A complex scene multi-source heterogeneous data collection and monitoring diagnosis method and device

Abstract

The present invention discloses a method and device for multi-source heterogeneous data acquisition and monitoring diagnosis in complex scenarios, the method comprising: acquiring multi-source heterogeneous data of a monitored diagnostic object, and performing data cleaning, data preprocessing, feature extraction and fusion on the data to obtain a feature data set; using the feature data set and the data after data cleaning, constructing and training a causal decoupling representation learning model based on input-output drive to extract the decoupling representation of the monitored diagnostic object and infer the causal relationship between variables, wherein the causal decoupling representation learning model comprises an exogenous control transformation network and a dual-view evolutionary modeling and causal inference network, and using the trained causal decoupling representation learning model as a heterogeneous signal monitoring and diagnosis model to perform state monitoring on the monitored diagnostic object and diagnose the root cause of its fault. The present invention can realize data acquisition, state monitoring and fault cause diagnosis in an integrated manner, save costs, and has broad application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment status monitoring and fault diagnosis, and specifically relates to status monitoring and diagnosis of industrial equipment running in complex industrial scenarios, and more particularly to a method and device for collecting, monitoring and diagnosing multi-source heterogeneous data in complex scenarios. Background Art

[0002] In the fields of electricity, water conservancy, mining, transportation, etc., there are a large number of complex industrial scenarios with harsh environments, high risks, numerous equipment and large distribution spans, such as thermal power plants, nuclear power plants and other high-temperature, high-pressure and noisy industrial scenarios. Once the industrial equipment running in such scenarios fails, it may cause serious economic losses and safety hazards. Therefore, it is very important to conduct effective status monitoring and fault cause diagnosis of industrial equipment in such scenarios.

[0003] For the multi-source data status monitoring and fault cause diagnosis tasks of industrial equipment in complex industrial scenarios, there are currently two mainstream implementation schemes based on hardware equipment: monitoring and diagnosis systems based on industrial computers and monitoring and diagnosis systems based on embedded systems. Both have their advantages and disadvantages. Among them, the monitoring and diagnosis system based on industrial computers has high performance, reliability and scalability, can adapt to various sensors, and can complete complex data analysis tasks. However, the system is expensive and large in size, and is not suitable for space-constrained environments. In contrast, there are currently few studies on monitoring and diagnosis systems based on embedded systems. This method has obvious advantages in cost and deployment flexibility, and can quickly respond to changes in sensor data and has strong real-time performance. However, existing monitoring and diagnosis solutions based on embedded systems generally have problems such as incompatibility with industrial field sensors, few channels, poor scalability, and no support for collecting high-frequency data, which greatly limits its application in industrial scenarios.

[0004] In order to more efficiently complete the tasks of equipment monitoring and diagnosis in complex industrial scenarios, it is necessary to explore a more flexible, economical and reliable multi-source data acquisition and monitoring hardware solution to meet the increasingly complex requirements of equipment monitoring and diagnosis tasks.

[0005] In the actual application of algorithm deployment of multi-source data acquisition and monitoring diagnosis solutions, in addition to some systems without external intervention, continuous systems regulated by discrete set values ​​are often encountered. Analyzing the causal relationships in these systems is crucial to understanding their operating principles. The extracted causal relationships can also be further applied to some downstream tasks, such as diagnosing the root causes of failures in industrial equipment.

[0006] Despite the promising prospects of time series causal discovery, most existing methods focus on studying causal relationships without external inputs, where external inputs specifically refer to factors that can change the statistical characteristics of the original time series. However, there are few isolated systems in the real world, and the research objects are usually affected by exogenous factors such as discrete set value adjustments, which limits the expansion and application of methods in practical scenarios. In addition, the ideal assumption that most current methods are based on, that is, time series data only contains one type of variable, such as variables with continuous values, is easily not true in the real world. In fact, a time series can be a mixture of continuous variables, discrete variables, and categorical variables, called a mixed time series. A specific example: In three-phase flow equipment in the industrial field, as an external input, there are only four set values ​​for air flow rate. Therefore, how to model the impact of discrete exogenous sources on continuous system observations to extract time series causal relationships is still a problem that needs to be solved.

[0007] The present invention provides a method and device for multi-source data collection and monitoring and diagnosis in complex industrial scenarios to solve the above problems. Summary of the invention

[0008] The purpose of the present invention is to provide a method and device for collecting, monitoring and diagnosing multi-source heterogeneous data in complex scenarios in view of the deficiencies in the prior art.

[0009] The objective of the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides a complex scene multi-source heterogeneous data collection and monitoring and diagnosis method, comprising the following steps:

[0010] (1) Collect multi-source heterogeneous data of the object to be monitored and diagnosed;

[0011] (2) Perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set;

[0012] (3) Using the feature data set obtained in step (2) and the data after data cleaning, construct a causal decoupling representation learning model based on input-output drive to extract the decoupling representation of the monitored diagnostic object and infer the causal relationship between variables, wherein the causal decoupling representation learning model includes an exogenous control transformation network and a dual-view evolutionary modeling and causal inference network; and train the causal decoupling representation learning model. During the training process, the model parameters are adjusted with minimizing the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as a heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored diagnostic object and diagnose the root cause of its fault;

[0013] (4) Input the normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and calculate the control limit of the square prediction error based on the prediction value corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, calculate the square prediction error based on the true value, and judge whether it exceeds the limit based on the calculated square prediction error and the control limit of the square prediction error to determine the fault diagnosis result; if the fault diagnosis result determines that the monitored diagnostic object has a fault, use the fault data and the heterogeneous signal monitoring and diagnosis model to obtain the cause-effect diagram of the variable and infer the root cause of the fault.

[0014] Furthermore, the multi-source heterogeneous data include discrete signals and continuous signals, wherein the discrete signal is a control signal whose input is a finite number of set values, and the discrete signal includes a rotational speed and a flow rate signal; the continuous signal is an output signal that changes under the influence of the set value input, and the continuous signal includes sound, vibration, voltage, current, and temperature signals.

[0015] Furthermore, the data cleaning specifically includes: using median filtering, mean filtering, Gaussian filtering or Kalman filtering method to clean the collected multi-source heterogeneous data to remove noise and outliers in the multi-source heterogeneous data;

[0016] The data preprocessing specifically includes: performing pre-emphasis, frame division, and windowing operations on the multi-source heterogeneous data after data cleaning;

[0017] The feature extraction and fusion specifically include: extracting the Mel-frequency cepstrum coefficients of the sound signal, the time domain, frequency domain and time-frequency domain features of the vibration signal, and the time domain features of the remaining low-frequency signals, and using a fusion method to fuse them to form a feature data set, wherein the fusion method includes feature splicing, weighted fusion, and adaptive fusion method.

[0018] Furthermore, the exogenous control transformation network specifically includes: comprehensively utilizing the feature data set and the discrete signal data and continuous signal data after data cleaning, and fully modeling the exogenous response of the input-output system by constructing a conditionally reversible mapping between discrete input and continuous control embedding; utilizing the feature data set as the conditional input of the exogenous control transformation network to restore the continuous control embedding for predicting the system input response, and at the same time reconstructing the continuous control embedding back to the original discrete input; the exogenous control transformation network introduces a control condition reference mechanism, uses internal observations as conditional inputs to guide the learning of the continuous control embedding, and expresses the different control effects of finite discrete values ​​under different system operating states;

[0019] The dual-perspective evolutionary modeling and causal inference network specifically includes: separately modeling the self-driven evolution and the controlled evolutionary processes to obtain decoupled representations containing different evolutionary information; designing joint prediction and independent prediction networks, using two evolutionary decoupling representations to predict the future values ​​of internal variables after evolutionary coupling and decoupling, respectively, iteratively updating network parameters according to the residual between the predicted value and the true value, and inferring the causal relationship between variables.

[0020] Furthermore, the step (3) specifically includes the following steps:

[0021] (3.1) Clean the data According to the discrete set value signal and the internal continuous signal, we can get ,in Indicates the total length The mixed time series include The observed output and External input; output variable are all continuous output values, input variables are all discrete set values. Represents the first continuous output value, Indicates the second consecutive output value, represents the pth consecutive output value, represents the first discrete input variable, represents the second discrete input variable, represents the qth discrete input variable;

[0022] (3.2) Given the length , the number of variables is A discrete input sequence fragment ,exist Under controlled conditions, the transformation network is obtained by exogenous control Embedding of continuous control of discrete variables ,in is a finite set, The condition is the feature dataset The conditional vector obtained by convolution, is the dimension of controlling embedding, t represents the position of the sequence fragment in the entire sequence, ; The network of 1×1 convolution and affine transformation cascade is used as the structure of the exogenous control transformation network. Its forward process obtains continuous control embedding from discrete input, and its expression is:

[0023]

[0024]

[0025]

[0026] in, , and , They are affine transformations The input and output vectors, represents the exponential function with base natural logarithm, represents the 4 sub-neural networks in the affine transformation, represents a 1×1 convolution process; the input and output dimensions of the exogenous control transformation network remain , The dimension of is half that of D; represents element-wise multiplication; represents a set of multidimensional potential random variables introduced; in the reverse process of the exogenous control transformation network, the input of each sub-neural network in the affine transformation remains unchanged, Through the inverse process of affine transformation and 1×1 convolution transformation Computational Reconstruction and , whose expression is:

[0027]

[0028]

[0029]

[0030] in, and They are and Reconstruction, is additive noise;

[0031] (3.3) The parameters of the forward and reverse processes of the exogenous control transformation network are shared by Indicates; alternately perform forward and reverse process calculations, accumulating The gradient is then used to update the parameters; the forward process extracts the control embedding for subsequent prediction, and the reverse process reconstructs the external input and hidden variables ,for and , respectively calculate the mean square error and the maximum mean difference , the mean square error and the maximum mean difference The sum is taken as the reconstruction loss , whose expression is:

[0032]

[0033] Among them, the kernel function , h is a constant, and They are and The i-th row vector of ;

[0034] (3.4) Dual-perspective evolutionary modeling and causal inference The network includes a self-driven evolutionary network and a controlled evolutionary network. Both the self-driven evolutionary network and the controlled evolutionary network use multi-layer perceptron as the backbone network. Different evolution laws of internal variables are established For structurally similar networks with non-shared parameters; each multilayer perceptron has a learned Dimensional self-driven evolution soft mask vector or Dimensional Controlled Evolution of Soft Mask Vectors , to extract sparse causal relationships; the time-delayed inputs of the self-driving evolution network and the controlled evolution network are firstly and Masked, and then mapped to the feature space through a multi-layer perceptron to capture the corresponding self-driving prediction features and controlled predictive features , expressed as:

[0035]

[0036]

[0037] in, and They represent the self-driving prediction characteristics and controlled prediction characteristics at the current time t respectively; Represents the observed value of the variable in L historical moments before time t; represents the control embedding obtained by transforming the external control of the input records at L historical moments; and are the network parameters of the self-driven evolution network and the controlled evolution network respectively; represents the jth self-driving evolution network, represents the jth controlled evolution network; by adding penalty loss , to use the learned sparse mask vector to reveal causal relationships, the penalty loss The expression is:

[0038]

[0039] in, It is a penalty loss middle The hyperparameters of It is a penalty loss middle Hyperparameters of

[0040] (3.5) Design a discriminator Used to distinguish self-driving prediction features and control prediction features , the input stream of the discriminator is the self-driving prediction feature and control prediction features The ordered splicing of and negative sequence characteristics , whose expressions are:

[0041]

[0042]

[0043] Among them, M represents and The dimension of the discriminator is ; the cross entropy loss of the discriminator training is calculated using the input stream of the discriminator, which is expressed as:

[0044]

[0045] in, represents the discriminant loss function, It is the probability estimate of the feature sample belonging to the positive sequence rather than the negative sequence; based on the cross entropy loss of discriminator training, a decoupled penalty loss is designed To approximate the joint distribution The product of the marginal distribution The KL divergence of is calculated as:

[0046]

[0047] in, and Represent the sample distribution of self-driven prediction features and controlled prediction features, respectively. represents KL divergence;

[0048] (3.6) Self-driving prediction features and control prediction features After splicing, it is input into the joint prediction network with ResNet as the skeleton to obtain the joint prediction result of the internal variable at the current time t; assuming that the value change of the internal variable at time t before and after adding the external source is , then the state of the system after decoupling is It is expressed as:

[0049]

[0050] in, represents the value of the internal variable at time t before adding the exogenous source; the self-driven prediction feature and control prediction features They are respectively input into independent prediction networks with ResNet as the skeleton, and the and The predicted estimated value of is superimposed to obtain the final predicted value of the internal variable; and Respectively represent the functions represented by the joint prediction network and the independent prediction network, As the network parameters corresponding to the two functions, the prediction loss function The calculation formula is:

[0051]

[0052] in, and They are joint prediction loss and independent prediction loss respectively; It means to concatenate two vectors into a new vector; yes Parameters, yes Parameters;

[0053] (3.7) According to the reconstruction loss , penalty loss , Decoupled Penalty Loss , prediction loss function Calculate the total loss function, expressed as:

[0054]

[0055] in, represents the total loss function, express weight parameters; taking minimizing the total loss function as the optimization goal, adjust the parameters of the exogenous control transformation network, dual-view evolutionary modeling and causal inference network, joint prediction network, independent prediction network and discriminator to obtain the trained causal decoupling representation learning model, and use it as a heterogeneous signal monitoring and diagnosis model to be used in subsequent steps to monitor the state of the monitored and diagnosed object and diagnose the root cause of its fault.

[0056] A second aspect of an embodiment of the present invention provides a complex scene multi-source heterogeneous data collection and monitoring and diagnosis device, which is used to implement the complex scene multi-source heterogeneous data collection and monitoring and diagnosis method mentioned above, and the device includes:

[0057] A multi-source data acquisition module is used to collect multi-source heterogeneous data of the object to be monitored and diagnosed. The multi-source data acquisition module includes a multi-source signal conditioner, an analog-to-digital converter and a Raspberry Pi, which together realize multi-source heterogeneous data acquisition of discrete set value signals and continuous output signals of the object to be monitored and diagnosed. The multi-source signal conditioner is connected to a variety of sensors on the device to be monitored, and is used to provide corresponding standard sensor interfaces for the various sensors, and simultaneously stimulate the sensors, isolate and protect, and condition the signals output by the sensors; the analog-to-digital converter is connected to the output end of the multi-source signal conditioner, and is used to convert the analog signal output after conditioning into a digital signal; the Raspberry Pi is connected to the analog-to-digital converter, and is used to control the analog-to-digital converter to perform synchronous acquisition of multi-source heterogeneous data;

[0058] The multi-source data processing module is used to perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set;

[0059] The heterogeneous signal monitoring and diagnosis model acquisition module is used to use the feature data set and the data after data cleaning to build an input-output driven causal decoupling representation learning model to extract the decoupling representation of the monitored and diagnosed object and infer the causal relationship between variables, wherein the causal decoupling representation learning model includes an exogenous control transformation network and a dual-view evolutionary modeling and causal inference network; and train the causal decoupling representation learning model. During the training process, the model parameters are adjusted with minimizing the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as the heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored and diagnosed object and diagnose the root cause of its failure; and

[0060] The fault monitoring and diagnosis module is used to input normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and calculate the control limit of the square prediction error according to the prediction value corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, calculate the square prediction error in combination with the true value, and judge whether it exceeds the limit based on the calculated square prediction error and the control limit of the square prediction error to judge the fault diagnosis result; if the fault diagnosis result determines that a fault has occurred in the monitored and diagnosed object, the fault data and the heterogeneous signal monitoring and diagnosis model are used to obtain the cause-effect diagram of the variable to infer the root cause of the fault.

[0061] Furthermore, the multi-source signal conditioner supports all analog sensors whose output is a voltage value, including integrated circuit type piezoelectric measurement microphones, integrated circuit type piezoelectric vibration sensors, thermal resistance sensors, pressure sensors, current sensors, voltage sensors, temperature sensors, and speed sensors;

[0062] The functions of the multi-source signal conditioner specifically include: providing corresponding standard sensor interfaces for various sensors, the standard sensor interfaces include bayonet nut connectors and small bayonet connectors; exciting sensors, specifically providing constant current power supply for integrated circuit type piezoelectric sensors; isolation protection, specifically isolating the signal source from the subsequent circuit to prevent the signal source from being affected by the load; conditioning the signal output by the sensor, specifically high-pass filtering, single-ended to differential conversion and gain amplification processing for high-frequency signals, and low-pass filtering and single-ended to differential conversion processing for low-frequency signals;

[0063] The functions of the above-mentioned multi-source signal conditioner are realized by an interface circuit, a constant current source circuit, a filter circuit, a voltage follower circuit, a single-ended to differential circuit, a gain amplifier circuit and a power conversion circuit.

[0064] Furthermore, the selection of the analog-to-digital converter is specifically achieved through the following requirement: the type and number of the analog-to-digital converters are dynamically selected according to the type and number of sensors on the object to be monitored and diagnosed.

[0065] Furthermore, the Raspberry Pi specifically includes: initializing hardware devices without changing the Linux kernel by dynamically overwriting the device tree, adding auxiliary serial peripheral interface communication hardware configuration to control the multi-threaded parallel and synchronous acquisition of multi-source heterogeneous data including discrete set value signals and continuous output signals by analog-to-digital converters; wherein the device tree refers to a data structure describing the hardware configuration, which is loaded during the Linux kernel startup process; and synchronously acquiring multi-source heterogeneous data including discrete set value signals and continuous output signals by controlling the analog-to-digital converters by multi-threaded parallel;

[0066] The Raspberry Pi controls the analog-to-digital converter to realize synchronous acquisition of source heterogeneous data, specifically including: firstly performing initial configuration of the serial peripheral interface, specifically: setting the serial peripheral interface channel, clock frequency, and serial peripheral interface working mode of the analog-to-digital converter; then initializing the operation, specifically: initializing all register contents and pin levels of the analog-to-digital converter to factory default values; then configuring the sampling channel and sampling sequence of the analog-to-digital converter; then configuring the output data format to ensure that the data transmission is output in the desired format; secondly reading the data, specifically: making the analog-to-digital converter repeat the reading operation under the above configuration to realize the cyclic reading of multi-source heterogeneous data; finally decoding the data, specifically: converting the reading result from a digital quantity to an analog voltage value according to the conversion logic of the register of the analog-to-digital converter; wherein, when multiple serial peripheral interfaces of the Raspberry Pi are connected to multiple analog-to-digital converters at the same time, multiple threads are created to control different analog-to-digital converters to read different data respectively, so as to control the analog-to-digital converters in parallel to perform synchronous acquisition of multi-source heterogeneous data.

[0067] Furthermore, the device also includes a shell, the shell is made of high temperature resistant material and a heat insulation layer and a buffer material are added inside the shell.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] (1) The present invention realizes monitoring and diagnosis in complex scenarios based on the input-output driven causal decoupling representation learning method (IOH-CDR), and solves the problem that discrete set value input leads to difficulty in heterogeneous data modeling and distortion of the internal operating state of the system in the monitoring and diagnosis process of industrial equipment with discrete control signals, which is common in complex scenarios. The present invention has effective state monitoring and fault cause diagnosis capabilities for industrial equipment with discrete set value control that is common in actual complex industrial scenarios. The present invention can effectively improve the accuracy and efficiency of equipment monitoring and diagnosis, has effective monitoring and diagnosis capabilities, and has broad application scenarios.

[0070] (2) The device described in the present invention can adapt to various standard sensors, and achieves good multi-source data acquisition performance through a multi-source data acquisition module based on an embedded system, reducing the size of the hardware equipment, while having deployment flexibility and scalability, reducing costs, and having effective data acquisition capabilities for various standard sensors on industrial equipment, and has a wide range of application scenarios.

[0071] (3) The multi-source data acquisition module of the present invention is composed of a multi-source signal conditioner, an analog-to-digital converter and a Raspberry Pi, wherein the multi-source signal conditioner supports various types of sensors, and is used to stimulate the sensors and condition their output signals; the analog-to-digital converter supports the acquisition of high-frequency, high-dynamic range sound and vibration signals, and is used to convert the conditioned output analog signals into digital signals; the Raspberry Pi adds auxiliary serial peripheral interface (SPI) communication hardware configuration by dynamically overlaying the device tree, and is used to control the analog-to-digital converter to perform multi-source data synchronous acquisition of discrete set value signals and internal continuous signals in the system.

[0072] (4) The present invention performs monitoring and diagnosis of multi-source heterogeneous signals on a Raspberry Pi, and performs data cleaning, data preprocessing, feature extraction and fusion based on the collected multi-source data; the present invention can solve the distribution difference problem of multi-source heterogeneous data, and restore the discrete set value signal to a continuous control embedding by learning conditional reversible mapping, thereby bridging the distribution difference; the present invention can solve the problem that the collected discrete set value signal distorts the internal continuous signal, and decouples the exogenous influence and internal operating state of the system through feature space constraint design and separate establishment of prediction models, thereby preventing statistics and causal analysis from being interfered with, thereby realizing state monitoring and fault cause diagnosis of industrial equipment regulated by discrete set values.

[0073] (5) The present invention comprehensively considers the characteristics of discrete exogenous source regulation of continuous systems in actual industrial scenarios and designs a heterogeneous signal monitoring and diagnosis algorithm. Compared with the traditional modeling input method, the present invention takes into account the limitations of inconsistent variable types and models discrete exogenous sources by restoring the continuity control effect of discrete set value input and designing decoupled feature space constraints. This avoids the input and output signal distribution differences and coupling effects interfering with statistical feature analysis, thereby achieving state monitoring and fault cause diagnosis.

[0074] (6) The present invention covers a multi-source data acquisition module and a multi-source heterogeneous signal monitoring and diagnosis, which can realize multi-source data acquisition, status monitoring and fault cause diagnosis in an integrated manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a flow chart of the architecture of the complex scenario multi-source heterogeneous data collection and monitoring and diagnosis method and device of the present invention;

[0076] Figure 2 It is a flow chart of the multi-source heterogeneous signal monitoring and diagnosis method of the present invention;

[0077] Figure 3 The architecture flow chart of the input-output driven causal decoupling representation learning model of the present invention;

[0078] Figure 4 A schematic diagram of a part of the circuit of the multi-source signal conditioner of the present invention;

[0079] Figure 5 is another part of the circuit principle diagram of the multi-source signal conditioner of the present invention;

[0080] Figure 6 It is a schematic diagram of the process of multi-source heterogeneous data collection of the present invention;

[0081] Figure 7 This is a data collection effect diagram of the multi-source heterogeneous data collection and monitoring diagnosis method of Example 1 of the present invention;

[0082] Figure 8 This is an overall effect diagram of fault monitoring and fault variable causal relationship diagnosis according to Example 1 of the present invention;

[0083] Fig. 9 This is an overall effect diagram of fault monitoring and fault variable causal relationship diagnosis according to Embodiment 2 of the present invention;

[0084] Fig.10 This is an overall effect diagram of fault monitoring and fault variable causal relationship diagnosis in Example 3 of the present invention. DETAILED DESCRIPTION

[0085] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0086] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0087] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0088] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0089] For industrial equipment operating in complex industrial scenarios with harsh environments, high risks, numerous equipment and a wide distribution span in the fields of electricity, water conservancy, mining, transportation, etc., the complex scenario multi-source heterogeneous data collection and monitoring and diagnosis method of the present invention can realize multi-source data collection, status monitoring and diagnosis of these equipment.

[0090] See also Figure 1 The complex scene multi-source heterogeneous data collection and monitoring diagnosis method of the present invention specifically includes the following steps:

[0091] (1) Collect multi-source heterogeneous data of the objects to be monitored and diagnosed in complex scenarios.

[0092] It should be noted that complex scenarios refer to complex industrial scenarios in the fields of electricity, water conservancy, mining, and transportation, which are characterized by harsh environments, high risks, numerous equipment, and large distribution spans, including but not limited to high-temperature, high-pressure, and noisy scenarios such as thermal power plants and nuclear power plants. The objects to be monitored and diagnosed refer to industrial equipment / systems that are commonly regulated by discrete set values ​​in complex scenarios, rather than isolated industrial equipment / systems that exist in actual scenarios and are not regulated externally.

[0093] Furthermore, heterogeneous signals refer to discrete signals and continuous signals existing in industrial equipment / systems, that is, multi-source heterogeneous data include discrete signals and continuous signals, among which discrete signals are control signals whose inputs are finite set values, and discrete signals include signals such as speed and flow; continuous signals are output signals that change under the influence of set value inputs, and continuous signals include signals such as sound, vibration, voltage, current, and temperature.

[0094] (2) Perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set.

[0095] Furthermore, data cleaning includes but is not limited to: using median filtering, mean filtering, Gaussian filtering or Kalman filtering methods to clean the collected multi-source heterogeneous data to remove noise and outliers in the multi-source heterogeneous data, improve the quality of the data, ensure the accuracy of subsequent analysis and modeling, and help improve the accuracy of subsequent feature extraction and heterogeneous signal monitoring and diagnosis model modeling including input-output driven causal decoupling representation learning methods. Data preprocessing includes but is not limited to: pre-emphasis, framing, windowing and other operations on the multi-source heterogeneous data after data cleaning. Feature extraction and fusion include but are not limited to: extracting the Mel frequency cepstral coefficients (MFCC) of the sound signal, the time domain, frequency domain and time-frequency domain features of the vibration signal, and the time domain features of the remaining low-frequency signals, and using fusion methods to fuse them to form a feature data set, where the fusion method includes but is not limited to feature splicing, weighted fusion, adaptive fusion and other methods.

[0096] like Figure 2 As shown, taking the following steps as an example, the specific steps of data cleaning, data preprocessing, feature extraction and fusion of the collected multi-source heterogeneous data are described in detail:

[0097] (2.1) In this embodiment, the Gaussian filtering method is first used to clean the collected sound, vibration, voltage, current, temperature, speed and other signals. Of course, other methods such as median filtering and mean filtering can also be used to clean the data to remove noise, outliers or inconsistencies in the data, smooth the signal, improve the quality of the data, and ensure the accuracy of subsequent feature extraction and heterogeneous signal monitoring and diagnosis modeling, including the input-output driven causal decoupling representation learning method, which is expressed as:

[0098]

[0099] in, Represents the data after Gaussian filtering; is the original time series data input The Sample points; is the value of the Gaussian weight function, indicating the distance from the current moment The offset is The weight of Represents the sliding window size of the filter. The expression of the Gaussian weight function is:

[0100]

[0101] According to the above formula, for the distance from the center point Data points that are farther away have a higher weight will decay rapidly, the standard deviation Controls the width of the Gaussian function, where the sliding window size of the filter is selected as , the standard deviation can be estimated using an empirical formula based on the filter size , is a larger odd number, such as 5, 7, 9, etc., and the standard deviation can be adjusted according to the size of the filter. Make adjustments.

[0102] (2.2) Design a high-pass filter to pre-emphasize the sound signal, increase its high-frequency components, and make the spectrum of the entire signal flatter. Assume that the input signal The sampling points are , the pre-emphasis formula is as follows:

[0103]

[0104] in, It is the signal output after pre-emphasis. is the pre-emphasis coefficient, and its value is between 0.9375 and 0.97. In this embodiment, .

[0105] (2.3) The pre-emphasized signal is divided into frames. This example uses the overlapping segmentation method to ensure that there is an overlapping part between two adjacent frames. The time interval between two adjacent frames is called frame shift, and the ratio of frame shift to frame length is between 0 and 0.5. Assume that the total length of each sample signal is , the original frame length is , the frame shift is , the overlapping part is: , the number of frames divided is , and its calculation formula is:

[0106]

[0107] It is necessary to add zeros to the end of each frame signal so that its length after zero padding is is a power of 2, usually 256 or 512; if It is not an integer and needs to be rounded down.

[0108] (2.4) Add a Hamming window to each frame signal to increase the continuity of the left and right ends of the frame, offset the influence of the data being infinite in time domain under the fast Fourier transform (FFT) assumption, and reduce spectrum leakage. Assume that the signal after framing is , the windowed output , the calculation formula is:

[0109]

[0110] in, represents the Hamming window function, and its calculation formula is:

[0111]

[0112] in, is the sampling point, Window length.

[0113] (2.5) Frame signals after frame division and windowing , perform fast Fourier transform to obtain the linear spectrum of each frame signal , whose fast Fourier transform is:

[0114]

[0115] in, is the frequency domain signal frequency components.

[0116] (2.6) The above linear spectrum pass dimensional Mel filter, Refers to the number of filters, The value of is between 24 and 40. Then calculate the logarithmic energy of the Mel spectrum , and its calculation formula is:

[0117]

[0118] Among them, the Mel filter bank is to apply a triangular filter on the Mel scale, from the frequency scale The formula for conversion to the Mel frequency scale is:

[0119]

[0120] The formula for converting from the Mel frequency scale to the frequency scale is:

[0121]

[0122] Frequency response of a triangular filter Defined as:

[0123]

[0124] in, For the The center frequency of the Mel filter.

[0125] (2.7) Logarithmic energy of Mel frequency by discrete cosine transform (DCT) Transform to obtain the Mel frequency cepstrum coefficients , the formula is:

[0126]

[0127] in, Refers to the Mel cepstral coefficient order, usually 2-13, here we take , Refers to Mel-order cepstral coefficients.

[0128] (2.8) Vibration signal, voltage, current, temperature and speed signal After normalization, framing, windowing and other preprocessing, normalization is performed. The normalization formula is:

[0129]

[0130] in, is the original signal, is the normalized signal, and the steps of framing and windowing are the same as the above steps (2.3) and (2.4).

[0131] (2.9) The extracted time domain features, mean The calculation formula is:

[0132]

[0133] variance The calculation formula is:

[0134]

[0135] Standard Deviation The calculation formula is:

[0136]

[0137] Peak The calculation formula is:

[0138]

[0139] Kurtosis The calculation formula is:

[0140]

[0141] in, is the time domain signal after framing and windowing, , is the window length, is the sampling point.

[0142] (2.10) The frequency components of the vibration signal are extracted by fast Fourier transform (FFT) to obtain frequency domain features, such as spectrum energy, in the same way as step (2.5), where the main frequency corresponds to the frequency of the maximum value in the spectrum, and the spectrum energy The formula is:

[0143]

[0144] in, is the linear spectrum of the fast Fourier transform, The frequency domain signal frequency components.

[0145] (2.11) Short-time Fourier transform (STFT) is used to extract the time-frequency characteristics of the vibration signal. The main formula is:

[0146]

[0147] in, is the windowing function, same as step (2.4); is the time parameter, is the frequency parameter.

[0148] (2.12) Integrate the extracted time domain, frequency domain and time-frequency domain features, and fuse and concatenate the Mel frequency cepstral coefficients (MFCC) to form a comprehensive feature data set, which is used to characterize the operating status of the equipment and serve as the conditional input of the subsequent monitoring and diagnosis algorithm to assist in modeling discrete control signals. Specifically, align the windowed samples of the above extracted features and concatenate them by column to form a feature data set , and standardize the integrated feature data set to eliminate the dimensional differences between different features. Integrate feature information to form a comprehensive feature set , the sample size is , the feature dimension is Specifically, first calculate the mean and standard deviation of all the features extracted above:

[0149]

[0150]

[0151] in, is the mean of the feature, is the standard deviation of the feature, is the sample size, is the value of the sample under this feature. Then standardize each feature:

[0152]

[0153] in, It is the standardized features, which are spliced ​​to form a new feature data set .

[0154] (3) Since the object of data collection is industrial equipment controlled by discrete set values, in order to solve the problem of difficulty in modeling heterogeneous signals caused by the introduction of exogenous discrete inputs and the distortion of the internal operating state of the system, the feature data set obtained in step (2) is used And the data after data cleaning , construct an input-output driven causal decoupled representation learning (IOH-CDR) model to extract the decoupled representation of the monitored diagnostic object and infer the causal relationship between variables. The causal decoupled representation learning model includes an exogenous control transformation network and a dual-perspective evolutionary modeling and causal inference network, such as Figure 3 As shown; and train the causal decoupling representation learning model. During the training process, the model parameters are adjusted with the minimization of the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as a heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored and diagnosed object and diagnose the root cause of its fault.

[0155] It should be understood that the monitoring and diagnosis process of industrial equipment with discrete set value control is widely present in the scene, rather than the isolated industrial equipment / system without external control that is less common in the real world. In order to solve the difficulties in modeling heterogeneous data brought by discrete set value input and the distortion of the internal operating state of the system, an input-output driven causal decoupled representation learning (IOH-CDR) model is proposed. It comprehensively considers mixed time series input and output data (including multi-source data of discrete set value signals and continuous output signals of industrial equipment) to extract the causal relationship of internal variables of industrial equipment. In order to solve the distribution difference problem of multi-source heterogeneous data, the learning conditional reversible mapping restores the discrete set value signal to continuous control embedding to bridge the distribution difference. In order to solve the problem that the collected discrete set value signal distorts the internal continuous signal, the external influence and internal operating state of the system are decoupled through feature space constraint design and the establishment of a separate causal decoupled representation learning model to prevent interference with statistics and causal analysis, thereby realizing the state monitoring and fault cause diagnosis of industrial equipment with discrete set value control.

[0156] In this embodiment, Figure 3 As shown in the figure, the exogenous control transformation network specifically includes: comprehensively utilizing the feature data set and the discrete signal data and continuous signal data after data cleaning, and fully modeling the exogenous response of the input-output system by constructing a conditionally reversible mapping between discrete input and continuous control embedding; utilizing the feature data set as the conditional input of the exogenous control transformation network to restore the high-dimensional continuous control embedding for predicting the system input response, and at the same time reconstructing the continuous control embedding back to the original discrete input, and transforming the discrete signal data with finite values ​​into the continuous control embedding of the system to characterize the rich control information of the input, which can effectively avoid the distortion of discrete numerical information in the high-dimensional embedding learning process; the exogenous control transformation network introduces a control condition reference mechanism, and uses internal observations as conditional inputs to guide the learning of continuous control embeddings, expressing the different control effects of finite discrete values ​​under different system operating conditions.

[0157] In this embodiment, Figure 3 As shown in the figure, the dual-perspective evolutionary modeling and causal inference network specifically include: separately modeling the self-driven evolution and controlled evolution processes to obtain decoupled representations containing different evolutionary information; designing joint prediction and independent prediction networks, using two types of evolutionary decoupling representations to predict the future values ​​of internal variables after evolutionary coupling and decoupling, respectively, iteratively updating network parameters according to the residual between the predicted value and the true value, and inferring the causal relationship between variables.

[0158] Specifically, Figure 3 As shown in the figure, the construction and training of the causal decoupling representation learning model specifically includes the following steps:

[0159] (3.1) The data after Gaussian filtering (i.e. data after data cleaning ) is divided according to the discrete set value signal and the internal continuous signal, and the ,in Indicates the total length It comes from an industrial equipment input-output system controlled by discrete setpoints, including The observed output and external input; due to the difference between natural measurement and manual setting, the output variable are all continuous output values, input variables are all discrete set values. Represents the first continuous output value, Indicates the second consecutive output value, represents the pth consecutive output value, represents the first discrete input variable, represents the second discrete input variable, represents the qth discrete input variable, that is can take values ​​in a continuous interval, and is restricted to a set of discrete values ​​with a finite number of set values.

[0160] (3.2) Given the length , the number of variables is A discrete input sequence fragment ,exist Under controlled conditions, the transformation network is obtained by exogenous control Embedding of continuous control of discrete variables ,in is a finite set, The condition is the feature dataset The conditional vector obtained by convolution, is the dimension of controlling embedding, t represents the position of the sequence fragment in the entire sequence, ; In order to fill the information gap between continuous control embedding and discrete input, a set of multidimensional potential random variables is introduced , expand the input dimension of the network and provide continuity initialization information. Referring to the existing reversible neural network framework, a 1×1 convolution and affine transformation cascade network is designed as an exogenous control transformation network; the forward process of the exogenous control transformation network obtains continuous control embedding from discrete input, and its expression is:

[0161]

[0162]

[0163]

[0164] in, , and , They are affine transformations The input and output vectors, represents the exponential function with base natural logarithm, Represents the 4 sub-neural networks in the affine transformation, which do not need to have reversible properties, thus improving the network model capacity and avoiding the strict inductive bias of reversibility that limits the control embedding learning ability of the network; represents a 1×1 convolution process; the input and output dimensions of the exogenous control transformation network remain ,in The dimension of is half that of D; In the reverse process of the exogenous control transformation network, the input of each sub-neural network in the affine transformation remains unchanged. Through the inverse process of affine transformation and 1×1 convolution transformation Computational Reconstruction and , whose expression is:

[0165]

[0166]

[0167]

[0168] in, and They are and Reconstruction, is additive noise that interferes with the inverse process to enhance the robustness of the network's learned mapping.

[0169] (3.3) The parameters of the forward and reverse processes of the exogenous control transformation network are shared by Indicates; alternately perform forward and reverse process calculations, accumulating The gradient is then used to update the parameters; the forward process extracts the control embedding for subsequent prediction, and the reverse process reconstructs the external input and hidden variables ,for and , respectively calculate the mean square error and the maximum mean difference , the mean square error and the maximum mean difference The sum is taken as the reconstruction loss , whose expression is:

[0170]

[0171] Among them, the kernel function , h is a constant, and They are and The i-th row vector of .

[0172] (3.4) Dual-perspective evolutionary modeling and causal inference The network includes self-driven evolutionary networks and controlled evolutionary networks. internal variables and The nonlinear relationship between external variables, both the self-driven evolution network and the controlled evolution network use Multi-Layer Perceptron (MLP) as the backbone network; Different evolution laws of internal variables are established For structurally similar networks with non-shared parameters; in order to extract sparse causal relationships, each MLP has a learned Dimensional self-driven evolution soft mask vector or Dimensional Controlled Evolution of Soft Mask Vectors , ; The time-delayed inputs of the self-driving evolution network and the controlled evolution network are firstly and Masked, and then mapped into the feature space through MLP to capture the corresponding self-driving prediction features and controlled predictive features , expressed as:

[0173]

[0174]

[0175] in, and They represent the self-driving prediction characteristics and controlled prediction characteristics at the current time t respectively; represents the time before time t The observed value of the variable at each historical moment; express The control embedding obtained by transforming the external control recorded in the historical moments; and are the network parameters of the self-driven evolution network and the controlled evolution network respectively; represents the jth self-driving evolution network, represents the jth controlled evolution network; by adding penalty loss , to use the learned sparse mask vector to reveal causal relationships, the penalty loss The expression is:

[0176]

[0177] in, It is a penalty loss middle The hyperparameters of It is a penalty loss middle Hyperparameters of .

[0178] (3.5) Since the self-driving evolution network and the controlled evolution network rely on independent internal states and external inputs, the two evolutions cannot explain each other's evolution results. Therefore, a discriminator is designed. Used to distinguish self-driving prediction features and control prediction features , the input stream of the discriminator is the self-driving prediction feature and control prediction features The ordered splicing of and negative sequence characteristics , whose expressions are:

[0179]

[0180]

[0181] Among them, M represents and The cross entropy loss of the discriminator training is calculated using the input stream of the discriminator, expressed as:

[0182]

[0183] in, represents the discriminant loss function, It is the probability estimate of the feature sample belonging to the positive sequence rather than the negative sequence. In order to decouple the features extracted from the evolution, a decoupling penalty loss is designed based on the cross entropy loss of the discriminator training. To approximate the joint distribution The product of the marginal distribution The KL (Kullback-Leibler) divergence is calculated as:

[0184]

[0185] in, and Represent the sample distribution of self-driven prediction features and controlled prediction features, respectively. represents the KL divergence.

[0186] (3.6) Self-driving prediction features and control prediction features After splicing, the input is fed into the joint prediction network with ResNet as the skeleton to obtain the joint prediction result of the internal variables at the current time t. The prediction result represents the dynamic characterization of the system internal variables after coupling the exogenous input. In addition, by decoupling the exogenous control and the internal self-driven evolution, the dynamic performance of the system internal variables when there is no exogenous input can be described. Assume that the value change of the internal variables at time t before and after the addition of the exogenous source is , then the state of the system after decoupling is It can be expressed as:

[0187]

[0188] in, represents the value of the internal variable at time t before adding the exogenous source. and control prediction features They are respectively input into independent prediction networks with ResNet as the skeleton, and the and The predicted estimated value of is superimposed to obtain the final predicted value of the internal variable; and Respectively represent the functions represented by the joint prediction network and the independent prediction network, As the network parameters corresponding to the two functions, the prediction loss function The calculation formula is:

[0189]

[0190] in, and They are joint prediction loss and independent prediction loss respectively; It means to concatenate two vectors into a new vector; yes Parameters, yes Parameters.

[0191] (3.7) According to the reconstruction loss , penalty loss , Decoupled Penalty Loss , prediction loss function Calculate the total loss function, expressed as:

[0192]

[0193] in, represents the total loss function, express The weight parameter of . Taking minimizing the total loss function as the optimization goal, the parameters of the exogenous control transformation network, dual-view evolutionary modeling and causal inference network, joint prediction network, independent prediction network and discriminator are adjusted to obtain the trained causal decoupling representation learning model, which is used as a heterogeneous signal monitoring and diagnosis model to monitor the state of the monitored and diagnosed object and diagnose the root cause of its failure in the subsequent steps.

[0194] It should be noted that the intervention of external input will change the state of the original system, so that the causal relationship cannot be discovered based on internal observations alone. With the help of joint prediction and independent prediction, the control evolution and self-driven evolution are learned separately using the gradient truncation method. Joint prediction models the dynamic process of the internal variables of coupled control evolution, which is used to infer the control evolution process and reveal the causal relationship between external variables and internal variables; while independent prediction describes the decoupling performance of self-driven evolution and control evolution, so it can exclude the interference of control evolution in inferring the causal relationship of self-driven evolution. The parameters of the control causal evolution network are updated in one round as follows:

[0195]

[0196]

[0197] in, is the learning rate, represents the number of iterations, represents the gradient operation. On the other hand, independent prediction describes the result when self-driven and controlled evolution are decoupled, thus revealing the causal relationship between internal variables. The parameters of the self-driven causal evolution network are updated as follows:

[0198]

[0199]

[0200] In order to ensure iterative stability, an alternating optimization strategy is used when updating the discriminator. can be written as:

[0201]

[0202] in, are the parameters of the discriminator, It is the probability estimate that the feature sample belongs to the negative sequence rather than the positive sequence.

[0203] (4) Input the normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and calculate the control limit of the square prediction error (SPE) based on the prediction value corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, calculate the square prediction error based on the true value, and judge whether it exceeds the limit based on the calculated square prediction error and the control limit of the square prediction error to determine the fault diagnosis result; if the fault diagnosis result determines that the monitored diagnostic object has a fault, use the fault data and the heterogeneous signal monitoring and diagnosis model to obtain the cause-effect diagram of the variable and infer the root cause of the fault.

[0204] Specifically, the normal operating data is first input into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and the control limit of SPE is determined according to the prediction value corresponding to the normal operating data; then the real-time collected data to be tested is input into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and the square prediction error is calculated in combination with the true value, and fault monitoring is performed based on the calculated square prediction error and the control limit of SPE. The calculation formula is:

[0205]

[0206] in, To collect the predicted value of samples in real time, is the true value of the sample. The control limit of SPE is determined by the square prediction error sample of historical normal data, and the value higher than 90% of the sample value is selected as the control limit of SPE If the sample to be tested , it means that the device is operating normally, otherwise, it can be considered that the device is faulty. Select the sampling value at the time of the fault to retrain the heterogeneous signal monitoring and diagnosis model according to the network parameters and The causal adjacency matrix is ​​obtained by binarization to capture the fault propagation path and locate the root cause of the fault.

[0207] It is worth mentioning that the embodiment of the present invention also provides a complex scene multi-source heterogeneous data collection and monitoring diagnosis device, which is used to implement the complex scene multi-source heterogeneous data collection and monitoring diagnosis method of the above embodiment. Figure 1 As shown, the device includes a multi-source data acquisition module, a multi-source data processing module, a heterogeneous signal monitoring and diagnosis model acquisition module and a fault monitoring and diagnosis module.

[0208] In this embodiment, the multi-source data acquisition module is used to collect multi-source heterogeneous data of the monitored diagnostic object in complex scenarios. The multi-source data acquisition module includes a multi-source signal conditioner, an analog-to-digital converter and a Raspberry Pi. Figure 1As shown, multi-source heterogeneous data acquisition of discrete set value signals and continuous output signals of the monitored diagnostic object is jointly realized, wherein the multi-source signal conditioner is connected to the various sensors on the monitored equipment to provide corresponding standard sensor interfaces for the various sensors, and simultaneously stimulates the sensors, isolates and protects and conditions the signals output by the sensors; the analog-to-digital converter is connected to the output end of the multi-source signal conditioner to convert the conditioned output analog signal into a digital signal; the Raspberry Pi is connected to the analog-to-digital converter to control the analog-to-digital converter to perform synchronous acquisition of multi-source heterogeneous data.

[0209] Furthermore, the multi-source signal conditioner supports all analog sensors whose output is voltage value, including but not limited to integrated circuit piezoelectric (IEPE) measurement microphones, IEPE vibration sensors, thermal resistor sensors, pressure sensors, current sensors, voltage sensors, temperature sensors, speed sensors, etc. Among them, the above discrete set value signals and continuous output signals can be measured by analog sensors whose output is voltage value.

[0210] Furthermore, the functions of the multi-source signal conditioner specifically include: providing corresponding standard sensor interfaces for various sensors, including but not limited to bayonet nut connectors (BNC), small bayonet connectors (SMA), etc.; exciting sensors, specifically referring to providing constant current source power supply for IEPE type sensors (such as IEPE measurement microphones, IEPE vibration sensors, etc.); isolation protection, specifically referring to isolating the signal source from the subsequent circuit to prevent the signal source from being affected by the load; conditioning the signal output by the sensor, specifically referring to high-pass filtering, single-ended to differential conversion and gain amplification of high-frequency signals, and low-pass filtering, single-ended to differential conversion of low-frequency signals, which can improve the quality and accuracy of the signal; wherein, high-frequency signals refer to signals such as sound and vibration, and low-frequency signals refer to signals such as voltage, current, temperature, and flow. The functions of the above-mentioned multi-source signal conditioner are realized by circuits such as interface circuits, constant current source circuits, filter circuits (including high-pass filter circuits and low-pass filter circuits), voltage follower circuits, single-ended to differential conversion circuits, gain amplifier circuits, and power conversion circuits, such as Figure 4 and Figure 5 As shown, these circuits are some commonly used circuits. For example, the voltage follower circuit can provide isolation protection function.

[0211] It should be understood that IEPE sensors are special compared to other sensors. The signal line and power line share the same cable, and the data acquisition system needs to be powered by a constant current source to have signal output and subsequent collection. Other sensors generally have a separate power supply part.

[0212] In this embodiment, the multi-source signal conditioner is used to access and stimulate the IEPE type sound and vibration sensors of industrial equipment controlled by discrete set values ​​in complex industrial scenarios, and perform DC isolation filtering, single-ended to differential conversion and gain amplification processing on the sound and vibration signals output by them to improve the quality and accuracy of the signal; at the same time, other types of sensors including discrete set value signals are connected to perform low-pass filtering, single-ended to differential conversion and other processing on the voltage, current, speed and temperature signals output by them, wherein the discrete set value signal can also be measured by an analog sensor with a voltage value output. These functions are realized by various circuits, and the design steps are as follows: the interface circuit is realized by parallel connection of BNC, SMA and other interfaces. The constant current source circuit is realized by a 24V-4mA constant current source circuit designed based on a constant current source chip to stimulate the output signal of the IEPE type sensor. The output signal of the IEPE type sensor is a superposition of a useful signal and a DC bias signal, wherein the DC component can be approximately a 0Hz signal, therefore, a filter circuit is designed based on a first-order RC high-pass filter circuit to filter out the DC bias signal, and a low-pass filter circuit is designed for low-frequency signals. After passing through the filtering circuit, the impedance of the signal changes. A voltage follower circuit is designed based on an operational amplifier to achieve signal buffering isolation and impedance matching to improve load capacity. A single-ended to differential circuit is further designed based on the operational amplifier to suppress common-mode interference in the signal. The gain amplifier circuit is designed based on a precision instrumentation amplifier to improve the accuracy of the collected data and amplify the original signal of the sound and vibration signal. The power conversion circuit is a 24V to + / -5V circuit designed based on a DC-DC power chip to power analog-to-digital converters, operational amplifiers, and instrumentation amplifiers.

[0213] Exemplarily, each circuit is obtained by the following steps:

[0214] ①The interface circuit is realized by serial connection of BNC interface.

[0215] ② Design a 24V-4mA constant current source circuit based on the LM334 constant current source chip to stimulate the IEPE sensor to output the signal.

[0216] ③The output signal of the IEPE sensor is the superposition of the useful signal and the DC bias signal, in which the DC component can be approximated as a 0Hz signal. Therefore, a first-order RC high-pass filter circuit is designed to filter out the DC bias signal; the high-pass cutoff frequency The calculation formula is:

[0217]

[0218] in, is the resistance value in the high-pass filter circuit, is the capacitance value in the high-pass filter circuit.

[0219] ④ The output signals of other sensors are low-frequency signals, so a first-order RC low-pass filter circuit is designed to filter out high-frequency noise signals; the low-pass cutoff frequency The calculation formula is:

[0220]

[0221] in, is the resistance value in the low-pass filter circuit, is the capacitance value in the low-pass filter circuit.

[0222] ⑤ After passing through the filtering circuit, the impedance of the signal changes. A voltage follower circuit is designed based on the LM358 operational amplifier to achieve signal buffering and isolation and improve the load capacity.

[0223] ⑥ Further design a single-ended to differential circuit based on the LM358 operational amplifier to suppress common-mode interference in the signal; the single-ended to differential circuit consists of a voltage follower circuit and an inverting amplifier circuit, expressed as:

[0224]

[0225]

[0226]

[0227] in, is the input signal, is the output signal of the voltage follower circuit, is the output signal of the inverting amplifier circuit, is the differential output signal, is the feedback resistor value, is the input resistance value.

[0228] ⑦ In order to match the acquisition range of the analog-to-digital converter and improve the accuracy of the collected data, a gain amplifier circuit is designed based on the INA128 precision instrument amplifier to amplify the original signal of the sound and vibration signal; gain amplifier The calculation formula is:

[0229]

[0230] in, is the feedback resistor value. According to the INA128 data sheet, the internal circuit design generates parameters .

[0231] ⑧ Design a 24V to + / -5V circuit based on the DC-DC power supply chip to power analog-to-digital converters, operational amplifiers, instrumentation amplifiers, etc.

[0232] ⑨ When selecting devices, pay attention to choosing high-temperature resistant components, and when making PCB boards, pay attention to choosing high-temperature resistant PCB materials, and apply a protective layer after soldering.

[0233] ⑩ Establish hardware connections between measurement microphones, vibration sensors and other sensors and multi-source signal conditioners to achieve effective signal transmission and processing.

[0234] Furthermore, the selection of analog-to-digital converters is specifically achieved through the following requirements: the type and number of analog-to-digital converters are dynamically selected according to the type and number of sensors on the object to be monitored and diagnosed (i.e., industrial equipment / system). Specifically, a multi-channel analog-to-digital converter with a high sampling rate and a high dynamic range is selected for the acquisition of high-frequency signals such as sound and vibration signals, wherein the high sampling rate and the high dynamic range are determined by the high-pass cutoff frequency of the high-pass filter circuit; a multi-channel analog-to-digital converter with a moderate sampling rate is selected for the acquisition of other signals including discrete set value signals. Among them, one channel of the analog-to-digital converter corresponds to the data collected from one sensor.

[0235] For example, if there are 5 IEPE measurement microphones, 3 IEPE vibration sensors, 3 voltage sensors, 3 current sensors, 1 temperature sensor and 1 speed sensor on the monitored diagnostic object, an 8-channel high-speed and high-precision register-type ADS8688 analog-to-digital converter is selected to jointly collect the sound signals of the 5 IEPE measurement microphones and the vibration signals of the 3 IEPE vibration sensors, and an 8-channel high-precision register-free ADS1278 analog-to-digital converter is selected to jointly collect the voltage, current, temperature and speed signals.

[0236] Furthermore, the Raspberry Pi specifically includes: initializing the hardware device without changing the Linux kernel by dynamically overwriting the device tree, adding auxiliary serial peripheral interface (SPI) communication hardware configuration to control the multi-threaded parallel and synchronous acquisition of multi-source heterogeneous data including discrete set value signals and continuous output signals by the analog-to-digital converter. Among them, the device tree refers to the data structure that describes the hardware configuration, which is loaded during the Linux kernel startup process; the analog-to-digital converter is controlled by multi-threaded parallel to perform synchronous acquisition of multi-source heterogeneous data including discrete set value signals and continuous output signals. The specific implementation method is to modify the configuration file (config.txt) of the Raspberry Pi terminal, add statements such as dtoverlay=spi1-1cs, dtoverlay=spi1-2cs, dtoverlay=spi1-3cs, and gradually enable other SPI communication hardware configurations.

[0237] It should be noted that the dynamic overlay device tree needs to refer to the Raspberry Pi SPI bus manual. Pay attention to the situation that some SPI channels conflict and cannot be opened at the same time. For details, you can refer to the official data manual of Raspberry Pi. For example, the core application processor chip BCM2711 used for Raspberry Pi 4B, Raspberry Pi 400 and Raspberry Pi CM4 can support up to 6 SPI input and output interfaces. Among them, Raspberry Pi 4B has some channels that cannot be opened at the same time, and only 4 can be opened at the same time. The core application processor chip BCM2711 used for Raspberry Pi 5 has up to 6 SPI input and output interfaces. Each SPI input and output interface has 2-3 chip select signal lines. Taking Raspberry Pi 4B with 8-channel analog-to-digital converter as an example, up to 72 (8×9) channels of data acquisition can be achieved; in addition, analog SPI communication can also be used synchronously, that is, SPI timing state is simulated through general function input and output (GPIO), and SPI communication with slave models such as analog-to-digital converters can also be achieved. Although the maximum communication rate supported by this analog SPI communication method is not as high as the SPI communication rate enabled by the above dynamic overlay device tree, it can maximize the number of acquisition channels. Specifically, a low-frequency analog-to-digital converter and a high-frequency analog-to-digital converter can be configured according to requirements to achieve effective collection of low-frequency signals and high-frequency signals.

[0238] Furthermore, the Raspberry Pi controls the analog-to-digital converter to realize the synchronous acquisition of source heterogeneous data. The specific steps vary according to the working mode of the analog-to-digital converter, but the overall process of realizing data acquisition includes SPI initial configuration, initialization operation, configuration of sampling order, configuration of output data format, reading data and decoding data, such as Figure 6As shown, it specifically includes: firstly, the SPI initial configuration is performed, specifically: the SPI channel, clock frequency, and SPI working mode of the analog-to-digital converter are set. Among them, in order to determine which SPI channels are turned on, the SPI channel of the analog-to-digital converter needs to be set, such as SPI1-1cs; the clock frequency determines the rate of data transmission in SPI communication, such as 7.8MHz; it is necessary to select according to the slave, that is, the working mode of the analog-to-digital converter, to determine the SPI working mode, such as mode 0 (CPOL=0, CPHA=0), where CPOL=0 means low level when idle, and CPHA means that the data is sampled on the second falling edge. Then the initialization operation is performed, specifically: all register contents and pin levels of the analog-to-digital converter are initialized to the factory default values. Then the sampling channel and sampling order of the analog-to-digital converter are configured, such as traversing 8 sampling channels in ascending order. Then the output data format is configured to ensure that the data transmission is output in the desired format, such as the data transmission is described as the most significant bit first (MSB) format. Secondly, the data is read, specifically: the analog-to-digital converter is configured to repeat the reading operation under the above configuration to realize the cyclic reading of multi-source heterogeneous data. Finally, the data is decoded, specifically: according to the conversion logic of the register of the analog-to-digital converter, the reading result is converted from a digital quantity to an analog voltage value. When multiple SPI interfaces of the Raspberry Pi are connected to multiple analog-to-digital converters at the same time,<pthread.h> Library file, create multiple threads, respectively control different analog-to-digital converters to read different data, and parallel control the analog-to-digital converters to perform synchronous acquisition of multi-source heterogeneous data, which can effectively improve the real-time performance of data acquisition.

[0239] For example, in this embodiment, an 8-channel high-speed and high-precision register-type ADS8688 analog-to-digital converter is selected to collect 5 sound signals and 3 vibration signals, and an 8-channel high-precision register-free ADS1278 analog-to-digital converter is selected to collect voltage, current, temperature and speed signals. Raspberry Pi opens another independent hardware SPI channel and uses multi-threaded parallel processing technology to synchronously collect sound, vibration signals and voltage, current, temperature and speed signals. The specific steps are as follows:

[0240] ① Establish the hardware connection between Raspberry Pi 4B, ADS8688 ADC, ADS1278 ADC and multi-source signal conditioner to ensure the normal SPI communication and signal flow.

[0241] ② Based on the / boot / firmware / config.txt file of the Raspberry Pi terminal, add dtoverlay =spi6-1cs and other operations to turn on the Raspberry Pi SPI6-1cs.

[0242] ③Based on<pthread.h> Library file, create two threads, one of which controls ADS8688 to collect sound and vibration signals, and the other controls ADS1278 to collect voltage, current, temperature, speed and other signals; because the sampling rates of the two analog-to-digital converters are similar, they can be used interchangeably, which is only used as an example here.

[0243] ④Control ADS8688 to collect sound signals. The specific steps are: Based on C language and its library files<wiringPi.h> and<wiringPiSPI.h> , write a program to set the SPI channel to SPI0, the clock frequency to 31.2MHz, and the sampling mode to mode 2 (CPOL=1, CPHA=0, that is, the serial clock (SCLK) idle state is high, and sampling is performed on the second falling edge of the SCLK clock); perform initialization operations to reset all program register contents to factory defaults; set the command register to automatic cycle (AUTO_RST) mode, in which the analog-to-digital converter cycles through the 8 sampling channels in ascending order to achieve synchronous sampling; configure the program register bit of the SDO data format to ensure that the data transmission adopts the format described as the most significant bit first (MSB); set the command register to no operation (NO_OP) mode, so that the analog-to-digital converter repeats the previous operation to achieve cyclic reading of 8-channel data; [24:9] in the extracted SDO output data is the conversion result of the selected channel, and the conversion result is converted from a digital quantity to a voltage value, expressed as:

[0244]

[0245] in, For the reference voltage, the onboard reference voltage is selected here. =2.5V; It is the digital signal read, that is, [24:9] in the extracted SDO output data, indicating that the MSB is first; is the converted voltage value.

[0246] It should be understood that SDO is the name of the interface. SPI communication requires four signal lines, namely SCLK, SDO, SDI and CS. MSB is the most significant bit of the data in a string of data output by SDO. The output data format usually has two types: MSB first and LSB first. According to the data sheet of the analog-to-digital converter, [24:9] in the SDO output data is the conversion result of the selected channel, that is, bits 24-9 in the SDO output data are the conversion results, and bits 0-8 are the current range and other information.

[0247] ⑤ Control ADS1278 to collect signals such as vibration, voltage, current, temperature and speed. The specific steps are: connect the jumper to set the analog-to-digital converter to high-precision mode, SPI communication mode, time division multiplexing mode (TDM), fixed position output mode, clock input distribution to high speed, set the clock frequency to 31.2MHz, and the sampling mode to mode 2; perform initialization operations, initialize SYNC and CLK settings; read the output data after the falling edge of DRDY; [23:0] in the extracted SDO output data is the conversion result of the selected channel and convert the read digital signal into an analog signal, which is the actual sampling signal. , expressed as:

[0248]

[0249] in, For the reference voltage, the onboard reference voltage is selected here. =2.5V; It is the digital signal read, indicating that MSB is first; is the converted voltage value; in TDM mode, all channel data are shifted out in serial format through the common pin SDO; in fixed mode, the data position of the channel remains fixed regardless of whether a channel is powered off.

[0250] In this embodiment, the multi-source data processing module is used to perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set.

[0251] In this embodiment, the heterogeneous signal monitoring and diagnosis model acquisition module is used to use the feature data set and the data after data cleaning to construct an input-output driven causal decoupling representation learning model, so as to extract the decoupled representation of the monitored and diagnosed object and infer the causal relationship between variables, wherein the causal decoupling representation learning model includes an exogenous control transformation network and a dual-perspective evolutionary modeling and causal inference network; and train the causal decoupling representation learning model. During the training process, the model parameters are adjusted with the minimization of the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as the heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored and diagnosed object and diagnose the root cause of its failure.

[0252] In this embodiment, the fault monitoring and diagnosis module is used to input normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain corresponding prediction values, and calculate the control limit of the square prediction error based on the prediction values ​​corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain corresponding prediction values, calculate the square prediction error in combination with the true value, and judge whether the calculated square prediction error and the control limit of the square prediction error are out of limit to judge the fault diagnosis result; if the fault diagnosis result determines that a fault has occurred in the monitored and diagnosed object, the fault data and the heterogeneous signal monitoring and diagnosis model are used to obtain a causal diagram of the variables to infer the root cause of the fault.

[0253] It should be noted that the above-mentioned multi-source data processing module, heterogeneous signal monitoring and diagnosis model acquisition module, and fault monitoring and diagnosis module are deployed on the Raspberry Pi, and can perform data cleaning, data preprocessing, feature extraction and fusion based on the collected multi-source heterogeneous data; through learning conditional reversible mapping, the discrete set value signal is restored to a continuous control embedding, bridging the distribution difference and solving the distribution difference problem of multi-source heterogeneous data; through feature space constraint design and separate model establishment, the problem of the collected discrete set value signal distorting the internal continuous signal is solved, the exogenous influence and internal operating state of the system are decoupled, and the statistics and causal analysis are prevented from being interfered with, thereby realizing the state monitoring and fault cause diagnosis of industrial equipment regulated by discrete set values.

[0254] In some other embodiments, the device described in the present invention also includes a shell, which is made of high-temperature resistant material and has an insulation layer and a buffer material inside the shell to wrap the above-mentioned module, protect it from water and dust intrusion, and reduce the impact of impact vibration and external heat on it.

[0255] Specifically, a high temperature resistant shell is designed to wrap the above hardware parts to protect them from water and dust intrusion and reduce the impact of shock vibration and external heat. The specific steps are as follows:

[0256] ① Hardware layout: Place each part reasonably according to the hardware function and heat dissipation requirements, and pay attention to placing the power supply at the edge.

[0257] ② Shell design: According to the hardware layout, reserve space for the insulation layer and buffer layer, design the outer dimensions of the shell, and choose high-temperature resistant engineering plastic as the material.

[0258] ③ Heat insulation layer design: Aluminum foil is pasted on the inner surface of the shell to form a heat insulation barrier.

[0259] ④Buffer layer design: Add silicone foam as a buffer material inside the shell and arrange it inside the shell.

[0260] ⑤Environmental testing: Conduct environmental testing on the designed casing to verify the performance of the casing.

[0261] The following detailed description of the complex scenario multi-source heterogeneous data collection and monitoring and diagnosis method and device of the present invention is based on the embodiments, and the purpose and effect of the present invention will become more obvious.

[0262] Embodiment 1: Multi-source heterogeneous data collection and monitoring diagnosis method and device for wind turbines

[0263] Taking the operation and maintenance task of wind turbines in wind power generation as an example, in this embodiment, the wind turbine has 1 measurement microphone, 3 vibration sensors, 3 voltage sensors, 3 speed sensors, 3 current sensors, 2 temperature sensors, 3 pressure sensors, 1 battery monitoring sensor, 1 wind speed sensor, and 1 humidity sensor, and the interface type of the sensors is BNC type. Among them, the data of the 3 speed sensors are control signals with a limited number of set values, and the data of the other 18 sensors are output signals that change under the influence of the set value input.

[0264] For this embodiment, the steps of the multi-source heterogeneous data collection and monitoring diagnosis method and device are as follows:

[0265] Step 1. Design a multi-source signal conditioner according to the method described above. Select two 8-channel high-speed and high-precision register-type ADS8688 analog-to-digital converters and one 8-channel high-precision non-register type ADS1278 analog-to-digital converter to collect sound, vibration, voltage, current, pressure, speed and other signals. Raspberry Pi opens another independent hardware SPI channel and uses multi-threaded parallel processing technology to synchronously collect sound, vibration, voltage, current, pressure, speed and other signals. The collected sample amplitude is as follows: Figure 7 shown.

[0266] Step 2: Use Gaussian filtering to clean the collected sound, vibration, voltage, current, pressure, speed, etc. to remove noise, outliers or inconsistencies in the data, smooth the signal, improve the quality of the data, and ensure the accuracy of subsequent analysis and modeling. Preprocess and extract features of the filtered signal. Integrate the extracted time domain, frequency domain, and time-frequency domain features, and fuse and splice the Mel frequency cepstral coefficients (MFCC) to form a comprehensive feature data set.

[0267] Step 3. Based on the filtered data, learn a conditionally reversible mapping relationship, use the extracted feature data set as the conditional input, and transform the limited discrete data into the continuous control embedding of the system to characterize the rich control information of the input. Design a dual-perspective causal inference network to separately model the self-driven evolution and control evolution processes to obtain a decoupled representation that contains different evolutionary information. Design a joint prediction and independent prediction network, and use the two evolutionary decoupling representations to predict the future values ​​of internal variables after evolutionary coupling and decoupling, respectively. Iterate and update the network parameters according to the residual between the predicted value and the true value to infer the causal relationship between the variables. Design an update strategy for the exogenous control transformation network, decoupled causal evolution network, prediction network, and discriminator parameters.

[0268] Step 4: Using the extracted comprehensive feature data set, we calculate the squared prediction error (SPE) and The control limits of the statistic are then substituted into the model to obtain SPE and Statistics for fault monitoring.

[0269] Step 5: Design a high temperature resistant casing to wrap the above hardware parts.

[0270] For this embodiment, the actual test results of the data acquisition hardware are that 21 channels of sound, vibration, voltage, current, pressure, speed and other signals are collected simultaneously, and the sampling rate is about 5800 sample points per second for a single channel, which meets the collection rate requirements.

[0271] The monitoring and diagnosis effects are as follows: the fault detection rate is 96.35%, the fault false alarm rate is 11.85%, and the monitoring curve and diagnosis causal relationship diagram are as follows: Figure 8 As shown in the figure, the variables are numbered, with continuous variables in front and discrete variables in the back. Variable 15 is diagnosed as the root variable of the fault. Figure 8 The data are marked with different colors from other data.

[0272] Example 2: Multi-source heterogeneous data collection and monitoring diagnosis method and device for gas-liquid reactor equipment in chemical production

[0273] Taking the operation and maintenance task of a gas-liquid reactor equipment in the chemical industry production process as an example, the gas-liquid reactor equipment to be monitored in this embodiment has 6 pressure sensors, 4 flow sensors, 3 position sensors, 3 density sensors, 3 temperature sensors, 2 differential pressure sensors, 2 liquid level sensors and 1 current sensor, and the interface type of the sensor is BNC type. Among them, the data of 2 flow sensors are control signals of a limited number of set values, and the data of the other 22 sensors are output signals that change under the influence of the set value input.

[0274] The actual test results of the data acquisition hardware are that 24 channels of pressure, flow, liquid level, temperature, density, position and current signals are collected simultaneously, and the sampling rate is about 5,800 sample points per second for a single channel, which meets the acquisition rate requirements.

[0275] The monitoring and diagnosis effects are as follows: the fault detection rate is 90.97%, the fault false alarm rate is 8.10%, and the monitoring curve and diagnosis causal relationship diagram are as follows: Fig. 9 As shown, the variables are numbered, with continuous variables in front and discrete variables in the back. Variable 2 is diagnosed as the root dependent variable of the fault. Fig. 9 The data are marked with different colors from other data.

[0276] Example 3: Multi-source heterogeneous data collection and monitoring diagnosis method and device for coal mill in thermal power equipment

[0277] Taking the coal mill operation and maintenance task in thermal power generation equipment as an example, in this embodiment, the coal mill includes 3 vibration sensors, 2 pressure sensors, 2 temperature sensors, 2 current sensors, 3 flow sensors, 2 voltage sensors, 2 speed sensors, 2 density sensors, 2 oil temperature sensors and 1 humidity sensor, and the interface type of the sensors is BNC type. Among them, the data of the 2 speed sensors are control signals with a limited number of set values, and the data of the other 17 sensors are output signals that change under the influence of the set value input.

[0278] The actual test results of the data acquisition hardware are that 19 channels of vibration, pressure, temperature, current, flow, voltage, speed and other signals are collected simultaneously, and the sampling rate is about 5,800 sample points per second for a single channel, which meets the collection rate requirements.

[0279] The monitoring and diagnosis effects are as follows: the fault detection rate is 87.66%, the fault false alarm rate is 0.26%, and the monitoring curve and diagnosis causal relationship diagram are as follows: Fig.10 As shown, the variables are numbered, with continuous variables in front and discrete variables in the back. Variable 12 is diagnosed as the root dependent variable of the fault. Fig.10 The data are marked with different colors from other data.

[0280] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A complex scene multi-source heterogeneous data collection and monitoring diagnosis method, characterized in that: The following steps are involved: (1) Collecting multi-source heterogeneous data of the object to be monitored and diagnosed; (2) Perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set; (3) using the feature data set obtained in step (2) and the data after data cleaning, constructing a causal decoupling representation learning model based on input-output drive to extract the decoupling representation of the monitored diagnostic object and infer the causal relationship between variables, wherein the causal decoupling representation learning model includes an exogenous control transformation network and a dual-view evolutionary modeling and causal inference network; and training the causal decoupling representation learning model. During the training process, the model parameters are adjusted with minimizing the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as a heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored diagnostic object and diagnose the root cause of its fault; The exogenous control transformation network specifically includes: comprehensively utilizing the feature data set and the discrete signal data and continuous signal data after data cleaning, and fully modeling the exogenous response of the input-output system by constructing a conditionally reversible mapping between discrete input and continuous control embedding; utilizing the feature data set as the conditional input of the exogenous control transformation network to restore the continuous control embedding for predicting the system input response, and at the same time reconstructing the continuous control embedding back to the original discrete input; the exogenous control transformation network introduces a control condition reference mechanism, uses internal observations as conditional inputs to guide the learning of the continuous control embedding, and expresses the different control effects of finite discrete values ​​under different system operating states; The dual-perspective evolution modeling and causal inference network specifically includes: separately modeling the self-driven evolution and the controlled evolution process to obtain decoupled representations containing different evolutionary information; designing joint prediction and independent prediction networks, using two evolutionary decoupling representations to respectively predict the future values ​​of internal variables after evolutionary coupling and decoupling, iteratively updating network parameters according to the residual between the predicted value and the true value, and inferring the causal relationship between variables; The step (3) specifically comprises the following steps: (3.1) The cleaned data x(n) is divided into discrete set value signals and internal continuous signals to obtain in represents a mixed time series with a total length of T, including p observation outputs and q external inputs; the output variable X = [x1, x2, ..., x p ] are all continuous output values, and the input variable Y=[y1,y2,...,y q ] are all discrete set values, x1 represents the first continuous output value, x2 represents the second continuous output value, and x p represents the pth continuous output value, y1 represents the first discrete input variable, y2 represents the second discrete input variable, and y q represents the qth discrete input variable; (3.2) Given a discrete input sequence segment Y with length L and number of variables q ＜t ∈S L×q , under C control conditions, the continuous control embedding of q discrete variables is obtained through the exogenous control transformation network Where S is a finite set and C is a feature data set The conditional vector obtained by convolution, D is the dimension of control embedding, t represents the position of the sequence fragment in the entire sequence, t∈[L+1,T]; the network of 1×1 convolution and affine transformation cascade is used as the structure of the exogenous control transformation network, and its forward process obtains continuous control embedding from discrete input, and its expression is: Among them, i1, i2 and o1, o2 are the input and output vectors of the affine transformation {exp(s1), exp(s2), t1, t2}, exp(·) represents the exponential function with the base being the natural logarithm, s1, s2, t1, t2 represent the four sub-neural networks in the affine transformation, Conv 1×1 (·) represents a 1×1 convolution process; the input and output dimensions of the exogenous control transformation network remain D, and the dimensions of i1, i2, o1, o2 are half of D; ⊙ represents element-by-element multiplication; represents a set of multidimensional potential random variables introduced; in the reverse process of the exogenous control transformation network, the input of each sub-neural network in the affine transformation remains unchanged, U ＜t Through the inverse process of affine transformation and 1×1 convolution transformation Calculate the reconstruction Y ＜t and Z ＜t , whose expression is: in, and Y ＜t and Z ＜t The reconstruction of , ε is the additive noise; (3.3) The forward and reverse processes of the exogenous control transformation network share the parameters of the network, denoted by θ; the forward and reverse process calculations are performed alternately, the gradient of θ is accumulated, and then the gradient is used to update the parameters; among them, the forward process extracts the control embedding for subsequent prediction, and the reverse process reconstructs the external input Y and the latent variable Z. For Y and Z, the mean square error is calculated respectively and the maximum mean difference The mean square error and the maximum mean difference The sum is taken as the reconstruction loss Its expression is: Among them, the kernel function h is a constant, z i＜t and Z ＜t and The i-th row vector of ; (3.4) Dual-perspective evolutionary modeling and causal inference The network includes a self-driven evolutionary network and a controlled evolutionary network. Both the self-driven evolutionary network and the controlled evolutionary network use multi-layer perceptrons as the backbone network. According to the different evolution laws of p internal variables, p pairs of structurally similar networks with non-shared parameters are established. Each multi-layer perceptron has a learned p-dimensional self-driven evolutionary soft mask vector or q-dimensional controlled evolution soft mask vector To extract sparse causal relationships; the time-delayed inputs of the self-driving evolution network and the controlled evolution network are firstly and Masked, and then mapped to the feature space through a multi-layer perceptron to capture the corresponding self-driving prediction feature F sd and the controlled predictive feature F ct , expressed as: Among them, F sd (t) and F ct (t) represent the self-driving prediction features and controlled prediction features at the current time t, respectively; Represents the observed value of the variable in L historical moments before time t; represents the control embedding obtained by transforming the external control of the input records at L historical moments; φ sd and φ ct are the network parameters of the self-driven evolution network and the controlled evolution network respectively; represents the jth self-driving evolution network, represents the jth controlled evolution network; by adding penalty loss To use the learned sparse mask vector to reveal causal relationships, the penalty loss The expression is: Among them, λ sd It is a penalty loss middle The hyperparameter, λ ct It is a penalty loss middle Hyperparameters of (3.5) Design a discriminator Used to distinguish the self-driving prediction feature F sd and control prediction feature F ct , the input stream of the discriminator is the self-driving prediction feature F sd and control prediction feature F ct The ordered concatenation of pos (t) and negative sequence characteristic F neg (t), their expressions are: Where M represents F sd and F ct The dimension of the discriminator is ; the cross entropy loss of the discriminator training is calculated using the input stream of the discriminator, which is expressed as: in, represents the discriminant loss function, It is the probability estimate of the feature sample belonging to the positive sequence rather than the negative sequence; based on the cross entropy loss of discriminator training, a decoupled penalty loss is designed To approximate the joint distribution q(F sd ,F ct ) and the product of the marginal distribution q(F sd )q(F ct ), the KL divergence of which is calculated as: Among them, q(F sd ) and q(F ct ) represent the sample distribution of self-driven prediction features and controlled prediction features, respectively, and KL() represents KL divergence; (3.6) The self-driving prediction feature F sd and control prediction feature F ct After splicing, it is input into the joint prediction network with ResNet as the skeleton to obtain the joint prediction result of the internal variable at the current time t; assuming that the value change of the internal variable at time t before and after adding the external source is δ t , then the state of the system after decoupling is It is expressed as: Among them, X t represents the value of the internal variable at time t before adding the exogenous source; the self-driven prediction feature F sd and control prediction feature F ct They are respectively input into independent prediction networks with ResNet as the skeleton, and the and δ t The predicted estimated value is superimposed to obtain the final predicted value of the internal variable; let h1 and h2 represent the functions represented by the joint prediction network and the independent prediction network respectively, and ψ1 and ψ2 are the network parameters corresponding to the two functions, then the prediction loss function The calculation formula is: in, and are the joint prediction loss and independent prediction loss respectively; Con(·,·) means connecting two vectors into a new vector; ψ1 is the parameter of h1, ψ2 is the parameter of h2; (3.7) According to the reconstruction loss Penalty loss Decoupled Penalty Loss Prediction loss function Calculate the total loss function, expressed as: in, represents the total loss function, λ represents weight parameters; taking minimization of the total loss function as the optimization goal, adjust the parameters of the exogenous control transformation network, dual-view evolutionary modeling and causal inference network, joint prediction network, independent prediction network and discriminator to obtain the trained causal decoupling representation learning model, and use it as a heterogeneous signal monitoring and diagnosis model to monitor the state of the monitored and diagnosed object and diagnose the root cause of its fault in the subsequent steps; (4) Input the normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and calculate the control limit of the square prediction error based on the prediction value corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, calculate the square prediction error based on the true value, and judge whether it exceeds the limit based on the calculated square prediction error and the control limit of the square prediction error to determine the fault diagnosis result; if the fault diagnosis result determines that the monitored diagnostic object has a fault, use the fault data and the heterogeneous signal monitoring and diagnosis model to obtain the cause-effect diagram of the variable and infer the root cause of the fault.

2. The complex scene multi-source heterogeneous data collection and monitoring diagnosis method according to claim 1 is characterized in that: The multi-source heterogeneous data include discrete signals and continuous signals, wherein the discrete signal is a control signal whose input is a finite number of set values, and the discrete signal includes a rotational speed and a flow rate signal; the continuous signal is an output signal that changes under the influence of the set value input, and the continuous signal includes sound, vibration, voltage, current, and temperature signals.

3. The complex scene multi-source heterogeneous data collection and monitoring diagnosis method according to claim 1 is characterized in that: The data cleaning specifically includes: using median filtering, mean filtering, Gaussian filtering or Kalman filtering method to clean the collected multi-source heterogeneous data to remove noise and outliers in the multi-source heterogeneous data; The data preprocessing specifically includes: performing pre-emphasis, frame division, and windowing operations on the multi-source heterogeneous data after data cleaning; The feature extraction and fusion specifically include: extracting the Mel-frequency cepstrum coefficients of the sound signal, the time domain, frequency domain and time-frequency domain features of the vibration signal, and the time domain features of the remaining low-frequency signals, and using a fusion method to fuse them to form a feature data set, wherein the fusion method includes feature splicing, weighted fusion, and adaptive fusion method.

4. A complex scene multi-source heterogeneous data collection and monitoring and diagnosis device, used to implement the complex scene multi-source heterogeneous data collection and monitoring and diagnosis method according to any one of claims 1 to 3, characterized in that: The device comprises: A multi-source data acquisition module is used to collect multi-source heterogeneous data of the object to be monitored and diagnosed. The multi-source data acquisition module includes a multi-source signal conditioner, an analog-to-digital converter and a Raspberry Pi, which together realize multi-source heterogeneous data acquisition of discrete set value signals and continuous output signals of the object to be monitored and diagnosed. The multi-source signal conditioner is connected to a variety of sensors on the device to be monitored, and is used to provide corresponding standard sensor interfaces for the various sensors, and simultaneously stimulate the sensors, isolate and protect, and condition the signals output by the sensors; the analog-to-digital converter is connected to the output end of the multi-source signal conditioner, and is used to convert the analog signal output after conditioning into a digital signal; the Raspberry Pi is connected to the analog-to-digital converter, and is used to control the analog-to-digital converter to perform synchronous acquisition of multi-source heterogeneous data; The multi-source data processing module is used to perform data cleaning, data preprocessing, feature extraction and fusion on the collected multi-source heterogeneous data to obtain a feature data set; The heterogeneous signal monitoring and diagnosis model acquisition module is used to use the feature data set and the data after data cleaning to build an input-output driven causal decoupling representation learning model to extract the decoupling representation of the monitored and diagnosed object and infer the causal relationship between variables, wherein the causal decoupling representation learning model includes an exogenous control transformation network and a dual-view evolutionary modeling and causal inference network; and train the causal decoupling representation learning model. During the training process, the model parameters are adjusted with minimizing the total loss function as the optimization goal, and the trained causal decoupling representation learning model is used as the heterogeneous signal monitoring and diagnosis model to perform status monitoring on the monitored and diagnosed object and diagnose the root cause of its failure; and The fault monitoring and diagnosis module is used to input normal operating condition data into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, and calculate the control limit of the square prediction error according to the prediction value corresponding to the normal operating condition data; input the data to be tested collected in real time into the heterogeneous signal monitoring and diagnosis model to obtain the corresponding prediction value, calculate the square prediction error in combination with the true value, and judge whether it exceeds the limit based on the calculated square prediction error and the control limit of the square prediction error to judge the fault diagnosis result; if the fault diagnosis result determines that a fault has occurred in the monitored and diagnosed object, the fault data and the heterogeneous signal monitoring and diagnosis model are used to obtain the cause-effect diagram of the variable to infer the root cause of the fault.

5. The complex scene multi-source heterogeneous data acquisition and monitoring and diagnosis device according to claim 4 is characterized in that: The multi-source signal conditioner supports all analog sensors whose output is voltage value, including integrated circuit type piezoelectric measurement microphone, integrated circuit type piezoelectric vibration sensor, thermal resistance type sensor, pressure sensor, current sensor, voltage sensor, temperature sensor, speed sensor; The functions of the multi-source signal conditioner specifically include: providing corresponding standard sensor interfaces for various sensors, the standard sensor interfaces include bayonet nut connectors and small bayonet connectors; exciting sensors, specifically providing constant current power supply for integrated circuit type piezoelectric sensors; isolation protection, specifically isolating the signal source from the subsequent circuit to prevent the signal source from being affected by the load; conditioning the signal output by the sensor, specifically high-pass filtering, single-ended to differential conversion and gain amplification processing for high-frequency signals, and low-pass filtering and single-ended to differential conversion processing for low-frequency signals; The functions of the above-mentioned multi-source signal conditioner are realized by an interface circuit, a constant current source circuit, a filter circuit, a voltage follower circuit, a single-ended to differential circuit, a gain amplifier circuit and a power conversion circuit.

6. The complex scene multi-source heterogeneous data acquisition and monitoring and diagnosis device according to claim 4 is characterized in that: The selection of the analog-to-digital converter is specifically achieved through the following requirement: the type and number of the analog-to-digital converters are dynamically selected according to the type and number of sensors on the object to be monitored and diagnosed.

7. The complex scene multi-source heterogeneous data acquisition and monitoring and diagnosis device according to claim 4 is characterized in that: The Raspberry Pi specifically includes: initializing hardware devices without changing the Linux kernel by dynamically overwriting the device tree, adding auxiliary serial peripheral interface communication hardware configuration to control the analog-to-digital converter to synchronously collect multi-source heterogeneous data including discrete set value signals and continuous output signals in parallel through multiple threads; wherein the device tree refers to a data structure describing the hardware configuration, which is loaded during the Linux kernel startup process; synchronously collecting multi-source heterogeneous data including discrete set value signals and continuous output signals by controlling the analog-to-digital converter in parallel through multiple threads; The Raspberry Pi controls the analog-to-digital converter to realize synchronous acquisition of source heterogeneous data, specifically including: firstly performing initial configuration of the serial peripheral interface, specifically: setting the serial peripheral interface channel, clock frequency, and serial peripheral interface working mode of the analog-to-digital converter; then initializing the operation, specifically: initializing all register contents and pin levels of the analog-to-digital converter to factory default values; then configuring the sampling channel and sampling sequence of the analog-to-digital converter; then configuring the output data format to ensure that the data transmission is output in the desired format; secondly reading the data, specifically: making the analog-to-digital converter repeat the reading operation under the above configuration to realize the cyclic reading of multi-source heterogeneous data; finally decoding the data, specifically: converting the reading result from a digital quantity to an analog voltage value according to the conversion logic of the register of the analog-to-digital converter; wherein, when multiple serial peripheral interfaces of the Raspberry Pi are connected to multiple analog-to-digital converters at the same time, multiple threads are created to control different analog-to-digital converters to read different data respectively, so as to control the analog-to-digital converters in parallel to perform synchronous acquisition of multi-source heterogeneous data.

8. The complex scene multi-source heterogeneous data acquisition and monitoring and diagnosis device according to claim 4 is characterized in that: The device also includes a shell, which is made of high temperature resistant material and has a heat insulation layer and buffer material added inside the shell.

Images