A data state detection and determination method based on a BP neural network model

Abstract

The application relates to the technical field of data state detection and judgment, in particular to a data state detection and judgment method based on a BP neural network model, which comprises the following steps: obtaining historical normal operation data of a target measuring point and associated measuring points, training a BP neural network model by using the data set, and taking the operation data of the associated measuring points as input and taking the operation data of the target measuring point as output; inputting the real-time collected operation data of the associated measuring points into the trained BP neural network model, and obtaining the predicted value of the operation data of the target measuring point through calculation; performing difference calculation on the real-time measured value of the target measuring point and the predicted value output by the model, and obtaining a prediction residual; extracting statistical characteristic quantities of the prediction residual in a sliding time window, and constructing a dynamic judgment threshold which can be updated in real time according to the statistical characteristic quantities; comparing the current prediction residual with the dynamic judgment threshold, combining the cumulative occurrence of the comparison result in a preset observation window, and outputting the data state judgment result of the target measuring point.

Description

Technical Field

[0001] This invention relates to the field of data status detection and determination technology, and more specifically to a data status detection and determination method based on a BP neural network model. Background Technology

[0002] In the actual operation of complex industrial systems such as heavy-duty gas turbines, aero engines, and steam turbines, various measuring points are distributed in different parts of the system. Real-time monitoring of the equipment's operating status is crucial for ensuring the safe and reliable operation of the system. When an anomaly or malfunction occurs at any point, its impact will propagate throughout the system through mass flow, energy flow, and control flow, causing some measuring points to deviate from their normal operating range. Therefore, accurate and timely detection and judgment of the status data at each measuring point are fundamental to industrial system condition monitoring, fault early warning, and health management.

[0003] Existing data status detection methods mainly include fixed threshold-based methods, rule-based reasoning methods, and physical model-based methods. Fixed threshold-based methods are simple to implement but struggle to adapt to situations involving shifting operating conditions and multivariate coupling, and have poor ability to distinguish between slowly changing anomalies and reasonable deviations caused by changes in operating conditions. Rule-based reasoning methods rely on the experience of domain experts and have limited ability to describe complex nonlinear relationships, making it difficult to cover all operating conditions. Physical model-based methods offer high accuracy but are difficult to model, with model parameters drifting as equipment ages, and the cost of detailed modeling for complex systems is high.

[0004] Backpropagation (BP) neural networks, as a classic multilayer feedforward neural network, possess excellent nonlinear mapping capabilities, self-learning ability, and fault tolerance. They can adaptively establish nonlinear mapping relationships between inputs and outputs using sample data and have been widely applied in industrial data modeling, fault diagnosis, and condition assessment. However, directly applying conventional BP neural networks to condition detection and judgment of industrial measurement point data still presents the following problems: First, the standard BP algorithm uses a fixed learning rate, resulting in slow convergence and a tendency to get trapped in local minima, making it difficult for the trained network to meet the needs of condition detection. Second, the single fixed threshold method struggles to balance false alarm and false negative rates and lacks robustness against operating condition drift and random noise. Third, it fails to adequately utilize the temporal statistical characteristics of prediction residuals, exhibiting weak identification capabilities for slowly changing anomalies. Therefore, to address these issues, this invention proposes a data condition detection method that integrates data preprocessing, improved BP neural network modeling, and dynamic adaptive judgment. Summary of the Invention

[0005] This invention provides a data state detection and determination method based on a BP neural network model, with the aim of improving the state monitoring level of complex industrial systems.

[0006] The above objectives are achieved through the following technical solutions:

[0007] A data state detection and determination method based on a BP neural network model includes:

[0008] S1: Obtain historical normal operation data of the target measurement point and its associated measurement points, and train a BP neural network model using the mentioned historical normal operation data. The mentioned BP neural network model takes the operation data of the mentioned associated measurement points as input and the operation data of the mentioned target measurement point as output.

[0009] S2: Input the real-time collected data from the associated measurement points into the trained BP neural network model to obtain the predicted value of the target measurement point.

[0010] S3: Calculate the prediction residual based on the measured values ​​of the mentioned target measurement points and the mentioned predicted values;

[0011] S4: Based on the statistical features of the predicted residuals mentioned within the sliding time window, construct a dynamic judgment threshold that is updated in real time with the mentioned statistical features.

[0012] S5: Based on the comparison results of the predicted residuals mentioned and the dynamic judgment thresholds mentioned, and combined with the cumulative occurrence of the comparison results mentioned within the preset observation window, output the judgment result of the target measurement point data status mentioned.

[0013] The historical normal operation data mentioned also includes coarse error removal and normalization preprocessing before being used to train the mentioned BP neural network model:

[0014] S11: Calculate the mean of the data after removing trend. and standard deviation :

[0015]

[0016] S12: Calculate the residuals of the data at each time point, we have:

[0017]

[0018] S13: Connect the residuals one by one with... If a comparison is made, The corresponding data is an outlier and should be removed or replaced with the mean; otherwise, the data should be retained.

[0019] S14: Standardize the data after removing outliers to [0,1].

[0020] The BP neural network model mentioned adopts a three-layer structure consisting of an input layer, a hidden layer, and an output layer;

[0021] The number of hidden layer nodes h is calculated as follows: Let m be the number of input layer nodes, n be the number of output layer nodes, and a be an adjustable constant, with a value ranging from 1 to 10.

[0022] The BP neural network model mentioned in the training section employs an improved gradient descent algorithm that combines an adaptive learning rate with an additional momentum term; the adaptive learning rate mentioned decays with the number of iterations t.

[0023]

[0024] in The initial learning rate, As the attenuation factor, The iteration number is given; the additional momentum update rule is:

[0025]

[0026] in Momentum factor This is the weight increment from the previous iteration.

[0027] The statistical features mentioned include the mean of the predicted residuals within the mentioned sliding time window. and standard deviation Let the length of the mentioned sliding time window be L, then:

[0028]

[0029] The summation range is from i = t - L + 1 to r. Let k be the prediction residual at time i; take the confidence coefficient k, and the dynamic decision threshold mentioned includes constructing the upper and lower bounds of the dynamic decision threshold:

[0030] .

[0031] The system combines the cumulative occurrence of comparison results mentioned within the preset observation window to output the judgment result of the target measurement point data status, specifically including: setting the voting window length. and abnormal count threshold ( The results of a single judgment are filtered twice, and the most recent ones are statistically analyzed. The total number of suspicious anomalies within each sampling time:

[0032]

[0033] in The indicator function is used; the final state determination result is:

[0034] .

[0035] The specific steps of S4 mentioned are as follows:

[0036] S41: Perform coarse error elimination and normalization preprocessing on the real-time collected measurement point data according to the method in S1;

[0037] S42: Feed the preprocessed input data into the trained BP neural network model, and obtain the predicted value of the target measurement point through forward propagation: ;

[0038] S43: Calculate the difference between the measured value and the predicted value to obtain the prediction residual: , and take the prediction residual as the basis for state judgment.

[0039] When training the mentioned BP neural network model, the Xavier method is adopted to initialize the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the biases b1 and b2 of each layer; in the forward propagation process, the output of the hidden layer and the output of the output layer are:

[0040] .

[0041] The mentioned hidden layer activation function f(·) is a Sigmoid function:

[0042]

[0043] The mentioned training takes the mean square error as the loss function:

[0044] .

[0045] According to the mentioned judgment result of the data state of the target measurement point, generate a maintenance confirmation request for the measurement point judged to be abnormal and send it to the maintenance terminal; receive the confirmation label returned by the maintenance terminal, wherein the mentioned confirmation label includes a first label representing a confirmed abnormality and a second label representing a false alarm; count the confirmation rate Rc=N_confirmed / N_total according to a preset evaluation period, where N_confirmed is the number of times the first label is confirmed, and N_total is the total number of times an abnormality is judged; adjust the mentioned confidence coefficient k and the mentioned abnormal count threshold N based on the comparison result of Rc with the preset confirmation rate lower limit Rc1 and the preset confirmation rate upper limit Rc2, wherein k and N are increased if Rc<Rc1, k and N are decreased if Rc≥Rc2, and k and N remain unchanged if Rc1≤Rc<Rc2.

[0046] The beneficial effects of the method for detecting and judging data state based on a BP neural network model of the present invention are as follows:

[0047] First, an improved BP training algorithm based on adaptive learning rate and additional momentum term is adopted. Compared with the standard BP algorithm, it has a faster convergence speed, higher final convergence accuracy, and can effectively avoid getting stuck in shallow local minima during the training process, thus improving the generalization ability of the model.

[0048] Second, by establishing a nonlinear mapping relationship between measurement points through a BP neural network, and then using the predicted residual as the criterion for state detection, compared with the method of directly determining the threshold of the measurement point amplitude, it can make full use of the inherent correlation between measurement points and has stronger robustness to changes in working conditions and reasonable fluctuations.

[0049] Third, the judgment strategy that combines the sliding window dynamic 3σ criterion with the continuous over-limit voting mechanism can not only adapt to the working condition drift so that the threshold is updated in real time with the working condition, but also suppress the instantaneous misjudgment caused by random noise, which significantly reduces the false alarm rate and false alarm rate of state detection.

[0050] Fourth, this method does not rely on a precise physical model; it only requires historical operating data for modeling. It is simple to implement in engineering and has a wide range of applications. It can be widely used for the status detection and fault early warning of measurement point data in complex industrial equipment such as gas turbines, steam turbines, and aero engines. Attached Figure Description

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0052] Figure 1 This is a schematic diagram of the data state detection and determination method based on the BP neural network model of the present invention;

[0053] Figure 2 This is a schematic diagram of the BP neural network topology of the present invention;

[0054] Figure 3 This is a schematic diagram of the original measurement point data of the present invention;

[0055] Figure 4 This is a schematic diagram of the measurement point data after coarse error filtering and normalization according to the present invention;

[0056] Figure 5 This is a schematic diagram comparing the training error convergence curves of the improved algorithm of this invention and the standard BP algorithm;

[0057] Figure 6 This is a schematic diagram comparing the predicted and measured values ​​of the BP neural network test set of the present invention;

[0058] Figure 7 This is a time series diagram of the predicted residuals of the present invention;

[0059] Figure 8This is a schematic diagram of the sliding window dynamic 3σ judgment threshold and suspicious anomaly points of the present invention;

[0060] Figure 9 This is a schematic diagram of the measurement point data status detection and judgment results of the present invention. Detailed Implementation

[0061] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0062] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0063] like Figures 1 to 9 As shown, a data state detection and determination method based on a BP neural network model includes the following steps:

[0064] S1: Collect historical normal operation data of the system measurement points as the raw dataset. The raw data is the operation data of the industrial system. Perform coarse error removal and normalization preprocessing on the operation data.

[0065] Will Figure 3 The original data shown was obtained by removing outliers and normalizing it according to the following steps. Figure 4 The data shown includes the following steps:

[0066] S11: Calculate the mean of the data after removing trend. and standard deviation :

[0067]

[0068] S12: Calculate the residuals of the data at each time point, we have:

[0069]

[0070] S13: Connect the residuals one by one with... If a comparison is made, The corresponding data is an outlier and should be removed or replaced with the mean; otherwise, the data should be retained.

[0071] S14: Standardize the data after removing outliers to [0,1]:

[0072]

[0073] S2: The preprocessed data is divided into input variable set and output variable set, and further divided into training set and test set according to the ratio. The target measurement point to be detected is used as the output quantity, and other physically related measurement points are used as the input quantity to construct a multi-input single-output regression learning sample. Among them, the target measurement point to be detected, such as the exhaust temperature of a heavy-duty gas turbine, is used as the output quantity, and other measurement points related to the operating mechanism of the measurement point, such as IGV angle, atmospheric humidity, load, and compressor inlet temperature, are used as the input quantity.

[0074] Specifically, Figure 2 This is a schematic diagram of the BP neural network topology, which consists of three layers: an input layer, a hidden layer, and an output layer.

[0075] S21: Use the measurement point whose state needs to be detected as the output quantity. Other measuring points related to the operating mechanism of this measuring point are used as input quantities. .

[0076] S22: Standardize all quantities according to the method in S1, processing them into [0,1].

[0077] S23: Randomly select 80% of the data as the training set and 20% as the test set.

[0078] S3: Construct a BP neural network model, train the network using an improved gradient descent algorithm based on adaptive learning rate and additional momentum term, and establish a nonlinear mapping relationship model between input measurement points and target measurement points;

[0079] S31: Determine the network structure, assuming the number of nodes in the input layer is... The number of output layer nodes is The number of hidden layer nodes h is determined by an empirical formula:

[0080]

[0081] S32: Initialize network parameters. The Xavier method is used to initialize the weight matrix from the input layer to the hidden layers. Weight matrix from hidden layer to output layer and the offset of each layer , This keeps the variance of activation values ​​in each layer stable, avoiding gradient vanishing and gradient explosion.

[0082] S33: Forward propagation, the hidden layer output and the output layer output are:

[0083]

[0084] Hidden layer activation function Using the Sigmoid function:

[0085]

[0086] S34: Using mean squared error as the loss function:

[0087]

[0088] S35: Backpropagation error, calculated using the chain rule to determine the local gradients of the output and hidden layers:

[0089]

[0090]

[0091] in Represents the Hadamard product. This is the derivative of the Sigmoid function.

[0092] S36: Considering that the standard BP algorithm uses gradient descent with a fixed learning rate, it has inherent drawbacks such as difficulty in balancing convergence speed and stability, and a tendency to get trapped in shallow local minima, making it difficult to meet the needs of high-precision nonlinear modeling of industrial measurement point data. Therefore, this invention introduces both an adaptive learning rate and an additional momentum term into the standard BP weight update, forming a new weight update rule. The adaptive learning rate decays with the number of iterations.

[0093]

[0094] in The initial learning rate, As the attenuation factor, The iteration number is given; the additional momentum update rule is:

[0095]

[0096] in Momentum factor This represents the weight increment from the previous iteration. The momentum term incorporates the inertia of historical gradient directions into the current update, helping the network overcome shallow local minima and smoothing the direction in regions where gradient directions change drastically.

[0097] Adaptive learning rate is essentially a simplified step-size annealing strategy, equivalent to performing stochastic gradient descent with annealing on the error surface. This results in faster convergence in the early stages of training, more stable parameter updates in the later stages, and avoids oscillations near minima, allowing it to probabilistically approach the global optimum. The added momentum term is equivalent to introducing a first-order low-pass filter into the weight space, incorporating the inertia of historical gradient directions into the current update. This smooths the direction of noisy gradients and helps the network overcome shallow local minima. The combination of these two approaches ensures both fast convergence and enhanced ability to escape local minima. Compared to the standard backpropagation algorithm, it offers faster convergence, smaller final training error, and stronger model generalization ability. Furthermore, it is simple to implement and has low additional computational overhead, making it particularly suitable for applications with limited computing power in industrial settings.

[0098] S37: When the training error is less than a set threshold Or reach the maximum number of iterations Training is stopped when the time is right, resulting in a BP neural network model that can characterize the nonlinear mapping relationship between measurement points. .

[0099] S4: Traditional data state detection methods based on fixed thresholds directly determine the high and low limits of the amplitude at the measuring point. This approach has drawbacks such as ignoring the physical correlation between multiple measuring points, lacking the ability to distinguish reasonable amplitude fluctuations caused by changes in operating conditions, and requiring a large amount of tuning work. The real-time data to be detected is input into a trained BP neural network model for forward propagation to obtain the predicted value of the target measuring point. The predicted residual is then calculated by subtracting the predicted value from the actual measured value, and this residual is used as the basis for state determination.

[0100] Specifically, S41: Perform coarse error removal and normalization preprocessing on the real-time collected measurement point data according to the method in S1;

[0101] S42: Construct a multi-input, single-output BP neural network regression model by using the target measurement point to be detected as the output and its physically related measurement points as the input:

[0102]

[0103] S43: After preprocessing the real-time data to be detected by coarse error removal and normalization as described in S1, input it into the trained BP neural network model for forward propagation to obtain the predicted value of the target measurement point. , the measured value Compared with the predicted value By subtracting the values, we obtain the predicted residuals:

[0104]

[0105] To predict residuals As the basis for state determination, rather than based on the amplitude of the measuring point. Direct determination. Its theoretical basis stems from system identification and multivariate statistical process control (MSPC) theory: in an industrial system, the various measuring points are coupled through mass flow, energy flow, and control flow. Under normal operating conditions, their correlation is stable. When an anomaly occurs in the system, even if the amplitude of a single measuring point remains within the normal range, the correlation between measuring points will change first, manifested as an increase in the predicted residual amplitude. Therefore, residual determination has higher sensitivity and a lower false alarm rate than amplitude determination, conforming to the industrial fault evolution pattern of first exhibiting abnormal correlation and then amplitude anomalies.

[0106] Specifically, Figure 7 This is a time series plot of the predicted residuals. As can be seen from the plot, when the measuring point is in a normal state, the residual fluctuates slightly around zero, and when the measuring point is abnormal, the residual amplitude increases significantly.

[0107] S5: The predicted residuals are tested using the sliding window dynamic 3σ criterion, and the single judgment result is filtered twice by combining the continuous over-limit counting voting mechanism, and the state detection judgment result of the measurement point data is output.

[0108] Traditional 3σ criteria use fixed mean and standard deviation calculated from the entire historical sample as the judgment threshold, which cannot adapt to the slow changes in residual distribution caused by operating condition drift and equipment aging; single-out-of-limit judgment methods are very sensitive to instantaneous residual exceeding the limit caused by random noise. In some embodiments, a sliding window dynamic 3σ is combined with continuous out-of-limit counting voting to form a two-level judgment strategy.

[0109] S51: Set the length to Sliding windows (e.g.) Online calculation of residual sequences within the window Mean and standard deviation:

[0110]

[0111] S52: Obtain the confidence coefficient Construct dynamic threshold upper and lower bounds:

[0112]

[0113] The sliding window advances over time, and the decision threshold is updated in real time, ensuring that the threshold remains consistent with the local statistical characteristics of the current operating condition. This method is based on the assumptions of "online learning" and "local stationarity," meaning that the residual sequence can be considered a locally stationary process within a sufficiently short time window; the window length... These are key parameters. If the value is too small, the variance of the statistic estimate will be large and the threshold will be unstable. If the value is too large, the tracking of changes in operating conditions will lag. In engineering practice, it is recommended to determine the value based on a combination of the system sampling period and the time constant of changes in operating conditions.

[0114] S53: Residual at each sampling time If an inspection is conducted, or If it is not, it is recorded as a suspicious anomaly; otherwise, it is recorded as a normal point.

[0115] S54: Set the voting window length (For example ) and abnormal count threshold (For example The results of a single judgment are filtered twice, and the most recent ones are statistically analyzed. The total number of suspicious anomalies within each sampling time:

[0116]

[0117] The final state detection result is given by the following formula:

[0118]

[0119] in The indicator function. The continuous over-limit counting voting mechanism is essentially a type of... -select- Voting -of- Filter: Statistically, the false alarm rate of a single decision. After being filtered by this mechanism, the exponential rate is reduced to approximately It only triggers the final decision in cases of persistent anomalies, thus possessing a strong ability to suppress noisy, transient limit violations. Voting window length and abnormal count threshold These are two mutually restrictive parameters that can be jointly tuned based on the sampling period, acceptable detection delay, and false alarm rate requirements.

[0120] Specifically, Figure 8 This is a schematic diagram of the dynamic 3σ judgment threshold and suspicious anomalies in a sliding window. In the diagram, the real-time residual curve is enveloped by the upper and lower bounds of the dynamic judgment threshold. Sampling points falling outside the envelope are identified as suspicious anomalies. Figure 9 This is a schematic diagram of the final measurement point data status detection and judgment results. As can be seen from the figure, this method can accurately identify the real abnormal time periods in the measurement point data, without misjudging normal operating periods and without missing real abnormal time periods, proving the effectiveness and practicality of the method of this invention.

[0121] In real-world industrial scenarios, sensor noise levels and coupling tightness vary across different measurement points and units. Fixed values ​​for k and N cannot be optimized over long-term data accumulation, leading to high false alarm rates at some measurement points and undetected false alarms at others. To address this practical issue, some embodiments also include:

[0122] S6. Based on the judgment result of S5, generate a maintenance confirmation request for the measurement point that is judged to be abnormal, and send the identification information of the measurement point and the residual curve data corresponding to the abnormal period to the maintenance terminal.

[0123] Specifically, the maintenance confirmation request mentioned in S6 is triggered based on the final state determination result output by S5. When S5 determines that the current data state of a certain measuring point is abnormal, S6 extracts the identification information of that measuring point, such as the measuring point number or name, and the residual curve data of that measuring point within the previously mentioned sliding window length L, including the current abnormal determination time, i.e., the e-sequence calculated in S4. This information is packaged into a maintenance confirmation request and sent to the maintenance terminal via industrial Ethernet or fieldbus. The maintenance terminal can be an engineering station installed in the control room or a handheld terminal used by maintenance personnel. It is understood that S6 only generates confirmation requests for situations determined to be abnormal, and does not generate confirmation requests for situations determined to be normal, in order to reduce the confirmation burden on maintenance personnel.

[0124] S7, Receive the confirmation tag returned by the maintenance terminal. The confirmation tag mentioned includes a first tag indicating that it is indeed an anomaly and a second tag indicating that it is a false alarm. Store the confirmation tag in association with the corresponding judgment time.

[0125] Specifically, in S7, after reviewing the residual curve on the maintenance terminal, maintenance personnel select a confirmation tag based on the actual situation on site. The first tag indicates that the maintenance personnel acknowledge the anomaly determination in S5, believing that the measurement data at that moment indeed reflects the abnormal state of the equipment or process. The second tag indicates that the maintenance personnel deny the anomaly determination in S5, believing that although the measurement data at that moment deviates from the prediction model statistically, the actual equipment state is normal, and the false alarm is caused by model error, instantaneous sensor noise, or sudden changes in operating conditions. The confirmation tag is transmitted back to the data processing device that executed steps S1 to S5 via the maintenance terminal and stored in the historical database in association with the corresponding determination time. It is not specifically limited here that the confirmation tag can be collected manually or automatically by the maintenance terminal based on subsequent maintenance records, as long as the first and second tags can be distinguished.

[0126] S8. Calculate the confirmation rate according to the preset evaluation cycle. Based on the comparison results of the confirmation rate with the preset lower limit and the preset upper limit of the confirmation rate, adjust the confidence coefficient k in the mentioned sliding window dynamic judgment criterion and the abnormal count threshold N in the mentioned continuous over-limit counting voting mechanism.

[0127] Specifically, S8 adjusts the judgment parameters based on the confirmation feedback.

[0128] First, an evaluation period T_eval is set, which is measured in units of the number of judgments. The value of T_eval ranges from 30 to 200 judgments; in this embodiment, it is set to 100 judgments. The value is based on the fact that, in a heavy-duty gas turbine operating scenario with a sampling period of 1Hz, 100 judgments correspond to approximately 100 seconds of operation time. Within this time, maintenance personnel have typically completed the verification and feedback of all maintenance confirmation requests generated during this period. When the cumulative number of judgments reaches T_eval, a confirmation rate calculation is triggered, and then the counter is reset to zero to enter the next evaluation period.

[0129] The confirmation rate R_c within the current evaluation period is calculated and defined as R_c = N_confirmed / N_total, where N_confirmed is the number of times the confirmation tag is the first tag within the current evaluation period, and N_total is the total number of times S5 determines an anomaly within the current evaluation period, i.e., the total number of maintenance confirmation requests generated by S6. When N_total = 0, meaning no anomaly determination is triggered within the current evaluation period, the parameter adjustment for this evaluation period is skipped, and the existing k and N remain unchanged.

[0130] Set a lower limit for the confirmation rate R_c1 and an upper limit for the confirmation rate R_c2. The value of R_c1 ranges from 0.40 to 0.60; in this embodiment, it is set to 0.50. The rationale is that when more than half of the anomaly detections are marked as false alarms by maintenance personnel, it indicates that the current detection threshold is too sensitive and generates an unacceptable false alarm load. The value of R_c2 ranges from 0.80 to 0.95; in this embodiment, it is set to 0.85. The rationale is that when more than 85% of the anomaly detections are confirmed as real-time by maintenance personnel, it indicates that the current detection criterion is effective and can be appropriately relaxed to capture more potential early anomalies. The specific values ​​of R_c1 and R_c2 are not limited here and can be adjusted according to the tolerance requirements of different measurement points.

[0131] Parameter adjustment is performed based on the comparison results of R_c with R_c1 and R_c2. If R_c<R_c1, which means an excessively low confirmation rate indicates a high false positive rate, the confidence coefficient k and the anomaly count threshold N are increased. The adjustment amplitudes are: k_new=k_old+Δk, where Δk ranges from 0.2 to 0.5, and is set to 0.3 in this embodiment; N_new=N_old+ΔN, where ΔN ranges from 3 to 10, and is set to 5 in this embodiment. The upper limit of k is 5, and the upper limit of N is 0.9M, where M is the mentioned voting window length. Increasing k makes a residual need to deviate from the window mean by more standard deviations to be marked as a suspicious anomaly point, and increasing N requires more suspicious anomaly points to be accumulated within a longer time window to trigger the final anomaly determination. The two measures together tighten the determination criteria and reduce the false positive rate. If R_c≥R_c2, which means a relatively high confirmation rate indicates that there is room for further relaxation of the current determination criteria to capture weak anomalies, the confidence coefficient k and the anomaly count threshold N are decreased. The adjustment amplitudes are: k_new=k_old-Δk', where Δk' ranges from 0.1 to 0.3, and is set to 0.2 in this embodiment; N_new=N_old-ΔN', where ΔN' ranges from 2 to 5, and is set to 3 in this embodiment. The lower limit of k is 2, and the lower limit of N is 0.5M. Decreasing k makes a residual be marked as soon as it deviates by fewer standard deviations, and decreasing N makes a suspicious pattern with a shorter duration trigger the determination. The two measures together relax the determination criteria and make the system more sensitive to gradual early anomalies. If R_c1≤R_c<R_c2, which means the confirmation rate is in a reasonable range, the current k and N are maintained unchanged.

[0132] The adjusted k and N are applied to steps S52 and S54 in the next evaluation cycle. The adjustment process is performed online, without the need to stop the normal detection process from S1 to S5. The adjusted parameters only affect subsequent determinations, and do not affect the completed determination records. It can be understood that since the lower limit of k is 2 and the upper limit is 5, and the lower limit of N is 0.5M and the upper limit is 0.9M, the parameters always fluctuate within a reasonable range, and will not diverge to extreme values due to statistical deviations in a few evaluation cycles.

[0133] It should be noted that the adjustment of k and N in S8 relies on the statistics of the confirmation labels in S7, S7 relies on the execution result of S6, and S6 relies on the determination output of S5. This serial dependency means that the input data of S8, which comes from the confirmation labels of S7, is essentially determined by the determination result of S5, while the output of S8, the adjusted k and N, in turn changes the subsequent determination conditions of S5. This is essentially different from the solution in which the maintenance record and the statistical determination are arranged as two mutually independent functions: in the mutually independent solution, the output of the maintenance record module does not change the behavior rules of the statistical determination module, and the two terminate respectively; while in this embodiment, the output of S8 is directly injected into the determination entry of S5 and is inseparable therefrom.

[0134] For example, in a complete evaluation cycle: S5 triggers 12 anomaly determinations out of the first 100 determinations. Maintenance personnel confirm 8 of these as first-label and 4 as second-label. Therefore, R_c = 8 / 12 ≈ 0.67, which is between R_c1 (0.50) and R_c2 (0.85), keeping k and N constant. In the next evaluation cycle, if a slight change in operating conditions causes S5 to trigger 15 anomaly determinations, and maintenance personnel only confirm 5 and mark 10 as second-label, then R_c = 5 / 15 ≈ 0.33.

Claims

1. A method for data state detection and determination based on a BP neural network model, characterized in that, include: S1: Obtain historical normal operation data of the target measuring point and its associated measuring points, and train a BP neural network model using the historical normal operation data. The BP neural network model takes the operation data of the associated measuring points as input and the operation data of the target measuring point as output. S2: Input the real-time collected data from the associated measurement points into the trained BP neural network model to obtain the predicted value of the target measurement point; S3: Calculate the prediction residual based on the measured value and the predicted value of the target measurement point; S4: Based on the statistical features of the predicted residuals within the sliding time window, construct a dynamic judgment threshold that is updated in real time with the statistical features. S5: Based on the comparison result between the predicted residual and the dynamic judgment threshold, and combined with the cumulative occurrence of the comparison result within the preset observation window, output the judgment result of the target measurement point data status.

2. The method according to claim 1, characterized in that, Before being used to train the BP neural network model, the historical normal operation data also includes coarse error removal and normalization preprocessing: S11: Calculate the mean of the data after removing trend. and standard deviation : S12: Calculate the residuals of the data at each time point, we have: S13: Connect the residuals one by one with... If a comparison is made, The corresponding data is an outlier and should be removed or replaced with the mean; otherwise, the data should be retained. S14: Standardize the data after removing outliers to [0,1].

3. The method according to claim 1, characterized in that, The BP neural network model adopts a three-layer structure consisting of an input layer, a hidden layer, and an output layer. The number of hidden layer nodes h is calculated as follows: Let m be the number of input layer nodes, n be the number of output layer nodes, and a be an adjustable constant, with a value ranging from 1 to 10.

4. The method according to claim 1, characterized in that, The BP neural network model is trained using an improved gradient descent algorithm that combines an adaptive learning rate with an additional momentum term; wherein the adaptive learning rate decays with the number of iterations t. in The initial learning rate, As the attenuation factor, The iteration number is given; the additional momentum update rule is: in Momentum factor This is the weight increment from the previous iteration.

5. The method according to claim 1, characterized in that, The statistical characteristic includes the mean of the prediction residuals within the sliding time window. and standard deviation Let the length of the sliding time window be L, then: The summation range is from i = t - L + 1 to r. Let k be the prediction residual at time i; take the confidence coefficient k, and the dynamic decision threshold includes constructing the upper and lower bounds of the dynamic decision threshold: 。 6. The method according to claim 5, characterized in that, The step of combining the cumulative occurrence of the comparison results within the preset observation window to output the determination result of the target measurement point data status specifically includes: setting the voting window length. and abnormal count threshold ( The results of a single judgment are filtered twice, and the most recent ones are statistically analyzed. The total number of suspicious anomalies within each sampling time: in The indicator function is used; the final state determination result is: 。 7. The method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Perform coarse error removal and normalization preprocessing on the real-time collected measurement point data according to the method in S1; S42: The preprocessed input data is fed into the trained BP neural network model, and the predicted value of the target measurement point is obtained through forward propagation. ; S43: Measured values Compared with the predicted value Difference, calculate the predicted residuals: And to predict residuals As the basis for state determination.

8. The method according to claim 1, characterized in that, When training the BP neural network model, the Xavier method is used to initialize the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, and the biases b1 and b2 of each layer; during the forward propagation, the outputs of the hidden layer and the output layer are: 。 9. The method according to claim 8, characterized in that, The hidden layer activation function f(·) is the Sigmoid function: The training process uses mean squared error as the loss function. 。 10. The method according to any one of claims 1 to 9, characterized in that, Also includes: Based on the determination result of the target measurement point data status, a maintenance confirmation request is generated for the measurement points determined to be abnormal and sent to the maintenance terminal; The system receives confirmation tags returned by the maintenance terminal. These tags include a first tag indicating a confirmed anomaly and a second tag indicating a false alarm. The system calculates the confirmation rate Rc = N_confirmed / N_total according to a preset evaluation cycle, where N_confirmed is the number of times the first tag is confirmed, and N_total is the total number of times anomalies are determined. Based on the comparison between Rc and preset confirmation rate lower limit Rc1 and preset confirmation rate upper limit Rc2, the system adjusts the confidence coefficient k and the anomaly count threshold N. If Rc < Rc1, k and N are increased; if Rc ≥ Rc2, k and N are decreased; if Rc1 ≤ Rc < Rc2, k and N remain unchanged.

Images