Cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation

Abstract

The application provides a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, relates to the field of transfer learning, and comprises the following steps: collecting historical sensor time series data, dividing a source domain and a target domain; calculating a heat transfer efficiency ratio, constructing a coupling feature matrix; performing cross-domain time series alignment processing, performing feature similarity measurement, and outputting the aligned feature representation; dynamically adjusting the source domain weight according to the divergence and water hardness information to obtain target fusion features; inputting the target fusion features into a trained fault classifier to output the fault diagnosis information and trend prediction information of the cooling tower. The application considers the thermodynamic parameters, water hardness and other physical states of the cooling tower, accurately differentiates the differences between different domains, avoids the interference of high-hardness source domains on low-hardness target domain diagnosis, considers the coupling feature matrix and domain-invariant features, enhances the recognition and diagnosis of the cooling tower fault mode, and improves the accuracy of the cooling tower composite fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transfer learning, in particular to a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation. BACKGROUND

[0002] With the rapid development of Internet, cloud computing, big data and other technologies, data centers, as the core infrastructure for information storage, processing and transmission, are expanding in number and scale, and play a vital role in the global economy and social operation. Cooling towers are the core components of the cooling system of data centers, responsible for dissipating the large amount of heat generated by servers and other equipment during operation, so that the data center can maintain a suitable temperature environment and ensure the stable operation of the equipment. Once the cooling tower fails, it may cause the temperature of the data center to rise, affecting the performance of servers and other equipment, and even causing equipment overheating damage, resulting in data loss, service interruption and other serious consequences, causing huge economic losses to enterprises and society.

[0003] In related technologies, most of the research on cooling tower fault diagnosis is based on a single data center. These researches have a high threshold in terms of data labeling requirements. Once the cooling tower of a new data center needs to be fault detected, it will face the problem of needing to re-label data and re-train the model. This undoubtedly wastes a lot of time and resources, including manpower for data labeling, computing resources for model training, and time cost.

[0004] Specifically, the related technology has the following defects: (1) The existing fault diagnosis model generally relies on full-amount labeled data of the target domain. However, in industrial scenarios, the acquisition of labeled data faces the problem of high cost, and in the cold start stage, there is not enough labeled data for the training of the fault diagnosis model; (2) The existing technology is based on the single-source domain homogeneity assumption, ignoring the distribution offset of sensor data caused by environmental conditions, load fluctuations, and device heterogeneity in different data centers; (3) Existing research is limited by the scale of single-domain data collection, and its training set only covers limited fault types and stable working conditions, making it difficult to cover complex and diverse fault modes and operating conditions, and limiting the model's ability to recognize and diagnose diversified faults; (4) In the existing multi-source domain fault diagnosis method, the domain difference is directly quantified using data-driven methods, ignoring the domain difference caused by physical factors, resulting in inaccurate domain difference quantification; (5) In the existing multi-source domain fault diagnosis method, the compound fault mode of the cooling tower (such as vibration anomaly caused by fouling) may be identified as an independent feature for processing, resulting in inaccurate fault diagnosis. SUMMARY

[0005] To overcome the defects of the prior art, the application provides a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, which solves the problems of dependence on labeled data, limitation on single-domain data, and insufficient fault recognition and diagnosis capability in current cooling tower fault diagnosis.

[0006] To achieve the above object, the application is implemented by the following technical solutions:

[0007] In a first aspect, the application provides a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, which includes: collecting historical sensor time series data of multiple data center cooling towers containing spatial position labels, dividing source and target domains and performing timestamp alignment, sampling frequency unification, and independent normalization processing on the data, learning and extracting domain-invariant features and label information of the source and target domains; extracting data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate the heat transfer efficiency ratio, and obtaining fan bearing feature band vibration acceleration to jointly construct a coupling feature matrix; based on the coupling feature matrix and the domain-invariant features, performing cross-domain time series alignment processing through contrast learning, combining the cooling tower unit area spray quantity difference and the filler type to construct positive sample pairs, and combining the spray density to separate the samples to construct negative sample pairs from different fault types, performing feature similarity measurement, and outputting the aligned feature representation; collecting circulating water conductivity through a circulating water conductivity sensor to determine water hardness information;

[0008] Based on the aligned feature representation, feature fusion is performed in combination with local time series patterns and global dependency relationships, the source domain weight is dynamically adjusted according to the KL divergence and the water hardness information, and the target fusion feature is obtained; the target fusion feature is input into the trained fault classifier, and the fault diagnosis information and trend prediction information of the cooling tower are output.

[0009] According to a first aspect of the embodiments of the present application, after the foregoing inputting the target fusion features into the trained fault classifier, outputting the fault diagnosis information and trend prediction information of the cooling tower, the multi-source adversarial domain adaptive cooling tower cross-domain fault diagnosis method specifically can further include the following steps: constructing a data distribution drift detection mechanism and a water quality parameter monitoring mechanism, monitoring the state change, and triggering the preset model retraining strategy when the state change reaches the preset target threshold; the data distribution drift detection mechanism includes: adopting a sliding window mechanism to continuously cache real-time sensor data streams of the target cooling tower; based on the cached data, periodically performing model parameter fine-tuning to make the model adapt to the slow drift of equipment operation; deploying a self-encoder to learn the feature representation of the normal operation data of the target domain, and calculating the reconstruction error of the real-time data; when the reconstruction error continuously exceeds the preset first threshold, it is determined that the current data distribution deviates significantly from the distribution when the model is trained, and it is determined that the state change reaches the preset target threshold. The water quality parameter monitoring mechanism includes: acquiring the makeup water conductivity through the makeup water conductivity sensor, monitoring the water quality parameters based on the makeup water conductivity and the circulating water conductivity, and determining the concentration multiple;

[0010] When the absolute deviation of the current concentration multiple and the concentration multiple when the model is trained exceeds the preset second threshold, it is determined that the state change reaches the preset target threshold; during the model retraining process, adaptive sampling is implemented, high concentration multiple period data corresponding to the scaling sensitive period is preferentially collected to optimize data efficiency and the sensitivity of the model to key faults; and the data sampling frequency or weight is increased to enable the model to focus on learning more discriminative features under conditions prone to scaling.

[0011] According to a first aspect of the embodiments of this application, the aforementioned cross-domain temporal alignment processing based on coupled feature matrices and domain-invariant features is performed through contrastive learning. Positive sample pairs are constructed by combining the difference in spray volume per unit area of ​​the cooling tower and the type of packing material. Negative sample pairs from different fault types are constructed by combining spray density for sample separation. Feature similarity is measured, and the aligned feature representation is output. Specifically, this may include the following steps: inputting coupled feature matrices and domain-invariant features; using a source domain pre-trained model to predict fault types; selecting high-confidence samples with a Softmax probability exceeding 0.9 as pseudo-labels; combining the pseudo-labels with the original labels of the target domain to determine new target domain labels; identifying time-series data segments from different domains but belonging to the same fault type, and those with small differences in spray volume per unit area. Samples with a value of 0.5 and the same filler type are considered positive sample pairs; the filler type includes film type and droplet type. Time series segments of different fault types are selected from the source domain and target domain to determine cross-domain different faults; and time series segments of different fault types within the same domain are determined as intra-domain different faults. Spray density greater than 1.2 is determined as high spray density, and spray density less than 0.8 is determined as low spray density. From the sample pairs constructed for cross-domain different faults and intra-domain different faults, the case where one sample has a high spray density and the other has a low spray density is removed to avoid the model using the difference in spray density as a feature to distinguish faults, resulting in negative sample pairs. Feature similarity is measured based on positive sample pairs and negative sample pairs, and the contrast loss value is calculated based on the contrast loss function. The aligned feature representation and contrast loss value are output.

[0012] According to a first aspect of the embodiments of this application, the aforementioned extraction of data characterizing the thermodynamic and vibration coupling characteristics from historical sensor time-series data to calculate the heat transfer efficiency ratio, and acquisition of the vibration acceleration of the characteristic frequency band of the fan bearing to jointly construct the coupling feature matrix, may specifically include the following steps: determining the heat transfer efficiency index data in the historical sensor time-series data, the heat transfer efficiency index data including: cooling tower inlet water temperature, cooling tower outlet water temperature, and circulating water mass flow rate;

[0013] The system obtains the preset heat transfer coefficient, effective packing area, and logarithmic mean temperature difference, and calculates the heat transfer efficiency ratio based on the heat transfer efficiency index data. It calculates the rate of change of heat transfer efficiency based on the heat transfer efficiency ratio, and constructs a coupled feature matrix based on the obtained vibration acceleration of the characteristic frequency band of the fan bearing. The heat transfer efficiency ratio and the rate of change of heat transfer efficiency are also used to trigger a scaling fault warning. When the heat transfer efficiency ratio is less than 0.85 and the rate of change of heat transfer efficiency decreases by more than 0.5 percentage points per minute, the scaling fault warning is triggered.

[0014] According to a first aspect of the embodiments of the present application, the spatial position marker includes water pipe segment position, filler layer height and fan radial coordinate; the historical sensor time series data includes time series data of corresponding vibration sensors, current sensors, water flow sensors, water pressure sensors, wind pressure sensors, temperature sensors, circulating water conductivity sensors and supplementary water conductivity sensors; wherein the time series data collected by the vibration sensors includes cooling tower fan bearing vibration data and speed reducer axial vibration data; the temperature sensors are used for inlet and outlet temperature measurement of the cooling tower. According to the first aspect of the embodiments of the present application, the fault diagnosis information includes fault type, spatial coordinate code and fouling probability; the trend prediction information includes fouling rate and fault development trend extrapolated by long short-term memory network (LSTM) hidden state, and the fault development trend includes vibration amplitude trend per unit time; the fault type includes fouling-related heat transfer efficiency decline, filler layer collapse, fan blade corrosion imbalance, water distributor pipe blockage, motor overload and bearing wear; the spatial coordinate code includes water pipe segment coordinate, filler layer height coordinate and fan radial coordinate corresponding to the spatial position marker.

[0015] According to the first aspect of the embodiments of the present application, the aforementioned KL divergence is used to quantify the difference between the multi-source domain and the target domain, dynamically adjust the source domain weight, and is related to the total loss function of the adversarial training. The aforementioned feature representation after alignment is combined with local time series patterns and global dependency relationships for feature fusion, and the target fusion feature is obtained according to the KL divergence and water quality hardness information dynamically adjusting the source domain weight. Specifically, it can include the following steps: based on the aligned feature representation, local feature extraction is performed through a one-dimensional convolutional neural network, and a convolution kernel is used to capture local patterns in the vibration signal, including transient impact, to obtain first target features; global feature extraction is performed based on a long short-term memory network (LSTM) to capture long-term dependency relationships of the signal to obtain second target features; the long-term dependency relationship includes a gradual trend; the first target features and the second target features are spliced to obtain multi-source fusion features; based on the multi-source fusion features, the source domain weight is dynamically adjusted through the KL divergence and the water quality hardness information to adjust the fusion weight of the multi-source features to obtain the target fusion feature.

[0016] According to a first aspect of the embodiments of the present application, the aforementioned collection of historical sensor time series data of multiple data center cooling towers containing spatial position labels, division of source domain and target domain and timestamp alignment, sampling frequency unification and independent normalization processing of data, learning and extracting domain-invariant features and label information of the source domain and the target domain, can specifically include the following steps: collecting historical sensor time series data of multiple data center cooling towers, determining a part of the data center cooling tower data of the historical sensor time series data as source domain data; determining another part of the data center cooling tower data of the historical sensor time series data as target domain data; using linear interpolation on all data to unify the time base to complete timestamp alignment; resampling all data to a unified frequency to align the sampling frequency, and filling in the missing values generated due to frequency conversion in the resampling process; wherein, in the case that the original sampling frequency is lower than the target frequency, the data points are increased by interpolation method; in the case that the original sampling frequency is higher than the target frequency, the data points are reduced by downsampling method; performing independent normalization processing on each type of sensor data to eliminate dimension differences; inputting the source domain data and the target domain data after timestamp alignment, sampling frequency unification and independent normalization processing into an adversarial training model; performing forward propagation based on a feature extraction network, a gradient reversal layer and a domain classifier to obtain domain prediction probability; wherein, the domain classifier is a 2-classification network; performing loss calculation and weighted processing based on a preset domain classification loss function , a preset fault classification loss function and domain prediction probability to obtain total loss; performing back propagation based on the total loss and updating the feature extraction network CNN, the domain classifier and the fault classifier to extract domain-invariant features and label information, and to minimize the domain classification loss to accurately distinguish the source domain and the target domain, and to minimize the fault classification loss to improve the fault classification accuracy; wherein, the fault classifier includes a fully connected layer and a Softmax activation layer and is used to map the aggregated features to fault category probability.

[0017] Secondly, the embodiments of the present application provide a cooling tower cross-domain fault diagnosis system based on multi-source adversarial domain adaptation, which comprises a first collection and analysis module, a feature extraction and analysis module, a cross-domain time series alignment module, a second collection and analysis module, a feature fusion module and a fault diagnosis and prediction module.

[0018] Specifically, the first collection and analysis module is configured to collect historical sensor time series data of multiple data center cooling towers containing spatial position labels, divide source and target domains, and perform timestamp alignment, sampling frequency unification, and independent normalization processing on the data, and learn to extract domain-invariant features and label information of the source and target domains; the feature extraction and analysis module is configured to extract data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate a heat transfer efficiency ratio, and obtain fan bearing feature band vibration acceleration to jointly construct a coupling feature matrix; the cross-domain time series alignment module is configured to perform cross-domain time series alignment processing based on the coupling feature matrix and the domain-invariant features through contrastive learning, construct positive sample pairs in combination with cooling tower unit area spray quantity differences and filler types, construct negative sample pairs from different fault types in combination with spray density, perform feature similarity measurement, and output aligned feature representations; the second collection and analysis module is configured to collect circulating water conductivity through a circulating water conductivity sensor to determine water hardness information; the feature fusion module is configured to perform feature fusion based on the aligned feature representations, in combination with local time series patterns and global dependency relationships, dynamically adjust source domain weights according to KL divergence and water hardness information, and obtain target fusion features; and the fault diagnosis and prediction module is configured to input the target fusion features into a trained fault classifier to output cooling tower fault diagnosis information and trend prediction information.

[0019] In a third aspect, an electronic device is provided, which includes a processor, a memory, and a program stored in the memory and executable on the processor, and the program, when executed by the processor, implements the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation in the first aspect.

[0020] In a fourth aspect, a computer-readable storage medium is provided, which stores a program or instructions, and the program or instructions, when executed by a processor, implement the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation in the first aspect.

[0021] The present application provides a method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation. Compared with the prior art, the following beneficial effects are achieved:

[0022] The application effectively reduces the distribution difference between the source domain and the target domain through the adversarial training mechanism, so that the domain-invariant features generated by the model are more focused on fault discrimination information, and the generalization ability of the model in different geographical regions and heterogeneous data centers is significantly improved; the application performs cross-domain time sequence alignment processing on the domain-invariant features of the source domain and the target domain based on contrast learning, triggers the model to retrain without the labeled data of the target domain, can significantly reduce the dependence on manual labeling, and reduces the labeling cost. Through the combination of local time sequence mode and global dependency relationship for multi-level time sequence feature fusion, and based on the divergence and water hardness information to dynamically adjust the source domain weight, multi-dimensional representation of the fault mode is realized, and the accuracy of the target domain fault diagnosis is improved. The application considers the thermodynamic parameters, water hardness and other physical states of the cooling tower, so that the difference between different domains is accurate, and the interference of the high-hardness source domain on the low-hardness target domain diagnosis is avoided. In the process of cross-domain time sequence alignment processing by contrast learning, the coupled feature matrix and the domain-invariant feature are considered comprehensively; the difference in the spraying amount per unit area of the cooling tower and the filler type are considered when constructing the positive sample pair; the spraying density is considered for sample separation when constructing the negative sample pair from different fault types, which enhances the recognition and diagnosis of the cooling tower fault mode and improves the accuracy of the cooling tower composite fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0024] Figure 1 is a flowchart of a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain self-adaption provided by an embodiment of the present application;

[0025] Figure 2 is a structural schematic diagram of a cooling tower cross-domain fault diagnosis system based on multi-source adversarial domain self-adaption provided by an embodiment of the present application;

[0026] Figure 3 is a schematic diagram of another cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain self-adaption provided by an embodiment of the present application;

[0027] Figure 4 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0029] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0030] The embodiments of the present application provide a cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, and solve the problems of dependence on labeled data, limitation on single-domain data, and insufficient fault recognition and diagnosis capability in current cooling tower fault diagnosis.

[0031] The technical solutions in the embodiments of the present application are as follows to solve the above technical problems:

[0032] With the rapid development of Internet, cloud computing, big data and other technologies, data centers, as the core infrastructure for information storage, processing and transmission, are expanding in number and scale, and play a vital role in global economic and social operation. The cooling tower is a core component of the cooling system of the data center, responsible for dissipating a large amount of heat generated during the operation of servers and other equipment, so that the data center can maintain a suitable temperature environment and ensure the stable operation of the equipment. Once the cooling tower fails, it may cause the temperature of the data center to rise, affecting the performance of servers and other equipment, and even causing equipment overheating damage, resulting in data loss, service interruption and other serious consequences, causing huge economic losses to enterprises and society.

[0033] In the related art, most of the researches on cooling tower fault diagnosis are based on a single data center. These researches have a high threshold in data labeling requirements. Once the cooling tower of a new data center is detected for fault, it needs to re-label the data and re-train the model. This undoubtedly wastes a lot of time and resources, including manpower for data labeling, computing resources for model training, and time cost.

[0034] Specifically, the related art has the following defects: (1) The existing fault diagnosis model generally relies on full-amount labeled data of the target domain. However, the acquisition of labeled data in industrial scenarios faces the problem of high cost, and in the cold start stage, there is not enough labeled data for the training of the fault diagnosis model. (2) The existing technology is based on the single-source domain homogeneity assumption, ignoring the distribution offset of sensor data caused by environmental conditions, load fluctuations, and device heterogeneity in different data centers. (3) The existing research is limited by the single-domain data collection scale, and the training set only covers limited fault types and stable working conditions, making it difficult to cover complex and diverse fault modes and operating conditions, and limiting the model's ability to recognize and diagnose diversified faults. (4) In the existing multi-source domain fault diagnosis method, the data-driven method is directly used to quantify the domain difference, ignoring the domain difference caused by physical factors, resulting in inaccurate domain difference quantification. (5) In the existing multi-source domain fault diagnosis method, the compound fault mode of the cooling tower (such as vibration anomaly caused by fouling) may be identified as an independent feature for processing, resulting in inaccurate fault diagnosis.

[0035] It should also be understood that adversarial domain adaptation is a method of aligning features between different domains through adversarial training. The core idea of adversarial domain adaptation is to use gradient reversal layers and domain discriminators to make the feature extractor generate domain-invariant features, thereby reducing the distribution difference between the source domain and the target domain. Through this method, the model can adapt to the sensor data distribution of the cooling tower in different data centers, so that the generalization ability and cross-domain adaptability of the model can be improved.

[0036] Multi-source transfer learning is a method of learning knowledge from multiple source domains and transferring it to the target domain. However, the data in different source domains differs in distribution, features, etc. For example, in the cooling tower fault diagnosis scenario, the cooling towers in each data center differ in environment, load, and equipment model, and the sensor data distribution of each data center is also different. Therefore, the weight needs to be dynamically adjusted according to the similarity between the source domain and the target domain, giving greater weight to the source domain with high similarity, and vice versa, so as to reduce the interference of the source domain with large differences on the learning of the target domain, and let the model focus better on the knowledge transfer related to the target domain. KL divergence is one of the commonly used methods for dynamically adjusting the weight. It can effectively quantify the feature distribution difference between different source domains and the target domain, providing a basis for weight allocation. This method enables the model to more reasonably utilize the information of multiple source domains, thereby improving the performance of the model in the target domain.

[0037] Therefore, in this case, it is particularly necessary to explore a cooling tower fault diagnosis method that can realize cross-domain. The cross-domain cooling tower fault diagnosis method can realize knowledge transfer and sharing between different data centers, thereby effectively avoiding repeated data labeling and model training work, improving diagnosis efficiency, reducing cost, and better adapting to the rapid development and diversification needs of data centers.

[0038] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the drawings in the specification and specific embodiments.

[0039] First, a multi-source anti-field adaptive cooling tower cross-domain fault diagnosis method provided by the embodiments of the present application will be introduced.

[0040] The flowchart of the multi-source anti-field adaptive cooling tower cross-domain fault diagnosis method provided by the embodiments of the present application is shown in Figure 1 and Figure 3 The multi-source anti-field adaptive cooling tower cross-domain fault diagnosis method can include the following steps S110-S160.

[0041] S110, collect historical sensor time series data of multiple data center cooling towers containing spatial position labels, divide source domains and target domains, and perform timestamp alignment, sampling frequency unification, and independent normalization processing on the data, and learn to extract domain-invariant features and label information of the source domains and the target domains.

[0042] S120, extract data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate the heat transfer efficiency ratio, and obtain the vibration acceleration of the fan bearing feature band to jointly construct a coupling feature matrix.

[0043] S130, based on the coupling feature matrix and the domain invariant feature, cross-domain time alignment processing is performed through contrastive learning, a positive sample pair is constructed in combination with the cooling tower unit area spray quantity difference and the filler type, and a negative sample pair from different fault types is constructed by combining the spray density for sample separation to perform feature similarity measurement, and output the aligned feature representation.

[0044] S140, the water quality hardness information is determined by collecting the circulating water conductivity through the circulating water conductivity sensor.

[0045] S150, based on the aligned feature representation, feature fusion is performed in combination with the local time sequence pattern and the global dependency relationship, the source domain weight is dynamically adjusted according to the KL divergence and the water quality hardness information, and the target fusion feature is obtained.

[0046] S160, input the target fusion feature into the trained fault classifier, and output the fault diagnosis information and trend prediction information of the cooling tower.

[0047] The above is a specific implementation of the cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation provided by the embodiment of the present application. It can be understood that the present application effectively reduces the distribution difference between the source domain and the target domain through the adversarial training mechanism, so that the domain invariant feature generated by the model focuses more on the fault discrimination information, and the generalization ability of the model in different geographical regions and heterogeneous data centers is significantly improved. The present application performs cross-domain time alignment processing on the domain invariant features of the source domain and the target domain based on contrastive learning, which can trigger the model to retrain without labeled data of the target domain, significantly reducing the dependence on manual labeling and reducing the labeling cost. The present application performs multi-level time sequence feature fusion by combining local time sequence patterns and global dependency relationships, and dynamically adjusts the source domain weight based on divergence and water quality hardness information, realizes multi-dimensional representation of fault patterns, and improves the accuracy of target domain fault diagnosis. The present application can quickly predict the fault development trend by real-time acquisition of multi-sensor data, and improves the overall operation and maintenance efficiency of the new data center cooling tower.

[0048] It should be noted that the present application considers the thermodynamic parameters, water quality hardness and other physical states of the cooling tower, so that the differences between different domains are accurate, and the high hardness source domain does not interfere with the diagnosis of the low hardness target domain. In the process of cross-domain time alignment processing through contrastive learning, the coupling feature matrix and the domain invariant feature are considered comprehensively; when constructing the positive sample pair, the cooling tower unit area spray quantity difference and the filler type are considered; when constructing the negative sample pair from different fault types, the spray density is considered for sample separation, which enhances the recognition and diagnosis of the cooling tower fault mode and improves the accuracy of the cooling tower composite fault diagnosis.

[0049] In one example, the spatial position label includes the water pipe segment position, the filler layer height and the fan radial coordinate;

[0050] The historical sensor time series data includes time series data of corresponding vibration sensors, current sensors, water flow sensors, water pressure sensors, wind pressure sensors, temperature sensors, circulating water conductivity sensors, and supplementary water conductivity sensors; wherein the time series data collected by the vibration sensors includes cooling tower fan bearing vibration data and speed reducer axial vibration data; the temperature sensors are used for inlet and outlet temperature measurement of the cooling tower.

[0051] It can be understood that the present application first carries out multi-source sensor data acquisition and domain division, selects multiple data centers (≥3), and all cooling tower historical data of each data center as a source domain, covering the time series data of various sensors described above, and labeling fault types, positions and timestamps. And select all cooling tower historical data of an independent data center as a target domain, covering the time series data of various sensors described above, which requires a small amount of fault labeling. The data frequency range collected by the vibration sensor is 5Hz-2000Hz.

[0052] In another example, the process of determining water quality hardness information according to circulating water conductivity satisfies the expression:

[0053]

[0054] In the formula, H represents water quality hardness information, and EC represents circulating water conductivity.

[0055] In some embodiments, after the foregoing inputting the target fusion features into the trained fault classifier, outputting the fault diagnosis information and trend prediction information of the cooling tower, the cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation can further include the following steps:

[0056] S170, a data distribution drift detection mechanism and a water quality parameter monitoring mechanism are constructed, state change monitoring is performed, and in the case that the state change reaches a preset target threshold, a preset model retraining strategy is triggered.

[0057] The foregoing data distribution drift detection mechanism can specifically include the following steps:

[0058] S171, a sliding window mechanism is used to continuously cache real-time sensor data streams of the target cooling tower;

[0059] S172, based on the cached data, periodically perform model parameter fine-tuning to make the model adapt to the slow drift of equipment operation;

[0060] S173, deploy an autoencoder to learn the feature representation of the normal operation data of the target domain, and calculate the reconstruction error of the real-time data;

[0061] S174, when the reconstruction error continues to exceed the preset first threshold value, it is determined that the current data distribution is significantly deviated from the distribution when the model is trained, and it is determined that the state change reaches the preset target threshold value.

[0062] In the embodiments of the present application, it can be understood that when the reconstruction error continues to exceed the preset first threshold value, it indicates that the current data distribution has deviated significantly from the distribution when the model is trained, triggering the model retraining.

[0063] The foregoing water quality parameter monitoring mechanism can specifically include the following steps:

[0064] S175, acquiring the makeup water conductivity through the makeup water conductivity sensor, monitoring the water quality parameters based on the makeup water conductivity and the circulating water conductivity, and determining the concentration multiple;

[0065] S176, when the absolute deviation of the current concentration multiple from the concentration multiple when the model is trained exceeds the preset second threshold value, it is determined that the state change reaches the preset target threshold value;

[0066] S177, during the model retraining process, adaptive sampling is implemented, high concentration multiple period data corresponding to the scaling sensitive period is preferentially collected, so as to optimize the data efficiency and the sensitivity of the model to key faults; and the data sampling frequency or weight is increased, so that the model focuses on learning more discriminative features under the condition of easy scaling.

[0067] In the embodiments of the present application, it can be understood that in order to ensure that the diagnostic model can continuously adapt to the equipment state change and new working conditions after being deployed in the target data center, the online adaptive updating strategy is adopted; the present application performs joint monitoring in two dimensions based on the data distribution drift detection mechanism and the water quality parameter monitoring mechanism. The window size of the sliding window mechanism can be set to 1 hour, and the calculation of the concentration multiple CT satisfies the expression:

[0068]

[0069] In the expression, represents the circulating water conductivity, represents the makeup water conductivity.

[0070] Regarding the water quality parameter monitoring mechanism, the trigger condition for determining that the state change reaches the preset target threshold value and triggering the preset model retraining strategy satisfies the expression:

[0071]

[0072] In the expression, represents the current concentration multiple, The concentration multiple during model training is represented. The second threshold is 1.5. When the absolute deviation of the current concentration multiple from the concentration multiple during model training exceeds the preset second threshold, it indicates that the water quality condition has changed significantly (such as pollution control strategy change, water source switching), which is easy to affect the scaling tendency and heat transfer efficiency, and the model needs to be retrained to adapt to the new water quality condition.

[0073] It should be noted that the online updating mechanism is optimized, and the concentration multiple monitoring and adaptive sampling strategy are added to ensure that the system can dynamically adapt to the working state of the cooling tower equipment in actual operation and adjust the monitoring frequency in time.

[0074] In some embodiments, the foregoing collection of historical sensor time series data of multiple data center cooling towers containing spatial position markers, division of source domain and target domain and time stamp alignment, sampling frequency unification and independent normalization processing of data, learning to extract domain-invariant features and label information of source domain and target domain, that is, S110 can specifically include the following steps:

[0075] S210, collecting historical sensor time series data of multiple data center cooling towers, determining a part of data center cooling tower data of the historical sensor time series data as source domain data; determining another part of data center cooling tower data of the historical sensor time series data as target domain data;

[0076] S220, using linear interpolation to unify the time base to complete time stamp alignment;

[0077] S230, resampling all data to a unified frequency to align the sampling frequency, and filling in the missing values generated in the resampling process; wherein, in the case that the original sampling frequency is lower than the target frequency, the data points are increased by interpolation method; in the case that the original sampling frequency is higher than the target frequency, the data points are reduced by downsampling method;

[0078] S240, independently normalizing each type of sensor data to eliminate dimension differences;

[0079] S250, inputting the source domain data and the target domain data after time stamp alignment, sampling frequency unification and independent normalization processing into the adversarial training model;

[0080] S260, performing forward propagation based on the feature extraction network, the gradient reversal layer and the domain classifier to obtain domain prediction probability; wherein, the domain classifier is a 2-classification network;

[0081] S270, performing loss calculation and weighted processing based on the preset domain classification loss function , the preset fault classification loss function and the domain prediction probability to obtain the total loss;

[0082] S280, based on the total loss, the feature extraction network CNN, the domain classifier and the fault classifier are updated to extract domain-invariant features and label information, and to minimize the domain classification loss to accurately distinguish the source domain and the target domain, and to minimize the fault classification loss to improve the fault classification accuracy;

[0083] Wherein, the fault classifier includes a fully connected layer and a Softmax activation layer and is used to map the aggregated features to fault class probabilities. The present application uses an Adam optimizer, and the learning rate decays with the training round; the present application jointly trains the fault classifier based on the contrast loss, the classification loss and the adversarial loss.

[0084] In the embodiments of the present application, it can be understood that the present application selects multiple data centers, determines all cooling tower historical data of each data center as a source domain, covers time series data of vibration, current, water flow and pressure sensors, and labels fault types, locations and timestamps. The present application additionally selects all cooling tower historical data of one or more independent data centers as a target domain, covers time series data of vibration, current, water flow and pressure sensors, and requires a small amount of fault labeling for the historical data.

[0085] The present application effectively aligns the feature distribution of different data centers through the adversarial training mechanism, so that the model fully considers the differences of multi-source domain data in the learning process, and generates domain-invariant features, thereby better adapting to the distribution characteristics of the target data center and improving the diagnostic performance. Adversarial training involves forward propagation, loss calculation, back propagation and parameter update.

[0086] In the forward propagation, the feature extraction network generates features based on the source domain data and the target domain data through a 3-layer convolution + max pooling feature extraction network CNN ; the gradient reversal layer involves forward propagation and back propagation; in the forward propagation, the feature is directly outputted, i.e. ; in the back propagation, the gradient is multiplied by the reversal coefficient (initial value 1.0, which can be dynamically adjusted); the domain classifier is a binary classification network (2-layer fully connected), and the input of the domain classifier is the feature , and the output is the domain prediction probability

[0087] In the loss calculation, the domain classification loss function satisfies the expression:

[0088]

[0089] In the formula, wherein, is the number of samples of the source domain, is the number of samples of the target domain; : domain label (1 represents the source domain, and 0 represents the target domain); : prediction probability of the sample by the domain classifier.

[0090] fault classification loss function satisfies the expression:

[0091]

[0092] wherein, is the number of samples of the source domain, is the true class index of the i-th sample (for example, 0 represents "piping blockage"), represents the probability value corresponding to the true class index i in the probability distribution output by the model.

[0093] The total loss function satisfies the expression:

[0094]

[0095] wherein, is the adversarial loss weight, and the initial value is 1.0, which is dynamically adjusted according to the validation set.

[0096] It should be noted that, based on the similarity between the source domain and the target domain, the transfer weight is dynamically adjusted in the present application, so that the model more reasonably utilizes the information of multiple source domains, and the generalization ability and adaptability of the model are further improved. Through data fusion across data centers, the model can more accurately identify the fault type, locate the fault position, and predict the fault development trend, thereby improving the accuracy and reliability of fault diagnosis. Under the multi-source adversarial field self-adaptive framework, after aligning the feature distribution through adversarial training, a small amount of labeled data of the target data center can quickly deploy a high-precision diagnosis model, greatly reducing the labeling cost and cold start difficulty, and improving the applicability of the model in actual industrial scenarios.

[0097] In some embodiments, the aforementioned extracting data representing thermodynamic and vibration coupling features from historical sensor time series data to calculate the heat transfer efficiency ratio, and obtaining the vibration acceleration of the fan bearing feature frequency band to jointly construct a coupling feature matrix, i.e., the aforementioned S120 can specifically include the following steps:

[0098] S310, determining heat transfer efficiency index data in historical sensor time series data, the heat transfer efficiency index data including: cooling tower inlet water temperature, cooling tower outlet water temperature, and circulating water mass flow rate;

[0099] S320, obtaining a preset heat transfer coefficient, effective area of filler, and logarithmic mean temperature difference, and combining the heat transfer efficiency index data to calculate and determine the heat transfer efficiency ratio;​​​​

[0100] S330, based on the heat transfer efficiency ratio, the heat transfer efficiency change rate is calculated, and the obtained fan bearing characteristic frequency band vibration acceleration is combined to construct a coupling characteristic matrix;

[0101] The heat transfer efficiency ratio and the heat transfer efficiency change rate are also used to trigger a fouling fault warning; in the case that the heat transfer efficiency ratio is less than 0.85 and the heat transfer efficiency change rate decreases by more than 0.5 percentage points per minute, the fouling fault warning is triggered.

[0102] In the embodiments of the present application, it can be understood that, in order to enhance the cross-domain discrimination ability of composite faults (such as structural accompanying vibration anomalies), the present application introduces thermodynamic and vibration coupling characteristics for fault detection. In view of the strong coupling characteristics of the cooling tower heat transfer efficiency and mechanical vibration, the heat transfer efficiency ratio is calculated in real time. The calculation process of the heat transfer efficiency ratio satisfies the expression:

[0103]

[0104] In the formula, represents the heat transfer efficiency ratio, represents the cooling tower inlet water temperature, represents the cooling tower outlet water temperature, represents the circulating water mass flow rate, K represents the heat transfer coefficient, and A represents the effective area of the filler, represents the logarithmic mean temperature difference.

[0105] The foregoing coupling characteristic matrix satisfies the expression:

[0106]

[0107] In the formula, represents the coupling characteristic matrix, represents the heat transfer efficiency change rate, is the vibration acceleration of the fan bearing characteristic frequency band, more specifically, refers to the effective value of the vibration acceleration of the fan bearing in the 200Hz-500Hz frequency band; The calculation steps of are as follows: the original vibration signal collected by the vibration sensor of the fan bearing is subjected to band-pass filtering at 200Hz-500Hz to retain the signal components in this frequency band, and the root mean square value of the filtered signal is calculated as the effective value of the vibration acceleration in this frequency band. When The numerical mixing of the three variables is changed to determine the fault, for example, in the case that the heat transfer efficiency ratio is less than 0.85 and the heat transfer efficiency change rate decreases by more than 0.5 percentage points per minute, the fouling fault warning is triggered. The thermodynamic efficiency of the cooling tower is combined with the vibration signal to form a new coupling feature matrix, which is particularly critical for fault diagnosis. In the early stage of fouling and fan failure, the early warning accuracy can be significantly improved.

[0108] In some embodiments, the aforementioned coupling feature matrix and domain-invariant feature are used for cross-domain time alignment processing by contrast learning, and the positive sample pairs are constructed by combining the cooling tower unit area spray amount difference and the filler type, and the negative sample pairs from different fault types are constructed by combining the spray density and sample separation, the feature similarity is measured, and the aligned feature representation is output, that is, the aforementioned S130 can include the following steps:

[0109] S410, input the coupling feature matrix and the domain-invariant feature, use the source domain pre-training model to predict the fault type, and select the high-confidence samples with Softmax probability exceeding 0.9 as pseudo-labels; the pseudo-labels are used in combination with the original labels of the target domain to determine new target domain labels;

[0110] S420, determine the time series data segments from different domains but belonging to the same fault type, and the samples with unit area spray amount difference less than 0.5 and the same filler type as positive sample pairs; wherein the filler type includes thin film and point drop;

[0111] S430, select time series segments of different fault types from the source domain and the target domain to determine cross-domain different faults; and determine the time series segments of different fault types in the same domain as same-domain different faults;

[0112] S440, determine that the spray density greater than 1.2 is high spray density, and the spray density less than 0.8 is low spray density;

[0113] S450, remove one sample as high spray density and the other sample as low spray density from the sample pairs constructed from cross-domain different faults and same-domain different faults, to avoid the model using the spray density difference as a feature to distinguish faults, to obtain negative sample pairs,

[0114] S460, measure the feature similarity based on the positive sample pairs and the negative sample pairs, and calculate the contrast loss value based on the contrast loss function, and output the aligned feature representation and the contrast loss value.

[0115] In the embodiments of the present application, it can be understood that the present application constructs positive and negative sample pairs by using contrastive learning, and through cosine similarity measurement and a contrastive loss function, forces the time sequence features of the same type of faults in the cross-domain to cluster in the embedding space, and different fault features to separate, thereby improving the robustness of the model to domain differences and the discrimination ability of the fault features. In the present application, only a small amount of labeled data is required in the target domain, and by pre-training the model in the source domain to generate pseudo-labels and combining contrastive learning, the cost of obtaining labeled data in an industrial scene can be greatly reduced, and the problem of insufficient data in the cold start stage can be solved.

[0116] To improve the quality of positive samples and the fault discrimination degree, the present application selects samples with a spray volume difference of less than 0.5 L / (min·m²) per unit area and the same type of filler (the type of filler is distinguished by thin film and point drop), so as to ensure that the positive samples not only match in fault type, but also have high similarity in key operating conditions (spray density) and core equipment structure (filler type), thereby reducing the feature deviation caused by physical condition differences and improving the effectiveness of contrastive learning.

[0117] When constructing negative sample pairs, samples with high and low spray densities are actively separated, thereby enhancing the discrimination ability of the model for the two types of easily confused faults. The total spray water volume of the cooling tower divided by the effective spray area of the cooling tower is the spray density. High spray density (>1.2 L / min·m²) can accelerate fouling, and low spray density (<0.8 L / min·m²) can cause cavitation; by adding spray density as a new constraint condition to screen positive and negative samples, the confusion problem of fouling and cavitation faults in the traditional method can be effectively solved. The fault discrimination based on the spray density can effectively enhance the robustness of the fault diagnosis model and avoid misjudgment.

[0118] In other words, the negative sample pair must come from different fault types; at the same time, if one of the two samples has high spray density and the other has low spray density, they are not allowed to form a negative sample pair (even if the fault types are different). Other conditions (such as both high, both low, or both in the middle interval, or one in the middle interval and the other in the high / low interval) are allowed to form a negative sample pair. The purpose of this is to avoid the model using the spray density difference as a feature to distinguish faults, and thus to focus on the real fault features. At the same time, by separating the spray density intervals, the model can more clearly distinguish between fouling and cavitation faults.

[0119] It should be noted that the present application performs cross-domain time sequence alignment processing on the domain-invariant feature and the coupled feature matrix of the source domain and the target domain based on contrastive learning, so as to align the time sequence features of the same fault in the cross-domain and enhance the robustness of the model to domain differences. The source domain time sequence feature is the time sequence feature after domain-invariant feature extraction:

[0120] ( is the number of source domain samples,T is a time step, D is a feature dimension);

[0121] The target domain timing feature is a timing feature extracted after domain-invariant feature extraction:

[0122] is the number of target domain samples);The label information is the fault type label of the source domain data and a small amount of labels of the target domain.

[0123] In addition, the cross-domain same fault is a timing segment of the same fault type selected from the source domain and the target domain (the fault label of the target domain is derived from the original small amount of labels and the pseudo-labels generated before);The cross-domain different fault is a timing segment of different fault types selected from the source domain and the target domain;The same-domain different fault is a timing segment of different fault types in the same domain;The feature similarity measure can use the cosine similarity calculation method. Regarding the aligned feature representation, the timing features of the cross-domain same fault are closer in the embedding space, and the different fault features are farther apart;The contrast loss value can be used to jointly optimize the model.

[0124] In one example, the contrast loss function satisfies the expression:

[0125]

[0126] In the formula, is the batch size, that is, the number of samples processed simultaneously in one training iteration, refers to the anchor sample, that is, the sample of the source domain, is the positive sample of the same fault as the anchor sample in the target domain, that is, the corresponding sample of the target domain; refers to the negative sample (different fault or different domain sample), and s is a similarity function, refers to the temperature parameter (usually set to 0.1-1.0), which is used to control the discrimination of the similarity; M is the number of negative samples.

[0127] In some embodiments, the aforementioned KL divergence is used to quantify the difference between the multi-source domain and the target domain, dynamically adjust the source domain weight, and is related to the total loss function of the adversarial training. Based on the aligned feature representation, the local timing pattern and the global dependency relationship are combined for feature fusion, the source domain weight is dynamically adjusted according to the KL divergence and the water hardness information, and the target fusion feature is obtained, that is, the aforementioned S150 can specifically include the following steps:

[0128] S510, based on the aligned feature representation, local feature extraction is performed through a one-dimensional convolutional neural network, a convolution kernel is used to capture the local pattern in the vibration signal including transient impact, and a first target feature is obtained; ​

[0129] S520, global feature extraction is performed based on a long short-term memory network (LSTM) to capture long-term dependence of the signals, and second target features are obtained; the long-term dependence includes a gradual trend;

[0130] S530, the first target features and the second target features are spliced to obtain multi-source fusion features;

[0131] S540, based on the multi-source fusion features, source domain weights are dynamically adjusted by KL divergence and water hardness information, and fusion weights of the multi-source features are adjusted to obtain target fusion features.

[0132] In the embodiments of the present application, it can be understood that the present application combines one-dimensional convolutional neural network to extract local transient features of the vibration signals and LSTM to capture global long-term dependence of the signals such as current and water flow, generates fusion features containing multi-dimensional information through feature splicing, and improves the comprehensiveness of fault mode recognition. Through local feature extraction and global feature extraction on the aligned feature representation, the model can cover complex and diverse fault modes. For example, motor overload can be manifested as a long-term slow rise of the current signal (global gradual trend, captured by LSTM), accompanied by an increase in high-frequency transient impact in the motor vibration signal (local burst feature, captured by one-dimensional CNN), and the two together constitute the coupled fault features of motor overload. For another example, pipeline blockage can cause a sudden drop in water flow (local mutation of the flow sensor), while causing abnormal rise of the water pump pressure (long-term trend change of the pressure sensor), and accompanied by changes in the vibration mode of the water pump, and the coupling of the three can fully reflect the essence of the blockage fault.

[0133] The present application dynamically adjusts the source domain weights by KL divergence and water hardness information, realizes optimal fusion of multi-source features, and obtains target fusion features. By introducing water hardness information to optimize weight distribution, the physical differences (such as water hardness) between the multi-source domain and the target domain are accurately quantified. Considering water hardness in the multi-source domain fault diagnosis of the cooling tower mainly because different water hardness has a significant impact on the heat transfer efficiency of the cooling tower, such as the formation of water scale and equipment corrosion. Dynamic hardness correction is performed according to the water quality difference in different regions, and weight distribution is combined with KL divergence, which significantly improves the accuracy of cross-domain transfer learning. In particular, the sensitive coefficient of the hardness difference is added as part of the dynamic adjustment, which can effectively alleviate the interference of water quality difference on the diagnosis model.

[0134] In one example, the fault diagnosis information includes a fault type, a spatial coordinate code, and a fouling probability; the trend prediction information includes a fouling rate and a fault development trend extrapolated by a long short-term memory network (LSTM) hidden state, and the fault development trend includes a vibration amplitude trend per unit time; the fault type includes a heat transfer efficiency decline related to fouling, a packing layer collapse, a fan blade corrosion imbalance, a pipeline blockage of a water distributor, a motor overload, and bearing wear; and the spatial coordinate code includes a water pipe segment coordinate corresponding to a spatial position marker, a packing layer height coordinate, and a fan radial coordinate.

[0135] In some embodiments, the present application provides a cooling tower cross-domain fault diagnosis system 600 based on multi-source adversarial domain adaptation, as shown in Figure 2 The cooling tower cross-domain fault diagnosis system 600 based on multi-source adversarial domain adaptation can include the following modules:

[0136] A first collection and analysis module 610 is configured to collect historical sensor time series data of a plurality of data center cooling towers containing spatial position markers, divide source and target domains, and perform timestamp alignment, sampling frequency unification, and independent normalization processing on the data, and learn and extract domain-invariant features and label information of the source and target domains;

[0137] A feature extraction and analysis module 620 is configured to extract data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate a heat transfer efficiency ratio, and obtain fan bearing feature band vibration acceleration to jointly construct a coupling feature matrix;

[0138] A cross-domain time series alignment module 630 is configured to perform cross-domain time series alignment processing based on the coupling feature matrix and the domain-invariant features through contrastive learning, construct positive sample pairs in combination with cooling tower unit area spray quantity differences and packing types, construct negative sample pairs from different fault types in combination with spray density, perform feature similarity measurement, and output aligned feature representations;

[0139] A second collection and analysis module 640 is configured to collect circulating water conductivity through a circulating water conductivity sensor to determine water hardness information;

[0140] A feature fusion module 650 is configured to perform feature fusion based on the aligned feature representations in combination with local time series patterns and global dependency relationships, dynamically adjust source domain weights according to KL divergence and water hardness information, and obtain target fusion features;

[0141] A fault diagnosis and prediction module 660 is configured to input the target fusion features into a trained fault classifier, and output fault diagnosis information and trend prediction information of the cooling tower.

[0142] According to embodiments of the present application, any multiple of the first acquisition analysis module 610, the feature extraction analysis module 620, the cross-domain time alignment module 630, the second acquisition analysis module 640, the feature fusion module 650 and the fault diagnosis and prediction module 660 can be combined in one module, or any one of them can be split into multiple modules. Alternatively, at least part of the function of one or more of these modules can be combined with at least part of the function of other modules and implemented in one module.

[0143] Figure 2 Each module in the system shown has the function of implementing each step in the foregoing cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, and can achieve its corresponding technical effects. For the sake of brevity, no further description is given here.

[0144] In some embodiments, the present application provides an electronic device, a structural schematic diagram of which is shown as Figure 4

[0145] The electronic device can include a processor 710 and a memory 720 storing computer program instructions.

[0146] Specifically, the processor 710 described above can include a central processing unit (CPU), or a specific integrated circuit (ASIC), or can be configured to implement one or more integrated circuits of the embodiments of the present application.

[0147] The memory 720 can include a mass storage for data or instructions. By way of example and not limitation, the memory 720 can include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. Where appropriate, the memory 720 can include removable or non-removable (or fixed) media. Where appropriate, the memory 720 can be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, the memory 720 is a non-volatile solid-state memory.

[0148] ​The memory 720 can include read-only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical / tangible memory storage devices. Thus, generally, the memory 720 includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software that, when executed (e.g., by the one or more processors), is configured to perform the operations described above in any of the embodiments of the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation.

[0149] The processor 710 implements any of the embodiments of the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation by reading and executing computer program instructions stored in the memory 720.

[0150] In one example, the electronic device can further include a communication interface 730 and the bus 700. As shown, the processor 710, the memory 720, and the communication interface 730 are connected through the bus 700 and complete communication with each other. Figure 4

[0151] The communication interface 730 is mainly used to realize the communication between the modules, devices, units, and / or equipment in the embodiments of the present application.

[0152] The bus 700 includes hardware, software, or both, which couples the components of the online data traffic billing device to each other. By way of example, and not limitation, the bus can include an accelerated graphics port (AGP) or other graphics bus, an enhanced industry standard architecture (EISA) bus, a front-side bus (FSB), a HyperTransport (HT) interconnect, an industry standard architecture (ISA) bus, an infiniband interconnect, a low pin count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a peripheral component interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a serial advanced technology attachment (SATA) bus, a video electronics standards association local (VLB) bus, or another suitable bus or a combination of two or more of these. Where appropriate, the bus 700 can include one or more buses. Although the present application describes and illustrates a particular bus, the present application contemplates any suitable bus or interconnect.

[0153] In addition, in combination with the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation described above, the embodiments of the present application can provide a computer storage medium to implement. The computer storage medium has computer program instructions stored thereon; the computer program instructions are executed by the processor to implement any of the embodiments of the method for cooling tower cross-domain fault diagnosis based on multi-source adversarial domain adaptation. ​

[0154] It is to be understood that the present application is not limited to the particular examples described and illustrated herein, and that the application includes a variety of configurations and processes. For simplicity, detailed descriptions of well-known methods and apparatuses are omitted so as not to obscure the description of the present application. In the above embodiments, several specific steps are described and illustrated as examples. However, the methods of the present application are not limited to the specific steps described and illustrated, and one skilled in the art can make various changes, modifications and additions, or can change the order of steps, after understanding the spirit of the present application.

[0155] The functional blocks shown in the structural block diagrams above can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, and the like. When implemented in software, the elements of the present application are program or code segments that are used to perform the required tasks. The program or code segments can be stored in a machine-readable medium or transmitted through a data signal carried in a carrier wave over a transmission medium or communication link. A "machine-readable medium" includes any medium that can store or transfer information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, and the like. The code segments can be downloaded via a computer network such as the Internet, an intranet, and the like.

[0156] It is also to be understood that the example embodiments described in the present application are based on a series of steps or apparatuses to describe some methods or systems. However, the present application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiments, or in an order different from the embodiments, or several steps can be performed simultaneously.

[0157] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0158] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation, characterized in that, The method comprises the following steps: Collecting historical sensor time series data of multiple data center cooling towers containing spatial position markers, dividing source and target domains and performing timestamp alignment, sampling frequency unification and independent normalization processing on the data, and learning to extract domain-invariant features and label information of the source and target domains; Extracting data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate the heat transfer efficiency ratio, and obtaining vibration acceleration in the bearing feature frequency band to jointly construct a coupling feature matrix; Based on the coupling feature matrix and the domain-invariant features, cross-domain time series alignment processing is performed through contrastive learning, positive sample pairs are constructed by combining the differences in cooling tower unit area spray quantity and the types of fillers, and negative sample pairs from different fault types are constructed by combining spray density to separate samples, feature similarity is measured, and the aligned feature representation is outputted; Collecting circulating water conductivity through a circulating water conductivity sensor to determine water hardness information; Based on the aligned feature representation, feature fusion is performed by combining local time series patterns and global dependency relationships, the source domain weight is dynamically adjusted according to the KL divergence and the water hardness information, and the target fusion feature is obtained; The target fusion feature is inputted into a trained fault classifier to output fault diagnosis information and trend prediction information of the cooling tower.

2. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to claim 1, wherein, After the target fusion feature is inputted into the trained fault classifier to output fault diagnosis information and trend prediction information of the cooling tower, the cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation further comprises: A data distribution drift detection mechanism and a water quality parameter monitoring mechanism are constructed to monitor state changes, and a preset model retraining strategy is triggered when the state change reaches a preset target threshold; The data distribution drift detection mechanism comprises: A sliding window mechanism is used to continuously cache real-time sensor data streams of the target cooling tower; Based on the cached data, model parameter fine-tuning is performed regularly to make the model adapt to the slow drift of equipment operation; A self-encoder is deployed to learn the feature representation of the target domain normal operation data, and the reconstruction error of real-time data is calculated; When the reconstruction error continuously exceeds a preset first threshold, it is determined that the current data distribution significantly deviates from the distribution during model training, and it is determined that the state change reaches the preset target threshold; The water quality parameter monitoring mechanism comprises: Supplement water conductivity is collected through a supplement water conductivity sensor, and water quality parameter monitoring is performed based on the supplement water conductivity and the circulating water conductivity to determine the concentration multiple; When the absolute deviation of the current concentration multiple from the concentration multiple during model training exceeds a preset second threshold, it is determined that the state change reaches the preset target threshold; During the model retraining process, adaptive sampling is implemented, high concentration multiple period data corresponding to the scaling sensitive period is preferentially collected to optimize data efficiency and the sensitivity of the model to key faults, and the data sampling frequency or weight is increased to enable the model to focus on learning more discriminative features under conditions prone to scaling.

3. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to claim 1, wherein, The cross-domain time sequence alignment processing is performed through contrast learning based on the coupling feature matrix and the domain-invariant feature, positive sample pairs are constructed by combining the cooling tower unit area spray quantity difference and the filler type, and negative sample pairs from different fault types are constructed by combining the spray density, feature similarity is measured, and aligned feature representations are output, including: The coupling feature matrix and the domain-invariant feature are input, a source domain pre-training model is used to predict the fault type, and high-confidence samples with Softmax probability exceeding 0.9 are selected as pseudo-labels; the pseudo-labels are used in combination with original labels of a target domain to determine new target domain labels; Time sequence data segments from different domains but belonging to the same fault type, and samples with unit area spray quantity difference less than 0.5 and the same filler type are determined as positive sample pairs; the filler type includes thin film type and point drop type; Time sequence segments of different fault types are selected from the source domain and the target domain to determine cross-domain different faults, and time sequence segments of different fault types in the same domain are determined as same-domain different faults; Spray density greater than 1.2 is determined as high spray density, and spray density less than 0.8 is determined as low spray density; In the sample pairs constructed from the cross-domain different faults and the same-domain different faults, one sample is the high spray density and the other sample is the low spray density, to avoid the model regarding the spray density difference as a feature for distinguishing faults, to obtain negative sample pairs, Feature similarity is measured based on the positive sample pairs and the negative sample pairs, a contrast loss value is calculated based on a contrast loss function, and aligned feature representations and the contrast loss value are output.

4. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to claim 1, characterized in that, The data representing thermodynamic and vibration coupling features are extracted from the historical sensor time sequence data to calculate the heat transfer efficiency ratio, and the vibration acceleration of the fan bearing feature frequency band is obtained to jointly construct a coupling feature matrix, including: The heat transfer efficiency index data in the historical sensor time sequence data are determined, including: cooling tower inlet water temperature, cooling tower outlet water temperature, and circulating water mass flow rate; A preset heat transfer coefficient, effective area of filler, and logarithmic mean temperature difference are obtained, and the heat transfer efficiency ratio is calculated and determined in combination with the heat transfer efficiency index data; The heat transfer efficiency change rate is calculated based on the heat transfer efficiency ratio, and the coupling feature matrix is constructed in combination with the obtained vibration acceleration of the fan bearing feature frequency band; The heat transfer efficiency ratio and the heat transfer efficiency change rate are also used to trigger a fouling fault warning; in the case that the heat transfer efficiency ratio is less than 0.85 and the heat transfer efficiency change rate decreases by more than 0.5 percentage points per minute, the fouling fault warning is triggered.

5. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to any one of claims 1-4, characterized in that, The spatial position markers include water pipe segmented positions, filler layer heights, and fan radial coordinates. The historical sensor time series data includes time series data of corresponding vibration sensors, current sensors, water flow sensors, water pressure sensors, wind pressure sensors, temperature sensors, circulating water conductivity sensors, and makeup water conductivity sensors; wherein the time series data collected by the vibration sensors includes cooling tower fan bearing vibration data and speed reducer axial vibration data; the temperature sensors are used for inlet and outlet temperature measurement of the cooling tower; The fault diagnosis information includes fault types, spatial coordinate encodings, and fouling probabilities; the trend prediction information includes a fouling rate and a fault development trend extrapolated by a long short-term memory network (LSTM) hidden state, and the fault development trend includes a vibration amplitude trend per unit time; The fault types include fouling-related heat transfer efficiency reduction, filler layer collapse, fan blade corrosion imbalance, water distributor pipeline blockage, motor overload, and bearing wear; The spatial coordinate encodings include water pipe segment coordinates, filler layer height coordinates, and fan radial coordinates corresponding to the spatial position markers.

6. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to any one of claims 1-4, characterized in that, The KL divergence is used to quantify the difference between the multi-source domain and the target domain, dynamically adjust the source domain weight, and is related to the total loss function of the adversarial training; Based on the aligned feature representation, local temporal patterns and global dependency relationships are combined for feature fusion, and the source domain weight is dynamically adjusted based on the KL divergence and the water hardness information to obtain target fusion features, including: Based on the aligned feature representation, local features are extracted by a one-dimensional convolutional neural network, and a convolution kernel is used to capture local patterns in the vibration signal, including transient impacts, to obtain first target features; Global features are extracted based on a long short-term memory network (LSTM) to capture long-term dependencies of the signal, and second target features are obtained; the long-term dependencies include gradual trends; The first target features and the second target features are spliced to obtain multi-source fusion features; Based on the multi-source fusion features, the source domain weight is dynamically adjusted based on the KL divergence and the water hardness information to adjust the fusion weight of the multi-source features, and target fusion features are obtained.

7. The cooling tower cross-domain fault diagnosis method based on multi-source adversarial domain adaptation according to any one of claims 1-4, characterized in that, The historical sensor time series data of multiple data center cooling towers containing spatial position markers is collected, the source domain and the target domain are divided, and the data is timestamp aligned, sampling frequency unified, and independently normalized, domain-invariant features and label information of the source domain and the target domain are learned and extracted, including: Collecting historical sensor time series data of multiple data center cooling towers, determining a part of the data center cooling tower data of the historical sensor time series data as source domain data; determining another part of the data center cooling tower data of the historical sensor time series data as target domain data; All data are linearly interpolated to unify the time base to complete timestamp alignment; All data are resampled to a unified frequency to align the sampling frequency, and missing values generated during frequency conversion are filled in the resampling process; wherein, in the case that the original sampling frequency is lower than the target frequency, data points are increased by interpolation method; in the case that the original sampling frequency is higher than the target frequency, data points are reduced by downsampling method; Each type of sensor data is independently normalized to eliminate dimensional differences; The timestamp-aligned, uniformly sampled frequency, and independently normalized source domain data and target domain data are input into the adversarial training model; Forward propagation is performed based on the feature extraction network, gradient reversal layer, and domain classifier to obtain domain prediction probability; the domain classifier is a 2-classification network; based on a preset domain classification loss function , a preset fault classification loss function and the domain prediction probability, loss calculation and weighting processing are performed to obtain a total loss; Backpropagation is performed based on the total loss to update the feature extraction network CNN, domain classifier, and fault classifier to extract domain-invariant features and label information, and to minimize domain classification loss to accurately distinguish between source and target domains and minimize fault classification loss to improve fault classification accuracy; The fault classifier includes a fully connected layer and a Softmax activation layer and is used to map aggregated features to fault class probabilities.

8. A cooling tower cross-domain fault diagnosis system based on multi-source adversarial domain adaptation, characterized in that, It comprises: A first acquisition and analysis module for acquiring historical sensor time series data of multiple data center cooling towers containing spatial position markers, dividing source and target domains, and performing timestamp alignment, uniform sampling frequency, and independent normalization on the data to learn and extract domain-invariant features and label information of the source and target domains; A feature extraction and analysis module for extracting data representing thermodynamic and vibration coupling features from the historical sensor time series data to calculate heat transfer efficiency ratio and obtain fan bearing feature frequency band vibration acceleration to jointly construct a coupling feature matrix; A cross-domain time series alignment module for cross-domain time series alignment processing based on the coupling feature matrix and the domain-invariant features through contrastive learning, constructing positive sample pairs combining cooling tower unit area spray quantity differences and filler types, and constructing negative sample pairs from different fault types by combining spray density to perform feature similarity measurement and output aligned feature representations; A second acquisition and analysis module for acquiring circulating water conductivity through a circulating water conductivity sensor to determine water hardness information; A feature fusion module for feature fusion based on the aligned feature representations, combining local temporal patterns and global dependency relationships, dynamically adjusting source domain weights according to KL divergence and the water hardness information, and obtaining target fusion features; A fault diagnosis and prediction module for inputting the target fusion features into the trained fault classifier to output cooling tower fault diagnosis information and trend prediction information.

9. An electronic device, comprising: It comprises: A processor, a memory, and a program stored on the memory and executable on the processor, which, when executed by the processor, implements the multi-source adversarial domain adaptation-based cooling tower cross-domain fault diagnosis method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores programs or instructions, which, when executed by the processor, implement the multi-source adversarial domain adaptation-based cooling tower cross-domain fault diagnosis method of any one of claims 1-7.

Images