Intelligent water pump control system

Abstract

The invention discloses an intelligent water pump control system, which belongs to the technical field of intelligent control and comprises a data acquisition module, a data preprocessing module, an operation state detection model construction module, an energy-saving efficiency prediction model construction module and an intelligent control module. According to the method, attention weights are calculated based on mutual information, time sequence fusion features are obtained, a multi-modal attention dynamic graph is generated, an initial feature matrix is extracted, time sequence trend features are obtained based on a trend slope, a residual error between the initial features and a window mean value is calculated and converted into a mutation probability, and the trend slope and the mutation probability are introduced to adjust gating parameters. According to the method, multi-frequency subsequences are obtained based on wavelet decomposition, time sequence features are extracted, a time feature tensor is obtained in combination with an attention mechanism, time-channel cross features are weighted through nonlinear channel weights, a channel feature tensor is obtained, an accurate motor rotating speed instruction is provided for intelligent control, and the safety and stability of water pump control are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent control, and particularly relates to an intelligent water pump control system. BACKGROUND

[0002] The water pump control system collects data during water pump operation, analyzes and processes the data by means of intelligent algorithms, timely warns potential equipment faults, reduces downtime maintenance time, simultaneously converts real-time operation data into precise control instructions, realizes automatic adjustment and optimized operation of the water pump, and assists operators to improve water supply efficiency and reduce energy consumption. However, the existing water pump control system has the problems that the water pump fault features are represented by multiple modal parameters, and the importance of each mode dynamically changes over time, the traditional method is difficult to effectively identify slow development gradual faults, and cannot timely capture sudden abnormalities, resulting in insufficient early warning capability and high fault missing rate, and further reducing the safety and stability of water pump control; the existing water pump control system has the problem that the fluctuation of water pump energy-saving related parameters contains multiple frequency characteristics, and there is a nonlinear correlation between parameters, the traditional method cannot separate multiple frequency characteristics, and it is difficult to capture nonlinear correlation, resulting in the problem that the water pump control cannot achieve the optimal energy-saving state. SUMMARY

[0003] In view of the above, in order to overcome the defects of the prior art, the application provides an intelligent water pump control system, which is aimed at the fact that the water pump fault characteristics of the existing water pump control system are represented by multiple modal parameters, and the importance of each mode changes dynamically over time, and the traditional method cannot effectively identify slow developing gradual faults and cannot timely capture sudden abnormalities, resulting in insufficient early warning capability and high fault omission rate, thereby reducing the safety and stability of water pump control. The present scheme calculates attention weights based on mutual information, obtains time sequence fusion features, generates a multi-modal attention dynamic graph, extracts an initial feature matrix, obtains time sequence trend features based on a trend slope, calculates the residual of the initial features and the window mean and converts it into a mutation probability, introduces the trend slope and the mutation probability to adjust the gating parameters, and obtains the running state detection result based on the time sequence key features, thereby improving the accuracy and timeliness of the running state detection, early warning when the water pump has early faults, ensuring continuous and stable operation of the water pump, providing a reliable state basis for intelligent control, and improving the safety and stability of water pump control. In view of the fact that the fluctuations of the water pump energy-saving related parameters in the existing water pump control system contain multiple frequency characteristics, and there is a nonlinear correlation between the parameters, and the traditional method cannot separate the multiple frequency characteristics and cannot capture the nonlinear correlation, resulting in the problem that the water pump control cannot achieve the optimal energy-saving state. The present scheme obtains multiple frequency sub-sequences based on wavelet decomposition, extracts time sequence features and obtains time feature tensors by combining the attention mechanism, weights the time-channel cross features by using nonlinear channel weight, obtains channel feature tensors, obtains energy efficiency prediction results based on time-channel fusion features, improves the accuracy of energy efficiency prediction, accurately predicts the optimal energy efficiency corresponding to the operating parameters under different conditions, provides accurate motor speed instructions for intelligent control, and realizes efficient energy-saving operation of the water pump under different conditions.

[0004] The intelligent water pump control system provided by the application comprises a data acquisition module, a data preprocessing module, a running state detection model construction module, an energy efficiency prediction model construction module and an intelligent control module.

[0005] The data acquisition module acquires historical water pump time sequence data.

[0006] The data preprocessing module performs data cleaning, data normalization, data encoding and dataset construction on the data to obtain detection time sequence datasets and prediction time sequence datasets.

[0007] The running state detection model construction module calculates attention weights based on mutual information, obtains time sequence fusion features, generates a multi-modal attention dynamic graph, extracts an initial feature matrix, obtains time sequence trend features based on a trend slope, calculates the residual of the initial features and the window mean and converts it into a mutation probability, introduces the trend slope and the mutation probability to adjust the gating parameters, and obtains the running state detection result of the water pump based on the time sequence key features.

[0008] The energy-saving efficiency prediction model construction module obtains multiple frequency subsequences based on wavelet decomposition, extracts time sequence features, and obtains a time feature tensor by combining an attention mechanism, weights the time-channel cross feature by using a nonlinear channel weight, obtains a channel feature tensor, and obtains the energy-saving efficiency prediction result of the water pump based on the time-channel fusion feature;

[0009] The intelligent control module realizes intelligent control of the water pump according to the outputs of the running state detection model and the energy-saving efficiency prediction model.

[0010] Further, the data acquisition module acquires historical water pump time sequence data;

[0011] The historical water pump time sequence data includes a timestamp, a water pump operating parameter, a water pump pipe network parameter, an environmental parameter, a water pump vibration parameter, a water pump noise parameter, a running state, and an energy-saving efficiency grade.

[0012] Further, the data preprocessing module performs data cleaning, data normalization, data encoding, and dataset construction on the data to obtain a detection time sequence dataset and a prediction time sequence dataset;

[0013] The dataset construction is to construct the detection time sequence dataset and the prediction time sequence dataset; the detection time sequence dataset is constructed based on the timestamp, the water pump operating parameter, the water pump vibration parameter, the water pump noise parameter, and the running state after data cleaning, data normalization, and data encoding, and the running state is taken as a data label in the detection time sequence dataset; the prediction time sequence dataset is constructed based on the timestamp, the water pump operating parameter, the water pump pipe network parameter, the environmental parameter, and the energy-saving efficiency grade after data cleaning, data normalization, and data encoding, and the energy-saving efficiency grade is taken as a data label in the prediction time sequence dataset.

[0014] Further, the running state detection model construction module constructs the running state detection model of the water pump by using a machine learning method based on the detection time sequence dataset, and is provided with a multi-modal attention dynamic graph generation unit, an initial feature extraction unit, a trend perception unit, a mutation perception unit, a trend-mutation weighting gate unit, and a detection unit, and specifically includes the following contents:

[0015] The multi-modal attention dynamic graph generation unit; attention weights are calculated based on the mutual information of each time step of each modal feature and the data label, the multi-modal features are weighted and summed, and the time sequence fusion feature of each time step is obtained; the time sequence fusion feature of each time step is taken as a graph node, the cosine similarity of the time sequence fusion feature between each two graph nodes is compressed to by using a Sigmoid function, as the weight of the edge between the two graph nodes, and a multi-modal attention dynamic graph is generated;

[0016] An initial feature extraction unit; using a graph convolutional neural network, based on the adjacency matrix of the multi-modal attention dynamic graph, performing graph convolution operation on the time sequence fusion feature matrix to extract an initial feature matrix containing time sequence correlation information;

[0017] A trend perception unit; setting a trend window size, performing linear regression on the time steps and their corresponding initial features within the trend window, calculating the trend slope, normalizing the trend slope through a tanh function, and then concatenating the normalized result with the initial features to obtain time sequence trend features containing trend information;

[0018] A mutation perception unit; setting a mutation detection window size, calculating the residual of the initial feature of each time step graph node and the average of the initial features of the graph nodes within the mutation detection window, and based on the residual mean and standard deviation of the time steps in the detection time sequence dataset, converting the residual into a mutation probability through a Sigmoid function;

[0019] A trend-mutation weighted gating unit; introducing trend slope and mutation probability to adjust gating parameters, controlling the fusion ratio of historical features and current features, updating the hidden state of each time step based on the gating parameters to obtain the gated time sequence features of each time step graph node, and concatenating the gated time sequence features and the initial features to obtain the time sequence key features of each time step graph node;

[0020] A detection unit; after dimensionality reduction and nonlinear transformation of the time sequence key features through a fully connected layer, outputting the running state detection result of the corresponding time step water pump.

[0021] Further, the energy efficiency prediction model construction module is based on the prediction time sequence dataset, and a machine learning method is used to construct the energy efficiency prediction model of the water pump, and is provided with an adversarial enhancement unit, a multi-frequency time sequence fusion unit, a nonlinear channel interaction unit and a prediction unit, and specifically includes the following contents:

[0022] An adversarial enhancement unit; using a gradient-penalty-based Wasserstein generative adversarial network WGAN-GP to perform data enhancement on the prediction time sequence dataset to obtain an enhanced prediction time sequence dataset B;

[0023] The multi-frequency time sequence fusion unit is used for wavelet decomposition of the enhanced prediction time sequence data set to obtain four frequency sub-sequences, bidirectional gate recurrent units are used for extracting time sequence features of each frequency sub-sequence, attention mechanisms are combined, a query matrix is constructed by aligning and stacking the time sequence features of all frequency sub-sequences along a time axis, a key matrix and a value matrix are constructed by aligning and stacking the time sequence features of all frequency sub-sequences along a frequency axis, for each time step and frequency sub-sequence, attention weights are calculated by similarity between a query vector and a key vector, and cross features of each time step are obtained by weighted fusion of value vectors based on the attention weights, the cross features of all time steps are spliced in time sequence to obtain a time feature tensor;

[0024] The nonlinear channel interaction unit comprises the following contents:

[0025] The nonlinear channel weight is calculated; B and M all are spliced along the channel axis, a residual vector is constructed after linear transformation, a radial basis function is used for nonlinear mapping of features of each channel at each time step in the residual vector, nonlinear correlation weights between different channels are calculated based on the mapped nonlinear features to obtain a nonlinear channel weight matrix W nl ;

[0026] Time-channel cross attention; for the residual vector, cosine similarity between feature vectors corresponding to each two time step-channel pairs is calculated to construct a similarity matrix, for each time step-channel pair, five time step-channel pairs with the highest similarity to the pair are selected as a neighborhood in the similarity matrix, feature vectors in the neighborhood are weighted and aggregated to obtain time-channel cross features of each time step-channel pair to form a time-channel cross feature matrix R cross ;

[0027] Dynamic channel screening; W nl is used to weight R cross to obtain a weighted time-channel cross feature matrix , and channel attention weights of each channel in R are calculated to obtain a channel feature tensor R filter after preliminary screening;

[0028] The prediction unit is used for splicing the time feature tensor and the channel feature tensor along a feature axis, linear transformation to obtain time-channel fusion features, dimension reduction and nonlinear transformation of the time-channel fusion features through a fully connected layer, and output of a prediction result of energy-saving efficiency of the water pump corresponding to the time step.

[0029] Further, the intelligent control module collects real-time water pump time series data; the real-time water pump time series data includes a timestamp, water pump operating parameters, water pump pipe network parameters, environmental parameters, water pump vibration parameters, and water pump noise parameters; the real-time water pump time series data is subjected to data cleaning, data normalization, and data encoding processing, the timestamp, water pump operating parameters, water pump vibration parameters, and water pump noise parameters are extracted from the processed real-time water pump time series data and input into the operating state detection model, if the operating state output by the operating state detection model is a fault, an early warning signal is immediately sent to relevant personnel; otherwise, the timestamp, water pump operating parameters, water pump pipe network parameters, and environmental parameters are extracted from the processed real-time water pump time series data and input into the energy saving efficiency prediction model, the motor speed corresponding to the highest energy saving efficiency level output by the energy saving efficiency prediction model is taken as a control instruction, and intelligent control of the water pump is realized.

[0030] The beneficial effects achieved by the above-mentioned scheme are as follows:

[0031] (1) The existing water pump control system has the problem that the water pump fault features are represented by multiple modal parameters, and the importance of each mode changes dynamically over time. Traditional methods cannot effectively identify slow developing gradual faults, and cannot timely capture sudden abnormalities, resulting in insufficient early warning capability and high fault omission rate, thereby reducing the safety and stability of water pump control. The scheme calculates the mutual information of each modal feature and label to obtain attention weights, obtains time series fusion features, generates a multi-modal attention dynamic graph, extracts an initial feature matrix, and reduces misjudgments caused by improper modal weights or ignored time series correlation. The time series trend features are obtained based on the trend slope to quantify the early development trend of gradual faults. The residual of the initial feature and the window mean is calculated and converted into a mutation probability to quantify the severity of instantaneous abnormalities. The trend slope and mutation probability are introduced to adjust the gating parameters to obtain time series key features, so that the model focuses on historical trend features and current mutation features according to specific conditions, making the time series key features more consistent with actual fault patterns. The output operating state detection result improves the accuracy and timeliness of operating state detection, can provide early warning when the water pump has early faults, ensures continuous and stable operation of the water pump, provides a reliable state basis for intelligent control, and improves the safety and stability of water pump control.

[0032] (2) In view of the fact that the existing water pump control system has fluctuations in energy-saving related parameters of the water pump, which contain multi-frequency characteristics, and there is a nonlinear correlation between the parameters, the traditional method cannot separate the multi-frequency characteristics and capture the nonlinear correlation, resulting in the problem that the water pump control cannot achieve the optimal energy-saving state, the scheme uses WGAN-GP for data enhancement, providing a more complete data basis for subsequent feature extraction; based on wavelet decomposition, multi-frequency subsequences are obtained, time sequence features are extracted, and time feature tensors are obtained by combining the attention mechanism, which can accurately capture the correlation between different frequency parameters and energy-saving efficiency, so that the water pump control can dynamically adjust the strategy according to the parameter fluctuations of different frequencies; the time-channel cross features are weighted by nonlinear channel weights to obtain channel feature tensors, which can accurately capture the nonlinear relationship between parameters and fully reflect the complex influence mechanism of energy-saving efficiency; the time feature tensor and the channel feature tensor are fused to output the energy-saving efficiency prediction result, which improves the accuracy of energy-saving efficiency prediction and can accurately predict the optimal energy-saving efficiency corresponding to the operating parameters under different conditions, providing accurate motor speed instructions for intelligent control, and realizing efficient energy-saving operation of the water pump under different conditions. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A schematic diagram of an intelligent water pump control system provided by the present application is shown in the figure.

[0034] Figure 2 A schematic diagram of a running state detection model construction module is shown in the figure.

[0035] Figure 3 A schematic diagram of an energy-saving efficiency prediction model construction module is shown in the figure.

[0036] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not 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 scope of the present application.

[0038] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0039] Embodiment one, refer to Figure 1 The present application provides an intelligent water pump control system, comprising a data acquisition module, a data preprocessing module, a running state detection model construction module, an energy saving efficiency prediction model construction module and an intelligent control module.

[0040] The data acquisition module acquires historical water pump time series data and sends the data to the data preprocessing module.

[0041] The data preprocessing module receives the data sent by the data acquisition module, performs data cleaning, data normalization, data encoding and dataset construction on the data, obtains detection time series dataset and prediction time series dataset, and sends the data to the running state detection model construction module.

[0042] The running state detection model construction module receives the data sent by the data preprocessing module, calculates attention weights based on mutual information, obtains time series fusion features, generates multi-modal attention dynamic graph, extracts initial feature matrix, obtains time series trend features based on trend slope, calculates the residual of initial features and window mean and converts it into mutation probability, introduces trend slope and mutation probability to adjust gating parameters, obtains the running state detection result of the water pump based on time series key features, and sends the data to the energy saving efficiency prediction model construction module.

[0043] The energy saving efficiency prediction model construction module receives the data sent by the running state detection model construction module, obtains multi-frequency subsequences based on wavelet decomposition, extracts time series features and obtains time feature tensor combined with attention mechanism, weights time-channel cross features with nonlinear channel weights, obtains channel feature tensor, and obtains the energy saving efficiency prediction result of the water pump based on time-channel fusion features, and sends the data to the intelligent control module.

[0044] The intelligent control module receives the data sent by the energy saving efficiency prediction model construction module, and realizes intelligent control of the water pump according to the output of the running state detection model and the energy saving efficiency prediction model.

[0045] Embodiment two, refer to Figure 1 This embodiment is based on the above embodiment, in the data acquisition module, the historical water pump time series data is collected, and the sampling frequency is 1Hz.

[0046] The historical water pump time series data includes timestamps, water pump operating parameters, water pump pipe network parameters, environmental parameters, water pump vibration parameters, water pump noise parameters, operating states, and energy-saving efficiency levels.

[0047] The water pump operating parameters include water pump flow, outlet pressure, inlet pressure, motor speed, motor power, current, voltage, medium temperature, motor temperature, bearing temperature, and operating time length.

[0048] The water pump pipe network parameters include pipe network pressure fluctuation values, pipe diameters, and valve openings.

[0049] The environmental parameters include environmental temperature, environmental humidity, and atmospheric pressure.

[0050] The water pump vibration parameters include vibration acceleration, vibration speed, vibration displacement, and vibration frequency spectrum characteristics.

[0051] The water pump noise parameters include noise decibel values and noise frequency spectrum characteristics.

[0052] The operating states include normal and fault.

[0053] The energy-saving efficiency levels include very poor, poor, general, good, and excellent.

[0054] Embodiment Three, see Figure 1 This embodiment is based on the above-mentioned embodiment, and in the data preprocessing module, data cleaning, data normalization, data encoding, and data set construction are performed on the data to obtain the detection time series data set and the prediction time series data set.

[0055] The data cleaning is to remove error values, missing values, and abnormal values in the data.

[0056] The data normalization is to use the maximum-minimum scaling method to unify the data into the same range.

[0057] The data encoding is to use One-Hot encoding to convert data labels into numerical data.

[0058] The data set construction is to construct the detection time series data set and the prediction time series data set; based on the timestamps, water pump operating parameters, water pump vibration parameters, water pump noise parameters, and operating states after data cleaning, data normalization, and data encoding, the detection time series data set is constructed, and the operating state is taken as the data label in the detection time series data set; based on the timestamps, water pump operating parameters, water pump pipe network parameters, environmental parameters, and energy-saving efficiency levels after data cleaning, data normalization, and data encoding, the prediction time series data set is constructed, and the energy-saving efficiency level is taken as the data label in the prediction time series data set.

[0059] Embodiment Four, see Figure 1 andFigure 2 The embodiment is based on the above-mentioned embodiment, and in the running state detection model construction module, a running state detection model of the water pump is constructed based on the detection time series data set by using a machine learning method, and a multi-modal attention dynamic graph generation unit, an initial feature extraction unit, a trend perception unit, a mutation perception unit, a trend-mutation weighted gate unit and a detection unit are arranged, and the embodiment specifically includes the following contents:

[0060] The multi-modal attention dynamic graph generation unit; the water pump fault is jointly reflected by multi-modal features, but the importance of each modality changes over time, and direct fusion ignores the weight difference. At the same time, the correlation between time steps is not effectively captured. By adaptively focusing on key modalities, the time series dependence is dynamically captured, and the pertinence and time series correlation of feature fusion are improved. The attention weight is calculated based on the mutual information of each time step of each modality feature and the data label, and the multi-modal features are weighted and summed to obtain the time series fusion features of each time step. The time series fusion features of each time step are taken as a graph node, and the cosine similarity of the time series fusion features between each two graph nodes is compressed to by using a Sigmoid function, as the weight of the edge between the two graph nodes, and a multi-modal attention dynamic graph is generated. The formula used is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] wherein, , and are the water pump running parameter features, the water pump vibration parameter features and the water pump noise parameter features of the tth time step in the detection time series data set, is the i th modality feature of the t th time step in the detection time series data set, and y t is the data label corresponding to the t th time step in the detection time series data set, , , and are the mutual information of , , , and y t , , , and are the attention weights of , , and , and Lt and L k are the time step fusion features of the t-th and k-th time steps, respectively, is the L2 norm, is the smoothing term, , is the weight of the edge between the graph node at the t-th time step and the graph node at the k-th time step, and t and k are the time step indices, and are the exponential function and the Sigmoid function, respectively, is the max function;

[0065] The initial feature extraction unit; the nodes and edges in the multi-modal attention dynamic graph contain time sequence correlation information, but traditional sequence models are difficult to use the graph structure to mine global correlation, and by fully utilizing the global information of the graph structure, the hidden correlation of non-continuous time steps is captured, which is more comprehensive than the pure time sequence model; the graph convolutional neural network is used, the adjacency matrix of the multi-modal attention dynamic graph is used to perform graph convolution operation on the time step fusion feature matrix, and the initial feature matrix containing time sequence correlation information is extracted; the formula used is as follows:

[0066] ;

[0067] In the formula, A and L are the initial feature matrix and the time step fusion feature matrix of all graph nodes in the multi-modal attention dynamic graph, respectively, I is the unit matrix, Z is the adjacency matrix of the multi-modal attention dynamic graph, D is the degree matrix of Z, W o and b o are the learnable weights and bias of the graph convolution, respectively, is the leaky linear rectifier function;

[0068] The trend perception unit; pump failure is often accompanied by a gradual trend, and relying only on instantaneous features may ignore early signs of failure. By identifying early features of gradual failure, potential failures can be warned in advance to avoid sudden shutdown. The trend window size p is set to 5, the time steps in the trend window and their corresponding initial features are linearly regressed, the trend slope is calculated, the trend slope is normalized by the tanh function, and then the normalized result is spliced with the initial feature to obtain the time step trend feature containing trend information; the formula used is as follows:

[0069] ;

[0070] ;

[0071] In the formula, c t is the trend slope at the t-th time step, s is the index used to traverse the time steps in the trend window, A t and A sare the initial features of the t-th and s-th time step graph nodes, respectively, and are the mean values of the time steps and the initial features within the trend window, respectively, X t is the time series trend feature of the t-th time step graph node, is the tanh function, is the concatenation operation;

[0072] The mutation awareness unit; water pump failure may suddenly occur, and it is difficult to capture instantaneous abnormalities by relying only on trend features. By quickly responding to sudden failure features, the missed detection rate of sudden failures is reduced; the mutation detection window size q = 3 is set, the residual of the initial feature of each time step graph node and the mean value of the initial feature of the graph node within the mutation detection window is calculated, based on the mean value and standard deviation of the residual of the time step in the detection time series data set, the residual is converted into a mutation probability through the Sigmoid function; the formula used is as follows:

[0073] ;

[0074] ;

[0075] In the formula, e t is the residual of the t-th time step, is the mean value of the initial feature within the mutation detection window, T is the transpose operation, is the covariance matrix of the initial feature within the mutation detection window, f t is the mutation probability of the t-th time step, μ e and σ e are the mean value and standard deviation of the residual, respectively; t0 is the sliding window reference value, t0 = 100; when t = 0 or 1, μ e = 0, σ e = 1; when t < t0, μ e and σ e take the mean value and standard deviation of the residual of the first t time steps in the detection time series data set; when t ≥ t0, μ e and σ e take the mean value and standard deviation of the residual of the first t0 time steps in the detection time series data set;

[0076] Trend-mutation weighted gate unit; the fusion ratio of traditional gate mechanism for historical and current features is fixed, which cannot adapt to the dynamic change of water pump state. By focusing on historical trends during gradual failure and focusing on current abnormalities during sudden failure, the model's adaptability to different failure modes is improved. Trend slope and mutation probability are introduced to adjust the gate parameters to control the fusion ratio of historical and current features. The hidden state of each time step is updated based on the gate parameters to obtain the gated time sequence features of each graph node. The gated time sequence features and initial features are concatenated to obtain the time sequence key features of each graph node. The formula is as follows:

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] where r t and z t are the adjusted reset gate and update gate respectively, is the candidate hidden state, is the element product, W h and b h are the learnable weights and biases of the hidden layer respectively, W r and b r are the learnable weights and biases of the reset gate respectively, W z and b z are the learnable weights and biases of the update gate respectively, H t and H t-1 are the gated time sequence features of the graph node at the tth and (t-1)th time steps respectively, and λ1 and λ2 are the learnable weights of the trend slope and mutation probability respectively;

[0082] Detection unit; after dimensionality reduction and nonlinear transformation of the time sequence key features through the fully connected layer, the running state detection result of the water pump corresponding to the time step is output.

[0083] By performing the above operations, in view of the fact that the fault characteristics of the existing water pump control system are jointly represented by multi-modal parameters, and the importance of each mode changes dynamically over time, and the traditional method is difficult to effectively identify slowly developing gradual faults, and cannot capture sudden anomalies in time, resulting in insufficient early warning capabilities and high fault missed detection rate, thereby reducing the safety and stability of water pump control, this scheme calculates the mutual information of each modal feature and label to obtain the attention weight, obtains the time series fusion feature, generates a multi-modal attention dynamic graph, extracts the initial feature matrix, and reduces the misjudgment caused by improper modal weights or neglect of time series correlation; based on the trend slope, Time series trend features are used to quantify the early development trend of gradual faults; the residual between the initial features and the window mean is calculated and converted into mutation probability to quantify the severity of transient anomalies; the trend slope and mutation probability are introduced to adjust the gating parameters to obtain time series key features, so that the model focuses on historical trend features and current mutation features according to specific circumstances, making the time series key features more in line with actual fault modes; the output of operation status detection results improves the accuracy and timeliness of operation status detection, can provide early warning when early faults occur in the water pump, ensure the continuous and stable operation of the water pump, provide a reliable status basis for intelligent control, and improve the safety and stability of water pump control.

[0084] Example 5, see Figure 1 and Figure 3 Based on the above embodiment, this embodiment uses a machine learning method to construct an energy-saving efficiency prediction model for a water pump in an energy-saving efficiency prediction model construction module based on a prediction time series data set. The model is provided with an anti-enhancement unit, a multi-frequency time series fusion unit, a nonlinear channel interaction unit, and a prediction unit. Specifically, the model includes the following contents:

[0085] Adversarial enhancement unit: In historical energy-saving efficiency data, some level samples may be relatively rare, resulting in low prediction accuracy for rare levels. By improving the generalization ability of the model for rare energy-saving levels, the impact of data imbalance on prediction is reduced. The Wasserstein generative adversarial network (WGAN-GP) based on gradient penalty is used to enhance the predicted time series dataset. The generator has a 2-layer fully connected + 1-layer GRU structure, and the discriminator has a 1-layer GRU + 2-layer fully connected structure. The generator input is standard Gaussian noise and outputs synthetic data with the same dimension as the original data. The discriminator outputs a authenticity score. The synthetic data is combined with the original data to obtain the enhanced predicted time series dataset B.

[0086] A multi-frequency time sequence fusion unit; the energy efficiency of the water pump is affected by the parameter fluctuation of different frequencies, and it is difficult to capture the full frequency domain characteristics through single frequency analysis. By separating and utilizing the key information of different frequencies, the prediction accuracy is improved; the enhanced prediction time sequence dataset is decomposed by db6 wavelet to obtain 4 frequency subsequences, the frequency ranges corresponding to the 4 frequency subsequences are 0-16Hz, 16-32Hz, 32-64Hz and 64-128Hz respectively, the time sequence features of each frequency subsequence are extracted by using a bidirectional gate recurrent unit respectively, the time sequence features of all frequency subsequences are aligned and stacked along the time axis to construct a query matrix, the time sequence features of all frequency subsequences are aligned and stacked along the frequency axis to construct a key matrix and a value matrix, for each time step and frequency subsequence, the attention weight is calculated by the similarity between the query vector and the key vector, and the cross feature of each time step is obtained based on the attention weight. Weighted fusion of value vectors, all time steps of cross features are spliced in chronological order to obtain a time feature tensor; the formula used is as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, B g is the gth frequency subsequence, g is the frequency index, and are the bidirectional gate recurrent unit function and the Softmax function respectively, and are the weights and biases of the forward GRU corresponding to the gth frequency subsequence, and are the weights and biases of the backward GRU corresponding to the gth frequency subsequence, E g is the time sequence feature of B g , Q, K and V are the query matrix, the key matrix and the value matrix respectively, E1, E2, E3 and E4 are the time sequence features of the 1st, 2nd, 3rd and 4th frequency subsequences respectively, and are the time axis alignment and stacking operation and the frequency axis alignment and stacking operation respectively, is the attention weight of the gth frequency subsequence at the tth time step, is the query vector corresponding to the t-th time step and the g-th frequency subsequence in Q, is the key vector corresponding to the g-th frequency subsequence and the t-th time step in K, is the value vector corresponding to the g-th frequency subsequence and the t-th time step in V, C is the feature dimension of a single frequency subsequence, M t and M N are the cross features of the 1st, t-th and N-th time steps, respectively, M all is the time feature tensor, and N is the total number of time steps;

[0094] a nonlinear channel interaction unit; fully captures the nonlinear, spatiotemporal cross-correlation between parameters, filters effective features, and improves the accuracy of energy efficiency prediction; including the following contents:

[0095] calculating nonlinear channel weights; solving the problem of nonlinear correlation between pump parameters which is difficult to model, capturing the nonlinear correlation between channels through radial basis function, improving the accuracy of feature interaction; B and M all are spliced along the channel axis, and after linear transformation, a residual vector is constructed, the features of each channel in the residual vector at each time step are nonlinearly mapped using radial basis function, based on the mapped nonlinear features, the nonlinear correlation weights between different channels are calculated, and a nonlinear channel weight matrix W nl is obtained; the formula used is as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] In the formula, O is the residual vector, W c and b c are the weight and bias of linear transformation, respectively, is the feature vector of the residual vector at the t-th time step and the u-th channel, is the feature vector of the residual vector at the t-th time step and the v-th channel, is the mean vector of the residual vector at the u-th channel, is the variance of the u-th channel feature, and are the nonlinear features obtained after radial basis function mapping of and , respectively, is the nonlinear correlation weight between the u-th channel and the v-th channel of the residual vector;

[0100] Time-channel cross attention; solve the problem of ignoring the cross characteristics of time steps and parameter channels, capture the spatio-temporal cross characteristics through neighborhood aggregation, and enrich the feature dimension; for the residual vector, calculate the cosine similarity between the feature vectors corresponding to each two time step-channel pairs, construct a similarity matrix, for each time step-channel pair, select the top 5 time step-channel pairs with the highest similarity to the pair in the similarity matrix as the neighborhood, and aggregate the feature vectors in the neighborhood to obtain the time-channel cross characteristics of each time step-channel pair, forming a time-channel cross characteristic matrix R cross ; The formula used is as follows:

[0101] ;

[0102] ;

[0103] In the formula, is the neighborhood set of the time step-channel pair , and are the jth and mth feature vectors in , is the cosine similarity between and , is the cosine similarity between and , is the normalized cosine similarity weight between and , is the time-channel cross characteristic of the time step-channel pair ;

[0104] Dynamic channel screening; solve the problem of irrelevant channel interference model, select key channels through channel attention, and reduce redundant information; use W nl to weight R cross , obtain the weighted time-channel cross characteristic matrix , calculate the channel attention weight of each channel in , and obtain the channel feature tensor R filter after preliminary screening by weighting; the formula used is as follows:

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, U pass is the number of channels in the residual vector, is a time-step-channel pair the weighted time-channel cross feature, is the attention weight of the u-th channel;

[0109] The prediction unit is used to splice the time feature tensor and the channel feature tensor along the feature axis, and then linearly transform the time-channel fusion feature to obtain the energy-saving efficiency prediction result of the corresponding time step water pump.

[0110] By performing the above operation, for the problem that the existing water pump control system has fluctuation of water pump energy-saving related parameters containing multi-frequency characteristics, and there is a nonlinear correlation between parameters, and the traditional method cannot separate the multi-frequency characteristics and is difficult to capture the nonlinear correlation, resulting in that the water pump control cannot reach the optimal energy-saving state, the scheme uses WGAN-GP for data enhancement, which provides a more complete data basis for subsequent feature extraction; based on wavelet decomposition, a multi-frequency subsequence is obtained, and the time feature tensor is obtained by extracting the time sequence feature and combining the attention mechanism, which can accurately capture the correlation between different frequency parameters and energy-saving efficiency, so that the water pump control can dynamically adjust the strategy according to the parameter fluctuation of different frequencies; the time-channel cross feature is weighted by the nonlinear channel weight to obtain the channel feature tensor, which can accurately capture the nonlinear relationship between parameters and comprehensively reflect the complex influence mechanism of energy-saving efficiency; the time feature tensor and the channel feature tensor are fused to output the energy-saving efficiency prediction result, which improves the accuracy of energy-saving efficiency prediction and can accurately predict the optimal energy-saving efficiency corresponding to the operating parameters under different conditions, provides accurate motor speed instructions for intelligent control, and realizes efficient energy-saving operation of the water pump under different conditions.

[0111] Embodiment six, refer to Figure 1 The embodiment is based on the above-mentioned embodiment, in the intelligent control module, real-time water pump time sequence data is collected; the real-time water pump time sequence data includes time stamp, water pump operating parameter, water pump pipe network parameter, environmental parameter, water pump vibration parameter and water pump noise parameter; the real-time water pump time sequence data is processed by data cleaning, data normalization and data encoding; the time stamp, water pump operating parameter, water pump vibration parameter and water pump noise parameter are extracted from the real-time water pump time sequence data after data cleaning, data normalization and data encoding processing, and input to the running state detection model; if the running state detection model outputs a running state of fault, an early warning signal is immediately sent to the relevant personnel; otherwise, the time stamp, water pump operating parameter, water pump pipe network parameter and environmental parameter are extracted from the real-time water pump time sequence data after data cleaning, data normalization and data encoding processing, and input to the energy-saving efficiency prediction model; the motor speed corresponding to the highest energy-saving efficiency level output by the energy-saving efficiency prediction model is taken as the control instruction to realize intelligent control of the water pump.

[0112] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary and that changes can be made in the embodiments without departing from the spirit and scope of the application.

[0113] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution should belong to the protection scope of the application.

Claims

1. An intelligent water pump control system, characterized by: It includes data acquisition module, data preprocessing module, operation status detection model building module, energy-saving efficiency prediction model building module and intelligent control module; The data acquisition module collects historical water pump time series data; The data preprocessing module performs data cleaning, data normalization, data encoding and data set construction on the data to obtain a detection time series data set and a prediction time series data set; The operating status detection model construction module calculates the attention weight based on mutual information, obtains the time series fusion feature, generates a multimodal attention dynamic graph, extracts the initial feature matrix, obtains the time series trend feature based on the trend slope, calculates the residual between the initial feature and the window mean and converts it into a mutation probability, introduces the trend slope and mutation probability to adjust the gating parameters, and obtains the operating status detection result of the water pump based on the key time series features; The energy-saving efficiency prediction model construction module obtains multi-frequency subsequences based on wavelet decomposition, extracts time series features and combines them with the attention mechanism to obtain a time feature tensor, uses nonlinear channel weights to weight the time-channel cross features to obtain a channel feature tensor, and obtains the energy-saving efficiency prediction result of the water pump based on the time-channel fusion features; The intelligent control module realizes intelligent control of the water pump according to the output of the operation status detection model and the energy-saving efficiency prediction model.

2. The intelligent water pump control system according to claim 1, characterized in that: The operation status detection model construction module is based on the detection time series data set and uses machine learning methods to build the operation status detection model of the water pump. It is equipped with a multimodal attention dynamic graph generation unit, an initial feature extraction unit, a trend perception unit, a mutation perception unit, a trend-mutation weighted gating unit and a detection unit. Specifically, it includes the following contents: Multimodal attention dynamic graph generation unit; Initial feature extraction unit: Using a graph convolutional neural network, based on the adjacency matrix of the multimodal attention dynamic graph, a graph convolution operation is performed on the time series fusion feature matrix to extract the initial feature matrix containing time series correlation information; trend sensing unit; Set the trend window size, perform linear regression on the time steps within the trend window and their corresponding initial features, calculate the trend slope, normalize the trend slope, and then concatenate the normalized result with the initial features to obtain the time series trend features containing trend information; Mutation perception unit: Set the size of the mutation detection window, calculate the residual between the initial features of the graph nodes at each time step and the mean of the initial features of the graph nodes within the mutation detection window, and convert the residual into a mutation probability based on the residual mean and standard deviation of the time steps in the detection time series dataset; Trend-mutation weighted gating unit; Detection unit: After the key features of the time series are reduced in dimension and transformed nonlinearly through the fully connected layer, the operating status detection results of the water pump at the corresponding time step are output.

3. The intelligent water pump control system according to claim 2, characterized in that: The multimodal attention dynamic graph generation unit calculates the attention weight based on the mutual information between each modal feature and the data label at each time step, performs weighted summation on the multimodal features, and obtains the temporal fusion feature of each time step; The temporal fusion feature of each time step is taken as a graph node, and the cosine similarity of the temporal fusion feature between each two graph nodes is compressed to Range, as the weight of the edge between two graph nodes, generates a multimodal attention dynamic graph.

4. The intelligent water pump control system according to claim 2, characterized in that: The trend-mutation weighted gating unit introduces trend slope and mutation probability to adjust gating parameters, controls the fusion ratio of historical features and current features, updates the hidden state of each time step based on the gating parameters, obtains the gated timing features of each time step graph node, and splices the gated timing features with the initial features to obtain the timing key features of each time step graph node.

5. The intelligent water pump control system according to claim 1, characterized in that: The energy-saving efficiency prediction model construction module is based on the prediction time series data set and uses machine learning methods to build the energy-saving efficiency prediction model of the water pump. It is equipped with an anti-enhancement unit, a multi-frequency time series fusion unit, a nonlinear channel interaction unit and a prediction unit. Specifically, it includes the following contents: Adversarial enhancement unit: Use the Wasserstein generative adversarial network WGAN-GP based on gradient penalty to perform data enhancement on the predicted time series dataset to obtain the enhanced predicted time series dataset B; Multi-frequency timing fusion unit; The enhanced prediction time series dataset is decomposed by wavelet to obtain 4 frequency subsequences. The time series features of each frequency subsequence are extracted separately using a bidirectional gated recurrent unit. Combined with the attention mechanism, the time series features of all frequency subsequences are aligned and stacked along the time axis to construct a query matrix. The time series features of all frequency subsequences are aligned and stacked along the frequency axis to construct a key matrix and a value matrix. For each time step and frequency subsequence, the attention weight is calculated by the similarity between the query vector and the key vector. The value vector is weightedly fused based on the attention weight to obtain the cross feature of each time step. The cross features of all time steps are spliced ​​in chronological order to obtain the time feature tensor M. all ; Nonlinear channel interaction unit; Prediction unit: After splicing the time feature tensor and the channel feature tensor along the feature axis, the time-channel fusion feature is obtained by linear transformation. After the time-channel fusion feature is reduced in dimension and nonlinearly transformed by the fully connected layer, the energy-saving efficiency prediction result of the water pump at the corresponding time step is output.

6. The intelligent water pump control system according to claim 5, characterized in that: The nonlinear channel interaction unit specifically includes the following contents: Calculate nonlinear channel weights; B and M all Splicing along the channel axis, constructing the residual vector after linear transformation, using radial basis function to perform nonlinear mapping on the features of each channel in the residual vector at each time step, and calculating the nonlinear association weights between different channels based on the nonlinear features after mapping, to obtain the nonlinear channel weight matrix W nl ; Time-channel cross attention; For the residual vector, the cosine similarity between the eigenvectors corresponding to each two time step-channel pairs is calculated to construct a similarity matrix. For each time step-channel pair, the five time step-channel pairs with the highest similarity to the pair are selected as the neighborhood in the similarity matrix. The eigenvectors in the neighborhood are weighted and aggregated to obtain the time-channel cross-features of each time step-channel pair, forming the time-channel cross-feature matrix R. cross ; Dynamic channel screening; use W nl R cross Perform weighting to obtain the weighted time-channel cross feature matrix ,calculate The channel attention weight of each channel in is weighted to obtain the channel feature tensor after preliminary screening.

7. The intelligent water pump control system according to claim 1, characterized in that: The data acquisition module collects historical water pump time series data; The historical water pump time series data includes timestamp, water pump operating parameters, water pump pipe network parameters, environmental parameters, water pump vibration parameters, water pump noise parameters, operating status and energy-saving efficiency level.

8. The intelligent water pump control system according to claim 1, characterized in that: The data preprocessing module performs data cleaning, data normalization, data encoding and data set construction on the data to obtain a detection time series data set and a prediction time series data set; The data set construction is to construct a detection time series data set and a prediction time series data set; the detection time series data set is constructed based on the timestamp, water pump operating parameters, water pump vibration parameters, water pump noise parameters and operating status after data cleaning, data normalization and data encoding, and the operating status is used as the data label in the detection time series data set; the prediction time series data set is constructed based on the timestamp, water pump operating parameters, water pump pipe network parameters, environmental parameters and energy-saving efficiency level after data cleaning, data normalization and data encoding, and the energy-saving efficiency level is used as the data label in the prediction time series data set.

9. The intelligent water pump control system according to claim 1, characterized in that: The intelligent control module collects real-time water pump time series data; the real-time water pump time series data includes timestamps, water pump operating parameters, water pump pipe network parameters, environmental parameters, water pump vibration parameters and water pump noise parameters; the real-time water pump time series data is cleaned, normalized and coded, and the timestamps, water pump operating parameters, water pump vibration parameters and water pump noise parameters are extracted from the processed real-time water pump time series data, and input into the operation status detection model; if the operation status output by the operation status detection model is a fault, an early warning signal is immediately sent to relevant personnel; Otherwise, the timestamp, pump operating parameters, pump network parameters and environmental parameters are extracted from the processed real-time pump time series data and input into the energy-saving efficiency prediction model. The motor speed corresponding to the highest energy-saving efficiency level output by the energy-saving efficiency prediction model is used as the control instruction to realize intelligent control of the water pump.

Images