Multi-terminal cooperative remote temperature control method and platform

Abstract

The application discloses a multi-terminal cooperative remote temperature control method and platform, and relates to the technical field of temperature control in bulk cargo inspection and detection. The application comprises the following steps: collecting real-time temperature data of the environment where a plurality of temperature control terminals are located to form a temperature state matrix; performing a dynamic time warping operation on the temperature state matrix to obtain temperature change lag coefficients between the temperature control terminals; constructing a cooperative control graph according to the temperature change lag coefficients, wherein a node represents a terminal, and an edge weight is assigned by the lag coefficient; performing random walk sampling in the cooperative control graph to generate a plurality of terminal access sequences; and inputting the sequences into a sequence-to-sequence deep recurrent network to output a target control temperature of each terminal. The application quantifies the temperature response delay relationship between terminals and realizes cooperative temperature setting by using graph sequence learning, thereby improving the overall coordination and adaptability of multi-region temperature control.

Description

Technical Field

[0001] This invention relates to the field of temperature control technology in bulk cargo inspection and testing, specifically to a remote temperature control method and platform with multi-terminal collaboration. Background Technology

[0002] In international bulk cargo trade, independent inspection and testing agencies, commissioned by buyers and sellers or traders, conduct weighing, sampling, sample preparation, moisture testing, and quality analysis on bulk cargoes such as copper concentrate and iron ore. The weight and quality certificates issued by these agencies serve as the basis for final trade settlement. The moisture test results directly determine the dry weight of the goods, thus affecting the settlement amount. Taking copper concentrate as an example, with the current copper price at approximately 100,000 yuan / ton, a copper concentrate containing 30% copper has a value of approximately 30,000 yuan / ton, and a shipment of 10,000 tons would be worth approximately 300 million yuan. A 0.3% deviation in the moisture test would result in a settlement amount discrepancy of approximately 900,000 yuan. Based on an annual inspection volume of 10 million tons of copper concentrate, such deviations could cause annual trade settlement discrepancies of up to 900 million yuan. Therefore, the accuracy of moisture testing is crucial for fair trade.

[0003] Hot air drying ovens are crucial equipment in moisture testing. During testing, samples must be dried within a target temperature range of 105°C ± 5°C, or other specially required temperature ranges, and the moisture content is calculated by gravimetric method. Currently, testing and inspection institutions typically have multiple independently operating hot air drying ovens, lacking centralized management and coordinated control. Existing oven temperature control relies on local independent adjustments, making it impossible to monitor temperature fluctuations in real time or identify interference factors such as abnormal door opening during sample drying. Due to differences in thermal response characteristics between different drying ovens, and the inability of operators to remotely and uniformly set and adjust temperatures, temperature consistency is difficult to guarantee, directly affecting the accuracy and repeatability of moisture test results. Furthermore, the lack of traceability of historical temperature data hinders quality control and anomaly detection.

[0004] Therefore, it is necessary to solve the problem of how to integrate multiple hot air drying ovens distributed in different sample preparation workshops into a unified management platform to achieve remote real-time monitoring and coordinated temperature control; it is also necessary to solve the problem of how to determine whether the drying process is stable and whether there are any abnormal operations by backtracking through historical temperature curves. Summary of the Invention

[0005] This invention provides a remote hot air drying oven temperature control method and platform with multi-terminal collaboration. The aim is to address the problems of existing hot air drying ovens operating independently, lacking centralized collaborative management and temperature anomaly tracking capabilities, and the inability of existing methods to dynamically adjust the collaborative target temperature based on the temperature response lag relationship between multiple drying ovens, leading to significant deviations in moisture test results. The purpose of this invention is to calculate the temperature change lag coefficient between each drying oven through dynamic time warping, construct a collaborative control graph, and utilize random walks and sequence-to-sequence networks to transform the collaborative relationships in the graph structure into the target control temperature of each drying oven. Simultaneously, it achieves the recording of historical temperature curves and the detection of abnormal door opening events, thereby improving the consistency of temperature control across multiple drying ovens and the accuracy of moisture testing.

[0006] The objective of this invention can be achieved through the following technical solutions: This invention provides a remote temperature control method for multi-terminal collaboration, comprising: collecting real-time temperature data of the environment where multiple temperature control terminals are located to form a temperature state matrix; performing dynamic time warping on the temperature state matrix to obtain temperature change lag coefficients between each temperature control terminal; constructing a collaborative control graph based on the temperature change lag coefficients, wherein nodes in the collaborative control graph represent temperature control terminals, and the edge weights between nodes are assigned by the temperature change lag coefficients; performing a random walk sampling operation on the collaborative control graph to generate multiple terminal access sequences; and inputting each terminal access sequence into a sequence-to-sequence deep recurrent network to output the target control temperature corresponding to each temperature control terminal. This method captures the sequential correlation and time lag of temperature changes between different terminals through a temperature state matrix. It uses dynamic time warping to quantify the degree of lag and constructs a weighted collaborative control graph, giving closely related terminals a stronger connection weight in the graph. Based on this, it generates a sequence reflecting the collaborative access relationship of terminals through random walks, and directly predicts the target control temperature of each terminal through a deep recurrent network from sequence to sequence. This fully integrates the temperature coupling characteristics between terminals into the control decision, improves the consistency and response timeliness of multi-region temperature collaborative regulation, and reduces overshoot and energy waste caused by ignoring lag relationships.

[0007] Preferably, a temperature state matrix is ​​constructed by collecting real-time temperature data of the environment where multiple temperature control terminals are located, including: assigning a unique terminal identifier to each temperature control terminal; simultaneously recording the real-time temperature value of each temperature control terminal under its local clock at a uniform time sampling interval; using the unique terminal identifier as a row index, the sampling time point as a column index, and the real-time temperature value as a matrix element to generate an initial temperature state matrix; detecting whether there are missing values ​​in the initial temperature state matrix, and if there are missing values, locating the target row index and target column index where the missing value is located; extracting the temperature values ​​of other time points corresponding to the target row index, and extracting the temperature values ​​of other temperature control terminals corresponding to the target column index; calculating the average value of the temperature values ​​of other time points corresponding to the target row index as a first time average, and calculating the average value of the temperature values ​​of other temperature control terminals corresponding to the target column index as a second time average; filling the position of the missing value with the arithmetic mean of the first time average and the second time average to generate a complete temperature state matrix. This method can repair missing values ​​by using the time dimension information of the same terminal and the spatial dimension information of different terminals at the same time when individual sampling data are lost, thus maintaining the continuity and integrity of the matrix and providing a reliable data foundation for subsequent hysteresis analysis and walk sampling.

[0008] As a technical solution of the present invention, a dynamic time warping operation is performed on the temperature state matrix to obtain the temperature change lag coefficient between each temperature control terminal, including: extracting a first temperature time series of the first terminal and a second temperature time series of the second terminal from the temperature state matrix; constructing a cumulative distance matrix between the first temperature time series and the second temperature time series, wherein the number of rows of the cumulative distance matrix is ​​equal to the length of the first temperature time series and the number of columns is equal to the length of the second temperature time series; initializing the first element of the cumulative distance matrix to zero, and initializing the first row element of the cumulative distance matrix to be horizontally moved from the first element to the current position. The cumulative cost is calculated by initializing the first column of the cumulative distance matrix as the cumulative cost of vertically moving from the first element to the current position. Starting from the second row and second column of the cumulative distance matrix, the value of each remaining element is calculated. The value of each remaining element is equal to the absolute value of the temperature difference between the two time points at the corresponding position plus the minimum value among its left neighbor, upper neighbor, and upper left neighbor. After completing the traversal, the lower right corner element of the cumulative distance matrix is ​​taken as the total cost of the normalized path. The total cost of the normalized path is divided by the sum of the lengths of the first temperature time series and the second temperature time series to obtain the temperature change lag coefficient. The lag coefficient obtained in this way can accurately measure the asynchronous degree of the temperature change waveforms of the two terminals and is robust to scaling and offset on the time axis, ensuring the objectivity of the edge weights of the collaborative graph.

[0009] Preferably, constructing a cooperative control graph based on the temperature change lag coefficient includes: initializing an undirected graph structure, treating each temperature control terminal as an independent node in the undirected graph structure; traversing all temperature control terminal pairs to obtain the temperature change lag coefficient between the current temperature control terminal pairs; determining whether the temperature change lag coefficient is less than a preset cooperative threshold; if less, adding an undirected edge between the two nodes corresponding to the current temperature control terminal pair; if the temperature change lag coefficient is greater than or equal to the cooperative threshold, not adding an undirected edge between the two nodes; assigning an edge weight to each undirected edge in the undirected graph, the value of which is equal to the reciprocal of the temperature change lag coefficient between the temperature control terminals corresponding to the two nodes connecting the edge; and using the assigned undirected graph structure as the cooperative control graph. By using the reciprocal of the lag coefficient as the weight, the weight between terminal pairs with tighter temperature change coupling and smaller lag is increased, guiding random walks to preferentially propagate along strong cooperative relationships and enhancing the cooperative expression characteristics of the sampling sequence. Furthermore, the edge weights between nodes in the collaborative control graph are dynamically updated with the lag coefficient of the real-time temperature change between the corresponding terminals. The update cycle is synchronized with the sampling interval of the temperature state matrix, ensuring that the graph structure always reflects the latest dynamics of environmental heat exchange and adapts to time-varying operating conditions.

[0010] In a preferred embodiment of the present invention, a random walk sampling operation is performed in the cooperative control graph to generate multiple terminal access sequences, including: adding all nodes in the cooperative control graph to a candidate starting node set; randomly selecting a node from the candidate starting node set as the current starting node and creating an empty sequence; adding the identifier of the current starting node as the first element of the empty sequence to the sequence; sampling from all neighboring nodes of the current starting node according to the weight ratio of the edges between the neighboring nodes and the current starting node, selecting a neighboring node as the next access node; adding the identifier of the next access node to the end of the sequence and updating the next access node to the new current starting node; repeating the sampling and adding operation from neighboring nodes until the length of the sequence reaches a preset sequence length threshold, closing the current sequence, and generating a complete terminal access sequence; randomly selecting a new starting node from the nodes in the candidate starting node set that were not selected as starting nodes, and repeating the generation of new terminal access sequences until the number of generated terminal access sequences reaches a preset sequence number threshold. This sampling method ensures that each terminal has an equal opportunity to be the starting point of a walk, and during the walk, it tends to visit neighbors with stronger cooperative relationships with the current terminal, so that the generated sequence naturally carries cooperative context information between terminals. In the above random walk sampling operation, a reset probability is set between two adjacent walks, with the probability of resetting back to the starting node being the reciprocal of the number of neighboring nodes of the current node. This can prevent the walk path from getting stuck in a local region or loop too early, improve the global coverage and diversity of the generated sequence, and help the downstream network learn richer terminal interaction patterns.

[0011] As an important technical solution of this invention, each terminal access sequence is input into a sequence-to-sequence deep recurrent network to output the target control temperature corresponding to each temperature control terminal. This includes: configuring an encoder and a decoder for the sequence-to-sequence deep recurrent network, where both the encoder and decoder are composed of stacked long short-term memory units; converting each node identifier in each terminal access sequence into a one-hot encoded vector, and sequentially inputting the one-hot encoded vector into the encoder; updating the encoder's internal hidden state at each time step, and outputting the final encoder hidden state as a sequence semantic vector after receiving the last node identifier of the terminal access sequence; using the sequence semantic vector as the initial hidden state of the decoder, and using a start symbol vector as the input to the decoder's first time step; the decoder... At each time step, a probability distribution vector is output. The dimension of the probability distribution vector is equal to the total number of temperature control terminals, and each dimension represents the probability that the corresponding terminal is predicted to be the target control target. The dimension index corresponding to the maximum probability value is selected from the probability distribution vector, and the temperature control terminal corresponding to that dimension index is used as the predicted output terminal for the current time step. The one-hot encoded vector of the predicted output terminal is fed back to the decoder as the input for the next time step, until the decoder outputs a terminator. All the predicted output terminals are arranged in chronological order to form a target terminal sequence. According to the arrangement order of the terminals in the target terminal sequence, a decreasing control priority value is assigned to each terminal in the sequence. The control priority value corresponding to each terminal is input into a preset linear mapping function, and the output of the linear mapping function is the target control temperature of that terminal. Therefore, the network does not provide the setpoint for each terminal in isolation, but determines the priority control object and the corresponding target temperature sequentially according to the context of the terminal access sequence. This makes the overall temperature field control hierarchical and planned, avoiding conflicts and over-adjustment caused by simple uniform control. In this sequence-to-sequence deep recurrent network, an attention mechanism is introduced. The attention mechanism performs a weighted summation of the hidden states of the encoder at each decoding time step and then participates in the calculation of the output probability distribution, thereby enhancing the utilization of key access node information in the sequence and improving the prediction accuracy of the target terminal sequence.

[0012] Before using the deep recurrent network, the following network training steps are performed: Multiple historical terminal access sequences generated by the cooperative control graph within a historical time period are collected, and each historical terminal access sequence is labeled with its corresponding real output terminal sequence; the multiple historical terminal access sequences are divided into a training sequence set and a validation sequence set; the sequences in the training sequence set are sequentially input into the encoder and the decoder, and the cross-entropy loss between the predicted probability distribution output by the decoder at each time step and the real probability distribution corresponding to the real output terminal is calculated; the gradient of the cross-entropy loss relative to each trainable parameter in the sequence-to-sequence deep recurrent network is calculated using the time backpropagation algorithm; based on the gradient, each trainable parameter in the sequence-to-sequence deep recurrent network is updated using the adaptive moment estimation optimizer; after each round of traversal of the training sequence set, the validation sequence set is input into the currently updated sequence-to-sequence deep recurrent network, and the validation set cross-entropy loss is calculated; when the validation set cross-entropy loss no longer decreases in consecutive rounds of traversal, the update operation is stopped, and the values ​​of all current trainable parameters are fixed. This supervised training and early stopping strategy enables the network to fully learn the order of terminal control and temperature mapping under different cooperative access modes, while preventing overfitting and ensuring good generalization and stability in actual deployment.

[0013] As a further improvement to the above method, after outputting the target control temperature for each temperature control terminal, the following steps are performed: A separate control instruction cache queue is created for each temperature control terminal; the target control temperature corresponding to the current temperature control terminal is obtained, along with the actual ambient temperature of the current temperature control terminal at the current moment; the absolute value of the temperature deviation between the target control temperature and the actual ambient temperature is calculated; it is determined whether the absolute value of the temperature deviation is greater than a preset trigger threshold. If it is greater, the control action direction is determined based on the sign of the absolute value of the temperature deviation, and the control action intensity is determined based on the magnitude of the absolute value of the temperature deviation; the control action direction and the control action intensity are combined into a control instruction, and this control instruction is pushed to the end of the control instruction cache queue of the current temperature control terminal; when the number of control instructions in the control instruction cache queue reaches a preset batch transmission threshold, all control instructions in the control instruction cache queue are packaged into a data packet and sent to the corresponding temperature control terminal via the network. This batch transmission mechanism significantly reduces the number of communications and network overhead without affecting the real-time performance of control, making it particularly suitable for large-scale terminal deployment scenarios. More preferably, after packaging control commands into data packets, a globally unique sequence number is assigned to each sent data packet, and the sending time is recorded. A timer is started to wait for the instruction execution receipt from the temperature control terminal. The instruction execution receipt contains the sequence number of the executed data packet and the execution result status code. If the instruction execution receipt is not received within a preset timeout period, the data packet is marked as a lost data packet. The same control command as in the lost data packet is re-extracted from the control command cache queue of the temperature control terminal corresponding to the lost data packet to form a retransmission data packet. The retransmission data packet is sent to the corresponding temperature control terminal again, and the timer is restarted to wait for the receipt. When the number of consecutive retransmissions reaches the preset maximum number of retransmissions and no receipt is received, a terminal offline alarm signal is generated, and the temperature control terminal is temporarily removed from the cooperative control graph. This retransmission and offline protection strategy ensures the instruction arrival rate while quickly detecting communication anomalies and adjusting the cooperative graph topology in a timely manner, preventing unreachable terminals from affecting the accuracy of subsequent walk sampling and network inference.

[0014] In an application scenario involving fault tolerance, after outputting the target control temperature for each temperature control terminal, the following steps are performed: receiving local ambient temperature and local device operating status data uploaded from any temperature control terminal; using the local ambient temperature as new real-time temperature data to update the corresponding time point and terminal position in the temperature status matrix; performing pattern matching between the local device operating status data and a pre-stored device fault feature library, which contains multiple fault feature rules; if a fault feature rule is matched, marking the temperature control terminal as a suspected fault terminal; locating the fault node corresponding to the suspected fault terminal from the collaborative control graph and extracting all neighbor nodes of the fault node; sending an auxiliary control request to the temperature control terminal corresponding to each neighbor node of the fault node, the auxiliary control request containing the historical set temperature information of the suspected fault terminal; receiving the auxiliary control response returned by each neighbor node, calculating a median temperature based on the suggested temperature values ​​in all auxiliary control responses, and using the median temperature as the alternative control temperature for the suspected fault terminal. When a terminal is suspected of malfunctioning, the temperature control of its area is temporarily replaced by the suggested temperature provided by adjacent collaborative terminals in the collaborative graph. By utilizing the regional thermal coupling effect, uninterrupted temperature maintenance is achieved, which significantly improves the overall resilience of the system and user comfort.

[0015] This invention also provides a multi-terminal collaborative remote temperature control platform, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the multi-terminal collaborative remote temperature control method. This platform integrates the control processes described above, which possess capabilities such as hysteresis sensing, collaborative graph construction, random walk sampling, sequence prediction, and robust instruction issuance, into a single implementation. It is suitable for complex temperature monitoring environments such as distributed industrial temperature control terminal groups.

[0016] The beneficial effects of this invention are: Temperature time series of different hot air drying ovens are extracted from the temperature state matrix. A cumulative distance matrix is ​​constructed through dynamic time warping, and the total cost of the warped path is calculated. Dividing this cost by the sum of the lengths of the two sequences yields the temperature change lag coefficient. This coefficient dynamically reflects the temperature response time series offset between any two drying ovens caused by factors such as equipment performance differences, placement, and ventilation conditions, rather than using a fixed delay constant or static grouping. A collaborative control graph is constructed using this coefficient, and undirected edges are added between nodes based on whether they are less than a preset collaborative threshold. The reciprocal of the lag coefficient is used as the edge weight, and the edge weight is updated synchronously with the sampling interval of the temperature state matrix. This graph construction method allows the management platform to capture the strength and propagation direction of thermal influence between drying ovens in real time. When environmental conditions change or equipment aging causes changes in thermal response characteristics, the connection relationships and weights of the edges in the graph are automatically adjusted to ensure that the collaborative temperature setting is always consistent with the actual thermal dynamics, avoiding temperature setting conflicts or inconsistencies caused by individual equipment differences.

[0017] Weighted random walk sampling is performed on the collaborative control graph to generate multiple drying oven access sequences. Each sequence is input into a sequence deep recurrent network composed of stacked long short-term memory units. The encoder transforms the drying oven identifier sequence into a sequence semantic vector, and the decoder outputs the probability distribution of each drying oven being predicted as the target control objective step by step. It then assigns decreasing control priority values ​​to the drying ovens in the sequence according to the output order, and obtains the target control temperature through a linear mapping function. This approach transforms the causal logic and spatial propagation chain of thermal effects between drying ovens inherent in the graph structure into a context-sensitive sequence learning task. The network can capture the temperature regulation dependencies between drying ovens from the walk paths. The generated target control temperature not only considers the current deviation of a single drying oven but also implicitly includes the adjustment needs of adjacent drying ovens connected along high-weight edges, making the overall temperature field of multiple drying ovens more uniform and stable. This effectively reduces moisture test deviations caused by temperature fluctuations and ensures fairness in trade settlement. Simultaneously, the platform records historical temperature curves, facilitating the verification of any abnormal door openings or other interference events during the drying process, thus improving quality traceability capabilities. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of a multi-terminal collaborative remote temperature control method; Figure 2 This is a flowchart of the method for filling missing values ​​in the temperature state matrix; Figure 3 This is a flowchart for calculating the temperature change hysteresis coefficient; Figure 4 It is a flowchart for constructing a collaborative control diagram; Figure 5 This is a flowchart of a collaborative temperature control method based on sequence-to-sequence deep recurrent networks. Detailed Implementation

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

[0021] See Figure 1 This invention provides a multi-terminal collaborative remote hot air drying oven temperature control method, applicable to the unified management and collaborative temperature control of multiple hot air drying ovens in the field of bulk cargo inspection and testing. The method includes the following steps: collecting real-time temperature data of the environment where each hot air drying oven is located to form a temperature state matrix; performing dynamic time warping on the temperature state matrix to obtain the temperature change lag coefficient between each drying oven; constructing a collaborative control graph based on the temperature change lag coefficient, where nodes in the collaborative control graph represent hot air drying ovens, and the edge weights between nodes are assigned by the temperature change lag coefficient; performing a random walk sampling operation on the collaborative control graph to generate multiple drying oven access sequences; inputting each drying oven access sequence into a sequence-to-sequence deep recurrent network to output the target control temperature corresponding to each hot air drying oven. Through this method, collaborative temperature setting of multiple drying ovens and centralized recording of historical temperature data can be achieved, ensuring accurate and stable temperature of each drying oven during moisture testing.

[0022] See Figure 2 Each temperature control terminal is assigned a unique terminal identifier. In practice, the unique terminal identifier can be the hardware serial number, MAC address, or UUID generated by the platform. The unique terminal identifier is used to uniquely mark the data row of the corresponding temperature control terminal in the temperature status matrix.

[0023] The system simultaneously records the real-time temperature value of each temperature control terminal under its local clock, following a uniform time sampling interval. In some embodiments, the uniform time sampling interval is preset according to the temperature change rate and system accuracy requirements, and the sampling times of all temperature control terminals are calibrated and aligned by a network time synchronization protocol. The real-time temperature value is acquired by the temperature sensor built into the temperature control terminal, and the data acquired by the temperature sensor is converted from analog to digital to form the real-time temperature value.

[0024] An initial temperature state matrix is ​​generated by using unique terminal identifiers as row indices, sampling time points as column indices, and real-time temperature values ​​as matrix elements. The number of rows in the initial temperature state matrix equals the total number of temperature control terminals, and the number of columns depends on the total number of sampling time points within the sampling period. The initial temperature state matrix... The row index corresponds to the first The unique terminal identifier of each temperature control terminal, the first The column index corresponds to the first Each sampling time point, matrix element Indicates the first The temperature control terminal is in the first Real-time temperature values ​​at each sampling time point.

[0025] The system checks for missing values ​​in the initial temperature state matrix. Missing values ​​can be caused by factors such as communication interruption at the temperature control terminal, sensor acquisition failure, or data packet loss. In practice, the system iterates through each element of the initial temperature state matrix, determining whether each element is null or an abnormal value exceeding a preset reasonable temperature range. Null and abnormal values ​​are marked as missing values.

[0026] When there are missing values ​​in the initial temperature state matrix, locate the target row index and target column index of the missing value. The target row index is the row number of the missing value in the initial temperature state matrix, and the target column index is the column number of the missing value in the initial temperature state matrix.

[0027] Extract the temperature values ​​at other time points corresponding to the target row index, and extract the temperature values ​​at other temperature control terminals corresponding to the target column index. Specifically, in the initial temperature state matrix, determine the complete row vector of the temperature control terminal corresponding to the target row index based on the target row index. After removing the positions of missing values ​​from the complete row vector, obtain the temperature values ​​at other time points corresponding to the target row index; determine the complete column vector of the sampling time point corresponding to the target column index based on the target column index. After removing the positions of missing values ​​from the complete column vector, obtain the temperature values ​​at other temperature control terminals corresponding to the target column index.

[0028] The average of the temperature values ​​at other time points corresponding to the target row index is calculated as the first time mean, and the average of the temperature values ​​at other temperature control terminals corresponding to the target column index is calculated as the second time mean. The first time mean reflects the average temperature level of the temperature control terminal corresponding to the missing value during the sampling period, excluding the time of the missing value. The second time mean reflects the average temperature level among multiple temperature control terminals in the same environment at the sampling time corresponding to the missing value.

[0029] The arithmetic mean of the first and second time-mean values ​​is used to fill in the missing values, generating a complete temperature state matrix. After the filling operation is completed, all missing values ​​in the initial temperature state matrix are replaced with reasonable estimated values, resulting in a complete temperature state matrix. Every element in the complete temperature state matrix is ​​a valid value, which is used for subsequent dynamic time warping operations.

[0030] See Figure 3 The first temperature time series of the first terminal and the second temperature time series of the second terminal are extracted from the temperature state matrix. In a specific implementation, the first terminal and the second terminal are any two different temperature control terminals in the temperature state matrix. The first temperature time series is a sequence composed of the temperature values ​​of all sampling time points of the corresponding row of the first terminal in the temperature state matrix, and the second temperature time series is a sequence composed of the temperature values ​​of all sampling time points of the corresponding row of the second terminal in the temperature state matrix.

[0031] Construct a cumulative distance matrix between the first and second temperature time series. The number of rows in the cumulative distance matrix equals the length of the first temperature time series, and the number of columns equals the length of the second temperature time series. Optionally, the length of the first temperature time series is denoted as... The length of the second temperature time series is denoted as Then the cumulative distance matrix is ​​a OK The two-dimensional matrix of columns, the cumulative distance matrix of the th column Line 1 The elements of the column are used to represent the first temperature time series. The elements and the first time series of the second temperature The cumulative cost of regularizing and aligning elements, where The value ranges from 1 to All integers, The value ranges from 1 to All integers.

[0032] The first element of the initial cumulative distance matrix is ​​zero, and it is located in the first row and first column of the matrix. The first row of the initial cumulative distance matrix represents the cumulative cost of moving horizontally from the first element to the current position; that is, for the element in the first row of the cumulative distance matrix... The elements of the column are equal to the first element of the first temperature time series and the second to third elements of the second temperature time series. The result of summing the absolute values ​​of the temperature differences of the elements sequentially is expressed by the formula: ; in: This represents the first row of the cumulative distance matrix. Column elements, This represents the first temperature value in the first temperature time series. The second temperature time series represents the first A temperature value.

[0033] The first column of the initial cumulative distance matrix is ​​the cumulative cost of moving vertically from the first element to the current position, i.e., for the element in the cumulative distance matrix... The element in the first column of the second temperature time series has a value equal to the first element of the second temperature time series and the second to third elements of the first temperature time series. The result of summing the absolute values ​​of the temperature differences of the elements sequentially is expressed by the formula: ; in: Represents the cumulative distance matrix. The element in the first column of the row, Represents the first temperature time series. A temperature value, This represents the first temperature value in the second temperature time series.

[0034] Starting from the second row and second column of the cumulative distance matrix, iterate through and calculate the value of each remaining element. The value of each remaining element is equal to the absolute value of the temperature difference between the two time points at the corresponding location, plus the minimum value among its left neighbor, its upper neighbor, and its upper left neighbor. The calculation formula is as follows: ; in: Represents the cumulative distance matrix. Line 1 Column elements, Represents the first temperature time series. A temperature value, The second temperature time series represents the first A temperature value, Represents the th element in the cumulative distance matrix Line 1 Column elements, Represents the th element in the cumulative distance matrix Line 1 Column elements, Represents the th element in the cumulative distance matrix Line 1 Column elements, This means taking the minimum value among all input values. Represents the first temperature time series. The temperature value and the second temperature time series The absolute value of the temperature difference between temperature values.

[0035] In some embodiments, the traversal calculation process starts from... and Starting from the position, calculate the element values ​​of each remaining position in the cumulative distance matrix in a row-by-row or column-by-column order, until... and The position has been calculated.

[0036] After completing the traversal, the bottom-right element of the cumulative distance matrix is ​​used as the total cost of the normalized path. The bottom-right element is the element at position (i) in the cumulative distance matrix. Line 1 Column elements The total cost of the normalized path represents the minimum cumulative cost obtained after dynamically normalizing and aligning the first temperature time series with the second temperature time series.

[0037] Dividing the total cost of the normalized path by the sum of the lengths of the first and second temperature time series yields the temperature change lag coefficient. The formula for calculating the temperature change lag coefficient is: ; The temperature change lag coefficient is used to quantify the degree of time lag in temperature changes between two temperature control terminals. The smaller the temperature change lag coefficient, the more synchronized the temperature change patterns of the two temperature control terminals are. The larger the temperature change lag coefficient, the greater the time delay in temperature changes between the two temperature control terminals.

[0038] See Figure 4 An undirected graph structure is initialized, with each temperature control terminal treated as an independent node in the graph. In the specific implementation, the undirected graph structure is stored using an adjacency list, where each node corresponds to a unique terminal identifier for a temperature control terminal, and the connections between nodes represent the collaborative control relationships between the temperature control terminals.

[0039] Iterate through all temperature control terminal pairs to obtain the temperature change lag coefficient between the current temperature control terminal pairs. All temperature control terminal pairs refer to all combinations formed by arbitrarily selecting two different temperature control terminals from the entire set of temperature control terminals. The traversal operation sequentially retrieves each group of temperature control terminal pairs and extracts the corresponding temperature change lag coefficient from the already calculated set of temperature change lag coefficients.

[0040] The system determines whether the temperature change lag coefficient is less than a preset coordination threshold. If it is, an undirected edge is added between the two corresponding nodes of the current temperature control terminal. If the temperature change lag coefficient is greater than or equal to the coordination threshold, no undirected edge is added between the two nodes. In some embodiments, the preset coordination threshold is set to 0.5. This value is determined based on the statistical analysis results of the temperature propagation delay of the temperature control terminal in a typical indoor environment. When the temperature change lag coefficient is less than 0.5, it indicates that the temperature change synchronization between the two temperature control terminals has reached the standard for coordinated control. In specific implementations, the coordination threshold is stored in the platform's configuration parameters and can be adjusted during the system initialization phase according to the deployment density of the temperature control terminals and the building's spatial structure.

[0041] Assign an edge weight to each undirected edge in the undirected graph. The edge weight is equal to the reciprocal of the temperature hysteresis coefficient between the temperature control terminals corresponding to the two nodes connecting the edge. The formula for calculating the edge weight is: ; in: Represents nodes in an undirected graph With nodes Edge weights between them Represents a node Corresponding temperature control terminals and nodes The temperature change hysteresis coefficient between the corresponding temperature control terminals.

[0042] The assigned undirected graph structure is used as a cooperative control graph. The cooperative control graph is used to provide node connectivity and edge weight information in subsequent random walk sampling operations.

[0043] In the collaborative control graph, the edge weights between nodes are dynamically updated based on the real-time temperature change lag coefficients between the corresponding terminals. The update cycle is synchronized with the sampling interval of the temperature state matrix. In some embodiments, whenever the temperature state matrix completes one sampling and generates a complete temperature state matrix, the platform re-extracts the temperature time series of all temperature control terminal pairs, re-executes dynamic time warping to obtain the updated temperature change lag coefficients, and then re-determines whether to add undirected edges based on the updated temperature change lag coefficients and the unchanged collaborative threshold, and recalculates the edge weights of existing undirected edges. During the dynamic update process, if a temperature change lag coefficient becomes greater than or equal to the collaborative threshold, the corresponding undirected edge will be removed; if there was no undirected edge between a pair of nodes originally, but the updated temperature change lag coefficient is less than the collaborative threshold, an undirected edge is added and assigned the corresponding edge weight.

[0044] See Figure 5All nodes in the collaborative control graph are added to the candidate starting node set. In specific implementations, the collaborative control graph is stored in the form of an adjacency list. All node identifiers are read one by one and placed into a list structure, which is the candidate starting node set. The candidate starting node set contains the nodes corresponding to all temperature control terminals in the collaborative control graph.

[0045] A node is randomly selected from the candidate starting node set as the current starting node, and an empty sequence is created. The empty sequence is a variable-length list with an initial length of 0, used to store the identifiers of nodes that will be visited subsequently. Each node in the candidate starting node set has an equal probability of being selected. After selecting a node, the selected node is marked as the current starting node.

[0046] Add the identifier of the current starting node as the first element of the empty sequence to the sequence. The identifier of the current starting node is the unique terminal identifier of the temperature control terminal, and the sequence now contains one element.

[0047] From all neighboring nodes of the current starting node, probabilistic sampling is performed based on the weight ratio of the edges between neighboring nodes and the current starting node, selecting one neighboring node as the next visited node. A neighboring node is a node in the collaborative control graph that has an undirected edge with the current starting node. The specific method of probabilistic sampling is as follows: All neighboring nodes of the current starting node in the collaborative control graph are obtained to form a neighboring node set. The edge weight between each neighboring node and the current starting node is extracted, and the edge weight value is the reciprocal of the corresponding temperature change lag coefficient. The probability of each neighboring node being sampled is calculated using the following formula: ; in: Representing neighboring nodes The probability of being selected as the next access node. Indicates the current starting node. Indicates the current starting node The set of all neighboring nodes. Indicates the current starting node with neighboring nodes Edge weights between them Indicates the current starting node with neighboring nodes Edge weights between, sign Represents the set of neighboring nodes The edge weights corresponding to all neighboring nodes are summed. Based on the calculated sampling probabilities of each neighboring node, a roulette wheel selection method is used to determine a neighboring node as the next node to be visited.

[0048] The identifier of the next visited node is appended to the end of the sequence, and the next visited node is updated to the new current starting node. After appending the identifier of the next visited node to the sequence, the sequence length is increased by 1.

[0049] The process of sampling from neighboring nodes and adding them to the sequence is repeated until the sequence length reaches a preset sequence length threshold. The current sequence is then closed, generating a complete terminal access sequence. In some embodiments, the preset sequence length threshold is set as the product of the total number of nodes in the collaborative control graph and a fixed multiple. When the total number of nodes in the collaborative control graph is... When the sequence length threshold is set to The fixed multiplier of 2 is set to ensure that a single sequence can cover a number of steps equivalent to twice the number of nodes in the graph, thereby fully capturing long-range dependencies between nodes. When the sequence length reaches the sequence length threshold, adding new node identifiers to the current sequence stops, the current sequence is saved as a complete terminal access sequence, and the current sequence is closed.

[0050] A new starting node is randomly selected from the candidate starting node set of nodes that were not chosen as starting nodes. New terminal access sequences are generated repeatedly until the number of generated terminal access sequences reaches a preset sequence number threshold. In some embodiments, the preset sequence number threshold is set as the product of the total number of nodes in the collaborative control graph and another fixed multiple. When the total number of nodes in the collaborative control graph is... When the sequence number threshold is set to The fixed multiplier of 50 is set to ensure that each node is used as a starting node a sufficient number of times, so that the generated terminal access sequences can provide sufficient sample coverage for the training of subsequent sequences to the deep recurrent network. When generating a new terminal access sequence, a new starting node is randomly selected from the set of nodes that have not been selected as starting nodes each time. After all nodes have been selected as starting nodes once, all nodes in the candidate starting node set are marked as not selected as starting nodes again, and the selection continues in a loop.

[0051] Before selecting the next access node from the current starting node, a reset step is set up, with the reset probability equal to the reciprocal of the number of neighboring nodes of the current starting node. In specific implementation, when the current starting node is determined... Then, obtain the current starting node. The set of neighboring nodes cardinality Reset probability Calculated as Generate a random number that follows a uniform distribution within the range of zero to one. Compare this random number with the reset probability. If the random number is less than the reset probability, ignore the normal probability sampling result, directly designate the next access node as the starting node of the current terminal access sequence, add the identifier of this starting node to the end of the sequence, and reset the current starting node to this starting node. If the random number is greater than or equal to the reset probability, then proceed according to the sampling probability. Probabilistic sampling is performed to select a neighboring node as the next node to be visited. The decision-making process is reset so that the random walk has a certain probability of returning to the visited starting point, avoiding the walk from spreading too far into the distant region of the cooperative control graph and improving the compactness of the generated terminal visit sequence in expressing local neighborhood cooperative relationships.

[0052] In practice, a sequence-to-sequence deep recurrent network is configured with an encoder and a decoder. The encoder consists of three stacked layers of long short-term memory (LSM) units. Each LSM unit comprises an input gate, a forget gate, an output gate, and a memory unit. The input to the first LSM unit is the one-hot encoded vector at the current time step. The hidden state of the first LSM unit serves as the input to the second LSM unit, and so on. The hidden state of the third LSM unit becomes the output hidden state of the encoder at that time step. The decoder also consists of three stacked LSM units. The input to the first LSM unit of the decoder is the input vector at the current time step. The three LSM units of the decoder pass the hidden state layer by layer in the same hierarchical connection manner.

[0053] Each node identifier in each terminal access sequence is converted into a one-hot encoded vector. The dimension of the one-hot encoded vector is equal to the total number of temperature control terminals. For each node identifier corresponding to a temperature control terminal, its one-hot encoded vector has a value of 1 in the dimension corresponding to that temperature control terminal and 0 in all other dimensions. These one-hot encoded vectors are sequentially input into the encoder. At each time step, the encoder receives one one-hot encoded vector, and the three layers of Long Short-Term Memory (LSTM) update their respective internal hidden states sequentially. After receiving the one-hot encoded vector corresponding to the last node identifier of the terminal access sequence, the encoder uses the hidden state output by the third LSM at this point as the sequence semantic vector. The sequence semantic vector is a fixed-length real-number vector used to summarize the temporal characteristics and cooperative relationship information of the entire terminal access sequence.

[0054] Before decoding begins, the sequence semantic vector is used as the initial hidden state of the three layers of long short-term memory (LSTM) in the decoder. Simultaneously, the initial memory state of each LTM layer in the decoder is set to an all-zero vector. A predefined start symbol vector is used as the input vector for the first time step of the decoder. The dimension of the start symbol vector is the same as the dimension of the one-hot encoded vector, and all elements of the start symbol vector are set to 0.

[0055] An attention mechanism is introduced into a sequence-to-sequence deep recurrent network. At each decoding time step, the attention mechanism performs a weighted summation of the third-layer hidden states output by the encoder at each time step, and participates in the calculation of the output probability distribution. The specific implementation of the attention mechanism is as follows: At the current decoding time step, the hidden state output by the third-layer long short-term memory unit of the decoder at the current time step is obtained as a query vector; the query vector is multiplied by the third-layer hidden states output by the encoder at each time step to obtain a set of raw attention scores; the raw attention scores are normalized using the softmax function to obtain a set of attention weights, the number of attention weights being equal to the total length of the encoder time steps; the third-layer hidden state of each encoder time step is multiplied by the corresponding attention weight, and all products are summed to obtain a context vector; the context vector is concatenated with the hidden state of the third-layer long short-term memory unit of the decoder at the current time step, and the concatenated vector is processed by a fully connected layer and the softmax function to obtain the probability distribution vector output at the current decoding time step.

[0056] The dimension of the probability distribution vector equals the total number of temperature control terminals, with each dimension representing the probability that the corresponding temperature control terminal is predicted as the target control object. The dimension index corresponding to the maximum probability value is selected from the probability distribution vector, and the temperature control terminal corresponding to that dimension index is used as the predicted output terminal for the current time step. If the dimension index corresponding to the maximum probability value is a predefined terminator index, then the terminator index indicates that decoding should end at the current time step. The dimension corresponding to the terminator index is an extra dimension added to the one-hot encoding vector; that is, the total dimension of the one-hot encoding vector is the total number of temperature control terminals plus one, with the extra dimension specifically used to identify the terminator. The decoding process stops when the predicted output terminal is a terminator index.

[0057] If the predicted output terminal is not a terminator index, the one-hot encoded vector of the predicted output terminal is fed back to the decoder as the input vector for the next time step. The decoder then proceeds to the next time step calculation until it outputs a terminator index. All predicted output terminals are then arranged into a target terminal sequence according to the temporal order of the decoding time steps. The target terminal sequence contains a unique terminal identifier for the temperature control terminal, and the sequence length is determined by the time step position of the terminator index output.

[0058] During the training phase of the sequence-to-sequence deep recurrent network, multiple historical terminal access sequences generated by the collaborative control graph within historical time periods are collected. Each historical terminal access sequence is labeled with its corresponding real output terminal sequence, which is determined based on the order in which the temperature control terminals were actually selected to perform temperature regulation in the historical control log. The multiple historical terminal access sequences are divided into a training sequence set and a validation sequence set according to a preset ratio. The preset ratio is set to 80% of the historical terminal access sequences in the training sequence set and 20% in the validation sequence set. The division method uses stratified random sampling to ensure that the distribution of terminal types in the two sets remains consistent.

[0059] The historical terminal access sequences from the training sequence set are sequentially input into the encoder and decoder. The encoder and decoder generate the predicted probability distribution for each decoding time step according to the aforementioned forward propagation method. For each decoding time step, the cross-entropy loss between the predicted probability distribution and the true probability distribution corresponding to the actual output terminal is calculated. The formula for calculating the cross-entropy loss is: ; in: This represents the cross-entropy loss value of a training sample; This indicates the length of the target terminal sequence, i.e., the number of decoding time steps before the decoder output terminator index; Indicates the total number of temperature control terminals; This represents the total dimension of the probability distribution vector, including... The dimension corresponding to each temperature control terminal and the dimension corresponding to a terminator; Indicates the first Each decoding time step, the actual output terminal in dimension The true probability distribution value on the actual output terminal corresponding to the dimension hour The value is 1, otherwise The value is 0; Indicates the first Each decoding time step, the dimension of the probability distribution vector output by the decoder. The predicted probability value on The range of values ​​for is a real number greater than 0 and less than 1; Represents the natural logarithm operation.

[0060] The backpropagation algorithm is used to calculate the gradient of the cross-entropy loss relative to each trainable parameter in the sequence-to-sequence deep recurrent network. Trainable parameters include the weight matrices and bias vectors of all long short-term memory units in the encoder and decoder, the weight matrices and bias vectors of the fully connected layers in the attention mechanism, and the weight matrices and bias vectors of the output layer. Based on the calculated gradients, each trainable parameter in the sequence-to-sequence deep recurrent network is updated using an adaptive moment estimation optimizer. The learning rate of the adaptive moment estimation optimizer is set to 0.001, the exponential decay rate of the first-order moment estimation is set to 0.9, the exponential decay rate of the second-order moment estimation is set to 0.999, and the numerical stability constant is set to 10⁻⁸. After each epoch of iterating through all training samples in the training sequence set (i.e., after one training epoch), all validation samples in the validation sequence set are input into the currently updated sequence-to-sequence deep recurrent network, and the validation set cross-entropy loss is calculated in the same way as in the training phase. When the validation set cross-entropy loss no longer decreases after five consecutive epochs, the update operation is stopped, and the values ​​of all trainable parameters are fixed.

[0061] After training, during the actual operation phase, the terminal access sequence generated in real time from the collaborative control graph is input into a deep recurrent network with fixed parameters, outputting the target terminal sequence. Based on the order of the temperature control terminals in the target terminal sequence, each temperature control terminal is assigned a decreasing control priority value. The first temperature control terminal in the target terminal sequence is assigned the highest control priority value, and the control priority values ​​decrease linearly according to the terminal's position in the sequence. The control priority value corresponding to each temperature control terminal is input into a preset linear mapping function. The expression of the linear mapping function is: the target control temperature equals the reference temperature minus the product of the mapping coefficient and the control priority value. The reference temperature is determined based on the set comfort temperature of the environment where the temperature control terminal is located, and the mapping coefficient is determined based on the value range of the control priority value and the temperature control range. The mapping coefficient is set as the difference between the reference temperature and the lowest target control temperature divided by the maximum value of the control priority value. The output of the linear mapping function is the target control temperature of the corresponding temperature control terminal.

[0062] After outputting the target control temperature for each temperature control terminal, an independent control command buffer queue is created for each terminal. This queue is a first-in, first-out (FIFO) data structure that stores control commands to be sent. The target control temperature for the current terminal is retrieved, along with the current ambient temperature, obtained from real-time temperature data uploaded by the terminal. The absolute value of the temperature deviation between the target and ambient temperatures is calculated. The calculation of the absolute deviation does not distinguish between positive and negative values; only the magnitude of the deviation is considered. It is then determined whether the absolute value of the temperature deviation exceeds a preset trigger threshold of 0.5 degrees Celsius. This threshold is set based on the assumption that minute temperature fluctuations within the temperature sensing accuracy range do not trigger control actions, and a 0.5-degree Celsius fluctuation will not cause significant discomfort. If the absolute value of the temperature deviation is not greater than the trigger threshold, no control command is generated. If the absolute value exceeds the threshold, the direction of the control action is determined by the sign of the difference between the target and ambient temperatures: a positive difference indicates heating, and a negative difference indicates cooling. The intensity of the control action is determined based on the magnitude of the absolute value of the temperature deviation. The intensity of the control action is proportional to the absolute value of the temperature deviation, and the proportional coefficient is pre-calibrated based on the maximum output power of the temperature control terminal and the maximum allowable range of temperature deviation.

[0063] The control action direction and intensity are combined into a single control command, which includes an action direction field and an intensity value field. This control command is then pushed to the end of the control command buffer queue of the current temperature control terminal. When the number of control commands in the buffer queue reaches a preset batch transmission threshold, all control commands in the buffer queue are packaged into a single data packet and sent to the corresponding temperature control terminal via the network. The batch transmission threshold is set to 10, a value determined by balancing network transmission efficiency and command timeliness; the cumulative delay of 10 control commands is within an acceptable range in temperature control scenarios.

[0064] Each sent data packet is assigned a globally unique sequence number, generated by an incrementing integer generator. The sending time of the data packet is recorded, taking the platform's current system time. A timer is started, beginning at the sending time, waiting for a command execution receipt from the temperature control terminal. The command execution receipt contains the sequence number of the executed data packet and the execution result status code, which indicates whether the control command was executed correctly. If no command execution receipt is received within a preset timeout period (set to 5 seconds), this value is based on the typical round-trip communication delay between the temperature control terminal and the control platform in a LAN environment multiplied by a safety margin factor; 5 seconds covers most normal communication scenarios. The data packet without a receipt is then marked as a lost data packet. The control commands identical to those in the lost data packet are retrieved from the control command cache queue of the temperature control terminal corresponding to the lost data packet and retransmitted in their original order. The retransmitted data packet is then sent back to the corresponding temperature control terminal, and the timer is restarted to wait for a receipt. If no acknowledgment is received after the preset maximum number of retransmissions, the maximum number of retransmissions is set to 3. This value is set based on providing a limited number of attempts within the typical recovery time window for network failures or terminal disconnections. Three retransmissions avoid prematurely determining offline status while also preventing excessive occupation of communication resources. A terminal offline alarm signal is then generated, containing the unique terminal identifier and offline timestamp of the offline temperature control terminal. Simultaneously, the temperature control terminal is temporarily removed from the collaborative control graph. The temporary removal operation includes deleting the node corresponding to the temperature control terminal in the collaborative control graph and deleting all undirected edges connected to that node.

[0065] After outputting the target control temperature for each temperature control terminal, the platform continuously receives local ambient temperature and local device operating status data uploaded from any temperature control terminal. The local device operating status data includes at least one of the following: compressor current, fan speed, circuit board temperature, and communication module signal strength. The received local ambient temperature is used as new real-time temperature data to update the elements in the temperature status matrix corresponding to the time point and the row containing the corresponding temperature control terminal. Specifically, the newly acquired real-time temperature data is written to the corresponding position in a new column of the temperature status matrix.

[0066] The system performs pattern matching between local device operating status data and a pre-stored device fault feature database. This database contains multiple fault feature rules. Each rule defines a multi-dimensional parameter range combination corresponding to a fault type. For example, the rule for a compressor overcurrent fault is that the compressor current exceeds 1.5 times the rated current and the circuit board temperature exceeds 85 degrees Celsius. Pattern matching uses a rule-triggered approach, traversing all fault feature rules in the database. If all relevant parameters in the local device operating status data fall within the parameter range specified by a fault feature rule, then a match is determined for that rule. Once a fault feature rule is matched, the temperature control terminal is marked as a suspected faulty terminal, and a fault alarm record containing the terminal identifier and the matched fault type is generated.

[0067] Locate the faulty node corresponding to the suspected faulty terminal in the collaborative control graph. The faulty node is the node in the collaborative control graph corresponding to the unique terminal identifier of the suspected faulty terminal. Extract all neighboring nodes of the faulty node. Neighboring nodes are all nodes in the collaborative control graph that have an undirected edge with the faulty node. Send an auxiliary control request to the temperature control terminal corresponding to each neighboring node of the faulty node. The data structure of the auxiliary control request contains the historical set temperature information of the suspected faulty terminal. The historical set temperature information is a sequence of target control temperatures received by the suspected faulty terminal in the past preset time period. Receive the auxiliary control response returned by each neighboring node. The auxiliary control response contains the alternative temperature value suggested by the neighboring node based on its current ambient temperature and the auxiliary control request. After all auxiliary control responses returned by all neighboring nodes have been received or the timeout has expired, collect the suggested temperature values ​​from all auxiliary control responses to form a temperature value set. Sort all values ​​in the temperature value set from smallest to largest, and take the value in the middle of the sorted sequence as the median temperature; if the number of elements in the temperature value set is even, take the arithmetic mean of the two middle values ​​as the median temperature. The median temperature is used as the alternative control temperature for suspected faulty terminals, and the alternative control temperature is generated and sent to the suspected faulty terminals or indirectly forwarded through neighboring nodes in accordance with the aforementioned control command generation and sending method.

[0068] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for remote temperature control of multi-terminal collaboration, characterized by, Includes the following steps: A temperature status matrix is ​​formed by collecting real-time temperature data of the environment where multiple temperature control terminals are located. Perform dynamic time warping on the temperature state matrix to obtain the temperature change hysteresis coefficient between each temperature control terminal; A collaborative control graph is constructed based on the temperature change lag coefficient. The nodes in the collaborative control graph represent temperature control terminals, and the edge weights between nodes are assigned by the temperature change lag coefficient. A random walk sampling operation is performed in the cooperative control graph to generate multiple terminal access sequences; Each terminal access sequence is input into a sequence-to-sequence deep recurrent network, and the target control temperature corresponding to each temperature control terminal is output.

2. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, The process of collecting real-time temperature data from the environment where multiple temperature control terminals are located to construct a temperature status matrix includes the following steps: Assign a unique terminal identifier to each temperature control terminal; Record the real-time temperature value of each temperature control terminal under its local clock at a uniform time sampling interval. The unique terminal identifier is used as the row index, the sampling time point is used as the column index, and the real-time temperature value is used as the matrix element to generate an initial temperature state matrix. Detect whether there are missing values ​​in the initial temperature state matrix. If there are missing values, locate the target row index and target column index where the missing value is located. Extract the temperature values ​​of other time points corresponding to the target row index, and extract the temperature values ​​of other temperature control terminals corresponding to the target column index; The average temperature values ​​at other time points corresponding to the target row index are calculated as the first time average, and the average temperature values ​​at other temperature control terminals corresponding to the target column index are calculated as the second time average. The arithmetic mean of the first time mean and the second time mean is used to fill the positions of the missing values ​​to generate a complete temperature state matrix.

3. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, Performing dynamic time warping on the temperature state matrix to obtain the temperature change hysteresis coefficient between each temperature control terminal includes the following steps: Extract the first temperature time series of the first terminal and the second temperature time series of the second terminal from the temperature state matrix; Construct a cumulative distance matrix between the first temperature time series and the second temperature time series, wherein the number of rows in the cumulative distance matrix is ​​equal to the length of the first temperature time series and the number of columns is equal to the length of the second temperature time series; The first element of the cumulative distance matrix is ​​initialized to zero, the first row of the cumulative distance matrix is ​​initialized to the cumulative cost of moving horizontally from the first element to the current position, and the first column of the cumulative distance matrix is ​​initialized to the cumulative cost of moving vertically from the first element to the current position. Starting from the second row and second column of the cumulative distance matrix, the value of each remaining element is calculated by iterating through the matrix. The value of each remaining element is equal to the absolute value of the temperature difference between the two time points at the corresponding position plus the minimum value among its left neighbor, its upper neighbor, and its upper left neighbor. After completing the traversal, the bottom right element of the cumulative distance matrix is ​​used as the total cost of the normalized path; The temperature change lag coefficient is obtained by dividing the total cost of the regularized path by the sum of the lengths of the first temperature time series and the second temperature time series.

4. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, Constructing a collaborative control chart based on the temperature change hysteresis coefficient includes the following steps: Initialize an undirected graph structure, treating each temperature control terminal as an independent node in the undirected graph structure; Traverse all temperature control terminal pairs and obtain the temperature change hysteresis coefficient between the current temperature control terminal pairs; Determine whether the temperature change hysteresis coefficient is less than a preset collaborative threshold. If it is less, add an undirected edge between the two corresponding nodes in the current temperature control terminal. If the temperature change hysteresis coefficient is greater than or equal to the collaborative threshold, then no undirected edge is added between the two nodes; Assign an edge weight to each undirected edge in the undirected graph. The value of the edge weight is equal to the reciprocal of the temperature change hysteresis coefficient between the temperature control terminals corresponding to the two nodes connecting the edge. The undirected graph structure after assignment is used as the cooperative control graph.

5. The multi-terminal collaborative remote temperature control method according to claim 4, characterized in that, The edge weights between nodes in the collaborative control graph are dynamically updated with the lag coefficient of the real-time temperature change between the corresponding terminals, and the update period is synchronized with the sampling interval of the temperature state matrix.

6. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, Performing a random walk sampling operation in the cooperative control graph to generate multiple terminal access sequences includes the following steps: Add all nodes in the collaborative control graph to the candidate starting node set; Randomly select a node from the set of candidate starting nodes as the current starting node, and create an empty sequence; Add the identifier of the current starting node as the first element of the empty sequence to the sequence; From all the neighboring nodes of the current starting node, a probability sampling is performed based on the weight ratio of the edges between the neighboring nodes and the current starting node, and a neighboring node is selected as the next visited node. The identifier of the next access node is appended to the end of the sequence, and the next access node is updated to the new current starting node; Repeat the operation of sampling from neighboring nodes and adding to the sequence until the length of the sequence reaches the preset sequence length threshold, then close the current sequence and generate a complete terminal access sequence; A new starting node is randomly selected from the nodes that were not selected as starting nodes in the candidate starting node set, and new terminal access sequences are generated repeatedly until the number of generated terminal access sequences reaches a preset sequence number threshold.

7. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, In the random walk sampling operation, a reset probability is set between two adjacent walks, with the reciprocal of the number of neighboring nodes of the current node serving as the probability of resetting back to the starting node.

8. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, Each terminal access sequence is input into a sequence-to-sequence deep recurrent network, and the target control temperature corresponding to each temperature control terminal is output, including the following steps: An encoder and a decoder are configured for the sequence-to-sequence deep recurrent network, both of which are composed of stacked long short-term memory units; Each node identifier in each terminal access sequence is converted into a one-hot encoded vector, and the one-hot encoded vectors are sequentially input into the encoder; The encoder updates its internal hidden state at each time step, and after receiving the last node identifier of the terminal access sequence, outputs the final encoder hidden state as a sequence semantic vector. The sequence semantic vector is used as the initial hidden state of the decoder, and a start symbol vector is used as the input of the decoder at the first time step. The decoder outputs a probability distribution vector at each time step. The dimension of the probability distribution vector is equal to the total number of temperature control terminals, and each dimension represents the probability that the corresponding terminal is predicted to be the target of the control. Select the dimension index corresponding to the maximum probability from the probability distribution vector, and use the temperature control terminal corresponding to the dimension index as the prediction output terminal for the current time step; The one-hot encoded vector of the predicted output terminal is fed back to the decoder as the input for the next time step until the decoder outputs a terminator. All the predicted output terminals are then arranged in chronological order to form a target terminal sequence. Based on the order of the terminals in the target terminal sequence, assign a decreasing control priority value to each terminal in the sequence; The control priority value corresponding to each terminal is input into a preset linear mapping function, and the output of the linear mapping function is the target control temperature of that terminal.

9. The remote temperature control method with multi-terminal collaboration according to claim 1, characterized in that, An attention mechanism is introduced into the sequence-to-sequence deep recurrent network. The attention mechanism performs a weighted summation of the hidden states of the encoder at each decoding time step and then participates in the calculation of the output probability distribution.

10. A multi-terminal collaborative remote temperature control platform, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-terminal collaborative remote temperature control method according to any one of claims 1 to 9.

Images