A fault diagnosis and self-recovery control method for a power plant pump cluster

Abstract

This disclosure provides a fault diagnosis and self-healing control method for a pump cluster in a thermal power plant, relating to the field of intelligent automatic control technology for thermal power plants. The method includes: acquiring full-band acoustic signals from the pumps using an array of microphones; performing hierarchical decomposition, adaptive filtering, target acoustic band separation, feature extraction and dimensionality reduction, and principal component dimensionality reduction to obtain preprocessed target feature data; fusing the target feature data with pump operating parameters to reconstruct the state space of the target network and optimizing the reward function; outputting the probability distribution of fault types of the pump cluster based on the state space through a decision network in the target network; evaluating the output of the decision network through an evaluation network in the target network; and updating the decision network and evaluation network in conjunction with the reward function; and realizing fault diagnosis and self-healing control of the pump cluster based on the updated target network. This disclosure achieves automatic fault diagnosis and corresponding control in thermal power plants, ensuring the safe and efficient operation of thermal power plants.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent automatic control technology for thermal power plants, specifically to a fault diagnosis and self-healing control method for a thermal power plant pump cluster. Background Technology

[0002] In the production and operation system of a thermal power plant, various pumps, such as circulating water pumps, feedwater pumps, and condensate pumps, form a large pump cluster that performs tasks such as steam and water circulation and energy transmission, supplying steam, heating, and electricity to the power plant. The operational stability of this pump cluster directly affects the power generation and safety of the thermal power plant.

[0003] To monitor faults in pump clusters, existing technologies mostly employ acoustic monitoring. However, current acoustic monitoring technologies suffer from several drawbacks in monitoring and handling faults in thermal power plant pump clusters, including difficulty in separating signal interference, weak collaborative diagnostic capabilities, disconnect between diagnosis and control, and poor adaptability. These limitations fail to meet the demands of thermal power plants for efficient, reliable, and intelligent monitoring and control of pump clusters. Summary of the Invention

[0004] This disclosure addresses the problems existing in the prior art by providing a fault diagnosis and self-healing control method for pump clusters in thermal power plants, thereby solving the problems of signal interference separation, weak collaborative diagnosis capability, disconnect between diagnosis and control, and poor adaptability of existing acoustic monitoring technologies in pump cluster applications.

[0005] To achieve the above objectives, the technical solution adopted in this disclosure is as follows: The first aspect of this disclosure provides a method for fault diagnosis and self-healing control of pump clusters in thermal power plants, including: An array microphone was used to synchronously acquire the full-band acoustic signals of the pump cluster in a thermal power plant, obtaining the raw full-band acoustic data of the pump cluster, which includes multiple pumps operating in parallel. The original full-band acoustic data is preprocessed, which includes hierarchical decomposition, adaptive filtering, target acoustic band separation, feature extraction and dimensionality reduction, and principal component dimensionality reduction to obtain preprocessed target feature data. The target feature data is fused with the pump operating parameters of the pump cluster to construct the state space of the target network; The decision network in the target network outputs the probability distribution of fault types of the pump cluster based on the state space, and the evaluation network in the target network evaluates the output of the decision network. A reward function is set based on the fault identification degree, and the decision network and evaluation network are updated in conjunction with the reward function. Fault diagnosis and self-healing control of the pump cluster are realized based on the updated target network.

[0006] In some embodiments of this disclosure, an array microphone is used to synchronously acquire the full-band acoustic signals of the pump cluster in a thermal power plant, obtaining the raw full-band acoustic data of the pump cluster, including: An array microphone is provided, which includes multiple microphones. The microphones are placed at the bearing end and inlet / outlet end of each pump body in the pump body cluster, and the microphones are attached to the surface of the pump body by magnetic mounting brackets. An array microphone was used to synchronously acquire the full-band acoustic signals of each pump in the pump cluster, thus obtaining the raw full-band acoustic data of the pump cluster.

[0007] In some embodiments of this disclosure, the hierarchical decomposition includes: The original whole-band data is decomposed using wavelet packets with a preset number of layers to obtain multiple corresponding sub-bands, where the preset number of layers includes any number of layers in the range of 3 to 5. Multiple sub-bands are merged into a first band, a second band, and a third band, where the first band includes low-frequency data, the second band includes mid-frequency data, and the third band includes high-frequency data.

[0008] In some embodiments of this disclosure, adaptive filtering includes: Adaptive filtering is performed on each of the multiple sub-bands based on an adaptive filtering threshold. The adaptive filtering threshold is dynamically adjusted based on the root mean square value of the first N sets of historical data for the corresponding sub-band, where N is a positive integer greater than 1. The formula for calculating the adaptive filtering threshold includes: , In the formula, Th M is the adaptive filtering threshold, M is the adjustment coefficient, and N is the number of historical data sets. x i For the first i Feature values ​​of a set of historical data.

[0009] In some embodiments of this disclosure, feature extraction and dimensionality reduction include: Feature extraction and dimensionality reduction are performed on the separated target acoustic bands to obtain a multidimensional feature matrix, wherein the multidimensional feature matrix includes at least one time-domain feature, at least one frequency-domain feature and at least one time-frequency-domain feature. The formulas for calculating the impulse factor in the time-domain features include: , In the formula, I p For pulse factor, x max The peak value of the time-domain waveform of the acoustic signal. L The number of data points in a single group of signals. xj For the first j One time-domain data point.

[0010] In some embodiments of this disclosure, principal component dimensionality reduction includes: Set the cumulative contribution rate of principal components, where the cumulative contribution rate of principal components includes any percentage within the range of 90% to 98%; Based on the cumulative contribution rate of principal components, principal component dimensionality reduction is performed on the multidimensional feature matrix to obtain the target feature data, where the dimension of the target feature data includes any dimension in the range of 3 to 7.

[0011] In some embodiments of this disclosure, a reward function is set based on the fault identification degree, including: The fault type with the highest probability in the fault type probability distribution is selected as the fault judgment result for each pump body. Calculate the cosine similarity between the preprocessed target feature data corresponding to the fault judgment result and the fault feature template corresponding to the fault judgment result; Based on cosine similarity, a reward function is set: when the cosine similarity is greater than or equal to the first threshold, a first positive reward is set; when the cosine similarity is greater than or equal to the second threshold and less than the first threshold, a second positive reward is set; when the cosine similarity is less than the second threshold, a negative penalty is set. The formula for calculating cosine similarity includes: , In the formula, S For cosine similarity, a k This is the k-th element in the fault feature template. b k Let m be the k-th element of the preprocessed target feature data, where m is the dimension of the feature vector.

[0012] In some embodiments of this disclosure, the fault types of the pump cluster include no faults and multiple actual faults; fault diagnosis and self-healing control of the pump cluster are implemented based on the updated target network, including: In the probability distribution of fault types of each pump in the pump cluster, if the probability of no fault is greater than or equal to the first threshold, the pump is judged to be normal and the pump is controlled not to operate. In the probability distribution of fault types of each pump in the pump cluster, if the probability of any actual fault is greater than or equal to the third threshold, it is determined that the pump has the actual fault, and the alarm and pump action are controlled. The third threshold is greater than the second threshold but less than the first threshold.

[0013] In some embodiments of this disclosure, the initial learning rate of the evaluation network is a first learning rate, wherein the evaluation network dynamically adjusts the learning rate by incorporating the real-time change rate of the full-band features: when the real-time change rate of the full-band features is less than or equal to a preset threshold, the learning rate is maintained at the first learning rate; when the real-time change rate of the full-band features is greater than the preset threshold, the learning rate is adjusted to a second learning rate; wherein the second learning rate is greater than the first learning rate. The formula for calculating the real-time rate of change of characteristics in the whole sound band includes: , In the formula, Δ A For the real-time rate of change of characteristics across the entire acoustic band, A t Let A be the characteristic amplitude at the current moment. t-1 Δt represents the characteristic amplitude at the previous moment, and Δt represents the time interval.

[0014] In some embodiments of this disclosure, preprocessing employs predetermined preprocessing parameters, including the number of wavelet packet decomposition layers, principal component dimensionality reduction dimension, and adaptive filtering threshold. The preprocessing parameters are determined based on a pre-stored preprocessing parameter mapping library, which includes the mapping relationship between the equipment type of the pump cluster and the preprocessing parameters. When the equipment type of the pump cluster does not exist in the preprocessing parameter mapping library, the preprocessing parameters are fine-tuned based on the pre-acquired sample data of the pump cluster through network search, and the equipment type of the pump cluster and the corresponding preprocessing parameters are stored in the preprocessing parameter mapping library. The objective function for fine-tuning includes: , In the formula, Acc The fault diagnosis accuracy is represented by TP, which is the number of true positive samples (the number of samples that are actually faulty and diagnosed as faulty), TN, which is the number of true negative samples (the number of samples that are actually normal and diagnosed as normal), FP, which is the number of false positive samples (the number of samples that are actually normal but diagnosed as faulty), and FN, which is the number of false negative samples (the number of samples that are actually faulty but diagnosed as normal).

[0015] In some embodiments of this disclosure, before constructing the state space of the target network, the method further includes: The target feature data and pump operating parameters are normalized. The formulas for normalization include: , In the formula, x norm These are the normalized parameter values. x These are the original parameter values. x min , x maxThese are the historical minimum and maximum values ​​of the parameter, respectively.

[0016] Compared with the prior art, this disclosure has the following beneficial effects: The fault diagnosis and self-healing control method for power plant pump clusters provided in this embodiment of the invention synchronously acquires full-band acoustic signals from multiple parallel pumps using an array of microphones. It then combines a preprocessing process involving hierarchical decomposition, adaptive filtering, target acoustic band separation, multi-dimensional feature extraction, and principal component dimensionality reduction. The purified target feature data is then deeply integrated with pump operating parameters to construct a state space. Relying on the division of labor and cooperation between the decision network and the evaluation network in the target network, and in conjunction with a reward function based on fault identification, the two networks are linked for updating. Finally, the updated target network completes fault diagnosis and self-healing control, achieving synergistic enhancement of multi-dimensional technical features. This method ensures the integrity and accuracy of fault characteristics through full-band acoustic acquisition and refined preprocessing, enhances the network's adaptability to complex operating conditions by fusing features with operating parameters, and strengthens the real-time performance and reliability of fault diagnosis through a dual-network linkage update mechanism. Ultimately, it achieves accurate identification, rapid response, and autonomous control of pump cluster faults, effectively avoiding delays caused by manual intervention, preventing the spread of single pump faults to the entire cluster, significantly reducing the risk of unplanned shutdowns in thermal power plants, and ensuring the safe, efficient, and stable operation of the pump cluster and the entire power generation system. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a fault diagnosis and self-healing control method for a pump cluster in a thermal power plant, provided according to an embodiment of this disclosure. Figure 2 This is a flowchart of a DDPG network fault diagnosis and self-healing control provided in an embodiment of this disclosure. Detailed Implementation

[0018] The present disclosure will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and should not be construed as limiting the scope of protection of the present disclosure. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application.

[0019] The acquisition, transmission, storage, use, and processing of data in this disclosed technical solution comply with relevant national laws and regulations. In the embodiments of this disclosure, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this disclosure, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] All terms used in this disclosure have the same meaning as understood by one of ordinary skill in the art to which this disclosure pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein.

[0021] In the production and operation system of a thermal power plant, a large pump cluster, including circulating water pumps, feedwater pumps, and condensate pumps, performs tasks such as steam and water circulation and energy transmission, supplying steam, heating, and electricity to the plant. Therefore, the stability of pump operation directly affects the power generation and safety of the thermal power plant. Because thermal power plants need to supply electricity and heat externally, they generally operate under high load and full capacity for extended periods. Simultaneously, the working environment of the pump cluster is extremely complex, involving multiple operating conditions, such as water, steam, and various chemical agents. Due to varying grid loads, there are variable operating conditions. This means that pump operation is exceptionally complex, making it highly susceptible to issues such as bearing wear, impeller cavitation, and seal damage. If such faults are not detected and addressed promptly and accurately: at best, pump efficiency decreases and energy consumption increases; at worst, pumps stop operating, or even the entire main steam and water system of the power plant shuts down, causing unplanned unit shutdowns, resulting in huge losses, wasted energy, and a serious threat to production safety.

[0022] To monitor faults in pump clusters, existing technologies primarily employ acoustic monitoring. The general workflow is as follows: First, acoustic sensors are deployed at key locations on the pump body to collect acoustic signals at a sampling rate of 20 kHz or higher. Simultaneously, ambient noise is collected as a baseline, and a pipeline pass-frequency noise level correlator is used to achieve preliminary monitoring of the pump's operating status. Next, the collected raw acoustic signals undergo preprocessing operations such as noise reduction and frequency division. Then, a Fourier transform is used to convert the time-domain signal to a frequency-domain signal, from which fault features are extracted. Afterward, the extracted features are compared with fault feature templates in the normal acoustic band. Indicators such as signal amplitude and frequency changes are used to determine the fault type and locate the fault point. Finally, when the feature value exceeds a threshold, an alarm is triggered, displaying the fault location and type, allowing subsequent maintenance personnel to intervene.

[0023] However, due to numerous sources of acoustic interference within thermal power plants—including pump noise generated by the plant itself, as well as noise from the generator units and fans where the pumps are located, and issues such as pipe vibration and resonance between layers or shells—all these factors significantly impact the final detection results. It is difficult to filter out the influence of environmental noise within the power plant on the fault acoustic frequency band. When using conventional filtering methods such as low-pass, high-pass, and band-pass filters to remove fault and interference signals, it is difficult to completely separate the spectrum of fault and interference signals, easily leading to false alarms or missed fault detections. For example, the acoustic signal generated by minor wear on the pump bearing is easily masked by high-frequency noise from nearby operating fans. In this case, the sensor cannot obtain the acoustic signal characterizing the bearing wear state, resulting in missed fault detection. Alternatively, occasional sudden noise signals from equipment in the environment can be mistaken for pump fault signals, causing unwarranted equipment shutdowns for inspection.

[0024] From the perspective of the connection between collaborative diagnosis and control, the existing technology can only diagnose and alarm pump body faults, but cannot automatically interpret the results to form solutions for regulation. This is because the pump body is a clustered working system. During operation, it does not work alone. Once one piece of equipment fails, its signal will cause all other equipment to fail together. However, manual judgment and analysis of signals, solution handling and execution control all have a certain delay. This determines that this working mode ultimately cannot achieve rapid, intelligent and automatic response and handling of equipment faults, and cannot fully realize the full-process application requirements of "monitoring-repair-stabilization" for the clustered operation of pump bodies. For example, if a feed pump of a certain unit has an impeller cavitation problem, if it is to be completely repaired with the existing technology, it is necessary to: (1) detect the problem and alarm; (2) manually judge the degree of impeller cavitation; (3) choose whether to increase or decrease the pump speed, increase or decrease the feed water pressure, or even directly shut down for maintenance. Only after completing the above work can the cavitation problem be guaranteed to be solved, but this process takes a long time and has a significant impact on the production schedule. Furthermore, due to the long repair time, impeller cavitation may become severe during the repair process, rendering the established repair plan insufficient to repair the equipment.

[0025] From the perspective of sample dependence and adaptability, current technical diagnostics rely on a complete sample library. Since power plant pumps come in different types, each with varying structures, dimensions, and operating parameters, the acoustic characteristics of different faults also differ. Therefore, if conditions such as start-stop transitions are not fully considered in the sample library, this method cannot accurately determine the occurrence of the fault. Furthermore, for different types of pumps, a series of processes involving sample collection, feature extraction, and template creation must be repeated, inevitably consuming manpower, resources, and time. This also hinders the rapid application of this technology to other new types of pumps, making it difficult to achieve efficient and unified monitoring and management of all pump clusters. For example, using existing acoustic monitoring technology requires collecting a large number of acoustic samples under various operating conditions of the condensate pump under normal and fault states, and then constructing corresponding fault feature templates based on these samples. This process is extremely time-consuming, sometimes taking weeks or even months, severely impacting the practical applicability and economic viability of the technology.

[0026] In summary, existing acoustic monitoring technologies for monitoring and handling pump cluster faults in thermal power plants suffer from several problems, including difficulty in separating signal interference, weak collaborative diagnostic capabilities, disconnect between diagnosis and control, and poor adaptability. These issues severely limit their effectiveness in the entire process of "monitoring-repair-stabilization" of pump clusters and fail to meet the needs of thermal power plants for efficient, reliable, and intelligent monitoring and control of pump clusters. Therefore, breakthroughs in new technologies are urgently needed to achieve intelligent self-healing and stable operation of pump clusters, ensuring the safe and efficient power generation and heating supply of thermal power plants.

[0027] Example 1; Figure 1 This is a flowchart illustrating a fault diagnosis and self-healing control method for a pump cluster in a thermal power plant, as provided in an embodiment of this disclosure. Figure 1 As shown, the specific steps include S11 to S16.

[0028] Step S11: Using an array microphone, synchronously acquire the full-band acoustic signal of the pump cluster of the thermal power plant to obtain the original full-band acoustic data of the pump cluster, wherein the pump cluster includes multiple pumps operating in parallel.

[0029] In some embodiments of this disclosure, an array microphone is used to synchronously acquire the full-band acoustic signals of a power plant's pump cluster, obtaining the raw full-band acoustic data of the pump cluster. Specifically, this may include: setting up an array microphone, wherein the array microphone includes multiple microphones, the microphones are placed at the bearing ends and inlet / outlet ends of each pump in the pump cluster, and the microphones are attached to the pump surface via magnetic mounting brackets; using the array microphone to synchronously acquire the full-band acoustic signals of each pump in the pump cluster, obtaining the raw full-band acoustic data of the pump cluster. It is understood that the microphones may also be placed in other key parts of the pump, or may be installed on the pump surface using other structures; this disclosure does not specifically limit this approach.

[0030] For example, an array of microphones is used to build a distributed acquisition network. Two microphones are configured for each pump in the pump cluster that operates in parallel. The microphones are tightly attached to the bearing end of the pump by magnetic mounting brackets to ensure that the acquisition signal-to-noise ratio is ≥35dB. All microphones work synchronously at a sampling rate of 48 kHz, generating a set of raw full-band data containing 4800 data points every 100 ms.

[0031] Step S12: Preprocess the original full-band acoustic data, which includes hierarchical decomposition, adaptive filtering, target acoustic band separation, feature extraction and dimensionality reduction, and principal component dimensionality reduction to obtain the preprocessed target feature data.

[0032] In some embodiments of this disclosure, the hierarchical decomposition may specifically include: decomposing the original whole-band data using wavelet packets of a preset number of layers to obtain multiple corresponding sub-bands, wherein the preset number of layers includes any number of layers in the range of 3 to 5; merging the multiple sub-bands into a first frequency band, a second frequency band, and a third frequency band, wherein the first frequency band includes low-frequency data, the second frequency band includes mid-frequency data, and the third frequency band includes high-frequency data.

[0033] For example, the hierarchical decomposition step uses db4 wavelet packets to decompose the original full-band acoustic data of each pump into four layers, dividing the complete frequency band into 16 sub-bands, and then merging them according to frequency range into low-frequency bands, such as 20~500 Hz, which mainly contain workshop background noise; mid-frequency bands, such as 500 Hz~5 kHz, which mainly contain signals such as gearbox and pump housing vibration; and high-frequency bands, such as 5~20 kHz, which mainly contain fault signals such as bearing wear and seal leakage.

[0034] In some embodiments of this disclosure, adaptive filtering may specifically include: performing adaptive filtering on each of the multiple sub-bands based on an adaptive filtering threshold, wherein the adaptive filtering threshold is dynamically adjusted based on the root mean square value of the first N sets of historical data of the corresponding sub-band, where N is a positive integer greater than 1.

[0035] For example, the root mean square value of the first 10 sets of historical data for the corresponding sub-band is dynamically adjusted to eliminate background noise in the workshop concentrated in the low frequency band of 100~300 Hz.

[0036] The formula for calculating the adaptive filtering threshold includes: , In the formula, Th The adaptive filtering threshold is defined by M, which is an adjustment coefficient ranging from 1.2 to 1.5, and N is the number of historical data sets. x i For the first i Feature values ​​of a set of historical data.

[0037] In some embodiments of this disclosure, target acoustic band separation may specifically include: target frequency band enhancement separation of the filtered signal based on the equipment type and common fault characteristics of the pump cluster.

[0038] For example, when monitoring bearing faults, the focus is on high-frequency signals in the 5~20 kHz range; when monitoring gearbox and impeller faults, the focus is on mid-frequency signals in the 500 Hz~5 kHz range; when monitoring issues such as loose bases or pipe resonance, the focus is on retaining effective signals in the 20~500 Hz low-frequency range, thus achieving precise focusing of fault signals.

[0039] In some embodiments of this disclosure, feature extraction and dimensionality reduction may specifically include: performing feature extraction and dimensionality reduction on the separated target acoustic bands to obtain a multidimensional feature matrix, wherein the multidimensional feature matrix includes at least one-dimensional time-domain feature, at least one-dimensional frequency-domain feature and at least one-dimensional time-frequency-domain feature.

[0040] For example, multi-dimensional feature extraction is performed on the separated target acoustic band signal to construct an 18-dimensional feature matrix, which specifically includes 8-dimensional time-domain features, 6-dimensional frequency-domain features, and 4-dimensional time-frequency-domain features.

[0041] The formulas for calculating the impulse factor in the time-domain features include: , In the formula, I p For pulse factor, x max The peak value of the time-domain waveform of the acoustic signal. L The number of data points in a single group of signals. x j For the first j One time-domain data point.

[0042] In some embodiments of this disclosure, principal component dimensionality reduction may specifically include: setting a cumulative contribution rate of principal components, wherein the cumulative contribution rate of principal components includes any percentage in the range of 90% to 98%; and performing principal component dimensionality reduction on the multidimensional feature matrix based on the cumulative contribution rate of principal components to obtain target feature data, wherein the dimension of the target feature data includes any dimension in the range of 3 to 7.

[0043] This disclosure provides a full-band signal preprocessing technique to ensure high-quality state input to the target network, and in one possible implementation, it includes the following steps.

[0044] Noise reduction: Wavelet thresholding noise reduction technology is adopted. By utilizing the multi-resolution analysis capability of wavelet transform, the original acoustic signal is decomposed into wavelet coefficients of different scales. An improved soft threshold function is used to process the high-frequency coefficients. When the absolute value of the coefficient is less than the threshold, the high-frequency wavelet coefficient is directly assigned a value of 0 to suppress the influence of noise on the system. When the absolute value of the coefficient is greater than the threshold, the effective signal components are retained through "coefficient shrinkage". The signal is then reconstructed through inverse wavelet transform to initially filter random noise.

[0045] Furthermore, considering the regular and periodic noise interference, an adaptive filter is constructed using the "cluster-wide ambient sound signal" as input. Under the minimum mean square error criterion, the filter coefficients are continuously adjusted according to the original signal to cancel out the regular interference terms in the original signal.

[0046] Time-frequency domain transformation: A longer time signal is divided into a series of short-time signal segments by using a sliding time window. Then, a Short-Time Fourier Transform (STFT) is performed on each segment to obtain the different frequency components at each instant, thus representing it as a two-dimensional time-frequency graph. This method allows us to observe the characteristic frequencies of a fault and the changes in energy during operation, thereby determining which stage of the fault development is currently in.

[0047] Furthermore, considering that traditional STFTs cannot meet the required frequency resolution in the low-frequency band, wavelet packet transform is used to divide the entire frequency band into multiple frequency band sub-intervals, and these sub-intervals are further subdivided. By calculating the proportion of energy in each sub-band, faults are classified and located, making it easier to distinguish faults in overlapping areas between frequency bands and preventing fault confusion.

[0048] Feature engineering: Find the peak point in the signal obtained after time-frequency domain processing, and take the peak value after quantization; take the root mean square of the signal as the output result after quantization; divide the peak value by the root mean square to obtain the pulse exponent; define the spectral entropy, i.e., the spectrum entropy, based on the signal energy entropy.

[0049] The quantized data values ​​are processed in the form of target network state input.

[0050] In some embodiments of this disclosure, preprocessing employs predetermined preprocessing parameters, including the number of wavelet packet decomposition layers, principal component dimensionality reduction dimension, and adaptive filtering threshold. The preprocessing parameters are determined based on a pre-stored preprocessing parameter mapping library, which includes a mapping relationship between the equipment type of the pump cluster and the preprocessing parameters.

[0051] When the equipment type of the pump cluster does not exist in the preprocessing parameter mapping library, the preprocessing parameters are fine-tuned based on the pre-acquired sample data of the pump cluster through a network search method, and the equipment type of the pump cluster and the corresponding preprocessing parameters are stored in the preprocessing parameter mapping library.

[0052] The objective function for fine-tuning includes: , In the formula, Acc The fault diagnosis accuracy is represented by TP, which is the number of true positive samples (the number of samples that are actually faulty and diagnosed as faulty), TN, which is the number of true negative samples (the number of samples that are actually normal and diagnosed as normal), FP, which is the number of false positive samples (the number of samples that are actually normal but diagnosed as faulty), and FN, which is the number of false negative samples (the number of samples that are actually faulty but diagnosed as normal).

[0053] Step S13: Integrate the target feature data with the pump operating parameters of the pump cluster to construct the state space of the target network.

[0054] Real-time operating parameters of each pump in the pump cluster are collected, including key operating parameters such as pump speed, inlet and outlet pressure, motor operating temperature, and medium flow rate, and then fused with the pre-processed target feature data.

[0055] In this embodiment of the disclosure, the target network can be a DDPG (Deep Deterministic Policy Gradient) network. In this embodiment of the disclosure, the target network can also be a PPO (Proximal Policy Optimization) network, a SAC (Soft Actor-Critic) network, a TD3 (Twin Delayed Deep Deterministic Policy Gradient) network, or other deep learning networks; this disclosure does not specifically limit its use.

[0056] In some embodiments of this disclosure, before fusion, the method further includes: normalizing the target feature data and pump operating parameters and mapping them to the [0,1] interval.

[0057] The formulas for normalization include: , In the formula, x norm These are the normalized parameter values. x These are the original parameter values. x min , x max These are the historical minimum and maximum values ​​of the parameter, respectively.

[0058] By eliminating the influence of dimensional differences on network training through normalization, the state space of the target network with 7 to 9 dimensions is finally constructed, such as 5-dimensional reduced features + rotational speed + inlet and outlet pressure + motor temperature, so that the network can learn the correlation between "acoustic band features and operating conditions" at the same time.

[0059] Step S14: The decision network in the target network outputs the failure type probability distribution of the pump cluster according to the state space, and the evaluation network in the target network evaluates the output of the decision network.

[0060] In some embodiments, the decision network can be an Actor network with a fully connected three-layer structure. The number of nodes in the input layer is equal to the dimension of the state space. The hidden layer has 64 nodes activated using the ReLU activation function, and the output layer has 5 nodes, corresponding to five types of faults: bearing outer ring wear, bearing inner ring crack, gear tooth breakage, seal leakage, and no fault. The output uses the Softmax activation function to output the probability distribution of each fault type. The evaluation network can be a Critic network, also a three-layer fully connected structure, but its input layer consists of "state space + fault probability distribution output by the Actor network," totaling 13 dimensions. It has 64 hidden layer nodes and a single output node representing the value of an action, or the action value of a certain state—that is, the value used to evaluate the output of the decision network.

[0061] In the embodiments disclosed herein, the decision network may also employ CNN (Convolutional Neural Network), LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), etc., and the evaluation network may also employ MLP (Multi-Layer Perceptron), GNN (Graph Neural Network), Transformer, etc., without specific limitations.

[0062] In some embodiments of this disclosure, the initial learning rate of the evaluation network is set to a first learning rate, wherein the evaluation network dynamically adjusts the learning rate by incorporating the real-time change rate of the whole-band features. Specifically, this may include: maintaining the learning rate as the first learning rate when the real-time change rate of the whole-band features is less than or equal to a preset threshold; and adjusting the learning rate to a second learning rate when the real-time change rate of the whole-band features is greater than the preset threshold, wherein the second learning rate is greater than the first learning rate.

[0063] For example, when A > 5%, i.e., in the case of sudden failure, the evaluation network increases the learning rate to 0.005, accelerates the iteration speed of the dual network parameters, and enables rapid response to sudden failures.

[0064] The formula for calculating the real-time rate of change of characteristics in the whole sound band includes: , In the formula, Δ A For the real-time rate of change of characteristics across the entire acoustic band, A t Let A be the characteristic amplitude at the current moment. t-1 Δt represents the characteristic amplitude at the previous moment, and Δt represents the time interval.

[0065] Step S15: Set a reward function based on the fault identification degree, and update the decision network and evaluation network in conjunction with the reward function.

[0066] In some embodiments of this disclosure, a reward function is set based on the fault identification degree. Specifically, this may include: selecting the fault type with the highest probability in the fault type probability distribution as the fault judgment result for each pump; calculating the cosine similarity between the preprocessed target feature data corresponding to the fault judgment result and the fault feature template corresponding to the fault judgment result; and setting a reward function based on the cosine similarity. Specifically, this may include: setting a first positive reward when the cosine similarity is greater than or equal to a first threshold; setting a second positive reward when the cosine similarity is greater than or equal to a second threshold and less than the first threshold; and setting a negative penalty when the cosine similarity is less than the second threshold.

[0067] For example, when S≥90%, a reward of R=5 is given to positively incentivize the network to make accurate decisions; when 70%≤S<90%, a reward of R=1 is given to weakly incentivize the network to optimize; when S<70% or there are faults, misjudgments, or omissions, a penalty of R=-3 is given to guide the network to correct errors.

[0068] The formula for calculating cosine similarity includes: , In the formula, S For cosine similarity, a k This is the k-th element in the fault feature template. b k Let m be the k-th element of the preprocessed target feature data, where m is the dimension of the feature vector.

[0069] Dual-network linkage update: A reward value is generated every 100 ms and fed back to the target network. The evaluation network updates the value evaluation parameters based on the reward value and decision results, and at the same time, the evaluation error is back-propagated to the decision network to adjust the weight coefficients of the decision network. At the same time, the real-time change rate A of the full-band features can be introduced. When A>5%, i.e., a sudden failure scenario, the evaluation network increases the learning rate to 0.005, which accelerates the iteration speed of the dual network parameters and realizes rapid response to sudden failures.

[0070] Step S16: Implement fault diagnosis and self-healing control of the pump cluster based on the updated target network.

[0071] In some embodiments of this disclosure, the fault types of the pump cluster may include no fault and multiple actual faults; the fault diagnosis and self-healing control of the pump cluster based on the updated target network may specifically include: in the fault type probability distribution of each pump in the pump cluster, if the probability of no fault is greater than or equal to a first threshold, then the pump is determined to be normal and the pump is controlled not to operate; in the fault type probability distribution of each pump in the pump cluster, if the probability of any actual fault is greater than or equal to a third threshold, then the pump is determined to have the actual fault, and an alarm and pump operation are controlled; wherein the third threshold is greater than the second threshold and less than the first threshold.

[0072] For example, based on the updated target network, the state space data of the pump cluster is received in real time, and the decision network outputs the probability distribution of the fault type for each pump. If the probability of "no fault" is ≥90%, the pump is judged to be operating normally; if the probability of a certain fault type is ≥75%, the pump is accurately diagnosed to have a corresponding fault, and the severity of the fault is identified at the same time. The higher the probability, the more serious the fault.

[0073] Furthermore, based on the diagnostic results, the system can automatically generate self-healing control schemes. For example, for minor faults, corresponding to a fault probability of 75%~85%, the system can mitigate the worsening of the fault by adjusting parameters such as pump speed, inlet and outlet valve opening, and medium flow rate. For moderate faults, corresponding to a fault probability of 85%~95%, the system can initiate cluster redundancy scheduling to distribute the load of the faulty pump to other normal pumps, while triggering an early warning to prompt maintenance personnel to plan maintenance. For severe faults, corresponding to a fault probability ≥95%, the system can immediately shut down the operation of the faulty pump to prevent the fault from spreading to the entire cluster and ensure the overall stable operation of the pump cluster. It should be noted that all diagnostic results, control commands, and cluster operating status are displayed in real time through an industrial touch screen, and the audible and visual alarm module triggers corresponding alarms according to the fault level.

[0074] Example 2; Figure 2 This is a flowchart of a DDPG network fault diagnosis and self-healing control provided in an embodiment of this disclosure.

[0075] like Figure 2 As shown, the specific steps include the following: Step S21: Integrate target feature data with pump operating parameters, and normalize and reconstruct the state range; Step S22: The Actor network outputs the probability distribution of fault types; Step S23: Calculate cosine similarity and design reward function; Step S24: The Critic network evaluates the decision value and calculates the real-time change rate of the full-band characteristics; Step S25: Determine if the rate of change is greater than or equal to 5% per second; if yes, proceed to step S26: Increase the learning rate; otherwise, proceed to step S27. Step S27: Determine the probability of failure; if the probability of no failure is greater than or equal to 90%, the pump body is determined to be normal, and step S28: Normal, no action is taken; if the probability of any actual failure is greater than or equal to 75%, the pump body is determined to send the actual failure, and step S29: Alarm, action is taken.

[0076] For a detailed description of the above steps, please refer to the step description in Example 1, which will not be repeated here.

[0077] This embodiment uses information obtained from full acoustic preprocessing to design a Deep Deterministic Policy Gradient (DDPG) network.

[0078] The core of this embodiment is to deeply integrate full-band acoustic signal preprocessing technology with Deep Deterministic Policy Gradient (DDPG) network. This solves the problems of high noise interference, low feature extraction efficiency, and insufficient fault diagnosis accuracy that exist when the traditional DDPG network directly processes raw acoustic band data. By providing high-quality input features for the DDPG network through band preprocessing, accurate identification and decision-making of equipment faults can be achieved.

[0079] In addition, a preprocessing feature and DDPG network coupling input mechanism were introduced. First, a dimensionality reduction algorithm was used to compress the preprocessing features of the entire acoustic band and combine them with equipment operating parameters, such as speed, load, and temperature, to form the state space of the DDPG network. This allows the network to learn the relationship between "acoustic band features and operating conditions" at the same time, avoiding the diagnostic bias that can easily occur when only a single feature is used.

[0080] Furthermore, by using full-band acoustic features as the reward function for fault identification, a reward is given when the fault diagnosis result output by the network matches the acoustic band features corresponding to the preprocessed fault category of the target with a degree of ≥90%; conversely, if the degree of matching is <70% or the fault is incorrectly identified as another fault category, a penalty is given, thereby enabling the network to better learn the correspondence between "acoustic band features and fault type".

[0081] Furthermore, based on the superior stability of the full-band preprocessing features, the multi-classification structure of the Actor network's output layer is optimized during the decision-making and iteration process of the DDPG algorithm. This allows the output values ​​to map several common equipment fault types, such as cracks, wear, and loosening. After learning, the Actor network can obtain preprocessed acoustic band features. For different equipment, different fault classifications and different categories within different ranges need to be set, and the probability value of belonging to the category is output. The category with the highest probability value is taken as the final diagnostic result. The change rate of the full-band preprocessing features, such as the instantaneous increase of the full-band feature values, is used to evaluate the Actor decision value. If the change rate of a certain band feature value is greater than a preset threshold, the Critic network parameters are triggered for emergency updates to cope with sudden faults.

[0082] Example 3; Signal Acquisition: The full-band acoustic pump cluster inspection equipment mentioned in this embodiment includes a full-band acoustic signal acquisition unit, a layered preprocessing unit, a DDPG fault diagnosis unit, a cross-scenario adaptation unit, and a result output unit. All units use industrial Ethernet for data communication, which can simultaneously perform online cluster detection on centrifugal pumps, diaphragm pumps, and other pumps of the same or different models. The overall response of the equipment is ≤0.5 s, which can achieve the purpose of dynamic monitoring of pump clusters at the workshop level.

[0083] The full-band acoustic signal acquisition unit uses an array microphone with a sampling rate of 48 kHz and a frequency response range of 20 Hz to 20 kHz. Each pump has two microphones, placed at the bearing end and the inlet / outlet end of the pump body, respectively, to collect friction sound, vibration sound, and fluid noise. The microphone mounting base has a magnetic structure that fits tightly against the pump body surface, ensuring that the signal-to-noise ratio of the acoustic signal acquisition is ≥35 dB. The result output unit includes an industrial touch screen and an acoustic-optical alarm module. The industrial touch screen is used to display the fault type, confidence level, and cluster health status heat map of a single pump body. The acoustic-optical alarm module is used for buzzer and red light flashing alarms when a fault occurs, and the alarm threshold can be set.

[0084] Furthermore, during the signal acquisition phase, after the device is powered on, the cross-scenario adaptation unit first reads the initial parameters corresponding to the pump body in the "Device Type - Band Preprocessing Parameters" mapping library, such as the high-frequency fault signal threshold of the centrifugal pump being 0.8 V and the mid-frequency filtering window size being 256 points, and sends them to the full-band signal acquisition unit and the layered preprocessing unit.

[0085] Furthermore, the acquisition unit collects acoustic signals from the bearing end and inlet / outlet end of each pump body at a sampling rate of 48 kHz, and generates a complete set of raw full-band acoustic data within 100 ms. This set of data has a length of 4800 points and is transmitted to the layered preprocessing unit via industrial Ethernet.

[0086] Further, in the layered preprocessing stage: In order to accurately determine the true state of the signal, the signal needs to be processed in layers beforehand. The layered preprocessing unit uses db4 wavelet packets to decompose the full-band data into 4 layers, resulting in 16 sub-bands, which are then combined according to frequency into a low-frequency band of 20~500 Hz, a mid-frequency band of 500 Hz~5 kHz, and a high-frequency band of 5~20 kHz.

[0087] In addition, each sub-band utilizes adaptive threshold filtering, where the adaptive filtering threshold is set based on the root mean square value of the first 10 historical data sets of the sub-band to eliminate background noise in the workshop. This reduces unnecessary interference and ensures that the system can accurately acquire the signal that needs to be extracted.

[0088] Furthermore, the target acoustic bands can be distinguished based on the characteristics of pump failures: for example, when monitoring a centrifugal pump bearing failure, signals with a frequency band of 5–20 kHz should be highlighted; when monitoring a pump gearbox failure, signals with a frequency band of 500 Hz–5 kHz should be highlighted.

[0089] Furthermore, to enhance data representation capabilities and simplify the data structure, feature enhancement and dimensionality reduction are required for the separated target acoustic bands. This involves extracting 8-dimensional time-domain features, such as peak value, root mean square, peak-to-peak value, kurtosis, skewness, waveform factor, impulse factor, and margin factor; 6-dimensional frequency-domain features, such as peak power spectral density, center frequency, mean square frequency, frequency variance, spectral kurtosis, and spectral entropy; and 4-dimensional time-frequency-domain features, such as wavelet entropy, wavelet energy, wavelet variance, and wavelet correlation, forming an 18-dimensional high-dimensional feature matrix. Furthermore, the dimensionality of the 18-dimensional feature matrix is ​​reduced to 4-6 dimensions through Principal Component Analysis (PCA). For example, centrifugal pump bearing faults are compressed to 5 dimensions, and gearbox faults are compressed to 4 dimensions, while ensuring that the cumulative contribution rate of the principal components is ≥95%. Then, the dimensionality-reduced feature information is input into the DDPG fault diagnosis unit.

[0090] After preprocessing the full acoustic band signal, the processed data needs to be deeply coupled. During the deep coupling process, the state space needs to be reconstructed and the reward function needs to be optimized.

[0091] Furthermore, state space reconstruction: The DDPG fault diagnosis unit performs data fusion on the dimensionality-reduced features after hierarchical preprocessing and pump operating parameters, such as speed, inlet and outlet pressure, and motor temperature, to obtain a 7-9 dimensional state space. For example, the state space for centrifugal pump bearing fault monitoring can be represented as: [5-dimensional high-frequency band dimensionality-reduced features + speed + inlet and outlet pressure + motor temperature], totaling 8 dimensions. All parameters are first normalized (mapped to the range of [0,1]) before data fusion to prevent errors caused by different dimensions of the parameters from affecting the training accuracy of the network.

[0092] Furthermore, the reward function is optimized: the reward function R for fault identification is designed, and its specific calculation process is based on the characteristics of the entire acoustic band.

[0093] Assuming the output of the DDPG network is Y, which represents "bearing outer ring wear", "bearing inner ring crack" or "no fault", the preprocessed target acoustic band fault feature template T is obtained by training 1000 sets of pump body fault samples, with one fault feature template for each type of fault.

[0094] In addition, the cosine similarity S can be calculated based on the features corresponding to Y and T. If S≥90%, a reward R=5 is set, which is a positive reward to encourage the network to make such a choice. If 70≤S<90%, then R=1, which is a weak positive reward to motivate the network to learn this skill. If S<70%, or Y is “fault-free” but a fault actually occurs, then R=-3, which is a negative penalty that can make the network correct its wrong behavior.

[0095] Furthermore, the formula for S is as follows: , That is, the fault feature vector output by the DDPG network and the preprocessed target acoustic band fault feature template vector are respectively represented as: , ; This indicates its modulus length.

[0096] Furthermore, during the parameter updates in this process, a reward value is generated every 100 ms and then transmitted to the DDPG network in real time.

[0097] Based on the understanding of the core mechanism of DDPG network fault diagnosis above, this embodiment deploys the DDPG network in a real industrial application scenario and makes improvements to its working efficiency.

[0098] First, the faults in the Actor network are categorized: The Actor network adopts a 3-layer fully connected structure. The input layer has n nodes, where n equals the dimension of the state space. The hidden layer has 64 nodes, with ReLU as the activation function. The number of nodes in the output layer is the number of common fault types in the pump body. In this embodiment, it is set to 5 categories: bearing outer ring wear, bearing inner ring crack, gear tooth breakage, seal leakage, and no fault. The output layer uses the Softmax activation function to output the probability distribution of each fault type.

[0099] Furthermore, the activation functions ReLU and Softmax are as follows: Activation function ReLU: , where x is the input value.

[0100] Activation function Softmax: , C represents the number of fault categories, and zi is the raw output score of the last layer of the Actor network.

[0101] Furthermore, the network selects the fault type with the highest probability as the final diagnostic result. If the probability of no fault is ≥90% among all types, the pump body is judged to be normal; if the probability of a certain fault type is ≥75%, the pump body is judged to have that type of fault, and the diagnostic result (including fault type and confidence level) is transmitted to the result output unit.

[0102] Continue updating the Critic network using linkage: Similar to the Actor network, this network also has a three-layer fully connected structure, but the input layer is "state space + fault probability distribution of the Actor network output", totaling 13 dimensions; the number of hidden layer nodes in this network is 64, while the output layer has 1 node, representing the value of the action, or the action value of a certain state.

[0103] Meanwhile, a real-time rate of change across the entire acoustic band is introduced, such as taking the instantaneous increment ΔA of the characteristic amplitude of the high-frequency band. When ΔA > a preset threshold (set to 5% per second in this embodiment) – a typical sudden fault (such as a bearing jamming causing a sudden increase in high-frequency friction noise) – an emergency update of the Critic network is initiated, and the learning rate is adjusted to 0.005 to accelerate the parameter iteration speed, achieving network update within 300 ms and enabling rapid response to sudden faults. When ΔA is less than or equal to 5% per second, the Critic network is maintained to learn and update at a normal speed to ensure network stability.

[0104] Furthermore, this disclosure only applies to experimental testing of a limited number of pump body clusters. When the equipment is first applied to the testing of a new type of pump body cluster, some parameters of the equipment must be adjusted first. This is its parameter adaptive adjustment process.

[0105] The first step is to manually collect 50 sets of full-band acoustic data for this model of diaphragm pump, specifically including 20 sets of normal data and 30 sets of fault data, and upload them to the cross-scenario adaptation unit. In the second step, a cross-scenario adaptation unit with the highest fault diagnosis accuracy is set up. For wavelet packet decomposition layers (3~5 layers), PCA dimensionality reduction dimensions (3~7 dimensions), and threshold filtering threshold (0.6~1.0 V), the optimal solution for different parameter combinations is selected using the grid search method. After layered preprocessing and training in the DDPG network, the parameter combinations are obtained. Then, 10 rounds of testing are performed and the average value is calculated to obtain the corresponding average diagnostic accuracy.

[0106] The third step is to select the parameter combination with the highest average diagnostic accuracy as a reference for adapting the new pump cluster, and enter it into the "Equipment Type - Band Preprocessing Parameters" mapping library; in the future, when testing this type of diaphragm pump cluster, the updated parameters can be directly called without manual re-adjustment, and it can be adapted in less than 1 hour, which greatly improves efficiency and saves a lot of time compared with the traditional method (8~12 hours).

[0107] It should be noted that the terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different parts. Terms such as "including" or "contains" mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility of covering other elements as well.

[0108] Although operations are described in a specific order in the accompanying drawings in this disclosure, it should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the shown operations to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0109] Finally, it should be noted that the above content is only used to illustrate the technical solution of this disclosure, and is not intended to limit the scope of protection of this disclosure. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this disclosure do not depart from the substance and scope of the technical solution of this disclosure.

Claims

1. A method for fault diagnosis and self-healing control of a pump cluster in a thermal power plant, characterized in that, include: An array microphone is used to synchronously acquire the full-band acoustic signal of the pump cluster of a thermal power plant, and the original full-band acoustic data of the pump cluster is obtained, wherein the pump cluster includes multiple pumps operating in parallel. The original full-band acoustic data is preprocessed, wherein the preprocessing includes hierarchical decomposition, adaptive filtering, target acoustic band separation, feature extraction and dimensionality reduction, and principal component dimensionality reduction, to obtain preprocessed target feature data. The target feature data is fused with the pump operating parameters of the pump cluster to construct the state space of the target network; The decision network in the target network outputs the probability distribution of the failure type of the pump cluster according to the state space, and the evaluation network in the target network evaluates the output of the decision network. A reward function is set based on the fault identification degree, and the decision network and the evaluation network are updated in conjunction with the reward function. Fault diagnosis and self-healing control of the pump cluster are achieved based on the updated target network.

2. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 1, characterized in that, The method employs an array microphone to synchronously acquire the full-band acoustic signals of the pump cluster in a thermal power plant, obtaining the raw full-band acoustic data of the pump cluster, including: An array microphone is provided, wherein the array microphone includes multiple microphones, the microphones are placed at the bearing end and the inlet / outlet end of each pump body in the pump body cluster, and the microphones are attached to the surface of the pump body by magnetic mounting brackets; Using the array microphone, the full-band acoustic signals of each pump in the pump cluster are collected synchronously to obtain the raw full-band acoustic data of the pump cluster.

3. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 1, characterized in that, The hierarchical decomposition includes: The original whole-band data is decomposed using wavelet packets with a preset number of layers to obtain multiple corresponding sub-bands, wherein the preset number of layers includes any number of layers in the range of 3 to 5. The multiple sub-bands are merged into a first frequency band, a second frequency band, and a third frequency band, wherein the first frequency band includes low-frequency data, the second frequency band includes mid-frequency data, and the third frequency band includes high-frequency data.

4. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 3, characterized in that, The adaptive filtering includes: Adaptive filtering is performed on each of the multiple sub-bands based on an adaptive filtering threshold. The adaptive filtering threshold is dynamically adjusted based on the root mean square value of the first N sets of historical data for the corresponding sub-band, where N is a positive integer greater than 1. The formula for calculating the adaptive filtering threshold includes: ， In the formula, Th M is the adaptive filtering threshold, M is the adjustment coefficient, and N is the number of historical data sets. x i For the first i Feature values ​​of a set of historical data.

5. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 1, characterized in that, The feature extraction and dimensionality reduction include: Feature extraction and dimensionality reduction are performed on the separated target acoustic bands to obtain a multidimensional feature matrix, wherein the multidimensional feature matrix includes at least one time-domain feature, at least one frequency-domain feature and at least one time-frequency-domain feature. The formula for calculating the impulse factor in the time-domain features includes: ， In the formula, I p For pulse factor, x max The peak value of the time-domain waveform of the acoustic signal. L The number of data points in a single group of signals. x j For the first j One time-domain data point.

6. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 5, characterized in that, The principal component dimensionality reduction includes: Set the cumulative contribution rate of principal components, where the cumulative contribution rate of principal components includes any percentage within the range of 90% to 98%. Based on the cumulative contribution rate of the principal components, principal component dimensionality reduction is performed on the multidimensional feature matrix to obtain the target feature data, wherein the dimension of the target feature data includes any dimension in the range of 3 to 7.

7. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 1, characterized in that, The reward function based on fault identification includes: The fault type with the highest probability in the fault type probability distribution is selected as the fault judgment result for each pump body. Calculate the cosine similarity between the preprocessed target feature data corresponding to the fault judgment result and the fault feature template corresponding to the fault judgment result; Based on the cosine similarity, a reward function is set: when the cosine similarity is greater than or equal to a first threshold, a first positive reward is set; when the cosine similarity is greater than or equal to a second threshold and less than the first threshold, a second positive reward is set; when the cosine similarity is less than the second threshold, a negative penalty is set. The formula for calculating the cosine similarity includes: ， In the formula, S For cosine similarity, a k This is the k-th element in the fault feature template. b k Let m be the k-th element of the preprocessed target feature data, where m is the dimension of the feature vector.

8. The fault diagnosis and self-healing control method for a thermal power plant pump cluster according to claim 7, characterized in that, The fault types of the pump cluster include no faults and multiple actual faults; the fault diagnosis and self-healing control of the pump cluster based on the updated target network includes: In the probability distribution of fault types of each pump in the pump cluster, if the probability of no fault is greater than or equal to the first threshold, the pump is determined to be normal and the pump is controlled not to operate. In the probability distribution of fault types of each pump in the pump cluster, if the probability of any actual fault is greater than or equal to the third threshold, it is determined that the pump has the actual fault, and the alarm and pump action are controlled. The third threshold is greater than the second threshold and less than the first threshold.

9. The fault diagnosis and self-healing control method for pump clusters in thermal power plants according to any one of claims 1-8, characterized in that, The initial learning rate of the evaluation network is a first learning rate, wherein the evaluation network incorporates a dynamic adjustment of the learning rate based on the real-time change rate of the full-band features: when the real-time change rate of the full-band features is less than or equal to a preset threshold, the learning rate is maintained at the first learning rate; when the real-time change rate of the full-band features is greater than the preset threshold, the learning rate is adjusted to a second learning rate; wherein the second learning rate is greater than the first learning rate. The formula for calculating the real-time rate of change of the full-band characteristics mentioned above. include: ， In the formula, Δ A For the real-time rate of change of characteristics across the entire acoustic band, A t Let A be the characteristic amplitude at the current moment. t-1 Δt represents the characteristic amplitude at the previous moment, and Δt represents the time interval.

10. The fault diagnosis and self-healing control method for pump clusters in thermal power plants according to any one of claims 1-8, characterized in that, The preprocessing uses predetermined preprocessing parameters, which include the number of wavelet packet decomposition layers, principal component dimensionality reduction dimension, and adaptive filtering threshold. The preprocessing parameters are determined based on a pre-stored preprocessing parameter mapping library, which includes the mapping relationship between the equipment type of the pump cluster and the preprocessing parameters. When the equipment type of the pump cluster does not exist in the preprocessing parameter mapping library, the preprocessing parameters are fine-tuned based on the pre-acquired sample data of the pump cluster using a network search method, and the equipment type of the pump cluster and the corresponding preprocessing parameters are stored in the preprocessing parameter mapping library. The objective function for fine-tuning includes: ， In the formula, Acc The fault diagnosis accuracy is represented by TP, which is the number of true positive samples (the number of samples that are actually faulty and diagnosed as faulty), TN, which is the number of true negative samples (the number of samples that are actually normal and diagnosed as normal), FP, which is the number of false positive samples (the number of samples that are actually normal but diagnosed as faulty), and FN, which is the number of false negative samples (the number of samples that are actually faulty but diagnosed as normal).

Images