A method, system, device and medium for detecting blockage of a flow measuring device

Abstract

A method, system, device, and medium for detecting a blockage in a flow measurement device relate to the field of blockage detection. In this method, the original pressure signal and original flow signal of the flow measurement device to be measured are obtained, and an eigenmode function is generated based on the original pressure signal and the original flow signal. The time vector, space vector, and frequency domain vector of the eigenmode function are extracted, and attention weights are obtained through a two-way cross-attention mechanism. The time vector, space vector, frequency domain vector, and attention weights are weighted and fused to obtain a multi-dimensional feature vector. The multi-dimensional feature vector is input into a pre-trained deep learning-based blockage probability prediction model to obtain a blockage probability value. A state space is constructed based on the blockage probability value, historical maintenance records, and equipment operating time. The execution operation is determined by balancing false alarm and missed detection penalties in combination with a reward function. Implementing the technical solution provided by this application improves the accuracy and reliability of blockage detection.

Description

Technical Field

[0001] The present application relates to the technical field of blockage detection, and in particular to a blockage detection method, system, equipment and medium for a flow measurement device. Background Art

[0002] Flow measurement is crucial in numerous fields, including industrial production and fluid transportation. Accurate flow measurement helps companies effectively monitor and control production processes, improve productivity, reduce energy consumption, and ensure product quality. With the continuous advancement of science and technology, flow measurement technology has also made significant progress. From traditional mechanical measuring instruments to today's intelligent sensors, measurement accuracy and stability have been greatly improved.

[0003] Blockage detection remains a key issue in the daily use of flow measurement devices. Currently, direct observation is a common method for detecting blockage in flow measurement devices. This method relies on operators regularly inspecting the device, visually inspecting the internal and external conditions to determine if there is any blockage. Another alternative is pressure differential detection, which monitors the pressure difference between the inlet and outlet of the device. Any abnormal change in the pressure differential indicates a possible blockage.

[0004] However, these existing blockage detection methods have significant drawbacks. Direct observation is highly subjective and susceptible to operator experience and visual errors. Furthermore, it can be difficult to detect blockages in hidden locations. Pressure differential detection and flow comparison methods can only reflect overall pressure and flow changes, failing to accurately determine the specific cause and extent of the blockage. Furthermore, they are susceptible to interference from external factors in practical applications, leading to frequent misjudgments and missed detections. Summary of the Invention

[0005] The present application provides a congestion detection method, system, device and medium for a flow measurement device, which improves the accuracy and reliability of congestion detection by combining multiple factors to construct a state space and using a reward function to balance the penalties for false alarms and missed detections.

[0006] In a first aspect of the present application, a method for detecting a blockage of a flow measurement device is provided, which is applied to a blockage detection platform. The method comprises:

[0007] Acquire an original pressure signal and an original flow signal of a flow measurement device to be measured, and generate an eigenmode function according to the original pressure signal and the original flow signal;

[0008] Extracting the time vector, space vector, and frequency domain vector of the intrinsic mode function, obtaining an attention weight through a two-way cross attention mechanism, and weightedly fusing the time vector, the space vector, the frequency domain vector, and the attention weight to obtain a multi-dimensional feature vector;

[0009] Inputting the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model to obtain a congestion probability value;

[0010] A state space is constructed based on the congestion probability value, historical maintenance records, and equipment operation time. The reward function is combined to balance the penalties for false alarms and missed detections to determine the execution action.

[0011] Optionally, generating an eigenmode function according to the original pressure signal and the original flow signal includes:

[0012] adding white noise of different amplitudes to the original pressure signal and the original flow signal multiple times to form multiple noisy signal sequences;

[0013] Decompose each noisy signal sequence separately to obtain multiple intrinsic mode function components;

[0014] Determine the sliding time window size and step size based on the signal sampling frequency and congestion detection requirements, and calculate the cross-correlation coefficient between the target intrinsic mode function component and the preset reference signal within each sliding time window, wherein the target intrinsic mode function component is any one of the multiple intrinsic mode function components;

[0015] Recording the number of sliding time windows in which the cross-correlation coefficient is lower than a preset coefficient threshold, when the number is greater than or equal to the preset number threshold, determining the target intrinsic mode function component as an abnormal mode component caused by blockage; when the number is less than the preset number threshold, determining the target intrinsic mode function component as a normal mode component;

[0016] The intrinsic mode functions are constructed based on normal mode components.

[0017] Optionally, decomposing each noisy signal sequence separately to obtain multiple intrinsic mode function components includes:

[0018] Identify all the maximum and minimum points from the noisy signal sequence, use interpolation methods to construct the upper envelope based on all the maximum points, and construct the lower envelope based on all the minimum points;

[0019] Calculating a mean curve of the upper envelope and the lower envelope, and subtracting the mean curve from the noisy signal sequence to obtain a new signal sequence;

[0020] It is determined whether the new signal sequence meets a preset condition of an eigenmode function; if the preset condition is met, the new signal sequence is determined as an eigenmode function component.

[0021] Optionally, extracting the time vector, space vector, and frequency domain vector of the intrinsic mode function includes:

[0022] Sampling the intrinsic mode function on the time axis, recording the function value corresponding to each sampling point, and arranging these function values ​​in chronological order to form a time vector;

[0023] quantizing the energy or amplitude of the intrinsic mode function at different frequency components, using the quantized values ​​as elements of a vector with frequency as a dimension, and constructing a space vector;

[0024] The intrinsic mode function is converted from the time domain to the frequency domain using Fourier transform to obtain a frequency domain representation of the signal in the frequency domain, the target features of the frequency domain representation are extracted, and the target features are arranged in a preset order to form a frequency domain vector.

[0025] Optionally, obtaining the attention weight through the two-way cross attention mechanism includes:

[0026] In a first path, the time vector is used as a query vector, and the space vector is used as a key vector and a value vector; in a second path, the time vector is used as a query vector, and the frequency domain vector is used as a key vector and a value vector;

[0027] Performing cross attention calculation on the first path and the second path to obtain a first attention weight and a second attention weight;

[0028] Different weights are assigned according to the importance of the first path and the second path, and the first attention weight and the second attention weight are linearly combined according to the weights to obtain attention weights, which include time vector weights, space vector weights and frequency domain vector weights.

[0029] Optionally, obtaining the attention weight through the two-way cross attention mechanism includes:

[0030] In the first path, the similarity matrix between the query vector and the key vector is calculated, normalized, and multiplied by the value vector to obtain the first attention output:

[0031] In the second path, a multi-head attention mechanism is used to split the query vector into multiple head subspaces, which are calculated in parallel with the key vector and value vector respectively. The second attention output is obtained by splicing the multi-head outputs.

[0032] The first attention output and the second attention output are input into the gated fusion module, and the contribution of the two-way information is dynamically adjusted through the weight parameter to obtain the attention weight.

[0033] Optionally, constructing the state space according to the congestion probability value, historical maintenance records, and equipment operation time includes:

[0034] The device's runtime is segmented, each segmented interval is assigned a unique code value, and a runtime state code vector is generated based on the interval to which the device's actual runtime belongs.

[0035] generating a maintenance cycle matrix according to the historical maintenance records, and performing singular value decomposition on the maintenance cycle matrix to obtain a maintenance feature vector;

[0036] Performing tensor splicing on the congestion probability value, the maintenance feature vector, and the running time state encoding vector, and generating a time series correlation feature according to the spliced ​​tensor;

[0037] A state space is generated according to the temporal correlation features.

[0038] In a second aspect of the present application, a blockage detection system for a flow measurement device is provided, comprising an acquisition module, a vector module, a prediction module, and an execution module, wherein:

[0039] an acquisition module configured to acquire an original pressure signal and an original flow signal of a flow measurement device to be measured, and generate an eigenmode function according to the original pressure signal and the original flow signal;

[0040] a vector module configured to extract a time vector, a space vector, and a frequency domain vector of the intrinsic mode function, obtain an attention weight through a two-way cross attention mechanism, and perform weighted fusion on the time vector, the space vector, the frequency domain vector, and the attention weight to obtain a multi-dimensional feature vector;

[0041] a prediction module configured to input the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model to obtain a congestion probability value;

[0042] The execution module is configured to construct a state space according to the congestion probability value, historical maintenance records and equipment operation time, balance false alarm and missed detection penalties in combination with a reward function, and determine an execution operation.

[0043] In the third aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs any of the methods described above.

[0044] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions. When the instructions are executed, any one of the methods described above is executed.

[0045] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0046] 1. By acquiring the original pressure signal and the original flow signal and generating the intrinsic mode function, it is possible to decompose the signal components with different characteristic scales from the complex original signal, and more accurately capture the intrinsic characteristics of the signal. The time vector, space vector, and frequency domain vector of the eigenmode function are further extracted to comprehensively characterize the signal characteristics from multiple dimensions, providing rich and effective feature information for subsequent analysis. Compared with single-dimensional feature extraction methods, it can more accurately reflect the operating status of the flow measurement device;

[0047] 2. A two-way cross-attention mechanism is used to obtain attention weights, which are then weighted and fused with the time vector, space vector, and frequency domain vector to form a multi-dimensional feature vector. This fusion method can highlight important information in features of different dimensions and suppress irrelevant or interfering information. This allows the multi-dimensional feature vector to more accurately characterize the key features of the flow measurement device in terms of blockage detection, enhances the expressiveness of the features, and helps improve the accuracy of subsequent blockage detection.

[0048] 3. Input the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model. Leveraging the deep learning model's powerful nonlinear fitting and learning capabilities, it can learn the complex mapping relationship between features and congestion probability from a large amount of training data, thereby more accurately predicting the congestion probability value of the flow measurement device. Compared with traditional congestion detection methods, the prediction results are more reliable and accurate.

[0049] 4. The state space is constructed based on the blockage probability value, historical maintenance records, and equipment operating time, comprehensively considering multiple relevant factors, making the decision-making process more comprehensive and scientific. Combined with the reward function to balance the penalties for false alarms and missed detections, it can find a reasonable balance between avoiding false alarms (i.e., incorrectly judging the device as blocked, resulting in unnecessary maintenance operations) and missed detections (i.e., failing to discover the device blockage in time, affecting normal production operations). The most appropriate execution action is determined based on the actual situation, such as whether maintenance is required and the urgency of maintenance, effectively reducing maintenance costs while ensuring the normal operation of the flow measurement device.

[0050] 5. The entire method, from signal acquisition and feature extraction to congestion probability prediction and decision-making, forms a complete and efficient congestion detection process. Through automated signal processing and model prediction, it can quickly and accurately determine whether the flow measurement device is at risk of congestion. Compared with traditional manual detection methods, this method greatly improves detection efficiency, reduces errors caused by human intervention and subjective judgment, and improves the reliability of congestion detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method for detecting blockage in a flow measurement device disclosed in an embodiment of the present application;

[0052] Figure 2 This is a module schematic diagram of a blockage detection system for a flow measurement device disclosed in an embodiment of the present application;

[0053] Figure 3 This is a schematic structural diagram of an electronic device disclosed in an embodiment of the present application.

[0054] Explanation of the accompanying drawings: 201, acquisition module; 202, vector module; 203, prediction module; 204, execution module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0056] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0057] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0058] This embodiment discloses a method for detecting blockage of a flow measurement device. Figure 1 FIG. 1 is a flow chart of a method for detecting a blockage in a flow measuring device disclosed in an embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0059] S101, obtaining an original pressure signal and an original flow signal of a flow measurement device to be measured, and generating an eigenmode function according to the original pressure signal and the original flow signal;

[0060] S102, extracting the time vector, space vector, and frequency domain vector of the intrinsic mode function, obtaining an attention weight through a two-way cross attention mechanism, and weightedly fusing the time vector, the space vector, the frequency domain vector, and the attention weight to obtain a multi-dimensional feature vector;

[0061] S103: Inputting the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model to obtain a congestion probability value;

[0062] S104: Construct a state space based on the congestion probability value, historical maintenance records, and equipment operation time, balance false alarm and missed detection penalties in combination with a reward function, and determine an execution operation.

[0063] During operation, flow measurement devices generate physical signals related to pressure and flow. By installing appropriate sensors, such as pressure and flow sensors, on the device, the raw pressure and flow signals from the flow measurement device can be collected in real time. These raw signals contain rich information about the device's operating status, but they can also be affected by factors such as noise and interference, exhibiting complex and nonlinear characteristics. Eigenmode functions are generated using the Empirical Mode Decomposition (EMD) algorithm. EMD is an adaptive signal processing method that decomposes complex raw signals into several Intrinsic Mode Functions (IMFs). Each IMF meets two conditions: the number of extreme points and the number of zero crossings must be equal or differ by at most one across the entire data range; and at any given moment, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points must be zero. Through EMD decomposition, the original signal is decomposed into IMFs of different frequency components. These IMFs reflect the characteristics of the signal at different time scales, which helps to analyze the essential characteristics of the signal more deeply and provide a basis for subsequent feature extraction and blockage detection.

[0064] The time vector reflects the temporal characteristics of a signal. Time series data can be extracted from intrinsic mode functions (IMFs), analyzing information such as the signal's amplitude and changing trends at different time points, and quantifying this information into a time vector. For example, statistics such as the average, maximum, and minimum values ​​of each IMF over different time periods can be calculated, or the signal's time-domain waveform characteristics, such as rise and fall times, can be analyzed to construct a time vector. In flow measurement devices, the spatial vector may be related to the distribution or correlation of the signal at different spatial locations. If the device has multiple measurement points or sensors distributed at different locations, the relationship between sensor signals at different locations, such as signal correlation and phase difference, can be analyzed and quantified into a spatial vector. Even if the device itself is relatively simple, spatial characteristics can be considered from the perspective of signal propagation and reflection, and spatial vectors can be constructed to describe the spatial characteristics of the signal. Frequency domain vectors are used to describe the characteristics of a signal in the frequency domain. By applying frequency domain analysis methods such as Fourier transforms to the IMFs, the energy distribution and phase information of the signal at different frequency components can be obtained. Signal characteristics such as the dominant frequency, frequency bandwidth, and spectral energy can be extracted to construct a frequency domain vector to reflect the signal's frequency characteristics. The two-way cross-attention mechanism is a deep learning technique for processing multimodal data (such as time vectors, spatial vectors, and frequency domain vectors). It consists of two input paths, each processing different types of vector data. In each path, an attention weight is assigned to each vector element by calculating the similarity or correlation between the vectors. These attention weights reflect the importance of each vector element in the overall feature. The two-way cross-attention mechanism automatically learns the correlations between different vectors, enabling the model to focus on information that is important for congestion detection and suppress irrelevant or interfering information. The extracted time vector, spatial vector, and frequency domain vector are multiplied by the corresponding attention weights obtained through the two-way cross-attention mechanism. The weighted vectors are then concatenated or summed to generate a multidimensional feature vector. This weighted fusion approach integrates important information from features across different dimensions, enabling the multidimensional feature vector to more comprehensively and accurately represent the operating status and congestion characteristics of the flow measurement device, providing richer feature information for subsequent congestion probability prediction.

[0065] Congestion probability prediction models are built based on deep learning algorithms, typically employing neural network architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), their variants (such as LSTMs and GRUs), or a combination thereof. During the model training phase, the model is trained using a large amount of labeled data (including multi-dimensional feature vectors of flow measurement devices under different operating conditions and their corresponding congestion labels). This allows the model to learn the complex mapping relationship between the multi-dimensional feature vectors and congestion probability. The resulting multi-dimensional feature vectors are then input into a pre-trained congestion probability prediction model. The model processes and analyzes the input feature vectors, performing feature extraction, transformation, and mapping through a series of neural network layers (such as convolutional layers, pooling layers, and fully connected layers), ultimately outputting a congestion probability value. This congestion probability value indicates the probability of congestion at the flow measurement device and typically ranges from 0 to 1, with values ​​closer to 1 indicating a greater likelihood of congestion and values ​​closer to 0 indicating a lower likelihood.

[0066] The state space is a multidimensional space that comprehensively considers factors such as the jam probability, historical maintenance records, and equipment operating hours. The jam probability reflects the probability of a current device jam. Historical maintenance records contain information such as past maintenance time, maintenance content, and fault conditions, providing a reference for determining the current device state. Equipment operating hours reflect the device's service life and wear; longer operating hours increase the likelihood of a jam or other fault. By quantifying these three factors and combining them into a multidimensional vector, a state space is constructed, where each state point corresponds to a specific device operating state. The reward function is a mathematical function used to evaluate the effectiveness of different execution operations. In the jam detection scenario, both false alarms (incorrectly determining a device jam, leading to unnecessary maintenance) and missed detections (failure to promptly detect a jam, impacting normal production operations) need to be considered and penalized accordingly. The reward function can assign different rewards or penalties to different execution operations based on the degree of false alarms and missed detections. Based on the constructed state space and reward function, reinforcement learning or other decision-making algorithms are used to search for the optimal execution operation at different state points. Actions taken might include immediate maintenance, scheduling regular inspections, or continuing monitoring. By comprehensively considering the probability of congestion, historical maintenance records, and equipment operating time, combined with a reward function that balances false alarms and missed detections, the most appropriate action is determined for the current state, achieving the goal of reducing maintenance costs while ensuring the normal operation of the device.

[0067] Optionally, generating an eigenmode function according to the original pressure signal and the original flow signal includes:

[0068] adding white noise of different amplitudes to the original pressure signal and the original flow signal multiple times to form multiple noisy signal sequences;

[0069] Decompose each noisy signal sequence separately to obtain multiple intrinsic mode function components;

[0070] Determine the sliding time window size and step size based on the signal sampling frequency and congestion detection requirements, and calculate the cross-correlation coefficient between the target intrinsic mode function component and the preset reference signal within each sliding time window, wherein the target intrinsic mode function component is any one of the multiple intrinsic mode function components;

[0071] Recording the number of sliding time windows in which the cross-correlation coefficient is lower than a preset coefficient threshold, when the number is greater than or equal to the preset number threshold, determining the target intrinsic mode function component as an abnormal mode component caused by blockage; when the number is less than the preset number threshold, determining the target intrinsic mode function component as a normal mode component;

[0072] The intrinsic mode functions are constructed based on normal mode components.

[0073] During actual acquisition, raw pressure and flow signals are often affected by various noise and interference, resulting in poor signal quality and making accurate analysis difficult. By adding white noise multiple times to the raw signals, the signals are evenly distributed in the time-frequency space, mitigating the modal aliasing problem associated with EMD. Modal aliasing occurs when a single IMF contains signal components of different time scales, or when different IMFs contain signal components of the same time scale. This can lead to inaccurate signal decomposition and hinder subsequent analysis and detection. White noise of varying amplitudes is added to the raw pressure and flow signals multiple times according to a specific rule (such as randomly or in a specific sequence). Each addition of white noise forms a new noisy signal sequence. The number and amplitude of white noise additions should be appropriately configured based on the signal characteristics and actual requirements. Generally speaking, a greater number of additions results in more stable decomposition results, but the computational effort also increases. The white noise amplitude should be neither too large, otherwise it will mask the characteristics of the original signal, nor too small, otherwise it will fail to effectively mitigate modal aliasing.

[0074] Each noisy signal sequence is decomposed using the empirical mode decomposition algorithm. EMD is an adaptive signal processing method that decomposes complex signals into several intrinsic mode function (IMF) components and a residual component. Each IMF satisfies two conditions: the number of extreme points and the number of zero crossings must be equal or differ by at most one over the entire data range; and at any instant, the average value of the upper envelope formed by the local maxima and the lower envelope formed by the local minima must be zero. For each noisy signal sequence, all local maxima and local minima of the signal are first found. The upper and lower envelopes are then fitted using cubic spline interpolation. The average value of the upper and lower envelopes is calculated and subtracted from the original signal to obtain a new signal. This process is repeated until the new signal meets the IMF conditions, resulting in the first IMF component. The first IMF component is subtracted from the original signal to obtain a new signal. The decomposition process is repeated on this new signal until the remaining signal becomes a monotonic function or meets a certain stopping condition, resulting in all IMF components and the residual component.

[0075] The sliding time window size and step size are determined based on the signal sampling frequency and congestion detection requirements. A sliding time window is used to perform localized signal analysis, enabling observation of changes in signal characteristics over different time segments. The sampling frequency determines the signal's temporal resolution. A higher sampling frequency results in higher temporal resolution, but also increases the amount of data. The congestion detection requirements determine the timescale of interest. For example, if congestion develops slowly, a larger time window may be needed to capture its characteristics; if congestion occurs suddenly, a smaller time window may be required for timely detection. Within each sliding time window, the cross-correlation coefficient is calculated between the target intrinsic mode function component (any one selected from multiple IMF components) and a preset reference signal. The preset reference signal is typically constructed based on the signal characteristics of a flow measurement device under normal operating conditions and represents the typical signal pattern of the device during normal operation. The cross-correlation coefficient measures the degree of similarity between two signals. Its value ranges from -1 to 1. Values ​​closer to 1 indicate greater similarity; values ​​closer to -1 indicate less similarity; and a value of 0 indicates no correlation between the two signals. By calculating the cross-correlation coefficient between the target IMF component and the preset reference signal, it can be determined whether the target IMF component deviates from the normal state within the time window.

[0076] A coefficient threshold is preset to determine whether the degree of similarity between the target IMF component and a preset reference signal is within a normal range. The number of sliding time windows in which the cross-correlation coefficient falls below the preset coefficient threshold is recorded. This number reflects the number of time segments in which the target IMF component deviates significantly from its normal state. A threshold is also preset. When the number of sliding time windows in which the cross-correlation coefficient falls below the preset coefficient threshold is greater than or equal to the threshold, the target IMF component deviates significantly from its normal state in a significant number of time segments and is identified as an abnormal modal component caused by congestion. When this number is less than the threshold, the target IMF component deviates only slightly from its normal state and is identified as a normal modal component. This method allows the abnormal modal component associated with congestion to be distinguished from multiple IMF components, providing a basis for subsequent analysis and detection. All IMF components identified as normal modal components are combined to construct an intrinsic mode function (IMF). These normal modal components reflect the signal characteristics of the flow measurement device under normal operation. By combining them, an IMF representing the normal operation of the device can be obtained. During the construction process, the normal mode components can be weighted and filtered as needed to further improve the quality and accuracy of the intrinsic mode function. The constructed intrinsic mode function will be used in subsequent steps such as feature extraction and blockage probability prediction, providing effective feature information for blockage detection in flow measurement devices.

[0077] The randomness of white noise allows for clearer separation of signal features at different scales, improving the stability and accuracy of signal decomposition. This allows for more effective extraction of meaningful intrinsic mode function components from complex signals, providing a more reliable feature foundation for subsequent analysis. The cross-correlation coefficient quantifies the degree of similarity between the target intrinsic mode function component and the reference signal. When a blockage occurs, the signal characteristics generated by the flow measurement device change, resulting in a decrease in the similarity between the relevant intrinsic mode function component and the reference signal under normal conditions. By recording the number of sliding time windows where the cross-correlation coefficient falls below a preset coefficient threshold and comparing it with a preset threshold, it is possible to accurately determine whether the target intrinsic mode function component is an abnormal mode component caused by blockage or a normal mode component. This method can accurately identify abnormal features related to blockage from among numerous intrinsic mode function components, facilitating a deeper analysis of the impact of blockage on the signal. This abnormal mode component identification process distinguishes signal changes caused by blockage from those during normal device operation. Normal mode components reflect the normal operation of the device in the absence of blockage, while abnormal mode components contain characteristic information related to blockage. By constructing the intrinsic mode function based on the normal mode component, interference information such as blockage noise can be removed, so that the constructed intrinsic mode function can more accurately reflect the signal characteristics of the device in a normal state. In subsequent blockage detection, by comparing the actual signal with the model or characteristics constructed based on the normal mode component, it is possible to more accurately determine whether the device is blocked, thereby improving the accuracy and reliability of blockage detection. The sliding time window size and step size, as well as the preset coefficient threshold and preset quantity threshold are determined according to the signal sampling frequency and blockage detection requirements, making this method highly flexible and adaptable. Different flow measurement devices and blockage detection scenarios may have different signal characteristics and detection requirements. By adjusting these parameters, optimization can be performed for specific situations to ensure that signal changes caused by blockage can be accurately captured to meet the blockage detection requirements in different application scenarios.

[0078] Optionally, decomposing each noisy signal sequence separately to obtain multiple intrinsic mode function components includes:

[0079] Identify all the maximum and minimum points from the noisy signal sequence, use interpolation methods to construct the upper envelope based on all the maximum points, and construct the lower envelope based on all the minimum points;

[0080] Calculating a mean curve of the upper envelope and the lower envelope, and subtracting the mean curve from the noisy signal sequence to obtain a new signal sequence;

[0081] It is determined whether the new signal sequence meets a preset condition of an eigenmode function; if the preset condition is met, the new signal sequence is determined as an eigenmode function component.

[0082] For a given noisy signal sequence, it is necessary to find all the maximum and minimum points in the entire signal sequence. A maximum point refers to a point where the signal value is larger than the values ​​of the adjacent points on its left and right sides; a minimum point refers to a point where the signal value is smaller than the values ​​of the adjacent points on its left and right sides. This process can be achieved by comparing the size relationship of adjacent data points in the signal sequence. For example, for a signal sequence x ( n ),like x ( n )> x ( n -1) and x ( n )> x ( n +1), then x ( n ) is the maximum point. The interpolation method is used to construct the upper envelope based on all the maximum points. There are many interpolation methods, the most common one is cubic spline interpolation. Cubic spline interpolation constructs cubic polynomials between adjacent maximum points so that these polynomials not only have equal function values ​​at the maximum points, but also have continuous first-order and second-order derivatives, thereby obtaining a smooth curve. This curve is the upper envelope, which can fit the distribution trend of the maximum points very well and is always located above the maximum points. Similarly, the lower envelope is constructed based on all the minimum points using the interpolation method. The upper envelope and the lower envelope jointly define the fluctuation range of the signal sequence in the vertical direction, providing a basis for the subsequent extraction of the main components of the signal. After obtaining the upper envelope and the lower envelope, the mean curve of the two envelopes is calculated. The mean curve can be obtained by averaging the values ​​of the upper envelope and the lower envelope at each corresponding point, that is, for each time point t , mean curve m ( t ) = 2 u ( t )+ l ( t ),in u ( t ) is the upper envelope at time point t The value of l ( t ) is the lower envelope at time point t The mean curve reflects the average change trend of the signal sequence between the upper and lower envelopes, which contains the low-frequency components and overall trend information in the signal sequence. x ( t ) minus the mean curve m ( t ), and obtain the new signal sequence h( t ) = x ( t )− m ( t The purpose of this step is to remove the low-frequency trend components in the signal sequence so that the new signal sequence h ( t ) focuses more on the high-frequency oscillation part of the signal, which may contain important characteristic information of the signal, such as signal changes caused by blockage.

[0083] The intrinsic mode function (IMF) needs to meet the following two conditions: Condition 1: In the entire data range, the number of extreme points and the number of zero crossings must be equal or differ by at most one. The extreme points reflect the local maximum and minimum values ​​of the signal, and the zero crossings are the points where the signal intersects the horizontal axis. This condition ensures that the signal is symmetrical on the time scale, that is, the positive and negative half cycles of the signal are relatively balanced in time, avoiding overly complex or irregular oscillation patterns in the signal. Condition 2: At any moment, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero. This means that the symmetry of the upper and lower envelopes of the signal is guaranteed in any local interval, and the local average trend of the signal is zero, which further emphasizes the high-frequency oscillation characteristics of the signal and removes the low-frequency trend components. For the new signal sequence obtained h ( t ) to check whether it meets the above two preset conditions. If it does, it means that the new signal sequence h ( t ) is an intrinsic mode function component, which can be determined as IMF. If it is not satisfied, the new signal sequence needs to be h ( t ) as a new noisy signal sequence, repeat the above process of identifying extreme points, constructing envelopes, calculating mean curves and obtaining new signal sequences until an IMF that meets the conditions is obtained.

[0084] Identifying all maximum and minimum points in a noisy signal sequence and constructing the upper and lower envelopes, respectively, using interpolation methods can accurately delineate the local fluctuation range of the signal. Interpolation methods (such as cubic spline interpolation) can smoothly and accurately fit the upper and lower boundaries of the signal based on limited information about the extreme points. This allows for a more comprehensive capture of the amplitude variation characteristics of the signal at different moments, laying the foundation for subsequent analysis of the signal's intrinsic components. By calculating the mean curves of the upper and lower envelopes and subtracting them from the noisy signal sequence to generate a new signal sequence, this step acts as a filtering process. The noise component in the original noisy signal typically manifests as random fluctuations, while the mean curve reflects the overall trend and key characteristics of the signal. Subtracting the mean curve reduces the relative influence of noise in the new signal sequence, making the key features more prominent. This helps improve the stability and accuracy of signal decomposition and reduces noise interference in the subsequent extraction of intrinsic mode function components. Determine whether the new signal sequence meets the preset conditions of the intrinsic mode function. Only new signal sequences that meet these conditions will be determined as intrinsic mode function components. This strict judgment standard ensures that the extracted intrinsic mode function components have clear physical meaning and mathematical properties, can accurately reflect the natural vibration mode of the signal at different frequency scales, avoids mistaking signal components that do not meet the conditions for intrinsic mode function components, and improves the rationality of component extraction. According to the above method, the signal is continuously decomposed, and each time an eigenmode function component that meets the conditions is extracted and separated from the original signal, and then the above process is repeated for the remaining signal. This step-by-step decomposition method can decompose complex noisy signals into a series of intrinsic mode function components with different frequency components. Each component represents the characteristics of the signal at different time scales, which helps to deeply analyze the composition structure of the signal and provide more detailed and effective feature information for subsequent tasks such as signal processing, feature extraction, and pattern recognition.

[0085] Optionally, extracting the time vector, space vector, and frequency domain vector of the intrinsic mode function includes:

[0086] Sampling the intrinsic mode function on the time axis, recording the function value corresponding to each sampling point, and arranging these function values ​​in chronological order to form a time vector;

[0087] quantizing the energy or amplitude of the intrinsic mode function at different frequency components, using the quantized values ​​as elements of a vector with frequency as a dimension, and constructing a space vector;

[0088] The intrinsic mode function is converted from the time domain to the frequency domain using Fourier transform to obtain a frequency domain representation of the signal in the frequency domain, the target features of the frequency domain representation are extracted, and the target features are arranged in a preset order to form a frequency domain vector.

[0089] A given intrinsic mode function (IMF) is a time-varying signal. Sampling on the time axis means selecting several time points at regular intervals (sampling intervals). The choice of this sampling interval depends on the signal characteristics and analysis requirements. If the signal changes rapidly, a smaller sampling interval is required to accurately capture signal details. If the signal changes relatively slowly, the sampling interval can be increased to reduce the amount of data. At each sampling point, the corresponding IMF value is recorded. These function values ​​reflect the amplitude of the signal at that moment. Then, these function values ​​are arranged in chronological order to form a time vector. The time vector intuitively displays the temporal trend of the IMF and contains dynamic information about the signal in the time domain, such as its rise, fall, and fluctuation characteristics. This provides basic data for subsequent analysis of the signal's temporal characteristics. Although the IMF is a time-domain signal, it can be considered to be composed of components of different frequencies. To construct the space vector, the energy or amplitude of the IMF at different frequency components must be analyzed. Energy generally reflects the power of a signal within a certain frequency range, while amplitude reflects the strength of the signal at that frequency component. Signal processing methods (such as short-time Fourier transform and wavelet transform) can be used to estimate the energy or amplitude of the IMF at different frequency components and quantize them to convert them into specific numerical values. With frequency as the dimension, the quantized energy or amplitude values ​​at different frequency components are used as elements of a vector. In other words, each element of the vector corresponds to a specific frequency, and the value of that element represents the energy or amplitude of the IMF at that frequency. This constructed spatial vector can describe the characteristics of the IMF from a frequency perspective, reflecting the distribution of the signal across different frequency spaces. This helps analyze the frequency characteristics of the signal and the relationships between different frequency components. By performing a Fourier transform on the IMF, the signal, which originally varies on the time axis, is converted to a signal represented on the frequency axis, resulting in a frequency domain representation of the signal. This frequency domain representation clearly displays the various frequency components in the signal, along with their corresponding amplitude and phase information, allowing for in-depth analysis of the signal's characteristics from a frequency perspective. Target features can be extracted from the frequency domain representation. These features can be selected based on specific application requirements and analysis objectives. For example, features such as the signal's dominant frequency (i.e., the frequency component with the highest signal energy or amplitude), frequency bandwidth (the frequency range in which the signal energy or amplitude is non-zero), and spectral energy (the sum of the signal's energy at each frequency) can be extracted. These target features are then arranged in a predetermined order to form a frequency domain vector. The frequency domain vector concentrates the key characteristics of the intrinsic mode function in the frequency domain, providing important feature basis for subsequent tasks such as signal analysis and pattern recognition based on frequency domain information.

[0090] Sampling the intrinsic mode function (IMF) on the time axis, recording the function value corresponding to each sampling point, and arranging it in time sequence to form a time vector, this process accurately captures the dynamic characteristics of the signal over time. Each element in the time vector represents the signal state at a specific moment. By analyzing these elements, we can understand the signal's time-domain characteristics, such as amplitude trends, periodicity, and mutation points. This provides a detailed and accurate data foundation for subsequent time-based analysis (such as determining the onset of congestion and studying the time-domain fluctuation patterns of signals). The resulting time vector is presented as an intuitive numerical sequence, facilitating various time-related analysis operations. The energy or amplitude of the IMF at different frequency components is quantified, and a space vector is constructed with frequency as the dimension. This provides a comprehensive and detailed representation of the signal's energy distribution in the frequency domain. Each element in the space vector corresponds to a specific frequency component, and its value represents the energy or amplitude at that frequency component. By analyzing spatial vectors, we can clearly understand the relative strengths of various frequency components in a signal, identify the primary and secondary frequency components, and help detect possible abnormal frequency characteristics in the signal, providing critical frequency-dimensional information for tasks such as congestion detection. The constructed spatial vectors represent the signal from the perspective of frequency energy, complementing feature vectors in other dimensions, such as the time vector. In subsequent feature fusion, spatial vectors can be organically combined with other feature vectors to provide more comprehensive and richer signal feature information. This helps improve the accuracy and robustness of congestion probability prediction models based on methods such as deep learning, enabling the models to better capture the relationship between signal characteristics in different dimensions and congestion. The Fourier transform is used to convert the intrinsic mode function from the time domain to the frequency domain, obtaining a frequency-domain representation of the signal. Target features (such as dominant frequency, frequency bandwidth, and spectral energy distribution) are then extracted to form frequency-domain vectors. The Fourier transform clearly reveals the frequency information hidden in the time-domain signal, and the extracted target features accurately describe the key characteristics of the signal in the frequency domain. The resulting frequency domain vector concisely captures the signal's key frequency domain characteristics, facilitating rapid analysis and processing. In applications such as congestion detection, directly analyzing the target features in the frequency domain vector allows for rapid determination of abnormal frequency variations in the signal, enabling timely detection of congestion and other fault conditions. This simplicity also helps reduce computational effort during subsequent model training and inference, improving system efficiency.

[0091] Optionally, obtaining the attention weight through the two-way cross attention mechanism includes:

[0092] In a first path, the time vector is used as a query vector, and the space vector is used as a key vector and a value vector; in a second path, the time vector is used as a query vector, and the frequency domain vector is used as a key vector and a value vector;

[0093] Performing cross attention calculation on the first path and the second path to obtain a first attention weight and a second attention weight;

[0094] Different weights are assigned according to the importance of the first path and the second path, and the first attention weight and the second attention weight are linearly combined according to the weights to obtain attention weights, which include time vector weights, space vector weights and frequency domain vector weights.

[0095] In the first path, the time vector serves as the query vector, and the spatial vector serves as the key and value vectors. In the attention mechanism, the query vector is used to query information related to itself, the key vector is used to query the information being queried, and the value vector contains the actual information content. This approach focuses on the relationship between the time vector and the spatial vector, calculating the attention distribution of the time vector on the spatial vector. In the second path, the time vector is also used as the query vector, but the frequency vector serves as the key and value vectors. This allows the second path to focus on the relationship between the time vector and the frequency vector, calculating the attention distribution of the time vector on the frequency vector. This approach uses the time vector as a bridge to explore the correlations between time and space, and between time and frequency, helping to explore the inherent connections between multi-dimensional features from different perspectives. In the first path, the similarity between the time vector (query vector) and the spatial vector (key vector) is calculated according to the attention mechanism formula (typically dot product attention, additive attention, etc.), resulting in an attention score. These attention scores are then normalized (e.g., using the Softmax function) to obtain the first attention weight. The first attention weight reflects the distribution of the time vector's attention on different elements of the spatial vector. Similarly, in the second path, the similarity between the time vector (query vector) and the frequency vector (key vector) is calculated and normalized to obtain the second attention weight. The second attention weight represents the distribution of attention of the time vector across different elements of the frequency vector, revealing the focus of the association between the time and frequency vectors. Because the first and second paths focus on different feature association dimensions, their importance in the overall feature fusion process may differ. Based on the actual application scenario and task requirements, the importance of the first and second paths can be determined through experimentation, experience, or specific algorithms, and different weights can be assigned to them. For example, if the relationship between time and space is considered more critical in the current task, a larger weight can be assigned to the first path; conversely, if the relationship between time and frequency is more important, a larger weight can be assigned to the second path. Based on the assigned weights, the first and second attention weights are linearly combined. The formula for this linear combination is generally: Final Attention Weight = Weight 1 × First Attention Weight + Weight 2 × Second Attention Weight, where Weight 1 and Weight 2 are the weights of the first and second paths, respectively. The final attention weights include time vector weights, space vector weights, and frequency domain vector weights. These weights reflect the importance of different vectors in the overall features. They can automatically learn and highlight feature information that is more critical to the current task (such as congestion detection) and suppress irrelevant or interfering information.

[0096] Through cross-attention calculation, the model can deeply explore the inherent connections between temporal and spatial features, as well as temporal and frequency domain features. The resulting attention weights include time vector weights, spatial vector weights, and frequency vector weights, reflecting the importance of different dimensional features within the overall feature set. By linearly combining the attention weights from different paths, the model effectively integrates information from the three dimensions of time, space, and frequency. This fusion approach avoids the limitations of single-dimensional feature analysis, enabling the model to comprehensively consider information from multiple dimensions, more accurately characterizing the operating status of the flow measurement device and providing richer feature information for subsequent tasks such as congestion detection. Different weights are assigned based on the importance of the first and second paths, and the final attention weights are linearly combined to obtain the final attention weights. This process enables the model to dynamically adjust its focus on temporal, spatial, and frequency domain features based on the information value provided by different paths. The dual-path cross-attention mechanism adaptively learns the importance of features from different dimensions without the need for manual pre-setting of fixed feature weights. This adaptive feature selection approach enables the model to better adapt to diverse data distributions and task requirements, accurately extracting the most valuable features for congestion detection in various flow measurement device operating scenarios, and improving the model's generalization and adaptability. In congestion detection tasks, temporal, spatial, and frequency domain features may all contain information related to congestion. By using the attention weights derived from the dual-path cross-attention mechanism, the model can more accurately locate features closely related to congestion. For example, the time vector may reflect the moment of signal mutation when congestion occurs, the spatial vector may indicate the sensor signal changes corresponding to the specific location of the congestion in the device, and the frequency domain vector may reveal the changes in frequency components caused by congestion. By combining these appropriately weighted features, the model can more accurately determine whether a device is congested, improving the accuracy of congestion detection.

[0097] Optionally, obtaining the attention weight through the two-way cross attention mechanism includes:

[0098] In the first path, the similarity matrix between the query vector and the key vector is calculated, normalized, and multiplied by the value vector to obtain the first attention output:

[0099] In the second path, a multi-head attention mechanism is used to split the query vector into multiple head subspaces, which are calculated in parallel with the key vector and value vector respectively. The second attention output is obtained by splicing the multi-head outputs.

[0100] The first attention output and the second attention output are input into the gated fusion module, and the contribution of the two-way information is dynamically adjusted through the weight parameter to obtain the attention weight.

[0101] In the first path, the query vector is usually a time vector, and the key vector is a space vector. The calculation of the similarity matrix is ​​used to measure the similarity between the query vector and each element in the key vector. Common similarity calculation methods include dot product and cosine similarity. Taking dot product as an example, if the query vector Q =[ q 1, q 2,......, qn ], key vector K =[ k 1, k 2,......, kn ], then the similarity matrix S Elements in Sij for qi · kj , i ∈[1, n ], j ∈[1, n By calculating the similarity matrix, the model can understand the degree of correlation between temporal features and spatial features, providing a basis for subsequent attention allocation. The calculated similarity matrix S The element values ​​in may have different dimensions and numerical ranges. In order to facilitate subsequent calculations and comparisons, they need to be normalized. The commonly used normalization method is the softmax function. The Softmax function converts each element in the similarity matrix into Sij Convert to probability value aij The normalized matrix A The sum of each row of elements in is 1, which indicates the probability of each element in the time feature to allocate attention to each element in the spatial feature. The value vector is also played by the spatial vector. A With value vector V Multiply to get the first attention output O 1. The specific calculation is O 1= A · V The purpose of this step is to perform weighted summation of spatial features according to the attention allocation probability, so that the model can pay more attention to the spatial feature parts that are highly correlated with the temporal features, thereby extracting more valuable feature information.

[0102] In the second path, the query vector is still a time vector, but a multi-head attention mechanism is used. The core idea of ​​the multi-head attention mechanism is to split the query vector, key vector, and value vector into multiple head subspaces. Assume that the query vector is split into h heads, and the dimension of each head subspace is dk , then the dimension of the original query vector is d model =h × dk . Through segmentation, the model can capture the relationship between features from different subspace perspectives, increase the model's expressive power and ability to recognize different feature patterns. For each head subspace, attention calculations are performed with the key vector and value vector respectively. The calculation process is similar to the first path, that is, the similarity matrix between the query vector head and the key vector head is first calculated, then normalized, and finally multiplied with the value vector head to obtain the attention output of the head. Since multiple heads are calculated in parallel, feature information can be extracted from different angles at the same time, greatly improving the calculation efficiency. The attention outputs of each head are spliced ​​to obtain the second attention output O 2. The concatenated vector dimension is restored to the dimension of the original query vector d model The multi-head attention mechanism enables the model to comprehensively consider the feature information of different subspaces through parallel calculation and splicing, thereby more comprehensively capturing the complex relationship between temporal features and frequency domain features.

[0103] Output the first attention O 1st and second attention output O 2 is input into the gated fusion module. These two outputs represent feature information extracted from different perspectives (time-space relationship and time-frequency relationship), and their contribution to the final attention weight may vary depending on task requirements and data characteristics. The gated fusion module dynamically adjusts the contribution of dual-path information by learning weight parameters. Specifically, the model learns a set of weight parameters w 1 and w 2, used to weight the first attention output and the second attention output respectively. Then, the weighted results are fused to obtain the final attention weight O ,Right now O = w 1. O 1+ w 2. O 2. Weight parameters w 1 and w 2 are automatically learned through the back-propagation algorithm during model training. They can adaptively adjust the relative importance of dual-path information according to the actual situation of the data and the task objectives, so that the model can better utilize the information of the two paths and improve the accuracy and effectiveness of the attention weight.

[0104] The dual-path cross-attention mechanism exploits the correlations between time and space, and time and frequency, through different computational paths. This is then integrated through a gated fusion module. This allows for a more comprehensive and in-depth capture of complex relationships between features, providing the model with richer feature information and improving performance on related tasks such as congestion detection. The use of a multi-head attention mechanism and gated fusion module enhances the model's flexibility and adaptability. The multi-head attention mechanism extracts features from multiple subspace perspectives, while the gated fusion module dynamically adjusts the contribution of dual-path information based on specific circumstances. This allows the model to better adapt to diverse data distributions and task requirements, reducing its reliance on specific data and tasks. By comprehensively considering information from both paths and dynamically weighting the fusion, the resulting attention weights more accurately reflect the importance of different feature dimensions within the overall feature set. This allows the model to more precisely focus on key features and suppress irrelevant or interfering information, thereby improving the model's decision accuracy and reliability.

[0105] Optionally, constructing the state space according to the congestion probability value, historical maintenance records, and equipment operation time includes:

[0106] The device's runtime is segmented, each segmented interval is assigned a unique code value, and a runtime state code vector is generated based on the interval to which the device's actual runtime belongs.

[0107] generating a maintenance cycle matrix according to the historical maintenance records, and performing singular value decomposition on the maintenance cycle matrix to obtain a maintenance feature vector;

[0108] Performing tensor splicing on the congestion probability value, the maintenance feature vector, and the running time state encoding vector, and generating a time series correlation feature according to the spliced ​​tensor;

[0109] A state space is generated according to the temporal correlation features.

[0110] Equipment operating time is a continuous variable. Directly using continuous operating time data can lead to complex model processing and less intuitive feature representation. By segmenting the equipment operating time, the continuous operating time can be converted into discrete intervals, facilitating subsequent feature extraction and model processing. Equipment operating time can be divided into several intervals based on the actual operating conditions of the equipment and business needs. For example, segmentation can be based on the equipment's lifespan phase or maintenance cycle. Assuming that the normal operating time of equipment ranges from 0 to 10,000 hours, it can be divided into four intervals: 0-1,000 hours, 1,001-3,000 hours, 3,001-6,000 hours, and 6,001-10,000 hours. Each interval is assigned a unique code value, typically using a simple numeric code. For example, the four intervals mentioned above can be coded as 0, 1, 2, and 3, respectively. The actual operating time of the equipment is used to determine the interval to which it belongs. For example, if the actual operating time of the equipment is 2,500 hours, it belongs to the interval 1,001-3,000 hours, with a corresponding code value of 1. The encoded value is converted to a vector, namely the runtime state encoding vector. In practical applications, one-hot encoding is often used to facilitate subsequent tensor operations. For the encoding of the four intervals mentioned above, if the device runtime code is 1, its one-hot encoding vector is [0, 1, 0, 0]. Historical maintenance records contain important information such as the device's maintenance time, maintenance type, and maintenance content. By analyzing these records, we can calculate the maintenance status of the device in different time periods and construct a maintenance cycle matrix. A two-dimensional matrix is ​​constructed with time as the horizontal axis and maintenance type as the vertical axis. Each element in the matrix represents the number or frequency of a specific type of maintenance performed at a specific point in time. For example, assuming there are three types of maintenance, the maintenance status for each month over the past year is counted, resulting in a maintenance cycle matrix with 12 rows (12 months) and 3 columns (three maintenance types). Singular value decomposition is a commonly used matrix decomposition method that decomposes a matrix into the product of three matrices, namely, A = UΣV T, where A is the original matrix, U and V are orthogonal matrices, and Σ is a diagonal matrix with singular values ​​on the diagonal. Singular values ​​reflect the importance of information in different directions within the matrix. After performing singular value decomposition on the maintenance cycle matrix, the first k largest singular values ​​and their corresponding left and right singular vectors are taken and linearly combined to obtain maintenance eigenvectors. These eigenvectors can extract key characteristic information from the maintenance cycle matrix, such as the maintenance patterns of the equipment and the changing trends in maintenance frequency. The congestion probability value, maintenance eigenvector, and runtime state encoding vector are concatenated into tensors. The congestion probability value is typically a scalar. It can be expanded to a vector with the same dimensions as the maintenance eigenvector and runtime state encoding vector (for example, by padding them with zeros where appropriate) and then concatenated with these two vectors. The concatenated tensor contains information about the equipment's congestion probability, maintenance characteristics, and runtime. Since equipment operation is a time-series process, there is correlation between state information at different time points. When generating time series correlation features, consider using time series models (such as recurrent neural networks (RNNs) or long short-term memory (LSTMs)) to process the concatenated tensors to capture the temporal dependencies between state information at different time points. Alternatively, simple time series feature extraction methods, such as calculating the difference between tensors at adjacent time points or using sliding window statistics, can be used to obtain feature vectors that reflect the temporal changes in device states, known as time series correlation features. The state space is a multidimensional space where each point represents a device state. Each element in a time series correlation feature vector corresponds to a dimension in the state space, and the value in the vector represents the device state value in that dimension. By treating the time series correlation feature vector as a point in the state space and continuously collecting time series correlation feature vectors for the device at different time points, the entire state space can be constructed. The state space can intuitively display the range and distribution of device state variations, providing a basis for subsequent device state analysis, fault prediction, and decision-making. For example, by observing the distribution of points in the state space, it is possible to determine whether the device is in a normal state, a warning state, or a faulty state.

[0111] By integrating multiple sources of information, such as congestion probability values, historical maintenance records, and equipment operating hours, the constructed state space can more comprehensively and accurately represent the equipment's operating status. These different types of information complement each other, avoiding the limitations of a single source and improving the accuracy and reliability of equipment status assessment. The process of generating temporal correlation features takes into account the temporal changes in equipment status, enabling the state space to reflect the changing trends of equipment status over time. This is crucial for predicting the future state of equipment and detecting potential faults in advance, helping to implement predictive maintenance, reduce equipment failure rates, and improve equipment operating efficiency and reliability. The constructed state space provides an important foundation for subsequent tasks such as equipment status analysis, fault diagnosis, and maintenance decision-making. Based on the state space, various machine learning and data analysis methods can be used to explore the potential relationships between equipment status and faults, develop more scientific and reasonable maintenance strategies, and optimize equipment operation management.

[0112] This embodiment also discloses a blockage detection system for a flow measurement device. Figure 2 This is a module diagram of a blockage detection system for a flow measurement device disclosed in an embodiment of the present application. Figure 2 As shown, the system includes an acquisition module 201, a vector module 202, a prediction module 203 and an execution module 204, wherein:

[0113] An acquisition module 201 is configured to acquire an original pressure signal and an original flow signal of a flow measurement device to be measured, and generate an eigenmode function according to the original pressure signal and the original flow signal;

[0114] a vector module 202 configured to extract a time vector, a space vector, and a frequency domain vector of the intrinsic mode function, obtain an attention weight through a two-way cross attention mechanism, and perform weighted fusion on the time vector, the space vector, the frequency domain vector, and the attention weight to obtain a multi-dimensional feature vector;

[0115] The prediction module 203 is configured to input the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model to obtain a congestion probability value;

[0116] The execution module 204 is configured to construct a state space based on the congestion probability value, historical maintenance records and equipment operation time, balance false alarm and missed detection penalties in combination with a reward function, and determine an execution operation.

[0117] Optionally, the acquisition module 201 is configured to:

[0118] adding white noise of different amplitudes to the original pressure signal and the original flow signal multiple times to form multiple noisy signal sequences;

[0119] Decompose each noisy signal sequence separately to obtain multiple intrinsic mode function components;

[0120] Determine the sliding time window size and step size based on the signal sampling frequency and congestion detection requirements, and calculate the cross-correlation coefficient between the target intrinsic mode function component and the preset reference signal within each sliding time window, wherein the target intrinsic mode function component is any one of the multiple intrinsic mode function components;

[0121] Recording the number of sliding time windows in which the cross-correlation coefficient is lower than a preset coefficient threshold, when the number is greater than or equal to the preset number threshold, determining the target intrinsic mode function component as an abnormal mode component caused by blockage; when the number is less than the preset number threshold, determining the target intrinsic mode function component as a normal mode component;

[0122] The intrinsic mode functions are constructed based on normal mode components.

[0123] Optionally, the acquisition module 201 is configured to:

[0124] Identify all the maximum and minimum points from the noisy signal sequence, use interpolation methods to construct the upper envelope based on all the maximum points, and construct the lower envelope based on all the minimum points;

[0125] Calculating a mean curve of the upper envelope and the lower envelope, and subtracting the mean curve from the noisy signal sequence to obtain a new signal sequence;

[0126] It is determined whether the new signal sequence meets a preset condition of an eigenmode function; if the preset condition is met, the new signal sequence is determined as an eigenmode function component.

[0127] Optionally, the vector module 202 is configured to:

[0128] Sampling the intrinsic mode function on the time axis, recording the function value corresponding to each sampling point, and arranging these function values ​​in chronological order to form a time vector;

[0129] quantizing the energy or amplitude of the intrinsic mode function at different frequency components, using the quantized values ​​as elements of a vector with frequency as a dimension, and constructing a space vector;

[0130] The intrinsic mode function is converted from the time domain to the frequency domain using Fourier transform to obtain a frequency domain representation of the signal in the frequency domain, the target features of the frequency domain representation are extracted, and the target features are arranged in a preset order to form a frequency domain vector.

[0131] Optionally, the vector module 202 is configured to:

[0132] In a first path, the time vector is used as a query vector, and the space vector is used as a key vector and a value vector; in a second path, the time vector is used as a query vector, and the frequency domain vector is used as a key vector and a value vector;

[0133] Performing cross attention calculation on the first path and the second path to obtain a first attention weight and a second attention weight;

[0134] Different weights are assigned according to the importance of the first path and the second path, and the first attention weight and the second attention weight are linearly combined according to the weights to obtain attention weights, which include time vector weights, space vector weights and frequency domain vector weights.

[0135] Optionally, the vector module 202 is configured to:

[0136] In the first path, the similarity matrix between the query vector and the key vector is calculated, normalized, and multiplied by the value vector to obtain the first attention output:

[0137] In the second path, a multi-head attention mechanism is used to split the query vector into multiple head subspaces, which are calculated in parallel with the key vector and value vector respectively. The second attention output is obtained by splicing the multi-head outputs.

[0138] The first attention output and the second attention output are input into the gated fusion module, and the contribution of the two-way information is dynamically adjusted through the weight parameter to obtain the attention weight.

[0139] Optionally, the execution module 204 is configured to:

[0140] The device's runtime is segmented, each segmented interval is assigned a unique code value, and a runtime state code vector is generated based on the interval to which the device's actual runtime belongs.

[0141] generating a maintenance cycle matrix according to the historical maintenance records, and performing singular value decomposition on the maintenance cycle matrix to obtain a maintenance feature vector;

[0142] Performing tensor splicing on the congestion probability value, the maintenance feature vector, and the running time state encoding vector, and generating a time series correlation feature according to the spliced ​​tensor;

[0143] A state space is generated according to the temporal correlation features.

[0144] It should be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0145] This embodiment also discloses an electronic device, referring to Figure 3 The electronic device may include: at least one processor 301 , at least one communication bus 302 , a user interface 303 , a network interface 304 , and at least one memory 305 .

[0146] The communication bus 302 is used to implement the connection and communication between these components.

[0147] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0148] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0149] The processor 301 may include one or more processing cores. Using various interfaces and circuits, the processor 301 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 305, as well as accesses data stored in the memory 305, to perform various server functions and process data. Optionally, the processor 301 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 301 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may not be integrated into the processor 301 but implemented as a separate chip.

[0150] Among them, the memory 305 may include a random access memory (RAM) or a read-only memory (Read-Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may also be optionally at least one storage device located away from the aforementioned processor 301. As Figure 3 As shown, the memory 305 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program of a method for detecting blockage of a flow measurement device.

[0151] exist Figure 3 In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and obtain data input by the user; and the processor 301 can be used to call an application program stored in the memory 305 for a blockage detection method for a flow measuring device. When executed by one or more processors 301, the electronic device executes one or more methods in the above-mentioned embodiments.

[0152] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0153] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0154] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0155] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0156] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0157] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory 305 includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a magnetic disk, or an optical disk.

[0158] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the technical field that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for detecting blockage of a flow measuring device, characterized in that: Applied to a blockage detection platform, the method comprises: Obtaining an original pressure signal and an original flow signal of a flow measurement device to be measured, and generating an eigenmode function according to the original pressure signal and the original flow signal; Extract the time vector, space vector and frequency domain vector of the intrinsic mode function, obtain the attention weight through the two-way cross attention mechanism, and perform weighted fusion of the time vector, space vector, frequency domain vector and attention weight to obtain a multi-dimensional feature vector; Input the multi-dimensional feature vector into the pre-trained deep learning-based congestion probability prediction model to obtain the congestion probability value; The state space is constructed based on the congestion probability value, historical maintenance records and equipment operation time. The reward function is combined to balance the false alarm and missed detection penalties to determine the execution action. Generating the eigenmode function according to the original pressure signal and the original flow signal includes: Adding white noise of different amplitudes to the original pressure signal and the original flow signal multiple times to form multiple noisy signal sequences; Decompose each noisy signal sequence separately to obtain multiple intrinsic mode function components; Determine the sliding time window size and step size based on the signal sampling frequency and congestion detection requirements, and calculate the cross-correlation coefficient between the target intrinsic mode function component and the preset reference signal within each sliding time window, where the target intrinsic mode function component is any one of the multiple intrinsic mode function components; Recording the number of sliding time windows whose mutual correlation coefficient is lower than a preset coefficient threshold, when the number is greater than or equal to the preset number threshold, determining the target intrinsic mode function component as an abnormal mode component caused by congestion, and when the number is less than the preset number threshold, determining the target intrinsic mode function component as a normal mode component; Construct the eigenmode functions based on the normal mode components, Extracting the time vector, space vector, and frequency domain vector of the intrinsic mode function includes: Sample the intrinsic mode function on the time axis, record the function value corresponding to each sampling point, and arrange these function values ​​in chronological order to form a time vector; Quantify the energy or amplitude of the intrinsic mode function at different frequency components, use the quantized values ​​as elements of the vector with frequency as the dimension, and construct a space vector; The Fourier transform is used to convert the intrinsic mode function from the time domain to the frequency domain to obtain the frequency domain representation of the signal in the frequency domain. The target features of the frequency domain representation are extracted and arranged in a preset order to form a frequency domain vector.

2. The method for detecting blockage of a flow measurement device according to claim 1, wherein: Each noisy signal sequence is decomposed separately to obtain multiple intrinsic mode function components including: Identify all the maximum and minimum points from the noisy signal sequence, use interpolation methods to construct the upper envelope based on all the maximum points, and construct the lower envelope based on all the minimum points; Calculate the mean curve of the upper envelope and the lower envelope, subtract the mean curve from the noisy signal sequence to obtain a new signal sequence; It is determined whether the new signal sequence meets the preset conditions of the intrinsic mode function. If the preset conditions are met, the new signal sequence is determined as the intrinsic mode function component.

3. The method for detecting blockage of a flow measurement device according to claim 1, wherein: The attention weights obtained through the two-way cross attention mechanism include: In the first path, the time vector is used as the query vector, and the spatial vector is used as the key vector and the value vector. In the second path, the time vector is used as the query vector, and the frequency domain vector is used as the key vector and the value vector. Perform cross attention calculation on the first path and the second path to obtain the first attention weight and the second attention weight; Different weights are assigned according to the importance of the first path and the second path, and the first attention weight and the second attention weight are linearly combined according to the weights to obtain the attention weights. The attention weights include time vector weights, space vector weights and frequency domain vector weights.

4. The method for detecting blockage of a flow measurement device according to claim 3, wherein: The attention weights obtained through the two-way cross attention mechanism include: In the first path, the similarity matrix between the query vector and the key vector is calculated, normalized, and multiplied by the value vector to obtain the first attention output: In the second path, a multi-head attention mechanism is used to split the query vector into multiple head subspaces, which are then computed in parallel with the key vector and value vector respectively. The second attention output is obtained by concatenating the multi-head outputs. The first attention output and the second attention output are input into the gated fusion module, and the contribution of the two-way information is dynamically adjusted through the weight parameter to obtain the attention weight.

5. The method for detecting blockage of a flow measurement device according to claim 1, wherein: Constructing the state space based on the jam probability value, historical maintenance records, and equipment operation time includes: The device's runtime is segmented, each segmented interval is assigned a unique code value, and a runtime state code vector is generated based on the interval to which the device's actual runtime belongs. Generate a maintenance cycle matrix based on historical maintenance records, and perform singular value decomposition on the maintenance cycle matrix to obtain maintenance eigenvectors; Concatenate the congestion probability value, maintenance feature vector, and runtime state encoding vector into tensors, and generate time series correlation features based on the concatenated tensors. Generate state space based on temporal correlation features.

6. A blockage detection system for a flow measurement device, characterized in that: It includes acquisition module, vector module, prediction module and execution module, among which: an acquisition module configured to acquire an original pressure signal and an original flow signal of a flow measurement device to be measured, and generate an eigenmode function according to the original pressure signal and the original flow signal; The vector module is configured to extract the time vector, space vector, and frequency domain vector of the intrinsic mode function, obtain the attention weight through a two-way cross-attention mechanism, and perform a weighted fusion of the time vector, space vector, frequency domain vector, and attention weight to obtain a multi-dimensional feature vector; a prediction module configured to input the multi-dimensional feature vector into a pre-trained deep learning-based congestion probability prediction model to obtain a congestion probability value; The execution module is configured to construct a state space based on the congestion probability value, historical maintenance records, and equipment operation time, and to determine the execution action by balancing the false alarm and missed detection penalties in combination with the reward function. The acquisition module is configured to: Adding white noise of different amplitudes to the original pressure signal and the original flow signal multiple times to form multiple noisy signal sequences; Decompose each noisy signal sequence separately to obtain multiple intrinsic mode function components; Determine the sliding time window size and step size based on the signal sampling frequency and congestion detection requirements, and calculate the cross-correlation coefficient between the target intrinsic mode function component and the preset reference signal within each sliding time window, where the target intrinsic mode function component is any one of the multiple intrinsic mode function components; Recording the number of sliding time windows whose mutual correlation coefficient is lower than a preset coefficient threshold, when the number is greater than or equal to the preset number threshold, determining the target intrinsic mode function component as an abnormal mode component caused by congestion, and when the number is less than the preset number threshold, determining the target intrinsic mode function component as a normal mode component; Construct the eigenmode functions based on the normal mode components, The vector module is configured to: Sample the intrinsic mode function on the time axis, record the function value corresponding to each sampling point, and arrange these function values ​​in chronological order to form a time vector; Quantify the energy or amplitude of the intrinsic mode function at different frequency components, use the quantized values ​​as elements of the vector with frequency as the dimension, and construct a space vector; The Fourier transform is used to convert the intrinsic mode function from the time domain to the frequency domain to obtain the frequency domain representation of the signal in the frequency domain. The target features of the frequency domain representation are extracted and arranged in a preset order to form a frequency domain vector.

7. An electronic device, characterized in that: The electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 5 is performed.

Images