Liquid cooling server control method and system

Abstract

The present invention discloses a liquid-cooled server control method and system, and belongs to the field of computer technology. The present invention first obtains the first multi-mode liquid-cooled state monitoring data through state attribute identification, integrates the sensor information, and provides a comprehensive data basis for decision-making. Then, physical and security knowledge are split by modality, and different elements can be focused on to mine problems respectively. Next, the second multi-mode liquid-cooled state monitoring data is obtained based on the knowledge interaction of confidence, which can improve the accuracy of the overall state assessment. The semantic set obtained by further residual connection processing can clarify the logical relationship between the state and the decision. Finally, a control decision annotation log is generated, which can be used to provide accurate instructions, optimize the control process, improve heat dissipation efficiency, ensure the safe and stable operation of the server, extend the hardware life, and reduce the risk of failure and operating costs.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to a liquid cooling server control method and system. Background Art

[0002] With the rapid development of information technology, liquid-cooled servers are increasingly used in data centers and other fields. However, the efficient and stable operation of liquid-cooled servers faces many challenges. Traditional monitoring and control methods for liquid-cooled servers often have limitations.

[0003] In terms of data acquisition and integration, traditional methods cannot comprehensively and systematically collect various status information during the operation of liquid-cooled servers. In terms of status assessment and decision-making, traditional methods lack scientific and effective data fusion and logical analysis methods. In addition, traditional methods are relatively weak in analyzing the logical relationship between the status of liquid-cooled servers and control decisions. It can be seen that existing technologies are difficult to effectively deal with various possible operating conditions, which increases the risks and maintenance costs of server operation. Summary of the invention

[0004] The present invention provides a liquid cooling server control method and system, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0005] The present invention provides a liquid cooling server control method, which is implemented by a liquid cooling server control system. The method comprises: performing state attribute recognition on an original sensor monitoring log of a liquid cooling server to be controlled to obtain first multi-mode liquid cooling state monitoring data; performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge; wherein the physical state includes temperature state, flow state and pressure state, and the safety state includes leakage state, coolant impurity state and communication state; based on the physical state confidence and the safety state confidence, performing knowledge interaction on the first physical state monitoring knowledge and the first safety state monitoring knowledge to obtain second multi-mode liquid cooling state monitoring data, wherein the physical state confidence and the safety state confidence are respectively The safety status confidence is obtained by performing AI quantitative mining and analysis based on sensor monitoring log samples and corresponding control instruction log samples; the second multi-mode liquid cooling status monitoring data is subjected to residual connection processing through a first residual embedding module to obtain a semantic set of trend descriptions to be controlled, the semantic set of trend descriptions to be controlled represents the control logic comparison feature between the original sensor monitoring log and the corresponding control instruction log, and the first residual embedding module is obtained by performing feature mining based on the sensor monitoring log samples and the control instruction log samples; in combination with the original sensor monitoring log and the semantic set of trend descriptions to be controlled, a liquid cooling control decision annotation log of the liquid cooling server to be controlled is determined, and the control processing instructions for the liquid cooling server to be controlled are annotated in the liquid cooling control decision annotation log.

[0006] The present invention provides a liquid cooling server control system, comprising at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the above method.

[0007] The present invention provides a readable storage medium, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0008] The present invention first obtains the first multi-mode liquid cooling state monitoring data through state attribute identification, integrates sensor information, and provides a comprehensive data basis for decision-making. Then, physical and security knowledge are split out by modality, and different elements can be focused on to mine problems respectively. Next, the second multi-mode liquid cooling state monitoring data is obtained based on confidence knowledge interaction, which can improve the accuracy of the overall state assessment. The semantic set obtained by further residual connection processing can clarify the logical relationship between state and decision. Finally, a control decision annotation log is generated, which can be used to provide accurate instructions, optimize the control process, improve heat dissipation efficiency, ensure safe and stable operation of the server, extend hardware life, and reduce failure risks and operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a flow chart of a liquid cooling server control method provided by the present invention.

[0010] Figure 2 This is a schematic structural diagram of a liquid cooling server control system provided by the present invention. DETAILED DESCRIPTION

[0011] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0013] Figure 1 A liquid cooling server control method is shown, which is applied to a liquid cooling server control system. The method includes the following S110-S150.

[0014] S110: The liquid cooling server control system identifies the state attributes of the original sensor monitoring log of the liquid cooling server to be controlled, and obtains first multi-mode liquid cooling state monitoring data.

[0015] S120: The liquid cooling server control system performs modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge; wherein the physical state includes temperature state, flow state and pressure state, and the safety state includes leakage state, coolant impurity state and communication state.

[0016] S130: The liquid cooling server control system performs knowledge interaction on the first physical state monitoring knowledge and the first safety state monitoring knowledge based on the physical state confidence and the safety state confidence to obtain second multi-mode liquid cooling state monitoring data. The physical state confidence and the safety state confidence are obtained by performing AI quantitative mining and analysis on sensor monitoring log samples and corresponding control indication log samples.

[0017] S140: The liquid cooling server control system performs residual connection processing on the second multi-mode liquid cooling status monitoring data through the first residual embedding module to obtain a semantic set of descriptions of the trend to be controlled, wherein the semantic set of descriptions of the trend to be controlled represents the control logic comparison features between the original sensor monitoring log and the corresponding control indication log, and the first residual embedding module is obtained by feature mining based on the sensor monitoring log samples and the control indication log samples.

[0018] S150: The liquid cooling server control system determines a liquid cooling control decision annotation log of the liquid cooling server to be controlled in combination with the original sensor monitoring log and the semantic set of the trend description to be controlled, wherein the liquid cooling control decision annotation log is annotated with control processing instructions for the liquid cooling server to be controlled.

[0019] Furthermore, the following is a detailed introduction to the implementation process involved in S110-S150.

[0020] In S110, the liquid cooling server control system processes the original sensor monitoring log obtained from the liquid cooling server to be controlled. The original sensor monitoring log contains information about the operating status of the liquid cooling server collected by many sensors, such as the temperature values ​​at different times recorded by the temperature sensor, the coolant flow value recorded by the flow sensor, the pressure reading of the pressure sensor, and sensor data related to leakage detection, coolant impurity detection and communication status.

[0021] In order to obtain the first multi-mode liquid cooling status monitoring data, it is necessary to identify the status attributes. This process may involve operations such as data cleaning and feature extraction. For example, for temperature data, some obviously abnormal outliers may need to be removed. For example, in the original temperature data, 95% of the data is between 0-80°C, but there are individual data points reaching 150°C, which is likely caused by sensor failure or interference. By setting a reasonable range (such as a statistical range based on the temperature during normal operation of the liquid cooling server), these abnormal values ​​can be excluded.

[0022] In terms of feature extraction, for flow data, in addition to the directly obtained flow values, it is also possible to extract features such as the flow rate of change. For example, over a period of time, the flow rate drops from 10L / min to 8L / min, and the flow rate change rate is calculated to be -20% / unit time. By performing such comprehensive processing on the data of various modes in the original sensor monitoring log, it is converted into the first multi-mode liquid cooling status monitoring data that can accurately reflect the multi-mode liquid cooling status of the liquid-cooled server.

[0023] Next, S120 involves the key step of modal splitting to obtain two types of condition monitoring knowledge.

[0024] 1. Physical condition monitoring knowledge

[0025] 1) Temperature status: Separate the temperature-related data from the first multi-mode liquid cooling status monitoring data. This data may include the temperature values ​​of various key components (such as CPU, GPU, etc.) and their changes over time. For example, in a liquid-cooled server with multiple CPU cores, the temperature sensor of each core provides the temperature readings of the core at different times. For example, the temperature of the CPU1 core rises from 30°C to 50°C within 10 minutes. This temperature change trend is part of the temperature status.

[0026] 2) Flow status: The splitting of flow status data involves the collation of information related to coolant flow. As mentioned above, it includes the specific value of flow, the rate of change of flow, etc. Taking a liquid cooling system as an example, the flow of the coolant pump is 15L / min during stable operation. When the server load increases, the flow is adjusted to 20L / min. The flow values ​​and their changes in these different situations constitute the knowledge of flow status monitoring.

[0027] 3) Pressure state: The decomposition of pressure state focuses on the pressure data in the liquid cooling pipeline. The pressure sensor data at the inlet, outlet and key branch points of the liquid cooling pipeline are extracted. For example, the normal inlet pressure of the pipeline is 200kPa. When a valve is partially blocked, the pressure may rise to 250kPa. These pressure values ​​and their changes reflect the knowledge of pressure state monitoring.

[0028] 2. Safety status monitoring knowledge

[0029] 1) Leakage status: Sensor data related to leak detection is separated. For example, when using electrical detection methods, the conduction state of the leak detection circuit (conduction indicates a possible leak) and related resistance value changes. If the resistance value of the circuit suddenly changes from a high resistance state to a low resistance state, it may indicate that the coolant has leaked into the area where the detection circuit is located.

[0030] 2) Coolant impurity status: For the coolant impurity status, it may involve special impurity detection sensor data. For example, optical detection methods are used to determine the impurity content based on the change in the propagation characteristics of light in the coolant. When the light scattering coefficient increases from the normal 0.1 to 0.2, it may indicate that the impurity content in the coolant has increased.

[0031] 3) Communication status: Communication status monitoring knowledge mainly comes from the monitoring data of the communication link between the liquid cooling control system and various sensors and devices. For example, the communication status is judged by monitoring the loss rate of communication data packets. If the communication data packet loss rate rises from the normal 0.1% to 5%, it means that there is a problem with the communication status.

[0032] Then, the technical solution involved in S130 focuses on obtaining the second multi-mode liquid cooling state monitoring data based on confidence knowledge interaction.

[0033] 1. Obtaining Confidence

[0034] The physical state confidence and safety state confidence are obtained by performing AI quantitative mining and analysis on sensor monitoring log samples and corresponding control instruction log samples.

[0035] 1) Data preparation: Collect a large number of sensor monitoring log samples and corresponding control instruction log samples. For example, collect 1,000 sets of sensor monitoring logs and corresponding control instruction logs of liquid cooling servers in different operating scenarios. These scenarios include the operation of servers under different loads (low load, medium load, high load) and different ambient temperatures (10℃-30℃).

[0036] 2) AI quantitative mining analysis: Neural network algorithms (such as deep neural networks) can be used for analysis. Taking the temperature state as an example, the temperature sensor reading, temperature change rate, etc. are used as input features, and the corresponding control instructions (such as whether to adjust the coolant flow, whether to change the fan speed, etc.) are used as output labels. Through training on a large amount of data, the neural network can learn the relationship between the temperature state and the control instructions, thereby obtaining the confidence of the temperature state. For example, when the temperature rises to 80°C, according to the trained model, the system predicts that the confidence level of the coolant flow adjustment is 0.8.

[0037] 2. Knowledge Interaction Process

[0038] 1) Confidence-based weighting: The first physical state monitoring knowledge and the first safety state monitoring knowledge are weighted according to the physical state confidence and the safety state confidence. For example, the physical state confidence is 0.7 and the safety state confidence is 0.6. For a data point in the temperature state (such as the CPU temperature rises to 70°C), when interacting with the leakage state in the safety state, the temperature state data is weighted according to its confidence of 0.7, and the leakage state data is weighted according to its confidence of 0.6.

[0039] 2) Knowledge fusion: The weighted physical state monitoring knowledge and safety state monitoring knowledge are integrated to obtain the second multi-mode liquid cooling state monitoring data. For example, the weighted temperature state, flow state, pressure state and weighted leakage state, coolant impurity state, communication state and other data are combined according to certain rules to form a new comprehensive data structure. This structure contains the interactive information of each modal state data and more comprehensively reflects the operating status of the liquid cooling server.

[0040] Next, S140 may be used to perform residual connection processing to obtain a semantic set of descriptions of the trend to be controlled.

[0041] 1. Construction of the first residual embedding module

[0042] The first residual embedding module is obtained by performing feature mining based on sensor monitoring log samples and control instruction log samples.

[0043] 1) Feature mining process: A large number of sensor monitoring log samples and control indication log samples are used again. For example, algorithms such as principal component analysis (PCA) are used to perform feature mining on the data in these logs. For temperature status, the principal component features of temperature changes may be mined; for traffic status, the correlation features between traffic and server load may be mined, etc.

[0044] 2) Constructing the residual embedding module

[0045] The first residual embedding module is constructed based on the mined features. This module contains multiple submodules, each of which is responsible for processing the residual connection of a modal state data. For example, there are submodules specifically processing the residual connection of the temperature state, submodules for the flow state, etc.

[0046] (II) Residual Connection Processing

[0047] 1) Calculate the residual: For each modal data in the second multi-mode liquid cooling state monitoring data, calculate the residual between it and the corresponding data in the original sensor monitoring log. Take the flow state as an example. For example, if the original flow data is 15L / min, and the flow data after processing in the previous step is 14L / min, then the residual is -1L / min.

[0048] 2) Residual connection and semantic set generation: The residuals of each mode are connected through the first residual embedding module to obtain the semantic set describing the trend to be controlled. This semantic set represents the control logic comparison characteristics between the original sensor monitoring log and the corresponding control indication log. For example, when the temperature residual is positive (indicating a temperature increase) and the flow residual is negative (indicating a flow decrease), the semantic set will contain the relationship between the temperature and flow change trends and the control logic, which may imply that the coolant flow needs to be increased to control the temperature rise.

[0049] Finally, the liquid cooling control decision annotation log is determined according to S150.

[0050] 1) Comprehensive analysis: The liquid cooling server control system combines the original sensor monitoring log and the semantic set of trend descriptions to be controlled for comprehensive analysis. For example, the original sensor monitoring log shows that the CPU temperature continues to rise, and the semantic set of trend descriptions to be controlled indicates that the temperature rise is associated with the flow reduction and the flow needs to be increased to control the temperature.

[0051] 2) Determine the control decision annotation log: Based on the results of the comprehensive analysis, determine the liquid cooling control decision annotation log for the liquid cooling server to be controlled. This annotation log is annotated with specific control processing instructions. For example, the annotation log may record control processing instructions such as "increase the speed of the coolant pump and increase the flow rate from 10L / min to 12L / min to control the CPU temperature below 70°C", or "check the leakage detection circuit because the communication status shows that the increase in the packet loss rate may be related to leakage", thereby providing a clear operating guide for the effective control of the liquid cooling server.

[0052] Through the above steps, the liquid-cooled server control system can effectively utilize the sensor monitoring data, and after complex processing and analysis, finally obtain a reasonable liquid cooling control decision annotation log to ensure the efficient and stable operation of the liquid-cooled server.

[0053] Based on the above content, the key feature vectors involved in the control decision annotation log are determined to include but are not limited to the first multi-mode liquid cooling state monitoring data, the first physical state monitoring knowledge, the first safety state monitoring knowledge, the second multi-mode liquid cooling state monitoring data, the semantic set of descriptions of trends to be controlled, and the control logic comparison features.

[0054] First, the first multi-mode liquid cooling state monitoring data, as a set reflecting the multi-mode states of the liquid cooling server, can be represented by a feature vector. For example, the liquid cooling server has n key temperature monitoring points, m flow monitoring points, and p pressure monitoring points, as well as discrete monitoring values ​​related to leakage, coolant impurities, and communication status. Then the feature vector of the first multi-mode liquid cooling state monitoring data can be expressed as (vec{X}=[x_1, x_2, cdots, x_{n+m+p+q+r}]), where (x_1-x_n) may represent the temperature values ​​of different temperature monitoring points, (x_{n+1}-x_{n+m}) represents the flow value, (x_{n+m+1}-x_{n+m+p}) represents the pressure value, (x_{n+m+p+1}-x_{n+m+p+q}) represents the state value related to leakage and coolant impurities (such as using 0-1 to indicate whether there is a sign of leakage, etc.), and (x_{n+m+p+q+1}-x_{n+m+p+q+r}) represents the value related to the communication state (such as the communication data packet loss rate, etc.).

[0055] For the first type of physical state monitoring knowledge, since it mainly involves temperature, flow and pressure states. For example, there are (n) monitoring points for temperature, (m) monitoring points for flow, and (p) monitoring points for pressure. Then its eigenvector can be expressed as (vec{Y}=[y_1, y_2, cdots, y_{n+m+p}]), where (y_1-y_n) is the temperature value, (y_{n+1}-y_{n+m}) is the flow value, and (y_{n+m+1}-y_{n+m+p}) is the pressure value. For example, in a small liquid-cooled server, (n=3) (temperature monitoring points for CPU, GPU and memory module respectively), (m=2) (coolant inlet and outlet flow monitoring points), (p=3) (pressure monitoring points for three key nodes of the liquid cooling pipeline), then (vec{Y}) contains the values ​​of these 8 physical quantities.

[0056] For the first safety status monitoring knowledge, the first safety status monitoring knowledge includes leakage status, coolant impurity status and communication status. It can be represented as a feature vector (vec{Z}=[z_1, z_2, z_3]), where (z_1) represents the leakage status (for example, 0 represents no leakage, 1 represents leakage signs), (z_2) represents the coolant impurity status (which can be the quantitative value of the impurity content or the 0-1 value converted according to the preset threshold), and (z_3) represents the communication status (such as the communication data packet loss rate or the 0-1 value converted according to the communication quality to indicate whether the communication is normal or not).

[0057] Based on the above technical features, the second multi-mode liquid cooling state monitoring data is obtained through confidence-based knowledge interaction. Its feature vector is obtained on the basis of the fusion of physical and safety state monitoring knowledge feature vectors. For example, the physical state monitoring knowledge feature vector is (vec{Y}), and the safety state monitoring knowledge feature vector is (vec{Z}), and weighted fusion is performed according to the confidence (w_1) (physical state confidence) and (w_2) (safety state confidence). Then the feature vector (vec{A}) of the second multi-mode liquid cooling state monitoring data can be expressed as (vec{A}=w_1vec{Y}+w_2vec{Z}), the addition of the present invention is the addition of corresponding elements, and in practical applications, reasonable normalization processing needs to be performed according to the nature and physical meaning of the data.

[0058] Furthermore, the semantic set of trend description to be controlled represents the control logic comparison features between the original sensor monitoring log and the corresponding control indication log. Its feature vector can be expressed as (vec{B}=[b_1, b_2, cdots, b_k]), where (b_i) represents different control logic comparison features. For example, (b_1) may represent the relationship between the temperature increase trend and flow adjustment, (b_2) may represent the relationship between pressure change and leak detection, etc. These features are obtained by residual connection processing of the second multi-mode liquid cooling state monitoring data, and each element contains the semantic information of the state change trend and control logic of the liquid cooling server.

[0059] Finally, the control logic comparison feature reflects the relationship between the state of the liquid cooling server and the control decision. Each element of its feature vector corresponds to a specific comparison relationship. For example, in a feature vector element, if it represents the temperature state and the coolant flow control logic comparison feature, when the temperature rises, the corresponding element value may be expressed as a quantitative relationship that requires an increase in flow (such as a numerical value obtained based on the historical data relationship between the temperature increase and the flow adjustment range). These feature vector elements together constitute the control logic comparison feature, which provides an important basis for determining liquid cooling control decisions.

[0060] To facilitate understanding of the above technical solution, an actual application scenario example is used to introduce it below. First, an exemplary introduction to the original sensor monitoring log is given.

[0061]

Original sensor monitoring log

[0062] [Time stamp: 2023-01-01 10:00:00]

[0063] Temperature monitoring: CPU temperature is 40°C, GPU temperature is 35°C, memory module temperature is 30°C. Liquid cooling system inlet temperature is 25°C, outlet temperature is 30°C.

[0064] Flow monitoring: coolant inlet flow is 10L / min, outlet flow is 9.5L / min.

[0065] Pressure monitoring: The inlet pressure of the liquid cooling pipeline is 150kPa, the outlet pressure is 120kPa, and the pressure of the intermediate key node is 130kPa.

[0066] Leak detection: No leakage signs (0 means).

[0067] Coolant impurity detection: The impurity content is low (quantitative value is 0.1).

[0068] Communication status: The communication data packet loss rate is 0.5%.

[0069] [Time stamp: 2023-01-01 10:05:00]

[0070] Temperature monitoring: CPU temperature is 42°C, GPU temperature is 36°C, memory module temperature is 31°C. Liquid cooling system inlet temperature is 26°C, outlet temperature is 31°C.

[0071] Flow monitoring: coolant inlet flow is 10L / min, outlet flow is 9.4L / min.

[0072] Pressure monitoring: The inlet pressure of the liquid cooling pipeline is 152kPa, the outlet pressure is 122kPa, and the pressure of the intermediate key node is 132kPa.

[0073] Leak detection: No leakage signs (0 means).

[0074] Coolant impurity detection: The impurity content is low (quantitative value is 0.1).

[0075] Communication status: The communication packet loss rate is 0.6%.

[0076] [Time stamp: 2023-01-01 10:10:00]

[0077] Temperature monitoring: CPU temperature is 45°C, GPU temperature is 38°C, memory module temperature is 33°C. Liquid cooling system inlet temperature is 27°C, outlet temperature is 33°C.

[0078] Flow monitoring: coolant inlet flow is 10L / min, outlet flow is 9.3L / min.

[0079] Pressure monitoring: The inlet pressure of the liquid cooling pipeline is 155kPa, the outlet pressure is 125kPa, and the pressure of the intermediate key node is 135kPa.

[0080] Leak detection: No leakage signs (0 means).

[0081] Coolant impurity detection: The impurity content is low (quantitative value is 0.1).

[0082] Communication status: The communication data packet loss rate is 0.7%.

[0083] Then, through the above technical solution, an example of a liquid cooling control decision annotation log corresponding to the exemplary introduction of the original sensor monitoring log can be generated.

[0084] [Liquid cooling control decision annotation log]

[0085] [Time stamp: 2023-01-01 10:00:00]

[0086] Control decision: The current server temperature, flow, pressure, and safety-related status are all within the normal range. Maintain the current coolant pump speed without adjusting the flow. Continue to monitor various status parameters and pay attention to the communication status. Although the packet loss rate of 0.5% is within the acceptable range, if it continues to rise, the communication link needs to be checked.

[0087] [Time stamp: 2023-01-01 10:05:00]

[0088] Control decision: The temperature has a slight upward trend, the flow rate has a slight downward trend, and the pressure has increased slightly. Since the overall change is small and still within the normal range, the coolant flow rate will not be adjusted for the time being. Pay close attention to temperature changes, especially CPU temperature. For communication status, the packet loss rate has risen to 0.6%. Arrange to check whether the communication equipment (such as network cards, switches, etc.) is loose during the next inspection, but no immediate measures are required.

[0089] [Time stamp: 2023-01-01 10:10:00]

[0090] Control decision: The temperature continues to rise, especially the CPU temperature has reached 45°C. Although it has not exceeded the critical value (for example, the critical value is 50°C), considering the continuous decrease in flow, it is decided to fine-tune the speed of the coolant pump and increase the inlet flow from 10L / min to 10.5L / min. As for the pressure increase, since the amplitude is within the normal fluctuation range, no processing is done for the time being. There is no abnormality in the leak detection, and the impurity content of the coolant is stable, and monitoring will continue. In terms of communication status, the packet loss rate has risen to 0.7%. In addition to checking the connection of the communication equipment, it is also necessary to check whether the communication line is subject to electromagnetic interference. Preliminary troubleshooting can be performed by using electromagnetic shielding equipment. If the packet loss rate exceeds 1% in the next 10 minutes, consider restarting the communication-related equipment.

[0091] It can be seen that the present invention has many significant beneficial effects in controlling liquid cooling servers. First, by identifying the state attributes of the original sensor monitoring log to obtain the first multi-mode liquid cooling state monitoring data, various sensor information during the operation of the liquid cooling server can be fully integrated, avoiding data fragmentation, so that the state information of the liquid cooling server can be completely acquired and processed by the system. This helps to more accurately grasp the operating status of the server and provide a comprehensive data basis for subsequent decision-making.

[0092] In the mode splitting stage, the first multi-mode liquid cooling state monitoring data is split into the first physical state monitoring knowledge and the first safety state monitoring knowledge. This classification helps to focus on the physical parameters and safety-related key elements of the liquid cooling server operation. Special analysis of physical states such as temperature, flow and pressure, as well as safety states such as leakage, coolant impurities and communication, can more carefully explore possible problems in different categories of states, thereby achieving targeted monitoring and processing.

[0093] The second multi-mode liquid cooling state monitoring data is obtained by knowledge interaction based on the physical state confidence and the safety state confidence. This process makes full use of the confidence information obtained by AI quantitative mining and analysis based on a large number of sensor monitoring log samples and control instruction log samples. This data-driven approach can accurately weigh the importance of different types of state knowledge, making the fused second multi-mode liquid cooling state monitoring data more representative and reliable, and effectively improving the accuracy of the overall state assessment of the liquid cooling server.

[0094] The first residual embedding module performs residual connection processing on the second multi-mode liquid cooling state monitoring data to obtain a semantic set of trend descriptions to be controlled, which can clearly represent the control logic comparison characteristics between the original sensor monitoring log and the control indication log. In this way, we can deeply understand the inherent logical relationship between the state change of the liquid cooling server and the control decision, providing a strong basis for accurately formulating control strategies.

[0095] Finally, the liquid cooling control decision annotation log is determined by combining the original sensor monitoring log and the semantic set of the trend description to be controlled, which can provide accurate and targeted control processing instructions for the liquid cooling server. This helps to optimize the control process of the liquid cooling server, improve the heat dissipation efficiency, ensure that the server runs in a safe and stable state, and at the same time extend the service life of the server hardware, reduce the risk of failures caused by overheating, leakage and other problems, and reduce operating costs.

[0096] Based on the above content, in an optional embodiment, S120 is designed to perform modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge, including: splitting the first multi-mode liquid cooling state monitoring data into second physical state monitoring knowledge and second safety state monitoring knowledge according to the modal category, performing knowledge interaction on the second physical state monitoring knowledge and the second safety state monitoring knowledge to obtain third multi-mode liquid cooling state monitoring data; performing knowledge mining based on time series features on the third multi-mode liquid cooling state monitoring data to obtain the first physical state monitoring knowledge, and performing knowledge mining based on decision trees on the third multi-mode liquid cooling state monitoring data to obtain the first safety state monitoring knowledge.

[0097] The exemplary expansion of the technical solutions described in the above embodiments is described as follows.

[0098] 1. Modal Decomposition and Knowledge Interaction

[0099] 1. Disassembly by modal category

[0100] 1) Data structure analysis: The first multi-mode liquid cooling state monitoring data is a collection of multiple state monitoring information of liquid cooling servers. When it is decomposed into the second physical state monitoring knowledge and the second safety state monitoring knowledge according to the modal category, it is necessary to have a clear understanding of the structure and composition of the data. For example, the first multi-mode liquid cooling state monitoring data is represented in the form of a feature vector. The elements in the vector are arranged in a certain order. The first few elements may correspond to temperature state data (such as CPU temperature, GPU temperature, etc.), followed by flow state data (such as coolant inlet flow, outlet flow), and then pressure state data (such as pressure values ​​at different positions of the liquid cooling pipeline), and the second half of the vector contains data related to leakage state (0 or 1 to indicate whether it is leaked), coolant impurity state (which may be a quantitative value of the impurity content) and communication state (such as communication data packet loss rate, etc.).

[0101] According to such a structure, physical and safety status data are distinguished through specific algorithms or rules. An index range or marking mechanism can be set, for example, elements with indexes of 1-n (n is the last element index of physical status data) are classified as the second physical status monitoring knowledge, and elements with indexes of n+1-m (m is the length of the entire feature vector) are classified as the second safety status monitoring knowledge.

[0102] 2) Taking an actual liquid cooling server monitoring data as an example, the feature vector of the first multi-mode liquid cooling state monitoring data is ([30, 35, 10, 9.5, 150, 120, 0, 0.1, 0.5%]), of which the first three values ​​(30) (CPU temperature), (35) (GPU temperature) and (10) (coolant inlet flow) belong to physical state data, and the last three values ​​(0) (no leakage), (0.1) (coolant impurity content) and (0.5%) (communication packet loss rate) belong to security state data. In this way, the first multi-mode liquid cooling state monitoring data is accurately decomposed into the second physical state monitoring knowledge ([30, 35, 10, 9.5, 150, 120]) and the second security state monitoring knowledge ([0, 0.1, 0.5%]).

[0103] (II) Knowledge interaction to obtain the third multi-mode liquid cooling state monitoring data

[0104] 1) Interaction principle: After obtaining the second physical state monitoring knowledge and the second safety state monitoring knowledge, knowledge interaction is performed. This knowledge interaction aims to explore the potential relationship between the two types of knowledge, so as to construct the third multi-mode liquid cooling state monitoring data that more comprehensively reflects the state of the liquid-cooled server. For example, the temperature increase in the physical state may be related to the increase of coolant impurities in the safety state (the increase of impurities may affect the heat dissipation performance of the coolant, thereby causing the temperature to rise).

[0105] The knowledge interaction can be performed by weighted summation. For example, the feature vector of the second physical state monitoring knowledge is (vec{A}=[a_1, a_2, cdots, a_n]), and the feature vector of the second safety state monitoring knowledge is (vec{B}=[b_1, b_2, cdots, b_m]). According to the predetermined weight vectors (vec{W}_A=[w_{a1}, w_{a2}, cdots, w_{an}]) and (vec{W}_B=[w_{b1}, w_{b2}, cdots, w_{bm}]), the interacted feature vector (vec{C}) is calculated, where the element (c_i) ((i=1, 2, cdots, n+m)) of (vec{C}) is calculated as follows: for (ileqn), (c_i=a_itimesw_{ai}); for (i>n), (c_i=(b_{in}timesw_{b(in)})).

[0106] 2) Numerical examples and weight determination: for example, the second physical state monitoring knowledge (vec{A}=[30, 35, 10, 9.5, 150, 120]), weight vector (vec{W}_A=[0.1, 0.1, 0.2, 0.2, 0.2, 0.2]); the second safety state monitoring knowledge (vec{B}=[0, 0.1, 0.5%]), weight vector (vec{W}_B=[0.3, 0.3, 0.4]). Some elements of the third multi-mode liquid cooling state monitoring data (vec{C}) are calculated: (c_1=30times0.1=3), (c_2=35times0.1=3.5), (c_3=10times0.2=2), (cdots), (c_7=(0.1times0.3)=0.03), (c_8=(0.5%times0.4)=0.002).

[0107] 2. Obtaining the first physical condition monitoring knowledge based on knowledge mining based on time series characteristics

[0108] 1.1 Timing feature analysis

[0109] 1) Data preparation: For the physical state part of the third multi-mode liquid cooling state monitoring data, knowledge mining based on time series features is performed. First, the third multi-mode liquid cooling state monitoring data within a period of time needs to be collected. For example, with a time interval of every 5 minutes, collect the liquid cooling server state data for 1 hour (a total of 12 time intervals). The third multi-mode liquid cooling state monitoring data corresponding to each time interval contains elements related to the physical state.

[0110] Arrange these data in chronological order to form a time series data set. For example, for the temperature state, a time series temperature data set ([T_1, T_2, cdots, T_{12}]) is obtained, where (T_i) represents the temperature data vector of the (i)th time interval (including CPU temperature, GPU temperature, etc.).

[0111] 2) Mining algorithm: The autoregressive moving average (ARMA) model can be used to mine time series features. The general form of the ARMA model is:

[0112] (y_t=c+sum_{i=1}^pvarphi_iy_{ti}+sum_{j=1}^qtheta_jepsilon_{tj}), where (y_t) is the observation value at the current moment (in the example of the present invention, such as the temperature value at the current moment), (c) is a constant term, (varphi_i) and (theta_j) are parameters of the model, (p) and (q) are the orders of autoregression and moving average, respectively, and (epsilon_{tj}) is a white noise sequence.

[0113] Take CPU temperature as an example. For example, if (p=2) and (q=1) have been determined, the model parameters (varphi_1), (varphi_2) and (theta_1) are estimated by fitting the time series temperature data set with an ARMA model. These parameters reflect the change pattern of CPU temperature over time. For example, if (varphi_1) is large and positive, it means that the current CPU temperature has a strong positive correlation with the temperature at the previous moment.

[0114] 2. Obtaining the first physical condition monitoring knowledge

[0115] 1) Knowledge representation: Through knowledge mining based on time series features, the first physical state monitoring knowledge obtained includes the changing rules and mutual relationships of physical states (temperature, flow, pressure) in the time dimension. For example, in addition to the specific values ​​of temperature, flow and pressure at each moment, we also know the temperature rise rate (calculated based on the ARMA model parameters), the fluctuation trend of flow and pressure over time, etc.

[0116] This knowledge can be represented in the form of an extended feature vector or knowledge graph. In the feature vector, in addition to the original physical state value, elements representing time series characteristics are added, such as temperature rise rate, flow fluctuation amplitude, etc. In the knowledge graph, nodes represent different physical states, and edges represent the time series relationship between them. For example, there is a directed edge between the temperature increase node and the flow reduction node, indicating that the temperature increase may lead to a decrease in flow (based on the mined time series relationship).

[0117] 3. Obtaining the first safety status monitoring knowledge based on knowledge mining of decision tree

[0118] 1. Decision Tree Construction

[0119] 1) Feature selection and data set division: For the safety status part of the third multi-mode liquid cooling status monitoring data, a decision tree is used for knowledge mining. First, the features in the safety status are selected as the input features of the decision tree, for example, leakage status (whether leakage), coolant impurity status (impurity content) and communication status (packet loss rate) are used as input features.

[0120] The data set containing these security status data is divided into a training set and a test set. For example, the ratio is 80:20, where 80% of the data is used to build the decision tree model and 20% of the data is used to test the accuracy of the model.

[0121] 2) Decision tree algorithm: The C4.5 algorithm can be used to build a decision tree. The C4.5 algorithm selects the best split attribute by calculating the information gain rate. For example, for a security status data set, the information gain rate of each attribute (leakage status, coolant impurity status, communication status) is calculated. For example, if the coolant impurity status has the highest information gain rate during the calculation process, then the coolant impurity status is used as the split attribute at the root node of the decision tree.

[0122] The data set is further divided according to the value of the split attribute. For example, if the coolant impurity state is divided into three levels: low, medium, and high, then the data set is divided into three subsets according to the different levels of the impurity state, and then the information gain rate is continued to be calculated on each subset, and the next split attribute is selected until the stopping condition is met (such as the number of samples in all subsets is less than a certain threshold or the information gain rate is less than a set value).

[0123] 2. Obtaining the first safety condition monitoring knowledge

[0124] 1) Decision rule extraction: Through the constructed decision tree, a series of decision rules can be extracted. For example, one of the decision rules may be: if the coolant impurity state is high and the communication packet loss rate is greater than 1%, there is a high possibility of security risks (such as possible damage to server hardware or communication failure).

[0125] These decision rules constitute the first safety status monitoring knowledge. This knowledge can be expressed in the form of a rule set, which is convenient for use in the safety monitoring and decision-making of liquid cooling servers. When new safety status data is input, these rules can be used to quickly determine whether there is a safety risk and take corresponding measures.

[0126] It can be seen that in the process of modal splitting, the third multi-mode liquid cooling state monitoring data is constructed through knowledge interaction. This method can mine the potential connection between physical and safety states, such as integrating the two types of knowledge through weighted summation, so that the data can more comprehensively and accurately reflect the server status. When mining the first physical state monitoring knowledge based on time series characteristics, the ARMA model is used to analyze the time series change rules of physical states such as temperature, flow, and pressure. For example, the temperature rise rate is clarified through model parameters, which helps to predict the change trend of the physical state in advance. Mining the first safety state monitoring knowledge based on the decision tree, such as the C4.5 algorithm to construct a decision tree to derive decision rules, can quickly and accurately judge safety risks, such as judging safety risks based on coolant impurities and communication status, providing guarantees for the safe and stable operation of liquid cooling servers, and overall improving the effectiveness of server monitoring, management, and control.

[0127] In some other replaceable embodiments, the method also includes: performing residual connection processing on the third multi-mode liquid cooling state monitoring data through a second residual embedding module to obtain a discriminant knowledge mining module, wherein the second residual embedding module is obtained by performing feature mining based on past sensor monitoring logs and the control indication logs; determining a temporal knowledge mining module using a long short-term memory module and the discriminant knowledge mining module, wherein the number of attention indicators of the long short-term memory module is consistent with the number of attention indicators of the discriminant knowledge mining module, and the attention weight of one of the attention indicators of the long short-term memory module is 1, and the attention weights of the remaining attention indicators are 0. Based on this, the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on time series features to obtain the first physical state monitoring knowledge, and the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on a decision tree to obtain the first safety state monitoring knowledge, including: performing knowledge mining based on time series features on the third multi-mode liquid cooling state monitoring data through the time series knowledge mining module to obtain the first physical state monitoring knowledge, and performing knowledge mining based on a decision tree on the third multi-mode liquid cooling state monitoring data through the discrimination knowledge mining module to obtain the first safety state monitoring knowledge.

[0128] In the next possible expansion step, the number of the discriminant knowledge mining modules is a, and the number of attention indicators of each discriminant knowledge mining module is b; the number of the time series knowledge mining modules is a, and the number of attention indicators of each time series knowledge mining module is b; the third multi-mode liquid cooling state monitoring data is subjected to time series feature-based knowledge mining by the time series knowledge mining module to obtain the first physical state monitoring knowledge, and the third multi-mode liquid cooling state monitoring data is subjected to decision tree-based knowledge mining by the discriminant knowledge mining module to obtain the first safety state monitoring knowledge, including: decomposing the third multi-mode liquid cooling state monitoring data into a third local liquid cooling state monitoring knowledge based on the attention indicator, wherein the discriminant knowledge mining The number of feature units corresponding to the mining module and the number of feature units corresponding to the timing knowledge mining module are consistent with the number of feature units corresponding to the third local liquid cooling state monitoring knowledge; through a said timing knowledge mining modules, the a said third local liquid cooling state monitoring knowledge are respectively subjected to knowledge mining based on timing features to obtain a first local physical state monitoring vectors, and the a said first local physical state monitoring vectors form the first physical state monitoring knowledge; through a said discriminant knowledge mining modules, the a said third local liquid cooling state monitoring knowledge are respectively subjected to knowledge mining based on decision trees to obtain a first-stage safety state monitoring vectors, and the a said first-stage safety state monitoring vectors form the first safety state monitoring knowledge.

[0129] In actual application, exemplary implementations of the above two embodiments are as follows.

[0130] 1. Second residual embedding module and discriminative knowledge mining module

[0131] 1) Construction of the second residual embedding module: The second residual embedding module is obtained by feature mining based on past sensor monitoring logs and control instruction logs. Past sensor monitoring logs contain various status monitoring data of the liquid-cooled server during past operation, such as temperature, flow, pressure, leakage, coolant impurities, and communication status. The control instruction log records the control measures taken for these states, such as adjusting the specific value of the coolant flow or changing the speed of the cooling fan when the temperature reaches a certain threshold.

[0132] When constructing the second residual embedding module, feature engineering techniques may be used. For example, the data in the past sensor monitoring logs and control indication logs are standardized to make data of different magnitudes comparable. For example, for temperature data, the original range is 0-100°C, and it is converted to the range of 0-1 through standardization. For flow data, the original unit is liters / minute, and corresponding standardization is also performed.

[0133] Then, the principal component analysis (PCA) algorithm is used for feature mining. The PCA algorithm projects high-dimensional data into a low-dimensional space by finding the principal component direction of the data while retaining the main information of the data. For example, a high-dimensional vector containing multiple modal state data (such as temperature, flow, etc.) is projected onto several principal component directions, and these principal components constitute part of the second residual embedding module.

[0134] 2) Obtaining the residual connection processing and discriminant knowledge mining module: The third multi-mode liquid cooling state monitoring data is subjected to residual connection processing using the second residual embedding module. The purpose of residual connection processing is to better capture subtle differences and feature changes in the data. For example, for the temperature data sequence (T={t_1, t_2, cdots, t_N}) and flow data sequence (F={f_1, f_2, cdots, f_N}) in the third multi-mode liquid cooling state monitoring data, the residuals (R_T=t_{n+1}-t_n) and (R_F=f_{n+1}-f_n) between them are calculated (here (n) represents a moment in the time series).

[0135] The residual information is combined with the original third multi-mode liquid cooling state monitoring data to obtain a discriminative knowledge mining module through a specific mapping relationship. This mapping relationship may be a neural network model, for example, using the residual information and the original data as the input layer, after several hidden layers of neuron calculations, the output of the discriminative knowledge mining module is finally obtained.

[0136] 2. Long- and short-term memory module and temporal knowledge mining module

[0137] 1) Characteristics of the Long Short-Term Memory Module (LSTM): The Long Short-Term Memory Module is a special recurrent neural network (RNN) structure that has the ability to process time series data. In this embodiment, the number of its attention indexes is consistent with the number of attention indexes of the discriminative knowledge mining module. The attention weight of one of the attention indexes of the Long Short-Term Memory Module is 1, and the attention weights of the remaining attention indexes are 0. This special attention weight setting enables the Long Short-Term Memory Module to focus on specific information.

[0138] For example, when processing the temperature time series data of a liquid-cooled server, the LSTM module has three attention indicators, one of which focuses on short-term fluctuations in temperature (attention weight is 1), while the other two focus on long-term trends and periodic changes (attention weight is 0). When a temperature data sequence is input, the LSTM module can selectively process the data based on this attention weight setting.

[0139] 2) Determine the time series knowledge mining module: Use the long short-term memory module and the discriminant knowledge mining module to determine the time series knowledge mining module. Specifically, by combining or associating the long short-term memory module and the discriminant knowledge mining module. For example, the output of the long short-term memory module can be used as part of the input of the time series knowledge mining module, and some features of the discriminant knowledge mining module can also be used as input. Then, a fusion function (such as weighted summation or nonlinear mapping function) is used to construct the time series knowledge mining module.

[0140] 3. Knowledge mining based on temporal knowledge mining module and discriminant knowledge mining module

[0141] 1) Decompose the third multi-mode liquid cooling state monitoring data: the number of discriminative knowledge mining modules is (a), the number of attention indicators of each discriminative knowledge mining module is (b), the number of temporal knowledge mining modules is also (a), and the number of attention indicators of each temporal knowledge mining module is (b). Decompose the third multi-mode liquid cooling state monitoring data into (a) third local liquid cooling state monitoring knowledge based on the attention indicators.

[0142] For example, the feature vector of the third multi-mode liquid cooling state monitoring data is (vec{X}=[x_1, x_2, cdots, x_n]), if (a=3), (b=5), then according to the pre-set rules, (vec{X}) can be decomposed into three third local liquid cooling state monitoring knowledge: (vec{X}_1=[x_1, x_2, cdots, x_{n / 3}]), (vec{X}_2=[x_{n / 3+1}, x_{n / 3+2}, cdots, x_{2n / 3}]) and (vec{X}_3=[x_{2n / 3+1}, x_{2n / 3+2}, cdots, x_n]), and ensure that the number of feature units corresponding to the discriminant knowledge mining module and the time series knowledge mining module is consistent with the number of feature units corresponding to the third local liquid cooling state monitoring knowledge.

[0143] 2) Obtain the first physical state monitoring knowledge based on time series feature knowledge mining: Through (a) time series knowledge mining modules, (a) third local liquid cooling state monitoring knowledge is respectively mined based on time series features. Taking one of the time series knowledge mining modules as an example, for the temperature data sequence in the third local liquid cooling state monitoring knowledge, the difference method can be used to mine time series features. For example, if the temperature data sequence is (T'={t'_1, t'_2, cdots, t'_m}), calculate the first-order difference (Deltat'_i=t'_{i+1}-t'_i)((i=1, 2, cdots, m-1)).

[0144] These differential data are used as feature elements and together with the original temperature data, a new feature vector is constructed (vec{T}''=[t'_1, Deltat'_1, t'_2, Deltat'_2, cdots, t'_{m-1}, Deltat'_{m-1}]). The same method can be applied to flow and pressure data. After being processed by (a) time series knowledge mining modules, (a) first local physical state monitoring vectors are obtained. For example, when (a=3), the three first local physical state monitoring vectors (vec{V}_1), (vec{V}_2) and (vec{V}_3) are obtained, which together form the first physical class state monitoring knowledge.

[0145] 3) Obtaining the first safety status monitoring knowledge through decision tree-based knowledge mining: Through (a) discriminant knowledge mining modules, (a) third local liquid cooling status monitoring knowledge is respectively mined based on decision trees. For example, for each leakage status, coolant impurity status, and communication status data in the third local liquid cooling status monitoring knowledge, a decision tree can be constructed.

[0146] Take the third local liquid cooling state monitoring knowledge processed by one of the discriminant knowledge mining modules as an example. For example, the leakage state is represented by (L) (value 0 or 1), the coolant impurity state is represented by (I) (quantitative value), and the communication state is represented by (C) (such as communication packet loss rate). When constructing a decision tree, the root node can be divided according to (L). If (L=0), it is further divided according to (I). If (I<0.1), it is divided according to (C), etc. Through such a decision tree construction and classification process, (a) first-stage safety state monitoring vectors are obtained. When (a=3), the three first-stage safety state monitoring vectors (vec{S}_1), (vec{S}_2) and (vec{S}_3) are obtained, which together form the first safety class state monitoring knowledge.

[0147] It can be seen that the discriminative knowledge mining module is constructed through the second residual embedding module, and the residual connection processing can better capture the changes in data features. The long short-term memory module and the discriminative knowledge mining module determine the temporal knowledge mining module, and the special attention weight setting makes the information processing more targeted. The data is broken down into multiple local knowledge for mining, which improves the accuracy and comprehensiveness of knowledge mining. When mining physical knowledge based on temporal features, the differential method can effectively extract temporal features. Mining security knowledge based on decision trees can accurately classify security status, which helps to grasp the status of liquid-cooled servers more accurately overall.

[0148] In some technical solutions, the first multi-mode liquid cooling state monitoring data is modally split to obtain first physical state monitoring knowledge and first safety state monitoring knowledge in S120, including: modally splitting the first multi-mode liquid cooling state monitoring data to obtain intermediate physical state monitoring knowledge and intermediate safety state monitoring knowledge, and based on the physical state confidence and the safety state confidence, performing knowledge interaction between the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge to obtain intermediate multi-mode liquid cooling state monitoring data; performing modal splitting of the intermediate multi-mode liquid cooling state monitoring data in the next cycle to obtain To the intermediate physical state monitoring knowledge of the next level and the intermediate safety state monitoring knowledge of the next level, according to the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge of the next level and the intermediate safety state monitoring knowledge of the next level are interacted to obtain the intermediate multi-mode liquid cooling state monitoring data of the next level, until the number of modal splittings of the current cycle reaches the first set number, the intermediate physical state monitoring knowledge obtained in the current cycle is determined as the first physical state monitoring knowledge, and the intermediate safety state monitoring knowledge obtained in the current cycle is determined as the first safety state monitoring knowledge.

[0149] Further, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge are interacted with each other according to the physical state confidence and the safety state confidence to obtain the intermediate multi-mode liquid cooling state monitoring data, which can be realized by two embodiments:

[0150] Embodiment 1: On the basis that the number of modal splits in the current cycle is not greater than the second set number, according to the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision to obtain first liquid cooling state monitoring collision data, and the second set number is less than the first set number; the original sensor monitoring log is subjected to semantic fine-grained adjustment, and the state attribute of the original sensor monitoring log after the semantic fine-grained adjustment is identified to obtain second liquid cooling state monitoring collision data; the state attribute of the first liquid cooling state monitoring collision data and the second liquid cooling state monitoring collision data is identified through a feature focusing strategy to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle;

[0151] Example 2: On the basis that the number of modal splits in the current cycle is greater than the third set number, according to the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision to obtain third liquid cooling state monitoring collision data, and the third set number is not greater than the first set number; at least one intermediate multi-mode liquid cooling state monitoring data obtained in the cycle before the current cycle is integrated with the third liquid cooling state monitoring collision data to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

[0152] The above technical solution focuses on modal splitting and knowledge interaction process, which can be divided into the following three design ideas.

[0153] 1. Overall modal splitting framework

[0154] In S120, the process of performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain the first physical state monitoring knowledge and the first safety state monitoring knowledge is not a simple splitting, but a multi-cycle, step-by-step optimization process.

[0155] First, the first multi-mode liquid cooling state monitoring data is modally split to obtain intermediate physical state monitoring knowledge and intermediate safety state monitoring knowledge. The intermediate state monitoring knowledge of the present invention is the result of the preliminary splitting. Then, based on the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge are knowledge-interacted to obtain intermediate multi-mode liquid cooling state monitoring data.

[0156] Then, the next round of modal splitting is performed, and the intermediate multi-mode liquid cooling state monitoring data is split into the next cycle of modal splitting to obtain the next-level intermediate physical state monitoring knowledge and the next-level intermediate safety state monitoring knowledge. The knowledge interaction is performed again based on the confidence level to obtain the next-level intermediate multi-mode liquid cooling state monitoring data. This process is repeated until the number of modal splits in the current cycle reaches the first set number. Finally, the intermediate physical state monitoring knowledge obtained in the current cycle is determined as the first physical state monitoring knowledge, and the intermediate safety state monitoring knowledge obtained in the current cycle is determined as the first safety state monitoring knowledge.

[0157] (II) Example 1 of knowledge interaction

[0158] 1) Knowledge collision to obtain the first liquid cooling state monitoring collision data

[0159] On the basis that the number of modal splittings in the current cycle is not greater than the second set number (the second set number is less than the first set number), according to the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision to obtain the first liquid cooling state monitoring collision data.

[0160] For example, the confidence of the physical state is 0.7, and the confidence of the safety state is 0.6. For the temperature state part of the intermediate physical state monitoring knowledge, if the temperature of a key component (such as the CPU) rises to 70°C, it is weighted according to its confidence of 0.7; for the leakage state part of the intermediate safety state monitoring knowledge, if the resistance value of the leakage detection circuit changes (indicating that there may be a leakage risk), it is weighted according to its confidence of 0.6. Then the two weighted types of knowledge are collided, for example, by combining the weighted temperature value with the leakage-related value (such as the leakage risk quantification value) according to a certain algorithm (such as addition, multiplication, etc., depending on the system settings and the physical meaning of the data), to obtain the first liquid cooling state monitoring collision data.

[0161] 2) Semantic fine-grained adjustment and state attribute recognition to obtain the second liquid cooling state monitoring collision data

[0162] Perform semantic fine-grained adjustments on the original sensor monitoring logs. Semantic fine-grained adjustments may involve semantic understanding and re-labeling of the data in the original logs. For example, for temperature data, the original temperature value was simply recorded. After semantic fine-grained adjustments, more detailed semantic information such as the trend of temperature increase or decrease and the degree of deviation from the normal range may be labeled.

[0163] Then, the original sensor monitoring log after semantic fine-grained adjustment is subjected to state attribute recognition to obtain the second liquid cooling state monitoring collision data. For example, the flow data in the original sensor monitoring log, after semantic fine-grained adjustment, the change trend of the flow is clearly marked (such as from stable to slowly declining). In the process of state attribute recognition, in addition to identifying the flow value, the attributes of this trend change are also identified, and these attribute information are integrated to obtain the second liquid cooling state monitoring collision data.

[0164] 3) Obtaining intermediate multi-mode liquid cooling status monitoring data through feature focusing strategy

[0165] The state attributes of the first liquid cooling state monitoring collision data and the second liquid cooling state monitoring collision data are identified through the feature focusing strategy to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle. The feature focusing strategy may use a certain algorithm to focus on key features. For example, the principal component analysis (PCA) algorithm is used to find the principal component features that have a greater impact on the state of the liquid cooling server in the first liquid cooling state monitoring collision data and the second liquid cooling state monitoring collision data.

[0166] For example, the first liquid cooling state monitoring collision data contains multiple features such as temperature, flow, and pressure, and the second liquid cooling state monitoring collision data contains safety-related features such as leakage and communication status. Through the PCA algorithm, the temperature change trend, key change points of flow (such as the moment of sudden drop), and leakage risk are determined as the main principal component features, and these features are combined to form the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

[0167] (III) Example 2 of knowledge interaction

[0168] 1) Knowledge collision to obtain the third liquid cooling state monitoring collision data

[0169] On the basis that the number of modal splittings in the current cycle is greater than the third set number (the third set number is not greater than the first set number), according to the physical state confidence and the safety state confidence, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision to obtain the third liquid cooling state monitoring collision data.

[0170] For example, in the pressure state part of the intermediate physical state monitoring knowledge, the inlet pressure of the liquid cooling pipeline increases from 200kPa to 250kPa, and is weighted according to its confidence (for example, 0.8); in the coolant impurity state part of the intermediate safety state monitoring knowledge, the impurity content increases from 0.1 to 0.2, and is weighted according to its confidence (for example, 0.7). Then, the weighted pressure state data and coolant impurity state data are collided, such as according to a specific formula (such as taking the logarithm after weighted summation, etc.) to obtain the third liquid cooling state monitoring collision data.

[0171] 2) Integrate and obtain the intermediate multi-mode liquid cooling status monitoring data

[0172] At least one intermediate multi-mode liquid cooling state monitoring data obtained in a cycle before the current cycle is integrated with the third liquid cooling state monitoring collision data to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

[0173] For example, the intermediate multi-mode liquid cooling state monitoring data obtained in the previous cycle includes state data of temperature, flow, pressure, leakage state, communication state, etc. These data are integrated with the third liquid cooling state monitoring collision data obtained currently (such as data related to pressure and impurity state) according to certain rules. This rule can be to merge data of the same type (such as taking the average or weighted average of pressure data of different cycles), and to arrange data of different types in a certain logical order (such as physical state data first and safety state data later), so as to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

[0174] In this way, through multi-cycle modal splitting and knowledge interaction, the information in the liquid cooling server status monitoring data can be mined more deeply than a one-time simple splitting. In the splitting and interaction process of different cycles, the acquisition of physical and safety status monitoring knowledge is continuously optimized. For example, by gradually performing knowledge collision, semantic fine-grained adjustment, and feature focusing operations based on confidence, the first physical status monitoring knowledge and the first safety status monitoring knowledge finally obtained can more accurately and comprehensively reflect the real status of the liquid cooling server.

[0175] In Example 1 of knowledge interaction, by combining data from different sources (knowledge collision and state attribute recognition after semantic adjustment of original logs), the feature focusing strategy can be used to highlight key features and avoid interference from irrelevant information. In Example 2, by integrating data from different periods, effective information in historical data can be fully utilized to enhance the consistency and integrity of data, thereby improving the accuracy of liquid cooling server status assessment, providing a more reliable basis for subsequent liquid cooling control decisions, and further optimizing the control process of the liquid cooling server, improving the operating efficiency and stability of the server, and extending the service life of the server hardware.

[0176] In actual application, artificial intelligence technology can be combined to realize the above-mentioned liquid cooling server control, that is, the liquid cooling server control method can be implemented based on an auxiliary control decision algorithm, and the auxiliary control decision algorithm includes a state attribute recognition branch, a modal splitting branch, a knowledge interaction branch and a control labeling branch. Furthermore, the state attribute identification branch is used to perform state attribute identification on the original sensor monitoring log to obtain the first multi-mode liquid cooling state monitoring data; the modal splitting branch is used to perform modal splitting on the first multi-mode liquid cooling state monitoring data to obtain the first physical state monitoring knowledge and the first safety state monitoring knowledge; the knowledge interaction branch is used to perform knowledge interaction between the first physical state monitoring knowledge and the first safety state monitoring knowledge based on the physical state confidence and the safety state confidence to obtain the second multi-mode liquid cooling state monitoring data, and the physical state confidence and the safety state confidence are branch algorithm variables of the knowledge interaction branch; the control annotation branch is used to perform residual connection processing on the second multi-mode liquid cooling state monitoring data through the first residual embedding module to obtain the semantic set of trend description to be controlled, and combine the original sensor monitoring log and the semantic set of trend description to be controlled to obtain the liquid cooling control decision annotation log, and the first residual embedding module is the branch algorithm variable of the control annotation branch.

[0177] On this basis, the debugging steps of the auxiliary control decision algorithm may include the following steps 1 to 5.

[0178] Step 1: Perform state attribute identification on the sensor monitoring log sample through the state attribute identification branch to obtain a first multi-mode liquid cooling state monitoring data sample.

[0179] Step 2: Perform modal splitting on the first multi-mode liquid cooling state monitoring data sample through the modal splitting branch to obtain a first physical state monitoring knowledge sample and a first safety state monitoring knowledge sample.

[0180] Step three: Through the knowledge interaction branch, based on the physical state confidence and the safety state confidence, the first physical state monitoring knowledge sample and the first safety state monitoring knowledge sample are knowledge-interacted to obtain a second multi-mode liquid cooling state monitoring data sample.

[0181] Step 4: Through the control annotation branch, the second multi-mode liquid cooling state monitoring data sample is subjected to residual connection processing through the first residual embedding module to obtain a sample of the semantic set of trend descriptions to be controlled, and the sensor monitoring log sample is integrated with the sample of the semantic set of trend descriptions to be controlled to obtain a control training log.

[0182] Step 5: Debug the auxiliary control decision algorithm according to the comparison result between the control training log and the control instruction log sample.

[0183] In actual application, the liquid cooling server control technology solution based on the auxiliary control decision algorithm is described in detail as follows.

[0184] 1. Overall architecture of auxiliary control decision algorithm

[0185] An auxiliary control decision algorithm is introduced in the liquid cooling server control. The algorithm contains four key branches: state attribute identification branch, modal splitting branch, knowledge interaction branch and control labeling branch.

[0186] 1) State attribute identification branch

[0187] The main function of this branch is to identify the state attributes of the original sensor monitoring log, so as to obtain the first multi-mode liquid cooling state monitoring data. The original sensor monitoring log contains information about the operating status of the liquid cooling server collected by many sensors, such as the temperature values ​​recorded by the temperature sensor at different times, the coolant flow value recorded by the flow sensor, the pressure reading of the pressure sensor, and sensor data related to leak detection, coolant impurity detection and communication status.

[0188] When performing state attribute identification, operations such as data cleaning and feature extraction may be involved. Take temperature data as an example. For example, in the original temperature data, 90% of the data are between 10-70°C, but there are individual data points reaching 120°C. This is likely caused by sensor failure or interference. By setting a reasonable range (based on the statistical range of the temperature during normal operation of the liquid-cooled server), these abnormal values ​​can be excluded. For flow data, in addition to the directly obtained flow value, it is also possible to extract features such as the rate of change of the flow. For example, over a period of time, the flow rate dropped from 15L / min to 12L / min, and the flow change rate was calculated to be -20% / unit time. By performing such comprehensive processing on the data of various modes in the original sensor monitoring log, it is converted into the first multi-mode liquid cooling state monitoring data that can accurately reflect the multi-mode liquid cooling state of the liquid-cooled server.

[0189] 2) Modal Split Branch

[0190] The modal splitting branch is responsible for performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge.

[0191] (1) Physical condition monitoring knowledge

[0192] Temperature status: Separate the temperature-related data from the first multi-mode liquid cooling status monitoring data. For example, in a liquid cooling server with multiple CPU cores, the temperature sensor of each core provides the temperature readings of the core at different times. For example, the temperature of the CPU1 core rises from 35°C to 55°C within 15 minutes. This temperature change trend is part of the temperature status.

[0193] Flow status: The splitting of flow status data involves the collation of information related to coolant flow, including specific flow values, flow rate of change, etc. For example, the flow of the coolant pump is 20L / min during stable operation, and when the server load increases, the flow is adjusted to 25L / min. The flow values ​​and their changes in these different situations constitute the knowledge of flow status monitoring.

[0194] Pressure state: The decomposition of pressure state focuses on the pressure data in the liquid cooling pipeline. The pressure sensor data at the inlet, outlet and key branch points of the liquid cooling pipeline are extracted. For example, the normal pressure at the pipeline inlet is 180kPa. When a valve is partially blocked, the pressure may rise to 230kPa. These pressure values ​​and their changes reflect the knowledge of pressure state monitoring.

[0195] (2) Safety status monitoring knowledge

[0196] Leakage status: Sensor data related to leak detection is separated. For example, when using electrical detection methods, the conduction state of the leak detection circuit (conduction indicates a possible leak) and related resistance value changes. If the resistance value of the circuit suddenly changes from a high resistance state to a low resistance state, it may indicate that the coolant has leaked into the area where the detection circuit is located.

[0197] Coolant impurity status: For the coolant impurity status, it may involve special impurity detection sensor data. For example, optical detection methods are used to determine the impurity content based on the change in the propagation characteristics of light in the coolant. When the light scattering coefficient increases from the normal 0.08 to 0.15, it means that the impurity content in the coolant has increased.

[0198] Communication status: Communication status monitoring knowledge mainly comes from the monitoring data of the communication link between the liquid cooling control system and various sensors and devices. For example, the communication status is judged by monitoring the loss rate of communication data packets. If the communication data packet loss rate rises from the normal 0.1% to 3%, it means that there is a problem with the communication status.

[0199] 3) Knowledge interaction branch

[0200] The knowledge interaction branch performs knowledge interaction on the first physical state monitoring knowledge and the first safety state monitoring knowledge according to the physical state confidence and the safety state confidence, and obtains the second multi-mode liquid cooling state monitoring data. The physical state confidence and the safety state confidence of the present invention are branch algorithm variables of the knowledge interaction branch.

[0201] (1) Obtaining confidence

[0202] The physical state confidence and safety state confidence are obtained by performing AI quantitative mining and analysis on sensor monitoring log samples and corresponding control instruction log samples.

[0203] Data preparation: Collect a large number of sensor monitoring log samples and corresponding control instruction log samples. For example, collect 800 sets of sensor monitoring logs and corresponding control instruction logs of liquid cooling servers in different operating scenarios. These scenarios include the operation of servers under different loads (low load, medium load, high load) and different ambient temperatures (5℃-35℃).

[0204] AI quantitative mining analysis: Use neural network algorithms (such as deep neural networks) for analysis. Taking the temperature state as an example, the temperature sensor reading, temperature change rate, etc. are used as input features, and the corresponding control instructions (such as whether to adjust the coolant flow, whether to change the fan speed, etc.) are used as output labels. Through training on a large amount of data, the neural network can learn the relationship between the temperature state and the control instructions, thereby obtaining the confidence of the temperature state. For example, when the temperature rises to 75°C, according to the trained model, the system predicts that the confidence that the coolant flow needs to be adjusted is 0.75.

[0205] (2) Knowledge interaction process

[0206] Confidence-based weighting: The first physical state monitoring knowledge and the first safety state monitoring knowledge are weighted according to the physical state confidence and the safety state confidence. For example, the physical state confidence is 0.8 and the safety state confidence is 0.7. For a data point in the temperature state (such as the CPU temperature rises to 65°C), when interacting with the leakage state in the safety state, the temperature state data is weighted according to its confidence of 0.8, and the leakage state data is weighted according to its confidence of 0.7.

[0207] Knowledge fusion: The weighted physical state monitoring knowledge and safety state monitoring knowledge are integrated to obtain the second multi-mode liquid cooling state monitoring data. For example, the weighted temperature state, flow state, pressure state, weighted leakage state, coolant impurity state, communication state and other data are combined according to certain rules to form a new comprehensive data structure. Specifically, for example, the weighted data of the temperature state is [65℃*0.8], the weighted data of the flow state is [22L / min*0.8], the weighted data of the pressure state is [200kPa*0.8], the weighted data of the leakage state is [0.3*0.7] (here 0.3 represents the quantitative value of the leakage risk), the weighted data of the coolant impurity state is [0.12*0.7] (here 0.12 represents the quantitative value of the impurity content), and the weighted data of the communication state is [2%*0.7] (here 2% represents the communication data packet loss rate). These data are combined into a new array in a certain order, and this array is the second multi-mode liquid cooling state monitoring data.

[0208] 4) Control annotation branches

[0209] The control annotation branch first performs residual connection processing on the second multi-mode liquid cooling state monitoring data through the first residual embedding module to obtain a semantic set of trend descriptions to be controlled, and then combines the original sensor monitoring log and the semantic set of trend descriptions to be controlled to obtain a liquid cooling control decision annotation log. The first residual embedding module of the present invention is a branch algorithm variable of the control annotation branch.

[0210] (1) Construction of the first residual embedding module and residual connection processing

[0211] Construction of the first residual embedding module: The first residual embedding module is obtained by feature mining based on sensor monitoring log samples and control indication log samples. For example, algorithms such as principal component analysis (PCA) are used to perform feature mining on the data in these logs. For temperature status, the principal component features of temperature changes may be mined; for flow status, the correlation features between flow and server load may be mined, etc. The first residual embedding module is constructed based on the mined features. This module contains multiple sub-modules, each of which is responsible for processing the residual connection of a modal state data. For example, there are sub-modules that specifically process residual connections of temperature status, sub-modules for flow status, etc.

[0212] Residual connection processing: For each modal data in the second multi-mode liquid cooling state monitoring data, calculate its residual with the corresponding data in the original sensor monitoring log. Take the flow state as an example. For example, if the original flow data is 18L / min, and the flow data after processing in the previous step is 16L / min, the residual is -2L / min. The residuals of each mode are connected through the first residual embedding module to obtain a semantic set of trend descriptions to be controlled. This semantic set represents the control logic comparison features between the original sensor monitoring log and the corresponding control indication log. For example, when the temperature residual is positive (indicating an increase in temperature) and the flow residual is negative (indicating a decrease in flow), the semantic set will contain information about the relationship between the temperature and flow change trends and the control logic, which may imply that the coolant flow needs to be increased to control the temperature rise.

[0213] (2) Determine the liquid cooling control decision annotation log

[0214] Combine the original sensor monitoring log and the semantic set of the trend description to be controlled for comprehensive analysis. For example, the original sensor monitoring log shows that the CPU temperature continues to rise, and the semantic set of the trend description to be controlled indicates that the temperature rise is associated with a decrease in flow and that the flow needs to be increased to control the temperature. Based on the results of the comprehensive analysis, determine the liquid cooling control decision annotation log for the liquid-cooled server to be controlled. Specific control processing instructions are annotated in this annotation log. For example, the annotation log may record control processing instructions such as "Increase the speed of the coolant pump and increase the flow from 12L / min to 14L / min to control the CPU temperature below 70°C", or "Check the leakage detection circuit, because the communication status shows that the increase in the packet loss rate may be related to leakage".

[0215] In some extended innovative embodiments, the auxiliary control decision algorithm includes u modal splitting branches and u knowledge interaction branches, where u is a positive integer; the modal splitting of the first multi-mode liquid cooling state monitoring data sample is performed through the modal splitting branches to obtain a first physical state monitoring knowledge sample and a first safety state monitoring knowledge sample, including: based on the first modal splitting branch, the first multi-mode liquid cooling state monitoring data is modally split to obtain an intermediate physical state monitoring knowledge sample and an intermediate safety state monitoring knowledge sample; based on the first knowledge interaction branch, according to the physical state confidence and the safety state confidence in the first knowledge interaction branch, the intermediate physical state monitoring knowledge sample and the intermediate safety state monitoring knowledge sample are knowledge interacted to obtain an intermediate multi-mode liquid cooling state monitoring data sample; according to the modal splitting branch of the latter level, the intermediate multi-mode liquid cooling state monitoring The measured data sample is subjected to modal splitting of the next cycle to obtain the intermediate physical state monitoring knowledge sample and the intermediate safety state monitoring knowledge sample of the next level; according to the knowledge interaction branch of the next level, according to the physical state confidence and the safety state confidence in the knowledge interaction branch of the next level, the intermediate physical state monitoring knowledge sample and the intermediate safety state monitoring knowledge sample of the next level are subjected to knowledge interaction to obtain the intermediate multi-mode liquid cooling state monitoring data sample of the next level, until the last modal splitting branch generates the intermediate physical state monitoring knowledge sample and the intermediate safety state monitoring knowledge sample, the intermediate physical state monitoring knowledge sample generated by the last modal splitting branch is determined as the first physical state monitoring knowledge sample, and the intermediate safety state monitoring knowledge sample generated by the last modal splitting branch is determined as the first safety state monitoring knowledge sample.

[0216] Based on this, the first physical state monitoring knowledge sample and the first safety state monitoring knowledge sample are subjected to knowledge interaction through the knowledge interaction branch according to the physical state confidence and the safety state confidence, so as to obtain the second multi-mode liquid cooling state monitoring data sample, including: through the last knowledge interaction branch, according to the physical state confidence and the safety state confidence in the last knowledge interaction branch, the first physical state monitoring knowledge sample and the first safety state monitoring knowledge sample are subjected to knowledge interaction, so as to obtain the second multi-mode liquid cooling state monitoring data sample.

[0217] In this extended innovative embodiment, the auxiliary control decision algorithm includes u modal splitting branches and u knowledge interaction branches (u is a positive integer). This multi-branch structure provides a more refined and hierarchical processing method for liquid cooling server control.

[0218] 1) First processing based on multiple modal split branches and knowledge interaction branches

[0219] (1) Modal splitting based on the first modal splitting branch

[0220] Based on the first modal splitting branch, the first multi-mode liquid cooling state monitoring data is modally split to obtain intermediate physical state monitoring knowledge samples and intermediate safety state monitoring knowledge samples.

[0221] Taking the physical quantities such as temperature, flow and pressure of liquid cooling servers and safety-related states such as leakage, impurities and communication as examples, in the first modal splitting branch, the splitting of the physical state monitoring knowledge sample is as follows:

[0222] Temperature status: If the liquid-cooled server has multiple key components, such as CPU, GPU, etc., separate the temperature data corresponding to each component from the first multi-mode liquid cooling status monitoring data. For example, for a server with 4 CPU cores, the temperature sensor of each core records the temperature value at different times, forming part of the temperature status monitoring data. For example, the temperature of the CPU1 core shows a specific value change over a period of time, such as rising from 30°C to 35°C.

[0223] Flow status: Extract information related to coolant flow from the overall monitoring data, including the actual value of the flow and its rate of change. For example, the coolant inlet flow is initially 15L / min, and after a period of time it changes to 13L / min. The calculated flow rate change rate is -13.33% (according to the formula: (13-15) / 15*100%).

[0224] Pressure status: Focus on the pressure data in the liquid cooling pipeline, and obtain data from pressure sensors at different locations (such as inlet, outlet, and key branch points). For example, the normal pressure at the pipeline inlet is 200kPa, and at a certain moment it becomes 220kPa due to some factors (such as local blockage).

[0225] For the splitting of intermediate safety status monitoring knowledge samples:

[0226] Leakage status: Information is obtained from sensor data related to leak detection. If an electrical detection method is used, for example, the resistance value of the leakage detection circuit is high under normal circumstances (such as 1000Ω), and when a possible leak occurs, the resistance value changes to a low resistance value (such as 100Ω). This change in resistance value is part of the leakage status monitoring data.

[0227] Coolant impurity status: For example, optical detection methods are used to detect coolant impurities, and the impurity content is determined by measuring the propagation characteristics of light in the coolant. Normally, the scattering coefficient of light is 0.1. When impurities increase, the scattering coefficient becomes 0.15. This change reflects the coolant impurity status.

[0228] Communication status: Obtain communication status information from the monitoring data of the communication link between the liquid cooling control system and each sensor and device. For example, by counting the loss rate of communication data packets, it is normally 0.1%. If it rises to 0.5% at a certain moment, it means that the communication status has fluctuated.

[0229] (2) Knowledge interaction based on the first knowledge interaction branch

[0230] Based on the first knowledge interaction branch, according to the physical state confidence and safety state confidence in this branch, the intermediate physical state monitoring knowledge sample and the intermediate safety state monitoring knowledge sample are interacted to obtain the intermediate multi-mode liquid cooling state monitoring data sample.

[0231] The physical state confidence and safety state confidence of the present invention are obtained by performing AI quantitative mining analysis on a large number of sensor monitoring log samples and corresponding control instruction log samples. For example, a neural network algorithm (such as a deep neural network) is used for analysis:

[0232] For the temperature state in the physical state, the temperature sensor reading, temperature change rate, etc. are used as input features, and the corresponding control instructions (such as whether to adjust the coolant flow, whether to change the fan speed, etc.) are used as output labels. For example, after a large amount of data training, when the temperature rises to a certain value (such as 60°C), the confidence level of the model prediction that the coolant flow needs to be adjusted is 0.7.

[0233] For the leakage state in the safety state, the relevant parameters of the leakage detection circuit (such as the resistance value change rate) are used as input features, and the corresponding control instructions (such as whether to issue a leakage alarm, whether to stop the operation of the liquid cooling system, etc.) are used as output labels. Through training, it is found that when the resistance value changes to a certain extent (such as from 1000Ω to 200Ω), the confidence level of determining the existence of leakage risk is 0.8.

[0234] In the process of knowledge interaction, the confidence-based weighting and fusion operations are as follows.

[0235] Confidence-based weighting: For example, the confidence of physical state is 0.7, and the confidence of security state is 0.8. For a data point in temperature state (such as CPU temperature is 60°C), when interacting with the leakage state in security state (such as the resistance value changes from 1000Ω to 200Ω), the temperature state data is weighted according to its confidence of 0.7, that is, 60°C*0.7=42°C (this is to illustrate the weighting operation, and the weighted numerical representation method may be different in actual applications); the leakage state data is weighted according to its confidence of 0.8, for example, the leakage risk quantification value is 0.5 (indicating a certain possibility of leakage), and the weighted value is 0.5*0.8=0.4.

[0236] Knowledge fusion: Fusion of weighted physical state monitoring knowledge and safety state monitoring knowledge. For example, weighted temperature state, flow state, pressure state, weighted leakage state, coolant impurity state, communication state and other data are combined according to certain rules. For example, the weighted data of flow state is [13L / min*0.7], the weighted data of pressure state is [220kPa*0.7], the weighted data of coolant impurity state is [0.15*0.8] (for example, the quantitative value of impurity content is 0.15), and the weighted data of communication state is [0.5%*0.8]. These weighted data are combined in a specific order to form an intermediate multi-mode liquid cooling state monitoring data sample.

[0237] 2) Modal splitting and knowledge interaction in subsequent cycles

[0238] (1) Modal splitting based on the next-level modal splitting branch

[0239] According to the next-level modal splitting branch, the intermediate multi-mode liquid cooling state monitoring data sample is subjected to the next-cycle modal splitting to obtain the next-level intermediate physical state monitoring knowledge sample and the next-level intermediate safety state monitoring knowledge sample.

[0240] This process is similar to the first modal splitting, but because it is based on the intermediate multi-mode liquid cooling state monitoring data sample of the previous cycle, the data has been preliminarily processed and fused. For example, for the temperature state in the next-level physical state monitoring knowledge sample, since the interaction in the previous cycle may have considered the impact of some control factors on the temperature, this cycle may pay more attention to the adjustment of temperature under new environmental factors or changes in the internal state of the server. For example, the server load changed in this cycle, causing the CPU temperature to rise from 35°C to 38°C under the new conditions. This temperature change is part of the next-level temperature state monitoring knowledge sample.

[0241] For the flow status, the flow adjustment decision in the previous cycle may have affected the current flow situation. For example, the coolant flow was 14L / min after adjustment in the previous cycle. Due to other factors (such as slight changes in the liquid cooling pipeline), the flow rate changed to 13.5L / min in this cycle. This flow change and the current flow value constitute the next-level flow status monitoring knowledge sample.

[0242] For the next-level intermediate safety status monitoring knowledge sample, taking the leakage status as an example, if the leakage detection in the previous cycle found a potential risk, this cycle may use a more accurate detection method or increase the detection frequency. For example, if the resistance value changes from 1000Ω to 200Ω in the previous cycle, it means there is a leakage risk. This cycle may further detect more detailed information related to the leakage location, and this information will become part of the next-level leakage status monitoring knowledge sample.

[0243] (2) Knowledge interaction based on the next level of knowledge interaction branch

[0244] According to the next-level knowledge interaction branch, based on the physical state confidence and safety state confidence in this branch, the next-level intermediate physical state monitoring knowledge sample and the next-level intermediate safety state monitoring knowledge sample are interacted to obtain the next-level intermediate multi-mode liquid cooling state monitoring data sample.

[0245] Similarly, the confidence degree of the present invention is obtained in a similar manner as before, by analyzing new sensor monitoring log samples and control indication log samples. For example, in a new server operation scenario, after retraining the neural network model, the confidence degree of adjusting the coolant flow rate when the temperature state rises to a certain value (such as 65°C) may become 0.8.

[0246] During the knowledge interaction process, confidence-based weighting and fusion operations are performed again. For example, the confidence of the next-level physical state is 0.8, and the confidence of the safety state is 0.7. For the temperature state of 65°C, the weighted value is 65°C*0.8=52°C; for the leakage state in the safety state (for example, the leakage risk quantification value is 0.6 after more precise detection), the weighted value is 0.6*0.7=0.42. Then the weighted next-level physical state monitoring knowledge (including temperature, flow, pressure, etc.) and the next-level safety state monitoring knowledge (including leakage, impurities, communication, etc.) are fused according to the rules to obtain the next-level intermediate multi-mode liquid cooling state monitoring data sample.

[0247] This process will be repeated until the last modal split branch generates an intermediate physical class state monitoring knowledge sample and an intermediate safety class state monitoring knowledge sample.

[0248] 3) Finalize the first physical and safety condition monitoring knowledge samples and the second multi-mode liquid cooling condition monitoring data samples

[0249] Finally, the intermediate physical state monitoring knowledge sample generated by the last modal splitting branch is determined as the first physical state monitoring knowledge sample, and the intermediate safety state monitoring knowledge sample generated by the last modal splitting branch is determined as the first safety state monitoring knowledge sample.

[0250] Then, through the last knowledge interaction branch, according to the physical state confidence and safety state confidence in this branch, the first physical state monitoring knowledge sample and the first safety state monitoring knowledge sample are interacted to obtain the second multi-mode liquid cooling state monitoring data sample. This process is similar to the previous knowledge interaction process, except that the confidence of the last branch and the finalized first physical and safety state monitoring knowledge samples are used.

[0251] It can be seen that the setting of multiple modal splitting branches and knowledge interaction branches makes the processing of liquid cooling server status monitoring data more meticulous and in-depth. Through multi-cycle modal splitting and knowledge interaction, the acquisition of physical and safety status monitoring knowledge can be gradually optimized. Each modal splitting can analyze the data from different angles and mine more comprehensive status information, while knowledge interaction integrates this information according to the confidence of different cycles, improving the accuracy and reliability of the data. For example, the physical states such as temperature, flow, pressure and the security state information such as leakage, impurities and communication in different cycles are continuously interacted and integrated, so that the first physical state monitoring knowledge sample and the first security state monitoring knowledge sample obtained can more accurately reflect the real state of the server. This accuracy is further transmitted to the generation process of the second multi-mode liquid cooling status monitoring data sample, providing a more reliable basis for subsequent liquid cooling control decisions, helping to optimize the control process of liquid cooling servers, improve the operating efficiency, stability and safety of servers, and reduce the risk of failures caused by inaccurate status monitoring.

[0252] In some further extended innovative embodiments, the auxiliary control decision algorithm includes v control annotation branches, v is a positive integer not greater than u; through the control annotation branches, the second multi-mode liquid cooling state monitoring data sample is residually connected through the first residual embedding module to obtain a sample of the semantic set of trend description to be controlled, and the sensor monitoring log sample is integrated with the sample of the semantic set of trend description to be controlled to obtain a control training log, including: based on the wth control annotation branch, the intermediate multi-mode liquid cooling state monitoring data sample generated by the u-v+wth knowledge interaction branch is residually connected through the first residual embedding module to obtain the wth sample of the semantic set of trend description to be controlled, and the sensor monitoring log sample is integrated with the wth sample of the semantic set of trend description to be controlled to obtain the wth control training log; wherein w is a positive integer not greater than v.

[0253] Based on this embodiment, the auxiliary control decision algorithm includes v control annotation branches (v is a positive integer not greater than u). This multi-control annotation branch structure further enriches the system of the liquid cooling server control decision algorithm.

[0254] Furthermore, the processing process based on multiple control annotation branches is as follows.

[0255] 1) Operation of the w-th control annotation branch

[0256] (1) Residual connection processing and generation of examples of semantic sets describing trends to be controlled

[0257] Based on the w-th control annotation branch, the intermediate multi-mode liquid cooling state monitoring data samples generated by the u-v+w-th knowledge interaction branch are residually connected through the first residual embedding module to obtain the w-th semantic set sample of the trend description to be controlled.

[0258] The first residual embedding module is obtained by feature mining based on sensor monitoring log samples and control indication log samples. For example, the principal component analysis (PCA) algorithm is used to mine the data in these logs. For temperature status, the main component characteristics of temperature change may be mined, such as the slope of temperature change or the frequency of temperature fluctuation. For flow status, the correlation characteristics between flow and other factors such as server load and temperature are mined. For example, there may be a linear relationship between flow and server load (flow = k*load + b, where k and b are coefficients obtained through data mining).

[0259] When performing residual connection processing, take the flow state as an example. For example, the flow value in the intermediate multi-mode liquid cooling state monitoring data sample generated by the u-v+wth knowledge interaction branch is 15L / min, while the flow value in the original sensor monitoring log sample is 18L / min. According to the residual calculation formula: residual = processed value - original value, the residual of the flow state is 15-18 = -3L / min. Similar residual calculations are performed on the data of each mode (such as temperature, pressure, leakage state, coolant impurity state, communication state, etc.), and then these residuals are connected through the first residual embedding module to obtain the wth sample of the semantic set of trend description to be controlled.

[0260] For example, in the temperature state, if the original temperature data shows an upward trend, and the temperature rise after the previous step is reduced, then the temperature residual may be a negative number, indicating that the temperature rise trend has been controlled to a certain extent. This temperature residual information, together with the residual information of other modes such as flow residual, constitutes an example of the semantic set of trend description to be controlled, which represents the control logic comparison characteristics between the original sensor monitoring log and the corresponding control indication log.

[0261] (2) Integration to obtain control training log

[0262] The sensor monitoring log samples are integrated with the w-th sample of the trend description semantic set to be controlled to obtain the w-th control training log.

[0263] The integration process may involve combining various data in the sensor monitoring log sample (such as original temperature, flow, pressure, leakage detection results, coolant impurity detection results, communication status data, etc.) with the control logic comparison features in the trend description semantic set sample to be controlled (such as the residual relationship between temperature, flow, etc. and the control trend implied by these relationships) according to a certain structure. For example, the data in the sensor monitoring log sample can be used as a vector, and the data in the trend description semantic set sample to be controlled can be used as another vector, and then these two vectors can be spliced ​​or combined into a new vector in a certain weighted manner. This new vector constitutes the wth control training log.

[0264] 2. Overall operation of multiple control annotation branches

[0265] Since w is a positive integer not greater than v, this means that the entire process will perform the above operations on v control annotation branches. Each control annotation branch will process the intermediate multi-mode liquid cooling state monitoring data samples generated by the corresponding knowledge interaction branch to obtain their own trend description semantic set samples to be controlled, and integrate them with the sensor monitoring log samples to obtain the corresponding control training log.

[0266] For example, when v=3, for w=1, the first control annotation branch will be used to operate on the intermediate multi-mode liquid cooling state monitoring data samples generated by the u-3+1=u-2 knowledge interaction branch; for w=2, the second control annotation branch will be used to operate on the intermediate multi-mode liquid cooling state monitoring data samples generated by the u-3+2=u-1 knowledge interaction branch; for w=3, the third control annotation branch will be used to operate on the intermediate multi-mode liquid cooling state monitoring data samples generated by the u-3+3=u knowledge interaction branch. Each branch independently completes the residual connection processing, the generation of the semantic set samples of the trend description to be controlled, and the integration with the sensor monitoring log samples, thereby obtaining three different control training logs.

[0267] With this design, the setting of multiple control annotation branches enables the control decision of the liquid cooling server state to be evaluated and optimized from multiple perspectives. By processing the intermediate multi-mode liquid cooling state monitoring data samples generated by different knowledge interaction branches through different control annotation branches, more diverse control logic comparison features can be mined. For example, in different branches, due to the different sources of processed data (corresponding to different knowledge interaction branches), different temperature-flow relationships, pressure-leakage relationships, etc. may be found. These different relationships can more comprehensively reflect the control requirements of liquid cooling servers under various states. Secondly, each control training log integrates the original sensor monitoring log samples and the samples of the semantic set of trend descriptions to be controlled, which makes the training log contain both the original state information and the processed control logic information, providing a richer and more accurate data basis for the subsequent debugging of the auxiliary control decision algorithm based on the comparison results between the control training log and the control indication log samples. This multi-branch structure helps to improve the adaptability of the algorithm to different server operating states, optimize the control process of the liquid cooling server, and thus improve the operating efficiency, stability and security of the server.

[0268] Further, Figure 2 FIG. 2 is a schematic diagram of a liquid cooling server control system 200 provided by the present invention. Figure 2 The liquid cooling server control system 200 shown includes a processor 210, and the processor 210 can call and run a computer program from a memory to implement the method of the present invention.

[0269] Alternatively, if Figure 2 As shown, the liquid cooling server control system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method of the present invention.

[0270] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0271] Alternatively, if Figure 2 As shown, the liquid cooling server control system 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, may send information or data to other devices, or receive information or data sent by other devices.

[0272] Optionally, the liquid cooling server control system 200 can implement the corresponding processes corresponding to the storage engine or the components (such as processing modules) in the storage engine or the device with the storage engine deployed in each method of the present invention, which will not be described here for the sake of brevity. It should be understood that the processor of the present invention may be an integrated circuit chip with signal processing capabilities.

[0273] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0274] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0275] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

Claims

1. A liquid cooling server control method, characterized in that: The method is implemented by a liquid cooling server control system, and the method comprises: Performing state attribute recognition on the original sensor monitoring log of the liquid cooling server to be controlled to obtain first multi-mode liquid cooling state monitoring data; Performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge; wherein the physical state includes temperature state, flow state and pressure state, and the safety state includes leakage state, coolant impurity state and communication state; According to the physical state confidence and the safety state confidence, the first physical state monitoring knowledge and the first safety state monitoring knowledge are interacted to obtain second multi-mode liquid cooling state monitoring data, wherein the physical state confidence and the safety state confidence are obtained by performing AI quantitative mining analysis based on sensor monitoring log samples and corresponding control indication log samples; Performing residual connection processing on the second multi-mode liquid cooling state monitoring data through a first residual embedding module to obtain a semantic set of trend descriptions to be controlled, wherein the semantic set of trend descriptions to be controlled represents a control logic comparison feature between the original sensor monitoring log and the corresponding control instruction log, and the first residual embedding module is obtained by performing feature mining based on the sensor monitoring log sample and the control instruction log sample; In combination with the original sensor monitoring log and the semantic set of trend descriptions to be controlled, a liquid cooling control decision annotation log of the liquid cooling server to be controlled is determined, wherein the liquid cooling control decision annotation log annotates control processing instructions for the liquid cooling server to be controlled.

2. The method according to claim 1, characterized in that The performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge includes: Decomposing the first multi-mode liquid cooling state monitoring data into second physical state monitoring knowledge and second safety state monitoring knowledge according to the mode category, performing knowledge interaction on the second physical state monitoring knowledge and the second safety state monitoring knowledge, and obtaining third multi-mode liquid cooling state monitoring data; The third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on time series features to obtain the first physical state monitoring knowledge, and the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on decision trees to obtain the first safety state monitoring knowledge.

3. The method according to claim 2, characterized in that The method further comprises: Performing residual connection processing on the third multi-mode liquid cooling state monitoring data through a second residual embedding module to obtain a discriminant knowledge mining module, wherein the second residual embedding module is obtained by performing feature mining based on past sensor monitoring logs and the control instruction logs; Determine a temporal knowledge mining module by using a long short-term memory module and the discriminative knowledge mining module, wherein the number of attention indicators of the long short-term memory module is consistent with the number of attention indicators of the discriminative knowledge mining module, and the attention weight of one of the attention indicators of the long short-term memory module is 1, and the attention weights of the remaining attention indicators are 0; The performing of knowledge mining based on time series features on the third multi-mode liquid cooling state monitoring data to obtain the first physical state monitoring knowledge, and performing of knowledge mining based on a decision tree on the third multi-mode liquid cooling state monitoring data to obtain the first safety state monitoring knowledge, include: The third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on time series features through the time series knowledge mining module to obtain the first physical state monitoring knowledge, and the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on decision trees through the discrimination knowledge mining module to obtain the first safety state monitoring knowledge.

4. The method according to claim 3, characterized in that: The number of the discriminant knowledge mining modules is a, and the number of attention indicators of each discriminant knowledge mining module is b; the number of the time series knowledge mining modules is a, and the number of attention indicators of each time series knowledge mining module is b; the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on time series features by the time series knowledge mining module to obtain the first physical state monitoring knowledge, and the third multi-mode liquid cooling state monitoring data is subjected to knowledge mining based on a decision tree by the discriminant knowledge mining module to obtain the first safety state monitoring knowledge, including: Decomposing the third multi-mode liquid cooling state monitoring data into a third local liquid cooling state monitoring knowledge based on the attention index, wherein the number of feature units corresponding to the discriminant knowledge mining module and the number of feature units corresponding to the temporal knowledge mining module are consistent with the number of feature units corresponding to the third local liquid cooling state monitoring knowledge; By using a said time series knowledge mining modules, respectively performing knowledge mining based on time series characteristics on a said third local liquid cooling state monitoring knowledge, to obtain a first local physical state monitoring vectors, and the a said first local physical state monitoring vectors form the first physical state monitoring knowledge; Through a of the discriminant knowledge mining modules, decision tree-based knowledge mining is performed on a of the third local liquid cooling state monitoring knowledge to obtain a first-stage safety state monitoring vectors, and a of the first-stage safety state monitoring vectors form the first safety class state monitoring knowledge.

5. The method according to claim 1, characterized in that The performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain first physical state monitoring knowledge and first safety state monitoring knowledge includes: Performing modal splitting on the first multi-mode liquid cooling state monitoring data to obtain intermediate physical state monitoring knowledge and intermediate safety state monitoring knowledge, and performing knowledge interaction on the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge according to the physical state confidence and the safety state confidence to obtain intermediate multi-mode liquid cooling state monitoring data; Perform modal splitting of the next cycle on the intermediate multi-mode liquid cooling state monitoring data to obtain the intermediate physical state monitoring knowledge of the next level and the intermediate safety state monitoring knowledge of the next level, and perform knowledge interaction between the intermediate physical state monitoring knowledge of the next level and the intermediate safety state monitoring knowledge of the next level according to the physical state confidence and the safety state confidence to obtain the intermediate multi-mode liquid cooling state monitoring data of the next level, until the number of modal splittings of the current cycle reaches the first set number, determine the intermediate physical state monitoring knowledge obtained in the current cycle as the first physical state monitoring knowledge, and determine the intermediate safety state monitoring knowledge obtained in the current cycle as the first safety state monitoring knowledge.

6. The method according to claim 5, characterized in that The step of performing knowledge interaction between the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge based on the physical state confidence and the safety state confidence to obtain intermediate multi-mode liquid cooling state monitoring data includes: On the basis that the number of modal splits in the current cycle is not greater than the second set number, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision according to the physical state confidence and the safety state confidence to obtain first liquid cooling state monitoring collision data, and the second set number is less than the first set number; Performing semantic fine-grained adjustment on the original sensor monitoring log, and performing state attribute recognition on the original sensor monitoring log after the semantic fine-grained adjustment to obtain second liquid cooling state monitoring collision data; The state attributes of the first liquid cooling state monitoring collision data and the second liquid cooling state monitoring collision data are identified through a feature focusing strategy to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

7. The method according to claim 5, characterized in that The step of performing knowledge interaction between the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge based on the physical state confidence and the safety state confidence to obtain intermediate multi-mode liquid cooling state monitoring data includes: On the basis that the number of modal splits in the current cycle is greater than a third set number, the intermediate physical state monitoring knowledge and the intermediate safety state monitoring knowledge obtained in the current cycle are subjected to knowledge collision according to the physical state confidence and the safety state confidence to obtain third liquid cooling state monitoring collision data, and the third set number is not greater than the first set number; At least one intermediate multi-mode liquid cooling state monitoring data obtained in a cycle before the current cycle is integrated with the third liquid cooling state monitoring collision data to obtain the intermediate multi-mode liquid cooling state monitoring data of the current cycle.

8. The method according to any one of claims 1 to 7, characterized in that: The liquid cooling server control method is implemented based on an auxiliary control decision algorithm, and the auxiliary control decision algorithm includes a state attribute identification branch, a modal splitting branch, a knowledge interaction branch, and a control annotation branch; The state attribute identification branch is used to perform state attribute identification on the original sensor monitoring log to obtain the first multi-mode liquid cooling state monitoring data; The modal splitting branch is used to perform modal splitting on the first multi-mode liquid cooling state monitoring data to obtain the first physical state monitoring knowledge and the first safety state monitoring knowledge; The knowledge interaction branch is used to perform knowledge interaction on the first physical state monitoring knowledge and the first safety state monitoring knowledge according to the physical state confidence and the safety state confidence, so as to obtain the second multi-mode liquid cooling state monitoring data, wherein the physical state confidence and the safety state confidence are branch algorithm variables of the knowledge interaction branch; The control annotation branch is used to perform residual connection processing on the second multi-mode liquid cooling state monitoring data through the first residual embedding module to obtain the semantic set of descriptions of the trend to be controlled, and combine the original sensor monitoring log and the semantic set of descriptions of the trend to be controlled to obtain the liquid cooling control decision annotation log. The first residual embedding module is a branch algorithm variable of the control annotation branch.

9. The method according to claim 8, characterized in that The debugging process of the auxiliary control decision algorithm includes: Through the state attribute identification branch, the state attribute of the sensor monitoring log sample is identified to obtain a first multi-mode liquid cooling state monitoring data sample; Through the modal splitting branch, the first multi-mode liquid cooling state monitoring data sample is modally split to obtain a first physical state monitoring knowledge sample and a first safety state monitoring knowledge sample; Through the knowledge interaction branch, according to the physical state confidence and the safety state confidence, the first physical state monitoring knowledge sample and the first safety state monitoring knowledge sample are knowledge-interacted to obtain a second multi-mode liquid cooling state monitoring data sample; Through the control annotation branch, the second multi-mode liquid cooling state monitoring data sample is subjected to residual connection processing through the first residual embedding module to obtain a sample of a semantic set of trend descriptions to be controlled, and the sensor monitoring log sample is integrated with the sample of the semantic set of trend descriptions to be controlled to obtain a control training log; The auxiliary control decision algorithm is debugged according to the comparison result between the control training log and the control instruction log sample.

10. A liquid cooling server control system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images