Machine learning based cold storage environment anomaly detection method and system

Abstract

The application provides a cold storage environment anomaly detection method and system based on machine learning, which comprises the following steps: acquiring a cold storage environment space-time correlation benchmark model; collecting a current cold storage environment parameter sequence, and calculating a space-time coordination deviation of the current cold storage environment parameter sequence relative to an environment parameter sequence under a normal operation state based on the space-time correlation benchmark model; generating a multi-dimensional deviation feature vector according to the space-time coordination deviation; inputting the multi-dimensional deviation feature vector into a hierarchical anomaly pattern reasoning model, performing nonlinear feature enhancement and multi-scale anomaly pattern recognition on the multi-dimensional deviation feature vector, and obtaining a preliminary confidence corresponding to each of a plurality of candidate anomaly patterns; generating a comprehensive anomaly quantitative index based on the preliminary confidence of the plurality of candidate anomaly patterns through a space-time correlation weighted fusion algorithm, and outputting a cold storage environment anomaly detection result according to the comprehensive anomaly quantitative index. The application improves the practicability and application value of the anomaly detection result.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a machine learning-based method and system for detecting anomalies in cold storage environments. Background Technology

[0002] With the development of cold chain logistics and warehousing technology, the detection of anomalies in cold storage environments has become increasingly important. Monitoring and analyzing environmental parameters within cold storage to identify abnormal states that may affect the quality of stored goods is crucial for ensuring the stable operation of the cold chain system. Currently, methods typically rely on preset static thresholds or fixed-dimensional benchmark models. This involves collecting environmental parameter sequences within the cold storage, comparing real-time parameter values ​​with static thresholds or fixed-dimensional benchmark values, calculating the deviation, and then using a conventional classification model to make a simple judgment on the deviation, outputting a result indicating whether an anomaly has occurred. However, static thresholds or fixed-dimensional benchmark models struggle to reflect the spatial dependence and temporal evolution of environmental parameters in different monitoring areas of the cold storage, leading to discrepancies between the benchmark representation of normal operation and the actual environmental changes. Furthermore, the simple deviation calculation fails to consider the propagation characteristics of deviations in the spatial dimension and their evolutionary trends in the temporal dimension, making it difficult to comprehensively capture the spatiotemporal distribution characteristics of anomalies in complex environments. Summary of the Invention

[0003] In view of this, the present invention provides a method and system for detecting anomalies in cold storage environments based on machine learning. The technical solution of the embodiments of the present invention is implemented as follows:

[0004] On one hand, embodiments of the present invention provide a machine learning-based method for detecting anomalies in cold storage environments, comprising: acquiring a spatiotemporal correlation benchmark model of the cold storage environment; collecting a current cold storage environment parameter sequence, and calculating the spatiotemporal co-deviation degree of the current cold storage environment parameter sequence relative to the environmental parameter sequence under normal operating conditions based on the spatiotemporal correlation benchmark model, and generating a multidimensional deviation feature vector based on the spatiotemporal co-deviation degree; inputting the multidimensional deviation feature vector into a hierarchical anomaly pattern inference model, performing nonlinear feature enhancement and multi-scale anomaly pattern recognition on the multidimensional deviation feature vector to obtain preliminary confidence levels corresponding to multiple candidate anomaly patterns; generating a comprehensive anomaly quantification index based on the preliminary confidence levels of the multiple candidate anomaly patterns through a spatiotemporal correlation weighted fusion algorithm, wherein the comprehensive anomaly quantification index is used to characterize the severity and propagation trend of the current cold storage environment anomaly; and outputting a cold storage environment anomaly detection result based on the comprehensive anomaly quantification index, wherein the cold storage environment anomaly detection result includes anomaly type judgment and anomaly source location information.

[0005] On the other hand, embodiments of the present invention provide a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0006] This invention provides a machine learning-based method for detecting anomalies in cold storage environments. It constructs a spatiotemporal correlation benchmark model of the cold storage environment to capture the spatial dependence and temporal evolution of environmental parameter sequences in different monitoring areas under normal operating conditions. This method overcomes the limitations of traditional static threshold or fixed-dimensional benchmark models. By integrating spatial correlation characteristics and temporal evolution patterns, the benchmark representation of normal operating conditions more closely matches the actual changes in cold storage environmental parameters over time and space, providing a more accurate reference basis for subsequent anomaly detection. After collecting the current cold storage environmental parameter sequence, the method calculates the spatiotemporal collaborative deviation degree based on this benchmark model and generates a multidimensional deviation feature vector. Compared to traditional methods that only compare single-point parameters with fixed thresholds or accumulate fixed-dimensional deviations, this method considers the propagation characteristics of deviations in the spatial dimension and the evolution trend in the temporal dimension, forming a high-dimensional feature expression that includes spatiotemporal correlation information. This more comprehensively captures the spatiotemporal distribution characteristics of anomalies in complex environments. By inputting multidimensional deviation feature vectors into a hierarchical anomaly pattern inference model, complex nonlinear correlations between features are captured through nonlinear feature enhancement processing. Combined with multi-scale anomaly pattern recognition to cover anomaly patterns across different spatiotemporal ranges, this effectively improves the ability to identify nonlinear, multi-scale anomaly patterns in cold storage environments, avoiding the difficulty of handling complex anomalies by traditional conventional classification models or fixed-scale recognition. Based on the preliminary confidence of multiple candidate anomaly patterns, a comprehensive anomaly quantification index is generated through a spatiotemporal correlation weighted fusion algorithm. This algorithm adjusts the weights of each pattern according to its spatial propagation range and temporal evolution speed, making the comprehensive index more accurate in representing the severity and propagation trend of the anomaly. Compared to fixed-weight fusion, this approach better reflects the development characteristics of anomalies. Finally, based on the comprehensive anomaly quantification index, the detection results, including anomaly type judgment and anomaly source location information, are output. This achieves complete inference from anomaly detection to type judgment and source location, expanding the functional boundaries of traditional anomaly detection which only outputs simple states or single-type judgments, and enhancing the practicality and application value of anomaly detection results. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a machine learning-based method for detecting anomalies in a cold storage environment, as provided in an embodiment of the present invention.

[0009] Figure 2This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on 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.

[0011] In the following description, references to "some embodiments" describe a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "first / second / third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permissible, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of the invention.

[0013] This invention provides a machine learning-based method for detecting anomalies in a cold storage environment, which can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers or personal PCs.

[0014] Figure 1 This is a schematic diagram illustrating the implementation process of a machine learning-based cold storage environment anomaly detection method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0015] Step S100: Obtain the spatiotemporal correlation benchmark model of the cold storage environment. The spatiotemporal correlation benchmark model is used to capture the spatial dependence and temporal evolution of environmental parameter sequences in different monitoring areas of the cold storage under normal operating conditions.

[0016] The spatiotemporal correlation benchmark model for cold storage environments is a model capable of accurately depicting the changing patterns of environmental parameters during normal operation of a cold storage facility. The environmental parameter sequence is an ordered sequence composed of environmental parameter values ​​collected at fixed time intervals from various monitoring areas within the cold storage facility over a certain time range. These environmental parameters include temperature, humidity, and gas concentration. Spatial dependence reflects the mutual influence and correlation of environmental parameters between different monitoring areas; for example, the temperatures of adjacent areas may exhibit a certain correlation due to physical processes such as heat conduction. Temporal evolution patterns reflect the dynamic changing trends of environmental parameters within the same monitoring area over time; for example, temperature may exhibit periodic fluctuations at different times of the day.

[0017] To obtain this model, it is first necessary to collect historical normal operation data of the cold storage over a relatively long period. This data should comprehensively and accurately cover the environmental parameter sequences of each monitoring area. After data collection, it needs to be preprocessed, including data cleaning to remove noise and erroneous data, and data normalization to ensure that different types of environmental parameters have the same scale range, facilitating subsequent model training. Next, a suitable machine learning algorithm is selected to build the model. Here, a Long Short-Term Memory (LSTM) network is used as an example for detailed explanation. LSTM is a special type of recurrent neural network with unique memory units and gating mechanisms, which can effectively process time series data and capture long-term dependencies.

[0018] An LSTM network mainly consists of an input layer, hidden layers, and an output layer. The input layer receives preprocessed historical environmental parameter sequences. The hidden layer contains multiple LSTM units, each consisting of an input gate, a forget gate, and an output gate. The input gate determines how much of the current input information enters the memory unit; the forget gate controls the retention or forgetting of old information in the memory unit; and the output gate determines how much information is output from the memory unit to the next layer. During training, input data sequentially passes through the input layer into the LSTM units in the hidden layer. The LSTM units update the memory units based on the current input and the hidden state from the previous time step, using a gating mechanism. The output of the hidden layer is passed to the output layer, which typically uses a linear layer or activation function (such as ReLU) to transform the output of the hidden layer, obtaining the predicted environmental parameter values.

[0019] The training process involves continuously adjusting the model's weight parameters to make the model's predictions as close as possible to the actual environmental parameter values. Gradient descent algorithms (such as stochastic gradient descent (SGD) and adaptive moment estimation (Adam) can be used to update the model's weights. In each iteration, the error between the predicted and actual values ​​(such as mean squared error, MSE) is calculated, and then the weight parameters are updated according to the gradient direction of the error until the error converges to a small value. At this point, the model training is complete, and the resulting trained LSTM model is the spatiotemporal correlation benchmark model for the cold storage environment.

[0020] As one implementation method, the spatiotemporal correlation benchmark model is constructed through the following steps: First, a historical normal operating environment parameter sequence of the cold storage is obtained, which includes historical environmental parameter values ​​continuously collected at fixed time intervals from each monitoring area within a preset time period. Then, the historical normal operating environment parameter sequence is spatiotemporally segmented, dividing the preset time period evenly into multiple continuous and equal-length time blocks, and dividing all monitoring areas of the cold storage into multiple non-overlapping spatial grid blocks according to their spatial location, resulting in a spatiotemporal segmented dataset. This dataset contains environmental parameter subsequences for each spatial grid block within each time block. Next, the temporal evolution characteristics of the environmental parameter subsequences for each spatial grid block in the spatiotemporal segmented dataset are extracted. These temporal evolution characteristics characterize the trend fluctuations and periodic repetitions of the environmental parameter subsequences over time. Finally, the spatial correlation characteristics of the environmental parameter subsequences of adjacent spatial grid blocks in the spatiotemporal segmented dataset are calculated. These spatial correlation characteristics characterize the environmental parameters of different spatial grid blocks. The study investigates the synchronicity and mutual influence of parameter subsequences within the same time block; it also examines how to dimensionally concatenate the temporal evolution features of each spatial grid block with the spatial correlation features of all its adjacent spatial grid blocks to generate a spatiotemporal joint feature matrix. The row dimensions of this matrix correspond to spatial grid blocks, the column dimensions to time blocks, and the matrix elements are the fused spatiotemporal feature values. An initial spatiotemporal correlation model is constructed based on this matrix. The weight parameters of the initial spatiotemporal correlation model are iteratively optimized by minimizing the prediction error of the initial model for each spatial grid block's environmental parameter subsequence within each time block in the historical normal operating environment parameter sequence. Finally, when the prediction error is less than a preset error threshold, optimization is stopped, and the current model is designated as the spatiotemporal correlation benchmark model. This benchmark model can adaptively update the spatiotemporal joint feature matrix by periodically inputting newly added historical normal operating environment parameter sequences and repeatedly performing spatiotemporal block processing, spatiotemporal feature extraction, and model parameter optimization steps.

[0021] Step S200: Collect the current cold storage environmental parameter sequence, and calculate the spatiotemporal coordination deviation of the current cold storage environmental parameter sequence relative to the environmental parameter sequence under normal operation based on the spatiotemporal correlation benchmark model, and generate a multidimensional deviation feature vector based on the spatiotemporal coordination deviation.

[0022] The current cold storage environmental parameter sequence is a sequence of environmental parameter values ​​collected at the current moment from various monitoring areas within the cold storage over a continuously preset time period. These environmental parameter values ​​are collected continuously at fixed time intervals. The spatiotemporal coordination deviation measure the overall degree of deviation of the current cold storage environmental parameter sequence from the environmental parameter sequence under normal operating conditions in both spatial and temporal dimensions. It comprehensively considers the spatial correlation between different monitoring areas and the changes in the same monitoring area at different time points. The multidimensional deviation feature vector represents this spatiotemporal coordination deviation in vector form, containing information from multiple dimensions, and is used for anomaly pattern recognition and analysis.

[0023] As one implementation method, step S200 involves collecting the current cold storage environmental parameter sequence and calculating the spatiotemporal coordination deviation of the current cold storage environmental parameter sequence relative to the environmental parameter sequence under normal operating conditions based on the spatiotemporal correlation benchmark model. A multidimensional deviation feature vector is then generated based on the spatiotemporal coordination deviation. Specifically, this can be implemented as follows: steps S210~S260:

[0024] Step S210: Collect the current cold storage environmental parameter sequence. The current cold storage environmental parameter sequence includes the environmental parameter values ​​of each monitoring node in the preset monitoring grid within the cold storage within a continuous preset time period. The environmental parameter values ​​are sequence data collected continuously at fixed time intervals.

[0025] The preset monitoring grid is a pre-defined grid structure within the cold storage facility, with each grid intersection forming a monitoring node. These monitoring nodes are distributed across different locations within the cold storage, enabling comprehensive monitoring of environmental parameters. The preset continuous duration is a pre-defined time frame, such as the length of a work shift. Continuous data collection at fixed time intervals means that environmental parameter values ​​are collected at fixed intervals (e.g., every minute, every five minutes, etc.) within the preset duration.

[0026] To collect the current environmental parameter sequence of the cold storage, corresponding sensors, such as temperature sensors and humidity sensors, can be installed at each monitoring node. These sensors will detect changes in environmental parameters in real time and transmit the data to the data acquisition system. The data acquisition system will sample the sensor data at preset time intervals and store the collected data. For example, in a cold storage, the preset monitoring grid is a regular two-dimensional grid with multiple monitoring nodes. Within a preset time period, the temperature sensor at each monitoring node collects the temperature value once per minute, and the humidity sensor collects the humidity value once per minute. Finally, these data are arranged according to the order of monitoring nodes and the collection time to obtain the current environmental parameter sequence of the cold storage.

[0027] Step S220: Based on the spatiotemporal correlation benchmark model, extract the normal environmental parameter benchmark sequence corresponding to the current cold storage environmental parameter sequence. The normal environmental parameter benchmark sequence includes the normal parameter reference values ​​of each monitoring node in the preset monitoring grid within a continuous preset time period. The normal parameter reference values ​​are predicted and generated by the spatiotemporal correlation benchmark model based on historical normal operation data.

[0028] The normal environmental parameter baseline sequence is a sequence of environmental parameters under normal operating conditions, corresponding to the current cold storage environmental parameter sequence. The normal parameter reference values ​​are predicted by the spatiotemporal correlation baseline model based on previously learned normal operating patterns, for the environmental parameters of each monitoring node within the current preset time period.

[0029] In practice, the time information of the current cold storage environmental parameter sequence (such as the start time and time interval of collection) and the monitoring node information (such as the location and number of the monitoring node) are input into a pre-trained spatiotemporal correlation benchmark model. Based on this input information, and combined with its internally learned spatial dependencies and temporal evolution patterns, the model predicts the normal parameter reference value for each monitoring node within a continuously preset time period. For example, if the model knows from historical data that the temperature of a certain monitoring node typically fluctuates within a certain range during a certain time period, then for the current same time period, the model will predict the normal temperature reference value for that monitoring node. Arranging the normal parameter reference values ​​of all monitoring nodes in order of monitoring node sequence and collection time sequence yields the normal environmental parameter benchmark sequence.

[0030] Step S230: Calculate the absolute deviation between the environmental parameter value of each monitoring node in the current cold storage environmental parameter sequence at each collection time and the normal parameter reference value of the corresponding monitoring node and the corresponding collection time in the normal environmental parameter baseline sequence, and obtain the node time deviation matrix. The row dimension of the node time deviation matrix corresponds to the monitoring node, and the column dimension corresponds to the collection time.

[0031] Absolute deviation is an indicator that measures the degree of difference between the current environmental parameter value and the normal reference value. It is obtained by calculating the absolute value of the difference between the two. The node-time deviation matrix is ​​a two-dimensional matrix used to comprehensively record the degree to which the environmental parameters of each monitoring node deviate from the normal state at each acquisition time.

[0032] As one implementation method, step S230 involves calculating the absolute deviation between the environmental parameter value of each monitoring node in the current cold storage environmental parameter sequence at each acquisition time and the normal parameter reference value of the corresponding monitoring node and the corresponding acquisition time in the normal environmental parameter baseline sequence, thus obtaining the node time deviation matrix. Specifically, this can be implemented as follows: steps S231~S236:

[0033] Step S231: Traverse each monitoring node in the preset monitoring grid and take the currently traversed monitoring node as the target monitoring node.

[0034] Traversing each monitoring node in the preset monitoring grid ensures that the environmental parameters of all monitoring nodes within the cold storage are processed. During the traversal, the currently processed monitoring node is defined as the target monitoring node, allowing for targeted operations on its environmental parameters later. This traversal can be implemented using loop structures in programming languages ​​(such as a for loop), starting with the first monitoring node and sequentially setting each node as the target monitoring node for processing, until all monitoring nodes have been traversed.

[0035] Step S232: Extract the environmental parameter values ​​of the target monitoring node within a continuous preset time period from the current cold storage environmental parameter sequence, and arrange them in the order of collection time to form a target node parameter subsequence.

[0036] The environmental parameter values ​​of the target monitoring nodes are selected from the current cold storage environmental parameter sequence and arranged in chronological order of collection time to form a subsequence. This step can be achieved through indexing operations. Data corresponding to the target monitoring node is extracted from the current cold storage environmental parameter sequence based on the node's number, and then reorganized according to the collection time. For example, in a two-dimensional array storing the current cold storage environmental parameter sequence, the data for the target monitoring node is found using the array's row index, and then the data is sorted according to the column index (corresponding to the collection time) to obtain the target node parameter subsequence.

[0037] Step S233: Extract the reference values ​​of normal parameters of the target monitoring node within a continuous preset time period from the normal environmental parameter baseline sequence, and arrange them in the order of collection time to form the target node baseline subsequence.

[0038] Similarly, normal parameter reference values ​​for the target monitoring nodes are extracted from the normal environmental parameter baseline sequence and arranged into subsequences according to the collection time order. This process is similar to step S232, which also involves obtaining the data of the target monitoring nodes from the normal environmental parameter baseline sequence through indexing operations and sorting them.

[0039] Step S234: Traverse each acquisition time in the target node parameter subsequence and take the currently traversed acquisition time as the target acquisition time.

[0040] Each acquisition time in the target node parameter subsequence is accessed sequentially, and the currently processed acquisition time is defined as the target acquisition time. This is done to calculate the absolute deviation of the target monitoring node at each acquisition time. A loop structure can be used to iterate through each acquisition time in the target node parameter subsequence, starting from the first acquisition time and sequentially setting each acquisition time as the target acquisition time for processing, until all acquisition times have been traversed.

[0041] Step S235: Calculate the absolute value of the difference between the environmental parameter value of the target node parameter subsequence at the target acquisition time and the normal parameter reference value of the target node baseline subsequence at the target acquisition time, to obtain the absolute deviation of the target node at the target acquisition time.

[0042] The absolute deviation of the target node at a given moment is obtained by subtracting the environmental parameter value at the target acquisition time from the normal parameter reference value at the corresponding time in the target node's baseline subsequence, and then taking the absolute value of the difference. For example, if the actual temperature of the target node at a certain moment is one value, while the normal reference temperature is another value, then the absolute deviation at that moment is the absolute value of the difference between these two values.

[0043] Step S236: Arrange the absolute deviations of all monitoring nodes at all acquisition times into a matrix according to the order of monitoring nodes and acquisition times to obtain the node time deviation matrix.

[0044] The absolute deviation of each monitoring node at each acquisition time is sequentially filled into a matrix according to the order of the monitoring nodes and the acquisition times. Rows in the matrix correspond to monitoring nodes, and columns correspond to acquisition times. This matrix can be stored using a two-dimensional array, and each absolute deviation is assigned to its corresponding position using nested loops. This yields a complete node-time deviation matrix, which visually reflects the degree to which the environmental parameters of each monitoring node deviate from the normal state at each acquisition time.

[0045] Step S240: Perform spatial dimension correlation analysis on the node time deviation matrix, extract the absolute deviation sequence of each monitoring node based on the preset time window length, calculate the spatial correlation coefficient between any two absolute deviation sequences of monitoring nodes, use the spatial correlation coefficient as the deviation transmission coefficient between nodes, and obtain the spatial deviation transmission matrix.

[0046] Spatial correlation analysis aims to study the spatial relationships between different monitoring nodes in the node time deviation matrix. The preset time window length is a pre-defined time range used to extract the absolute deviation sequence of each monitoring node from the node time deviation matrix. The spatial correlation coefficient measures the similarity and correlation strength between the absolute deviation sequences of any two monitoring nodes, and serves as the deviation transmission coefficient between nodes to reflect the impact of one monitoring node's deviation on other monitoring nodes. The spatial deviation transmission matrix is ​​a two-dimensional matrix used to store the deviation transmission coefficients between any two monitoring nodes.

[0047] As one implementation method, step S240 involves performing spatial dimension correlation analysis on the node time deviation matrix, extracting the absolute deviation sequence of each monitoring node based on a preset time window length, calculating the spatial correlation coefficient between the absolute deviation sequences of any two monitoring nodes, and using the spatial correlation coefficient as the deviation transmission coefficient between nodes to obtain the spatial deviation transmission matrix. Specifically, this can be implemented as the following steps S241~S247:

[0048] Step S241: Based on the regular adjustment cycle of the cold storage equipment and the sampling interval of environmental parameters, set a preset time window length. The preset time window length is a set of multiple consecutive sampling moments, and the preset time window length is not less than two sampling moments.

[0049] The routine adjustment cycle of cold storage equipment refers to the time required for a complete adjustment operation of cold storage equipment (such as refrigeration equipment, ventilation equipment, etc.). The sampling interval of environmental parameters refers to the time interval between two consecutive samplings when collecting environmental parameters. The setting of the preset time window length needs to take both of these factors into account to ensure that effective change information of environmental parameters can be captured. The preset time window length should be a set of multiple consecutive sampling moments, and should include at least two sampling moments to ensure effective correlation analysis. For example, if the routine adjustment cycle of cold storage equipment is one hour and the sampling interval of environmental parameters is five minutes, then the preset time window length can be set to twelve sampling moments (i.e., the number of samplings within one hour).

[0050] Step S242: Determine the spatial adjacency relationship of each monitoring node in the preset monitoring grid, and construct a node adjacency list based on the spatial adjacency relationship. The node adjacency list contains each monitoring node and its adjacent monitoring nodes.

[0051] Spatial adjacency describes the spatial proximity of different monitoring nodes in a pre-defined monitoring grid. For example, in a two-dimensional grid, a monitoring node's neighbors might be nodes in its four cardinal directions (up, down, left, and right). Constructing a node adjacency list facilitates subsequent analysis of the relationships between adjacent monitoring nodes. This can be done by traversing each monitoring node in the pre-defined grid, determining its neighbors based on its position within the grid, and storing the information of each monitoring node and its neighbors in a list. For instance, in a regular two-dimensional grid, the neighboring nodes can be determined by calculating the node's coordinates, and this information can then be stored in a dictionary or list structure.

[0052] Step S243: Traverse each monitoring node pair in the node adjacency list, and take the currently traversed monitoring node pair as the target node pair. The target node pair includes the first target node and the second target node.

[0053] Traversing each monitoring node pair in the node adjacency list is for analyzing the spatial correlation between any two monitoring nodes. During the traversal, the currently processed monitoring node pair is defined as the target node pair, which includes a first target node and a second target node. Nested loops can be used to traverse each monitoring node pair in the node adjacency list, processing each monitoring node pair as the target node pair in turn, until all monitoring node pairs have been traversed.

[0054] Step S244: Extract all absolute deviations of the first target node within a preset time window from the node time deviation matrix, and arrange them in the order of acquisition time to form the first deviation sequence.

[0055] The absolute deviation of the first target node within a preset time window is selected from the node time deviation matrix and arranged in chronological order of acquisition time to form a sequence. This process can be achieved through indexing operations, extracting the corresponding data from the node time deviation matrix based on the first target node number and the preset time window length, and then reorganizing the data according to the acquisition time order.

[0056] Step S245: Extract all absolute deviations of the second target node within a preset time window from the node time deviation matrix, and arrange them in the order of acquisition time to form the second deviation sequence.

[0057] Similarly, the absolute deviation of the second target node within a preset time window is extracted from the node time deviation matrix and arranged into a sequence according to the acquisition time order. This process is similar to step S244, which also involves obtaining the data of the second target node from the node time deviation matrix through indexing operations and sorting it.

[0058] Step S246: Calculate the spatial correlation coefficient between the first deviation sequence and the second deviation sequence. The spatial correlation coefficient is used to characterize the overall correlation strength of the deviation between the two monitoring nodes within a preset time window.

[0059] There are several methods for calculating spatial correlation coefficients, such as the Pearson correlation coefficient method. The Pearson correlation coefficient is an indicator that measures the linear correlation between two variables. Its value ranges from -1 to 1; the closer the value is to 1, the stronger the positive correlation between the two sequences; the closer the value is to -1, the stronger the negative correlation; and the closer the value is to 0, the weaker the correlation. Specifically, the calculation involves first calculating the means of the first and second deviation sequences, then calculating the difference between each element in each sequence and the mean, then summing the products of these differences, and finally dividing this sum by the product of the standard deviations of the two sequences to obtain the Pearson correlation coefficient, which is the spatial correlation coefficient.

[0060] Step S247: Use the spatial correlation coefficient of the target node pair as the deviation transmission coefficient between nodes, arrange them in matrix form according to the order of the monitored node pairs, and obtain the spatial deviation transmission matrix.

[0061] The spatial correlation coefficient of each target node pair is used as the inter-node deviation transmission coefficient, and these coefficients are sequentially filled into the matrix according to the order of the monitored node pairs. The rows and columns of the matrix correspond to different monitored nodes, and the elements in the matrix are the inter-node deviation transmission coefficients between the corresponding monitored node pairs. This matrix can be stored using a two-dimensional array, and nested loops are used to assign each inter-node deviation transmission coefficient to its corresponding position in the matrix. This yields a complete spatial deviation transmission matrix, which reflects the deviation transmission relationship between any two monitored nodes.

[0062] Step S250: Perform time-dimensional evolution analysis on the node time deviation matrix, calculate the change trend coefficient of the absolute deviation of each monitoring node at two consecutive acquisition times, use the change trend coefficient as the time-time deviation evolution coefficient, and obtain the time deviation evolution matrix.

[0063] Temporal evolution analysis studies the trend of the absolute deviation of each monitoring node in the time-time deviation matrix over time. The trend coefficient measures the degree of change in the absolute deviation of each monitoring node between two consecutive acquisition times, and its use as the time-time deviation evolution coefficient reflects the rate of evolution of the deviation over time. The time-time deviation evolution matrix is ​​a two-dimensional matrix used to store the time-time deviation evolution coefficient of each monitoring node at each consecutive acquisition time.

[0064] As one implementation method, step S250 involves performing a time-dimensional evolution analysis on the node time deviation matrix, calculating the trend coefficient of the absolute deviation of each monitoring node at two consecutive acquisition times, and using the trend coefficient as the time-time deviation evolution coefficient to obtain the time deviation evolution matrix. Specifically, this can be implemented as the following steps S251~S256:

[0065] Step S251: Traverse each monitoring node in the preset monitoring grid and take the currently traversed monitoring node as the target time node.

[0066] Traversing each monitoring node in the preset monitoring grid is for analyzing the temporal evolution of each node. During the traversal, the currently processed monitoring node is defined as the target time node so that the absolute deviation of that node can be processed in the time dimension later. A loop structure can be used to implement the traversal process, starting from the first monitoring node and setting each monitoring node as the target time node for processing in turn, until all monitoring nodes have been traversed.

[0067] Step S252: Extract all absolute deviations of the target time node within a continuous preset duration from the node time deviation matrix, and arrange them in the order of collection time to form the target node time deviation sequence.

[0068] The absolute deviation of the target time node within a continuous preset duration is selected from the node time deviation matrix and arranged in chronological order of acquisition time to form a sequence. This process can be achieved through indexing operations, extracting the corresponding data from the node time deviation matrix based on the target time node number and reorganizing the data according to the acquisition time order.

[0069] Step S253: Traverse each acquisition time in the target node time deviation sequence except for the last acquisition time, take the currently traversed acquisition time as the current time, and take the next acquisition time of the current time as the next time.

[0070] The process involves sequentially accessing each acquisition moment in the target node's time deviation sequence, excluding the last acquisition moment. The currently processed acquisition moment is defined as the current moment, and the next acquisition moment is defined as the next moment. This is done to calculate the trend coefficient of the absolute deviation of each monitoring node between two consecutive acquisition moments. A loop structure can be used to iterate through each acquisition moment in the target node's time deviation sequence, starting from the first acquisition moment and sequentially setting each acquisition moment as the current moment for processing, until all acquisition moments except the last one have been traversed.

[0071] Step S254: Extract the absolute deviation of the target node time deviation sequence at the current time as the current deviation, and extract the absolute deviation at the next time as the next deviation.

[0072] The absolute deviations at the current and next moments are obtained from the target node's time deviation sequence and defined as the current deviation and the next deviation, respectively. This process can be achieved through indexing operations, extracting the corresponding data based on the positions of the current and next moments in the target node's time deviation sequence.

[0073] Step S255: Calculate the difference between the next deviation and the current deviation, and use the ratio of the difference to the current deviation as the trend coefficient. The trend coefficient is used to characterize the rate of evolution of the degree of deviation over time.

[0074] The difference is obtained by subtracting the current deviation from the next deviation, and then dividing this difference by the current deviation to obtain the trend coefficient. This coefficient reflects the degree of change in the absolute deviation of each monitoring node between two consecutive acquisition times, that is, the rate of evolution of the deviation over time.

[0075] Step S256: Arrange the trend coefficients of the target time node at continuous acquisition time into a matrix form according to the order of monitoring nodes and time, and obtain the time deviation evolution matrix.

[0076] The trend coefficients of each target time node at each continuously collected time point are sequentially filled into the matrix according to the order of the monitoring nodes and the time sequence. The rows of the matrix correspond to the monitoring nodes, the columns to the collection times, and the elements in the matrix are the deviation evolution coefficients of the corresponding monitoring node at the corresponding collection time. This matrix can be stored using a two-dimensional array, and nested loops are used to assign the deviation evolution coefficients between each time point to the corresponding positions in the matrix. This yields a complete time deviation evolution matrix, which reflects the evolution of the deviation of each monitoring node over time.

[0077] Step S260: Flatten the node time deviation matrix into a one-dimensional vector in row-major order to obtain the node time deviation vector; flatten the spatial deviation transfer matrix into a one-dimensional vector in row-major order to obtain the spatial deviation transfer vector; flatten the temporal deviation evolution matrix into a one-dimensional vector in row-major order to obtain the temporal deviation evolution vector; concatenate the node time deviation vector, spatial deviation transfer vector, and temporal deviation evolution vector in sequence to generate a multi-dimensional deviation feature vector. The dimension of the multi-dimensional deviation feature vector is the sum of the dimensions of the node time deviation vector, spatial deviation transfer vector, and temporal deviation evolution vector.

[0078] Flattening the matrix into a one-dimensional vector in row-major order converts the data in the matrix into a one-dimensional form, facilitating concatenation. Row-major order means arranging the elements of the matrix into a one-dimensional vector row by row. After flattening the node time-time deviation matrix, spatial deviation transitivity matrix, and temporal deviation evolution matrix into one-dimensional vectors respectively, these three vectors are concatenated in sequence to form a new multi-dimensional deviation feature vector. This vector contains deviation information at node time, spatial correlation information, and temporal evolution information, and its dimension is the sum of the dimensions of the three flattened vectors.

[0079] Step S300: Input the multidimensional deviation feature vector into the hierarchical anomaly pattern inference model, perform nonlinear feature enhancement and multi-scale anomaly pattern recognition on the multidimensional deviation feature vector, and obtain the preliminary confidence scores corresponding to each of the multiple candidate anomaly patterns.

[0080] The hierarchical anomaly pattern inference model is used to identify anomaly patterns in cold storage environments. It processes the input multi-dimensional deviation feature vector, extracts its nonlinear features, and identifies anomaly patterns at different scales. Nonlinear feature enhancement aims to extract the hidden nonlinear relationships in the multi-dimensional deviation feature vector, enabling the model to better capture the features of anomaly patterns. Multi-scale anomaly pattern recognition refers to identifying anomaly patterns at different scales to improve the accuracy and comprehensiveness of anomaly identification. The initial confidence score represents the probability of each candidate anomaly pattern.

[0081] As one implementation method, step S300 involves inputting the multidimensional deviation feature vector into the hierarchical anomaly pattern inference model, performing nonlinear feature enhancement and multi-scale anomaly pattern recognition on the multidimensional deviation feature vector, and obtaining the preliminary confidence scores corresponding to each of the multiple candidate anomaly patterns. Specifically, this can be implemented as follows: steps S310~S350:

[0082] Step S310: Input the multidimensional deviation feature vector into the feature enhancement layer of the hierarchical anomaly pattern inference model. The feature enhancement layer performs nonlinear mapping on the multidimensional deviation feature vector to obtain the enhanced deviation feature vector. The enhanced deviation feature vector contains the nonlinear correlation features after the original deviation features are mapped in a high-dimensional space.

[0083] The feature enhancement layer is a component of the hierarchical anomaly pattern inference model. Its function is to perform nonlinear mapping on the input multidimensional deviation feature vectors to enhance the correlation and expressive power between features. Nonlinear mapping maps the original multidimensional deviation feature vectors to a high-dimensional space, making the nonlinear relationships between features more apparent. The enhanced deviation feature vector is a new vector obtained after processing by the feature enhancement layer, containing the nonlinear correlation features of the original deviation features in the high-dimensional space.

[0084] As one implementation method, step S310 involves inputting the multidimensional deviation feature vector into the feature enhancement layer of the hierarchical anomaly pattern inference model. The feature enhancement layer performs nonlinear mapping processing on the multidimensional deviation feature vector to obtain the enhanced deviation feature vector. Specifically, this can be implemented as follows: steps S311~S313.

[0085] Step S311: Expand the feature dimension of the multidimensional deviation feature vector by adding feature cross terms and higher-order feature terms to expand the original dimension to a preset high-dimensional space to obtain a high-dimensional extended feature vector.

[0086] Feature dimension expansion aims to increase the expressive power of features. By adding feature interaction terms and higher-order feature terms, latent relationships between features can be uncovered. Feature interaction terms refer to new features obtained by combining different features; for example, the interaction term of two features x and y could be x×y. Higher-order feature terms refer to higher powers of features, such as x... 2 x 3 The pre-defined high-dimensional space is a pre-defined dimensional space. Extending multi-dimensional deviation feature vectors into this space can make the relationships between features more complex and richer. Feature dimension expansion can be achieved through methods such as polynomial feature transformation. By inputting multi-dimensional deviation feature vectors into a polynomial feature transformation function and setting the order of the polynomial, the function will automatically generate feature interaction terms and higher-order feature terms, thereby obtaining high-dimensional extended feature vectors.

[0087] Step S312: Perform nonlinear transformation on the high-dimensional extended feature vector. Use an activation function to perform nonlinear mapping on each feature component in the high-dimensional extended feature vector to enhance the nonlinear correlation between features and obtain the nonlinear transformed feature vector.

[0088] Nonlinear transformation is used to further enhance the nonlinear correlation between features, enabling the model to better capture features of anomalous patterns. Activation functions can include ReLU (Revised Linear Unit), Sigmoid, Tanh, etc. The high-dimensional extended feature vector is input into the activation function, which performs a nonlinear mapping on each feature component, resulting in a nonlinearly transformed feature vector. For example, when using the ReLU activation function, for the input feature component x, if x > 0, the output is x; if x <= 0, the output is 0.

[0089] Step S313: Perform feature filtering on the nonlinear transformation feature vector, and remove feature components that do not meet the requirements based on the feature importance score to obtain the enhanced deviation feature vector. The feature importance score is learned by the feature enhancement layer through historical training data.

[0090] Feature selection aims to remove unimportant feature components from nonlinear transformation feature vectors, reducing feature dimensionality and improving the model's computational efficiency and generalization ability. Feature importance scores, learned by the feature enhancement layer from historical training data, represent the importance of each feature component to the model's prediction results. Feature selection algorithms (such as recursive feature elimination and random forest feature importance evaluation) can be used to calculate feature importance scores, and feature components that do not meet the requirements can be removed based on a score threshold. For example, using a random forest algorithm to train on historical training data yields an importance score for each feature; features with scores below a certain threshold are then removed, ultimately resulting in an enhanced deviation feature vector.

[0091] Step S320: Input the enhanced deviation feature vector into the pattern recognition layer of the hierarchical anomaly pattern inference model. The pattern recognition layer performs multi-scale feature decomposition on the enhanced deviation feature vector to obtain deviation feature sub-vectors at different scales. Different scales correspond to deviation features at different time granularities.

[0092] The pattern recognition layer is another important component of the hierarchical anomaly pattern inference model. It is used to perform multi-scale feature decomposition on the enhanced deviation feature vectors to identify anomaly patterns at different scales. Multi-scale feature decomposition divides the enhanced deviation feature vectors according to different time granularities, resulting in deviation feature sub-vectors at different scales. Different scales correspond to different time ranges; for example, time can be divided into short-term, medium-term, and long-term time scales. The deviation feature sub-vectors at each scale reflect the anomaly pattern features within different time ranges.

[0093] As one implementation method, in step S320, the enhanced deviation feature vector is decomposed into multiple scales by the pattern recognition layer to obtain deviation feature sub-vectors at different scales. Specifically, this can be implemented as follows: S321~S326:

[0094] Step S321: Based on the historical normal operation data of cold storage environmental parameters, extract the fluctuation frequency characteristics of the environmental parameter sequence and generate a frequency feature spectrum. The frequency feature spectrum contains multiple feature frequency values, and each feature frequency value corresponds to the typical fluctuation frequency of the environmental parameters under normal conditions.

[0095] Extracting the fluctuation frequency characteristics of environmental parameter sequences is crucial for understanding the fluctuation patterns of cold storage environmental parameters under normal operating conditions. A frequency characteristic spectrum is a set of multiple characteristic frequency values, each corresponding to a typical fluctuation frequency of the environmental parameter under normal conditions. Signal processing techniques (such as Fourier transform) can be used to process historical normal operating data of cold storage environmental parameters, converting the time-domain environmental parameter sequence into a frequency-domain frequency characteristic spectrum. Fourier transform decomposes a signal into a superposition of sine and cosine waves of different frequencies; by analyzing the amplitude and phase of these frequency components, the fluctuation frequency characteristics of the environmental parameters can be obtained.

[0096] Step S322: Calculate the corresponding time window length based on the characteristic frequency value in the frequency characteristic spectrum. The time window length is the product of the reciprocal of the characteristic frequency value and the preset sampling interval, resulting in multiple different target window lengths. The length of each target window increases as the corresponding characteristic frequency value decreases.

[0097] Calculating the time window length based on the characteristic frequency values ​​in the frequency spectrum is to determine the time range at different scales. The time window length is the product of the reciprocal of the characteristic frequency value and the preset sampling interval. This is because frequency and time are interrelated; the higher the frequency, the shorter the corresponding time period; and the lower the frequency, the longer the corresponding time period. The preset sampling interval refers to the time interval between two adjacent samples when collecting environmental parameters. By calculating the time window lengths corresponding to different characteristic frequency values, multiple different target window lengths are obtained. These target window lengths reflect the time granularity at different scales. For example, if a certain characteristic frequency value is f and the preset sampling interval is t, then the corresponding time window length is 1 / f*t.

[0098] Step S323: Traverse multiple different target window lengths and use the current target window length as the current decomposition window.

[0099] Traversing multiple target window lengths allows for the decomposition of the augmented deviation feature vectors at different scales. During the traversal, the currently processed target window length is defined as the current decomposition window, enabling subsequent processing of the augmented deviation feature vectors based on this window length. This traversal can be implemented using a loop, starting with the first target window length and sequentially setting each target window length as the current decomposition window for processing, until all target window lengths have been traversed.

[0100] Step S324: Perform continuous non-overlapping truncation of the enhanced deviation feature vector according to the current decomposition window to obtain multiple window feature segments. Each window feature segment contains the enhanced deviation feature component corresponding to the current decomposition window.

[0101] The purpose of performing continuous, non-overlapping truncation of the augmented deviation feature vector according to the current decomposition window is to divide the augmented deviation feature vector into multiple window feature segments, each corresponding to a time range. Continuous, non-overlapping truncation means that there is no overlap between the truncated window feature segments. This truncation process can be implemented through indexing operations, sequentially extracting feature components of appropriate length from the augmented deviation feature vector according to the length of the current decomposition window to obtain multiple window feature segments.

[0102] Step S325: Perform feature statistical aggregation processing on each window feature segment, calculate the central trend feature value and dispersion feature value of each enhanced deviation feature component within the window feature segment, and combine the central trend feature value and dispersion feature value into window aggregated features.

[0103] Feature statistical aggregation is used to statistically analyze the enhanced deviation feature components within each window feature segment and extract key features. Central tendency features reflect the concentration of feature components within a window feature segment; common central tendency features include the mean and median. Dispersion features reflect the dispersion of feature components within a window feature segment; common dispersion features include standard deviation and variance. Combining central tendency and dispersion features forms a window aggregation feature, which better represents the characteristics of the window feature segment. For example, for a window feature segment, its mean and standard deviation can be calculated, and these two values ​​can be combined into a two-dimensional vector as the window aggregation feature for that feature segment.

[0104] Step S326: Arrange all window aggregation features under the same current decomposition window in chronological order to form a deviation feature sub-vector of the corresponding scale. The dimension of the deviation feature sub-vector is positively correlated with the length of the current decomposition window and the number of feature value types processed by feature statistical aggregation.

[0105] All window aggregated features under the same current decomposition window are arranged in chronological order to form a new vector, namely the deviation feature sub-vector at the corresponding scale. This vector reflects the feature changes of the enhanced deviation feature vector at the current scale. The dimension of the deviation feature sub-vector is positively correlated with the current decomposition window length and the number of feature value types processed by feature statistical aggregation. This is because the longer the current decomposition window length, the more window feature segments are extracted, and the more window aggregated features there are; the more feature value types processed by feature statistical aggregation, the larger the dimension of each window aggregated feature, thus leading to an increase in the dimension of the deviation feature sub-vector. For example, if the current decomposition window length is n and the number of feature value types processed by feature statistical aggregation is k, then the dimension of the deviation feature sub-vector at the corresponding scale is n*k.

[0106] Step S330: Based on the preset abnormal pattern library, match the deviation feature sub-vectors at each scale with the abnormal pattern feature templates in the abnormal pattern library, and calculate the matching degree between each deviation feature sub-vector and each abnormal pattern feature template.

[0107] The pre-defined anomaly pattern library is a database containing multiple anomaly pattern feature templates. Each anomaly pattern feature template corresponds to a typical type of cold storage environment anomaly and includes standard feature sub-vectors at different scales. Matching the deviation feature sub-vectors at each scale with the anomaly pattern feature templates in the anomaly pattern library is to identify possible anomaly patterns in the current cold storage environment. The matching degree represents the similarity between the deviation feature sub-vectors and the anomaly pattern feature templates.

[0108] As one implementation method, step S330 involves matching the deviation feature vectors at each scale with the abnormal pattern feature templates in the abnormal pattern library based on a preset abnormal pattern library, and calculating the matching degree between each deviation feature vector and each abnormal pattern feature template. Specifically, this can be implemented as the following steps S331~S335:

[0109] Step S331: Obtain a preset abnormal pattern library. The abnormal pattern library contains multiple abnormal pattern feature templates. Each abnormal pattern feature template corresponds to a typical type of cold storage environment abnormality and contains standard feature sub-vectors at different scales.

[0110] Obtaining the preset anomaly pattern library involves loading a database storing anomaly pattern feature templates into memory for subsequent matching operations. These anomaly pattern feature templates are derived from the analysis and summarization of a large amount of historical cold storage environmental anomaly data. Each template corresponds to a typical anomaly type, such as refrigeration equipment failure or ventilation system anomaly. Each template contains standard feature sub-vectors at different scales, reflecting the characteristic performance of that anomaly type at different time granularities.

[0111] Step S332: Traverse the deviation feature vectors at different scales, and take the currently traversed deviation feature vector as the target scale vector, whose corresponding scale is the target scale.

[0112] Traversing the deviation feature vectors at different scales is for feature matching analysis at each scale. During the traversal, the deviation feature vector currently being processed is defined as the target scale subvector, and its corresponding scale is the target scale. A loop structure can be used to implement the traversal process, starting with the deviation feature vector at the first scale, and sequentially setting each subvector as the target scale subvector for processing, until all scales of deviation feature vectors have been traversed.

[0113] Step S333: Extract the standard feature sub-vector of each abnormal pattern feature template at the target scale from the abnormal pattern library, and use it as the target standard sub-vector.

[0114] The standard feature vectors of each anomaly pattern feature template at the target scale are extracted from the anomaly pattern library and used as target standard vectors for matching with target scale vectors. Standard feature vectors of the corresponding scale can be selected from the anomaly pattern library based on the target scale information and stored in a list or array to form a set of target standard vectors.

[0115] Step S334: Calculate the vector similarity between the target scale sub-vector and each target standard sub-vector. The vector similarity is calculated using the vector cosine similarity algorithm.

[0116] The vector cosine similarity algorithm measures the similarity between two vectors by calculating the cosine of the angle between them. A cosine value closer to 1 indicates greater similarity, while a value closer to 0 indicates less similarity. The target scale sub-vector and each target standard sub-vector are input into the vector cosine similarity algorithm to calculate their similarity value. Specifically, the dot product of the two vectors is calculated first, and then the dot product is divided by the product of the magnitudes of the two vectors to obtain the cosine similarity value.

[0117] Step S335: Use the vector similarity as the matching degree between the target scale sub-vector and the corresponding abnormal pattern feature template, and store the matching degree for subsequent preliminary confidence calculation.

[0118] The calculated vector similarity is used as the matching degree between the target scale sub-vector and the corresponding anomaly pattern feature template, and these matching degree values ​​are stored for subsequent calculation of the preliminary confidence of each candidate anomaly pattern. A two-dimensional array or dictionary can be used to store the matching degree values, where rows correspond to different anomaly pattern feature templates and columns correspond to deviation feature sub-vectors at different scales.

[0119] Step S340: Determine the scale confidence of each candidate anomaly pattern at different scales based on the matching degree, and perform a weighted average of the scale confidence of the same candidate anomaly pattern at different scales to obtain the preliminary confidence of the candidate anomaly pattern.

[0120] Scale confidence represents the probability of each candidate anomaly pattern at different scales, and it is determined based on the matching degree. A higher matching degree results in a higher scale confidence. Weighted averaging of the scale confidence of the same candidate anomaly pattern at different scales is used to comprehensively consider information from different scales and obtain a more accurate preliminary confidence. A weight value can be assigned to each scale, reflecting its importance in anomaly identification. The scale confidence at each scale is multiplied by its corresponding weight value, then summed and divided by the total weight to obtain the preliminary confidence of the candidate anomaly pattern.

[0121] Step S350: Integrate the preliminary confidence scores of all candidate anomaly patterns into a preliminary confidence vector, where each element of the preliminary confidence vector corresponds to the preliminary confidence score of a candidate anomaly pattern.

[0122] The initial confidence scores of all candidate anomaly patterns are arranged sequentially into a vector, known as the initial confidence vector. Each element of this vector corresponds to the initial confidence score of a candidate anomaly pattern. This vector provides a clear understanding of the likelihood of each candidate anomaly pattern. A one-dimensional array can be used to store the initial confidence vector, where the array index corresponds to the candidate anomaly pattern number, and the array element corresponds to the initial confidence score of the corresponding candidate anomaly pattern.

[0123] Step S400: Based on the preliminary confidence of multiple candidate anomaly patterns, a comprehensive anomaly quantification index is generated through a spatiotemporal correlation weighted fusion algorithm. The comprehensive anomaly quantification index is used to characterize the severity and spread trend of the current cold storage environment anomaly.

[0124] The spatiotemporal correlation weighted fusion algorithm is used to comprehensively consider the preliminary confidence levels and spatiotemporal correlation information of multiple candidate anomaly patterns. It can fuse the information of multiple candidate anomaly patterns to generate a comprehensive quantitative index. The comprehensive anomaly quantitative index is a numerical value whose magnitude is positively correlated with the severity and propagation trend of the current cold storage environment anomaly. This index provides a direct understanding of the anomaly situation in the cold storage environment.

[0125] As one implementation method, step S400, based on the preliminary confidence of multiple candidate anomaly patterns, generates a comprehensive anomaly quantification index through a spatiotemporal correlation weighted fusion algorithm, which can be specifically implemented as the following steps S410~S460:

[0126] Step S410: Extract the spatial deviation transmission vector and the temporal deviation evolution vector from the multidimensional deviation feature vector. Reconstruct the spatial deviation transmission vector into a spatial deviation transmission matrix according to the original matrix dimensions, and reconstruct the temporal deviation evolution vector into a temporal deviation evolution matrix according to the original matrix dimensions.

[0127] Extracting the spatial deviation transfer vector and temporal deviation evolution vector from the multidimensional deviation eigenvector is to recover the previously calculated spatial and temporal dimensional information. In step S260, the spatial deviation transfer matrix and temporal deviation evolution matrix are flattened into one-dimensional vectors and concatenated to form the multidimensional deviation eigenvector. Now, these need to be separated and reconstructed according to the original matrix dimensions. Based on the previous flattening order and dimensional information, corresponding parts can be extracted from the multidimensional deviation eigenvector to obtain the spatial deviation transfer vector and temporal deviation evolution vector, respectively. Then, based on the number of rows and columns of the original matrix, these two vectors are rearranged into a matrix form to obtain the spatial deviation transfer matrix and the temporal deviation evolution matrix.

[0128] Step S420: Calculate the spatial weight coefficient corresponding to each candidate anomaly pattern based on the spatial deviation transfer matrix. The spatial weight coefficient is positively correlated with the spatial propagation range of the anomaly pattern.

[0129] The spatial weighting coefficient reflects the spatial propagation range and influence of each candidate anomaly pattern, and it is calculated based on the spatial deviation transfer matrix. The spatial deviation transfer matrix stores the inter-node deviation transfer coefficients between any two monitoring nodes. By analyzing these coefficients, the spatial propagation of anomaly patterns can be understood.

[0130] As one implementation method, step S420, calculating the spatial weight coefficients corresponding to each candidate anomaly mode based on the spatial deviation transfer matrix, can be specifically implemented as the following steps S421~S424:

[0131] Step S421: Extract the inter-node deviation transmission coefficient of each monitoring node pair from the spatial deviation transmission matrix at all acquisition times, calculate the average deviation transmission coefficient of each monitoring node pair, and obtain the node pair average transmission coefficient matrix.

[0132] The deviation transfer coefficients between nodes for each monitoring node pair are extracted from the spatial deviation transfer matrix at all acquisition times. These coefficients are then averaged to obtain the average deviation transfer coefficient for each monitoring node pair. Alternatively, a loop can be used to iterate through each element of the spatial deviation transfer matrix, summing the deviation transfer coefficients between nodes for the same monitoring node pair at different acquisition times and dividing by the number of acquisition times to obtain the average deviation transfer coefficient. Finally, the average deviation transfer coefficients of all monitoring node pairs are arranged in matrix form to obtain the node pair average transfer coefficient matrix.

[0133] Step S422: For each candidate anomaly pattern, obtain the typical impact area corresponding to the anomaly pattern from the preset anomaly pattern library. The typical impact area contains multiple monitoring nodes.

[0134] In addition to containing anomalous pattern feature templates, the pre-defined anomalous pattern library also stores information on the typical impact area for each anomalous pattern. The typical impact area refers to the range of monitoring nodes that the anomalous pattern typically affects within the cold storage facility. For each candidate anomalous pattern, its corresponding typical impact area information is extracted from the anomalous pattern library. This information can be stored as a list or set, containing the numbers or locations of multiple monitoring nodes.

[0135] Step S423: Based on the node pair average transmission coefficient matrix, calculate the sum of the average deviation transmission coefficients of all monitoring node pairs in the typical influence area, as the sum of regional spatial correlations.

[0136] Based on the information of typical influence areas, the average deviation transmission coefficients of all monitoring node pairs within the typical influence area are extracted from the average transmission coefficient matrix of node pairs. These coefficients are then summed to obtain the total spatial correlation of the region. Nested loops can be used to iterate through each monitoring node pair within the typical influence area, obtaining the corresponding average deviation transmission coefficients from the average transmission coefficient matrix of node pairs and accumulating them. For example, if the typical influence area contains nodes A, B, and C, then the sum of the average deviation transmission coefficients of node pairs (A,B), (A,C), and (B,C) needs to be calculated.

[0137] Step S424: The ratio of the total regional spatial correlation to the preset spatial correlation benchmark value is used as the spatial weight coefficient of the candidate anomaly mode. The spatial correlation benchmark value is the sum of the average deviation transmission coefficients of all monitoring node pairs under normal operating conditions.

[0138] The preset spatial correlation benchmark value is the sum of the average deviation transmission coefficients of all monitoring node pairs under normal operating conditions of the cold storage, obtained through analysis of historical normal operating data. Dividing the sum of regional spatial correlations by the spatial correlation benchmark value yields the spatial weight coefficient of the candidate anomaly pattern. A larger spatial weight coefficient indicates a wider spatial propagation range and a greater degree of impact for the anomaly pattern.

[0139] Step S430: Calculate the time weight coefficient corresponding to each candidate anomalous mode based on the time deviation evolution matrix. The time weight coefficient is positively correlated with the evolution speed of the anomalous mode over time.

[0140] The time weighting coefficient reflects the evolution speed and development trend of each candidate anomaly pattern over time, and it is calculated based on the time deviation evolution matrix. The time deviation evolution matrix stores the time deviation evolution coefficient of each monitoring node at each consecutive acquisition time. By analyzing these coefficients, we can understand the changes of the anomaly pattern over time.

[0141] As one implementation method, step S430, calculating the time weight coefficient corresponding to each candidate anomaly mode based on the time deviation evolution matrix, can be specifically implemented as the following steps S431~S438:

[0142] Step S431: Extract the deviation evolution coefficient of each monitoring node from the time deviation evolution matrix at all consecutive acquisition times, and arrange them in the order of monitoring nodes and time to form a node time evolution sequence.

[0143] The deviation evolution coefficients of each monitoring node across all consecutive acquisition times are extracted from the time deviation evolution matrix. These coefficients are then arranged into a sequence according to the order of the monitoring nodes and the time sequence, i.e., the node time evolution sequence. A loop structure can be used to iterate through each element in the time deviation evolution matrix, extracting the deviation evolution coefficients of the same monitoring node at different acquisition times in chronological order. Finally, the sequences of all monitoring nodes are concatenated in the order of the monitoring nodes to obtain the node time evolution sequence.

[0144] Step S432: Collect a dataset of historical environmental anomaly cases in the cold storage. The dataset contains multiple historical anomaly events, and each historical anomaly event corresponds to a complete sequence of time-to-time deviation evolution coefficients and the duration of the anomaly.

[0145] The purpose of collecting a historical environmental anomaly case dataset for cold storage facilities is to analyze the temporal evolution of anomaly patterns. This dataset contains information on multiple historical anomaly events, each corresponding to a complete sequence of time-series deviation evolution coefficients and the duration of the anomaly. By filtering and organizing historical operational data from the cold storage facility, the time periods in which the anomalies occurred can be extracted, and the time-series deviation evolution coefficients and the duration of the anomaly for each monitoring node within those periods can be recorded. This information is then stored in a database or file to form the historical environmental anomaly case dataset.

[0146] Step S433: Perform statistical analysis on the duration of all anomalies in the historical environmental anomaly case dataset to generate a duration distribution model. The duration distribution model is used to characterize the probability distribution characteristics of the duration of anomalies corresponding to different anomaly types.

[0147] Statistical analysis is performed on the duration of all anomalies in the historical environmental anomaly case dataset, using statistical methods (such as histograms and kernel density estimation) to generate a duration distribution model. The duration distribution model can be a probability density function or a cumulative distribution function, describing the probability distribution characteristics of the duration of anomalies corresponding to different anomaly types. For example, kernel density estimation can be used to smooth the anomaly duration data, resulting in a continuous probability density function that reflects the probability of occurrence of different anomaly durations.

[0148] Step S434: For each candidate anomaly pattern, extract the probability distribution features corresponding to the anomaly pattern from the duration distribution model, and determine the duration of the anomaly with the highest probability density as the typical evolution duration.

[0149] For each candidate anomaly pattern, its corresponding probability distribution features are extracted from the duration distribution model, representing the probability of the anomaly pattern occurring under different anomaly durations. Then, by analyzing the probability distribution features, the anomaly duration with the highest probability density is determined and used as the typical evolution duration of that candidate anomaly pattern. Optimization algorithms or search algorithms can be used to find the point with the highest probability density; for example, the probability density function can be iterated to find the anomaly duration corresponding to the maximum value.

[0150] Step S435: Traverse each monitoring node in the node time evolution sequence and take the currently traversed monitoring node as the time analysis node.

[0151] The process iterates through each monitoring node in the node time evolution sequence, defining the currently processed monitoring node as a time analysis node for subsequent analysis of its time evolution. A loop can be used to implement this process, starting with the first monitoring node and sequentially setting each node as a time analysis node for processing, until all monitoring nodes have been traversed.

[0152] Step S436: Extract an evolution coefficient segment with the same length as the typical evolution duration from the node time evolution sequence of the time analysis node, calculate the cumulative sum of the deviations from the evolution coefficients at all times in the evolution coefficient segment, and obtain the node time cumulative value.

[0153] Based on typical evolution duration information, evolution coefficient segments of appropriate length are extracted from the node time evolution sequences of the time analysis nodes. Slicing operations can be used to extract subsequences of a specified length from the node time evolution sequences. Then, the inter-time deviation evolution coefficients of all time points in the evolution coefficient segment are summed to obtain the node time cumulative value. For example, if the typical evolution duration is n acquisition times, a subsequence of length n is extracted from the node time evolution sequence, and the inter-time deviation evolution coefficients of all time points in the subsequence are summed to obtain the node time cumulative value.

[0154] Step S437: Calculate the average of the cumulative node time values ​​of all monitored nodes as the sum of time correlations for abnormal modes.

[0155] The sum of the cumulative node times of all monitored nodes is obtained by adding them together and dividing by the number of monitored nodes. This sum reflects the overall temporal correlation of the candidate anomaly pattern. Alternatively, a loop can be used to iterate through the cumulative node times of all monitored nodes, sum them, and divide by the total number of monitored nodes to obtain the sum of the temporal correlations of the anomaly pattern.

[0156] Step S438: The ratio of the sum of time correlations of the abnormal modes to the preset time correlation benchmark value is used as the time weight coefficient of the candidate abnormal mode. The preset time benchmark value is the average value of the cumulative sum of the time deviation evolution coefficients of coefficient segments of the same length in the historical normal operation data.

[0157] The preset time correlation benchmark is the average of the cumulative sum of the deviations of coefficient segments of the same length over time in the historical normal operation data of the cold storage. Dividing the sum of the time correlations of the anomaly patterns by the time correlation benchmark yields the time weight coefficient of that candidate anomaly pattern. A larger time weight coefficient indicates a faster evolution rate and a greater impact of the anomaly pattern over time.

[0158] Step S440: Multiply the initial confidence level of each candidate anomaly pattern by the corresponding spatial weight coefficient and temporal weight coefficient to obtain the weighted confidence level.

[0159] The initial confidence score of each candidate anomaly pattern is multiplied by its corresponding spatial and temporal weighting coefficients to obtain the weighted confidence score. The weighted confidence score comprehensively considers factors such as the initial probability of the anomaly pattern, its spatial propagation range, and its temporal evolution speed, thus more accurately reflecting the actual probability of each candidate anomaly pattern. A loop structure can be used to iterate through each candidate anomaly pattern, multiplying its initial confidence score, spatial weighting coefficient, and temporal weighting coefficient to obtain the weighted confidence score.

[0160] Step S450: Sum the weighted confidence scores of all candidate anomaly patterns to obtain the confidence fusion value.

[0161] The weighted confidence scores of all candidate anomaly patterns are summed to obtain the combined confidence score. The combined confidence score is the overall confidence level of all candidate anomaly patterns, reflecting the overall probability of anomalies existing in the current cold storage environment. A loop can be used to iterate through the weighted confidence scores of all candidate anomaly patterns and sum them to obtain the combined confidence score.

[0162] Step S460: Based on the confidence fusion value and the preset anomaly severity grading standard, calculate the comprehensive anomaly quantification index. The value of the comprehensive anomaly quantification index is positively correlated with the anomaly severity.

[0163] The preset anomaly severity grading standard is a pre-defined rule that maps different confidence level fusion value ranges to different anomaly severity levels. Based on the confidence level fusion value and the anomaly severity grading standard, a comprehensive anomaly quantification index is calculated. The confidence level fusion value can be mapped to a numerical range that is positively correlated with the anomaly severity. For example, a higher confidence level fusion value indicates a greater probability of an anomaly, and the corresponding comprehensive anomaly quantification index is also higher. Piecewise functions or linear mapping methods can be used to convert the confidence level fusion value into a comprehensive anomaly quantification index.

[0164] Step S500: Output the cold storage environment anomaly detection results based on the comprehensive anomaly quantification index. The cold storage environment anomaly detection results include anomaly type judgment and anomaly source location information.

[0165] The comprehensive anomaly quantification index reflects the severity and propagation trend of current cold storage environmental anomalies. Based on this index, anomaly type identification and anomaly source location can be performed. Anomaly type identification is determined based on the comprehensive anomaly quantification index and the preliminary and weighted confidence scores of previously calculated candidate anomaly patterns, selecting the candidate anomaly pattern with the highest confidence score as the likely anomaly type. Anomaly source location can be achieved by combining information from the spatial deviation transmission matrix and the temporal deviation evolution matrix to analyze the spatial and temporal propagation path of the anomaly, identifying the monitoring node or area where the anomaly first appeared as the possible location of the anomaly source. The anomaly type identification results and anomaly source location information are compiled into cold storage environmental anomaly detection results and output, enabling relevant personnel to take timely measures to address the issue.

[0166] As one implementation method, step S500 outputs the cold storage environment anomaly detection result based on the comprehensive anomaly quantification index. The cold storage environment anomaly detection result includes anomaly type judgment and anomaly source location information. Specifically, it can be implemented as follows: compare the comprehensive anomaly quantification index with a preset anomaly judgment threshold. When the comprehensive anomaly quantification index is greater than the preset anomaly judgment threshold, start the anomaly type judgment and anomaly source location process; based on the preliminary confidence of multiple candidate anomaly patterns, select candidate anomaly patterns with a preliminary confidence greater than a preset confidence threshold as high-probability anomaly patterns, and determine the candidate anomaly pattern with the highest preliminary confidence among the high-probability anomaly patterns as the anomaly type of the current cold storage environment; extract the feature components of the spatial deviation transmission dimension from the multi-dimensional deviation feature vector to obtain the node deviation transmission coefficient of each monitoring node pair, select monitoring node pairs with a node deviation transmission coefficient greater than a preset transmission threshold as strongly correlated node pairs, and based on the strongly correlated nodes... A point-to-point anomaly propagation correlation network is constructed, where nodes are monitoring nodes and edges represent the connections between strongly correlated node pairs. Feature components representing the temporal deviation evolution dimension are extracted from the multidimensional deviation feature vector to obtain the time-to-time deviation evolution coefficients of each monitoring node. The first acquisition time where the time-to-time deviation evolution coefficient exceeds a preset evolution threshold is determined as the anomaly initiation time. The absolute deviation of all monitoring nodes at the anomaly initiation time is extracted. The absolute deviation of each monitoring node at the anomaly initiation time is input into the anomaly propagation correlation network to calculate the propagation probability of each monitoring node as an anomaly source. The propagation probability is the sum of the products of the absolute deviation of the monitoring node at the anomaly initiation time and the deviation transmission coefficients between all strongly correlated node pairs in the anomaly propagation correlation network. The monitoring node with the highest propagation probability is identified as the anomaly source monitoring node in the anomaly source location information. The anomaly type and the anomaly source monitoring node are integrated to generate anomaly detection results for the cold storage environment.

[0167] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0168] Figure 2 A hardware entity diagram of a computer system provided as an embodiment of the present invention, such as... Figure 2As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0169] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0170] When processor 1001 executes a program, it implements the steps of any of the above-mentioned machine learning-based cold storage environment anomaly detection methods. Processor 1001 typically controls the overall operation of computer system 1000.

[0171] This invention provides a computer storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the machine learning-based cold storage environment anomaly detection method as described in any of the above embodiments.

[0172] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based method for detecting anomalies in a cold storage environment, characterized in that, include: Obtain a spatiotemporal correlation benchmark model of the cold storage environment; Collect the current cold storage environment parameter sequence, and calculate the spatiotemporal coordination deviation of the current cold storage environment parameter sequence relative to the environmental parameter sequence under normal operation based on the spatiotemporal correlation benchmark model. Generate a multidimensional deviation feature vector based on the spatiotemporal coordination deviation. The multidimensional deviation feature vector is input into the hierarchical anomaly pattern inference model, and nonlinear feature enhancement and multi-scale anomaly pattern recognition are performed on the multidimensional deviation feature vector to obtain the preliminary confidence of each of the multiple candidate anomaly patterns. Based on the preliminary confidence of the multiple candidate anomaly patterns, a comprehensive anomaly quantification index is generated through a spatiotemporal correlation weighted fusion algorithm. The comprehensive anomaly quantification index is used to characterize the severity and propagation trend of the current cold storage environment anomaly. Based on the comprehensive anomaly quantification index, the cold storage environment anomaly detection result is output, which includes anomaly type judgment and anomaly source location information. Based on the spatiotemporal correlation benchmark model, a normal environmental parameter benchmark sequence corresponding to the current cold storage environmental parameter sequence is extracted. The normal environmental parameter benchmark sequence includes the normal parameter reference values ​​of each monitoring node in the preset monitoring grid within a continuous preset time period. The normal parameter reference values ​​are predicted and generated by the spatiotemporal correlation benchmark model based on historical normal operation data. Calculate the absolute deviation between the environmental parameter value of each monitoring node in the current cold storage environmental parameter sequence at each collection time and the normal parameter reference value of the corresponding monitoring node and the corresponding collection time in the normal environmental parameter baseline sequence, and obtain the node time deviation matrix. The row dimension of the node time deviation matrix corresponds to the monitoring node, and the column dimension corresponds to the collection time. Spatial dimension correlation analysis is performed on the node time deviation matrix. Based on the preset time window length, the absolute deviation sequence of each monitoring node is extracted. The spatial correlation coefficient between the absolute deviation sequences of any two monitoring nodes is calculated. The spatial correlation coefficient is used as the deviation transmission coefficient between nodes to obtain the spatial deviation transmission matrix. A time-dimensional evolution analysis is performed on the node time deviation matrix. The change trend coefficient of the absolute deviation of each monitoring node at two consecutive acquisition times is calculated. The change trend coefficient is used as the time-time deviation evolution coefficient to obtain the time deviation evolution matrix. Flatten the node time deviation matrix into a one-dimensional vector in row-major order to obtain the node time deviation vector; The spatial deviation transfer matrix is ​​flattened into a one-dimensional vector in row-major order to obtain the spatial deviation transfer vector; the temporal deviation evolution matrix is ​​flattened into a one-dimensional vector in row-major order to obtain the temporal deviation evolution vector; the node time deviation vector, the spatial deviation transfer vector, and the temporal deviation evolution vector are concatenated in sequence to generate a multi-dimensional deviation feature vector, the dimension of which is the sum of the dimensions of the node time deviation vector, the spatial deviation transfer vector, and the temporal deviation evolution vector.

2. The method according to claim 1, characterized in that, The process involves collecting the current cold storage environmental parameter sequence and calculating the spatiotemporal coordination deviation of the current cold storage environmental parameter sequence relative to the environmental parameter sequence under normal operating conditions based on the spatiotemporal correlation benchmark model. A multidimensional deviation feature vector is then generated based on the spatiotemporal coordination deviation, including: Collect the current cold storage environmental parameter sequence, which includes the environmental parameter values ​​of each monitoring node in the preset monitoring grid within the cold storage within a continuous preset time period. The environmental parameter values ​​are sequence data collected continuously at fixed time intervals.

3. The method according to claim 2, characterized in that, The absolute deviation between the environmental parameter value of each monitoring node in the current cold storage environmental parameter sequence at each acquisition time and the normal parameter reference value of the corresponding monitoring node and the corresponding acquisition time in the normal environmental parameter baseline sequence is calculated to obtain the node-time deviation matrix, including: Traverse each monitoring node in the preset monitoring grid and take the currently traversed monitoring node as the target monitoring node; Extract the environmental parameter values ​​of the target monitoring node within the continuous preset time period from the current cold storage environmental parameter sequence, and arrange them in the order of collection time to form a target node parameter subsequence; Extract the normal parameter reference values ​​of the target monitoring node within the continuous preset time period from the normal environmental parameter baseline sequence, and arrange them in the order of collection time to form the target node baseline subsequence; Iterate through each acquisition time in the target node parameter subsequence and take the currently traversed acquisition time as the target acquisition time; The absolute deviation of the target node at the target acquisition time is obtained by calculating the absolute value of the difference between the environmental parameter value of the target node parameter subsequence at the target acquisition time and the normal parameter reference value of the target node baseline subsequence at the target acquisition time. The absolute deviations of all monitoring nodes at all acquisition times are arranged in matrix form according to the order of monitoring nodes and acquisition times to obtain the node time deviation matrix.

4. The method according to claim 2, characterized in that, The step of performing spatial dimension correlation analysis on the node time deviation matrix involves extracting the absolute deviation sequence of each monitoring node based on a preset time window length, calculating the spatial correlation coefficient between the absolute deviation sequences of any two monitoring nodes, and using the spatial correlation coefficient as the deviation transmission coefficient between nodes to obtain the spatial deviation transmission matrix, including: Based on the regular adjustment cycle of cold storage equipment and the sampling interval of environmental parameters, a preset time window length is set. The preset time window length is a set of multiple consecutive sampling moments, and the preset time window length is not less than two sampling moments. Determine the spatial adjacency relationship of each monitoring node in the preset monitoring grid, and construct a node adjacency list based on the spatial adjacency relationship. The node adjacency list includes each monitoring node and its adjacent monitoring nodes. Traverse each monitoring node pair in the node adjacency list, and take the currently traversed monitoring node pair as the target node pair, wherein the target node pair includes a first target node and a second target node; Extract all absolute deviations of the first target node within the preset time window length from the node time deviation matrix, and arrange them in the order of collection time to form a first deviation sequence; Extract all absolute deviations of the second target node within the preset time window length from the node time deviation matrix, and arrange them in the order of collection time to form a second deviation sequence; Calculate the spatial correlation coefficient between the first deviation sequence and the second deviation sequence. The spatial correlation coefficient is used to characterize the overall correlation strength of the deviation between the two monitoring nodes within a preset time window. The spatial correlation coefficient of the target node pair is used as the deviation transmission coefficient between nodes, and the nodes are arranged in matrix form according to the order of the monitored node pairs to obtain the spatial deviation transmission matrix.

5. The method according to claim 2, characterized in that, The step involves performing a time-dimensional evolution analysis on the node time-time deviation matrix, calculating the trend coefficient of the absolute deviation of each monitoring node over two consecutive acquisition times, and using this trend coefficient as the time-time deviation evolution coefficient to obtain the time deviation evolution matrix, including: Traverse each monitoring node in the preset monitoring grid and take the currently traversed monitoring node as the target time node; Extract all absolute deviations of the target time node within the continuous preset duration from the node time deviation matrix, and arrange them in the order of collection time to form a target node time deviation sequence; Iterate through each acquisition time in the target node time deviation sequence except for the last acquisition time, take the currently traversed acquisition time as the current time, and take the next acquisition time of the current time as the next time. Extract the absolute deviation of the target node time deviation sequence at the current time as the current deviation, and extract the absolute deviation at the next time as the next deviation; Calculate the difference between the next deviation and the current deviation, and use the ratio of the difference to the current deviation as a trend coefficient. The trend coefficient is used to characterize the rate of evolution of the degree of deviation over time. The change trend coefficients of the target time node at continuous acquisition time are arranged in matrix form according to the order of monitoring nodes and time, to obtain the time deviation evolution matrix.

6. The method according to claim 1, characterized in that, The process involves inputting the multidimensional deviation feature vector into a hierarchical anomaly pattern inference model, performing nonlinear feature enhancement and multi-scale anomaly pattern recognition on the multidimensional deviation feature vector, and obtaining the preliminary confidence scores corresponding to multiple candidate anomaly patterns, including: The multidimensional deviation feature vector is input into the feature enhancement layer of the hierarchical anomaly pattern inference model. The feature enhancement layer performs nonlinear mapping on the multidimensional deviation feature vector to obtain the enhanced deviation feature vector. The enhanced deviation feature vector contains the nonlinear correlation features after the original deviation features are mapped to a high-dimensional space. The enhanced deviation feature vector is input into the pattern recognition layer of the hierarchical anomaly pattern inference model. The pattern recognition layer performs multi-scale feature decomposition on the enhanced deviation feature vector to obtain deviation feature sub-vectors at different scales. The different scales correspond to deviation features at different time granularities. Based on a pre-defined abnormal pattern library, the deviation feature sub-vectors at each scale are matched with the abnormal pattern feature templates in the abnormal pattern library, and the matching degree between each deviation feature sub-vector and each abnormal pattern feature template is calculated. The scale confidence of each candidate anomaly pattern at different scales is determined based on the matching degree. The scale confidence of the same candidate anomaly pattern at different scales is weighted and averaged to obtain the preliminary confidence of the candidate anomaly pattern. The initial confidence scores of all candidate anomaly patterns are integrated into an initial confidence vector, where each element of the initial confidence vector corresponds to the initial confidence score of a candidate anomaly pattern.

7. The method according to claim 6, characterized in that, The step of inputting the multidimensional deviation feature vector into the feature enhancement layer of the hierarchical anomaly pattern inference model, and performing nonlinear mapping processing on the multidimensional deviation feature vector through the feature enhancement layer to obtain the enhanced deviation feature vector includes: The feature dimension of the multidimensional deviation feature vector is expanded by adding feature cross terms and higher-order feature terms to extend the original dimension to a preset high-dimensional space, thereby obtaining a high-dimensional extended feature vector. The high-dimensional extended feature vector is subjected to nonlinear transformation processing. Each feature component in the high-dimensional extended feature vector is nonlinearly mapped by an activation function to enhance the nonlinear correlation between features, thereby obtaining a nonlinear transformed feature vector. Feature filtering is performed on the nonlinear transformation feature vector, and feature components that do not meet the requirements are removed based on the feature importance score to obtain the enhanced deviation feature vector. The feature importance score is learned by the feature enhancement layer through historical training data.

8. The method according to claim 6, characterized in that, The method based on a preset anomaly pattern library matches the deviation feature sub-vectors at each scale with the anomaly pattern feature templates in the library, and calculates the matching degree between each deviation feature sub-vector and each anomaly pattern feature template, including: Obtain a preset abnormal pattern library, which contains multiple abnormal pattern feature templates. Each abnormal pattern feature template corresponds to a typical type of cold storage environment abnormality and contains standard feature sub-vectors at different scales. Traverse the deviation feature subvectors at different scales, and take the currently traversed deviation feature subvector as the target scale subvector, whose corresponding scale is the target scale; Extract the standard feature sub-vector of each abnormal pattern feature template at the target scale from the abnormal pattern library, and use it as the target standard sub-vector; Calculate the vector similarity between the target scale sub-vector and each target standard sub-vector, wherein the vector similarity is calculated using the vector cosine similarity algorithm; The vector similarity is used as the matching degree between the target scale sub-vector and the corresponding abnormal pattern feature template, and the matching degree is stored for subsequent preliminary confidence calculation.

9. The method according to claim 1, characterized in that, The preliminary confidence level based on the multiple candidate anomaly patterns is used to generate a comprehensive anomaly quantification index through a spatiotemporal correlation weighted fusion algorithm, including: From the multidimensional deviation feature vector, the spatial deviation transmission vector and the temporal deviation evolution vector are extracted. The spatial deviation transmission vector is reconstructed into a spatial deviation transmission matrix according to the original matrix dimension, and the temporal deviation evolution vector is reconstructed into a temporal deviation evolution matrix according to the original matrix dimension. The spatial weight coefficient corresponding to each candidate anomaly pattern is calculated based on the spatial deviation transfer matrix. The spatial weight coefficient is positively correlated with the spatial propagation range of the anomaly pattern. The time weight coefficient corresponding to each candidate anomaly mode is calculated based on the time deviation evolution matrix. The time weight coefficient is positively correlated with the evolution speed of the anomaly mode over time. The initial confidence level of each candidate anomaly pattern is multiplied by the corresponding spatial weight coefficient and temporal weight coefficient to obtain the weighted confidence level. The weighted confidence scores of all candidate anomaly patterns are summed to obtain the confidence fusion value; Based on the confidence fusion value and the preset anomaly severity grading standard, a comprehensive anomaly quantification index is calculated, and the value of the comprehensive anomaly quantification index is positively correlated with the anomaly severity.

10. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 9.

Images