Temperature monitoring method and system for seafood frozen food

Abstract

The invention relates to the technical field of cold-chain logistics monitoring, and discloses a temperature monitoring method and system for frozen seafood. The method comprises the following steps: acquiring an original temperature sequence of multiple monitoring points in a cold chain space, and processing the original temperature sequence to form a standard temperature sequence; identifying a temperature mode unit, and constructing a directed temperature network for describing the dynamic evolution of the temperature field according to the space-time relationship of the temperature mode unit; dividing different temperature feature regions based on a network, and converting network structure features of the different temperature feature regions into multi-dimensional feature vectors; obtaining a feature coding vector by using a depth auto-encoder, and calculating a predicted coding vector of a next time period through a prediction model in combination with environmental parameters and historical data; generating temperature evolution trend information by comparing the difference between the predicted vector and the current vector; and adaptively adjusting the acquisition frequency of the sensor and a network structure analysis strategy according to the trend information. According to the invention, deep analysis of dynamic correlation of the cold chain temperature field and prospective active monitoring based on prediction are realized.

Description

Technical Field

[0001] This invention relates to the field of cold chain logistics monitoring technology, specifically to a method and system for temperature monitoring of frozen seafood. Background Technology

[0002] Maintaining a stable low-temperature environment is crucial for frozen seafood during cold chain storage and transportation. Currently, commonly used temperature monitoring technologies rely primarily on multiple temperature sensors deployed within storage or transportation spaces. These sensors periodically measure temperature at fixed locations, and the monitoring system processes the collected discrete data points independently. Common processing methods include comparing real-time data from each monitoring point with preset safety thresholds to trigger alarms, or performing basic aggregated statistics such as averaging and extreme values ​​on the data from all points. This approach treats the temperature field of the entire space as a collection of isolated monitoring points.

[0003] Existing solutions based on fixed-point threshold judgments and simple statistics have the drawback of failing to characterize the dynamic structure and interrelationships within the temperature field. The system can only determine whether a specific point exceeds limits, but it cannot understand the actual propagation paths of heat and cold within space, the degree of temperature interaction between different regions, or the overall trend of temperature uniformity. Existing monitoring strategies are statically configured, with sensor sampling frequencies typically fixed and unable to be flexibly adjusted according to the actual state and changing trends of the temperature field. This results in a lack of foresight in monitoring, making it difficult to provide early warnings and pre-allocate resources before temperature anomalies occur.

[0004] There is a need for an intelligent method that can analyze the dynamic correlation structure inside the temperature field and autonomously adjust the monitoring strategy based on the prediction of the future state of the temperature field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for temperature monitoring of frozen seafood, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for temperature monitoring of frozen seafood, the method comprising: The original temperature sequence of seafood products in the cold chain space is collected. The original temperature sequence is obtained by periodic measurement by sensors distributed at different monitoring locations in the cold chain space. The original temperature sequence is subjected to outlier removal and missing value repair to form a standard temperature sequence; Identify temperature pattern units with periodic or trend-like changes from the standard temperature sequence; Based on the spatial proximity and temporal continuity of the temperature pattern units, a directed temperature network describing the dynamic evolution of the temperature field is constructed. Based on the temperature flow direction and intensity between nodes in the directed temperature network, different temperature characteristic regions are divided. The network structure features of different temperature regions are converted into multi-dimensional feature vectors; The multidimensional feature vector is compressed and reconstructed using a deep autoencoder network to obtain the feature encoding vector; By combining environmental parameters with historical feature encoding vectors, the predicted encoding vector of the temperature field for the next time period is calculated through a prediction model. The difference between the predicted encoding vector and the current feature encoding vector is compared to generate temperature evolution trend information; Based on temperature evolution trend information, an adaptive monitoring strategy is driven to adjust the sensor acquisition frequency and network structure.

[0007] Preferably, the step of removing outliers and repairing missing values ​​in the original temperature sequence to form a standard temperature sequence includes: A moving average is calculated based on the data from each monitoring location in the original temperature sequence; Temperature points that deviate from the moving average line by more than a preset deviation threshold are marked as candidate outliers; The candidate anomalies were verified by combining temperature readings from nearby monitoring locations at the same time point; Verified outliers are removed from the sequence, creating missing data points at the removal locations; A weighted average method with time proximity and spatial proximity as weights is used to interpolate and repair the missing data points; The repaired sequence is smoothed and filtered to obtain a continuous temperature sequence with noise suppression, which is then used as the standard temperature sequence.

[0008] Preferably, identifying temperature pattern units with periodic or trend-like changes from the standard temperature sequence includes: Slide a time window across the standard temperature sequence and extract the temperature subsequence within the window; Autocorrelation analysis was performed on each temperature subsequence to detect the presence of periodic temperature fluctuations; For temperature subsequences that do not exhibit significant periodicity, trend fitting is performed to calculate the slope of their temperature changes; Merge adjacent time windows with similar periods or similar slopes; A temperature pattern is defined for each merged time period, and the temperature pattern is described by the pattern type, duration, average rate of change and amplitude range. All identified patterns are arranged in order of their start time to form a temperature pattern unit sequence.

[0009] Preferably, the step of constructing a directed temperature network describing the dynamic evolution of the temperature field based on the spatial proximity and temporal continuity of the temperature model units includes: The status of each monitoring location during the duration of each temperature mode unit is defined as a network node. If the monitoring locations corresponding to two nodes are spatially adjacent, then a spatial connection edge is established between the two nodes; If two nodes correspond to two consecutive mode units at the same monitoring location in time, then a time connection edge is established between the two nodes. Each spatial connection edge is assigned a weight, which is calculated based on the average temperature difference between the mode cells of the two nodes. Each temporal connection edge is assigned a weight, which is calculated based on the temporal probability of mode transition from the previous node to the next node. All nodes, spatial connections, temporal connections, and their weights are combined to form a weighted directed graph structure, which serves as the directed temperature network describing the dynamic evolution of the temperature field.

[0010] Preferably, the step of dividing different temperature characteristic regions based on the temperature flow direction and intensity between nodes in the directed temperature network includes: Calculate the difference between the sum of outgoing edge strength and the sum of incoming edge strength of each node in the directed temperature network to obtain the net temperature flow intensity of each node; Nodes with high net temperature flow intensity are used as seed points; Starting from the seed point, perform region growth along the directed edge, and include adjacent nodes that have the same temperature flow direction as the seed point and whose connection strength exceeds the connection strength threshold into the same region. When no new nodes can be incorporated, an initial temperature characteristic region is formed; For nodes not included in any initial region, calculate their average temperature similarity with the core of each initial region and assign them to the region with the highest similarity. Boundary optimization is performed on all regions, merging regions with too small an area or too few nodes, ultimately forming stable regions with different temperature characteristics.

[0011] Preferably, the step of converting the network structure features of different temperature characteristic regions into multi-dimensional feature vectors includes: For each temperature characteristic region, count the total number of nodes it contains; Calculate the average weight of the connection edges between all nodes within the temperature characteristic region, and use it as the average connection strength of the region. The ratio of the total weight of outgoing edges from nodes in the temperature characteristic region to nodes outside the region to the total weight of incoming edges from nodes outside the region to nodes in the temperature characteristic region is calculated as the net outflow ratio of the region temperature. Extract the distribution histogram of temperature pattern unit types in the temperature feature region over the most recent complete time period; The distribution histograms of the total number of nodes, the average regional connection strength, the regional net temperature outflow ratio, and the temperature pattern unit type are normalized and spliced ​​to form a multi-dimensional feature vector representing the network structure characteristics of the temperature feature region.

[0012] Preferably, the step of compressing and reconstructing the multidimensional feature vector using a deep autoencoder network to obtain a feature encoding vector includes: A deep autoencoder network containing a symmetric encoder and a decoder is constructed. The encoder consists of multiple fully connected layers, each followed by a nonlinear activation function. The multidimensional feature vector is input into the encoder, and after layer-by-layer nonlinear transformation, a low-dimensional compressed code is generated at the bottleneck layer. The compressed code is input into the decoder, and after layer-by-layer nonlinear transformation, an output vector with the same dimension as the input is reconstructed. The deep autoencoder network is trained by minimizing the mean square error between the multidimensional feature vector and the reconstructed output vector. After training is completed, the multidimensional feature vector is input into the encoder, and the compressed code output by the encoder at the bottleneck layer is the feature encoding vector.

[0013] Preferably, the step of combining environmental parameters and historical feature encoding vectors to calculate the predicted encoding vector of the temperature field for the next time period through a prediction model includes: Collect environmental parameters that are recorded synchronously with the standard temperature sequence, including external ambient temperature, cold storage equipment operating power, and door opening / closing event records; Environmental parameters are standardized to form an environmental parameter vector; The feature encoding vector at the current moment is stacked with the feature encoding vectors from multiple consecutive historical moments in chronological order to form a historical feature encoding sequence. The environmental parameter vector and the historical feature encoding sequence are fused along the feature dimension. The fused features are then input into a prediction model built on a long short-term memory network. The prediction model outputs a vector with the same dimension as the feature encoding vector, which serves as the prediction encoding vector for the temperature field in the next time period.

[0014] Preferably, the step of comparing the difference between the predicted encoding vector and the current feature encoding vector to generate temperature evolution trend information includes: Calculate the difference between the predicted encoding vector and the current feature encoding vector in each corresponding dimension; The differences across each dimension are summed in a weighted manner, with the weights pre-set based on the contribution of each dimension to the overall stability of the temperature field. The weighted summation result is compared with a preset trend intensity threshold. If it exceeds the trend intensity threshold, it is determined that there is a significant temperature evolution trend. Based on the indices of the dimensions with the largest differences, the corresponding network structure features are mapped back to determine whether the main temperature changes occur in a specific region, connection strength, or pattern distribution. By combining the trend determination results and change analysis results, a structured text description is generated as the temperature evolution trend information driving the adaptive monitoring strategy.

[0015] Preferably, the present invention also includes a temperature monitoring system for frozen seafood, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the temperature monitoring method for frozen seafood as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By constructing a directed network from spatiotemporally correlated temperature pattern units, the dynamic topology of the temperature field in the cold chain space was modeled. The nodes and directed edges in the network can quantify the intensity and direction of temperature influence between different regions, thus revealing global correlation characteristics that static point monitoring cannot obtain. This enables the monitoring system to analyze the structural stability within the temperature field and identify regions that have a critical impact on the overall temperature distribution.

[0017] By encoding and predicting network features using a deep learning model, a predictive encoding vector for the temperature field state in future time periods is generated. The difference between the predicted state and the current state is quantified as trend information, and based on this, the data acquisition frequency of sensors in each region and the allocation of computational resources for network analysis are dynamically adjusted. The monitoring system can then proactively optimize its operating status and resource allocation based on predictions of temperature field evolution.

[0018] The networked representation based on the dynamic correlation structure of the temperature field provides a global analytical foundation for monitoring systems that transcends point-based data. Furthermore, adaptive strategy adjustments based on predicted trends enable the system to proactively regulate future states. Together, these two aspects realize the transformation of the monitoring mode from passive response to proactive decision-making. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the working principle of the temperature monitoring method for frozen seafood according to the present invention. Figure 2 A flowchart for anomaly handling and data repair to form a standard temperature sequence; Figure 3 A flowchart for constructing a directed temperature network; Figure 4 The training loss curve of a deep autoencoder; Figure 5 The difference distribution map of the encoding of multi-dimensional network structure features. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides a temperature monitoring method for frozen seafood. The method includes: periodically measuring the temperature using sensors distributed at different monitoring locations within the cold chain space to form an original temperature sequence; removing outliers and repairing missing values ​​in the original temperature sequence to form a standard temperature sequence; identifying temperature pattern units with periodic or trend-like changes from the standard temperature sequence; constructing a directed temperature network describing the dynamic evolution of the temperature field based on the spatial proximity and temporal continuity of the temperature pattern units; dividing different temperature feature regions based on the temperature flow direction and intensity between nodes in the directed temperature network; converting the network structure features of different temperature feature regions into multi-dimensional feature vectors; compressing and reconstructing the multi-dimensional feature vectors using a deep autoencoder network to obtain a low-dimensional feature encoding vector; combining environmental parameters and historical feature encoding vectors, calculating the predicted encoding vector of the temperature field for the next time period using a prediction model; comparing the difference between the predicted encoding vector and the current feature encoding vector to generate temperature evolution trend information; and driving an adaptive monitoring strategy based on this trend information to dynamically adjust the sensor acquisition frequency and the structure of the directed temperature network.

[0022] In one embodiment of the present invention, see [reference] Figure 2 A moving average is calculated based on data from each monitoring location in the original temperature sequence. Temperature points deviating from the moving average by more than a preset deviation threshold are marked as candidate outliers. These candidate outliers are validated by combining temperature readings from neighboring monitoring locations at the same time point. Validated outliers are removed from the sequence, creating data gaps at their removal locations. A weighted averaging method, with weights based on temporal and spatial proximity, is used to interpolate and repair the data gaps. The repaired sequence is then smoothed and filtered to obtain a noise-suppressed continuous temperature sequence, which serves as the standard temperature sequence.

[0023] A sliding time window is used to extract temperature subsequences within a standard temperature series. Autocorrelation analysis is performed on each temperature subsequence to detect the presence of periodic temperature fluctuations. For temperature subsequences without significant periodicity, trend fitting is performed and their temperature change slopes are calculated. Adjacent time windows with similar periods or similar slopes are merged. A temperature pattern is defined for each merged time period, described by pattern type, duration, average rate of change, and amplitude range. All identified patterns are arranged in chronological order of their start times to form a temperature pattern unit sequence.

[0024] In some embodiments, a moving average is calculated based on data from each monitoring location in the original temperature sequence. The window length of the moving average is set according to the temperature data acquisition frequency and the typical thermal inertia period of the cold chain environment. Temperature points that deviate from the moving average by more than a preset deviation threshold are marked as candidate anomalies. Candidate anomalies are verified by combining temperature readings from neighboring monitoring locations at the same time point. If the difference between the temperature reading of a candidate anomaly and the average temperature readings of all neighboring monitoring locations at the same time exceeds the verification threshold, the candidate anomaly is determined to be a true anomaly. Verified anomalies are removed from the original temperature sequence, creating data gaps at the removal locations.

[0025] Optionally, a weighted averaging method based on temporal and spatial proximity is used to interpolate and repair missing data points. For each missing data point, a set of reference points is selected in the original temperature sequence. These reference points include several normal data points at the same monitoring location before and after the missing data point's time, and several normal data points at neighboring monitoring locations near the missing data point's timestamp. Each reference point is assigned a weight, determined by the time difference and spatial distance between it and the missing data point; reference points with smaller time differences and closer spatial distances are given higher weights. The repaired value for the missing data point is obtained by calculating the weighted average of the temperature readings of all reference points. The repaired sequence is then smoothed using convolution operations. The parameters of the convolution kernel are designed to suppress high-frequency noise and preserve the true trend of temperature changes, ultimately yielding a noise-suppressed continuous temperature sequence as the standard temperature sequence.

[0026] The process involves identifying temperature pattern units with periodic or trend-like changes from a standard temperature sequence. A fixed-length time window is slid across the standard temperature sequence, and temperature subsequences within each window are extracted. Autocorrelation analysis is performed on each temperature subsequence, calculating the correlation coefficient between the subsequence and itself at different time lags to detect periodic temperature fluctuations. If one or more time lags cause the correlation coefficient to exceed a significant periodicity threshold, the temperature subsequence is considered periodic. For temperature subsequences without significant periodicity, trend fitting is performed using linear regression to calculate the slope of temperature change over time. Adjacent time windows with similar periods or similar slopes are merged. Similarity is determined based on the difference in period values ​​being less than a period merging threshold, or the difference in slope values ​​being less than a slope merging threshold. A temperature pattern is defined for each merged time period, described by pattern type, duration, average rate of change, and amplitude range. Pattern types include periodic fluctuations, upward trends, downward trends, or steady states. All identified patterns are arranged in chronological order of their start times to form a sequence of temperature pattern units.

[0027] The following formula can be used to combine temporal and spatial proximity in the specific calculation of weights in weighted average interpolation:

[0028] in: This represents the weight of the reference point j used to repair the missing point i. It is the time interval between the missing point i and the reference point j. It is the spatial distance between the monitoring positions corresponding to the missing point i and the reference point j. It is an adjustment coefficient used to balance the influence of temporal proximity and spatial proximity. and These are the scale parameters that control the temporal decay rate and the spatial decay rate, respectively.

[0029] In one embodiment of the present invention, see [reference] Figure 3Each monitoring location's state within each temperature model unit's duration is defined as a network node. If two nodes correspond to spatially adjacent monitoring locations, a spatial connection edge is established between them. If two nodes correspond to two consecutive model units at the same monitoring location, a temporal connection edge is established between them. Each spatial connection edge is assigned a weight based on the average temperature difference between the model units of the two nodes. Each temporal connection edge is also assigned a weight based on the temporal probability of a mode transition from one node to the next. All nodes, spatial connections, temporal connections, and their weights are combined to form a weighted directed graph structure, serving as a directed temperature network describing the dynamic evolution of the temperature field.

[0030] In some embodiments, the state of each monitoring location within the duration of each temperature mode unit is defined as a network node. The network node includes a location identifier, mode type, and average temperature attribute. For example, in a rectangular cold storage with twelve monitoring locations in three rows and four columns, if a monitoring location P is identified as exhibiting a "periodic fluctuation pattern" within the time interval [T1, T2], a network node representing "location P is in a periodic fluctuation state during the time interval T1 to T2" is generated. If the monitoring locations corresponding to two nodes are spatially adjacent, a spatial connection edge is established between the two nodes. Optionally, a weight is assigned to each spatial connection edge, calculated based on the average temperature difference between the mode units of the two nodes. The weight value of the spatial connection edge reflects the temperature gradient or heat exchange intensity of adjacent locations within the corresponding time period. If two nodes correspond to two consecutive mode units of the same monitoring location in time, a temporal connection edge is established between the two nodes. Temporal continuity means that the end time of the previous mode unit is the same as or separated from the start time of the next mode unit by a very short data acquisition interval. Each temporal connection edge is assigned a weight, which is calculated based on the temporal probability of the mode transition from the previous node to the next node.

[0031] It can be understood that combining all nodes, spatial connections, temporal connections, and their weights to form a weighted directed graph structure serves as a directed temperature network describing the dynamic evolution of the temperature field. In this directed temperature network, nodes represent spatiotemporal states, spatial connections represent heat transfer relationships at different locations within the same time period, and temporal connections represent the state evolution relationships at the same location at different time periods. The method for calculating the weights of spatial connections based on the average temperature difference can be standardized using the following formula:

[0032] in: This represents the weight of the spatial connection edge linking node u and node v. It is the average temperature of the mode cell represented by node u. It is the average temperature of the mode cell represented by node v. and These refer to the maximum and minimum average temperatures of all nodes in the current directed temperature network, respectively. Weights The value range is normalized to the [0,1] interval. A larger value indicates a greater average temperature difference between the two adjacent spatial locations represented by node u and node v within the corresponding time period, and a stronger potential driving force for thermal flow. The weight of the time-connected edge... Then, the value is directly assigned to the mode conversion probability value obtained from historical data statistics, and its value range is [0,1].

[0033] In practice, the process of constructing a directed temperature network includes traversing all temperature pattern units and instantiating nodes. Based on a preset spatial adjacency matrix and the temporal continuity judgment rules of the pattern units, connecting edges are established between nodes, and the weight of each edge is calculated. The resulting weighted directed graph structure has an adjacency matrix that simultaneously implies the correlation strength of the temperature field in the spatial dimension and the evolution possibility in the temporal dimension.

[0034] In one embodiment of the invention, the difference between the sum of outgoing edge strengths and the sum of incoming edge strengths of each node in the directed temperature network is calculated to obtain the net temperature flow intensity of each node. Nodes with high net temperature flow intensity are used as seed points. Region growth is performed along directed edges from the seed point, incorporating adjacent nodes with the same temperature flow direction as the seed point and whose connection strength exceeds the connection strength threshold into the same region. When no new nodes can be incorporated, an initial temperature feature region is formed. For nodes not incorporated into any initial region, their average temperature similarity to the core of each initial region is calculated, and they are assigned to the region with the highest similarity. Boundary optimization is performed on all regions, merging regions with too small an area or too few nodes, ultimately forming stable, partitioned temperature feature regions.

[0035] For each temperature characteristic region, the total number of nodes is counted, and the average weight of the edges connecting all nodes within the region is calculated as the region's average connection strength. The ratio of the total weight of outgoing edges from a node in the region to the total weight of incoming edges from a node outside the region to a node in the region is calculated as the region's net temperature outflow ratio. A histogram of the temperature pattern unit type distribution for this region is extracted over the most recent complete time period. The histograms of the total number of nodes, the region's average connection strength, the region's net temperature outflow ratio, and the temperature pattern unit type distribution are normalized and concatenated to form a multidimensional feature vector representing the network structure characteristics of this temperature characteristic region.

[0036] In some embodiments, the net temperature flow intensity of each node is obtained by calculating the difference between the sum of the outgoing edge strengths and the sum of the incoming edge strengths of each node in the directed temperature network. A positive net temperature flow intensity indicates that the spatiotemporal state represented by the node tends to output heat to the surroundings, while a negative value indicates that it tends to absorb heat from the surroundings. Nodes with high net temperature flow intensity are selected as seed points, and the selection criterion for seed points is that the absolute value of the net temperature flow intensity is greater than a preset seed point threshold. Region growth is performed along the directed edges from the seed points, and adjacent nodes with the same temperature flow direction as the seed point and whose connection strength exceeds the connection strength threshold are included in the same region. The same temperature flow direction means that the sign of the net temperature flow intensity of the adjacent nodes is the same as the sign of the net temperature flow intensity of the seed point. When no new nodes can be included, an initial temperature feature region is formed. For nodes not included in any initial region, the average temperature similarity between them and the core of each initial region is calculated, and they are assigned to the region with the highest similarity. The core of each initial region is defined by the average temperature vector of all nodes in the region. Boundary optimization is performed on all regions, merging regions with too small an area or too few nodes. Regions with an area or number of nodes below the merging threshold will be merged into the neighboring region with the most shared connection edges and the closest average temperature, ultimately forming stable regions with different temperature characteristics.

[0037] It is understandable that for each temperature feature region, its network structure features need to be converted into a multi-dimensional feature vector, and the total number of nodes contained in the temperature feature region needs to be counted. The average weight of the edges connecting all nodes within the temperature feature region is calculated as the region's average connection strength, which reflects the closeness of the correlation and influence between temperature states within the region. The ratio of the total weight of outgoing edges from nodes in the temperature feature region to nodes outside the region to the total weight of incoming edges from nodes outside the region to nodes in the temperature feature region is calculated as the region's net temperature outflow ratio, which describes whether the region as a whole acts as a heat source or a cold source. A histogram of the distribution of temperature pattern unit types in the temperature feature region within the most recent complete time period is extracted. A complete time period refers to the length of time from the current moment to the recurrence of temperature pattern units. The horizontal axis of the distribution histogram represents the pattern type, and the vertical axis represents the cumulative duration percentage of the corresponding type of pattern unit within the period. The distribution histograms of the total number of nodes, the average regional connection strength, the regional temperature net outflow ratio, and the temperature pattern unit type are normalized and spliced ​​together. The total number of nodes, the average regional connection strength, and the regional temperature net outflow ratio are scaled to the [0,1] interval. Each bucket value of the distribution histogram is also normalized. All values ​​are connected in a fixed order to form a multidimensional feature vector representing the network structure characteristics of the temperature feature region.

[0038] Optionally, the net outflow ratio of the zone temperature can be calculated using the following formula:

[0039] in: This indicates the net outflow ratio of the region's temperature. This represents the set of nodes belonging to the current temperature characteristic region. This represents the weight of the directed connection edge from node i to node j. It is a very small positive constant used to prevent the denominator from being zero. The numerator summarizes the total weight of all edges flowing from nodes inside the region to nodes outside the region, and the denominator summarizes the total weight of edges flowing from nodes outside the region to nodes inside the region. A value greater than 1 indicates a net heat outflow from the region as a whole, while a value less than 1 indicates a net heat inflow. During normalized stitching... It will be mapped to the [0,1] interval, for example, by processing it through a sigmoid function.

[0040] In some embodiments, the region growth process is iterative. In each iteration, all adjacent nodes of the current region boundary node are checked to determine whether they meet the inclusion criteria. The inclusion criteria include that the adjacent node has not yet belonged to any region, the net temperature flow direction intensity sign of the adjacent node and the region seed point is the same, and the absolute value of the weight of the directed edge connecting the current region boundary node and the adjacent node to be examined is greater than the connection strength threshold. The adjacent nodes that meet the criteria are added to the region and become new boundary nodes for the next round of growth judgment. The iteration continues until there are no adjacent nodes that meet the criteria for the region boundary node.

[0041] In one embodiment of the present invention, a deep autoencoder network comprising a symmetric encoder and a decoder is constructed. The encoder consists of multiple fully connected layers, each followed by a nonlinear activation function. A multidimensional feature vector is input into the encoder and subjected to layer-by-layer nonlinear transformation, generating a low-dimensional compressed code at the bottleneck layer. This compressed code is then input into the decoder and subjected to layer-by-layer nonlinear transformation to reconstruct an output vector with the same dimension as the input. The deep autoencoder network is trained by minimizing the mean square error between the multidimensional feature vector and the reconstructed output vector. After training, the multidimensional feature vector is input into the encoder, and the compressed code output at the bottleneck layer is the feature encoding vector.

[0042] Environmental parameters collected synchronously with the standard temperature sequence include external ambient temperature, cold storage equipment operating power, and door opening / closing event records. These environmental parameters are standardized to form an environmental parameter vector. The feature encoding vector of the current moment is stacked chronologically with the feature encoding vectors of multiple consecutive historical moments to form a historical feature encoding sequence. The environmental parameter vector and the historical feature encoding sequence are fused along the feature dimension. The fused features are then input into a prediction model built on a Long Short-Term Memory (LSTM) network. This prediction model outputs a vector with the same dimension as the feature encoding vector, which serves as the predicted encoding vector for the temperature field in the next time period.

[0043] A deep autoencoder network with a symmetric encoder and decoder is constructed. The encoder consists of multiple fully connected layers, each followed by a non-linear activation function. The number of layers and neurons in each layer of the encoder are pre-defined based on the dimensions of the input multidimensional feature vector. For example, when the input dimension is 128, the encoder structure can be 128-64-32-16, ultimately producing a 16-dimensional compressed code at the bottleneck layer. The decoder has a symmetric structure with the encoder, progressively mapping the compressed code back to the original dimensions, for example, using a 16-32-64-128 structure. The multidimensional feature vector is input into the encoder and undergoes layer-by-layer non-linear transformation, producing a low-dimensional compressed code at the bottleneck layer. The non-linear activation function can be either ReLU or Sigmoid. The compressed code is then input into the decoder and undergoes layer-by-layer non-linear transformation to reconstruct an output vector with the same dimensions as the input. The deep autoencoder network is trained by minimizing the mean squared error between the multidimensional feature vector and the reconstructed output vector, using backpropagation and gradient descent optimizer. After training, the multidimensional feature vector to be processed is input into the encoder, and the compressed code output by the encoder at the bottleneck layer is the feature encoding vector.

[0044] It is understandable that combining environmental parameters with historical feature encoding vectors and calculating the predicted encoding vector of the temperature field for the next time period through a prediction model requires collecting environmental parameters that are synchronously recorded with the standard temperature sequence. These environmental parameters include external ambient temperature, cold storage equipment operating power, and door opening / closing event records. The door opening / closing event records are binary sequences recording the times when doors open and close. In the specific implementation of the prediction model, the collection of environmental parameters must maintain strict time synchronization with the standard temperature sequence. The external ambient temperature is collected in real time by external sensors, the cold storage equipment operating power is read from the refrigeration system control unit, and the door opening / closing event records are generated as binary sequences by access control sensors. The environmental parameter vectors are standardized to eliminate the influence of dimensions; for example, the Z-score method is used to convert each parameter value to zero mean and unit variance. The historical feature encoding sequence is formed by stacking the feature encoding vectors of the current time and multiple consecutive previous time periods, with the time step set according to the inertial period of temperature field changes. The fusion operation uses vector concatenation to connect the environmental parameter vectors and the historical feature encoding sequence along the feature dimension, forming extended input features. A prediction model based on a Long Short-Term Memory (LSTM) network comprises an input layer, hidden layers, and an output layer. The number of hidden layer units is configured according to the dimension of the fused features. The output layer uses a fully connected layer to generate a prediction encoding vector with the same dimension as the feature encoding vector. This model is trained on historical data to learn the evolution patterns of the temperature field. Environmental parameters are standardized to form an environmental parameter vector, scaling the values ​​of each environmental parameter sequence to zero mean and unit variance. The feature encoding vector at the current moment is stacked chronologically with the feature encoding vectors from multiple consecutive historical moments to form a historical feature encoding sequence, with the number of historical moments constituting the time step. The environmental parameter vector and the historical feature encoding sequence are fused along the feature dimension; this fusion operation can be a simple vector concatenation. The fused features are input into the prediction model based on the LSM network, with the number of hidden layer units configured according to the total dimension of the input features. The prediction model outputs a vector with the same dimension as the feature encoding vector, serving as the prediction encoding vector for the temperature field in the next time period.

[0045] In some embodiments, the training loss function of a deep autoencoder network is defined as the mean square error between the input multidimensional feature vector and the reconstructed output vector, and its specific calculation formula is as follows:

[0046] in: Let represent the reconstruction loss of the deep autoencoder network on a single batch of training data, where m represents the number of multidimensional feature vectors contained in that batch. This represents the k-th true multidimensional feature vector of the input. This represents the k-th vector reconstructed from the deep autoencoder network. Let L2 norm be the squared value of a vector. The training objective is to minimize L2 norm by optimizing the network parameters. See Table 1.

[0047] Table 1: Example fragments of environmental parameter vector composition

[0048] In some embodiments, the fusion of the environmental parameter vector and the historical feature encoding sequence occurs before inputting into the Long Short-Term Memory (LSTM) network. For each historical time step, the feature encoding vector at that moment is concatenated with the corresponding environmental parameter vector fragment to form an expanded feature vector. The expanded feature vectors from all time steps are then sequentially arranged and input into the LTM network. The output layer of the LTM network prediction model is a fully connected layer with the number of neurons equal to the dimension of the feature encoding vector. The activation function is a linear function, and the output of this layer is the prediction encoding vector for the next time step.

[0049] See Figure 4 During the training process of the deep autoencoder, the relationship between the number of training epochs (Phase 3) and the mean squared error loss is presented. Specifically, the purple solid line in the figure represents the training loss, and the orange dashed line is the loss fitting curve. Both show a trend of continuous decrease and gradual convergence of the loss as the number of training epochs increases: in the initial stage (Epoch < 10), the loss decreases rapidly, dropping quickly from about 1.2 to below 0.5; as the number of training epochs increases (10 ≤ Epoch ≤ 30), the rate of loss decrease slows down; when Epoch ≥ 30, the loss tends to stabilize, eventually converging to a small range around 0.05. This change characteristic is consistent with the training pattern of deep autoencoders: the model optimizes the network parameters through backpropagation, minimizing the mean squared error (i.e., the loss in the figure) between the input multidimensional feature vector and the reconstructed output vector. The effect of parameter adjustment on loss optimization is more significant in the early stage. In the later stage, the model gradually fits the data distribution, and the loss converges to a lower level, indicating that the deep autoencoder has learned the effective compression and reconstruction capabilities of multidimensional feature vectors through training.

[0050] In one embodiment of the present invention, the difference between the predicted encoding vector and the current feature encoding vector in each corresponding dimension is calculated. The differences in each dimension are then weighted and summed, where the weights are pre-set based on the contribution of each dimension to the overall stability of the temperature field. The weighted summation result is compared with a preset trend intensity threshold; if it exceeds the threshold, a significant temperature evolution trend is determined. The indices of the dimensions with the largest differences are mapped back to the corresponding network structure features to analyze whether the main temperature changes occur in a specific region, connection strength, or pattern distribution. A structured text description is generated by combining the trend determination result and the change analysis result, serving as the temperature evolution trend information driving the adaptive monitoring strategy.

[0051] The differences between the predicted encoding vector and the current feature encoding vector are calculated along each corresponding dimension. The predicted encoding vector and the current feature encoding vector have the same dimensional structure, with each dimension corresponding to a compressed network structure feature. The differences along each dimension are then weighted and summed. The weights are pre-set based on the contribution of each dimension to the overall stability of the temperature field. The contribution is evaluated based on historical data analysis, with dimensions having a greater impact on stability receiving higher weights. The weighted summation result is compared to a preset trend intensity threshold. If the weighted summation result exceeds the trend intensity threshold, a significant temperature evolution trend is identified. The trend intensity threshold is a configurable parameter used to control the sensitivity of the trend alarm.

[0052] In some embodiments, the indices of several dimensions with the largest differences are mapped back to the corresponding network structure features. The dimension indices point to specific positions in the feature encoding vector, corresponding to specific feature components in the multi-dimensional feature vector before normalization and concatenation. The analysis determines whether the main temperature changes occur in a specific region, in connectivity strength, or in pattern distribution. For example, if the dimension index with the largest difference corresponds to the regional net temperature outflow ratio feature, the analysis suggests that the main evolution trend of the temperature field reflects a change in the balance of heat exchange between regions; if the index corresponds to the regional average connectivity strength feature, the analysis suggests that the thermal coupling relationship within the region is changing; if the index corresponds to a bucket in the histogram of temperature pattern unit types, the analysis suggests that a certain temperature fluctuation pattern is strengthening or weakening. Combining the trend determination results and change analysis results, a structured text description is generated. This structured text description includes the trend intensity level, the identifier of the main change dimensions, the direction of change, and a summary of the spatial range of impact.

[0053] Optionally, the specific process of weighted summation and calculating the trend strength value can be expressed by the following formula:

[0054] in: This represents the calculated intensity value of the temperature evolution trend, where N is the total dimension of the feature encoding vector. It is the difference between the predicted encoding vector and the current feature encoding vector in the i-th dimension. These are pre-defined weighting coefficients corresponding to the i-th dimension, used to reflect the contribution of that dimension's characteristics to the overall temperature field stability. This indicates an absolute value operation. Trend strength value. The calculation result is a non-negative scalar, and the larger the value, the greater the deviation of the prediction.

[0055] It is understandable that generating structured text descriptions serves as temperature evolution trend information to drive adaptive monitoring strategies. The format of these structured text descriptions can be machine-readable key-value pairs or a specific markup language. For example, a structured description might include fields such as "Trend Intensity: HIGH," "Main Change Characteristics: REGION_NET_OUTFLOW_RATIO," "Change Direction: INCREASING," and "Estimated Affected Area: ZONE_A, ZONE_C." This temperature evolution trend information will be sent to the monitoring system's strategy scheduling module. Based on the content of this information, the strategy scheduling module will determine whether to adjust the sensor acquisition frequency and network structure. For instance, when the trend intensity is high and the main change characteristics involve a specific area, the acquisition frequency of sensors within that area will be increased, and the network subgraph corresponding to that area will undergo more frequent reconstruction analysis.

[0056] See Figure 5 In the early warning stage of temperature monitoring for frozen seafood, the distribution of differences between the predicted encoding vector and the current feature encoding vector across various network structural feature dimensions is presented. Specifically, each dimension in the figure corresponds to the network structural features compressed by the deep autoencoder (such as the total number of nodes in a region, connection strength, and the proportion of temperature patterns), and the difference is the calculated result of "prediction - current". From the distribution, the differences in dimensions such as the average connection strength of the region (0.400) and the total number of nodes in region A (0.350) are positive and relatively high, indicating that the deviation of these features in the prediction is more significant; while the differences in dimensions such as the proportion of periodic patterns (-0.280) and the total number of nodes in region C (-0.220) are negative, reflecting the reverse deviation of their predicted features relative to the current situation. The magnitude of these dimensional differences directly corresponds to the tendency of the feature to influence the stability of the temperature field: positive difference dimensions represent an strengthening (or weakening) trend of the feature in the prediction, while negative difference dimensions represent the opposite. Subsequently, the intensity of the temperature evolution trend can be determined by summing the weights of these differences (based on the contribution of the dimension to the stability of the temperature field). The specific network structure characteristics can be mapped back through the dimension with prominent differences (such as the average connection strength of the region) to analyze the core direction of temperature change (such as the change of regional thermal coupling relationship), providing a basis for adaptive monitoring strategies.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for temperature monitoring of frozen seafood, characterized in that, include: The original temperature sequence of seafood products in the cold chain space is collected. The original temperature sequence is obtained by periodic measurement by sensors distributed at different monitoring locations in the cold chain space. The original temperature sequence is subjected to outlier removal and missing value repair to form a standard temperature sequence; Identify temperature pattern units with periodic or trend-like changes from the standard temperature sequence; Based on the spatial proximity and temporal continuity of the temperature pattern units, a directed temperature network describing the dynamic evolution of the temperature field is constructed. Based on the temperature flow direction and intensity between nodes in the directed temperature network, different temperature characteristic regions are divided. The network structure features of different temperature regions are converted into multi-dimensional feature vectors; The multidimensional feature vector is compressed and reconstructed using a deep autoencoder network to obtain the feature encoding vector; By combining environmental parameters with historical feature encoding vectors, the predicted encoding vector of the temperature field for the next time period is calculated through a prediction model. The difference between the predicted encoding vector and the current feature encoding vector is compared to generate temperature evolution trend information; Based on temperature evolution trend information, an adaptive monitoring strategy is driven to adjust the sensor acquisition frequency and network structure.

2. The temperature monitoring method for frozen seafood according to claim 1, characterized in that, The process of removing outliers and repairing missing values ​​in the original temperature sequence to form a standard temperature sequence includes: A moving average is calculated based on the data from each monitoring location in the original temperature sequence; Temperature points that deviate from the moving average line by more than a preset deviation threshold are marked as candidate outliers; The candidate anomalies were verified by combining temperature readings from nearby monitoring locations at the same time point; Verified outliers are removed from the sequence, creating missing data points at the removal locations; A weighted average method with time proximity and spatial proximity as weights is used to interpolate and repair the missing data points; The repaired sequence is smoothed and filtered to obtain a continuous temperature sequence with noise suppression, which is then used as the standard temperature sequence.

3. The temperature monitoring method for frozen seafood according to claim 1, characterized in that, The step of identifying temperature pattern units with periodic or trend-like changes from the standard temperature sequence includes: Slide a time window across the standard temperature sequence and extract the temperature subsequence within the window; Autocorrelation analysis was performed on each temperature subsequence to detect the presence of periodic temperature fluctuations; For temperature subsequences that do not exhibit significant periodicity, trend fitting is performed to calculate the slope of their temperature changes; Merge adjacent time windows with similar periods or similar slopes; A temperature pattern is defined for each merged time period, and the temperature pattern is described by the pattern type, duration, average rate of change and amplitude range. All identified patterns are arranged in order of their start time to form a temperature pattern unit sequence.

4. The temperature monitoring method for frozen seafood according to claim 3, characterized in that, The construction of a directed temperature network describing the dynamic evolution of the temperature field based on the spatial proximity and temporal continuity of the temperature model units includes: The status of each monitoring location during the duration of each temperature mode unit is defined as a network node. If the monitoring locations corresponding to two nodes are spatially adjacent, then a spatial connection edge is established between the two nodes; If two nodes correspond to two consecutive mode units at the same monitoring location in time, then a time connection edge is established between the two nodes. Each spatial connection edge is assigned a weight, which is calculated based on the average temperature difference between the mode cells of the two nodes. Each temporal connection edge is assigned a weight, which is calculated based on the temporal probability of mode transition from the previous node to the next node. All nodes, spatial connections, temporal connections, and their weights are combined to form a weighted directed graph structure, which serves as the directed temperature network describing the dynamic evolution of the temperature field.

5. The temperature monitoring method for frozen seafood according to claim 1, characterized in that, Based on the temperature flow direction and intensity between nodes in the directed temperature network, different temperature characteristic regions are divided, including: Calculate the difference between the sum of outgoing edge strength and the sum of incoming edge strength of each node in the directed temperature network to obtain the net temperature flow intensity of each node; Nodes with high net temperature flow intensity are used as seed points; Starting from the seed point, perform region growth along the directed edge, and include adjacent nodes that have the same temperature flow direction as the seed point and whose connection strength exceeds the connection strength threshold into the same region. When no new nodes can be incorporated, an initial temperature characteristic region is formed; For nodes not included in any initial region, calculate their average temperature similarity with the core of each initial region and assign them to the region with the highest similarity. Boundary optimization is performed on all regions, merging regions with too small an area or too few nodes, ultimately forming stable regions with different temperature characteristics.

6. The temperature monitoring method for frozen seafood according to claim 5, characterized in that, The process of converting network structure features in different temperature regions into multidimensional feature vectors includes: For each temperature characteristic region, count the total number of nodes it contains; Calculate the average weight of the connection edges between all nodes within the temperature characteristic region, and use it as the average connection strength of the region. The ratio of the total weight of outgoing edges from nodes in the temperature characteristic region to nodes outside the region to the total weight of incoming edges from nodes outside the region to nodes in the temperature characteristic region is calculated as the net outflow ratio of the region temperature. Extract the distribution histogram of temperature pattern unit types in the temperature feature region over the most recent complete time period; The distribution histograms of the total number of nodes, the average regional connection strength, the regional net temperature outflow ratio, and the temperature pattern unit type are normalized and spliced ​​together to form a multi-dimensional feature vector representing the network structure characteristics of the temperature feature region.

7. The temperature monitoring method for frozen seafood according to claim 1, characterized in that, The process of compressing and reconstructing the multidimensional feature vector using a deep autoencoder network to obtain a feature encoding vector includes: A deep autoencoder network containing a symmetric encoder and a decoder is constructed. The encoder consists of multiple fully connected layers, each followed by a nonlinear activation function. The multidimensional feature vector is input into the encoder, and after layer-by-layer nonlinear transformation, a low-dimensional compressed code is generated at the bottleneck layer. The compressed code is input into the decoder, and after layer-by-layer nonlinear transformation, an output vector with the same dimension as the input is reconstructed. The deep autoencoder network is trained by minimizing the mean square error between the multidimensional feature vector and the reconstructed output vector. After training is completed, the multidimensional feature vector is input into the encoder, and the compressed code output by the encoder at the bottleneck layer is the feature encoding vector.

8. The temperature monitoring method for frozen seafood according to claim 1, characterized in that, The process of combining environmental parameters and historical feature encoding vectors to calculate the predicted encoding vector of the temperature field for the next time period through a prediction model includes: Collect environmental parameters that are recorded synchronously with the standard temperature sequence, including external ambient temperature, cold storage equipment operating power, and door opening / closing event records; Environmental parameters are standardized to form an environmental parameter vector; The feature encoding vector at the current moment is stacked with the feature encoding vectors from multiple consecutive historical moments in chronological order to form a historical feature encoding sequence. The environmental parameter vector and the historical feature encoding sequence are fused along the feature dimension. The fused features are then input into a prediction model built on a long short-term memory network. The prediction model outputs a vector with the same dimension as the feature encoding vector, which serves as the prediction encoding vector for the temperature field in the next time period.

9. A method for temperature monitoring of frozen seafood according to claim 1, characterized in that, The step of comparing the difference between the predicted encoding vector and the current feature encoding vector to generate temperature evolution trend information includes: Calculate the difference between the predicted encoding vector and the current feature encoding vector in each corresponding dimension; The differences across each dimension are summed in a weighted manner, with the weights pre-set based on the contribution of each dimension to the overall stability of the temperature field. The weighted summation result is compared with a preset trend intensity threshold. If it exceeds the trend intensity threshold, it is determined that there is a significant temperature evolution trend. Based on the indices of the dimensions with the largest differences, the corresponding network structure features are mapped back to determine whether the main temperature changes occur in a specific region, connection strength, or pattern distribution. By combining the trend determination results and change analysis results, a structured text description is generated as the temperature evolution trend information driving the adaptive monitoring strategy.

10. A temperature monitoring system for frozen seafood, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the temperature monitoring method for frozen seafood as described in any one of claims 1 to 9.

Images