time sequence alignment sequence respectively: third enhanced feature vector and third time sequence fusion feature: the time sequence encoder (such as the LSTM network) of the sub-module inputs the noise sequence, and outputs a 128-dimensional "third enhanced feature vector" (containing abnormal information of the interference node); then through the deep feature fusioner (such as the attention network), a "third time sequence fusion feature" is obtained, which contains abnormal signals such as "temperature sudden rise" and "position offset" due to the interference; fourth enhanced feature vector and fourth time sequence fusion feature: the sub-module inputs the original sequence, and outputs a 128-dimensional "fourth enhanced feature vector" (normal features without interference); after deep fusion, a "fourth time sequence fusion feature" is obtained, which reflects the normal time sequence law of the transportation link (such as stable temperature and continuous change of position).The server calculates the deviation information of the link interference detection task by comparing the third time sequence fusion feature and the fourth time sequence fusion feature: feature similarity calculation: the cosine similarity is used to measure the difference between the two fusion features (the ideal similarity of normal features and interference features should be close to 0, because the interference will significantly change the feature distribution); deviation calculation: if the original sub-module is not optimized (the initial parameters are random), the cosine similarity of the third and fourth features may be as high as 0.8 (the model does not recognize the interference, and the feature does not change significantly), and the deviation is 1-0.8=0.2; if the sub-module is optimized (the parameters are adjusted), the abnormal signal of the interference feature is amplified, and the cosine similarity is reduced to 0.2 (the model is sensitive to the interference), and the deviation is 1-0.2=0.8 (the greater the deviation, the stronger the model's ability to recognize interference). Through the link interference detection task, the server simulates data anomalies caused by device failure or signal interference in the whole link (such as temperature false alarm and position deviation in the transportation link), and through the closed loop of interference injection-feature comparison-deviation calculation, the parameters of the original time sequence alignment sub-module are optimized, so that it can more sensitively recognize link interference, and improve the robustness and reliability of the whole link perception data.

[0078] In the embodiments of the present application, the plurality of analysis tasks includes a link quality self-supervised task.

[0079] The plurality of analysis tasks includes a link quality self-supervised task.

[0080] The plurality of analysis tasks includes a link quality self-supervised task.

[0081] The plurality of analysis tasks includes a link quality self-supervised task.

[0082] The plurality of analysis tasks includes a link quality self-supervised task.

[0083] The plurality of analysis tasks includes a link quality self-supervised task.

[0084] determine bias information of the link quality self-supervised task based on the dynamic anomaly confidence and the baseline anomaly confidence.

[0085] In the embodiments of the present application, an example is taken as a whole link tracking system of a certain aquatic product enterprise. The server optimizes the process of the original time sequence alignment sub-module through the "link quality self-supervision task" as follows: the server selects a batch of data from the "first training track set" (1000 batches of unannotated whole link data of soft-shelled turtles) as the "first traceability instance data". The data contains water temperature in the breeding link (0-180 days, collected every hour), temperature and humidity and location in the transportation link (181-183 days, collected every minute), and storage temperature in the sales link (184-185 days, collected every day). The server calls the original time sequence alignment sub-module and processes according to the following steps: identifying the multi-source perception information of each data (such as "environmental sensors" for the breeding link and "vehicle-mounted equipment" for the transportation link); dividing the original data into multi-source time sequence segments (such as breeding water temperature segments and transportation temperature and humidity segments); aligning the time stamps through the time sequence-link association space (such as the transportation link data starting from the 181st day 00:00), and filling in the missing values (such as interpolating to fill in the missing data of the transportation link at a certain minute), and finally generating "the first sample multi-source time sequence alignment sequence", a time-continuous, format-unified whole link data chain (such as [breeding-1st day 08:00-25℃…transportation-181st day 00:00-28℃ / 60%…sales-184th day 00:00-4℃]). The server calls the original time sequence alignment sub-module to perform feature processing on the first sample multi-source time sequence alignment sequence: time sequence feature embedding mapping: inputting the alignment sequence into the time sequence encoder (such as the LSTM network) of the sub-module. The encoder extracts time sequence features layer by layer: the first layer learns the local pattern of adjacent time points (such as "the breeding water temperature is stable ±1℃ every hour"), the second layer captures the trend across the time period (such as "the water temperature rises in the first 30 days of breeding and drops sharply on the 31st day"), and finally outputs a "fourth enhanced feature vector" of 128 dimensions (each dimension represents a key feature such as "water temperature fluctuation standard deviation" and "transportation temperature and humidity over-standard duration"). Deep feature fusion: the fourth enhanced feature vector is fused through the deep feature fusioner (such as the attention network) of the sub-module to generate "the fourth time sequence fusion feature", which integrates the time sequence correlation of the whole link data (such as "breeding water temperature drop—transportation temperature and humidity sensitive fluctuation—sales storage temperature anomaly" chain influence). The server calls the "detector module" corresponding to the link quality self-supervision task (part of the original sub-module, used to judge abnormal patterns) and inputs the fourth time sequence fusion feature into the module. The detector module analyzes the abnormal signals in the feature (such as water temperature drop amplitude exceeding 3 times the historical mean value and transportation temperature and humidity fluctuation frequency anomaly), and outputs a "dynamic abnormal confidence" (a probability value between 0 and 1, such as 0.1 indicating a 10% probability of abnormality). The confidence reflects the evaluation result of the original sub-module on the abnormal risk of the current whole link data. The server obtains a "benchmark traceability network", a simpler and fixed parameter comparison model (such as containing only the basic LSTM layer without complex optimization).The first traceability instance data of the same batch is input into a benchmark network, which processes the data through a similar process (multi-source time alignment, feature embedding, and simple fusion) and outputs a "benchmark anomaly confidence" (for example, 0.05 represents a 5% probability of abnormality). The benchmark confidence represents the basic judgment of the unoptimized model on the risk of data abnormality. The server compares the dynamic anomaly confidence (0.1) with the benchmark anomaly confidence (0.05), and calculates the absolute difference (0.1-0.05=0.05) between the two as the "deviation information of the link quality self-supervision task". If the original sub-module is not optimized (the initial parameters are random), the dynamic confidence may be highly consistent with the benchmark confidence (such as both being 0.05), and the deviation is 0 (indicating that the model has not captured the potential abnormality in the data); after training and optimization, the dynamic confidence may increase to 0.2 (the model more sensitively identifies the abnormal signal of "rapid drop in aquaculture water temperature"), and the deviation increases to 0.15 (indicating that the model's ability to judge link quality has improved). Through the link quality self-supervision task, the server uses unannotated data to evaluate the model performance through the "difference between the original sub-module judgment result and the benchmark model result", optimizes the parameters of the original time alignment sub-module, and enables it to more accurately capture potential abnormal patterns in the full-link data, providing more reliable abnormality evaluation capability for full-link tracking of soft-shelled turtles and other aquatic products.

[0086] In the embodiment of the application, the fusion deep feature is obtained by deep feature fusion processing of the perception feature vector by a deep feature fusioner in the time alignment sub-module; the time alignment sub-module belongs to a traceability anomaly classification network for full-link traceability anomaly classification of the tracked object; the traceability anomaly classification network further comprises a link quality analysis network arranged after the deep feature fusioner;

[0087] The link quality analysis network in the traceability anomaly classification network comprises an anomaly discriminator of each traceability link;

[0088] The link quality analysis network is determined from the traceability anomaly classification network; the link quality analysis network comprises an anomaly discriminator of each traceability link;

[0089] The anomaly discriminators of each traceability link are used to perform link anomaly evaluation on the fusion deep feature to obtain link anomaly confidence of each traceability link; one anomaly discriminator is used to determine the link anomaly confidence of one traceability link.

[0090] In an embodiment of the present invention, illustratively, taking the full-link tracking scenario of soft-shelled turtles in a certain aquatic enterprise as an example, the process of the server evaluating the confidence level of abnormalities in each link through the link quality analysis network in the traceability abnormality classification network is as follows: After completing the feature embedding and deep fusion of the soft-shelled turtle RFID full-link perception data (generating fused deep features), the server calls the pre-trained "traceability abnormality classification network". This network is a deep learning model constructed by the server for the full-link abnormality classification task of aquatic products. It contains three parts: a time series alignment submodule (responsible for multi-source data alignment and feature embedding), a deep feature fusion module (generating fused deep features), and a link quality analysis network (responsible for link abnormality assessment). The server locates the "link quality analysis network" from the network through the model structure index. The network is a module composed of multiple sub-models (anomaly discriminators), and each sub-model corresponds to a full-link link (such as breeding, transportation, and sales). The link quality analysis network internally integrates "anomaly discriminators" that correspond one-to-one to the full-link links. For example, the entire soft-shelled turtle production chain is divided into three stages: breeding, transportation, and sales. Therefore, the link quality analysis network includes three anomaly detectors: the breeding anomaly detector, which consists of one fully connected layer with a sigmoid activation function. Its input dimension is the length of the fused deep features (e.g., 128 dimensions), and its output is the "anomaly confidence level for the breeding stage" (a probability value between 0 and 1). The transportation anomaly detector, which consists of two fully connected layers with a ReLU activation function, also has a 128-dimensional input dimension and outputs the "anomaly confidence level for the transportation stage." The sales anomaly detector, which consists of one convolutional layer (for extracting local features) and a fully connected layer, has a 128-dimensional input dimension and outputs the "anomaly confidence level for the sales stage." The server inputs the fused deep features (a 128-dimensional vector that comprehensively reflects the status of the entire chain) into the link quality analysis network. The anomaly detectors in each stage independently process these features and output the anomaly confidence level for the corresponding stage. The fused deep features include sub-features such as the "standard deviation of aquaculture water temperature fluctuation" and the "percentage of time the dissolved oxygen level is below the threshold" (extracted from the perception feature vector by the deep fusion processor). The aquaculture anomaly discriminator calculates the weighted sum of these sub-features (weights determined by training) through a fully connected layer. For example, the standard deviation of water temperature fluctuation (0.3) × weight (2.0) + duration of low dissolved oxygen (0.8) × weight (1.5) = 0.3 × 2 + 0.8 × 1.5 = 0.6 + 1.2 = 1.8. After Sigmoid activation, the output aquaculture anomaly confidence score is σ(1.8) = 0.85 (85% probability of anomaly). The fused deep features include sub-features such as "duration of transport temperature and humidity exceeding standards," "number of emergency stops," and "position deviation rate."The transport anomaly discriminator extracts features step by step through two fully connected layers: the first layer calculates "temperature and humidity over-standard duration (0.6) x weight (1.2) + emergency braking times (0.4) x weight (1.0) = 0.6x1.2 + 0.4x1 = 0.72 + 0.4 = 1.12"; the second layer combines "position offset rate (0.2) x weight (0.8)" to obtain a comprehensive value of 1.12 + 0.2x0.8 = 1.28; after ReLU activation (retaining positive values), the output transport anomaly confidence is 0.3 (30% probability of anomaly). The sales anomaly discriminator extracts local features (such as "storage temperature continuously higher than 5℃ for 3 days") through a convolutional layer, and then calculates through a fully connected layer: temperature mean (0.5, representing 5℃) x weight (1.5) + over-standard duration (0.1) x weight (2.0) = 0.5x1.5 + 0.1x2 = 0.75 + 0.2 = 0.95; the output sales anomaly confidence is 0.1 (10% probability of anomaly). Finally, the server analyzes the link quality of each anomaly discriminator of the network to obtain the anomaly confidence of the breeding (0.85), transportation (0.3), and sales (0.1) links. The results show that the breeding link has a high abnormal risk (possibly due to large water temperature fluctuations and insufficient dissolved oxygen, resulting in weak turtle body), and the temperature control and oxygenation equipment of the breeding pond needs to be prioritized for troubleshooting; the transportation and sales links have low risk, and the monitoring frequency can be appropriately reduced. Through this process, the server realizes the accurate quantification of the abnormal risk of each link in the whole chain, providing a clear link positioning basis for quality control of aquaculture enterprises.

[0091] In the embodiments of the present application, the following implementation modes are further provided.

[0092] A second training trajectory set is obtained; the second training trajectory set includes second traceability instance data configured with a link identifier; the link identifier includes a plurality of link code values of the plurality of traceability links, one traceability link corresponding to one link code value, and the plurality of link code values being determined based on the traceability link to which the second traceability instance data belongs;

[0093] A pre-trained time sequence alignment submodule is obtained, the second traceability instance data is subjected to feature embedding processing through the time sequence alignment submodule to obtain a fifth enhanced feature vector of the second traceability instance data, and the fifth enhanced feature vector is subjected to deep feature fusion processing to obtain a fifth time sequence fusion feature of the fifth enhanced feature vector;

[0094] obtain a plurality of initial anomaly discriminators of the plurality of traceability links, respectively identify the fifth time sequence fusion feature based on the plurality of initial anomaly discriminators, and obtain a plurality of link stage identification confidence degrees of the plurality of initial anomaly discriminators; one traceability link corresponds to one initial anomaly discriminator, and one initial anomaly discriminator is used to determine one link stage identification confidence degree;

[0095] determine second target deviation information based on the plurality of link stage identification confidence degrees and the plurality of link encoding values, perform weight optimization on configuration parameters of the plurality of initial anomaly discriminators based on the second target deviation information, determine the plurality of initial anomaly discriminators after weight optimization as a plurality of anomaly discriminators, and determine the link quality analysis network based on the plurality of anomaly discriminators.

[0096] In the embodiment of the present application, an example is taken from the whole link tracking system of a certain aquatic product enterprise. The server trains the process of each link anomaly discriminator in the link quality analysis network as follows: the server extracts the "second training trajectory set" from the enterprise database. This data set contains the whole link RFID sensing data of 500 batches of soft-shelled turtles in the past year (such as water temperature in the breeding link, temperature and humidity in the transportation link, and storage temperature in the sales link), and each data has been labeled with "link identifier" (marked by artificial or system automatic marking). For example: the sensing data of the breeding link (such as "08:00-25℃ on the first day") is labeled with a link code value of "0"; the sensing data of the transportation link (such as "08:00-28℃ / 60% on the 181st day") is labeled with a link code value of "1"; the sensing data of the sales link (such as "09:00-4℃ on the 184th day") is labeled with a link code value of "2". The server calls the "time alignment sub-module" (such as the model optimized by the first training trajectory set) that has been optimized through unsupervised training, and processes each "second traceability instance data" in the second training trajectory set: feature embedding processing: the sub-module aligns the scattered data of the breeding, transportation, and sales links into a time-continuous sequence, and then maps it into a "fifth enhanced feature vector" (128-dimensional, containing the time sequence features of each link) through a time sequence encoder (such as an LSTM network); deep feature fusion: the fifth enhanced feature vector is fused through the deep feature fusioner (such as an attention network) of the sub-module to generate a "fifth time sequence fusion feature" (128-dimensional, comprehensively reflecting the correlation features of each link in the whole link). The server obtains "multiple initial anomaly discriminators", which are randomly initialized sub-models (such as fully connected networks) corresponding to the breeding, transportation, and sales links. For example: - breeding initial discriminator (structure: 128-dimensional input-64-dimensional hidden layer-1-dimensional output); transportation initial discriminator (structure: 128-dimensional input-32-dimensional hidden layer-1-dimensional output); sales initial discriminator (structure: 128-dimensional input-16-dimensional hidden layer-1-dimensional output). The server inputs the fifth time sequence fusion feature into each initial discriminator for "link stage recognition": the breeding initial discriminator analyzes the "water temperature fluctuation" and "dissolved oxygen" sub-features in the feature, and outputs the "breeding link stage recognition confidence" (such as 0.3, indicating a 30% probability of belonging to the breeding link); the transportation initial discriminator analyzes the "temperature and humidity exceeding time" and "position deviation" sub-features, and outputs the "transportation link stage recognition confidence" (such as 0.5, indicating a 50% probability of belonging to the transportation link); the sales initial discriminator analyzes the "storage temperature mean" and "shelf time" sub-features, and outputs the "sales link stage recognition confidence" (such as 0.2, indicating a 20% probability of belonging to the sales link). The server calculates the "second target deviation information". For example, a certain second traceability instance data belongs to the transportation link (link code value "1"), and the initial discriminator outputs the confidence as breeding 0.3, transportation 0.5, and sales 0.2.The server calculates the deviation using a cross-entropy loss function: the true label is [0, 1, 0] (the transportation link is the positive class); the predicted confidence is [0.3, 0.5, 0.2]; the cross-entropy loss = -(0*log0.3 + 1*log0.5 + 0*log0.2) = -log0.5 ≈ 0.693. Based on the deviation, the server adjusts the parameters of the initial discriminator (such as the weight matrix and bias term of the fully connected layer) through the backpropagation algorithm. For example, increase the weight of the "temperature and humidity exceeding duration" feature in the transportation discriminator (from 0.2 to 0.5) to enhance its recognition ability of the transportation link. After multiple rounds of training (such as 200 rounds), the deviation of each initial discriminator gradually decreases (such as the loss of the transportation discriminator decreases from 0.693 to 0.1), and the model parameters converge. The server determines the optimized discriminator as the "abnormal discriminator" and integrates it into the "link quality analysis network". For example: the breeding abnormal discriminator can accurately identify "water temperature fluctuation > 2℃ / hour" "dissolved oxygen < 5mg / L" and other breeding link features, with an output confidence error < 5%; the transportation abnormal discriminator can capture "temperature and humidity exceeding > 10 minutes" "emergency braking times > 5 times / hour" and other transportation link abnormalities, with a confidence error < 3%; the sales abnormal discriminator can detect "storage temperature > 5℃" "not sold after shelving > 3 days" and other sales link problems, with a confidence error < 2%. Through the supervised learning of the second training track set, the server optimizes the parameters of each initial abnormal discriminator, enabling it to accurately identify the link stage to which the data belongs, and ultimately builds a reliable link quality analysis network. The network can be used for subsequent soft-shelled turtle full-link abnormality evaluation, outputting accurate abnormal confidence for each link to support enterprise quality control decisions.

[0097] In the embodiment of the application, the perception feature vector is obtained by time sequence feature embedding mapping of a multi-source time sequence alignment sequence by a time sequence encoder in a time sequence alignment sub-module; the multi-source time sequence alignment sequence is obtained by multi-source time sequence alignment processing of the RFID full-link perception data by a time sequence data input module in the time sequence alignment sub-module; the time sequence alignment sub-module belongs to a trace abnormality classification network for full-link traceability of the tracked article; the trace abnormality classification network further comprises a link influence degree analysis network arranged after the time sequence encoder;

[0098] The link influence degree analysis processing of the perception feature vector to obtain the link influence coefficients of the traceability links can be implemented by the following examples.

[0099] The link influence degree analysis network is determined from the trace abnormality classification network;

[0100] The perception feature vector is subjected to link influence degree analysis processing by the link influence degree analysis network to obtain the link influence coefficients of the traceability links.

[0101] In the embodiment of the present application, an example is taken from the whole link tracking scene of soft-shelled turtles in a certain aquatic product enterprise. The server calculates the influence coefficient of each link through the link influence degree analysis network in the traceability anomaly classification network as follows: In the whole link of soft-shelled turtles from cultivation to sales, RFID sensing data is collected by multiple devices: underwater temperature and humidity sensors in the cultivation link record water temperature every hour (such as "08:00-25℃ on day 1" and "09:00-26℃ on day 1"), vehicle-mounted RFID devices in the transportation link collect temperature and humidity and sudden braking times every minute (such as "08:00-28℃ / 60% / 0 times of sudden braking on day 181" and "08:01-28℃ / 61% / 1 time of sudden braking on day 181"), and supermarket RFID terminals in the sales link record storage temperature every day (such as "09:00-4℃ on day 184" and "09:00-3℃ on day 185"). The server calls the "time series data input module" in the traceability anomaly classification network (which has been optimized through unsupervised training) to process these data: multi-source time series segment extraction: according to data labels (such as "cultivation-water temperature", "transportation-temperature and humidity", and "sales-storage temperature"), the original data is divided into three independent time series segments of cultivation, transportation, and sales; time series segment alignment: through a pre-defined "time series-link correlation space" (cultivation 0-180 days, transportation 181-183 days, and sales 184-185 days), the time stamp is corrected (such as a data in the transportation link mislabeled as "23:50 on day 180", which is corrected to "00:00 on day 181"), and the missing values are completed (such as no data at 08:05 on day 181 in the transportation link, which is interpolated as "28℃ / 61% / 0 times of sudden braking" through the previous and next values); whole link time series arrangement: the aligned segments are concatenated by time to generate a "multi-source time series alignment sequence" (such as [cultivation-08:00-25℃ on day 1… transportation-00:00-28℃ / 60% on day 181… sales-00:00-4℃ on day 184]). The server calls the "time series encoder" (such as an LSTM network, which has been optimized through unsupervised training) in the time series alignment submodule to process the multi-source time series alignment sequence: input processing: the alignment sequence (length N, such as 4320 data for 180 days of cultivation, 4320 data for 3 days of transportation, and 2 data for 2 days of sales, total length 8642) is input into the input layer of the LSTM; feature extraction: the hidden layer of the LSTM learns time series dependence layer by layer: the first layer captures local patterns at adjacent time points (such as "cultivation water temperature ±1℃ per hour"), the second layer captures cross-link trends (such as "sudden drop in water temperature in the later cultivation period - fluctuation in temperature and humidity in the early transportation period"), and finally outputs a 128-dimensional "sensing feature vector". Each dimension of the vector corresponds to a key feature (such as "cultivation water temperature fluctuation standard deviation", "transportation sudden braking times mean", and "sales storage temperature mean").The server locates from the trace abnormality classification network to a "link influence degree analysis network" (located after the time sequence encoder, which has been optimized through supervised training), and inputs the perception feature vector into the network. The network is composed of an attention mechanism layer and a fully connected layer, and the core function is to calculate the contribution degree of each link to the whole link abnormality: the attention layer traverses the 128 dimensions of the perception feature vector, and calculates the importance weight of each feature to the whole link abnormality. For example, the weight of "cultivation water temperature fluctuation standard deviation" (dimension 10) is 0.3 (the greater the fluctuation, the higher the risk of subsequent link abnormality); the weight of "transportation sudden braking frequency average" (dimension 25) is 0.2 (frequent sudden braking may cause stress of soft-shelled turtles); and the weight of "sales storage temperature average" (dimension 50) is 0.1 (the higher the temperature, the greater the risk of corruption). The server aggregates the feature weights by link through the fully connected layer: the cultivation link influence coefficient: the weights of all cultivation-related features (such as 0.3 of dimension 10 and 0.2 of dimension 15 "dissolved oxygen below threshold value duration") are accumulated, and the sum is 0.5; the transportation link influence coefficient: the weights of transportation-related features (such as 0.2 of dimension 25 and 0.2 of dimension 30 "temperature and humidity exceed standard duration") are accumulated, and the sum is 0.4; and the sales link influence coefficient: the weights of sales-related features (such as 0.1 of dimension 50 and 0.05 of dimension 55 "duration of not being sold after being placed on the shelf") are accumulated, and the sum is 0.15. Finally, the server outputs the influence coefficients of each link: cultivation 0.5, transportation 0.4, and sales 0.15. The results show that the abnormality of the cultivation link has the greatest influence on the whole link abnormality (accounting for 50%), followed by the transportation link (40%), and the sales link has the smallest influence (15%). The enterprise can accordingly prioritize strengthening the monitoring of the cultivation link (such as optimizing the water temperature regulation equipment), and appropriately pay attention to the sudden braking and temperature and humidity control of the transportation link, so as to reduce the whole link abnormality risk. Through this process, the server realizes the quantitative evaluation of the influence degree of each link, and provides data support for the accurate management and control of the aquatic whole link.

[0102] In the embodiment of the present application, the trace abnormality classification network further includes a link quality analysis network for determining the link abnormality confidence of each trace link.

[0103] The embodiment of the present application also provides the following implementation.

[0104] A third training trajectory set is obtained; the third trace instance data included in the third training trajectory set is configured with a whole link abnormality identifier;

[0105] A pre-trained time sequence alignment submodule is obtained, the third trace instance data is subjected to feature embedding processing through the time sequence alignment submodule, a sixth enhanced feature vector of the third trace instance data is obtained, and the sixth enhanced feature vector is subjected to deep feature fusion processing, and a sixth time sequence fusion feature of the sixth enhanced feature vector is obtained.

[0106] obtain a training link abnormal confidence of each traceable link by acquiring a pre-trained link quality analysis network, performing link abnormality evaluation on the sixth time sequence fusion feature through a plurality of anomaly discriminators in the link quality analysis network, wherein one anomaly discriminator is used to determine the training link abnormal confidence of one traceable link;

[0107] obtain a training link influence coefficient of each traceable link by performing link influence degree analysis processing on the sixth enhanced feature vector through an original link influence degree analysis network;

[0108] determine a training full-link abnormal probability of a sample tracking object of the third traceable instance data association based on the training link abnormal confidence of each traceable link and the training link influence coefficient of each traceable link;

[0109] determine third target deviation information based on the full-link abnormality identifier and the training full-link abnormal probability, perform weight optimization on configuration parameters of the original link influence degree analysis network based on the third target deviation information, determine the original link influence degree analysis network after weight optimization as a link influence degree analysis network, and determine the traceable abnormality classification network based on the time sequence alignment submodule, the link quality analysis network and the link influence degree analysis network.

[0110] In the embodiment of the present application, an example is taken as a whole link tracking system of soft-shelled turtles of a certain aquatic product enterprise. The server influences the process of the network through the third training track set optimization link as follows: the server extracts the "third training track set" from the enterprise quality database. The data set contains the whole link RFID sensing data (such as water temperature in cultivation, transportation temperature and humidity, and storage temperature in sales) of 500 batches of soft-shelled turtles in the past two years, and each data is labeled with "whole link abnormality identifier" (determined by the quality inspection result: 0 represents no abnormality, and 1 represents that quality problems such as corruption and death finally occur). For example: batch A: water temperature fluctuation is large (abnormal) in the cultivation link, temperature and humidity are stable (normal) in the transportation link, and storage temperature exceeds the standard (abnormal) in the sales link, and the final quality inspection is corruption (whole link abnormality identifier = 1); batch B: cultivation, transportation and sales links are all normal (whole link abnormality identifier = 0). The server calls the "time alignment sub-module" (such as the model optimized in steps 5-9) that has been optimized through unsupervised training, and processes each "third traceability instance data" in the third training track set: feature embedding processing: the sub-module aligns the data in multiple sources (such as aligning the scattered data in cultivation, transportation and sales links into time-continuous sequences), and then maps the data into "sixth enhanced feature vector" (128 dimensions, containing time sequence features of each link) through a time sequence encoder (such as an LSTM network); deep feature fusion: the sixth enhanced feature vector is fused through the deep feature fusioner (such as an attention network) of the sub-module to generate "sixth time sequence fusion feature" (128 dimensions, comprehensively reflecting the correlation features of each link in the whole link). The server calls the "link quality analysis network" (such as the model trained in step 11) that has been optimized through supervised training, and inputs the sixth time sequence fusion feature into each abnormality discriminator (cultivation, transportation and sales discriminators) in the network. Each discriminator outputs "training link abnormality confidence" (probability value of 0-1): in the sixth time sequence fusion feature of batch A, the cultivation discriminator detects "water temperature fluctuation standard deviation > 3℃" (abnormal feature) and outputs the cultivation link confidence of 0.8; the transportation discriminator detects "temperature and humidity exceed the standard for < 10 minutes" (normal feature) and outputs the transportation link confidence of 0.2; and the sales discriminator detects "storage temperature > 5℃ for 3 consecutive days" (abnormal feature) and outputs the sales link confidence of 0.6. The server calls the "original link influence degree analysis network" (the initial model that is not optimized, the parameters are randomly initialized), and inputs the sixth enhanced feature vector into the network. The network calculates "training link influence coefficient" (the contribution weight of each link to the whole link abnormality) through the fully connected layer and the attention mechanism: in the sixth enhanced feature vector of batch A, the feature weight of "cultivation water temperature fluctuation" is 0.3, the feature weight of "sales storage temperature" is 0.2, and the feature weight of "transportation temperature and humidity" is 0.5; after link aggregation, the training link influence coefficient is output: cultivation 0.3, transportation 0.5, and sales 0.2.The server calculates the "training full-link abnormality probability" based on the training link abnormality confidence (0.8 for breeding, 0.2 for transportation, and 0.6 for sales) and the training link influence coefficient (0.3 for breeding, 0.5 for transportation, and 0.2 for sales): training full-link abnormality probability = (0.8 x 0.3) + (0.2 x 0.5) + (0.6 x 0.2) = 0.24 + 0.1 + 0.12 = 0.46. The actual full-link abnormality label for this batch is 1 (abnormal), and the server calculates the "third target deviation information" (mean square error): deviation = (1-0.46)² = 0.54² = 0.2916. Based on the third target deviation (e.g., 0.2916 for batch A, with an average deviation of 0.35 for all batches), the server adjusts the parameters of the original link influence degree analysis network (e.g., the weight matrix of the attention layer and the bias term of the fully connected layer) through the backpropagation algorithm. For example: increase the weight of the "breeding water temperature fluctuation" feature (from 0.3 to 0.5), because the breeding link abnormality has a greater actual impact on the full-link abnormality; reduce the weight of the "transportation temperature and humidity" feature (from 0.5 to 0.3), because the transportation link abnormality has a smaller actual impact on this batch. After multiple rounds of training (e.g., 150 rounds), the deviation of the original link influence degree analysis network is reduced to less than 0.1 (the error between the training full-link abnormality probability and the actual label is less than 10%). The server determines the optimized network as the "link influence degree analysis network" and integrates it with the trained time alignment submodule and link quality analysis network to form the final "trace abnormality classification network". This network can accurately calculate the link influence coefficient and output the full-link abnormality probability (e.g., the predicted probability of batch A is improved from 0.46 to 0.85, with a deviation of only 0.02 from the actual label 1). Through the supervision learning of the third training track set, the server optimizes the parameters of the link influence degree analysis network, enabling it to more accurately quantify the contribution of each link to the full-link abnormality. The finally constructed trace abnormality classification network can realize end-to-end analysis from multi-source perception data input to full-link abnormality probability output, providing reliable technical support for quality control of aquatic product enterprises.

[0111] In the embodiment of the present application, the link influence degree analysis network comprises a link interaction network and an attention network.

[0112] The link influence degree analysis network is used to analyze the influence degree of each link based on the perception feature vector, and the link influence coefficient of each trace link is obtained.

[0113] The multi-source dynamic feature vector of the perception feature vector is obtained by the link interaction network.

[0114] The attention network is used to perform attention allocation on the interaction feature vector, to obtain a link influence coefficient of each traceability link.

[0115] In the embodiments of the present application, an example is taken from the whole-link tracking scenario of soft-shelled turtles in a certain aquatic product enterprise. The server calculates the influence coefficient of each link through the link influence degree analysis network (including the link interaction network and the attention network) as follows: After completing the multi-source time alignment of the soft-shelled turtle RFID whole-link sensing data (generating a multi-source time alignment sequence) and the time sequence feature embedding mapping (outputting a sensing feature vector through the time sequence encoder), the server obtains a 128-dimensional "sensing feature vector". Each dimension of the vector corresponds to the key features of each link in the whole link, for example: dimensions 1-30: aquaculture link features (such as "water temperature fluctuation standard deviation", "dissolved oxygen below threshold duration proportion", "feeding frequency"); dimensions 31-60: transportation link features (such as "temperature and humidity exceeding standard duration", "emergency braking frequency average", "position deviation rate"); dimensions 61-90: sales link features (such as "storage temperature average", "time duration after shelving without sales", "query frequency by scanning code"); dimensions 91-128: cross-link interaction features (such as "correlation between aquaculture water temperature sudden drop and transportation temperature and humidity fluctuation"). The server calls the "link interaction network" (composed of 2 layers of fully connected layers and a gating mechanism) in the link influence degree analysis network to perform "multi-source dynamic feature aggregation" on the sensing feature vector. The specific operation is as follows: the server inputs the 128-dimensional sensing feature vector into the first layer of fully connected layers (128-256 dimensions). This layer extracts the preliminary interaction relationship between the features through linear transformation and ReLU activation function. For example: the "water temperature fluctuation standard deviation" (dimension 1) of the aquaculture link and the "temperature and humidity exceeding standard duration" (dimension 31) of the transportation link are calculated as a new feature "water temperature fluctuation-transportation temperature and humidity sensitivity coefficient" (dimension 150); the "emergency braking frequency average" (dimension 32) of the transportation link and the "storage temperature average" (dimension 61) of the sales link are calculated as "emergency braking stress-sales corruption risk coefficient" (dimension 200). The server filters the 256-dimensional preliminary interaction features through the gating mechanism (such as the update gate of GRU), and retains the features that have a greater impact on the whole-link anomaly. For example: if the weight of "water temperature fluctuation-transportation temperature and humidity sensitivity coefficient" (dimension 150) is 0.8 (indicating that the aquaculture water temperature fluctuation significantly affects the transportation temperature and humidity stability), the feature is retained; if the weight of "emergency braking stress-sales corruption risk coefficient" (dimension 200) is 0.2 (indicating that the emergency braking has a smaller impact on the sales corruption), the weight is reduced or discarded. The filtered features are input into the second layer of fully connected layers (256-128 dimensions), and a "interaction feature vector" is generated through linear transformation and Tanh activation function. This vector focuses on the feature interactions that have a significant impact on the whole-link anomaly (such as "aquaculture water temperature fluctuation-transportation temperature and humidity sensitivity" and "sales storage temperature-corruption acceleration").The server calls the "attention network" (composed of a query vector and a Softmax function) in the influence degree analysis network to perform "attention allocation" on the 128-dimensional interaction feature vector and calculate the influence coefficients of each link: the attention network calculates the attention score of each feature dimension through the dot product of the learnable query vector (128 dimensions) and the interaction feature vector. For example, the score of the "cultivation water temperature fluctuation-transport humidity sensitivity" feature (dimension 150 mapped to dimension 50 of the interaction feature vector) is 0.7 (indicating that the interaction has a greater impact on the whole-link abnormality); the score of the "sales storage temperature-corruption acceleration" feature (dimension 61 mapped to dimension 80 of the interaction feature vector) is 0.5 (indicating that the influence of the interaction is secondary). The server performs Softmax normalization on the attention scores to convert the scores into weights between 0 and 1 (the sum is 1). For example, the total weight of the cultivation-related interaction features is 0.6 (including the "water temperature fluctuation-transport sensitivity" feature); the total weight of the transport-related interaction features is 0.3 (including the "sudden braking stress-sales corruption" feature); and the total weight of the sales-related interaction features is 0.1 (including the "storage temperature-corruption acceleration" feature). The server takes the attention weights of each link as the "link influence coefficient". For example, the cultivation link influence coefficient = 0.6 (indicating that the cultivation link abnormality contributes 60% to the whole-link abnormality); the transport link influence coefficient = 0.3 (the transport link abnormality contributes 30%); and the sales link influence coefficient = 0.1 (the sales link abnormality contributes 10%). Finally, the server outputs the influence coefficients of each link (e.g., cultivation 0.6, transport 0.3, and sales 0.1) through the link interaction network and the attention network. The results show that the abnormality of the cultivation link (such as large water temperature fluctuation) has the greatest impact on the whole-link abnormality, and the enterprise can accordingly prioritize optimizing the environmental control of the cultivation link (such as upgrading the temperature control equipment) to reduce the overall quality risk. Through this process, the server accurately quantifies the influence of each link in the whole chain and provides a scientific basis for the abnormality control of the whole chain of aquatic products.

[0116] In the embodiment of the present application, the plurality of traceability links includes a target traceability link;

[0117] The method further comprises:

[0118] An traceability abnormality classification network for classifying traceability abnormality of a tracked product is obtained, and an adaptive base model for adaptive migration from the traceability abnormality classification network is obtained;

[0119] A fourth training trajectory set corresponding to the target traceability link is obtained; the fourth training trajectory set includes fourth traceability instance data;

[0120] The fourth traceability instance data is subjected to trajectory feature analysis by the traceability anomaly classification network, to obtain a full-link anomaly probability of the fourth traceability instance data, and the full-link anomaly probability of the fourth traceability instance data is determined as a link anomaly guide signal for training the adaptive base model;

[0121] The fourth traceability instance data is subjected to trajectory feature analysis by the adaptive base model, to obtain an adaptive full-link anomaly probability of the fourth traceability instance data;

[0122] Based on the adaptive full-link anomaly probability and the link anomaly guide signal, the model parameters of the adaptive base model are subjected to weight optimization, and the adaptive base model after weight optimization is determined as an optimized traceability model of the traceability anomaly classification network; the optimized traceability model is used for full-link traceability anomaly classification of RFID full-link sensing data under the target traceability link.

[0123] In the embodiments of the present application, an example is taken as an example of a whole-link tracking system of soft-shelled turtles of a certain aquatic product enterprise. The server optimizes the process of the traceability abnormality classification network for the "transportation link" (target traceability link) as follows: the server first loads the "traceability abnormality classification network" trained from the local model library. The network has been trained through multiple rounds (such as steps 5-14) and can evaluate the abnormal risk of the whole link of breeding, transportation, and sales. At the same time, the server obtains an "adaptation base model", which is the same as the original network structure (including a time sequence alignment submodule, a deep feature fusioner, a link quality analysis network, and a link influence degree analysis network) but with random initialization of parameters, which is used for transfer learning to optimize the performance of the target link. The server extracts the "fourth training track set" from the special database of the enterprise transportation link, which contains the transportation link RFID sensing data (such as temperature and humidity, sudden braking times, position deviation rate, etc.) of 300 batches of soft-shelled turtles in the past year, and the corresponding whole-link abnormality results (such as stress death and corruption of soft-shelled turtles caused by improper transportation). For example: Batch C: the temperature and humidity of the transportation link exceeded the standard for 2 hours (30°C / 70%), the sudden braking times were 5 times per hour, and the final quality inspection was corruption (whole-link abnormality); Batch D: the temperature and humidity of the transportation link was stable (28°C / 60%), the sudden braking times were 0 times per hour, and the final quality inspection was normal (no abnormality). The server calls the original traceability abnormality classification network to perform "track feature analysis" on each "fourth traceability instance data" (such as the transportation link data of batches C and D) in the fourth training track set: multi-source time sequence alignment: align the transportation link data with the breeding and sales link data (such as the breeding data ending on the 180th day and the transportation data starting on the 181st day); feature embedding and fusion: generate a sensing feature vector through a time sequence encoder and generate a fused deep feature through a deep feature fusioner; link abnormality evaluation and influence degree analysis: the link quality analysis network outputs the transportation link abnormality confidence (such as batch C is 0.9 and batch D is 0.1), and the link influence degree analysis network outputs the transportation link influence coefficient (such as batch C is 0.6 and batch D is 0.2); whole-link abnormality probability calculation: combine the abnormality confidence and the influence coefficient to output the whole-link abnormality probability (such as batch C is 0.9 x 0.6 = 0.54 and batch D is 0.1 x 0.2 = 0.02). The server determines these probability values as "link abnormality guidance signals" (such as the guidance signal of batch C is 0.54 and the guidance signal of batch D is 0.02) as the supervision target for the adaptation base model training.The server inputs the fourth traceability instance data of the same batch into the adaptive base model (parameters are random), and performs the same trajectory feature analysis process: because the parameters of the adaptive base model are not optimized, the timing alignment submodule may not accurately align the transportation link data (for example, misaligning "08:00 on the 181st day" as "23:00 on the 180th day"); the deep feature fusioner may ignore the correlation between "excessive temperature and humidity" and "frequent hard braking" (for example, failing to capture the rule that "hard braking causes temperature and humidity fluctuations"); the transportation link abnormal confidence output by the link quality analysis network may deviate from the original network (for example, the confidence of batch C is only 0.3, and the confidence of batch D is 0.4); the transportation link influence coefficient output by the link influence degree analysis network may be incorrect (for example, the influence coefficient of batch C is 0.2, and the influence coefficient of batch D is 0.7); the self-adaptive full-link abnormal probability calculated finally and the guidance signal differ significantly (for example, the self-adaptive full-link abnormal probability of batch C is 0.3 x 0.2 = 0.06, and the self-adaptive full-link abnormal probability of batch D is 0.4 x 0.7 = 0.28). The server calculates the mean square error of the "self-adaptive full-link abnormal probability" and the "link abnormality guidance signal" (for example, the deviation of batch C is (0.54-0.06)2=0.2304, and the deviation of batch D is (0.02-0.28)2=0.0676), and adjusts the parameters of the adaptive base model through the back propagation algorithm: optimizes the time calibration parameter of the timing alignment submodule (for example, corrects the time offset of the transportation link to "00:00 on the 181st day"); adjusts the attention weight of the deep feature fusioner (for example, increases the correlation weight of "excessive temperature and humidity duration" and "frequent hard braking"); optimizes the fully connected layer parameters of the transportation abnormality discriminator in the link quality analysis network (for example, increases the weight of "excessive temperature and humidity duration" from 0.3 to 0.7); adjusts the attention allocation of the link influence degree analysis network (for example, increases the influence coefficient weight of the transportation link from 0.2 to 0.6). After 50 rounds of training, the deviation of the adaptive base model is reduced to below 0.05 (for example, the self-adaptive probability of batch C is increased from 0.06 to 0.52, and the deviation from the guidance signal 0.54 is only 0.0004). The server determines the optimized adaptive base model as the "optimized traceability model", and the accuracy rate of the abnormal classification of the transportation link is increased from the initial 50% to 90%. The optimized traceability model is specially strengthened for the abnormal classification of the transportation link (the target traceability link), and can more accurately identify the temperature and humidity exceeding the standard, frequent hard braking and other abnormalities in the transportation process (for example, the abnormal probability of batch C is increased from 0.54 of the original network to 0.92, which is closer to the actual quality inspection result). Enterprises can monitor the key indicators of the transportation link through the model to reduce the quality loss caused by improper transportation. Through this process, the server realizes self-adaptive migration optimization of the target traceability link, and improves the pertinence and accuracy of the full-link abnormal classification.

[0124] In the embodiment of the application, the fourth traceability instance data comprises link operation instance data generated under the target traceability link;

[0125] The fourth training track set corresponding to the target traceability link can be executed by the following examples.

[0126] Obtain link operation instance data under the target traceability link;

[0127] Determine the to-be-identified traceability instance data from the traceability instance data set for track data screening;

[0128] Perform link feature matching analysis on the to-be-identified traceability instance data and the link operation instance data through the feature filter, to obtain a link matching degree index between the link operation instance data and the to-be-identified traceability instance data;

[0129] If the link matching degree index meets the track enhancement criterion, the to-be-identified traceability instance data is determined as target traceability instance data corresponding to the target traceability link;

[0130] The target traceability instance data and the link operation instance data are determined as fourth traceability instance data corresponding to the target traceability link.

[0131] In the embodiment of the present application, an example is taken from the whole-link tracking system of a certain aquatic product enterprise. The server is an abnormal classification model for optimizing the "transportation link" (target traceability link). The specific process of obtaining the fourth training trajectory set is as follows: the server first extracts "link operation instance data" from the special database of the enterprise transportation link. These data are the RFID sensing records actually generated by the transportation link, including temperature and humidity, number of sudden stops, position deviation rate and other key indicators. For example: data 1: transportation time "day 181 08:00-day 183 24:00", temperature and humidity "28℃ / 60% (stable)", number of sudden stops "0 times / hour", position deviation rate "0.1%" (normal transportation); data 2: transportation time "day 182 09:00-day 182 10:00", temperature and humidity "35℃ / 75% (over standard)", number of sudden stops "5 times / hour", position deviation rate "5%" (abnormal transportation). The server accesses the enterprise whole-link traceability database (including historical data of the cultivation, transportation and sales links), extracts "to-be-identified traceability instance data" from it, and all batch records containing transportation link data. For example: batch X: normal cultivation link (water temperature 25±1℃), transportation link data "day 181 08:00-day 183 24:00" (to be identified), normal sales link (storage temperature 4℃); batch Y: abnormal cultivation link (water temperature sudden drop 5℃), transportation link data "day 181 08:00-day 183 24:00" (to be identified), abnormal sales link (storage temperature 8℃). The server calls the "feature filter" (a pre-trained feature matching model) to perform "link feature matching analysis" on each to-be-identified data and the link operation instance data. The core of the feature filter is to extract the key features of the transportation link (such as temperature and humidity average, sudden stop number standard deviation, and position deviation rate maximum value), and calculate the similarity between them. Taking the to-be-identified data of batch X as an example: the feature vector of the link operation instance data is temperature and humidity average (28℃ / 60%), sudden stop number standard deviation (0 times), and position deviation rate maximum value (0.1%)—vector V1=[28, 60, 0, 0.1]; the feature vector of the to-be-identified data is temperature and humidity average (27℃ / 58%), sudden stop number standard deviation (1 time), and position deviation rate maximum value (0.2%)—vector V2=[27, 58, 1, 0.2]; the matching degree is calculated by using the cosine similarity formula to calculate the similarity between V1 and V2: similarity=(28×27+60×58+0×1+0.1×0.2) / (√(28²+60²+0²+0.1²)×√(27²+58²+1²+0.2²))≈0.98 (close to 1, high matching degree). The server sets the "trajectory enhancement criterion", and the link matching degree indicator ≥0.8 (indicating high feature similarity). If the matching degree of the to-be-identified data meets the criterion, it is determined as "target traceability instance data".For example, the matching degree of batch X is 0.98 (≥ 0.8), which meets the criteria and is selected as the target traceability instance data; the average temperature and humidity of the transportation link in the to-be-identified data of batch Y are 32℃ / 70% (abnormal characteristics), and the similarity with the link operation instance data is only 0.6 (< 0.8), which does not meet the criteria and is excluded. The server combines the selected target traceability instance data (such as batch X) and the original link operation instance data (such as data 1 and data 2) to form a "fourth training track set". This data set includes: normal transportation scenario data (such as data 1 and batch X); abnormal transportation scenario data (such as data 2); and other high-matching-degree transportation link data (such as to-be-identified data with a matching degree ≥ 0.8). Finally, the fourth training track set focuses on the typical scenarios (normal and abnormal) of the transportation link, and the data characteristics are highly consistent with the actual operation. The server will use this data set to train the adaptive base model in the future, optimize the abnormal classification performance of the transportation link (such as improving the recognition accuracy of temperature and humidity exceeding the standard and frequent emergency braking), and provide data support for the enterprise to accurately control the transportation risk. Through this process, the server realizes the accurate screening and enhancement of the target traceability link training data, ensuring the effectiveness and pertinence of subsequent model migration training.

[0132] In the embodiments of the present application, the determination of the full-link abnormal probability of the tracked article based on the link abnormal confidence of each traceability link and the link influence coefficient of each traceability link can be implemented through the following examples.

[0133] The link abnormal confidence of each traceability link is dynamically aggregated by the link influence coefficient of each traceability link to obtain a fusion abnormal evaluation value of each traceability link; and the weight of one traceability link is used to dynamically aggregate the link abnormal confidence of the corresponding traceability link.

[0134] Based on the fusion abnormal evaluation value of each traceability link, the full-link abnormal probability of the tracked article is determined.

[0135] In the embodiments of the present application, an example is taken from the whole-link tracking scene of soft-shelled turtles in a certain aquatic product enterprise. The server calculates the whole-link abnormal probability based on the abnormal confidence and influence coefficient of each link as follows: a batch of soft-shelled turtles has completed the whole-link circulation (cultivation-transportation-sale), and the server has obtained the key indicators of each link through the traceability abnormal classification network: link abnormal confidence (output by the link quality analysis network): cultivation link 0.8 (80% probability of abnormality, due to large water temperature fluctuation and insufficient dissolved oxygen), transportation link 0.3 (30% probability of abnormality, due to temporary over-standard temperature and humidity), and sales link 0.1 (10% probability of abnormality, due to slightly high storage temperature); link influence coefficient (output by the link influence degree analysis network): cultivation link 0.6 (contribution to the whole-link abnormality accounts for 60%), transportation link 0.3 (contribution 30%), and sales link 0.1 (contribution 10%). The server needs to combine the abnormal confidence and influence coefficient of each link through "dynamic aggregation" to calculate the "fusion abnormality evaluation value". The "dynamic" here means that the influence coefficient is not a fixed weight, but is dynamically calculated according to the current batch of perception characteristics (such as the fluctuation amplitude of cultivation water temperature, the number of emergency braking times during transportation, etc.), to ensure that the weight is related to the actual risk. Cultivation link fusion abnormality evaluation value: cultivation abnormal confidence (0.8) x cultivation influence coefficient (0.6) = 0.8 x 0.6 = 0.48; transportation link fusion abnormality evaluation value: transportation abnormal confidence (0.3) x transportation influence coefficient (0.3) = 0.3 x 0.3 = 0.09; sales link fusion abnormality evaluation value: sales abnormal confidence (0.1) x sales influence coefficient (0.1) = 0.1 x 0.1 = 0.01. The server adds up the fusion abnormality evaluation values of each link to obtain the "whole-link abnormal probability" of the batch of soft-shelled turtles: whole-link abnormal probability = cultivation fusion value + transportation fusion value + sales fusion value = 0.48 + 0.09 + 0.01 = 0.58 (i.e. 58%). The result (58% abnormal probability) indicates that there is a high abnormal risk in the whole-link of the batch of soft-shelled turtles. Further analysis of the contribution of each link: the fusion value of the cultivation link (0.48) accounts for 82.8% (0.48 / 0.58) of the whole-link probability, which is the main risk source, and the temperature control and oxygenation equipment of the cultivation pond need to be prioritized for investigation; the contribution of the transportation link (0.09) and the sales link (0.01) is lower, indicating that the risk is mainly concentrated in the cultivation stage, and the abnormalities of the transportation and sales links have less impact on the final result. The key of "dynamic aggregation" is that the influence coefficient adjusts with the change of data characteristics. For example, if the number of emergency braking times of another batch of soft-shelled turtles increases sharply (the influence coefficient may increase to 0.5), even if its abnormal confidence is only 0.2, the transportation fusion value will change to 0.2 x 0.5 = 0.1, becoming an important factor of the whole-link abnormality. This dynamic adjustment ensures the strong correlation between the evaluation result and the actual risk, and avoids the deviation caused by fixed weight.By dynamically aggregating the abnormal confidence and influence coefficient of each link, the server finally outputs the full-link abnormal probability (such as 58% in this example), providing the enterprise with a quantitative risk assessment result. The enterprise can accordingly optimize the high-risk link (such as the breeding link in this example) to reduce the full-link quality loss and achieve precise control.

[0136] The embodiment of the present application provides a computer device 100, which comprises a processor and a nonvolatile memory storing computer instructions, and when the computer instructions are executed by the processor, the computer device 100 executes the tracking and tracing method based on the RFID Internet of Things technology. As shown in the figure, Figure 2 Figure 2 The embodiment of the present application provides a structural block diagram of the computer device 100. The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113. In order to realize the transmission or interaction of data, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected with each other. For example, the electrical connection between these elements can be realized by one or more communication buses or signal lines.

[0137] The foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. The embodiments are chosen and described in order to best explain the principles of the disclosure and its practical application to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.​

Claims

1. The tracking and tracing method based on RFID Internet of Things technology is characterized by: include: Performing feature embedding processing on the RFID full-link perception data of the tracked item to obtain a perception feature vector of the RFID full-link perception data; Performing a deep feature fusion process on the perception feature vector to obtain a fused deep feature of the perception feature vector; Determine multiple traceability links corresponding to the RFID full-link perception data, perform link anomaly assessment on the fusion deep features based on each traceability link, and obtain link anomaly confidence of each traceability link; Performing link influence analysis on the perception feature vector to obtain link influence coefficients of each traceability link; Determining the abnormality probability of the entire link of the tracked item based on the link abnormality confidence of each traceability link and the link influence coefficient of each traceability link; The performing feature embedding processing on the RFID full-link perception data of the tracked item to obtain the perception feature vector of the RFID full-link perception data includes: When the RFID full-link sensing data of the tracked item is obtained, a traceability anomaly classification network for performing full-link traceability anomaly classification on the tracked item is obtained; the traceability anomaly classification network includes a timing alignment submodule; Determine, by the timing alignment submodule, the multi-source perception data information corresponding to the RFID full-link perception data, and extract the multi-source time sequence segment data of the RFID full-link perception data based on the multi-source perception data information; Performing time sequence segment alignment on the multi-source time sequence segment data by the time sequence alignment submodule to obtain aligned time sequence segment data of the multi-source time sequence segment data; Performing full-link timing arrangement on the aligned timing segment data of the multi-source timing segment data by the timing alignment submodule to obtain a multi-source timing alignment sequence of the RFID full-link perception data; Performing time series feature embedding mapping on the multi-source time series alignment sequence to obtain a perceptual feature vector of the multi-source time series alignment sequence; The fused deep features are obtained by performing deep feature fusion processing on the perception feature vector by the deep feature fusion device in the timing alignment submodule; the timing alignment submodule belongs to a traceability anomaly classification network used to perform full-link traceability anomaly classification on the tracked items.

2. The method according to claim 1, characterized in that The timing alignment submodule includes a timing data input module; the multi-source timing segment data includes a plurality of multi-source timing segment data, and the plurality of multi-source timing segment data includes target multi-source timing segment data; The step of aligning the multi-source time sequence segment data by the time sequence alignment submodule to obtain aligned time sequence segment data of the multi-source time sequence segment data includes: Determining a time sequence-link association space of the target multi-source time sequence segment data based on multi-source perception data information to which the target multi-source time sequence segment data belongs; the time sequence-link association space including a plurality of time sequence-link association values, each time sequence-link association value having a time sequence-link characteristic segment; The timing-link characteristic segment to which the target multi-source timing segment data belongs is determined by the timing data input module, and the timing-link association value of the timing-link characteristic segment to which the target multi-source timing segment data belongs is determined as the aligned timing segment data of the target multi-source timing segment data.

3. The method according to claim 1, characterized in that The method further comprises: Obtaining a first training trajectory set; wherein the first traceability instance data included in the first training trajectory set is traceability instance data without abnormal annotation; Obtain multiple analysis tasks for the original timing alignment submodule; The original timing alignment submodule performs trajectory feature analysis on the first traceability instance data to obtain timing fusion features corresponding to each analysis task, and determines multiple deviation information of the multiple analysis tasks based on the timing fusion features; one analysis task corresponds to one deviation information; the multiple analysis tasks include a link interruption detection task, a link change detection task, a link interference detection task, and a link quality self-supervision task; First target deviation information is determined based on the multiple deviation information, configuration parameters of the original timing alignment submodule are weight-optimized based on the first target deviation information, and the weight-optimized original timing alignment submodule is determined as the timing alignment submodule.

4. The method according to claim 1, wherein The traceability anomaly classification network also includes a link quality analysis network arranged after the deep feature fusion device; The step of performing link anomaly assessment on the fused deep features based on each traceability link to obtain the link anomaly confidence of each traceability link includes: Determining a link quality analysis network from the traceability anomaly classification network; the link quality analysis network includes an anomaly discriminator for each traceability link; Through the anomaly discriminator of each traceability link, the fusion deep feature is evaluated for link anomaly to obtain the link anomaly confidence of each traceability link; an anomaly discriminator is used to determine the link anomaly confidence of a traceability link; The method further comprises: Obtaining a second training trajectory set; the second traceability instance data included in the second training trajectory set is configured with a link identifier; the link identifier includes multiple link code values ​​of the multiple traceability links, one link code value corresponding to each traceability link, and the multiple link code values ​​are determined based on the traceability link to which the second traceability instance data belongs; Obtain a pre-trained timing alignment submodule, perform feature embedding processing on the second traceability instance data through the timing alignment submodule to obtain a fifth enhanced feature vector of the second traceability instance data, perform deep feature fusion processing on the fifth enhanced feature vector, and obtain a fifth timing fusion feature of the fifth enhanced feature vector; Obtain multiple initial anomaly discriminators for the multiple traceability links, perform link stage identification on the fifth time series fusion feature based on the multiple initial anomaly discriminators, and obtain multiple link stage identification confidences of the multiple initial anomaly discriminators; one traceability link corresponds to one initial anomaly discriminator, and one initial anomaly discriminator is used to determine one link stage identification confidence; Based on the multiple link stage identification confidences and the multiple link code values, second target deviation information is determined; based on the second target deviation information, the configuration parameters of the multiple initial abnormality discriminators are weight-optimized; the multiple initial abnormality discriminators after weight optimization are determined as multiple abnormality discriminators; and the link quality analysis network is determined based on the multiple abnormality discriminators.

5. The method according to claim 1, wherein The perception feature vector is obtained by performing timing feature embedding mapping on the multi-source timing alignment sequence by the timing encoder in the timing alignment submodule; the multi-source timing alignment sequence is obtained by performing multi-source timing alignment processing on the RFID full-link perception data by the timing data input module in the timing alignment submodule; the timing alignment submodule belongs to the traceability anomaly classification network for performing full-link traceability anomaly classification on the tracked item; The traceability anomaly classification network also includes a link impact analysis network arranged after the timing encoder; The performing link influence analysis on the perception feature vector to obtain the link influence coefficient of each traceability link includes: Determining a link impact analysis network from the traceability anomaly classification network; The link influence analysis network is used to perform link influence analysis on the perception feature vector to obtain the link influence coefficient of each traceability link.

6. The method according to claim 5, characterized in that The traceability anomaly classification network also includes a link quality analysis network for determining the link anomaly confidence level of each traceability link; The method further comprises: Obtaining a third training trajectory set; wherein the third traceability instance data included in the third training trajectory set is configured with a full-link anomaly identifier; Obtain a pre-trained timing alignment submodule, perform feature embedding processing on the third traceability instance data through the timing alignment submodule to obtain a sixth enhanced feature vector of the third traceability instance data, perform deep feature fusion processing on the sixth enhanced feature vector, and obtain a sixth timing fusion feature of the sixth enhanced feature vector; Obtain a pre-trained link quality analysis network, perform link anomaly assessment on the sixth time series fusion feature through multiple anomaly discriminators in the link quality analysis network, and obtain the training link anomaly confidence of each traceability link; an anomaly discriminator is used to determine the training link anomaly confidence of a traceability link; Performing link influence analysis on the sixth enhanced feature vector through the original link influence analysis network to obtain the training link influence coefficient of each traceability link; Determining a full-link abnormality probability of training for the sample tracking item associated with the third traceability instance data based on the training link abnormality confidence of each traceability link and the training link influence coefficient of each traceability link; Based on the full-link anomaly identification and the trained full-link anomaly probability, the third target deviation information is determined; based on the third target deviation information, the configuration parameters of the original link impact analysis network are weight optimized; the original link impact analysis network after weight optimization is determined as the link impact analysis network; based on the timing alignment submodule, the link quality analysis network and the link impact analysis network, the traceability anomaly classification network is determined.

7. The method according to claim 1, characterized in that The multiple traceability links include a target traceability link; The method further comprises: Acquire a traceability anomaly classification network for performing full-link traceability anomaly classification on tracked items, and acquire an adaptation base model for adaptive migration from the traceability anomaly classification network; Acquire a fourth training trajectory set corresponding to the target traceability link; the fourth training trajectory set includes fourth traceability instance data; performing trajectory feature analysis on the fourth traceability instance data through the traceability anomaly classification network to obtain a full-link anomaly probability of the fourth traceability instance data, and determining the full-link anomaly probability of the fourth traceability instance data as a link anomaly guidance signal for training the adaptive base model; Performing trajectory feature analysis on the fourth traceability instance data using the adaptive base model to obtain an adaptive full-link anomaly probability of the fourth traceability instance data; Based on the adaptive full-link anomaly probability and the link anomaly guidance signal, the model parameters of the adaptation base model are weighted optimized, and the weight-optimized adaptation base model is determined as the optimized tracing model of the traceability anomaly classification network; the optimized tracing model is used to perform full-link traceability anomaly classification on the RFID full-link perception data under the target traceability link.

8. The method according to claim 1, characterized in that The determining of the abnormality probability of the entire link of the tracked item based on the abnormality confidence of each traceability link and the link influence coefficient of each traceability link includes: Dynamically aggregate the link anomaly confidence of each traceability link through the link influence coefficient of each traceability link to obtain the fusion anomaly assessment value of each traceability link; the weight of a traceability link is used to dynamically aggregate the link anomaly confidence of the corresponding traceability link; Based on the fused anomaly evaluation values ​​of the traceability links, the full-link anomaly probability of the tracked item is determined.

9. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 8.

Images