Industrial manufacturing intelligent early warning method and system based on electric power data

Abstract

The invention relates to the technical field of electric power data analysis, in particular to an industrial manufacturing intelligent early warning method and system based on electric power data, and the method comprises the steps: obtaining industrial manufacturing electric power data, and carrying out the preprocessing of the industrial manufacturing electric power data, and obtaining electric power feature data; inputting the electric power characteristic data into a pre-trained electric power data prediction model to obtain an electric power data prediction value in a prediction time period; on the basis of the electric power data predicted value in the prediction time period, a health index function constructed in advance is used for calculation to obtain an industrial equipment health index; and comparing the health index of the industrial equipment with a preset health threshold value, and if the health index of the industrial equipment is lower than the health threshold value, sending an equipment fault alarm to a control terminal. According to the method, the dynamic mode and the long-term dependency relationship in the data are deeply mined, the health state of the equipment is comprehensively evaluated, accurate fault early warning is realized, and the accuracy and the reliability of fault early warning are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power data analysis, and in particular to an industrial manufacturing intelligent early warning method and system based on power data. Background Art

[0002] The patent application with the publication number CN116707141A discloses a method and system for analyzing power operation data. The system obtains the equipment status information of the target power equipment by the power equipment monitoring server, and performs power information feature analysis on the equipment status information to obtain the power equipment features of the target power equipment; the power equipment features are input into the trained equipment status analysis model, and the predicted operation status value of the target power equipment under each preset operation cycle is output; the power equipment monitoring server analyzes each predicted operation status value to determine the equipment status information of the target power equipment, screens out the over-threshold state involved in the operation status information, determines the address information of the power equipment in the over-threshold state, and issues an alarm prompt, which can summarize the equipment status information, timely discover the power consumption hazards in the power equipment and issue early warnings, thereby realizing the supervision, management and control of the entire power consumption process, ensuring timeliness and scientificity.

[0003] However, existing technologies have limitations in utilizing power data of industrial equipment. The rich information contained in the power data of industrial equipment cannot be fully utilized, resulting in low accuracy and reliability of fault warning. Traditional methods often rely on simple threshold judgments or limited rules, which not only limits the effectiveness of the early warning system, but also has obvious blind spots, which can easily lead to missed reports or false alarms, causing unnecessary economic losses to enterprises. In addition, existing technologies have limited ability to deal with complex interference in industrial environments, and it is difficult to effectively improve the quality of power data. Interference factors such as noise, vibration, and temperature at industrial sites seriously affect the accuracy of power data, resulting in a large amount of noise and outliers in the data, which affects subsequent data analysis and modeling effects. In addition, power data has a strong time series correlation, and a single feature analysis method is one-sided and limited, which limits the discovery of complex patterns and long-term trends in the data. Due to insufficient fault warning capabilities, existing technologies are difficult to effectively improve the operating reliability and production efficiency of industrial equipment.

[0004] In view of the shortcomings of existing technologies, there is an urgent need to provide an intelligent early warning method for industrial manufacturing to solve these problems, so as to improve the accuracy of fault warning, reduce false alarms and missed alarms, and at the same time enhance the ability to handle complex interference in the industrial environment, improve the quality of power data, and thereby improve the operating reliability and production efficiency of industrial equipment. Summary of the invention

[0005] The purpose of the present invention is to provide an industrial manufacturing intelligent early warning method and system based on power data, which can comprehensively consider the interrelated time domain characteristics and health status of industrial manufacturing power data, so as to make fault detection more comprehensive and accurate.

[0006] In a first aspect, the present invention provides an industrial manufacturing intelligent early warning method based on power data, the method comprising the following steps:

[0007] Acquire industrial manufacturing power data, and pre-process the industrial manufacturing power data to obtain power characteristic data;

[0008] Inputting the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period;

[0009] Based on the predicted value of power data within the prediction period, the health index of industrial equipment is calculated using the pre-built health index function;

[0010] The industrial equipment health index is compared with a preset health threshold, and if the industrial equipment health index is lower than the health threshold, an equipment failure alarm is sent to the control terminal.

[0011] In a second aspect, the present invention provides an industrial manufacturing intelligent early warning system based on power data, the system comprising:

[0012] A data processing module is used to obtain industrial manufacturing power data and pre-process the industrial manufacturing power data to obtain power characteristic data;

[0013] A data prediction module, used for inputting the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period;

[0014] An index fitting module is used to calculate the health index of industrial equipment based on the predicted value of power data within the prediction period using a pre-built health index function;

[0015] The early warning module is used to compare the health index of the industrial equipment with a preset health threshold, and if the health index of the industrial equipment is lower than the health threshold, an equipment failure alarm is issued to the control terminal.

[0016] The present invention provides an industrial manufacturing intelligent early warning method and system based on power data, wherein the method obtains power feature data by preprocessing industrial manufacturing power data; inputs the power feature data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period; based on the power data prediction value within the prediction period, the health index of the industrial equipment is calculated using a pre-constructed health index function; the health index of the industrial equipment is compared with a preset health threshold, and if the health index of the industrial equipment is lower than the health threshold, an equipment fault alarm is issued to the control terminal. Compared with the prior art, the method can fully mine the rich information contained in the power data of industrial equipment, deeply mine the dynamic patterns and long-term dependencies in the data, comprehensively evaluate the health status of the equipment, and achieve accurate fault early warning, avoiding the blind spots of traditional methods and the one-sidedness of single feature analysis, and can effectively overcome the complex interference of the industrial environment, greatly improving the accuracy and reliability of fault early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of an industrial manufacturing intelligent early warning method based on power data provided by an embodiment of the present invention;

[0018] Figure 2 It is a block diagram of an industrial manufacturing intelligent early warning system based on power data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0020] refer to Figure 1 , the embodiment of the present invention provides an industrial manufacturing intelligent early warning method based on power data, such as Figure 1 As shown, the method comprises the following steps:

[0021] S1. Acquire industrial manufacturing power data, and preprocess the industrial manufacturing power data to obtain power characteristic data.

[0022] The industrial manufacturing power data collected in this embodiment includes voltage data, current data, power data, power factor and harmonic data; wherein the power data includes active power, reactive power and apparent power; the harmonic data includes voltage harmonic data and current harmonic data. The following will describe in detail the acquisition method of each data. For the acquisition of voltage data and current data, this embodiment acquires the voltage data and current data of industrial equipment respectively through a voltage transformer and a current transformer; for power data, the acquisition process of the power data is specifically as follows:

[0023] Performing analog-to-digital conversion (A / D conversion) on the voltage data and the current data to obtain a digitized voltage sampling value sequence and a current sampling value sequence;

[0024] According to the voltage sampling value sequence and the current sampling value sequence, a corresponding windowed voltage sequence and a corresponding windowed current sequence are obtained by using an adaptive windowing strategy;

[0025] The windowed voltage sequence and the windowed current sequence are respectively subjected to interpolation fast Fourier transform to obtain corresponding spectrum data;

[0026] Identifying the amplitude of the corresponding fundamental frequency component according to the frequency spectrum data, recording the amplitude of the fundamental frequency component as the fundamental amplitude, and calculating the voltage fundamental effective value and the current fundamental effective value according to the fundamental amplitude;

[0027] Active power, reactive power and apparent power are calculated according to the voltage fundamental effective value and the current fundamental effective value, wherein the calculation formula of the active power is:

[0028] PY=YU·YI·cos(φ)

[0029] The calculation formula of the reactive power is:

[0030] PW=YU·YI·sin(φ)

[0031] The calculation formula of the apparent power is:

[0032] SP=YU·YI

[0033]

[0034]

[0035] Where PY is active power; PW is reactive power; SP is apparent power; φ is the power factor angle; YU is the voltage fundamental effective value; YI is the current fundamental effective value; JF is the fundamental amplitude obtained by adding a window voltage sequence; JI is the fundamental amplitude obtained by adding a window current sequence.

[0036] This embodiment uses an adaptive windowing strategy to perform adaptive windowing processing according to the voltage sampling value sequence and the current sampling value sequence to obtain a windowed voltage sequence and a windowed current sequence. The adaptive windowing strategy is specifically as follows:

[0037] Performing interpolation operations on the voltage sampling value sequence and the current sampling value sequence respectively to generate a new voltage sampling value sequence and a new current sampling value sequence with a sampling rate increased by N times;

[0038] Define M fixed-length detection windows for the new voltage sampling value sequence and the new current sampling value sequence, respectively, calculate the mean and variance of the new voltage sampling value sequence and the new current sampling value sequence in each detection window, and use the mean and variance of the new voltage sampling value sequence and the new current sampling value sequence in each detection window as the mean and variance of the detection window;

[0039] Detecting a window mutation point according to the mean and variance of the detection window, and obtaining statistical features of several sequence points before and after the window mutation point according to the position of the window mutation point in the new voltage sampling value sequence and the new current sampling value sequence; the statistical features are mean, variance or autocorrelation;

[0040] Based on the statistical feature analysis, the mutation feature of the detection window where the window mutation point is located is obtained; wherein the mutation feature includes a slow mutation and a steep mutation;

[0041] Select a rectangular window as the windowing function for the detection window with a slow-changing mutation, and enlarge the main window bandwidth parameter by 2 times;

[0042] The Blackman window is selected as the windowing function for the detection window with abrupt changes, and the main window bandwidth parameter is reduced by 0.5 times. At this point, the adaptive windowing process is completed.

[0043] In this embodiment, the step of detecting the window mutation point according to the mean and variance of the detection window includes:

[0044] The mean values ​​of the two detection windows are equal as the null hypothesis, the mean values ​​of the two detection windows are unequal as the alternative hypothesis, and any detection window is selected as the first window, and the previous detection window of the first window is selected as the second window;

[0045] According to the variance S of the first window 2 The variance S of X and the second window 2 Y, calculate the joint sample variance S 2 P, joint sample variance S 2 The calculation formula of P is:

[0046]

[0047] Where n is the number of sequence points in the first window; m is the number of sequence points in the second window;

[0048] According to the joint sample variance S 2 P, the mean of the first window and the mean of the second window Calculate the statistics Among them, the statistics The calculation formula is:

[0049]

[0050] Calculate the common degrees of freedom of the first window and the second window, and find the corresponding critical value in the t distribution table based on the pre-defined significance level and common degrees of freedom df; the calculation formula of the common degrees of freedom df is:

[0051] df = n + m - 2

[0052] The mutation point is detected based on the statistic and the critical value. If the statistic is greater than the critical value, the null hypothesis is rejected and the mutation point is detected; if the statistic is less than or equal to the critical value, the null hypothesis is accepted and it is determined that there is no mutation point.

[0053] Therefore, for the power factor, the present embodiment can calculate the power factor NV according to the ratio of the active power to the apparent power. The mathematical expression of the power factor NV is:

[0054]

[0055] For harmonic data, the voltage harmonic data and current harmonic data are obtained in the following ways:

[0056] The voltage sampling value sequence and the current sampling value sequence are processed by wavelet transform method to decompose the voltage sampling value sequence and the current sampling value sequence into fundamental wave component and each harmonic component;

[0057] Obtain the effective value of the fundamental component and the effective value of each harmonic component, and perform normalization processing;

[0058] The effective value of each harmonic component is divided by the effective value of the fundamental component to obtain the corresponding voltage harmonic data / current harmonic data.

[0059] After the industrial manufacturing power data is acquired by the above method, the present embodiment obtains power characteristic data by preprocessing the industrial manufacturing power data. The specific steps include:

[0060] Using the improved isolation forest algorithm to identify outlier data points in industrial manufacturing power data, and removing the outlier data points to obtain the underlying power data;

[0061] All underlying power data are synchronized with timestamps and arranged in chronological order to form underlying power time series data;

[0062] Using a predefined time window, different types of underlying power data in the underlying power time series data are converted into corresponding underlying power data time series;

[0063] The mutual correlation time domain features of any two underlying power data time series are calculated by using the mutual correlation function, and the mutual correlation time domain features are normalized to obtain power feature data; the mutual correlation time domain features include the maximum mutual correlation value, the minimum mutual correlation value, the mutual correlation mean value and the mutual correlation autocorrelation value; specifically, for the acquisition process of the mutual correlation time domain features, this embodiment defines a time window of a preset length, and constructs the corresponding underlying power data time series for all types of data in the underlying power data, for any two underlying power data time series x and y, this embodiment defines a mutual correlation function, calculates the mutual correlation value, and uses the statistics of the mutual correlation value as the mutual correlation time domain feature, the statistics include the mutual correlation maximum value, the mutual correlation minimum value, the mutual correlation mean value and the mutual correlation autocorrelation value, in this embodiment, the mathematical expression of the mutual correlation function is:

[0064]

[0065] In the formula, R xy (τ) is the cross-correlation value between the underlying power data time series x and the underlying power data time series y at a time lag τ; τ is a predefined time lag, and the interval of the time lag τ is [-max(A, B)+1, max(A, B)-1]. It should be noted that when the time lag τ is greater than or equal to 0, the cross-correlation function is R xy (τ); when the time lag τ is less than 0, the cross-correlation function is R xy (τ) = R xy (-τ); A is the length of the underlying power data time series x; B is the length of the underlying power data time series y; X(r+τ) is the value of the underlying power data time series x at τ time units after time r, and Y(r) is the value of the underlying power data time series y at time r; μ x is the mean of the underlying power data time series x; μ y is the mean of the underlying power data time series y;

[0066] The calculation formula for the cross-correlation mean is:

[0067]

[0068] In the formula, mean(R xy ) is the cross-correlation mean;

[0069] The formula for calculating the autocorrelation value of the cross-correlation is:

[0070]

[0071] In the formula, mz(R xy ) is the autocorrelation value of the cross-correlation.

[0072] For the above preprocessing process, when the improved isolation forest algorithm is used to identify outlier data points in industrial manufacturing power data, the identification method of outlier data points includes:

[0073] In constructing the improved isolation forest algorithm, the parameters of the improved isolation forest are defined. The parameters of the improved isolation forest include the number of isolation trees, the subsampling size, and the limiting tree height hyperparameter;

[0074] For the i-th isolation tree, a number of samples are randomly sampled from the industrial manufacturing power data, which are recorded as the sample set M;

[0075] For each sample in the sample set M, this embodiment randomly selects a feature q, and for the feature q, randomly generates a threshold p, so that the threshold p falls between the minimum value and the maximum value of the feature q;

[0076] A counter C(i, ku) is defined for each isolation tree. The counter is used to count the number of times the kuth feature is used as a split node. Initially, the values ​​of all counters are set to 0. In the process of constructing the i-th isolation tree, each time the jth feature is selected as a split node, the counter C(i, ku) is increased by 1. After all isolation trees are constructed, the total number of splits of each feature is calculated for all isolation trees, and the total number of splits of each feature is normalized to obtain the isolation weight c(ku) of the feature. The isolation weight reflects the importance of the feature in the splitting process. When randomly selecting feature q, the isolation weight c(ku) is used as the probability of selecting the eth feature, so as to increase the diversity of feature selection.

[0077] According to the comparison result between the feature q and the threshold p, the present embodiment divides the sample g into the left child node or the right child node, and repeats this process until the node contains only one sample or the height of the isolation tree reaches the preset limit tree height super parameter;

[0078] For each sample, record the path length required for sample g to reach the leaf node, and construct U isolation trees by repeating the above process, and finally form an improved isolation forest. Through the above isolation forest construction method, this embodiment can construct an isolation forest model that can effectively identify outliers. At the same time, through the introduction of isolation weights, the model's ability to identify feature importance is enhanced;

[0079] For sample g, calculate its average path length c(g) in all U isolation trees and convert the average path length into anomaly score Among them, c(Z) is the average path length when the data size of industrial manufacturing power data is Z; it should be noted that the path length is used to measure the degree of isolation of a sample point in the tree. For an isolation tree, the calculation formula of the average path length is:

[0080]

[0081] H(Z-1)=ln(Z-1)+0.5772156649

[0082] This embodiment sets an anomaly score threshold and marks samples whose anomaly scores s(g) are greater than the anomaly score threshold as outlier data points.

[0083] S2. Input the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period.

[0084] In this embodiment, the basic architecture of the power data prediction model includes an input layer, a CNN extraction layer, an LSTM encoding layer, a fully connected layer and an output layer connected in sequence, wherein the network structure of the CNN extraction layer is a stack of a one-dimensional convolutional layer, a pooling layer and an activation layer; the network structure of the LSTM encoding layer is an LSTM layer and an attention layer; specifically, the input layer takes the sequence composed of power feature data as input, the CNN extraction layer uses a one-dimensional convolutional layer to extract the local feature sequence of the sequence composed of the input power feature data, and the LSTM encoding layer inputs the local feature sequence extracted by CNN into the LSTM to capture the long-term dependency of the sequence, that is, to obtain the hidden state of the LSTM; the hidden state output of the LSTM is used as a feature, mapped to a predicted value through a fully connected layer, and the output layer outputs the predicted value of the power data for the next t time steps.

[0085] In this embodiment, the training method of the pre-trained power data prediction model includes the following steps:

[0086] Utilize historical power characteristic data within a preset historical period, and arrange the historical power characteristic data according to the collected timestamps to form a historical characteristic sequence;

[0087] The training target of the CNN extraction layer is defined as capturing the local feature sequence of the historical feature sequence, that is, taking the local feature sequence that captures the historical feature sequence as the training target of the extraction layer, according to the pre-defined CNN extraction layer reconstruction loss, the CNN extraction layer is trained in batches using the historical feature sequence until the reconstruction loss is minimized, and the convolution kernel parameters of the CNN extraction layer are output; it should be noted that the CNN extraction layer reconstruction loss requires that the output of the CNN extraction layer can reconstruct the original input historical feature sequence;

[0088] The loss of the LSTM encoding layer is defined as the sequence-to-sequence prediction loss, and the long-term dependency of the local feature sequence is captured as the training target of the LSTM encoding layer, that is, the hidden state of the LSTM is obtained, and the LSTM encoding layer is trained in batches using the historical feature sequence until the sequence-to-sequence prediction loss is minimized, and the weight parameter of the LSTM encoding layer at this time is output; it should be noted that in the machine learning task of sequence-to-sequence prediction, the model needs to convert one sequence into another sequence. For the sequence-to-sequence prediction task, this embodiment can be implemented by using a loss function such as a cross entropy loss to measure the difference between the sequence generated by the model and the target sequence in the sequence-to-sequence prediction. Those skilled in the art can select a sequence-to-sequence prediction loss function according to the specific implementation situation, which is not limited to the embodiment of the present invention;

[0089] The infrastructure loss function is introduced. Based on the convolution kernel parameters of the CNN extraction layer and the weight parameters of the LSTM encoding layer, the infrastructure of the power data prediction model is trained in batches using the historical feature sequence to obtain the trained power data prediction model. The mathematical expression of the infrastructure loss function is:

[0090] Loss=∑(w f ·l f )+λ·R

[0091] in,

[0092]

[0093] Where Loss is the basic architecture loss function; w f is the loss weight coefficient used in the fth batch; l f is the loss term for the prediction of the fth batch; λ is the regularization coefficient; R is the regularization term; y f is the true value of the sequence point on the historical feature sequence of the fth batch; is the predicted value output by the power data prediction model of the fth batch; α is the square loss weight; β is the quantile loss weight; ρ is the quantile; I() is the indicator function.

[0094] In this embodiment, an optimizer can be defined, and the optimizer is used to optimize the parameters of the infrastructure during the training process of the power data prediction model. The historical feature sequences are input into the power data prediction model infrastructure in batches. If the function value of the loss function does not decrease for L consecutive batches, the training is stopped and the parameters of the power data prediction model infrastructure are saved.

[0095] S3. Based on the predicted value of power data within the prediction period, the health index of industrial equipment is calculated using a pre-built health index function.

[0096] In this embodiment, the construction process of the health indicator function is:

[0097] Define power data factors of industrial manufacturing power data, and obtain historical industrial manufacturing power data samples for each power data factor within a preset time period;

[0098] Calculate the health status membership of each power data factor according to the health status membership function;

[0099] According to the health status membership of each power data factor and the historical industrial manufacturing power data samples, the corresponding single factor health index is calculated;

[0100] Calculate the entropy value of each power data factor, and obtain the corresponding power data factor weight according to the entropy value of each power data factor;

[0101] According to the single factor health index of each power data factor and the power data factor weight, a health index function is constructed; in this embodiment, the mathematical expression of the health index function is:

[0102]

[0103] in,

[0104]

[0105] Where HG is the health index of industrial equipment; HF j is the single factor health index of the jth power data factor; w j is the power data factor weight of the jth power data factor; e j is the entropy value of the jth power data factor; J is the total number of power data factors; G is the total number of historical industrial manufacturing power data samples; D j,k It is the value of the j-th power data factor in the k-th historical industrial manufacturing power data sample.

[0106] The calculation formula of the single factor health index is:

[0107]

[0108] in,

[0109]

[0110] In the formula, θ j (D j ) is the health status membership; D j is the jth power data factor; a j is the lower limit parameter required to fit the jth power data factor; b j is the intermediate limit parameter required to fit the jth power data factor; c j It is the upper limit parameter required to fit the j-th power data factor; it should be noted that the upper limit parameter, middle limit parameter and lower limit parameter required to fit the power data factor refer to the threshold parameters set for the power data factor in the process of power data analysis and model construction, such as: the lower limit parameter refers to the minimum value or minimum effective value of a certain power data factor; corresponding to the lower limit parameter, the upper limit parameter defines the maximum value or maximum effective value of a certain power data factor; the middle limit parameter can be the central value or average value of the power data factor. Those skilled in the art can set the corresponding parameter values ​​according to the specific implementation situation, which is not limited to the embodiments of the present invention.

[0111] S4. Compare the health index of the industrial equipment with a preset health threshold. If the health index of the industrial equipment is lower than the health threshold, send an equipment failure alarm to the control terminal.

[0112] The embodiment of the present invention provides an intelligent early warning method for industrial manufacturing based on power data, wherein the method obtains power characteristic data by preprocessing the industrial manufacturing power data; inputs the power characteristic data into a pre-trained power data prediction model to obtain the power data prediction value within the prediction period; based on the power data prediction value within the prediction period, the health index of the industrial equipment is calculated using a pre-built health index function; the health index of the industrial equipment is compared with a preset health threshold, and if the health index of the industrial equipment is lower than the health threshold, an equipment fault alarm is issued to the control terminal. Compared with the prior art, the method can fully mine the rich information contained in the power data of industrial equipment, and comprehensively evaluate the health status of the equipment by deeply mining the dynamic patterns and long-term dependencies in the data, and can more accurately distinguish the boundaries between normal and fault states, reduce the large errors caused by improper setting of the health threshold, can adapt to the complexity of the industrial environment, improve the accuracy and reliability of fault warning, and greatly improve the operating reliability and production efficiency of industrial equipment.

[0113] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0114] In one embodiment, Figure 2 As shown, an embodiment of the present invention provides an industrial manufacturing intelligent early warning system based on power data, the system comprising:

[0115] The data processing module 101 is used to obtain industrial manufacturing power data and pre-process the industrial manufacturing power data to obtain power characteristic data;

[0116] The data prediction module 102 is used to input the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period;

[0117] An index fitting module 103 is used to calculate the health index of the industrial equipment using a pre-built health index function based on the power data prediction value within the prediction period;

[0118] The early warning module 104 is used to compare the health index of the industrial equipment with a preset health threshold, and if the health index of the industrial equipment is lower than the health threshold, send an equipment failure alarm to the control terminal.

[0119] For the specific definition of an industrial manufacturing intelligent early warning system based on power data, please refer to the above-mentioned definition of an industrial manufacturing intelligent early warning method based on power data, which will not be repeated here. A person of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0120] The embodiment of the present invention provides an industrial manufacturing intelligent early warning system based on power data, wherein the system pre-processes the industrial manufacturing power data through a data processing module to obtain power characteristic data; inputs the power characteristic data into a pre-trained power data prediction model through a data prediction module to obtain a power data prediction value within a prediction period; calculates the health index of industrial equipment by using a pre-constructed health index function and the power data prediction value within the prediction period through an index fitting module; compares the health index of industrial equipment with a preset health threshold through an early warning module, and if the health index of industrial equipment is lower than the health threshold, sends an equipment fault alarm to a control terminal. Compared with the prior art, the system can fully mine the rich information contained in the power data of industrial equipment, and by deeply mining the dynamic patterns and long-term dependencies in the data, it can comprehensively evaluate the health status of the equipment, thereby more accurately distinguishing the boundaries between normal and fault states, reducing the large errors caused by improper setting of the health threshold, being able to adapt to the complexity of the industrial environment, improving the accuracy and reliability of fault early warning, and greatly improving the operating reliability and production efficiency of industrial equipment.

[0121] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. An industrial manufacturing intelligent early warning method based on power data, characterized in that: The following steps are involved: Acquire industrial manufacturing power data, and pre-process the industrial manufacturing power data to obtain power characteristic data; Inputting the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period; Based on the predicted value of power data within the prediction period, the health index of industrial equipment is calculated using the pre-built health index function; The industrial equipment health index is compared with a preset health threshold, and if the industrial equipment health index is lower than the health threshold, an equipment failure alarm is sent to the control terminal.

2. The industrial manufacturing intelligent early warning method based on power data according to claim 1 is characterized in that: The industrial manufacturing power data includes voltage data, current data, power data, power factor and harmonic data. The power data includes active power, reactive power and apparent power. The acquisition process of the power data is specifically as follows: Perform analog-to-digital conversion on the voltage data and the current data to obtain a digitized voltage sampling value sequence and a current sampling value sequence; According to the voltage sampling value sequence and the current sampling value sequence, a corresponding windowed voltage sequence and a corresponding windowed current sequence are obtained by using an adaptive windowing strategy; The windowed voltage sequence and the windowed current sequence are respectively subjected to interpolation fast Fourier transform to obtain corresponding spectrum data; Identify the corresponding fundamental wave amplitude according to the frequency spectrum data, and calculate the voltage fundamental wave effective value and the current fundamental wave effective value according to the fundamental wave amplitude; Active power, reactive power and apparent power are calculated according to the voltage fundamental effective value and the current fundamental effective value.

3. The industrial manufacturing intelligent early warning method based on power data as claimed in claim 2 is characterized in that: The adaptive windowing strategy is specifically as follows: Performing interpolation operations on the voltage sampling value sequence and the current sampling value sequence respectively to generate a new voltage sampling value sequence and a new current sampling value sequence; Defining detection windows of new voltage sampling value sequences and new current sampling value sequences, and taking the mean and variance of the new voltage sampling value sequences and new current sampling value sequences in each detection window as the mean and variance of the detection window; Detecting a window mutation point according to the mean and variance of the detection window, and obtaining statistical features of several sequence points before and after the window mutation point according to the position of the window mutation point in the new voltage sampling value sequence and the new current sampling value sequence; Based on the statistical feature analysis, the mutation feature of the detection window where the window mutation point is located is obtained; wherein the mutation feature includes a slow mutation and a steep mutation; The rectangular window is selected as the windowing function for the detection window with a slow-changing mutation, and the Blackman window is selected as the windowing function for the detection window with a steep-changing mutation.

4. The industrial manufacturing intelligent early warning method based on power data according to claim 1 is characterized in that: The step of preprocessing the industrial manufacturing power data to obtain power characteristic data includes: Using improved isolation forests to identify outlier data points in industrial manufacturing power data, and removing the outlier data points to obtain underlying power data; All underlying power data are synchronized with timestamps and arranged in chronological order to form underlying power time series data; Using a predefined time window, different types of underlying power data in the underlying power time series data are converted into corresponding underlying power data time series; The cross-correlation function is used to calculate the cross-correlation time domain features of any two underlying power data time series, and the cross-correlation time domain features are normalized to obtain power feature data; the cross-correlation time domain features include the maximum cross-correlation value, the minimum cross-correlation value, the cross-correlation mean and the cross-correlation autocorrelation value.

5. The industrial manufacturing intelligent early warning method based on power data according to claim 1 is characterized in that: The power data prediction model includes an input layer, a CNN extraction layer, an LSTM encoding layer, a fully connected layer and an output layer connected in sequence, wherein the network structure of the CNN extraction layer is a stack of a one-dimensional convolution layer, a pooling layer and an activation layer; the network structure of the LSTM encoding layer is an LSTM layer and an attention layer; the training method of the power data prediction model includes: Using historical power characteristic data within a preset historical period, the historical power characteristic data is arranged according to the timestamp of collection to form a historical characteristic sequence; Taking the local feature sequence that captures the historical feature sequence as the training target of the extraction layer, according to the pre-defined CNN extraction layer reconstruction loss, the CNN extraction layer is trained in batches using the historical feature sequence until the reconstruction loss is minimized, and the convolution kernel parameters of the CNN extraction layer are output; The loss of the LSTM encoding layer is defined as the sequence-to-sequence prediction loss. The long-term dependency of the local feature sequence is captured as the encoding layer training objective. The LSTM encoding layer is trained in batches until the sequence-to-sequence prediction loss is minimized, and the weight parameters of the LSTM encoding layer are output. The infrastructure loss function is introduced, and based on the convolution kernel parameters of the CNN extraction layer and the weight parameters of the LSTM encoding layer, the power data prediction model is trained in batches using the historical feature sequence to obtain the trained power data prediction model.

6. The industrial manufacturing intelligent early warning method based on power data according to claim 1, characterized in that: The construction process of the health indicator function is: Define power data factors of industrial manufacturing power data, and obtain historical industrial manufacturing power data samples for each power data factor within a preset time period; Calculate the health status membership of each power data factor according to the health status membership function; According to the health status membership of each power data factor and the historical industrial manufacturing power data samples, the corresponding single factor health index is calculated; Calculate the entropy value of each power data factor, and obtain the corresponding power data factor weight according to the entropy value of each power data factor; According to the single factor health index of each power data factor and the power data factor weight, a health indicator function is constructed.

7. An industrial manufacturing intelligent early warning system based on power data, characterized in that: The system comprises: A data processing module is used to obtain industrial manufacturing power data and pre-process the industrial manufacturing power data to obtain power characteristic data; A data prediction module, used for inputting the power characteristic data into a pre-trained power data prediction model to obtain a power data prediction value within a prediction period; An index fitting module is used to calculate the health index of industrial equipment based on the predicted value of power data within the prediction period using a pre-built health index function; The early warning module is used to compare the health index of the industrial equipment with a preset health threshold, and if the health index of the industrial equipment is lower than the health threshold, an equipment failure alarm is issued to the control terminal.

8. The industrial manufacturing intelligent early warning system based on power data as claimed in claim 7, characterized in that: The preprocessing of industrial manufacturing power data to obtain power characteristic data specifically includes: Using improved isolation forests to identify outlier data points in industrial manufacturing power data, and removing the outlier data points to obtain underlying power data; All underlying power data are synchronized with timestamps and arranged in chronological order to form underlying power time series data; Using a predefined time window, different types of underlying power data in the underlying power time series data are converted into corresponding underlying power data time series; The cross-correlation function is used to calculate the cross-correlation time domain features of any two underlying power data time series, and the cross-correlation time domain features are normalized to obtain power feature data; the cross-correlation time domain features include the maximum cross-correlation value, the minimum cross-correlation value, the cross-correlation mean and the cross-correlation autocorrelation value.

9. The industrial manufacturing intelligent early warning system based on power data as claimed in claim 7, characterized in that: The power data prediction model includes an input layer, a CNN extraction layer, an LSTM encoding layer, a fully connected layer and an output layer connected in sequence, wherein the network structure of the CNN extraction layer is a stack of a one-dimensional convolution layer, a pooling layer and an activation layer; the network structure of the LSTM encoding layer is an LSTM layer and an attention layer; the training method of the power data prediction model includes: Using historical power characteristic data within a preset historical period, the historical power characteristic data is arranged according to the timestamp of collection to form a historical characteristic sequence; Taking the local feature sequence that captures the historical feature sequence as the training target of the extraction layer, according to the pre-defined CNN extraction layer reconstruction loss, the CNN extraction layer is trained in batches using the historical feature sequence until the reconstruction loss is minimized, and the convolution kernel parameters of the CNN extraction layer are output; The loss of the LSTM encoding layer is defined as the sequence-to-sequence prediction loss. The long-term dependency of the local feature sequence is captured as the encoding layer training objective. The LSTM encoding layer is trained in batches until the sequence-to-sequence prediction loss is minimized, and the weight parameters of the LSTM encoding layer are output. The infrastructure loss function is introduced, and based on the convolution kernel parameters of the CNN extraction layer and the weight parameters of the LSTM encoding layer, the power data prediction model is trained in batches using the historical feature sequence to obtain the trained power data prediction model.

10. The industrial manufacturing intelligent early warning system based on power data according to claim 7, characterized in that: The construction process of the health indicator function is: Define power data factors of industrial manufacturing power data, and obtain historical industrial manufacturing power data samples for each power data factor within a preset time period; Calculate the health status membership of each power data factor according to the health status membership function; According to the health status membership of each power data factor and the historical industrial manufacturing power data samples, the corresponding single factor health index is calculated; Calculate the entropy value of each power data factor, and obtain the corresponding power data factor weight according to the entropy value of each power data factor; According to the single factor health index of each power data factor and the power data factor weight, a health indicator function is constructed.

Images