Intelligent valve life prediction method based on big data analysis

Abstract

The present application relates to the technical field of data processing, in particular to a smart valve life prediction method based on big data analysis, which comprises: obtaining a plurality of data sequences at a target time, and constructing a dynamic base element according to the local trend slope and the local fluctuation degree of each data; calculating the coherence of two data sequences at each index according to the difference between the dynamic base elements of any two data sequences, obtaining a coherence sequence, determining the instability degree according to the coherence sequence, and aggregating the instability degree to generate a potential anomaly index at the target time; in the process of training the LSTM model, the hidden state of the last target time is optimized according to the potential anomaly index, the evolution trend of the reconstruction error with time is obtained through the optimized LSTM model, and the life of the smart valve is predicted. The method can effectively capture early failures caused by multi-variable relationship mismatch, and improve the accuracy of life prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. Specifically, it relates to a method for predicting the lifespan of intelligent valves based on big data analytics. Background Technology

[0002] In production processes with extremely high requirements for safety and continuity, such as petrochemicals and nuclear power, intelligent valves, as the core control units of fluid transport systems, are crucial for stable operation. Sudden failures of intelligent valves, such as internal leakage, jamming, or actuator malfunction, can not only lead to costly unplanned downtime and interrupt production processes, but may even trigger serious safety accidents. To prevent such incidents, modern industrial operation and maintenance systems are transitioning from traditional periodic maintenance to more efficient and predictive condition-based maintenance. By deploying multi-dimensional sensor arrays for vibration, temperature, pressure, and flow on intelligent valves, massive amounts of real-time operational data are collected, laying the foundation for accurate lifespan prediction and fault early warning.

[0003] In existing technologies, the LSTM (Long Short-Term Memory) model, as a deep learning model adept at processing time-series data, is widely used in equipment fault prediction based on sensor data. Its core idea is to train a predictive model using multi-dimensional sensor data collected under normal operating conditions. This model can learn and reproduce the complex intrinsic patterns of the evolution of various physical parameters over time when the equipment is in a healthy state. In actual monitoring, by comparing the real-time collected data with the model's predicted values, a reconstruction error can be calculated. Based on the magnitude of this reconstruction error's deviation from the normal range, the lifespan of the intelligent valve can be predicted.

[0004] However, when applying LSTM models, existing technologies typically involve simply concatenating time-series data from different sensors at the model's input. This approach essentially treats multiple physical quantities as independent, parallel information channels, with the model primarily learning the temporal evolution of each physical quantity, while ignoring the deep physical coupling and causal relationships between different physical quantities.

[0005] In valve systems, there is a close, pattern-driven linkage between valve core actuation commands, fluid pressure and flow rate, and actuator vibration. When a valve experiences early wear or slight jamming, this linkage pattern is disrupted first, potentially manifesting as a pressure response lagging behind valve core actuation, or the appearance of anomalous harmonic components in the vibration spectrum that are mismatched with the current pressure and flow rate. Traditional LSTM models, because they do not explicitly model this intervariate correlation, are insensitive to these subtle fault characteristics that only manifest through anomalies in multivariate relationships. This makes it impossible for the model to capture crucial information about the gradual degradation process from health to failure. Because this early degradation trend cannot be quantified, it is difficult to establish a reliable health state evolution model, leading to inaccurate lifespan predictions for smart valves. Summary of the Invention

[0006] To address the problem that existing technologies, when applying LSTM models, neglect the physical coupling relationships between multi-dimensional sensor data, leading to inaccurate remaining life predictions due to insensitivity to characterizing early progressive degradation trends, propose a smart valve life prediction method based on big data analysis. This method includes:

[0007] Taking any moment in the operation of the intelligent valve as the target moment, multiple data sequences at the target moment are acquired; among them, multiple data sequences correspond to data of different dimensions; for each data in each data sequence, a local window is constructed with the data as the center, and based on all the data within the local window, the local trend slope and local fluctuation degree of the data are determined to jointly constitute the dynamic primitive of the data;

[0008] For any two data sequences, the coherence of the two data sequences at each index is determined based on the difference between the dynamic primitives of the data in the same index, thus forming a coherence sequence of the two data sequences. Based on the change in coherence of adjacent data sequences in the coherence sequence, the degree of instability between the two data sequences is determined. All degrees of instability are aggregated to generate a potential anomaly index at the target time. The potential anomaly index is used to reflect the stability of the correlation pattern between all data sequences at the target time.

[0009] The data sequences corresponding to multiple target times when the smart valve is in a healthy state are used as samples to train the LSTM model. During the training process, the hidden state of the previous target time is optimized according to the potential anomaly index of each target time to obtain the optimized LSTM model. The evolution trend of reconstruction error over time is obtained through the optimized LSTM model to predict the lifespan of the smart valve.

[0010] This technical solution first generates dynamic primitives for each data point by constructing local windows, elevating isolated static values ​​into dynamic features capable of characterizing their local trends and fluctuations, thus providing richer feature inputs for subsequent relationship analysis. Second, in the feature fusion and analysis stage, the coherence between any two data sequences is quantified by calculating the difference in dynamic primitives at the same index, and the degree of instability is further calculated by analyzing the changes in coherent sequences. Finally, by aggregating all instability levels into a latent anomaly index, this method macroscopically and quantitatively assesses the stability of the correlation patterns between all data sequences at the current moment, effectively identifying latent early faults where individual variables appear normal, but the physical relationships between variables have been disrupted. Finally, in the prediction stage, this latent anomaly index is used as a key regulating factor, directly affecting the hidden states of the LSTM model. This means that the LSTM model's decisions not only depend on the temporal history but are also dynamically adjusted by the real-time stability of multivariate relationships. Therefore, the resulting reconstruction error is no longer a simple prediction bias, but an indicator that can simultaneously reflect the temporal anomaly of a single variable and the degree of mismatch in the relationship of multiple variables. The evolution trend of the reconstruction error can reveal the true degradation trajectory of the equipment earlier and more accurately, thus improving the accuracy of life prediction for smart valves.

[0011] Preferably, the local trend slope and the degree of local fluctuation are determined based on the following method:

[0012] For any data point in each data sequence, a local window is constructed based on that data. A weighted least squares method is used to linearly fit all data points within this local window, and the slope of the fitted line is taken as the local trend slope of that data point. Simultaneously, the weighted standard deviation of all data points within this local window is calculated as the degree of local volatility of that data point. When performing the weighted least squares method and calculating the weighted standard deviation, the weight of each data point within the local window is determined based on its distance from the center point. The weight distribution of all data points within the local window is set so that the center point has the highest weight, decreasing at equal intervals towards both sides.

[0013] Preferably, the coherence of any two data sequences at each index satisfies the following relationship:

[0014] ;

[0015] In the formula, and All are data sequences. for and index, for and In the index coherence, for In the index Data, for In the index Data, for The local trend slope, for The local trend slope, for The degree of local fluctuation, for The degree of local fluctuation, It is a natural exponential function. It is the absolute value symbol.

[0016] This technical solution calculates the difference between the local trend slope and the degree of local fluctuation of the data at the same index in two data sequences, and uses a negative exponential function for nonlinear mapping. It cleverly transforms this difference into a standardized value between 0 and 1, and through coherence, it can sensitively capture the consistency of the physical correlation patterns between multi-dimensional data.

[0017] Preferably, the degree of instability between any two data sequences is calculated based on the following method:

[0018] Obtain the coherence sequence of any two data sequences. In the coherence sequence, take the absolute value of the difference between adjacent coherence as the change between adjacent coherence. Sum the changes between all adjacent coherence and calculate the average value, and take the average value as the average fluctuation intensity. Take the normalized value of the average fluctuation intensity as the degree of instability between any two data sequences.

[0019] Preferably, the potential anomaly index at the target time is generated by aggregating all instabilities and calculating the mean of the instabilities between any two data sequences, and this mean is used as the potential anomaly index at the target time.

[0020] Preferably, during training, the optimization of the hidden state at the previous target time step based on the potential anomaly index at each target time step is performed according to the following relationship:

[0021]

[0022] In the formula, It is the first The hidden state of a target at any given time. It is the optimized version , This is an element-wise multiplication operation. For the first Potential anomaly index at each target time, and These are learnable scaling and translation parameters, used to adaptively adjust during model training. .

[0023] This technical solution incorporates a potential anomaly index reflecting the stability of correlations between multiple data sequences into the update mechanism of the LSTM model. Through element-wise multiplication, it amplifies anomalous features when the correlation is unstable and preserves normal patterns when the correlation is stable. This achieves deep integration between time-series information and correlation patterns, enabling the capture of weak signals in the early stages of equipment degradation where no single variable shows obvious anomalies but the correlation pattern has already mismatched. By optimizing hidden states, it amplifies these latent fault features, thereby improving the accuracy and timeliness of intelligent valve life prediction.

[0024] Preferably, the reconstruction error is determined based on the following method:

[0025] Multiple predicted data points for each target time step are obtained using an optimized LSTM model, with each predicted data point corresponding to one dimension. The multiple predicted data points for each target time step are combined into a predicted data set, and the multiple actual data points for each target time step are combined into an actual data set. The mean square error between the predicted data set and the actual data set is calculated, and this mean square error is determined as the reconstruction error for that target time step.

[0026] This technical solution first outputs the predicted values ​​of data in each dimension through the optimized LSTM model and forms a prediction set. Then, it is compared with the actual data set at the corresponding time. Finally, the mean square error is used as the reconstruction error. This method can not only integrate the prediction deviations of all sensor dimensions and avoid misjudgments caused by fluctuations in single-dimensional data, but also amplify the systematic deviations caused by the mismatch of multivariate correlation modes through the calculation characteristics of mean square error, so that the reconstruction error can more comprehensively reflect the transformation process of the valve from a healthy state to a degraded state.

[0027] Preferably, the method for obtaining the evolution trend of reconstruction error over time to predict the lifespan of the smart valve is as follows: A reconstruction error threshold that characterizes the end of the smart valve's lifespan is set; for any target time, the reconstruction error at that target time and the reconstruction errors at all target times within a preset time window preceding that target time are obtained to form a reconstruction error sequence; by analyzing the reconstruction error sequence, the evolution trend of the reconstruction error over time is obtained; if the evolution trend is an increasing trend, the time required for the reconstruction error at that target time to reach the reconstruction error threshold according to the increasing trend is determined as the remaining lifespan of the smart valve at that target time; if the evolution trend is a decreasing trend or a constant trend, the remaining lifespan of the smart valve at the previous target time is taken as the remaining lifespan of the smart valve at that target time.

[0028] This technical solution achieves accurate prediction of the remaining life of intelligent valves through a logical chain from threshold calibration to trend analysis and then to dynamic life calculation. This dynamic adjustment mechanism not only accurately captures the gradual degradation characteristics of equipment from health to failure, but also avoids misjudgment, providing a reliable decision reference for predictive maintenance.

[0029] Preferably, the evolution trend of the reconstruction error over time is obtained by performing linear regression fitting on the reconstruction error sequence to obtain a trend slope that characterizes the growth rate of the reconstruction error. If the trend slope is greater than 0, the evolution trend is an increasing trend; if the trend slope is less than 0, the evolution trend is a decreasing trend; and if the trend slope is equal to 0, the evolution trend is a constant trend.

[0030] Preferably, the multiple data sequences at the target time are obtained based on the following method:

[0031] Multiple sensors are deployed on the smart valve for synchronous data acquisition, including vibration sensors, temperature sensors, pressure sensors, and flow sensors. The data collected by each sensor at the target time, as well as the data within a preset time window before the target time, are combined to form a data sequence for the target time, thereby obtaining multiple data sequences for the target time.

[0032] The present invention has the following effects:

[0033] This invention generates a potential anomaly index to reflect the stability of the correlation patterns between different dimensions of data in a smart valve by quantifying the stability of the correlation patterns. The potential anomaly index is then used to dynamically modulate the internal hidden state of the LSTM model to optimize the LSTM model. This enables the LSTM model to keenly capture early, subtle faults where individual sensor data is normal but the physical coupling relationship has been disrupted, thus improving the accuracy of smart valve life prediction. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] This embodiment provides a method for predicting the lifespan of intelligent valves based on big data analysis, referring to... Figure 1 This includes the following steps:

[0037] S1: Acquire multiple data sequences of the smart valve at each target time.

[0038] First, multiple sensors are deployed on the smart valve, including vibration sensors, temperature sensors, pressure sensors, and flow sensors. The acquisition frequency is set to once every millisecond, and the data from multiple sensors are collected synchronously, so that data in these four dimensions can be obtained at any given moment.

[0039] During the operation of the intelligent valve, taking any given moment as the target moment, acquire the total data up to that target moment and the total data up to that point. The data from various sensors in milliseconds constitute multiple data sequences for that target time point, and each data sequence contains... Data points. Vibration sensors, temperature sensors, pressure sensors, and flow sensors in... The data collected in milliseconds form data sequences. Data sequence Data sequence and data sequence .

[0040] in, The empirical value is 500 milliseconds to 2000 milliseconds, for example, setting... The time range is 500 milliseconds (499 milliseconds before each target moment). This time range ensures that a complete valve switching action or fault characteristic can be captured with a high probability, while avoiding the computational burden and information interference caused by excessively long data sequences.

[0041] Considering that the collected data is easily contaminated with random noise caused by electromagnetic interference, outliers, and dimensional differences due to different sensor types, these defects in the raw data can seriously interfere with the accuracy of subsequent feature extraction and the stability of model training.

[0042] Therefore, for any data sequence, this step first removes noise and outliers through median filtering, then ensures data continuity through interpolation algorithms, and finally eliminates dimensional differences between different data streams through standardization. This process purifies and standardizes all data sequences, providing high-quality, consistent data input for subsequent steps, which is fundamental to ensuring the reliability of the entire prediction method. Specifically, this includes:

[0043] S11: Use a median filter to remove noise and outliers from each data sequence. Median filtering is a non-linear processing technique that is not sensitive to large-amplitude impulse noise or outliers caused by instantaneous electromagnetic interference. It can effectively filter out such contamination and at the same time, it better protects the inherent step or edge information in the data sequence than mean filtering, thus avoiding data distortion.

[0044] S12: For missing data in each data sequence, interpolation algorithms are used to fill in the gaps. After noise removal, for data loss that may be caused by network transmission or hardware failure, interpolation algorithms (such as linear interpolation) are used to fill in the gaps. This operation uses known data before and after the missing points to estimate unknown values, effectively restoring the continuity of the data sequence.

[0045] S13: Due to the significant differences in units and numerical ranges among different physical quantities, this step performs standardization on each data sequence to eliminate dimensional differences between them. Specifically, each data sequence is standardized using the Z-score standardization method. Standardization maps all data to a uniform scale, thereby eliminating the influence of dimensions on the calculation.

[0046] S2: Determine the dynamic primitives for each data in each data sequence.

[0047] Although the preprocessed data is clean, it is still isolated scalar values. In order to further increase the contextual information describing its dynamic behavior, this step constructs dynamic primitives for each data point by analyzing the data distribution in its local neighborhood, providing richer feature inputs for subsequent multivariate relationship analysis.

[0048] This step uses a sliding, weighted local window to extract two key dynamic dimensions for each data point: the local trend slope and the local fluctuation level. Together, these two dimensions constitute the dynamic primitives of the data.

[0049] Specifically:

[0050] For each data point in each data sequence, a local window is formed using that data point as the center, along with the four data points preceding and following it, for a total of nine data points. Four is an empirical value that can be adjusted as needed.

[0051] For example, with data sequences For example, As Index of data in In the index The data is ,by Centered on, and The previous four data points, and The following four data points constitute A local window.

[0052] Weighted least squares is used to perform linear fitting on all data within the local window, and the slope of the fitted line is taken as the local trend slope of the data. Simultaneously, the weighted standard deviation of all data within the local window is calculated as the degree of local volatility of the data. For example, ... The local trend slope is denoted as ,Will The degree of local fluctuation is denoted as .

[0053] Specifically, when performing weighted least squares fitting and calculating the weighted standard deviation, the weight of each data point within this local window is determined based on the following method:

[0054] Within a local window, the center point has the highest weight, and the weight decreases at equal intervals from the center point outwards. For example, the weight of the center point is set to 0.5, and the decrease increment is set to 0.1. Then, in... Within the local window, the weight of the center point is 0.5, the weights of the four data points forward from the center point are 0.4, 0.3, 0.2 and 0.1 respectively, and the weights of the four data points backward from the center point are 0.4, 0.3, 0.2 and 0.1 respectively.

[0055] By assigning higher weights to the center point, this approach focuses on the nearest neighbor information that best reflects the dynamic changes at the current moment, while effectively suppressing random disturbances from distant data. This ensures that the data most representative of the local trend slope is used when calculating the local trend slope. The trend of data changes at any given time is reflected in the data of its nearest neighbors, so it is given a larger weight, while the data located at the edge of the local window has a smaller impact and is given a smaller weight. When calculating the weighted standard deviation, we focus on the stability of the nearby data, which effectively suppresses the random disturbances of the data at the edge of the local window and provides a more reliable measure of local volatility.

[0056] Ultimately, the local trend slope and local volatility of each data point in each data sequence are used to construct the dynamic primitives of that data. For example, for The dynamic primitive is .

[0057] In summary, by calculating the dynamic primitives for each data point, static sensor data is transformed into tuples that describe its instantaneous dynamic behavior. These dynamic primitives include the local trend slope and the degree of local fluctuation. The local trend slope characterizes whether the data change trend within a local window is upward, downward, or stable; the degree of local fluctuation characterizes whether the data within a local window is stable. Thus, in subsequent multivariate relationship analysis, the comparison is no longer based on simple numerical values, but on the more informative dynamic behavior.

[0058] S3: Based on the dynamic primitives of the data, determine the degree of instability between data sequences and aggregate them into a potential anomaly index.

[0059] By evaluating the synchronicity and stability of the dynamic changes of different physical quantities, as well as the coupling relationship between different data sequences, a potential anomaly index is finally generated, which can macroscopically characterize the stability of the correlation patterns between different data sequences.

[0060] Specifically, it includes:

[0061] S31: Calculate the coherence of every two data sequences at the same index to obtain a coherent sequence.

[0062] Because the dynamic changes of different physical quantities (such as pressure and vibration) in a smart valve should follow certain inherent physical laws during normal operation, exhibiting a high degree of coherence, this coherence is disrupted when a fault occurs. Therefore, constructing an index that can measure this coherence in real time is the key to identifying relationship mismatch faults.

[0063] Therefore, for any two data sequences, the coherence of each data sequence at each index is determined based on the differences in the dynamic primitives of the data at the same index, thus forming a coherence sequence of the two data sequences. Since the data sequences are collected synchronously, the lengths and indices of any two data sequences are the same; the only difference is that they represent data of different dimensions.

[0064] For example, for and , In the index The data is , In the index The data is ,but and In the index The coherence is calculated based on the following formula:

[0065]

[0066] In the formula, and All are data sequences. for and index, for and In the index coherence, for In the index Data, for In the index Data, for The local trend slope, for The local trend slope, for The degree of local fluctuation, for The degree of local fluctuation, It is a natural exponential function. It is the absolute value symbol.

[0067] When the smart valve is operating normally, the dynamic primitives of different physical quantities tend to be the same, that is... The smaller the difference, The smaller the difference, the more... If the function constructs a negative correlation, then both the parts before and after the multiplication sign will be larger, leading to... and In the index coherence The larger the value, the closer it is to 1; when a smart valve malfunctions, the difference between the dynamic primitives of different physical quantities increases, i.e. The greater the difference, The greater the difference, the smaller both parts before and after the multiplication sign will be, leading to... and In the index coherence The smaller it is, the closer it is to 0.

[0068] because and All One data point, that is One index, and A coherence score is calculated for each index, resulting in a total of [amount missing]. A coherence, constituting and coherence sequence .

[0069] S32: Determine the degree of instability between data sequences by analyzing coherence sequences.

[0070] When a smart valve is in a stable operating state, the correlation pattern between its various physical quantities should be relatively stable. If the correlation pattern between the various physical quantities changes drastically or continuously, it usually indicates that the health status of the smart valve has deteriorated.

[0071] Therefore, this step analyzes the numerical changes in the coherence sequence of any two data sequences to determine whether the correlation pattern between the physical quantities represented by these two data sequences is stable, in order to preliminarily assess the health status of the smart valve.

[0072] Specifically, the coherence sequence of any two data sequences is obtained. In the coherence sequence, the absolute value of the difference between adjacent coherence sequences is taken as the change between adjacent coherence sequences. The changes between all adjacent coherence sequences are summed and the average value is calculated. This average value is taken as the average fluctuation intensity. The average fluctuation intensity is normalized and the normalized value is taken as the degree of instability between any two data sequences.

[0073] For example, based on calculate and The degree of instability between them:

[0074]

[0075] In this formula, for and The degree of instability between them For normalization operations, the Sigmoid function can be used to perform normalization, making... The value is in Within the range, for and The corresponding maximum index is also equal to and The total number of data contained in each for and In the index coherence, for and In the index coherence, It is the absolute value symbol.

[0076] In this formula, It is the absolute value of the difference between adjacent coherence values. It is the cumulative value of the changes between all adjacent coherences, divided by... This yields the average change in coherence across all adjacent pairs, reflecting... and The average volatility intensity of coherence across all indices; the higher the average volatility intensity, the stronger the correlation. and The greater the instability between two data sequences, the lower the instability, and vice versa. In short, by following the method in this step, we can obtain the degree of instability between any two data sequences.

[0077] S33: Aggregate all instabilities to determine potential anomaly indices.

[0078] Specifically, the mean of the instability between any two data sequences at the target time is used as the potential anomaly index at the target time. This index is used to characterize the stability of the correlation pattern between all data sequences at the target time. The larger the potential anomaly index, the worse the stability of the correlation pattern between all data sequences, indicating that the smart valve is more likely to malfunction at the target time. Conversely, the smaller the index, the more likely the smart valve is to be in a healthy state at the target time.

[0079] S4: Incorporate the potential anomaly index of each target time step into the training process of the LSTM model to optimize the hidden state of the previous target time step, so as to obtain the optimized LSTM model.

[0080] Since the potential anomaly index can characterize the stability of the correlation patterns between various physical quantities of a smart valve as a whole, this step further dynamically integrates the potential anomaly index into the LSTM process, thereby constructing an enhanced prediction model that is more sensitive to relational mismatch faults.

[0081] Based on the general knowledge of LSTM models, a standard LSTM model mainly updates and transmits information at each time step through three gates (forget gate, input gate, and output gate) and a cell state (a vector composed of multiple elements) to perform data prediction. The cell state is the memory core of LSTM, responsible for long-term information storage, while the hidden state is the working memory and external interface of LSTM, responsible for making specific judgments and outputs at each time step using long-term memory and current input. The two work together.

[0082] In one example, the update mechanism of the LSTM model is as follows:

[0083] exist At time t, the input vector is The hidden state at the previous moment was The cell state at the previous moment was The update process of the LSTM model is as follows:

[0084] S41: The forget gate determines which information to discard from the cell state of the previous moment.

[0085] The model views the hidden state from the previous time step. (Short-term memory) and current input and the cell state at the previous moment. Each element in the algorithm outputs a value between 0 and 1, where 1 represents complete retention and 0 represents complete discard. The output determines what to forget, and this decision is then applied to the cell state at the previous time step. (Long-term memory) refers to its application in updating cell states.

[0086] The core calculation formula for the forgetting gate is:

[0087]

[0088] In this formula, Representing the Gate of Oblivion The output vector at time step 1, where each element is a value between 0 and 1, determines the cell state at the previous time step. Should the information of the corresponding element in the code be retained or discarded? This represents the Sigmoid activation function, which is used to... The calculation result is used as input, and each element of the calculation result is compressed to between 0 and 1, thus obtaining the final output vector. , yes The hidden state at all times yes The input vector at time t, Indicates will Hidden state at any moment and The input vectors at time t are concatenated to obtain a concatenated vector. The weight matrix, representing the forget gate, is a set of parameters learned by the model during training. It determines the importance of the input information. The weight matrix defines a linear transformation that maps the concatenated vector to a low-dimensional output vector (usually the same dimension as the hidden state at the previous time step). In this transformation, each element of the concatenated vector contributes to each element of the output vector according to the corresponding weight in the matrix. The bias vector representing the forget gate is a parameter learned during model training, providing additional adjustment capabilities for the model's computation. The multiplication of a matrix and a vector results in a new vector, where the sum of the number of elements in the new vector is... The number of elements is the same. It is to combine the new vector and By performing element-wise addition, we obtain the forget gate. The output vector at time t, Is the Gate of Oblivion in Each element of the output vector at time step 1 is transformed by the Sigmoid activation function to... Within the range.

[0089] In summary, the core operation of the forget gate can be broken down into two steps:

[0090] The first step is a linear transformation: This step includes a learnable weight matrix. and bias vector The second step is to perform a nonlinear transformation using an activation function. The result from the first step is input into the Sigmoid function for the nonlinear transformation.

[0091] In neural networks, a layer typically refers to a complete computational unit that contains not only an activation function but also learnable parameters. This structural unit, composed of linear transformations and activation functions, is usually called a layer. Because the activation function of this layer is the sigmoid function, it is simply called a sigmoid layer. If the activation function of this layer is... If the function is used, then this layer can also be simply referred to as... layer.

[0092] S42: The input gate determines which new information is stored in the cell state.

[0093] Specifically, it includes two parallel computing processes:

[0094] The first process involves using a Sigmoid layer to determine which values ​​will be updated. The core calculation formula is as follows:

[0095]

[0096] In this formula, The input gate represents The output vector at time step 1, where each element is a value between 0 and 1, is used to determine... Candidate vector at time step The importance of the corresponding elements, Represents the Sigmoid activation function. yes The hidden state at all times yes The input vector at time t, Indicates will Hidden state at any moment and The input vectors at time t are concatenated to obtain a concatenated vector. The weight matrix representing the input gates is a set of parameters learned by the model during training, which determines the importance of the input information. The bias vector representing the input gate is also a parameter learned during model training, providing additional adjustment capabilities for computation.

[0097] The second process is through a The layer creates a new candidate vector, and the core calculation formula is:

[0098]

[0099] In this formula, represent The candidate vector at time step contains all the information to be updated. It is the hyperbolic tangent function, used here as the activation function. Its function is to compress the elements of the candidate vector to a specific size. Within the range, It is the weight matrix set to generate candidate vectors; these are the parameters learned by the model during training. It is a bias vector set to generate candidate vectors, and it is also a parameter learned by the model during training. yes The hidden state at all times yes The input vector at time t, Indicates will Hidden state at any moment and The input vectors at time t are concatenated to obtain a concatenated vector.

[0100] S43: Update cell state.

[0101] Specifically, the old cellular state With new candidate vectors Combine, generate Cellular state at any given moment The core calculation formula is:

[0102]

[0103] In this formula, represent Cellular state at any given moment The forgetting gate is the part that stores the cell state from the previous moment. With the Gate of Oblivion Output vector at time step By performing element-wise multiplication, we determine which parts of the old memory to forget. This indicates an element-wise multiplication operation. The function of the input gate is to... Candidate vectors generated at time step With input gate Output vector at time step Element-wise multiplication is performed to determine which information to incorporate from the new information. This involves adding the old memories (after forgetting) and the new memories (after adoption) element by element to obtain the final updated cell state. .

[0104] S44: Output the prediction result through the output gate.

[0105] Specifically, it includes two sequential calculation processes:

[0106] The first process uses a sigmoid layer to determine which part of the cell state will be output. The core calculation formula is as follows:

[0107]

[0108] In this formula, Represents the output gate in The output vector at each time step is used to control the degree of output of corresponding information in the cell state. This represents the Sigmoid activation function, which ensures that each element in the output vector is between 0 and 1. The weight matrix representing the output gate consists of parameters learned by the model during training. The bias vector representing the output gate is also a parameter learned by the model during training. yes The hidden state at all times yes The input vector at time t, Indicates will Hidden state at any moment and The input vectors at each time step are concatenated.

[0109] The second process is to generate the final output, which yields... The prediction result at time, that is The hidden state at time t, the core calculation formula is:

[0110]

[0111] In this formula, for The prediction results for the time period, Represents the output gate in The output vector at time t, for Cellular state at any given moment The hyperbolic tangent function is used here as the activation function to... Each element is compressed to Within the range, This indicates an element-wise multiplication operation.

[0112] In LSTM, Hidden state of time It has two important functions:

[0113] One is the internal short-term memory, which contains all the condensed information from the current moment in the historical sequence. This information is passed to the next time step as one of the inputs to the computation process at that time step. Without this passing process, the LSTM model loses its short-term memory function and cannot process sequential data; this is why it is called "hidden." The other is the current output representing the external context. This represents the prediction result obtained by the prediction model after analyzing the historical sequence at the current moment, reflecting the state at the current moment, which is why it is called "state".

[0114] The analysis of steps S41 to S44 demonstrates the core update mechanism of the LSTM model. Multivariate prediction is the basic capability of LSTM. The core principle of LSTM is to learn the patterns in a vector sequence. At each time step, LSTM receives an input vector, which can be one-dimensional or multi-dimensional. After the core update mechanism, LSTM outputs the hidden state at that time step, which is also a vector. Each element in the vector is a one-dimensional prediction value.

[0115] This step incorporates the potential anomaly index at each moment during the operation of the smart valve into the internal state update mechanism of the LSTM model, so as to optimize the hidden state of the LSTM model in real time and obtain the optimized model.

[0116] Specifically, the optimization process is as follows:

[0117] The smart valve collects data sequences at 1000 target moments in a healthy state using multiple sensors. Specifically, referring to step S1, for each target moment in the healthy state, multiple data sequences for that target moment are acquired, including one data sequence. A data sequence A data sequence and a data sequence Then, 1000 target moments in a healthy state will yield a total of 1000. 1000 1000 and 1000 For each of the four data sequences at each target time, the potential anomaly index for each target time is calculated according to steps S1 to S3, resulting in a total of 1000 potential anomaly indices. These data are then used as training samples to train the LSTM model.

[0118] During training, for each target time step in a healthy state, the data collected by all sensors at that target time step are used to construct a multi-dimensional vector, resulting in a sequence of 1000 multi-dimensional vectors as input vectors. This input vector sequence is then fed into the LSTM model. Based on the input vectors at each target time step and following the model update mechanism in steps S41 to S44, the LSTM model learns the weight matrices and bias vectors of each gate unit and determines the prediction result for each target time step—that is, the hidden state at each target time step.

[0119] The optimization of this step is as follows:

[0120] During training, when determining the prediction result for each target time step according to the model update mechanism from steps S41 to S44, the potential anomaly index for that target time step is obtained. This potential anomaly index is then incorporated into the model update mechanism. Specifically, when executing steps S41 to S44, the potential anomaly index for that target time step is used to optimize the "hidden state of the previous time step" involved in the model update mechanism. Here, each time step from S41 to S44 represents a specific target time step, for example... Indicates the first At the target moment, Indicates the first The target time point. Specific optimization is based on the following relationship:

[0121]

[0122] In the formula, It is the first The hidden state of a target at any given time. It is the optimized version , This is an element-wise multiplication operation. For the first Potential anomaly index at each target time and These are learnable scaling and translation parameters, used to adaptively adjust during model training. Exemplary and The initial value is .

[0123] in, The meaning is to The calculated value (a scalar) and the first The hidden state at each target time Multiply all elements one by one, and Updated to This achieved the goal of The optimization is equivalent to a dynamic gating, which allows information representing stability to directly regulate the flow intensity of short-term memory within the LSTM, thereby enhancing the dynamic coherence of the model.

[0124] During the training process, each time steps S41 to S44 are executed, the involved steps will be... Replace all with Finally, the number was obtained. The prediction result for the target time, i.e., the first target time. The hidden state at the nth target time, the th The prediction result for the target time step is a 4-dimensional vector, where each element corresponds to a one-dimensional predicted data point, thus yielding the prediction result for the target time step. Multiple prediction data for a target time.

[0125] Through this operation, a potential anomaly index calculated by external dynamic coherence is established for each gating unit from the hidden state of the previous target time to the current target time. The weights of short-term memory can be adaptively adjusted in real time according to the potential anomaly index of the system, so that each state update of each gating unit in the entire LSTM model is highly sensitive to the changes in dynamic coherence.

[0126] In this way, the model's ability to perceive and utilize dynamic coherence is enhanced. Instead of allowing the LSTM network to passively and implicitly learn the relationship between variables from the data, a real-time optimization mechanism is introduced. The potential anomaly index is an indicator calculated by an external module. This indicator is specifically used to measure the degree of mismatch between the relationship between multiple sensor data. When the system is stable and the data change pattern is normal, the potential anomaly index is smaller. When the relationship is abnormal (for example, pressure and flow no longer match), the potential anomaly index is larger.

[0127] Through learnable parameters and In addition, element-wise multiplication operations were implemented to achieve a dynamic gating system. and These are learnable parameters, meaning the model can learn during training how best to utilize the latent anomaly index. The element-wise multiplication operation is equivalent to using... The calculated scalar is used to directly control the influence of each element in the short-term memory vector.

[0128] Specifically, during training, to minimize prediction error, the model suppresses the influence of short-term memory from the previous time step on the current prediction when the potential anomaly index is high, and strengthens the influence of short-term memory from the previous time step on the current prediction when the potential anomaly index is low. When the potential anomaly index increases, the model automatically reduces it through backpropagation. and The value of the potential anomaly index is adjusted to suppress or weaken the influence of short-term memory from the previous target time on the current prediction result, as its reliability is reduced; when the potential anomaly index decreases, the model automatically increases it through the backpropagation algorithm. and The value of this value is increased to amplify the influence of short-term memory from the previous target time on the current prediction result, because its reliability is improved. The model learns this autonomously during training. and The value is used to intelligently achieve the best training effect.

[0129] Meanwhile, during training, the reconstruction error is set as the loss function to determine the termination condition for model training. Specifically:

[0130] Based on the prediction results and input vectors at each target time step, the reconstruction error for that target time step is calculated. Model training is complete when the reconstruction error is minimized. Each target time step's prediction result and input vector are both 4-dimensional vectors. Each element of the prediction result corresponds to one dimension of predicted data, and each element of the input vector corresponds to one dimension of actual value data. All predicted data for that target time step are combined into a prediction dataset, and all actual data for that target time step are combined into an actual dataset. The mean squared error between the prediction dataset and the actual dataset is calculated and used as the reconstruction error for that target time step. The model parameters corresponding to the minimum reconstruction error are then selected as the optimal parameters, and model training is complete.

[0131] In summary, through the optimization of this step, the LSTM model learns the collaborative working mode between physical quantities such as vibration, pressure and flow of the smart valve in a healthy state. The reconstruction error it calculates is not only a mathematical deviation, but also an indicator that can clearly reflect the health status of the smart valve.

[0132] When early wear of a smart valve causes the valve core action and pressure response to become uncoordinated, even if the readings of each sensor are not exceeded, the trained model will capture those hidden faults that cannot be detected by a single parameter threshold by analyzing whether the cooperative working mode has been disrupted, and obtain accurate prediction results, thus providing a more comprehensive method for assessing the health status of the equipment.

[0133] S5: Use the optimized LSTM model to obtain the evolution trend of reconstruction error over time in order to predict the lifespan of smart valves.

[0134] Since the reconstruction error of the LSTM model at each target time can reflect the degree of mismatch between the various physical quantities of the smart valve at that target time in real time, the larger the reconstruction error, the greater the degree of mismatch, and the less ideal the health status of the smart valve at that target time, and vice versa. Therefore, this step analyzes the evolution trend of the reconstruction error over a continuous period of time, and accurately predicts the lifespan of the smart valve based on the evolution trend.

[0135] S51: Set a reconstruction error threshold that can characterize the end of the lifespan of the smart valve.

[0136] Since the LSTM model is trained using sensor data from multiple target time points within a healthy state of the smart valve, and the reconstruction error is calculated for each target time point during training, after model training is complete, the reconstruction errors from all target time points during training are obtained. Based on the statistical distribution characteristics of these reconstruction errors, a reconstruction error threshold that characterizes the end of the smart valve's lifespan is set. For example, the 75th percentile of these reconstruction errors is first obtained as the baseline reconstruction error. And in accordance with safety requirements The reconfiguration error threshold serves as a characterization of the end of the lifespan of a smart valve.

[0137] S52: Obtain the evolution trend of reconstruction error over time and perform lifetime prediction.

[0138] In the process of using a trained LSTM model to predict the lifespan of a smart valve, for any target time, the reconstruction error at that target time and the reconstruction error sequence consisting of the reconstruction errors of all target times within a preset time window before that target time are obtained. An example preset time window is 24 hours.

[0139] Then, by analyzing the reconstruction error sequence, the evolution trend of the reconstruction error over time is obtained. Specifically, the least squares method is used to perform linear regression fitting on the reconstruction error sequence to obtain a trend slope that characterizes the growth rate of the reconstruction error. That is, the linear regression fitting yields a straight line equation, and the slope of this straight line is used as the trend slope here.

[0140] If the trend slope is greater than 0, the evolution trend is an increasing trend, indicating that all reconstruction errors contained in the reconstruction error sequence show an increasing trend over time. This means the difference between the predicted and actual values ​​of the data collected by the multiple sensors of the smart valve is continuously increasing. Since the predicted values ​​are based on data trained on the smart valve in a healthy state, the continuously increasing difference between the predicted and actual values ​​indicates that the health of the smart valve is deteriorating. At this point, the remaining lifespan of the smart valve at the target time is:

[0141]

[0142] In this formula, It represents the remaining lifespan of the smart valve at that target moment. It is the reconstruction error (mean square error) at that target time. It is the reconstruction error threshold. The trend slope represents the rate of error growth. This method of calculating remaining life transforms the abstract equipment degradation process into a quantifiable mathematical model, realizing the transition from condition monitoring to life prediction. By establishing a trend model of the reconstruction error's growth over time, it estimates how much time is left before the preset lifespan end. This prediction method based on linear extrapolation provides a direct decision-making basis for implementing predictive maintenance strategies, enabling maintenance plans to shift from reactive, reactive responses or fixed periodic maintenance to proactive intervention based on model predictions.

[0143] If the trend slope is less than 0, the evolution trend is decreasing, indicating that all reconstruction errors contained in the reconstruction error sequence are decreasing over time, suggesting that the health of the smart valve is recovering. If the trend slope is equal to 0, the evolution trend is unchanged, indicating that the health of the smart valve remains stable. However, since mechanical wear of equipment is usually irreversible, equipment performance gradually degrades over time, and the remaining lifespan of the equipment gradually shortens. When the evolution trend is decreasing or unchanged, the update of the remaining lifespan is usually paused, and the lifespan prediction result of the smart valve at the previous target time is directly used as the lifespan prediction result for this time. That is, the remaining lifespan of the smart valve at the previous target time is taken as the remaining lifespan of the smart valve at this target time, thus avoiding the unreasonable situation of incorrectly predicting an increase in lifespan due to short-term data fluctuations.

[0144] In summary, this step establishes an intelligent decision-making system for the health status of smart valves, combining sensitive diagnosis with robust judgment. From a physical perspective, when irreversible wear or aging occurs inside the smart valve, its operating mode will continuously deviate from the healthy baseline. This is manifested in the data as a stable increase in reconstruction error over time, i.e., a positive trend slope. This step can sensitively capture this degradation information and calculate the remaining lifespan based on its growth rate. More importantly, this step also fully considers the pseudo-healthy state that the equipment may exhibit in the short term due to changes in operating conditions, i.e., a negative or zero trend slope. In this case, it will not mistakenly judge that the equipment lifespan has increased, but will instead adopt a conservative strategy, using the previous effective prediction results. This design can provide forward-looking warnings when the equipment clearly deteriorates, and can also avoid misjudgments due to short-term data fluctuations when the state is stable or recovering, thus improving the accuracy and reliability of smart valve lifespan prediction.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the lifespan of intelligent valves based on big data analysis, characterized in that, include: Taking any moment in the operation of the intelligent valve as the target moment, multiple data sequences at the target moment are acquired; among them, multiple data sequences correspond to data of different dimensions; for each data in each data sequence, a local window is constructed with the data as the center, and based on all the data within the local window, the local trend slope and local fluctuation degree of the data are determined to jointly constitute the dynamic primitive of the data; For any two data sequences, the coherence of the two data sequences at each index is determined based on the difference between the dynamic primitives of the data in the same index, thus forming a coherence sequence of the two data sequences. Based on the change in coherence of adjacent data sequences in the coherence sequence, the degree of instability between the two data sequences is determined. All degrees of instability are aggregated to generate a potential anomaly index at the target time. The potential anomaly index is used to reflect the stability of the correlation pattern between all data sequences at the target time. The data sequences corresponding to multiple target times when the smart valve is in a healthy state are used as samples to train the LSTM model. During the training process, the hidden state of the previous target time is optimized according to the potential anomaly index of each target time to obtain the optimized LSTM model. The evolution trend of reconstruction error over time is obtained through the optimized LSTM model to predict the lifespan of the smart valve. The coherence of any two data sequences at each index satisfies the following relationship: In the formula, and All are data sequences. for and index, for and In the index coherence, for In the index Data, for In the index Data, for The local trend slope, for The local trend slope, for The degree of local fluctuation, for The degree of local fluctuation, It is a natural exponential function. It is the absolute value symbol; Multiple data sequences at the target time are obtained by deploying multiple sensors on the smart valve for synchronous data acquisition, including vibration sensors, temperature sensors, pressure sensors, and flow sensors; the data collected by each sensor at the target time, as well as the data within a preset time window before the target time, constitute a data sequence at the target time, thereby obtaining multiple data sequences at the target time.

2. The intelligent valve life prediction method according to claim 1, characterized in that, The local trend slope and the degree of local fluctuation are determined based on the following method: For any data point in each data sequence, obtain the local window constructed from that data point; We use weighted least squares to perform linear fitting on all data within the local window, and use the slope of the fitted line as the local trend slope of the data. At the same time, we calculate the weighted standard deviation of all data within the local window as the local volatility of the data. In the process of performing weighted least squares and calculating weighted standard deviation, the weight of each data point within the local window is determined based on its distance from the center point. The weight distribution of all data within the local window is set such that the center point has the highest weight, and the weight decreases at equal intervals to both sides.

3. The intelligent valve life prediction method according to claim 1, characterized in that, The degree of instability between any two data sequences is calculated based on the following method: Obtain the coherence sequence of any two data sequences. In the coherence sequence, take the absolute value of the difference between adjacent coherence as the change between adjacent coherence. The summation of the changes between all adjacent coherences and the calculation of the average value are used as the average fluctuation intensity. The normalized value of the average fluctuation intensity is used as the degree of instability between any two data sequences.

4. The intelligent valve life prediction method according to claim 1, characterized in that, The potential anomaly index at the target time is generated by aggregating all instabilities. This is achieved by calculating the mean of the instabilities between any two data sequences and using this mean as the potential anomaly index at the target time.

5. The intelligent valve life prediction method according to claim 1, characterized in that, During training, the optimization of the hidden state at the previous target time step based on the potential anomaly index at each target time step is performed according to the following relationship: ； In the formula, It is the first The hidden state of a target at any given time. It is the optimized version , This is an element-wise multiplication operation. For the first Potential anomaly index at each target time and These are learnable scaling and translation parameters, used to adaptively adjust during model training. .

6. The intelligent valve life prediction method according to claim 1, characterized in that, The reconstruction error is determined based on the following method: Multiple prediction data points for each target time step are obtained using the optimized LSTM model, with each prediction data point corresponding to one dimension; Multiple predicted data points at each target time are combined into a predicted data set, and multiple actual data points at each target time are combined into an actual data set. Calculate the mean square error between the predicted data set and the actual data set, and determine the mean square error as the reconstruction error at the target time.

7. The intelligent valve life prediction method according to claim 1, characterized in that, The method for predicting the lifespan of smart valves by obtaining the evolution trend of reconstruction error over time is as follows: Set a reconstruction error threshold that can characterize the end of the life of the smart valve; for any target time, obtain the reconstruction error at that target time and the reconstruction error of all target times within a preset time window before that target time to form a reconstruction error sequence; by analyzing the reconstruction error sequence, obtain the evolution trend of the reconstruction error over time. If the evolution trend is an increasing trend, the time required for the reconstruction error at the target time to reach the reconstruction error threshold according to the increasing trend is determined as the remaining lifespan of the smart valve at that target time; if the evolution trend is a decreasing trend or a constant trend, the remaining lifespan of the smart valve at the previous target time is taken as the remaining lifespan of the smart valve at that target time.

8. The intelligent valve life prediction method according to claim 7, characterized in that, The evolution trend of reconstruction error over time is obtained by linear regression fitting based on the reconstruction error sequence, which yields a trend slope that characterizes the growth rate of reconstruction error. If the trend slope is greater than 0, the evolution trend is an increasing trend; if the trend slope is less than 0, the evolution trend is a decreasing trend; and if the trend slope is equal to 0, the evolution trend is a constant trend.

Images