Tracking and tracing method and system based on RFID Internet of Things technology

Abstract

The invention discloses a tracking and tracing method and system based on an RFID Internet of Things technology. The method 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: carrying out feature embedding processing on RFID full-link sensing data of a tracked article, and extracting a sensing feature vector; generating fusion depth features reflecting full-link association through depth feature fusion; evaluating the fusion depth feature based on a multi-traceability link to obtain the abnormal confidence of each link; analyzing the perception feature vector to obtain an influence coefficient of each link on the whole link abnormity; and dynamically aggregating the abnormal confidence coefficient and the influence coefficient, and calculating a full-link abnormal probability. According to the method, the problems of multi-source data time sequence alignment and feature fusion are solved, accurate positioning of abnormities of all links and dynamic evaluation of full-link risks are achieved, the method is suitable for full-link tracking scenes such as agricultural products and aquatic products, and the quality control efficiency is improved.

Description

Technical Field

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

[0002] Traditional tracking and tracing methods often rely on manual recording or single-sensor data, resulting in data dispersion, chaotic timing, and ambiguous anomaly location. While the development of RFID IoT technology has enabled the automated collection of sensor data across the entire chain, multi-source data can suffer from timing misalignment due to differences in collection frequency and device type. Furthermore, traditional methods struggle to efficiently integrate multi-source features and quantify the impact of each link on anomalies across the entire chain. This results in inaccurate anomaly assessments and makes it difficult to support dynamic quality control requirements. Summary of the Invention

[0003] The purpose of the present invention 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 invention provides a tracking and tracing method based on RFID Internet of Things technology, comprising: 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; Based on the link abnormality confidence of each traceability link and the link influence coefficient of each traceability link, the full-link abnormality probability of the tracked item is determined.

[0005] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0006] Compared with the existing technology, the beneficial effects provided by the present invention include: the present invention discloses a tracking and tracing method and system based on RFID Internet of Things technology, including: first, quantifying the risk of abnormalities in the entire link through multi-source data fusion and dynamic evaluation. The method includes: performing feature embedding processing on the RFID full-link perception data of the tracked items and extracting the perception feature vector; generating a fusion deep feature reflecting the association of the entire link through deep feature fusion; evaluating the fusion deep feature based on multiple traceability links to obtain the confidence level of abnormalities in each link; analyzing the perception feature vector to obtain the influence coefficient of each link on the abnormalities in the entire link; dynamically aggregating the abnormality confidence and influence coefficient to calculate the probability of abnormalities in the entire link. This method solves the problem of time series alignment and feature fusion of multi-source data, realizes the precise positioning of abnormalities in each link and the dynamic evaluation of risks in the entire link, is applicable to full-link tracking scenarios such as agricultural products and aquatic products, and improves the efficiency of quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0008] Figure 1 A schematic diagram of the steps of a tracking and tracing method based on RFID Internet of Things technology provided by an embodiment of the present invention; Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0010] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0011] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of a tracking and tracing method based on RFID Internet of Things technology provided in an embodiment of the present disclosure. The tracking and tracing method based on RFID Internet of Things technology is introduced in detail below.

[0012] Step S101, 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; Step S102, performing a depth feature fusion process on the perception feature vector to obtain a fused depth feature of the perception feature vector; Step S103: determining multiple traceability links corresponding to the RFID full-link perception data, performing link anomaly assessment on the fusion deep features based on each traceability link, and obtaining link anomaly confidence of each traceability link; Step S104: performing link influence analysis on the perception feature vector to obtain the link influence coefficient of each traceability link; Step S105 : 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.

[0013] In the embodiment of the present invention, for example, taking the full-link tracking and tracing scenario of aquatic products (such as soft-shelled turtles) of a certain aquaculture enterprise as an example, this embodiment uses the server as the execution subject to describe in detail the specific implementation process of the tracking and tracing method based on RFID Internet of Things technology. In the full-link process of aquatic products from breeding to sales, RFID reading and writing equipment and supporting sensors are deployed in each link (such as breeding, transportation, warehousing, and sales) to collect environmental data, operation records and other perception information associated with aquatic products in real time to form RFID full-link perception data. The server needs to analyze the data to assess the abnormal risk of the entire link. The specific execution steps are as follows: First, the server needs to perform trajectory feature analysis on the RFID full-link perception data and extract the perception feature vector. Since the full-link perception data usually comes from multi-source devices (such as temperature and humidity sensors in the breeding pond, positioning and temperature control equipment of the transport vehicle, and code scanning records of the sales terminal), the data has time sequence dislocation (such as hourly collection in the breeding link and minute collection in the transportation link) and dimensional differences (such as numerical temperature data and counting number of operations). To this end, the server calls the pre-trained traceability anomaly classification network, which contains a time series alignment submodule. The time series alignment submodule first identifies the device type and link attributes corresponding to the multi-source perception data (such as environmental sensors corresponding to the breeding link and vehicle-mounted equipment corresponding to the transportation link), and divides the original data into multi-source time series segments (such as "water temperature time series segments" in the breeding link and "temperature and humidity time series segments" in the transportation link); then, through the time series data input module, each segment is mapped to the predefined "time series-link association space", which pre-stores the time feature range of the data of each link (such as the breeding link timestamp is 0-180 days and the transportation link is 181-183 days). By matching the target The timestamp of the fragment and the "time series-link feature segment" in the associated space (such as the transportation link corresponds to 181-183 days) are used to align multi-source time series fragments; the aligned fragments are arranged into a continuous multi-source time series alignment sequence (a full-link data chain arranged in actual time order), and then feature embedding mapping is performed through a time series encoder (such as an LSTM network) to convert the original sequence of length N into a low-dimensional, semantically rich perceptual feature vector (for example, a 128-dimensional vector, each dimension represents key features such as "frequency of aquaculture water temperature fluctuations" and "duration of transportation temperature and humidity exceeding the standard").

[0014] Next, the server performs deep feature fusion processing on the perception feature vectors to extract fused deep features. Although the perception feature vectors are aligned, each dimension only reflects the local status of a single link or device (such as "aquaculture water temperature" or "number of sudden stops during transportation") and cannot directly express the correlated impact of the entire chain. To this end, the server calls the deep feature fusion device (such as a Transformer model or an attention mechanism network) in the traceability anomaly classification network. Through the self-attention mechanism, it calculates the correlation weights of each feature dimension (for example, "persistent abnormal aquaculture water temperature" is more correlated with "sudden changes in transportation temperature and humidity"). The server then fuses the scattered feature dimensions into a fused deep feature that comprehensively reflects the status of the entire chain (for example, generating a chain feature that simultaneously includes "abnormal aquaculture environment - transportation stress response - changes in storage quality").

[0015] The server then assesses the anomaly confidence level for each traceability link based on the fused deep features. The entire traceability chain is typically divided into multiple stages, such as aquaculture, transportation, warehousing, and sales. Each stage may have specific anomalies (e.g., low water temperature in the aquaculture stage, resulting in weak aquatic product, or excessive temperature and humidity in the transportation stage, accelerating spoilage). To this end, the traceability anomaly classification network integrates a link quality analysis network, which contains anomaly discriminators corresponding to each stage (e.g., a breeding discriminator and a transportation discriminator). The server inputs the fused deep features into each stage's anomaly discriminator (a pre-trained CNN or fully connected network). The discriminator then analyzes the key features of that stage (e.g., the aquaculture discriminator analyzes the "water temperature" and "dissolved oxygen" sub-features, while the transportation discriminator analyzes the "temperature and humidity" and "number of sudden braking" sub-features). It then outputs a confidence level for the stage anomaly (a probability value between 0 and 1, e.g., a confidence level of 0.8 for the aquaculture stage indicates an 80% probability of an anomaly).

[0016] At the same time, the server needs to analyze the impact of each link on the overall quality chain. Anomalies in different links contribute to the final result to varying degrees (for example, anomalies in the aquaculture process may be mitigated by timely adjustments in subsequent transportation, while chronic over-temperature storage in the sales process may directly lead to spoilage). To this end, the server invokes the link impact analysis network (such as an attention-based regression model) within the traceability anomaly classification network to analyze the perceived feature vector. This network, trained on historical data (e.g., the full-chain data for the past 1,000 batches of aquatic products), calculates the correlation between each link's anomaly and the final quality failure (i.e., the link impact coefficient. For example, a 0.6 impact coefficient for the aquaculture process indicates that the anomaly contributes 60% to the overall quality chain result).

[0017] Finally, the server combines the anomaly confidence and impact coefficient of each link to calculate the anomaly probability for the entire chain. This is done through a dynamic aggregation method: the anomaly confidence and impact coefficient for each link are multiplied (for example, 0.8 × 0.6 = 0.48 for the aquaculture link, 0.3 × 0.3 = 0.09 for the transportation link, and 0.1 × 0.1 = 0.01 for the sales link). The sum of these sums yields the anomaly probability for the entire chain (0.48 + 0.09 + 0.01 = 0.58, or 58%). This probability directly reflects the risk of anomalies occurring throughout the entire aquatic product chain, providing a basis for quality control by businesses (for example, focusing on inspecting water temperature control equipment in the aquaculture link and strengthening temperature and humidity monitoring in the transportation link).

[0018] In summary, this method collects full-link perception data through RFID Internet of Things technology, and combines it with a deep learning network to perform time series alignment, feature fusion, link anomaly assessment and impact analysis on the data, ultimately achieving accurate quantification of the abnormal risk of the entire link of aquatic products, providing efficient and reliable technical support for agricultural product traceability.

[0019] In an embodiment of the present invention, the feature embedding processing of the RFID full-link perception data of the tracked item to obtain the perception feature vector of the RFID full-link perception data can be implemented through the following examples.

[0020] Performing multi-source timing alignment processing on the RFID full-link perception data 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.

[0021] In an embodiment of the present invention, for example, a full-chain tracking scenario for soft-shelled turtles at an aquatic product company requires a server to perform feature embedding processing on RFID sensing data from the breeding, transportation, and sales stages. The specific implementation process is as follows: First, the server performs multi-source time series alignment on the RFID sensing data from the entire chain. Throughout the entire chain from breeding to sales, RFID sensing data is collected by multiple types of equipment: underwater temperature and humidity sensors in the breeding stage record water temperature on an hourly basis (e.g., one data point per hour from 0 to 180 days); on-board RFID readers and thermometers in the transportation stage collect location and temperature on a minute-by-minute basis (e.g., one data point per minute from 181 to 183 days); and supermarket RFID terminals in the sales stage record shelf time and storage temperature on a daily basis (e.g., one data point per day from 184 to 185 days). Due to varying collection frequencies and device types, these data have misaligned raw timestamps (for example, the "water temperature at 20:00 on day 180" for the aquaculture process and the "temperature and humidity at 08:00 on day 181" for the transportation process are discontinuous on the timeline). The data also has mixed formats (numeric temperature values ​​and textual location descriptions). The server invokes the time series alignment submodule within the pre-trained traceability anomaly classification network. It first identifies the "multi-source perception data information" corresponding to each data point. It uses data labels (such as "aquaculture - water temperature," "transportation - temperature and humidity," and "sales - storage temperature") to determine the data's associated process and device type. The raw data is then segmented into independent multi-source time series segments (for example, the "water temperature time series segment" for the aquaculture process is [Day 1 08:00, 25°C; Day 1 09:00, 26°C...], while the "temperature and humidity time series segment" for the transportation process is [Day 181 08:00, 28°C / 60%; Day 181 08:01, 28°C / 62%...]). Next, the time series alignment submodule maps each fragment to a predefined "time series-link association space", which stores the time feature range of the data of each link (such as 0-180 days for the breeding link, 181-183 days for the transportation link, and 184-185 days for the sales link). Through the time series data input module, the server matches the timestamp of each time series segment with the "time series-link feature segment" in the associated space (for example, matching "08:00 on Day 181" in the transportation segment to the "181-183" segment). It also adjusts the timestamp offset of each segment (for example, unifying "08:00 on Day 181" in the transportation segment data to "08:00 on Day 181" in the full-link timeline). It also fills in missing time points (for example, if there is no data for a certain minute in the transportation segment, it is filled in by interpolating the previous and next data). Ultimately, the multi-source time series segments are arranged into a continuous "multi-source time series alignment sequence" (for example, [breeding - Day 1 08:00-25°C, breeding - Day 1 09:00-26°C...Transportation - Day 181 08:00-28°C / 60%...Sales - Day 184 09:00-4°C]). After completing the multi-source time series alignment, the server performs time series feature embedding mapping on the sequence.Although multi-source time series alignment sequences are arranged in chronological order, they can be thousands or even tens of thousands of data points long (for example, 180 days x 24 hours = 4,320 data points in the aquaculture process). Direct analysis is inefficient and redundant. The server invokes a time series encoder (such as an LSTM network) in the time series alignment submodule and inputs the sequence into the encoder. The encoder extracts time series features layer by layer using a recurrent neural network. The first layer learns local patterns between adjacent time points (e.g., "water temperature rises by 1°C per hour"), while the second layer captures trends across time periods (e.g., "water temperature is stable for the first 30 days of aquaculture, then plummets by 5°C on the 31st"). Ultimately, the entire sequence is compressed into a 128-dimensional perceptual feature vector. Each dimension of this vector corresponds to a key feature (e.g., "standard deviation of aquaculture water temperature fluctuation," "percentage of time spent exceeding temperature and humidity standards during transportation," and "average sales and storage temperature"). This preserves the time series information of the entire data chain while removing redundant noise, providing efficient input for subsequent in-depth analysis. The server thus completes the conversion of multi-source, misaligned RFID perceptual data into a structured perceptual feature vector, laying the foundation for full-chain anomaly assessment.

[0022] In an embodiment of the present invention, the multi-source timing alignment processing is performed on the RFID full-link perception data to obtain a multi-source timing alignment sequence of the RFID full-link perception data, which can be implemented through the following examples.

[0023] 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; The timing alignment submodule performs full-link timing arrangement on the aligned timing segment data of the multi-source timing segment data to obtain a multi-source timing alignment sequence of the RFID full-link perception data.

[0024] In an embodiment of the present invention, taking the full-link tracking scenario of soft-shelled turtles at an aquatic product enterprise as an example, the server processes the multi-source temporal alignment of its RFID full-link perception data as follows: When the server receives the full-link RFID perception data of the soft-shelled turtles (including environmental and operational data from the breeding, transportation, and sales stages), it first calls a pre-trained "source traceability anomaly classification network" from a local model library or the cloud. This network is a deep learning model pre-trained by the server for the full-link traceability task of aquatic products. Its core function is to identify anomaly risks along the entire chain. It also integrates a "temporal alignment submodule" specifically designed to address temporal disarray in multi-source perception data. For example, this submodule has been trained using historical data (e.g., full-link data from the past 1,000 batches of soft-shelled turtles) and is familiar with the time characteristic ranges and data formats of stages such as breeding (0-180 days), transportation (181-183 days), and sales (184-185 days). The server uses the temporal alignment submodule to first analyze the metadata of the full-link RFID perception data (e.g., data tags, device ID, and timestamps) to determine the "multi-source perception data information," namely, the stage and device type to which each piece of data belongs. For example, data labeled "Aquaculture - Temperature and Humidity Sensor" belongs to the aquaculture process and is collected by underwater sensors. The timestamp range is 0-180 days, and the collection frequency is once per hour. Data labeled "Transportation - Vehicle-mounted RFID" belongs to the transportation process and is collected by vehicle-mounted readers and thermometers and hygrometers. The timestamp range is 181-183 days, and the collection frequency is once per minute. Data labeled "Sales - Terminal RFID" belongs to the sales process and is collected by supermarket RFID terminals. The timestamp range is 184-185 days, and the collection frequency is once per day. Based on this information, the server segments the raw end-to-end data into independent "multi-source time series segments": temperature and humidity data from the farming phase is extracted as one time series segment (e.g., [Day 1 08:00 - 25°C / 60%, Day 1 09:00 - 26°C / 58%…Day 180 20:00 - 24°C / 62%]); location and temperature and humidity data from the transportation phase are extracted as another time series segment (e.g., [Day 181 08:00 - 112°E longitude / 28°N latitude - 28°C / 60%, Day 181 08:01 - 112.1°E longitude / 28.1°N latitude - 28°C / 62%…]); and storage temperature data from the sales phase is extracted as a third time series segment (e.g., [Day 184 09:00 - 4°C, Day 185 09:00 - 3°C]). The server aligns these multi-source time series segments using the "Time Series Data Input Module" of the Time Series Alignment submodule. The specific operations are as follows: Matching time series-link association space: This space pre-stores the standard time characteristic segments of each link (such as breeding link is 0-180 days, transportation is 181-183 days, and sales is 184-185 days).The server matches the timestamp of each fragment with the corresponding segment (for example, "08:00 on the 181st day" of the transportation link is matched to the "181-183 days" segment). Adjust the timestamp offset: If the timestamp of a fragment is offset (for example, the timestamp of a piece of data in the transportation link is mistakenly marked as "23:00 on the 180th day"), the server corrects it to "00:00 on the 181st day" based on the segment range to ensure consistency with the link time range. Fill in missing data: If a fragment is missing on the timeline (for example, there is no data at 08:05 on the 181st day of the transportation link), the server uses linear interpolation to fill in the gaps between the two data points (08:04-28℃ / 62%, 08:06-28℃ / 63%) and the data points before and after (08:05-28℃ / 62.5%). Ultimately, each segment is aligned into "aligned time series segments" with continuous, complete time series (e.g., the farming segment ends at 24:00 on day 180, the transportation segment begins at 00:00 on day 181, and the sales segment begins 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, forming a "multi-source time series alignment sequence." For example, the last data item in the farming phase is "24:00 on day 180 - 24°C / 62%"; the first data item in the transportation phase is "00:00 on day 181 - 112.2°E longitude / 28.2°N latitude - 28°C / 60%"; and the first data item in the sales phase is "00:00 on day 184 - 4°C." This data, arranged chronologically, forms a continuous data chain covering days 0-185 (e.g., [breeding - day 1 08:00 - 25°C / 60%...breeding - day 180 24:00 - 24°C / 62%, transportation - day 181 00:00 - 28°C / 60%...transportation - day 183 24:00 - 29°C / 65%, sales - day 184 00:00 - 4°C...]). The server now completes multi-source time series alignment, transforming the originally scattered and misaligned RFID sensor data into a clearly structured, temporally continuous multi-source time series alignment sequence, providing standardized input for subsequent feature embedding and anomaly analysis.

[0025] In an embodiment of the present invention, 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 may be implemented through the following examples.

[0026] 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.

[0027] In an embodiment of the present invention, illustratively, taking the full-link tracking scenario of soft-shelled turtles in a certain aquatic product enterprise as an example, the server processes the time series segment alignment process of the "target multi-source time series segment data" (i.e., the temperature, humidity and position data segments of the transportation link) in the transportation link as follows: the server first analyzes the "multi-source perception data information" of the target multi-source time series segment data, and determines that the segment belongs to the transportation link through the data tag (such as "transportation-on-board RFID-temperature and humidity"), the device ID (such as "on-board device 007") and the timestamp (such as "181st day 08:00-183rd day 24:00"), and the standard time range should be 181-183 days. Based on this, the server calls the predefined "time series-link association space" in the time series alignment submodule. This space is a structured database that stores the "time series-link association values" of each link in the entire link, and each association value corresponds to a "time series-link feature segment" (i.e., the time range and format requirements that the link data should cover). For example, the associated value for the breeding link is "breeding-environmental sensor," and the corresponding feature segment is "0-180 days, timestamp accuracy at the hour level, data format [time, water temperature, dissolved oxygen content]"; the associated value for the transportation link is "transportation-vehicle-mounted RFID," and the corresponding feature segment is "181-183 days, timestamp accuracy at the minute level, data format [time, longitude, latitude, temperature, humidity]"; the associated value for the sales link is "sales-terminal RFID," and the corresponding feature segment is "184-185 days, timestamp accuracy at the day level, data format [time, storage temperature]." The server uses the "Time Series Data Input Module" of the Time Series Alignment submodule to match the target multi-source time series segment data (temperature, humidity, and location data from the transportation phase) with the characteristic segments in the associated space. The server checks the timestamp of the target segment. A record in the original data has a timestamp of "Day 180, 23:50" (which falls within the time range of the farming phase, 0-180 days). However, based on the multi-source perception data (for the transportation phase), this timestamp should fall within the time range of days 181-183 of the transportation phase. The server determines that this timestamp is offset (possibly due to device clock errors) and corrects it to "Day 181, 00:00" (the start time of the transportation phase). The target segment contains discontinuous time intervals (for example, data exists at 08:00 on Day 181, but not at 08:01, and data exists at 08:02). The server uses linear interpolation to complete the missing points based on the minute-level accuracy requirements for the transport segment's characteristic segments: Given that the temperature at 08:00 is 28°C and the humidity is 60%; the temperature at 08:02 is 28°C and the humidity is 62%; the temperature at 08:01 is 28°C (unchanged) and the humidity is 61% (taking the middle value). Some data in the target segment is not formatted correctly (for example, one record's location information is "near 112 East longitude" instead of 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.

[0028] 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; The following embodiments are also provided.

[0029] 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; 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; 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.

[0030] In this embodiment of the present invention, the server extracts a "first training trajectory set" from the enterprise database. This dataset contains full-chain RFID sensing data for 1,000 batches of soft-shelled turtles over the past six months (e.g., water temperature and dissolved oxygen content during the breeding phase, temperature, humidity, and location during transportation, and storage temperature during sales). However, no abnormality tags are annotated (due to the low frequency of abnormal events in real-world scenarios and the high cost of manual labeling). Each data entry contains only a timestamp, device ID, and the raw sensing value (e.g., "breeding - day 1 08:00 - 25°C"; "transportation - day 181 08:00 - 28°C / 60%"). The server sets two analysis tasks for the original time series alignment submodule (the initial model before training) to evaluate its feature extraction capabilities: Time series reconstruction task: Input the perception data of the first N time points of a certain link (such as the water temperature data of the 1st to 10th days in the breeding link), and require the model to output the predicted value of the N+1th time point (the water temperature on the 11th day) to verify whether the model captures the time series pattern; Feature consistency task: The data collected by multi-source equipment in the same link (such as the water temperature sensor and dissolved oxygen sensor in the breeding link) should reflect the same environmental state. The feature vectors (water temperature features, dissolved oxygen features) output by the model are required to have high similarity to verify whether the model integrates the correlation of multi-source data. The server calls the original time series alignment submodule (with randomly initialized parameters) to analyze the trajectory features of each data point in the first training trajectory set, obtaining the "time series fusion features" corresponding to each analysis task and calculating the deviation. For example, for a data point from the aquaculture process, the original data consists of "08:00-25°C on day 1, 09:00-26°C on day 1, and 20:00-24°C on day 10" (a total of 240 time points). The original submodule extracts the perceived feature vectors and generates time series fusion features (including the water temperature trend). The server asks the model to predict the water temperature at 21:00 on day 10 (the actual value is 24°C). The model outputs a predicted value of 22°C (due to poor initial parameter optimization, the "water temperature is stable at night" pattern is not captured). The mean squared error ((24-22)² = 4) is calculated as the deviation information for this task. For example, consider the water temperature sensor (data A) and dissolved oxygen sensor (data B) used in aquaculture. The original submodule extracts the feature vector V1 (e.g., [0.2, 0.5, 0.3]) for data A and the feature vector V2 (e.g., [0.1, 0.8, 0.1]) for data B. The server calculates the cosine similarity between the two vectors (0.2 × 0.1 + 0.5 × 0.8 + 0.3 × 0.1 = 0.45). Since water temperature and dissolved oxygen are positively correlated in practice (increasing water temperature can reduce dissolved oxygen), the ideal similarity should be higher (e.g., 0.7). Therefore, 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.

[0031] In an embodiment of the present invention, the plurality of analysis tasks include a link interruption detection task; 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: 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; 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; 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; Performing time series feature embedding mapping on the first enhanced time series trajectory set by the original time series alignment submodule to obtain a first enhanced feature vector of the first enhanced time series trajectory set, performing deep feature fusion processing on the first enhanced feature vector to obtain a first time series fusion feature of the first enhanced feature vector; the first time series fusion feature belongs to the time series fusion feature; Performing noise link identification processing on the first time series fusion feature by a detector module corresponding to the link interruption detection task to obtain predicted aligned time series segment data of the noise link markers in the first enhanced time series trajectory set; Based on the sample alignment timing segment data and the predicted alignment timing segment data on the first timing node in the first sample multi-source timing alignment sequence, deviation information of the link interruption detection task is determined.

[0032] In an embodiment of the present invention, for example, taking a turtle full-link tracking system of a certain aquatic product company as an example, the process of optimizing the original timing alignment submodule through the "link interruption detection task" by the server is as follows: the server first extracts a batch of data from the "first training trajectory set" (1000 batches of unlabeled turtle full-link data) and performs multi-source timing alignment processing through the original timing alignment submodule. This batch of data contains perception data from the breeding (0-180 days), transportation (181-183 days), and sales (184-185 days) links. After alignment, a "first sample multi-source timing alignment sequence" is generated, a time-continuous full-link data chain containing thousands of "time series nodes" (each node corresponds to the aligned fragment data at a specific time point). For example, the data for the "08:00 AM on Day 181" node in the transportation process is "112.0°E / 28.0°N-28°C / 60%," and the data for the "08:01 AM on Day 181" node is "112.1°E / 28.1°N-28°C / 61%." To detect a link interruption, the server dynamically selects a "first time series node" (i.e., the time point at which the link interruption is simulated) from the thousands of time series nodes in the aligned sequence. The selection rule is based on data distribution: Prioritizing periods of relatively stable data changes in the transportation process (such as 08:00-08:30 on Day 181, when the turtles are in transit and temperature and humidity fluctuate minimally) to highlight anomalies after noise injection. Ultimately, the "08:00 AM on Day 181" node was selected as the target. The server performs "link noise injection" on the original data at the "Day 181, 08:00" node ("112.0°E / 28.0°N - 28°C / 60%"). The noise type simulates a transportation disruption (e.g., equipment failure) by replacing the temperature at that node with an abnormal value (from 28°C to 40°C) and removing the location information (leaving the East Longitude / North Latitude fields blank). The noise tag adds a "noise link tag" (e.g., "Transport - Day 181, 08:00 - Noise") to the node and records the original data ("28°C / 60%" and "112.0°E / 28.0°N") for reference. This process generates the "first enhanced time series trajectory set." In the original alignment sequence, only the "Day 181, 08:00" node is contaminated, while the remaining nodes remain unchanged. The server calls the original time series alignment submodule to perform feature processing on the enhanced trajectory set: Time series feature embedding mapping: The enhanced alignment sequence is input into the submodule's time series encoder (such as an LSTM network), and a 128-dimensional "first enhanced feature vector" (containing node information contaminated by noise) is output; deep feature fusion: The enhanced feature vector is fused through the submodule's deep feature fuser (such as an attention network) to extract the "first time series fusion feature", which integrates the time series correlation of the entire link data, especially focusing on the transportation links disturbed by noise.The server invokes the "detector module" corresponding to the link interruption detection task (part of the original submodule, responsible for identifying abnormal nodes) and inputs the first time series fusion feature into it. The detector module analyzes the features for abnormal signals (such as a sudden temperature rise and missing location) to locate the node "08:00 on Day 181" as a noise link and attempts to predict the original aligned segment data for this node (i.e., the correct value uncontaminated by noise). For example, based on the temperatures of the preceding and following nodes (28°C / 61% at 08:01 and 28°C / 59% at 07:59), the detector module predicts that the temperature at "08:00 on Day 181" should be 28°C, the humidity should be 60%, and the location should be "112.0° East Longitude / 28.0° North Latitude." The server compares the detector module's prediction ("28°C / 60%" and "112.0°E / 28.0°N") with the original, uncontaminated data ("28°C / 60%" and "112.0°E / 28.0°N") and calculates the deviations: Temperature deviation: The predicted value of 28°C matches the original value of 28°C, resulting in a deviation of 0; Humidity deviation: The predicted value of 60% matches the original value of 60%, resulting in a deviation of 0; Position deviation: The predicted longitude and latitude exactly match the original values, resulting in a deviation of 0. If the original submodule is not optimized (e.g., if the initial parameters are random), the detector module might predict a temperature of 35°C (missing the pattern of stable temperature during transportation), resulting in a temperature deviation of (28-35)² = 49. After training and optimization, the deviations gradually decrease, indicating that the submodule's ability to identify link interruptions has improved. Through the link interruption detection task, the server simulates an abnormal interruption in a certain link in the entire link (such as equipment failure), and optimizes the parameters of the original timing alignment sub-module through a closed loop of noise injection-prediction-deviation calculation, enabling it to more accurately identify and repair data interruptions, thereby improving the reliability of the entire link perception data.

[0033] In an embodiment of the present invention, the plurality of analysis tasks include a link change detection task; The method of 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 can be implemented through the following example.

[0034] 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; Dynamically selecting a second timing node to be injected with the link disturbance from a plurality of timing nodes of the multi-source timing alignment sequence of the first sample; performing link perturbation injection processing on the sample aligned time series segment data at the second time series node in the first sample multi-source time series alignment sequence based on a random link perturbation factor, to obtain a second enhanced time series trajectory set; the random link perturbation factor is different from the sample aligned time series segment data at the second time series node; Performing time series feature embedding mapping on the second enhanced time series trajectory set by the original time series alignment submodule to obtain a second enhanced feature vector of the second enhanced time series trajectory set, performing deep feature fusion processing on the second enhanced feature vector to obtain a second time series fusion feature of the second enhanced feature vector; the second time series fusion feature belongs to the time series fusion feature; Performing disturbance location analysis on the second time series fusion feature through a detector module corresponding to the link anomaly detection task to obtain disturbance link prediction results of the multiple time series nodes; Based on the disturbance link prediction results of the second timing node and the multiple timing nodes, the deviation information of the link abnormality detection task is determined.

[0035] In an embodiment of the present invention, taking a turtle full-link tracking system for an aquatic product company as an example, the server optimizes the original time series alignment submodule through the "link change detection task" as follows: the server selects a batch of data from the "first training trajectory set" (1,000 batches of unlabeled turtle full-link data) and performs multi-source time series alignment processing through the original time series alignment submodule. This batch of data covers the breeding (0-180 days), transportation (181-183 days), and sales (184-185 days). After alignment, the "first sample multi-source time series alignment sequence" is generated: a time-continuous full-link data chain containing thousands of "time series nodes" (each node corresponds to a segment of aligned data at a specific time point). For example, the data for the "09:00 on Day 181" node in the transportation link is "112.2°E / 28.2°N-28°C / 65%" (stable temperature and humidity, within the normal range during transportation). To detect link anomalies, the server dynamically selects a "second time series node" from the thousands of time series nodes in the aligned sequence. This node is chosen during a period of stable data during the transportation process (e.g., 09:00 on day 181, during which the turtles experienced no sudden braking or sudden changes in temperature or humidity) to highlight any anomalies after the disturbance. Ultimately, the "09:00 on day 181" node is selected as the target. The server performs a "link perturbation injection" on the raw data at 09:00 on day 181 ("112.2°E / 28.2°N - 28°C / 65%"). The perturbation factor is randomly generated, varying the original data (e.g., increasing temperature by 5°C to 33°C, decreasing humidity by 10% to 55%). The perturbation flag is recorded, along with the original data at that node ("28°C / 65%", "112.2°E / 28.2°N"), as a reference. This generates a "second enhanced time series trajectory set." Only the "09:00 on day 181" node is perturbed, while the remaining nodes remain unchanged. The server invokes the original time series alignment submodule to perform feature processing on the second enhanced trajectory set: Time series feature embedding and mapping: The enhanced aligned sequence is input into the submodule's time series encoder (such as an LSTM network), which outputs a 128-dimensional "second enhanced feature vector" (containing information about the perturbed node). Deep feature fusion: The submodule's deep feature fuser (such as an attention network) fuses the enhanced feature vector to extract the "second time series fusion feature." This feature integrates the temporal correlations of the entire link data and focuses on abnormal signals (such as a sudden temperature rise or humidity drop) at the perturbed node. The server invokes the "detector module" corresponding to the link change detection task (part of the original submodule, used to locate abnormal perturbations) and inputs the second time series fusion feature into it. The detector module analyzes abnormal patterns in the features (such as a temperature deviation of 5°C from the historical mean or a humidity deviation of 2 times the standard deviation), traverses all time series nodes, and outputs a "perturbed link prediction result." It identifies the node "09:00 on Day 181" as a perturbed link (with a probability of 0.95), while the remaining nodes are not perturbed (with probabilities below 0.1).The server compares the detector module's prediction ("Day 181, 09:00" as the disturbance node) with the actual second time series node ("Day 181, 09:00") where the disturbance was injected, and calculates the deviation. Positioning accuracy: If the predicted node and the actual node are completely consistent, the deviation is 0. If the original submodule is not optimized (e.g., if the initial parameters are random), the detector module may mistakenly identify the node as "Day 181, 09:30" (because it fails to capture the local characteristics of the sudden temperature change), resulting in a positioning deviation of 30 minutes (the square of the time difference). After training and optimization, the deviation gradually decreases, indicating that the submodule's ability to locate link anomalies has improved. Through the link anomaly detection task, the server simulates an abnormal disturbance at a specific link in the entire chain (such as temporary high temperatures during transportation). Through a closed loop of disturbance injection, positioning, and deviation calculation, the parameters of the original time series alignment submodule are optimized, enabling it to more accurately identify and locate data anomalies, improving the anomaly detection capabilities of data perception throughout the entire chain.

[0036] In an embodiment of the present invention, the multiple analysis tasks include a link interference detection task; The method of 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 can be implemented through the following example.

[0037] Dynamically injecting link interference features into multi-source time series segment data of the first traceability instance data to obtain interference traceability instance data, performing multi-source time series alignment processing on the interference traceability instance data to obtain a noise multi-source time series alignment sequence of the interference traceability instance data; Performing multi-source time series alignment processing on the first traceability instance data to obtain a first sample multi-source time series alignment sequence of the first traceability instance data; Performing time series feature embedding mapping on the noise multi-source time series alignment sequence through the original time series alignment submodule to obtain a third enhanced feature vector of the noise multi-source time series alignment sequence, performing deep feature fusion processing on the third enhanced feature vector to obtain a third time series fusion feature of the third enhanced feature vector; the third time series fusion feature belongs to the time series fusion feature; Performing time series feature embedding mapping on the first sample multi-source time series alignment sequence through the original time series alignment submodule to obtain a fourth enhanced feature vector of the first sample multi-source time series alignment sequence, performing deep feature fusion processing on the fourth enhanced feature vector to obtain a fourth time series fusion feature of the fourth enhanced feature vector; the fourth time series fusion feature belongs to the time series fusion feature; Based on the third timing fusion feature and the fourth timing fusion feature, deviation information of the link interference detection task is determined.

[0038] In an embodiment of the present invention, for example, taking the full-link tracking system of soft-shelled turtles of a certain aquatic product enterprise as an example, the process of the server optimizing the original timing alignment submodule through the "link interference detection task" is as follows: the server selects a batch of original data (including perception information of the breeding, transportation, and sales links) from the "first training trajectory set" (1,000 batches of unlabeled full-link data of soft-shelled turtles), and dynamically injects "link interference features" into the transportation link. The specific operations are as follows: Interference link selection: select the period when the data is stable in the transportation link (such as 10:00-10:30 on the 181st day, during which the turtles are in a uniform transportation state and the temperature and humidity fluctuate little); Interference feature injection: For the "181st day 10:00" node data within this period (the original value is "112.3° East longitude / 28.3° North latitude-28℃ / 65%"), randomly generate interference features, modify the temperature value to the abnormal 35℃ (5-8℃ higher than the normal range), and add an erroneous position offset (112.3° East longitude is changed to 115.0° to simulate GPS signal interference) to generate "interference tracing instance data". The server performs multi-source timing alignment processing on the interference tracing instance data and the original data respectively: Noise multi-source timing alignment sequence: After the interference data is processed by the timing alignment submodule, a "noise multi-source timing alignment sequence" is generated. In this sequence, the node data of "10:00 on the 181st day" is "115.0° East longitude / 28.3° North latitude-35°C / 65%" (data after interference), and the rest of the nodes remain normal (for example, 09:59 on the 181st day is "112.3 ° / 28.3°N-28℃ / 64%", and at 10:01 on the 181st day it is "112.3°E / 28.3°N-28℃ / 66%"); the first sample multi-source timing alignment sequence: after the original data is aligned, the "first sample multi-source timing alignment sequence" is generated. The node data of "10:00 on the 181st day" is the original correct value ("112.3°E / 28.3°N-28℃ / 65%"), and the remaining nodes are consistent with the interference data. The server calls the original time series alignment submodule to process the noisy multi-source time series alignment sequence and the first sample multi-source time series alignment sequence respectively: the third enhanced feature vector and the third time series fusion feature: the noise sequence is input into the time series encoder of the submodule (such as the LSTM network), and outputs a 128-dimensional "third enhanced feature vector" (containing abnormal information of the interference node); then it is fused through a deep feature fusion device (such as an attention network) to obtain the "third time series fusion feature", which is affected by interference and contains abnormal signals such as "temperature rise" and "position offset"; the fourth enhanced feature vector and the fourth time series fusion feature: the original sequence is input into the submodule, and outputs a 128-dimensional "fourth enhanced feature vector" (normal feature without interference); after deep fusion, the "fourth time series fusion feature" is obtained, which reflects the normal time series law of the transportation link (such as stable temperature and continuous position change).The server compares the third and fourth time series fusion features to calculate the deviation information for the link interference detection task: Feature similarity calculation: Cosine similarity is used to measure the difference between the two fused features (the ideal similarity between normal and interference features should be close to 0, as interference can significantly change the feature distribution). Deviation calculation: If the original submodule is not optimized (initial parameters are random), the cosine similarity between the third and fourth features may be as high as 0.8 (the model does not recognize the interference and the features do not change significantly), and the deviation is 1-0.8=0.2. If the submodule is optimized (parameter adjustment), the abnormal signal of the interference feature is amplified, and the cosine similarity drops to 0.2 (the model sensitively captures the interference), and the deviation is 1-0.2=0.8 (the larger the deviation, the stronger the model's ability to identify interference). Through the link interference detection task, the server simulates data anomalies in the entire link caused by equipment failure or signal interference (such as temperature misreporting and position offset in the transportation link), and optimizes the parameters of the original timing alignment sub-module through the closed loop of interference injection-feature comparison-deviation calculation, making it more sensitive to link interference and improving the robustness and reliability of the entire link perception data.

[0039] In an embodiment of the present invention, the multiple analysis tasks include a link quality self-supervision task; The method of 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 can be implemented through the following example.

[0040] Performing multi-source time series alignment processing on the first traceability instance data to obtain a first sample multi-source time series alignment sequence of the first traceability instance data; Performing time series feature embedding mapping on the first sample multi-source time series alignment sequence through the original time series alignment submodule to obtain a fourth enhanced feature vector of the first sample multi-source time series alignment sequence, performing deep feature fusion processing on the fourth enhanced feature vector to obtain a fourth time series fusion feature of the fourth enhanced feature vector; the fourth time series fusion feature belongs to the time series fusion feature; Based on the detector module corresponding to the link quality self-supervision task, performing abnormal pattern discrimination processing on the fourth time series fusion feature to obtain a dynamic abnormality confidence of the first traceability instance data; Acquire a reference traceability network corresponding to the original time series alignment submodule, perform abnormal pattern discrimination processing on the first traceability instance data based on the reference traceability network, and obtain a reference abnormality confidence of the first traceability instance data; Based on the dynamic anomaly confidence and the baseline anomaly confidence, deviation information of the link quality self-supervision task is determined.

[0041] In this embodiment of the present invention, taking a turtle full-link tracking system from an aquatic product company as an example, the server optimizes the original time series alignment submodule through the "link quality self-supervision task" as follows: the server selects a batch of data from the "first training trajectory set" (1000 batches of unlabeled turtle full-link data) as the "first traceability instance data." This data includes water temperature during the breeding phase (collected hourly from day 0 to day 180), temperature, humidity, and location during transportation (collected minutely from day 181 to day 183), and storage temperature during sales (collected daily from day 184 to day 185). The server calls the raw time series alignment submodule and processes the data in the following steps: identifying the multi-source sensor information for each data point (e.g., "environmental sensors" for the aquaculture stage and "on-board equipment" for the transportation stage); segmenting the raw data into multi-source time series segments (e.g., aquaculture water temperature segments and transportation temperature and humidity segments); aligning timestamps through time series-stage associations (e.g., transportation stage data starts at 00:00 on day 181); and filling missing values ​​(e.g., interpolating when no data occurs at a certain minute in the transportation stage). This ultimately generates the "first sample multi-source time series alignment sequence," a time-continuous, uniformly formatted, full-link data chain (e.g., [Aquaculture - Day 1 08:00-25°C...Transportation - Day 181 00:00-28°C / 60%...Sales - Day 184 00:00-4°C]). The server calls the raw time series alignment submodule and performs feature processing on the first sample multi-source time series alignment sequence: Time series feature embedding mapping: Inputting the aligned sequence into the submodule's time series encoder (e.g., an LSTM network). The encoder extracts temporal features layer by layer: the first layer learns local patterns at adjacent time points (e.g., "aquaculture water temperature stabilizes at ±1°C per hour"), while the second layer captures cross-period trends (e.g., "water temperature rises in the first 30 days of aquaculture and plummets on the 31st day"). The encoder ultimately outputs a 128-dimensional "fourth enhanced feature vector" (each dimension representing key features such as "standard deviation of water temperature fluctuation" and "duration of transport temperature and humidity exceeding standards"). Deep feature fusion: The submodule's deep feature fuser (e.g., an attention network) fuses the fourth enhanced feature vector to generate a "fourth time series fusion feature," which integrates the temporal relationships of data across the entire chain (e.g., the chain effect of "sudden drop in aquaculture water temperature, sensitive fluctuations in transportation temperature and humidity, and abnormal sales and storage temperatures"). The server calls the "detector module" corresponding to the link quality self-supervision task (part of the original submodule, responsible for identifying abnormal patterns) and inputs the fourth time series fusion feature into it. The detector module analyzes abnormal signals in the features (such as a sudden drop in water temperature exceeding three times the historical average or an abnormal frequency of temperature and humidity fluctuations during transportation) and outputs a "dynamic anomaly confidence score" (a probability value between 0 and 1, for example, 0.1 indicates a 10% probability of an anomaly). This confidence score reflects the original submodule's assessment of the anomaly risk of the current full-link data. The server retrieves the "baseline traceability network," a comparison model with a simpler structure and fixed parameters (for example, consisting only of a basic LSTM layer without complex optimization).The first traceability instance data from the same batch is fed into the baseline network. This network processes the data through a similar process (multi-source time series alignment, feature embedding, and simple fusion) and outputs a "baseline anomaly confidence level" (e.g., 0.05 indicates a 5% probability of anomaly). The baseline confidence level represents the unoptimized model's baseline assessment of the risk of data anomaly. The server compares the dynamic anomaly confidence level (0.1) with the baseline anomaly confidence level (0.05) and calculates the absolute difference between the two (0.1 - 0.05 = 0.05), which serves as the "deviation information for the link quality self-supervision task." If the original submodule is unoptimized (with random initial parameters), the dynamic confidence level may be highly consistent with the baseline confidence level (e.g., both are 0.05), resulting in a deviation of 0 (indicating that the model failed to capture potential anomalies in the data). After training and optimization, the dynamic confidence level may rise to 0.2 (the model more sensitively identifies the abnormal signal of "sudden drop in aquaculture water temperature"), and the deviation increases to 0.15 (indicating that the model's ability to discern link quality has improved). Through the link quality self-supervision task, the server uses unlabeled data to evaluate the model performance by "the difference between the original sub-module discrimination results and the baseline model results", and optimizes the parameters of the original timing alignment sub-module, so that it can more accurately capture potential abnormal patterns in the full-link data, providing more reliable abnormality assessment capabilities for the full-link tracking of aquatic products such as soft-shelled turtles.

[0042] 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; the traceability anomaly classification network also includes a link quality analysis network arranged after the deep feature fusion device; The process of performing link anomaly assessment on the fused deep features based on each traceability link to obtain link anomaly confidence of each traceability link can be implemented through the following examples.

[0043] 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 link anomaly assessment is performed on the fused deep features 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.

[0044] 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 detector progressively extracts features through two fully connected layers. The first layer calculates "duration of temperature and humidity exceeding standards (0.6) × weight (1.2) + number of emergency brakes (0.4) × weight (1.0) = 0.6 × 1.2 + 0.4 × 1 = 0.72 + 0.4 = 1.12." The second layer combines "position deviation rate (0.2) × weight (0.8)" to obtain a composite value of 1.12 + 0.2 × 0.8 = 1.28. After ReLU activation (positive values ​​are retained), the output transport anomaly confidence score is 0.3 (30% probability of anomaly). The fused deep features include sub-features such as "average sales and storage temperature" and "duration of unsold items after listing." The sales anomaly detector extracts local features (e.g., "storage temperature has been above 5°C for three consecutive days") through a convolutional layer. Then, through a fully connected layer, it calculates: mean temperature (0.5, representing 5°C) × weight (1.5) + timeout period (0.1) × weight (2.0) = 0.5 × 1.5 + 0.1 × 2 = 0.75 + 0.2 = 0.95; the output sales anomaly confidence score is 0.1 (10% probability of anomaly). Finally, the server uses the anomaly detectors in the link quality analysis network to obtain anomaly confidence scores for the aquaculture (0.85), transportation (0.3), and sales (0.1) links. This result indicates a high anomaly risk in the aquaculture link (possibly due to large water temperature fluctuations and insufficient dissolved oxygen, resulting in weak turtles). Prioritize checking the temperature control and aeration equipment in the aquaculture ponds. The transportation and sales links are less risky, so monitoring frequency can be appropriately reduced. Through this process, the server accurately quantifies the anomaly risk at each link in the entire supply chain, providing a clear basis for link identification and quality control for aquatic products companies.

[0045] In the embodiments of the present invention, the following implementation modes are also provided.

[0046] 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.

[0047] In an embodiment of the present invention, for example, taking a full-link tracking system for soft-shelled turtles of a certain aquatic product company as an example, the process of the server training the abnormality discriminators of each link in the link quality analysis network is as follows: the server extracts a "second training trajectory set" from the company database. This data set contains the full-link RFID perception data of 500 batches of soft-shelled turtles in the past year (such as the water temperature in the breeding link, the temperature and humidity in the transportation link, and the storage temperature in the sales link), and each data is marked with a "link identifier" (manually or automatically marked by the system). For example: the perception data of the breeding link (such as "08:00-25℃ on the first day") is marked with a link code value of "0"; the perception data of the transportation link (such as "08:00-28℃ / 60% on the 181st day") is marked with a link code value of "1"; the perception data of the sales link (such as "09:00-4℃ on the 184th day") is marked with a link code value of "2". The server invokes the "time series alignment submodule" optimized through unsupervised training (e.g., the model optimized using the first training trajectory set) to process each "second traceability instance data" in the second training trajectory set: Feature Embedding: The submodule performs multi-source time series alignment on the data (e.g., aligning scattered data from the breeding, transportation, and sales stages into a time-continuous sequence). This is then mapped into a "fifth enhanced feature vector" (128-dimensional, containing the time series features of each stage) using a time series encoder (e.g., an LSTM network). Deep Feature Fusion: The submodule's deep feature fuser (e.g., an attention network) fuses the fifth enhanced feature vector to generate a "fifth time series fusion feature" (128-dimensional, comprehensively reflecting the correlation features of each stage in the entire chain). The server obtains "multiple initial anomaly discriminators," which are randomly initialized submodels (e.g., fully connected networks) corresponding to the breeding, transportation, and sales stages, respectively. For example: the initial discriminator for aquaculture (structure: 128-dimensional input - 64-dimensional hidden layer - 1-dimensional output); the initial discriminator for transportation (structure: 128-dimensional input - 32-dimensional hidden layer - 1-dimensional output); and the initial discriminator for sales (structure: 128-dimensional input - 16-dimensional hidden layer - 1-dimensional output). The server inputs the fifth time series fusion feature into each initial discriminator to perform "stage identification": the initial discriminator for aquaculture analyzes the "water temperature fluctuation" and "dissolved oxygen content" sub-features within the feature set and outputs a "stage identification confidence level for aquaculture" (e.g., 0.3, indicating a 30% probability of belonging to the aquaculture stage); the initial discriminator for transportation analyzes the "duration of temperature and humidity exceeding the standard" and "position offset" sub-features and outputs a "stage identification confidence level for transportation" (e.g., 0.5, indicating a 50% probability of belonging to the transportation stage); and the initial discriminator for sales analyzes the "average storage temperature" and "shelf time" sub-features and outputs a "stage identification confidence level for sales" (e.g., 0.2, indicating a 20% probability of belonging to the sales stage). The server calculates the "second target deviation information". Taking a second traceability instance data as an example (actually belonging to the transportation link, the link coding value is "1"), the confidence level output by the initial discriminator is 0.3 for breeding, 0.5 for transportation, and 0.2 for sales.The server uses the cross-entropy loss function to calculate the deviation: the true label is [0, 1, 0] (the transport link is the positive class); the prediction 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 this deviation, the server uses the backpropagation algorithm to adjust the parameters of the initial discriminator (such as the weight matrix and bias term of the fully connected layer). For example, the weight of the "duration of temperature and humidity exceeding the standard" feature in the transport discriminator is increased (from 0.2 to 0.5), enhancing its ability to identify the transport link. After multiple rounds of training (e.g., 200 rounds), the deviations of each initial discriminator gradually decrease (for example, the loss of the transport discriminator drops from 0.693 to 0.1), and the model parameters converge. The server identifies the optimized discriminator as the "anomaly discriminator" and integrates it into the "link quality analysis network." For example, the aquaculture anomaly detector accurately identifies aquaculture characteristics such as "water temperature fluctuation > 2°C / hour" and "dissolved oxygen < 5 mg / L," with an output confidence error of <5%. The transportation anomaly detector captures transportation anomalies such as "temperature and humidity exceeding standards for >10 minutes" and "number of sudden brakes >5 times / hour" with a confidence error of <3%. The sales anomaly detector detects sales issues such as "storage temperature > 5°C" and "unsold for >3 days after being put on the shelf," with a confidence error of <2%. Through supervised learning using the second training trajectory set, the server optimized the parameters of the initial anomaly detectors for each link, enabling it to accurately identify the link stage to which the data belongs, ultimately building a reliable link quality analysis network. This network will subsequently be used for full-link anomaly assessment of soft-shelled turtles, outputting accurate anomaly confidence scores for each link to support enterprise quality control decisions.

[0048] In an embodiment of the present invention, the perception feature vector is obtained by performing timing feature embedding mapping on a multi-source timing alignment sequence by a timing encoder in a 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 a timing data input module 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; the traceability anomaly classification network also includes a link influence analysis network arranged after the timing encoder; The link influence analysis processing of the perception feature vector to obtain the link influence coefficient of each traceability link can be implemented through the following example.

[0049] 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.

[0050] In an embodiment of the present invention, illustratively, taking the full-link tracking scenario of soft-shelled turtles in a certain aquatic product enterprise as an example, the process by which the server calculates the impact coefficient of each link through the link impact analysis network in the traceability anomaly classification network is as follows: In the entire link from soft-shelled turtle breeding to sales, RFID perception data is collected by multiple types of equipment: underwater temperature and humidity sensors in the breeding link record the water temperature every hour (such as "08:00-25℃ on the first day" and "09:00-26℃ on the first day"), the on-board RFID equipment in the transportation link collects temperature, humidity and the number of emergency brakes every minute (such as "08:00-28℃ / 60% / 0 emergency brakes on the 181st day" and "08:01-28℃ / 61% / 1 emergency brake on the 181st day"), and the supermarket RFID terminal in the sales link records the storage temperature every day (such as "09:00-4℃ on the 184th day" and "09:00-3℃ on the 185th day"). The server calls the "time series data input module" (which has been optimized through unsupervised training) in the traceability anomaly classification network to process these data: Multi-source time series segment extraction: According to data labels (such as "breeding-water temperature", "transportation-temperature and humidity", and "sales-storage temperature"), the original data is divided into three independent time series segments: breeding, transportation, and sales; Time series segment alignment: Through the predefined "time series-link association space" (breeding 0-180 days, transportation 181-183 days, sales 184-185 days), the timestamps (such as transportation link) are corrected. For example, a data item in the transportation link was mistakenly marked as "23:50 on the 180th day", which was corrected to "00:00 on the 181st day"; missing values ​​were filled in (for example, there was no data at 08:05 on the 181st day in the transportation link, and the value was interpolated to "28℃ / 61% / 0 emergency brakes" through the previous and next values); full-link time series orchestration: the aligned segments were connected in time to generate a "multi-source time series alignment sequence" (for example, [breeding - day 1 08:00-25℃... transportation - day 181 00:00-28℃ / 60%... sales - day 184 00:00-4℃]). The server calls the "Time Series Encoder" (e.g., an LSTM network optimized through unsupervised training) in the Time Series Alignment submodule to process the multi-source time series alignment sequences. Input processing: The aligned sequences (length N, e.g., 180 days of breeding × 24 hours = 4,320 data points, 3 days of transportation × 1,440 minutes = 4,320 data points, 2 days of sales × 1 data point = 2 data points, total length 8,642) are fed into the LSTM input layer. Feature extraction: The LSTM's hidden layers learn temporal dependencies layer by layer. The first layer captures local patterns at adjacent time points (e.g., "breeding water temperature ±1°C per hour"), while the second layer captures cross-link trends (e.g., "water temperature drops sharply in the late breeding period, temperature and humidity fluctuate in the early transportation period"). The final output is a 128-dimensional "perception feature vector." Each dimension of this vector corresponds to a key feature (e.g., "standard deviation of breeding water temperature fluctuation," "mean number of sudden stops during transportation," "mean sales and storage temperature").The server locates the "Link Impact Analysis Network" (located after the time series encoder and optimized through supervised training) from the source-tracing anomaly classification network and inputs the perception feature vector into it. This network consists of an attention mechanism layer and a fully connected layer. Its core function is to calculate the contribution of each link to the overall chain anomaly. The attention layer traverses the 128 dimensions of the perception feature vector and calculates the importance weight of each feature to the overall chain anomaly. For example, the "Standard Deviation of Aquaculture Water Temperature Fluctuation" (dimension 10) has a weight of 0.3 (greater fluctuations indicate a higher risk of anomalies in subsequent links); the "Average Number of Sudden Stops During Transportation" (dimension 25) has a weight of 0.2 (frequent sudden stops can cause stress to turtles); and the "Average Sales and Storage Temperature" (dimension 50) has a weight of 0.1 (higher temperatures indicate a greater risk of spoilage). The server aggregates feature weights by link using a fully connected layer: The impact coefficient for the farming link is calculated by adding the weights of all farming-related features (e.g., 0.3 for dimension 10 and 0.2 for "dissolved oxygen levels below threshold" in dimension 15), totaling 0.5. The impact coefficient for the transportation link is calculated by adding the weights of transportation-related features (e.g., 0.2 for dimension 25 and 0.2 for "duration of temperature and humidity exceeding standards" in dimension 30), totaling 0.4. The impact coefficient for the sales link is calculated by adding the weights of sales-related features (e.g., 0.1 for dimension 50 and 0.05 for "duration of unsold items after being put on the shelf" in dimension 55), totaling 0.15. Finally, the server outputs the impact coefficients for each link: farming 0.5, transportation 0.4, and sales 0.15. This result shows that anomalies in the farming link have the greatest impact on all chain anomalies (50%), followed by transportation (40%), and the sales link has the least impact (15%). Based on this information, companies can prioritize strengthening monitoring in the aquaculture process (such as optimizing water temperature control equipment) and pay appropriate attention to emergency braking and temperature and humidity control during transportation to reduce the risk of abnormalities throughout the entire supply chain. Through this process, the server achieves a quantitative assessment of the impact of each link, providing data support for precise management and control of the entire aquaculture supply chain.

[0051] In an embodiment of the present invention, the traceability anomaly classification network further includes a link quality analysis network for determining the link anomaly confidence level of each traceability link; The embodiments of the present invention also provide the following implementation methods.

[0052] 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.

[0053] In an embodiment of the present invention, illustratively, taking the full-link tracking system for soft-shelled turtles of a certain aquatic enterprise as an example, the process of the server optimizing the link impact analysis network through the third training trajectory set is as follows: the server extracts the "third training trajectory set" from the enterprise quality database. This data set contains the full-link RFID perception data of 500 batches of soft-shelled turtles in the past two years (such as aquaculture water temperature, transportation temperature and humidity, sales and storage temperature), and each data is marked with a "full-link abnormality flag" (determined by the quality inspection results: 0 means no abnormality, 1 means that quality problems such as corruption and death will eventually occur). For example: Batch A: The water temperature in the aquaculture link fluctuates greatly (abnormal), the temperature and humidity in the transportation link are stable (normal), the storage temperature in the sales link exceeds the standard (abnormal), and the final quality inspection shows corruption (full-link abnormality flag = 1); Batch B: The aquaculture, transportation, and sales links are all normal (full-link abnormality flag = 0). The server invokes the "time series alignment submodule" optimized through unsupervised training (e.g., the model optimized in steps 5-9) to process each "third traceability instance data" in the third training trajectory set: Feature Embedding: The submodule performs multi-source time series alignment on the data (e.g., aligning the scattered data from the breeding, transportation, and sales stages into a time-continuous sequence). This is then mapped into a "sixth enhanced feature vector" (128-dimensional, containing the time series features of each stage) using a time series encoder (e.g., an LSTM network). Deep Feature Fusion: The submodule's deep feature fuser (e.g., an attention network) fuses the sixth enhanced feature vector to generate a "sixth time series fusion feature" (128-dimensional, comprehensively reflecting the correlation features of each stage in the entire chain). The server invokes the "link quality analysis network" optimized through supervised training (e.g., the model trained in step 11) and inputs the sixth time series fusion feature into each anomaly discriminator in the network (breeding, transportation, and sales discriminators). Each discriminator outputs a "confidence score for an abnormality in the training phase" (a probability value between 0 and 1). In the sixth time series fusion feature of batch A, the aquaculture discriminator detects "water temperature fluctuation standard deviation > 3°C" (an abnormal feature) and outputs a confidence score of 0.8 for the aquaculture phase. The transportation discriminator detects "temperature and humidity exceeded the standard for < 10 minutes" (a normal feature) and outputs a confidence score of 0.2 for the transportation phase. The sales discriminator detects "storage temperature > 5°C for three consecutive days" (an abnormal feature) and outputs a confidence score of 0.6 for the sales phase. The server calls the "original phase influence analysis network" (an unoptimized initial model with randomly initialized parameters) and inputs the sixth enhanced feature vector into it. The network calculates the "training phase influence coefficient" (the weight of each phase's contribution to the overall process anomaly) through a fully connected layer and an attention mechanism. In the sixth enhanced feature vector of batch A, the weight of the "aquaculture water temperature fluctuation" feature is 0.3, the weight of the "sales and storage temperature" feature is 0.2, and the weight of the "transportation temperature and humidity" feature is 0.5. After phase aggregation, the training phase influence coefficients are output as follows: 0.3 for aquaculture, 0.5 for transportation, and 0.2 for sales.Based on the training phase anomaly confidence levels for batch A (farming 0.8, transportation 0.2, sales 0.6) and the training phase influence coefficients (farming 0.3, transportation 0.5, sales 0.2), the server calculated the "training full-link anomaly probability": training full-link anomaly probability = (0.8 × 0.3) + (0.2 × 0.5) + (0.6 × 0.2) = 0.24 + 0.1 + 0.12 = 0.46. The actual full-link anomaly flag for this batch was 1 (abnormal). The server calculated 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 adjusted the parameters of the original phase influence analysis network (such as the weight matrix of the attention layer and the bias term of the fully connected layer) using the backpropagation algorithm. For example, the weight of the "aquaculture water temperature fluctuation" feature was increased (from 0.3 to 0.5) because aquaculture anomalies have a greater impact on overall chain anomalies. The weight of the "transportation temperature and humidity" feature was decreased (from 0.5 to 0.3) because transportation anomalies have a smaller impact on the batch. After multiple rounds of training (e.g., 150 rounds), the deviation of the original link impact analysis network dropped below 0.1 (the error between the trained overall chain anomaly probability and the actual identification was <10%). The server designated this optimized network as the "link impact analysis network" and integrated it with the trained time series alignment submodule and link quality analysis network to form the final "source tracing anomaly classification network." This network accurately calculates the impact coefficient of each link and, combined with the link anomaly confidence level, outputs an overall chain anomaly probability. For example, the predicted probability for batch A increased from 0.46 to 0.85, with a deviation of only 0.02 from the actual identification of 1. Through supervised learning on the third training trajectory set, the server optimized the parameters of the link impact analysis network, enabling it to more accurately quantify the contribution of each link to overall chain anomalies. The traceability anomaly classification network finally constructed can realize end-to-end analysis from multi-source perception data input to full-link anomaly probability output, providing reliable technical support for quality control of aquatic products enterprises.

[0054] In an embodiment of the present invention, the link influence analysis network includes a link interaction network and an attention network; The link influence analysis processing of the perception feature vector by the link influence analysis network to obtain the link influence coefficient of each traceability link can be implemented through the following example.

[0055] Performing multi-source dynamic feature aggregation on the perception feature vector through the link interaction network to obtain an interaction feature vector of the perception feature vector; Attention is allocated to the interaction feature vector through the attention network to obtain the link influence coefficient of each traceability link.

[0056] In an embodiment of the present invention, taking the full-link tracking scenario of soft-shelled turtles at an aquatic product company as an example, the server calculates the influence coefficient of each link using a link influence analysis network (including a link interaction network and an attention network) as follows: After completing multi-source time series alignment (generating a multi-source time series alignment sequence) and time series feature embedding mapping (outputting a perception feature vector via a time series encoder) of the soft-shelled turtle RFID full-link perception data, the server generates a 128-dimensional "perception feature vector." Each dimension of this vector corresponds to a key feature of each link in the full link, for example: dimensions 1-30: aquaculture link characteristics (such as "water temperature fluctuation standard deviation," "percentage of dissolved oxygen content below threshold," and "feeding frequency"); dimensions 31-60: transportation link characteristics (such as "duration of temperature and humidity exceeding standards," "average number of sudden stops," and "position deviation rate"); dimensions 61-90: sales link characteristics (such as "average storage temperature," "duration of unsold items after being put on the shelf," and "frequency of code scanning queries"); and dimensions 91-128: cross-link interaction characteristics (such as "correlation between sudden drops in aquaculture water temperature and fluctuations in transportation temperature and humidity"). The server invokes the "link interaction network" (composed of two fully connected layers and a gating mechanism) in the link impact analysis network to perform "multi-source dynamic feature aggregation" on the perception feature vector. The specific operation is as follows: the server inputs the 128-dimensional perception feature vector into the first fully connected layer (128-256 dimensions). This layer uses linear transformation and ReLU activation function to extract preliminary interactions between features. For example, the "standard deviation of water temperature fluctuation" (dimension 1) in the aquaculture stage and the "duration of temperature and humidity exceeding standards" (dimension 31) in the transportation stage are calculated as a new feature, "water temperature fluctuation-transportation temperature and humidity sensitivity coefficient" (dimension 150); the "average number of emergency braking" (dimension 32) in the transportation stage and the "average storage temperature" (dimension 61) in the sales stage are calculated as the "emergency braking stress-sales corruption risk coefficient" (dimension 200). The server uses a gating mechanism (such as the update gate of a GRU) to filter the 256-dimensional preliminary interaction features, retaining those with the greatest impact on abnormalities across the entire chain. For example, if the weight of "Water Temperature Fluctuation - Transportation Temperature and Humidity Sensitivity Coefficient" (dimension 150) is 0.8 (indicating that aquaculture water temperature fluctuation significantly affects transportation temperature and humidity stability), then this feature is retained. If the weight of "Sudden Braking Stress - Sales Corruption Risk Coefficient" (dimension 200) is 0.2 (indicating that sudden braking has little impact on sales corruption), then its weight is reduced or discarded. The filtered features are input into the second fully connected layer (256-128 dimensions), where a linear transformation and Tanh activation function are applied to generate an "interaction feature vector." This vector focuses on feature interactions that significantly influence anomalies across the entire chain (e.g., "Aquaculture Water Temperature Fluctuation - Transportation Temperature and Humidity Sensitivity" and "Sales Storage Temperature - Accelerated Corruption").The server calls the "attention network" (composed of a query vector and a Softmax function) in the impact analysis network. It distributes attention to the 128-dimensional interaction feature vector and calculates the influence coefficient of each link. The attention network calculates the attention score for each feature dimension by taking the dot product of the learnable query vector (128 dimensions) and the interaction feature vector. For example, the "aquaculture water temperature fluctuation - transport temperature and humidity sensitivity" feature (dimension 150 mapped to dimension 50 of the interaction feature vector) scores 0.7 (indicating that this interaction has a significant impact on overall link anomalies); the "sales storage temperature - accelerated spoilage" feature (dimension 61 mapped to dimension 80 of the interaction feature vector) scores 0.5 (indicating that this interaction has a secondary impact). The server performs Softmax normalization on the attention scores, converting them to weights ranging from 0 to 1 (summing to 1). For example, the total weight of aquaculture-related interaction features is 0.6 (including features such as "water temperature fluctuation - transportation sensitivity"); the total weight of transportation-related interaction features is 0.3 (including features such as "sudden braking stress - sales corruption"); and the total weight of sales-related interaction features is 0.1 (including features such as "storage temperature - accelerated corruption"). The server uses the attention weight of each link as the "link influence coefficient." For example, the influence coefficient of the aquaculture link is 0.6 (indicating that aquaculture-related anomalies contribute 60% to overall chain anomalies); the influence coefficient of the transportation link is 0.3 (transportation-related anomalies contribute 30%); and the influence coefficient of the sales link is 0.1 (sales-related anomalies contribute 10%). Ultimately, the server outputs the influence coefficient of each link (e.g., 0.6 for aquaculture, 0.3 for transportation, and 0.1 for sales) through the link interaction network and attention network. This result indicates that aquaculture-related anomalies (such as large water temperature fluctuations) have the greatest impact on overall chain anomalies. Based on this, companies can prioritize optimizing environmental controls in the aquaculture link (e.g., upgrading temperature control equipment) to reduce overall quality risks. Through this process, the server can accurately quantify the impact of each link in the entire chain, providing a scientific basis for abnormal control of the entire aquatic product chain.

[0057] In an embodiment of the present invention, 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.

[0058] In an embodiment of the present invention, using a turtle full-link tracking system at an aquatic product company as an example, the server optimizes the traceability anomaly classification network for the "transportation link" (the target traceability link) as follows: The server first loads a trained "traceability anomaly classification network" from its local model library. This network has undergone multiple rounds of training (e.g., steps 5-14) and can assess anomaly risks across the entire chain of breeding, transportation, and sales. Simultaneously, the server obtains an "adaptive base model," a model with the same structure as the original network (including a time alignment submodule, a deep feature fusion module, a link quality analysis network, and a link impact analysis network), but with randomly initialized parameters. This model is used for transfer learning to optimize the performance of the target link. The server extracts a "fourth training trajectory set" from the company's specialized database for the transportation link. This dataset contains RFID-based transport data (such as temperature and humidity, number of sudden stops, and position deviation rate) for 300 batches of turtles over the past year, along with corresponding full-link anomaly results (such as stress mortality and spoilage due to improper transportation). For example: Batch C: The temperature and humidity in the transportation process exceeded the standard for 2 hours (30℃ / 70%), emergency brakes occurred 5 times / hour, and the final quality inspection showed corruption (abnormality in the entire chain); Batch D: The temperature and humidity in the transportation process were stable (28℃ / 60%), emergency brakes occurred 0 times / hour, and the final quality inspection showed normal (no abnormality). The server calls the original traceability anomaly classification network to perform "trajectory feature analysis" on each "fourth traceability instance data" in the fourth training trajectory set (such as the transportation link data of batches C and D): Multi-source time series alignment: Align the transportation link data with the breeding and sales link data (for example, the breeding data ends on the 180th day and the transportation data starts on the 181st day); Feature embedding and fusion: Generate perception feature vectors through the time series encoder, and generate fused deep features through the deep feature fusion device; Link anomaly assessment and impact analysis: The link quality analysis network outputs the transportation link anomaly confidence (for example, batch C is 0.9, batch D is 0.1), and the link influence analysis network outputs the transportation link impact coefficient (for example, batch C is 0.6, batch D is 0.2); Full-link anomaly probability calculation: Combine the anomaly confidence and impact coefficient to output the full-link anomaly probability (for example, batch C is 0.9×0.6=0.54, batch D is 0.1×0.2=0.02). The server determines these probability values ​​as "link abnormality guidance signals" (such as the guidance signal for batch C is 0.54, and for batch D is 0.02) as supervision targets for training the adaptive base model.The server inputs the fourth traceability instance data of the same batch into the adaptive base model (with random parameters) and performs the same trajectory feature analysis process: Because the parameters of the adaptive base model are not optimized, its timing alignment submodule may not be able to accurately align the transportation link data (for example, "08:00 on the 181st day" is mistakenly aligned to "23:00 on the 180th day"); the deep feature fusion device may ignore the correlation features between "temperature and humidity exceeding the standard" and "number of sudden brakes" (for example, it fails to capture the pattern of "sudden brakes causing temperature and humidity fluctuations"); the transportation link anomaly 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 that of batch D is 0.4); the transportation link influence coefficient output by the link influence analysis network may be incorrect (for example, 0.2 for batch C and 0.7 for batch D); the final calculated adaptive full-link anomaly probability is significantly different from the guidance signal (for example, 0.3×0.2=0.06 for batch C and 0.4×0.7=0.28 for batch D). The server calculates the mean square error between the "adaptive full-link anomaly probability" and the "link anomaly guidance signal" (for example, the deviation of batch C = (0.54-0.06)²=0.2304, and the deviation of batch D = (0.02-0.28)²=0.0676), and adjusts the parameters of the adaptive base model through the backpropagation algorithm: optimizing the time calibration parameters of the timing alignment submodule (for example, correcting the time offset of the transportation link to "00:00 on the 181st day"); adjusting the attention weight of the deep feature fusion device (for example, increasing the association weight between "duration of temperature and humidity exceeding the standard" and "number of emergency brakes"); optimizing the fully connected layer parameters of the transportation anomaly discriminator in the link quality analysis network (for example, increasing the weight of "duration of temperature and humidity exceeding the standard" from 0.3 to 0.7); and adjusting the attention allocation of the link influence analysis network (for example, increasing the influence coefficient weight of the transportation link from 0.2 to 0.6). After 50 rounds of training, the deviation of the adapted base model dropped below 0.05 (for example, the adaptive probability of Batch C increased from 0.06 to 0.52, with a deviation of only 0.0004 from the guidance signal of 0.54). The server designated the optimized adapted base model as the "Optimized Traceability Model," and its accuracy in anomaly classification during transportation increased from the initial 50% to 90%. The optimized traceability model specifically enhances anomaly classification during transportation (the target traceability link), enabling more accurate identification of anomalies such as excessive temperature and humidity, and frequent sudden braking during transportation. (For example, the anomaly probability for Batch C increased from 0.54 in the original network to 0.92, more closely resembling actual quality inspection results). Enterprises can use this model to focus on monitoring key indicators during transportation and reduce quality losses caused by improper transportation. Through this process, the server achieves adaptive migration optimization of the target traceability link, improving the targeted and accurate anomaly classification throughout the entire process.

[0059] In an embodiment of the present invention, the fourth traceability instance data includes link operation instance data generated under the target traceability link; The obtaining of the fourth training trajectory set corresponding to the target tracing link may be implemented through the following example.

[0060] Obtaining the link operation instance data under the target traceability link; Determining the traceability instance data to be identified from the traceability instance data set used for trajectory data screening; Performing link feature matching analysis on the to-be-identified traceability instance data and the link operation instance data through a feature filter to obtain a link matching index between the link operation instance data and the to-be-identified traceability instance data; If the link matching index satisfies the trajectory enhancement criterion, the to-be-identified traceability instance data is determined as the target traceability instance data corresponding to the target traceability link; The target traceability instance data and the link operation instance data are determined as fourth traceability instance data corresponding to the target traceability link.

[0061] In an embodiment of the present invention, taking a turtle full-link tracking system from an aquatic products company as an example, the server obtains the fourth training trajectory set to optimize the anomaly classification model for the "transportation link" (the target traceability link) as follows: The server first extracts "link operation instance data" from the company's dedicated database for the transportation link. This data is RFID sensing records of actual transportation links, including key indicators such as temperature and humidity, number of emergency stops, and position deviation rate. For example: Data 1: Transportation time "Day 181 08:00 - Day 183 24:00", temperature and humidity "28°C / 60% (stable)", number of emergency stops "0 times / hour", and position deviation rate "0.1%" (normal transportation); Data 2: Transportation time "Day 182 09:00 - Day 182 10:00", temperature and humidity "35°C / 75% (exceeding the standard)", number of emergency stops "5 times / hour", and position deviation rate "5%" (abnormal transportation). The server accesses the enterprise's full-chain traceability database (containing historical data from each stage of the breeding, transportation, and sales process) and extracts "traceability instance data to be identified," including all batch records containing transportation data. For example, Batch X: Normal breeding (water temperature 25±1°C), transportation data from "Day 181 08:00-Day 183 24:00" (to be identified), and normal sales (storage temperature 4°C); Batch Y: Abnormal breeding (water temperature drops 5°C), transportation data from "Day 181 08:00-Day 183 24:00" (to be identified), and abnormal sales (storage temperature 8°C). The server invokes a "feature filter" (a pre-trained feature matching model) to perform a "link feature matching analysis" between each piece of data to be identified and the operational instance data. The core of the feature filter is to extract key features of the transportation process (such as mean temperature and humidity, standard deviation of emergency braking times, and maximum position deviation rate) and calculate the similarity between the two. Taking the data to be identified in batch X as an example: the feature vector of the link operation instance data: mean temperature and humidity (28°C / 60%), standard deviation of emergency braking (0 times), maximum position offset rate (0.1%) - vector V1 = [28, 60, 0, 0.1]; the feature vector of the data to be identified: mean temperature and humidity (27°C / 58%), standard deviation of emergency braking (1 time), maximum position offset rate (0.2%) - vector V2 = [27, 58, 1, 0.2]; Matching degree calculation: Using the cosine similarity formula, 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 a "trajectory enhancement criterion" with a link matching index of ≥ 0.8 (indicating high feature similarity). If the matching degree of the data to be identified meets this criterion, it is determined to be the "target traceability instance data."For example, Batch X has a matching degree of 0.98 (≥0.8), meeting the criteria and being selected as the target traceability instance data. However, the target data for Batch Y, whose average temperature and humidity during the transportation phase are 32°C / 70% (an abnormal characteristic), has a similarity of only 0.6 (<0.8) with the operational data during this phase, failing to meet the criteria and being discarded. The server then merges the selected target traceability instance data (e.g., Batch X) with the original operational data (e.g., Data 1 and Data 2) to form the "fourth training trajectory set." This dataset includes: data from normal transportation scenarios (e.g., Data 1 and Batch X); data from abnormal transportation scenarios (e.g., Data 2); and other highly matching transport data (e.g., target data with a matching degree ≥0.8). Ultimately, the fourth training trajectory set focuses on typical scenarios (both normal and abnormal) during the transportation phase, and its data features are highly consistent with actual operational conditions. The server will subsequently use this dataset to train an adaptive base model to optimize the anomaly classification performance for the transportation phase (e.g., improving the accuracy of identifying anomalies such as excessive temperature and humidity and frequent sudden braking), providing data support for enterprises to precisely manage transportation risks. Through this process, the server achieves accurate screening and enhancement of the training data for the target traceability link, ensuring the effectiveness and pertinence of subsequent model migration training.

[0062] In an embodiment of the present invention, the determination of the full-link abnormality probability of the tracked item based on the link abnormality confidence of each traceability link and the link influence coefficient of each traceability link can be implemented through the following example.

[0063] 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.

[0064] In an embodiment of the present invention, taking the full-chain tracking scenario of soft-shelled turtles at an aquatic product company as an example, the server calculates the probability of an anomaly in the entire chain based on the anomaly confidence and impact coefficient of each link as follows: A batch of soft-shelled turtles has completed the entire chain (breeding, transportation, and sales). The server has obtained key indicators for each link through the traceability anomaly classification network: link anomaly confidence (output by the link quality analysis network): breeding link 0.8 (80% probability of anomaly due to large water temperature fluctuations and insufficient dissolved oxygen), transportation link 0.3 (30% probability of anomaly due to temporary temperature and humidity exceeding standards), sales link 0.1 (10% probability of anomaly due to slightly elevated storage temperature); link impact coefficient (output by the link impact analysis network): breeding link 0.6 (contributing 60% to the full chain anomaly), transportation link 0.3 (contributing 30%), and sales link 0.1 (contributing 10%). The server needs to combine the anomaly confidence and impact coefficient of each link through "dynamic aggregation" to calculate the "fused anomaly assessment value." The "dynamic" factor here is reflected in the fact that the impact coefficient is not a fixed weight. Instead, it is dynamically calculated based on the perceived characteristics of the current batch (such as the fluctuation of aquaculture water temperature and the number of sudden stops during transportation), ensuring that the weight is correlated with the actual risk. The integrated anomaly assessment value for the aquaculture stage is: aquaculture anomaly confidence (0.8) × aquaculture impact coefficient (0.6) = 0.8 × 0.6 = 0.48; the integrated anomaly assessment value for the transportation stage is: transportation anomaly confidence (0.3) × transportation impact coefficient (0.3) = 0.3 × 0.3 = 0.09; and the integrated anomaly assessment value for the sales stage is: sales anomaly confidence (0.1) × sales impact coefficient (0.1) = 0.1 × 0.1 = 0.01. The server sums the integrated anomaly assessment values ​​for each stage to obtain the "full-chain anomaly probability" for the batch of soft-shelled turtles: Full-chain anomaly probability = aquaculture integrated value + transportation integrated value + sales integrated value = 0.48 + 0.09 + 0.01 = 0.58 (i.e., 58%). This result (58% anomaly probability) indicates that this batch of soft-shelled turtles presents a high risk of anomalies across the entire supply chain. Further analysis of the contributions of each link reveals that the aquaculture link (0.48) accounts for 82.8% of the total risk (0.48 / 0.58), making it the primary source of risk. Prioritizing inspection of the temperature control and aeration equipment in the aquaculture ponds is crucial. The contributions of the transportation link (0.09) and the sales link (0.01) are relatively low, indicating that the risk is primarily concentrated in the aquaculture phase, and anomalies in these two links have a minimal impact on the final results. The key to "dynamic aggregation" is that the impact coefficient adjusts as data characteristics change. For example, if another batch of soft-shelled turtles experiences a significant increase in the number of sudden stops during transportation (potentially increasing the impact coefficient to 0.5), even if its anomaly confidence level is only 0.2, the transportation integration value will increase to 0.2 × 0.5 = 0.1, becoming a significant factor in the overall anomaly. This dynamic adjustment ensures a strong correlation between the assessment results and actual risk, avoiding bias caused by fixed weights.By dynamically aggregating the anomaly confidence and impact coefficients of each link, the server ultimately outputs the probability of anomalies across the entire chain (e.g., 58% in this example), providing businesses with a quantitative risk assessment. Based on this, businesses can optimize high-risk links (e.g., the farming process in this example), reduce quality losses across the entire chain, and achieve precise management and control.

[0065] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned tracking and tracing method based on RFID Internet of Things technology. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0066] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application 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; Based on the link abnormality confidence of each traceability link and the link influence coefficient of each traceability link, the full-link abnormality probability of the tracked item is determined.

2. The method according to claim 1, characterized in that 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.

3. The method according to claim 2, 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.

4. The method according to claim 1, wherein 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 time sequence alignment submodule; the time sequence alignment submodule belongs to the traceability anomaly classification network for performing full-link traceability anomaly classification on the tracked item; 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.

5. The method according to claim 1, wherein 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 for performing full-link traceability anomaly classification on the tracked items; 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.

6. 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.

7. The method according to claim 6, 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.

8. 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.

9. 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.

10. 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 9.

Images