High-pressure gas filling equipment temperature anomaly detection method based on wireless sensing

Abstract

The invention relates to the technical field of high-pressure gas filling, in particular to a high-pressure gas filling equipment temperature anomaly detection method based on wireless sensing. Comprising the steps of constructing a state transition equation and an observation equation; collecting temperature data of a key area in the high-pressure gas filling equipment; the acquired temperature data is preprocessed by using an improved robust adaptive unscented Kalman filter; carrying out local fusion on the preprocessed temperature data based on a serialization fast covariance overlapping algorithm; global fusion is carried out on the temperature data after local fusion based on an improved zebra optimization algorithm; and comparing the global fusion result with a preset threshold value, and outputting a temperature detection result in the high-pressure filling equipment. The method can cope with the extreme environment in the high-pressure gas filling equipment, and ensures the accuracy and reliability of temperature monitoring. Based on the design of the wireless sensor network, the system has the characteristics of flexible deployment, low cost and strong expansibility, and can be widely applied to multi-point temperature monitoring in the fields of energy and chemical industry.

Description

Technical Field

[0001] The present invention relates to the technical field of high-pressure gas filling, and in particular to a method for detecting temperature anomaly of high-pressure gas filling equipment based on wireless sensing. Background Art

[0002] In the energy and chemical industries, high-pressure gas filling equipment, such as hydrogen storage and transportation tanks and natural gas compressors, is central to gas storage, transportation, and filling. They are widely used in fuel cells, oil and gas extraction, and other fields. This equipment must operate in extreme environments of high temperature and pressure for extended periods of time. Temperature fluctuations directly impact gas sealing, equipment safety, and process stability. Abnormal temperatures can lead to seal failure, gas leakage, and even explosions, seriously threatening production safety and economic profitability.

[0003] Traditional temperature monitoring for high-pressure gas filling equipment relies on wired sensors, but this approach has significant limitations. The cramped interior of the filling equipment and the harsh environment make wiring difficult and maintenance expensive. Furthermore, wired sensors are susceptible to mechanical vibration and electromagnetic noise in high-temperature, high-pressure, and strong electromagnetic fields, leading to data distortion. Furthermore, wired systems are difficult to flexibly add monitoring points, limiting the application of multi-point temperature monitoring and limiting scalability, further increasing safety risks.

[0004] To address these issues, wireless sensor networks, due to their flexible deployment, low cost, and strong scalability, have become an important technology for temperature monitoring in purification processes. However, in high-temperature and high-pressure environments, wireless sensor networks still face the following challenges:

[0005] 1. Sensor durability: High temperature, high pressure, corrosive gas and mechanical vibration place extremely high demands on the durability of the sensor;

[0006] 2. Data quality: Strong electromagnetic noise and equipment operation noise exist around the filling equipment, which mixes a large amount of non-Gaussian noise with the sensor data, affecting the monitoring accuracy.

[0007] 3. Data loss: Sensors may fail due to overheating or communication interruption, resulting in the loss of critical temperature data;

[0008] 4. Temperature Sudden Change Detection: The temperature in the process may change rapidly due to raw material fluctuations or equipment failures, requiring the monitoring system to have millisecond-level response capabilities;

[0009] 5. Multi-source data fusion: Large-scale high-pressure gas filling equipment deploys multiple sensors, and the data sources are heterogeneous and unevenly distributed, making fusion difficult.

[0010] Existing technologies fall short in addressing these challenges. Traditional Kalman filtering struggles to effectively handle nonlinear and non-Gaussian noise; Bayesian estimation suffers from high computational complexity when fusing multi-source data, making it difficult to meet real-time requirements; and traditional neural networks are prone to overfitting and poor robustness in the presence of strong noise or missing data. These deficiencies can lead to inaccurate temperature monitoring, potentially posing operational risks and safety hazards to equipment. Summary of the Invention

[0011] The present invention discloses a method for detecting temperature anomalies of high-pressure gas filling equipment based on wireless sensing. The specific method is as follows:

[0012] S1. Construct the state transfer equation and observation equation of high-pressure gas filling equipment;

[0013] S2. Collect temperature data of key areas inside the high-pressure gas filling equipment through wireless temperature sensors;

[0014] S3, preprocessing the collected temperature data using an improved robust adaptive unscented Kalman filter;

[0015] S4, locally fusion the pre-processed temperature data based on the serialized fast covariance overlap algorithm;

[0016] S5. Based on the improved zebra optimization algorithm, the temperature data after local fusion is globally fused;

[0017] S6. Compare the global fusion result with the preset threshold and output the temperature detection result inside the high-pressure filling equipment. Further, construct the state transfer equation and observation equation of the high-pressure gas filling equipment. The specific formula is as follows:

[0018] x a+1 =f(x a )+ω a ,a=0,1,...

[0019] z a =h(x a )+v a ,a=0,1,...

[0020] in, represents the temperature state vector of the high-pressure gas filling equipment system at time a+1; represents the sensor measurement output at time a; f(·) and h(·) represent the state transfer function and observation function of the system, respectively; ω a and v a They represent state noise and measurement noise respectively, both are zero-mean Gaussian white noise, and their covariances are Q a and R a .

[0021] Furthermore, the collected temperature data of key areas inside the high-pressure gas filling equipment includes: a timestamp, a sensor ID, a temperature value, and a check code.

[0022] Furthermore, step S3 pre-processes the collected temperature data using an improved robust adaptive unscented Kalman filter. The specific method is as follows:

[0023] S31. Using a Bernoulli process, missing data is identified for the collected temperature data, and the missing data is replaced with a one-step predicted value. The specific method is as follows;

[0024] First, according to the state x of the system at the previous moment a-1 And the state transition function f(·) predicts the current state in, represents the state prediction value at time a;

[0025] Then, using the system observation equation h(·) and the predicted state To predict the observation value at the current moment

[0026]

[0027] Predicted observations It is the one-step forecast value, which is used to fill the data gap when the actual observation value is missing:

[0028]

[0029] Among them, γ a =1 means the data is normal, γ a =0 means data is lost;

[0030] S32. Estimate and adjust the system noise covariance Q and the measurement noise covariance R to correct the state transfer equation and observation equation of the system. The specific method is as follows:

[0031] The system noise estimation formula is:

[0032]

[0033] in, is the Kalman gain, which is used to adjust the update amount of the state estimation according to the state error covariance matrix at the current moment The observation matrix H and the measurement noise covariance matrix R u Calculated; is the observation error, z u is the actual observed value, It is the observation prediction value obtained based on the state prediction at the previous moment; is the unscented transformation weight; x s are sampling points used to approximate the state distribution of the nonlinear system, generated by unscented transformation; is the final estimate of the system state, reflecting the mean of the state distribution; P u is the covariance matrix of the system state at the current moment;

[0034] The measurement noise estimation formula is:

[0035]

[0036] S33. Introduce the traceless transformation based on QR decomposition to ensure that the covariance matrix always remains positive definite;

[0037] S34. A dynamic adjustment robust factor mechanism is introduced. The dynamic adjustment formula of the robust factor is as follows:

[0038]

[0039] in, Represents the residual at the current moment, defined as the deviation between the actual observation value and the predicted value, ‖r a ‖ represents the norm of the residual, reflecting the size of the residual. a When λ is large, a Approaching 0 indicates that the filter has low confidence in the current observation value; when the residual r a When λ is small, a It approaches 1, indicating that the filter has a high degree of trust in the current observation.

[0040] Furthermore, in step S33, an unscented transformation based on QR decomposition is introduced, and the specific method is as follows:

[0041] S331. Construct an augmented matrix A containing the state vector x and the covariance matrix P, where A contains the state variables and their disturbance terms. Assume that the current state vector is The covariance matrix is Then the augmented matrix A can be expressed as:

[0042]

[0043] S332. Perform QR decomposition on the augmented matrix A to obtain the orthogonal matrix Q and the upper triangular matrix R:

[0044] A=QR

[0045] S333. Extract Sigma points based on the upper triangular matrix R for subsequent state prediction and observation update:

[0046] x0=x

[0047]

[0048] Among them, x i is the i-th Sigma point, represents the i-th column extracted from the upper triangular matrix R and multiplied by the scaling factor

[0049] Furthermore, in step S4, the pre-processed temperature data is locally fused, and the specific method is as follows:

[0050] S41, using the estimated state value of each sensor and its corresponding error covariance matrix P i As input data, is the state estimation value of the i-th sensor, P i is the error covariance matrix representing the estimated value of the i-th sensor state, reflecting the uncertainty of the sensor data;

[0051] S42, fusing the sensor data in the order of arrival;

[0052] S43. In the process of serialization fusion, the dynamic weight w i , the fusion weight w of the i-th sensor i The calculation formula is:

[0053]

[0054] Where trace(·) represents the trace of the matrix, i.e., the sum of the main diagonal elements; P i and P fused Represent the error covariance matrix of the current sensor and the fusion result respectively;

[0055] S44, when the error covariance matrix P of the sensor data i When the weight w is small i Increases, on the contrary, when P i When the weight w is large, i Decrease.

[0056] Furthermore, in step S42, the sensor data are fused in the order in which they arrive. The specific method is as follows:

[0057] S421, set the state estimation value of the first sensor And the error covariance matrix P1 is used as the initial fusion result:

[0058]

[0059] S422: For each subsequent sensor data Combine it with the current result The fusion formula is as follows:

[0060]

[0061] P fused =ω i ·P i +(1-ω i )·P fused

[0062] Among them, w i is the dynamically assigned fusion weight, which indicates the importance of the i-th sensor data in the fusion;

[0063] S423. According to the above formula, all sensor data are sequentially integrated into the current fusion result, and finally a globally consistent local fusion result is obtained.

[0064] Furthermore, in step S5, the temperature data after local fusion is globally fused. The specific method is as follows:

[0065] S51. Introduce contrastive learning mechanism in the population initialization stage by constructing positive and negative samples;

[0066] S52. Introduce the weight factor θ during the optimization process. The update formula of the dynamic weight factor is:

[0067]

[0068] Where α is the adjustment step size, t is the current number of iterations, and T is the maximum number of iterations. As the number of iterations increases, the global search weight gradually decreases and the local search weight gradually increases, thus achieving a smooth transition from global to local;

[0069] S53. Introduce the Levy flight mechanism to enhance the algorithm's exploration ability in the solution space. The Levy flight update formula is:

[0070]

[0071] in, represents the position of the i-th individual in the j-th dimension; x best represents the position of the current optimal solution in the jth dimension; Levy(λ) is a random step size that follows the Levy distribution and is used to simulate the jumping behavior of animals when foraging;

[0072] S54. Introduce Gaussian distribution to generate new solutions to avoid the population falling into local optimality. The Gaussian mutation formula is:

[0073] x new =x best +σ·Gaussian(0,1)

[0074] Among them, x new is a new solution generated by Gaussian mutation, x best is the optimal solution in the current population, and σ is the standard deviation of the control variation range.

[0075] Furthermore, the model for global fusion in step S5 is a CNN-BiLSTM hybrid neural network model, and the specific workflow is as follows:

[0076] A convolutional neural network module is used to extract the spatial distribution characteristics of temperature data. The input layer receives the locally fused temperature data and has a shape of [N, T, F], where N is the number of samples, T is the time step, and F is the feature dimension. The convolutional layer contains three layers of one-dimensional convolution with kernel sizes of 5, 3, and 3, respectively. Each layer is followed by a ReLU activation function and a BatchNormalization layer to accelerate the training process and prevent gradient vanishing. Each convolution layer is followed by a max pooling layer to reduce the dimensionality while retaining important features. The output shape of the output layer is [N, T', C], where T' is the length of the time step after pooling and C is the number of feature channels.

[0077] A bidirectional long short-term memory network module is used to capture the time series dependencies of temperature data. The input layer receives the output of the CNN module and has a shape of [N, T', C]. The BiLSTM layer contains two layers of bidirectional LSTM with 64 hidden units, processing the time series data in the forward and backward directions respectively. A dropout layer is added after each BiLSTM layer to prevent overfitting. The output shape of the output layer is [N, H], where H is the dimension of the time series feature vector.

[0078] The output feature vector of BiLSTM is mapped to the final temperature prediction value through a fully connected layer. The fully connected layer consists of two layers. The first layer uses ReLU as the activation function and the output dimension is 128. The second layer uses linear activation function and the output dimension is 1.

[0079] Model training uses mean square error as the loss function. The specific formula is as follows:

[0080]

[0081] Among them, y i is the true temperature value, is the predicted temperature value.

[0082] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0083] 1. Improved Data Quality: Using an improved unscented Kalman filter preprocessing mechanism, the quality of raw data is enhanced through a noise adaptation mechanism, QR decomposition optimization, and dynamic adjustment of robustness factors. Under high temperature and high pressure environments, the effects of non-Gaussian noise, outliers, and missing data are filtered out, providing high-quality input data for subsequent fusion.

[0084] 2. Improved Computational Efficiency and Consistency: Local fusion based on the serialized fast covariance crossover algorithm avoids the high computational complexity of processing all sensor data at once, making it suitable for real-time applications in large-scale sensor networks. A dynamic weight allocation mechanism adjusts weights based on the reliability of sensor data, ensuring the consistency and accuracy of fusion results.

[0085] 3. High-Precision Temperature Monitoring: Based on a global fusion model of a convolutional neural network and a bidirectional long short-term memory network optimized using an improved Zebra optimization algorithm, the neural network parameters are optimized to make the model more robust in complex industrial environments. The convolutional neural network module extracts spatial features, while the bidirectional long short-term memory network module captures time series dependencies. The combination of the two improves the accuracy of temperature prediction.

[0086] 4. Industrial Applicability: This method can withstand the extreme environments inside high-pressure gas filling equipment, ensuring accurate and reliable temperature monitoring. This wireless sensor network-based design offers flexible deployment, low cost, and strong scalability, making it widely applicable for multi-point temperature monitoring in the energy and chemical industries.

[0087] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The accompanying drawings of the present invention are described below.

[0089] Figure 1 A schematic diagram of the overall process.

[0090] Figure 2 A schematic diagram of the algorithm structure.

[0091] Figure 3 This is a flow chart of the serialized fast protocol difference crossover algorithm. DETAILED DESCRIPTION

[0092] The present invention will be further described below with reference to the accompanying drawings and examples.

[0093] A method for detecting abnormal temperature of high-pressure gas filling equipment based on wireless sensing, such as Figure 1and Figure 2 The specific steps are as follows:

[0094] S1. Construct the state transfer equation and observation equation of high-pressure gas filling equipment.

[0095] Assume that the state transfer equation and observation equation of the high-pressure gas filling equipment are:

[0096] x a+1 =f(x a )+ω a ,a=0,1,...

[0097] z a =h(x a )+v a ,a=0,1,...

[0098] in, represents the temperature state vector of the high-pressure gas filling equipment system at time a+1; represents the sensor measurement output at time a; f(·) and h(·) represent the state transfer function and observation function of the system, respectively; ω a and v a They represent state noise and measurement noise respectively, both are zero-mean Gaussian white noise, and their covariances are Q a and R a .

[0099] S2. Collect temperature data of key areas inside the high-pressure gas filling equipment through wireless temperature sensors.

[0100] High-precision wireless temperature sensors are deployed in key areas within high-pressure gas filling equipment to form a redundant, multi-point monitoring network. The sensors utilize a high-temperature, corrosion-resistant packaging design, and a star-shaped network topology is established using an ad hoc networking protocol to ensure stable data transmission. Each sensor node is equipped with redundant power modules and electromagnetic shielding to mitigate the risk of communication interruptions in extreme environments.

[0101] High-precision wireless temperature sensors are deployed in key areas within high-pressure gas filling equipment to form a redundant, multi-point monitoring network. The sensors utilize a high-temperature, corrosion-resistant packaging design, and a star-shaped network topology is established using an ad hoc networking protocol to ensure stable data transmission. Each sensor node is equipped with redundant power modules and electromagnetic shielding to mitigate the risk of communication interruptions in extreme environments.

[0102] S3. Preprocessing the collected temperature data with an improved robust adaptive unscented Kalman filter.

[0103] The improved unscented Kalman filter is used to preprocess the collected raw temperature data to remove the non-Gaussian noise and outliers in the industrial environment. The specific mechanism includes:

[0104] S31, Random Missing Data Identification

[0105] Missing data are identified through the Bernoulli process, and the missing data are replaced with one-step predicted values.

[0106] First, according to the state x of the system at the previous moment a-1 And the state transition function f(·) predicts the current state

[0107]

[0108] in, Represents the state prediction value at time a.

[0109] Then, using the system observation equation h(·) and the predicted state To predict the observation value at the current moment

[0110]

[0111] The predicted observed value It is the one-step forecast value, which is used to fill the data gap when the actual observation value is missing:

[0112]

[0113] Among them, γ a =1 means the data is normal, γ a =0 indicates data loss.

[0114] S32, Noise Adaptive Mechanism

[0115] The system noise covariance Q describes the impact of process fluctuations on temperature states, while the measurement noise covariance R reflects the intensity of sensor data fluctuations under high-temperature electromagnetic interference. To improve filtering accuracy, the system noise covariance Q and measurement noise covariance R are estimated and adjusted online to modify the system's state transition equations and observation equations to adapt to noise changes caused by process fluctuations or environmental interference during temperature monitoring.

[0116] The system noise estimation formula is:

[0117]

[0118] in, is the Kalman gain, which is used to adjust the update amount of the state estimation according to the state error covariance matrix at the current moment The observation matrix H and the measurement noise covariance matrix R u Calculated; is the observation error, z u is the actual observed value, It is the observation prediction value obtained based on the state prediction at the previous moment; is the unscented transformation weight; x s are sampling points used to approximate the state distribution of the nonlinear system, generated by unscented transformation; is the final estimate of the system state, reflecting the mean of the state distribution; P u is the covariance matrix of the system state at the current moment.

[0119] The measurement noise estimation formula is:

[0120]

[0121] S33, Unscented Transformation Based on QR Decomposition

[0122] In high-temperature and high-pressure environments, sensor data is contaminated with a large amount of non-Gaussian noise and outliers. The covariance matrix may lose positive definiteness due to numerical instability, such as a negative determinant, leading to filtering failure. To address this issue, this method introduces an unscented transformation based on QR decomposition. Using QR decomposition ensures that the covariance matrix remains positive definite, ensuring stable operation even in the presence of strong noise or missing data. This method is particularly suitable for industrial applications in extreme environments such as high temperature and high pressure, improving the reliability of temperature monitoring systems.

[0123] The specific steps are as follows:

[0124] S331, construct augmented matrix

[0125] Construct an augmented matrix A containing the state vector x and the covariance matrix P, where A contains the state variables and their disturbance terms. Assume that the current state vector is The covariance matrix is Then the augmented matrix A can be expressed as:

[0126]

[0127] S332, perform QR decomposition

[0128] Perform QR decomposition on the augmented matrix A to obtain the orthogonal matrix Q and the upper triangular matrix R:

[0129] A=QR

[0130] S333, extract Sigma points

[0131] Extract Sigma points based on the upper triangular matrix R for subsequent state prediction and observation update:

[0132] x0=x

[0133]

[0134] Among them, x i is the i-th Sigma point, represents the i-th column extracted from the upper triangular matrix R and multiplied by the scaling factor

[0135] S34. Dynamic adjustment of robustness factor

[0136] In high-temperature and high-pressure environments, sensor data is susceptible to sudden changes in values ​​(such as transient interference and equipment failures). These outliers can distort the filtering results. To mitigate the impact of outliers on filter performance, this paper introduces a dynamically adjusted robustness factor mechanism.

[0137] The robustness factor λ is a weight coefficient used to dynamically adjust the filter's confidence in the current observation. When the residual is large, the robustness factor is reduced to reduce the impact of outliers on the filtering results. When the residual is small, the robustness factor is increased to enhance the filter's responsiveness to normal data.

[0138] The dynamic adjustment formula of the robust factor is as follows:

[0139]

[0140] in, Represents the residual at the current moment, defined as the deviation between the actual observation value and the predicted value, ‖r a ‖ represents the norm of the residual, reflecting the size of the residual. a When λ is large, a Approaching 0 indicates that the filter has low confidence in the current observation value; when the residual r a When λ is small, a It approaches 1, indicating that the filter has a high degree of trust in the current observation.

[0141] During the dynamic adjustment of the robust factor, the initial value of the robust factor is set to 1, indicating full trust in the observed value. a , dynamically adjust the robustness factor λ a , ensuring that the filter can quickly adapt to environmental changes. When an outlier is detected (such as a sudden increase in the residual), the robust factor is quickly reduced to reduce the impact of the outlier on the filtering result.

[0142] By dynamically adjusting the robustness factor, the impact of sensor failures or transient interference on the filtering results is effectively suppressed. Automatic adjustments based on the size of the residual error are possible without manual intervention, adapting to complex and changing industrial environments. Even in the presence of strong noise or missing data, the filter maintains high accuracy and stability.

[0143] S4. Locally fuse the pre-processed temperature data based on the serialized fast covariance overlap algorithm.

[0144] In a high-temperature and high-pressure environment, multiple wireless temperature sensors are deployed inside high-pressure gas filling equipment. The data collected by these sensors are heterogeneous and unevenly distributed. In order to extract reliable and consistent temperature information from multi-source data, local fusion methods such as Figure 3 As shown in the figure, through dynamic weight allocation and fast covariance calculation, the computational complexity is significantly reduced while ensuring the consistency and accuracy of the fusion results.

[0145] At the local fusion center, the error covariance matrix reflects each sensor node's measure of the uncertainty of its state estimate. To handle the unknown correlation matrix in sensor state estimation, the present invention uses a fast diagonal weight coefficient covariance cross-fusion algorithm. By iteratively updating the weight coefficients, the computational cost is reduced and the influence of different node fusion order on the results is eliminated.

[0146] The specific steps of local fusion are as follows:

[0147] S41. Input data

[0148] The input is the estimated value of the state of each sensor and its corresponding error covariance matrix P i .in, is the state estimation value of the i-th sensor, P i is the error covariance matrix of the i-th sensor state estimate, reflecting the uncertainty of sensor data. Because sensors may be affected by noise or data loss, the error covariance matrices of different sensors can vary significantly. Therefore, a dynamic weighting mechanism is needed to properly assess the reliability of each sensor's data.

[0149] S42, serialization fusion process

[0150] Traditional covariance crossover algorithms typically require processing all sensor data at once, which leads to high computational complexity in large-scale sensor networks and makes it difficult to meet real-time requirements. To this end, this paper adopts a serialized fusion strategy, fusing sensor data in the order they arrive. The specific steps are as follows:

[0151] S421, Initialization

[0152] Let the state estimate of the first sensor be And the error covariance matrix P1 is used as the initial fusion result:

[0153]

[0154] S422, successive fusion

[0155] For each subsequent sensor data Combine it with the current result Fusion is performed. The fusion formula is as follows:

[0156]

[0157] P fused =ω i ·P i +(1-ω i )·P fused

[0158] Among them, w i is the dynamically assigned fusion weight, which indicates the importance of the i-th sensor data in the fusion.

[0159] S423, repeated updates

[0160] According to the above formula, all sensor data are integrated into the current fusion result in turn, and finally a globally consistent local fusion result is obtained.

[0161] Through serialization processing, the high computational burden of processing all data at once is avoided, while ensuring the real-time and consistency of the fusion results.

[0162] S43. Weight calculation

[0163] In the process of serialized fusion, the dynamic weight w i The distribution of is the key link. The present invention dynamically calculates weights based on the diagonal elements of the error covariance matrix, so that more reliable sensor data occupies a larger proportion in the fusion. The specific implementation is as follows:

[0164] S431, weight formula

[0165] The fusion weight w of the i-th sensor i The calculation formula is:

[0166]

[0167] Where trace(·) represents the trace of the matrix, i.e., the sum of the main diagonal elements; P i and P fused Represent the error covariance matrices of the current sensor and the fusion result respectively.

[0168] S432, Dynamic Adjustment

[0169] When the error covariance matrix P of the sensor data i When it is smaller (i.e., data reliability is higher), the weight w i Increases, indicating that the sensor data occupies a higher proportion in the fusion; on the contrary, when P i When the weight w is large, i Reduce, reduce the impact of the sensor data on the fusion result. In extreme cases, such as sensor failure or severe data distortion, the weight range can be limited by setting a threshold to ensure the stability of the fusion result. i When a preset threshold is exceeded, the sensor data is directly ignored.

[0170] Compared with the traditional covariance crossover algorithm, the serialized fast covariance crossover algorithm (SFDCI-CHNFE) proposed in this paper avoids high-complexity operations such as matrix inversion through serialization processing, significantly reducing the computational burden and making it suitable for real-time applications in large-scale sensor networks. Even when the sensor data sources are heterogeneous and unevenly distributed, the consistency of the fusion results can still be guaranteed, avoiding fusion errors caused by unknown data correlation. The dynamic weight allocation mechanism can adjust the fusion weights in real time according to the reliability of the sensor data, adapting to complex and changing industrial environments. Through the dynamic evaluation of the error covariance matrix, the influence of noise and outliers on the fusion results is effectively suppressed, improving the robustness of the system.

[0171] S5. Based on the improved zebra optimization algorithm, the temperature data after local fusion is globally fused.

[0172] In high-temperature, high-pressure environments, multiple wireless temperature sensors are deployed within high-pressure gas filling equipment. Local fusion centers already maintain consistency in multi-source data. However, to further enhance the accuracy and robustness of temperature monitoring, this paper proposes a global fusion method based on a hybrid convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM) model optimized with an improved Zebra Optimization Algorithm (IZOA). By optimizing network parameters and combining the extraction of spatial features and time series dependencies, this method achieves high-precision and robust temperature fusion in complex industrial environments.

[0173] The specific steps are as follows:

[0174] S51, improved zebra optimization algorithm

[0175] To solve the problem that traditional optimization algorithms are prone to falling into local optimality in complex solution spaces, this paper improves the classic Zebra Optimization Algorithm (ZOA) and proposes IZOA. The specific steps are as follows:

[0176] S511, Random Contrastive Learning

[0177] During the population initialization phase, a contrastive learning mechanism is introduced to improve the diversity and quality of initial solutions by constructing positive and negative samples. Positive samples are solutions that are close to the current individual in the solution space, while negative samples are solutions that are far away from the current individual. The goal is to maximize the similarity between positive samples and the current individual while minimizing the similarity between negative samples and the current individual. This method effectively prevents premature population convergence and enhances the algorithm's global search capabilities.

[0178] S512, dynamic weight factor

[0179] In order to balance the capabilities of global search and local search and ensure that the algorithm gradually transitions from global search to local search, a weight factor θ is introduced in the optimization process. The update formula of the dynamic weight factor is:

[0180]

[0181] Where α is the adjustment step size, t is the current number of iterations, and T is the maximum number of iterations. As the number of iterations increases, the global search weight gradually decreases and the local search weight gradually increases, thus achieving a smooth transition from global to local.

[0182] S513, Levy flight strategy

[0183] The Levy flight mechanism is introduced to enhance the algorithm's ability to explore the solution space, especially to quickly jump in complex solution spaces to find potential optimal solutions. The Levy flight update formula is:

[0184]

[0185] in, represents the position of the i-th individual in the j-th dimension; x best represents the position of the current optimal solution in the jth dimension; Levy(λ) is a random step size that obeys the Levy distribution and is used to simulate the jumping behavior of animals when foraging.

[0186] S514, Gaussian mutation and greedy selection

[0187] By introducing Gaussian distribution to generate new solutions, the population can avoid falling into local optimality. The Gaussian mutation formula is:

[0188] x new =x best +σ·Gaussian(0,1)

[0189] Among them, x new is a new solution generated by Gaussian mutation, x best is the optimal solution in the current population, and σ is the standard deviation of the control variation range.

[0190] A greedy selection strategy is used to retain high-quality solutions, retaining only solutions with higher fitness in each iteration, ensuring a gradual improvement in population quality. Through these improvements, IZOA demonstrates stronger exploration capabilities and faster convergence in complex solution spaces, providing reliable support for subsequent neural network parameter optimization.

[0191] S52, CNN-BiLSTM hybrid neural network model

[0192] In the global fusion center, the present invention uses a CNN-BiLSTM hybrid neural network model to further process the local fusion results and extract the spatial features and time series dependencies of the temperature data. The specific structure is as follows:

[0193] S521, CNN module

[0194] The convolutional neural network (CNN) module is used to extract the spatial distribution characteristics of temperature data.

[0195] The input layer is used to receive the temperature data after local fusion, and its shape is [N, T, F], where N is the number of samples, T is the time step, and F is the feature dimension.

[0196] The convolutional layer consists of three layers of one-dimensional convolutions with kernel sizes of 5, 3, and 3, respectively. Each layer is followed by a ReLU activation function and a BatchNormalization layer to accelerate training and prevent gradient vanishing. Each convolution layer is followed by a max pooling layer to reduce dimensionality while retaining important features.

[0197] The output shape of the output layer is [N, T', C], where T' is the length of the time step after pooling and C is the number of feature channels.

[0198] S522, BiLSTM module

[0199] The BiLSTM module is used to capture the time series dependency of temperature data. Its structure is designed as follows:

[0200] The input layer receives the output of the CNN module and has a shape of [N, T', C].

[0201] The BiLSTM layer contains two layers of bidirectional LSTM with 64 hidden units, which process time series data from the forward and backward directions respectively.

[0202] A Dropout layer is added after each BiLSTM layer to prevent overfitting.

[0203] The output shape of the output layer is [N,H], where H is the dimension of the time series feature vector.

[0204] S523, fully connected layer

[0205] The BiLSTM output feature vector is mapped to the final temperature prediction value through a fully connected layer. The fully connected layer consists of two layers. The first layer uses the ReLU activation function and has an output dimension of 128. The second layer uses the linear activation function and has an output dimension of 1, which is the predicted temperature value.

[0206] S524. Loss function

[0207] Model training uses mean square error (MSE) as the loss function to calculate the deviation between the predicted temperature value and the true value:

[0208]

[0209] Among them, y i is the true temperature value, is the predicted temperature value.

[0210] Compared to traditional neural networks, the proposed method uses a CNN module to extract spatial features and a BiLSTM module to capture time series dependencies. The combination of these two significantly improves temperature prediction accuracy. IZOA's optimized network parameters are adaptable to noise and missing data in complex industrial environments, preventing overfitting. Gaussian mutation and greedy selection strategies further enhance the model's robustness. By dynamically adjusting network parameters, the model can adapt to temperature variations under varying process conditions, demonstrating excellent generalization capabilities.

[0211] S6. Compare the global fusion result with the preset threshold and output the temperature detection result inside the high-pressure filling equipment.

[0212] Based on the global fusion results, the system outputs real-time temperature monitoring values ​​inside the high-pressure gas filling equipment. This fused temperature value is compared with a preset threshold. If it exceeds the threshold, an alarm is triggered and the control system is activated (for example, adjusting the feed rate or cooling water flow) to ensure safe and stable process operation.

[0213] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for detecting temperature anomaly of high-pressure gas filling equipment based on wireless sensing, characterized in that: The specific method is as follows: S1. Construct the state transfer equation and observation equation of high-pressure gas filling equipment; S2. Collect temperature data of key areas inside the high-pressure gas filling equipment through wireless temperature sensors; S3, preprocessing the collected temperature data using an improved robust adaptive unscented Kalman filter; S4, locally fusion the pre-processed temperature data based on the serialized fast covariance overlap algorithm; S5. Based on the improved zebra optimization algorithm, the temperature data after local fusion is globally fused; S6. Compare the global fusion result with the preset threshold and output the temperature detection result inside the high-pressure filling equipment.

2. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 1, characterized in that: Construct the state transfer equation and observation equation of the high-pressure gas filling equipment. The specific formulas are as follows: x a+1 =f(x a )+ω a ,a=0,1,... z a =h(x a )+v a ,a=0,1,... in, represents the temperature state vector of the high-pressure gas filling equipment system at time a+1; represents the sensor measurement output at time a; f(·) and h(·) represent the state transfer function and observation function of the system, respectively; ω a and v a They represent state noise and measurement noise respectively, both are zero-mean Gaussian white noise, and their covariances are Q a and R a .

3. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 1, characterized in that: The collected temperature data of key areas inside the high-pressure gas filling equipment includes: timestamp, sensor ID, temperature value and check code.

4. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 2, wherein: Step S3 pre-processes the collected temperature data using an improved robust adaptive unscented Kalman filter. The specific method is as follows: S31. Using a Bernoulli process, missing data is identified for the collected temperature data, and the missing data is replaced with a one-step predicted value. The specific method is as follows; First, according to the state x of the system at the previous moment a-1 And the state transition function f(·) predicts the current state in, represents the state prediction value at time a; Then, using the system observation equation h(·) and the predicted state To predict the observation value at the current moment Predicted observations It is the one-step forecast value, which is used to fill the data gap when the actual observation value is missing: Among them, γ a =1 means the data is normal, γ a =0 means data is lost; S32. Estimate and adjust the system noise covariance Q and the measurement noise covariance R to correct the state transfer equation and observation equation of the system. The specific method is as follows: The system noise estimation formula is: in, is the Kalman gain, which is used to adjust the update amount of the state estimation according to the state error covariance matrix at the current moment The observation matrix H and the measurement noise covariance matrix R u Calculated; is the observation error, z u is the actual observed value, It is the observation prediction value obtained based on the state prediction at the previous moment; is the unscented transformation weight; x s are sampling points used to approximate the state distribution of the nonlinear system, generated by unscented transformation; is the final estimate of the system state, reflecting the mean of the state distribution; P u is the covariance matrix of the system state at the current moment; The measurement noise estimation formula is: S33. Introduce the traceless transformation based on QR decomposition to ensure that the covariance matrix always remains positive definite; S34. A dynamic adjustment robust factor mechanism is introduced. The dynamic adjustment formula of the robust factor is as follows: in, Represents the residual at the current moment, defined as the deviation between the actual observation value and the predicted value, ‖r a ‖ represents the norm of the residual, reflecting the size of the residual. a When λ is large, a Approaching 0 indicates that the filter has low confidence in the current observation value; when the residual r a When λ is small, a It approaches 1, indicating that the filter has a high degree of trust in the current observation.

5. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 3, characterized in that: In step S33, an unscented transformation based on QR decomposition is introduced. The specific method is as follows: S331. Construct an augmented matrix A containing the state vector x and the covariance matrix P, where A contains the state variables and their disturbance terms. Assume that the current state vector is The covariance matrix is Then the augmented matrix A can be expressed as: S332. Perform QR decomposition on the augmented matrix A to obtain the orthogonal matrix Q and the upper triangular matrix R: A=QR S333. Extract Sigma points based on the upper triangular matrix R for subsequent state prediction and observation update: x0=x Among them, x i is the i-th Sigma point, represents the i-th column extracted from the upper triangular matrix R and multiplied by the scaling factor 6. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 2, characterized in that: In step S4, the pre-processed temperature data is locally fused. The specific method is as follows: S41, using the estimated state value of each sensor and its corresponding error covariance matrix P i As input data, is the state estimation value of the i-th sensor, P i is the error covariance matrix representing the estimated value of the i-th sensor state, reflecting the uncertainty of the sensor data; S42, fusing the sensor data in the order of arrival; S43. In the process of serialization fusion, the dynamic weight w i , the fusion weight w of the i-th sensor i The calculation formula is: Where trace(·) represents the trace of the matrix, i.e., the sum of the main diagonal elements; P i and P fused Represent the error covariance matrix of the current sensor and the fusion result respectively; S44, when the error covariance matrix P of the sensor data i When the weight w is small i Increases, on the contrary, when P i When the weight w is large, i Decrease.

7. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 6, characterized in that: In step S42, the sensor data are fused in the order in which they arrive. The specific method is as follows: S421, set the state estimation value of the first sensor And the error covariance matrix P1 is used as the initial fusion result: P fused =P1 S422: For each subsequent sensor data Combine it with the current result The fusion formula is as follows: P fused =ω i ·P i +(1-ω i )·P fused Among them, w i is the dynamically assigned fusion weight, which indicates the importance of the i-th sensor data in the fusion; S423. According to the above formula, all sensor data are sequentially integrated into the current fusion result, and finally a globally consistent local fusion result is obtained.

8. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 2, characterized in that: In step S5, the temperature data after local fusion is globally fused. The specific method is as follows: S51. Introduce contrastive learning mechanism in the population initialization stage by constructing positive and negative samples; S52. Introduce the weight factor θ during the optimization process. The update formula of the dynamic weight factor is: Where α is the adjustment step size, t is the current number of iterations, and T is the maximum number of iterations. As the number of iterations increases, the global search weight gradually decreases and the local search weight gradually increases, thus achieving a smooth transition from global to local; S53. Introduce the Levy flight mechanism to enhance the algorithm's exploration ability in the solution space. The Levy flight update formula is: in, represents the position of the i-th individual in the j-th dimension; x best represents the position of the current optimal solution in the jth dimension; Levy(λ) is a random step size that follows the Levy distribution and is used to simulate the jumping behavior of animals when foraging; S54. Introduce Gaussian distribution to generate new solutions to avoid the population falling into local optimality. The Gaussian mutation formula is: x new =x best +σ·Gaussian(0,1) Among them, x new is a new solution generated by Gaussian mutation, x best is the optimal solution in the current population, and σ is the standard deviation of the control variation range.

9. The method for detecting abnormal internal temperature of high-pressure gas filling equipment based on a wireless sensor according to claim 2, wherein: The model for global fusion in step S5 is a CNN-BiLSTM hybrid neural network model. The specific workflow is as follows: A convolutional neural network module is used to extract the spatial distribution characteristics of temperature data. The input layer receives the locally fused temperature data and has a shape of [N, T, F], where N is the number of samples, T is the time step, and F is the feature dimension. The convolutional layer contains three layers of one-dimensional convolution with kernel sizes of 5, 3, and 3, respectively. Each layer is followed by a ReLU activation function and a BatchNormalization layer to accelerate the training process and prevent gradient vanishing. Each convolution layer is followed by a max pooling layer to reduce the dimensionality while retaining important features. The output shape of the output layer is [N, T', C], where T' is the length of the time step after pooling and C is the number of feature channels. A bidirectional long short-term memory network module is used to capture the time series dependencies of temperature data. The input layer receives the output of the CNN module and has a shape of [N, T', C]. The BiLSTM layer contains two layers of bidirectional LSTM with 64 hidden units, processing the time series data in the forward and backward directions respectively. A dropout layer is added after each BiLSTM layer to prevent overfitting. The output shape of the output layer is [N, H], where H is the dimension of the time series feature vector. The output feature vector of BiLSTM is mapped to the final temperature prediction value through a fully connected layer. The fully connected layer consists of two layers. The first layer uses ReLU as the activation function and the output dimension is 128. The second layer uses linear activation function and the output dimension is 1. Model training uses mean square error as the loss function. The specific formula is as follows: Among them, y i is the true temperature value, is the predicted temperature value.

Images