Multi-sensor fusion environment anomaly detection method under communication limited condition

Abstract

The invention relates to the technical field of fire monitoring, in particular to a multi-sensor fusion environment anomaly detection method under a communication limited condition. According to the invention, multi-dimensional environmental parameters such as temperature, humidity, methane, carbon monoxide and carbon dioxide concentration are collected in real time by sensor nodes, multi-source data are fused through a Dempster-Shafer evidence theory, and sensor conflicts and uncertainty are solved; and a one-dimensional convolutional neural network 1DCNN classifier is used to identify fire abnormality based on the fused data. According to the invention, an edge computing technology is adopted, and a DSET filter and a 1DCNN classifier are deployed on a microcontroller. The dependence on cloud computing is reduced, and the real-time performance and the data security of the system are improved. Through local processing, the system can still operate efficiently even if the network is interrupted or delayed.

Description

Technical Field

[0001] The present invention relates to the technical field of fire monitoring, and particularly to a method for detecting environmental anomalies through multi-sensor fusion under communication-constrained conditions. Background Art

[0002] Traditional fire monitoring methods usually rely on a single sensor such as only temperature or gas concentration or a rule system based on fixed thresholds. These methods cannot comprehensively utilize multi-source data for cross-verification, resulting in a high false alarm rate; and they are sensitive to single-sensor failures and prone to false alarms; also, the original data needs to be transmitted to the cloud for processing, and the delay is significant in communication-constrained environments, making it difficult to respond to abnormal events in a timely manner.

[0003] With the rapid development of Internet of Things technology, sensor networks are widely used for real-time monitoring of environmental parameters to achieve fire monitoring. However, in areas such as underground mines, remote areas, and industrial plants, the environment often faces challenges of communication constraints, including problems such as insufficient network bandwidth, signal interference, communication delay, or a complete lack of communication infrastructure, which seriously affect the real-time performance and reliability of traditional centralized or cloud-computing-based monitoring systems.

[0004] To solve these problems, some multi-sensor data fusion methods have been proposed in the prior art. For example, fuzzy logic, statistical methods, and deep learning-based models are used to improve the accuracy of data processing. However, existing methods usually only target single abnormal types, lack the ability to identify complex anomalies, and do not clarify the specific determination logic of abnormal events. These methods perform poorly in the face of sensor failures or data uncertainties, especially in communication-constrained environments and cannot effectively meet real-time requirements. In addition, existing cloud-computing-based machine learning methods are difficult to implement in communication-constrained environments and have data security and delay problems. Summary of the Invention

[0005] The present invention discloses a method for detecting environmental anomalies through multi-sensor fusion under communication-constrained conditions, and the specific method is as follows:

[0006] Divide the overall monitoring area into n sub-areas, and each sub-area is equipped with a sensor node array. Each sensor node array includes k identical sensor nodes, and each sensor node is installed with m different types of sensors;

[0007] According to the data collected by the sensors, calculate the trust degree of each sensor, and determine whether the data collected by each sensor is trustworthy;

[0008] For a sensor node including untrustworthy sensors, calculate the combined trust degree of all sensors within the sensor node through the Dempster combination rule and perform weighting;

[0009] For untrustworthy sensors, a spatio-temporal fusion method is adopted to correct the data they collect;

[0010] According to the combined trustworthiness, calculate the trustworthiness adjustment factor for the data of untrustworthy sensors;

[0011] Input the data collected by all trustworthy sensors and the corrected data of untrustworthy sensors into the 1DCNN classification model to complete fire anomaly recognition; the weights of the neurons in the 1DCNN classification model are adjusted by the trustworthiness adjustment factor.

[0012] Furthermore, calculate the trustworthiness of each sensor, and the specific method is as follows:

[0013] Let the sensor node array be A 1 ~A n , and each sensor node is respectively N 1 ~N k , and the sensors installed on each sensor node are respectively G 1 ~G m ;

[0014] For any sensor node x in a sensor array ∈ [G 1 ,..., G m , define the regional environmental state discrimination framework as follows:

[0015] FOD x ={{H}, {N}}

[0016] where H represents the dangerous state and N represents the non-dangerous state;

[0017] Then its event set POW x is:

[0018]

[0019] The basic belief assignment function m x (e) maps the elements in the event set POW x to the interval [0, 1] and satisfies:

[0020]

[0021] The basic belief assignment function m x (e) of each sensor depends on the set alarm threshold Ref x and the difference Diff x between the sensor reading Obs x = Ref x - Obs x , and the specific definition is as follows:

[0022]

[0023] where m x (N), m x (H), m x (H, N) represent the degrees of belief assigned by sensor x to the non - dangerous state N, the dangerous state H, and the uncertain state, respectively; represents that the sensor reading is less than the alarm threshold Obs x <Ref x The difference factor at this time reflects the degree to which the sensor reading is lower than the alarm threshold; while represents that the sensor reading is greater than the alarm threshold Obs x >Ref x The difference factor at this time reflects the degree to which the sensor reading exceeds the alarm threshold; SF1 x and SF2 x are two scaling factors, which are used to adjust the values of the difference factors DF1 x and DF2 x respectively, linearly mapping the difference between the sensor reading Obs x and the alarm threshold Ref x into a standardized interval;

[0024] Define a degree - of - belief threshold τ. For sensor x, if its degree of belief m x (N) or m x (H) is less than the degree - of - belief threshold τ, then the data of this sensor is regarded as low - confidence data.

[0025] Furthermore, calculate the combined degree of belief, and the specific method is as follows:

[0026] For the k nodes in each sensor node array, there are k belief assignment functions for each type of data. Calculate the combined degree of belief of sensor x through the Dempster combination rule, and the specific formula is as follows:

[0027]

[0028] where is the conflict factor, indicating the degree of inconsistency between different nodes; M x {H} and M x {N} represent the comprehensive degrees of belief of the sensor array in the dangerous state and the non - dangerous state after fusing all sensor data, respectively; m1 x , m2 x ,..., mk x represent the degrees of belief of sensor x in other related nodes N 1 , N 2 ,..., Nk Trust level; weight w x = max(m x (H), m x (N)) is dynamically adjusted according to the trust level of the sensor.

[0029] Furthermore, calculate the trust level adjustment factor γ of the untrustworthy sensor data, and the specific method is as follows:

[0030] Compare M x {H} and M x {N} to determine the decision result of sensor x, and the specific formula is as follows:

[0031]

[0032] Define a trust level adjustment factor γ. When the trust level is less than the trust level threshold, that is, M x {H} < τ or M x {N} < τ, apply this factor to adjust the update amplitude of the 1DCNN neuron weight:

[0033] γ = min(max(m x (H), m x (N)), τ).

[0034] Furthermore, correct the data it uses, and the specific method is as follows:

[0035] When the data of sensor node x is determined to have a low trust level, that is, M x {N} < τ, perform time-domain supplementary value correction, space-domain supplementary value correction, spatio-temporal fusion supplementation, and credibility adjustment of the supplementary data.

[0036] Furthermore, the time-domain supplementary value, and the specific calculation formula is as follows:

[0037]

[0038] Among them, is the supplementary value based on the time domain, X t-1 and X t-2 are the two most recent valid historical data of this sensor respectively, and α is the time weight coefficient.

[0039] Furthermore, the space-domain supplementary value, and the specific method is as follows:

[0040] Calculate the distance-based weight coefficient:

[0041] S j = exp(-γd ij )

[0042] Among them, d ijis the actual physical distance between node i and adjacent node j, and γ is the distance attenuation coefficient, which is used to control the influence degree of distance on the weight;

[0043] Calculate the spatial domain supplement value based on the weight coefficient:

[0044]

[0045] Among them, is the supplement value based on the spatial domain, X j is the current data of adjacent normally working sensor nodes, S j is the weight coefficient based on distance, N i is the set of adjacent nodes of node i.

[0046] Further, spatio-temporal fusion supplement, the specific method is as follows:

[0047]

[0048] Among them, is the final supplementary data, and β is the spatio-temporal fusion weight coefficient.

[0049] Further, the credibility adjustment of the supplementary data, the specific method is as follows:

[0050] In order to reduce the influence of the supplementary data on the model decision-making, the trust degree of the supplementary data is attenuated, and the specific formula is as follows:

[0051]

[0052] Among them, is the trust degree of the supplementary data, and μ is the trust degree attenuation coefficient.

[0053] Further, the neurons of the 1DCNN classification model adjust their weights through the trust degree adjustment factor, and the specific formula is as follows:

[0054]

[0055] Among them, η is the learning rate, is the gradient of the loss function L with respect to the weight W.

[0056] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects:

[0057] 1. Effectively handle the uncertainty of sensor data and sensor fault problems through Dempster-Shafer evidence theory, evaluate the trust degree of sensor data, and filter out low-trust or faulty data.

[0058] 2. Utilize the automatic feature extraction ability of 1DCNN to efficiently classify the filtered data, which is applicable to the processing of time series data and can accurately detect abnormal events in the environment.

[0059] 3. Deploy the DSET, 1DCNN model, and spatio-temporal fusion-based data supplementation algorithm on edge devices to reduce dependence on cloud computing and ensure real-time performance and data security in communication-constrained environments.

[0060] 4. For low-trust sensor data, propose a spatio-temporal fusion-based data supplementation strategy to ensure the consistency of the model input dimension and improve the reliability of the system in case of sensor failures.

[0061] 5. The present invention adopts edge computing technology to deploy the DSET filter and 1DCNN classifier on a microcontroller. This not only reduces dependence on cloud computing but also improves the real-time performance and data security of the system. Through local processing, the system can still operate efficiently even in case of network interruption or latency.

[0062] Other advantages, objectives, and features of the present invention will be elaborated to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The drawings of the present invention are described as follows.

[0064] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0065] Figure 2 It is a schematic diagram of the 1DCNN model structure.

[0066] Figure 3 It is a schematic diagram of the hardware structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] The present invention will be further described below in conjunction with the drawings and embodiments.

[0068] A method for abnormal detection in a multi-sensor fusion environment under communication-constrained conditions, as Figure 1 shown, the specific steps are as follows:

[0069] S1. Install sensors

[0070] To achieve fire abnormal detection under communication-constrained conditions, multi-sensor data monitoring is adopted in the area to be monitored, and the installation structure is as Figure 3 shown.

[0071] The overall monitoring area is divided into n sub - areas, and each sub - area is equipped with a sensor node array (A 1 ~A n ). Each sensor node array contains k identical sensor nodes (N 1 ~N k ), and m different types of sensors (G 1 ~G m ) are installed on each node to monitor the changes in the local environment in real - time. By setting multiple sensor nodes in each area, the system can not only improve the accuracy of data collection but also effectively handle single - sensor failures or data uncertainty problems.

[0072] S2. Use the DSET filter to judge the trustworthiness

[0073] In this step, DSET evaluates the trustworthiness of each sensor observation by defining the basic belief assignment and uses the Dempster combination rule to make a comprehensive decision on the data of multiple sensors. The specific steps are as follows:

[0074] S21. Basic belief assignment. For a sensor node x∈[G 1 ,...,G m in any sensor array, first define a frame of discernment for the regional environmental state:

[0075] FOD x ={{H},{N}}

[0076] where H represents the dangerous state and N represents the non - dangerous state;

[0077] Then its power set POW x is:

[0078]

[0079] The basic belief assignment function m x (e) maps the elements in the power set POW x to the interval [0,1] and satisfies:

[0080]

[0081] The basic belief assignment function m x (e) of each sensor depends on the set alarm threshold Ref x and the difference Diff x between the sensor reading Obs x =Ref x -Obs x , and the specific definition is as follows:

[0082]

[0083]

[0084] where m x (N), m x (H), m x (H, N) represent the degrees of belief that the sensor x is assigned to the non - dangerous state N, the dangerous state H, and the uncertain state, respectively; represents that the sensor reading is less than the alarm threshold Obs x <Ref x The difference factor at this time reflects the degree to which the sensor reading is lower than the alarm threshold. then represents that the sensor reading is greater than the alarm threshold Obs x >Ref x The difference factor at this time reflects the degree to which the sensor reading exceeds the alarm threshold; SF1 x and SF2 x are two scaling factors used to adjust the values of the difference factors DF1 x and DF2 x respectively, linearly mapping the difference between the sensor reading Obs x and the alarm threshold Ref x into a standardized interval for subsequent belief assignment calculations.

[0085] Define a degree - of - belief threshold τ. For the sensor x, if its degree of belief m x (N) or m x (H) is less than the degree - of - belief threshold τ, then the data of this sensor is regarded as low - confidence data and needs special processing, which helps to screen out high - confidence data and reduce the influence of noise and outliers.

[0086] S21. DS combination rule based on degree - of - belief weighting. For the k nodes in each sensor node array, there are k belief assignment functions for each type of data. Calculate the combined degree of belief of the sensor x through the Dempster combination rule and introduce a weighting mechanism to make the influence of high - confidence sensors on the final decision greater:

[0087]

[0088] where is the conflict factor, representing the degree of inconsistency between different nodes; M x {H} and M x {N} represent the comprehensive degrees of belief of the sensor array for the dangerous state and the non - dangerous state after fusing all sensor data, respectively; m1 x , m2 x,..., mk x respectively represent the trust degrees of sensor x at other relevant nodes N 1 , N 2 ,..., N k ; the weight w x = max(m x (H), m x (N)) is dynamically adjusted according to the trust degree of the sensor.

[0089] S22. Fault node filtering and trust degree adjustment. The combined trust degree is used to determine the combined decision result, and a trust degree adjustment mechanism is introduced. For sensors with low trust degrees, their influence on the subsequent 1DCNN training is reduced.

[0090] First, compare M x {H} and M x {N} to determine the decision result of sensor x:

[0091]

[0092] Define a trust degree adjustment factor γ. When the trust degree is less than the trust degree threshold, that is, M x {H} < τ or M x {N} < τ, apply this factor to adjust the update amplitude of the 1DCNN neuron weights:

[0093] γ = min(max(m x (H), m x (N)), τ)

[0094] For the data of sensors with low trust degrees, in the subsequent 1DCNN training, their influence will be significantly weakened, thus avoiding the negative impact of low-quality data on model training.

[0095] Based on the above decision result, the nodes where sensor x does not obey the decision are identified as fault nodes, and the wrong observations of the fault nodes will be removed.

[0096] S3. Correct the data with spatio-temporal fusion.

[0097] When the data of sensor node x is determined to have a low trust degree (M x {H} < τ or M x {N} < τ), the following method is used to supplement the missing data:

[0098] S31. Calculation of the supplementary value in the time domain.

[0099]

[0100] Among them, is the supplementary value based on the time domain, Xt-1 and X t-2 are the two most recent valid historical data of the sensor, and α is the time weight coefficient, where 0 < α < 1.

[0101] S32. Spatial domain supplementary value calculation.

[0102] Calculate the distance-based weight coefficient:

[0103] S j = exp(-γd ij )

[0104] where d ij is the actual physical distance between node i and adjacent node j, and γ is the distance attenuation coefficient, which is used to control the influence degree of distance on the weight. The farther the distance, the more the weight decays exponentially.

[0105] Calculate the spatial domain supplementary value based on the weight coefficient:

[0106]

[0107] where is the supplementary value based on the spatial domain, X j is the current data of adjacent normally working sensor nodes, S j is the distance-based weight coefficient, and N i is the set of adjacent nodes of node i.

[0108] S33. Spatiotemporal fusion supplement.

[0109]

[0110] where is the final supplementary data, and β is the spatiotemporal fusion weight coefficient, where 0 < β < 1.

[0111] S34. Credibility adjustment of supplementary data.

[0112] To reduce the impact of supplementary data on model decision-making, the trust degree of supplementary data is attenuated:

[0113]

[0114] where is the trust degree of supplementary data, and μ is the trust degree attenuation coefficient, where 0 < μ < 1.

[0115] S4. The 1DCNN classification model completes fire anomaly recognition.

[0116] 1DCNN is a deep learning model specifically designed for processing sequence data, suitable for time series data and signal processing tasks, such as Figure 2As shown in the figure. The 1D CNN is used to classify the healthy node data after DSET filtering to identify abnormal events in the environment. The 1D CNN model adopts a multi-input and multi-output architecture, including a 1D CNN layer, a max pooling layer, a Dropout layer, and a fully connected layer.

[0117] Each input sample in the input layer of the model contains m×k data points within T seconds, that is, m sensors for each node. These data points are organized into a matrix with the shape of T×m×k to represent the m sensor readings per second within T seconds.

[0118] The 1D CNN layer uses a one-dimensional convolutional kernel to automatically extract features. Assume the convolutional kernel size is k f , and the stride is s. Then the l-th layer convolutional operation can be expressed as:

[0119] O l =σ(W l *X + b l )

[0120] where O l is the output feature map of the l-th layer, W l is the convolutional kernel weight matrix of the l-th layer, b l is the bias term, * represents the convolutional operation, and σ is the activation function. Multiple convolutional layers can be stacked together to capture more complex features.

[0121] After each convolutional layer, a max pooling layer is added to reduce the data dimension and retain important features. Assume the pooling window size is k p , and the stride is s p . Then the max pooling operation can be expressed as:

[0122] P l =MaxPool(O l , k p , s p )

[0123] Finally, based on the extracted features, a fully connected layer is used for classification decision-making. Assume the weight matrix of the fully connected layer is W f , and the bias term is b f . Then the operation of the fully connected layer can be expressed as:

[0124] Y=softmax(W f ·D l +b f )

[0125] The trust adjustment factor γ is used to dynamically adjust the amplitude of gradient update during the backpropagation process, thereby reducing the negative impact of low-quality data on model training:

[0126]

[0127] Among them, η is the learning rate, is the gradient of the loss function L with respect to the weight W.

[0128] In this way, the confidence adjustment factor γ dynamically adjusts the amplitude of the gradient update, ensuring that high-quality data contributes more to model training while the impact of low-quality data is weakened, thereby improving the robustness and accuracy of the model.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the specific implementation manners of the present invention can still be modified or equivalently replaced, and any modification or equivalent replacement without departing from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A multi-sensor fusion environment anomaly detection method under communication restricted conditions, characterized in that: The specific method is as follows: The overall monitoring area is divided into n sub-areas, each sub-area is equipped with a sensor node array, each sensor node array includes k identical sensor nodes, and each sensor node is equipped with m sensors of different types; Based on the data collected by the sensor, the trustworthiness of each sensor is calculated to determine whether the data collected by each sensor is trustworthy; For sensor nodes that include untrustworthy sensors, the joint trust of all sensors in the sensor node is calculated and weighted by the Dempster combination rule; For untrustworthy sensors, the time-space fusion method is used to correct the data they use; Calculate the trust adjustment factor of the untrustworthy sensor data according to the joint trust; All data collected by trustworthy sensors and corrected data from untrustworthy sensors are input into the 1DCNN classification model to complete fire anomaly identification; the neurons of the 1DCNN classification model adjust their weights through the trust adjustment factor.

2. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 1, characterized in that: Calculate the trust of each sensor as follows: Let the sensor node array A1~A n , each sensor node is N1~N k , the sensors installed on each sensor node are G1~G m ; For any sensor node x∈[G1,...,G m ], the regional environmental state resolution framework is defined as follows: BE x ={{H},{N}} Among them, H represents a dangerous state, and N represents a non-dangerous state; Then its event set POW x for: By the basic belief distribution function m x (e) The event set POW x The elements in are mapped to the interval [0,1] and satisfy: The basic belief distribution function m of each sensor x (e) Depends on the set alarm threshold Ref x With sensor readings Obs x Diff x =Ref x -Obs x , specifically defined as follows: Among them, m x (N),m x (H),m x (H, N) represents the trust of sensor x assigned to non-dangerous state N, dangerous state H and uncertain state respectively; Indicates that the sensor reading is less than the alarm threshold Obs x <Ref x The difference factor at , reflects the degree to which the sensor reading falls below the alarm threshold; It means that the sensor reading is greater than the alarm threshold Obs x >Ref x The difference factor at the time of the alarm reflects the degree to which the sensor reading exceeds the alarm threshold; SF1 x and SF2 x are two scaling factors, used to adjust the difference factor DF1 x and DF2 x The value of the sensor reading Obs x With alarm threshold Ref x The difference between them is linearly mapped into a standardized interval; Define a trust threshold τ, for sensor x, if its trust m x (N) or m x (H) is less than the trust threshold τ, then the data of the sensor is regarded as low-trust data.

3. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 2, characterized in that: The joint trust is calculated as follows: For each k node in the sensor node array, each type of data has k belief distribution functions, and the joint trust of the sensor x is calculated by the Dempster combination rule. The specific formula is as follows: in, is the conflict factor, indicating the degree of inconsistency between different nodes; M x {H} and M x {N} represents the comprehensive trust of the sensor array for dangerous and non-dangerous states after integrating all sensor data; m1 x ,m2 x ,...,mk x Respectively represent sensor x in other related nodes N1, N2, ..., N k The trust degree of x =max(m x (H),m x (N)) is dynamically adjusted according to the trustworthiness of the sensor.

4. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 4, characterized in that: The trust adjustment factor γ of the untrustworthy sensor data is calculated as follows: Compare M x {H} and M x {N} determines the decision result of sensor x. The specific formula is as follows: Define a trust adjustment factor γ. When the trust is less than the trust threshold, M x {H}<τ or M x {N}<τ, this factor is used to adjust the update amplitude of 1DCNN neuron weights: γ=min(max(m x (H),m x (N)),τ).

5. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 4, characterized in that: Correct the data used. The specific method is as follows: When the data of sensor node x is judged to be of low trust, that is, M x When {N}<τ, the time domain supplementary value correction, space domain supplementary value correction, space-time fusion supplement and credibility adjustment of the supplementary data are performed.

6. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 5, characterized in that: Time domain supplementary value, the specific calculation formula is as follows: in, is a supplementary value based on the time domain, X t-1 and X t-2 are the two most recent valid historical data of the sensor, and α is the time weight coefficient.

7. The method for detecting anomalies in a multi-sensor fusion environment under communication-restricted conditions as claimed in claim 5, characterized in that: Spatial domain supplementary value, the specific method is as follows: Calculate the distance-based weight coefficient: S j =exp(-γd ij ) Among them, d ij is the actual physical distance between node i and its adjacent node j, and γ is the distance attenuation coefficient, which is used to control the influence of distance on weight; Calculate the spatial domain supplementary value based on the weight coefficient: in, is a supplementary value based on the spatial domain, X j is the current data of the adjacent sensor nodes that are working normally, S j is the weight coefficient based on distance, N i is the set of adjacent nodes of node i.

8. The method for detecting anomalies in a multi-sensor fusion environment under communication-constrained conditions as claimed in claim 5, characterized in that: Time and space fusion supplement, the specific method is as follows: in, is the final supplementary data, and β is the spatiotemporal fusion weight coefficient.

9. The method for detecting anomalies in a multi-sensor fusion environment under communication-constrained conditions as claimed in claim 5, characterized in that: The credibility adjustment of supplementary data is as follows: In order to reduce the impact of supplementary data on model decision-making, the trust in the supplementary data is attenuated. The specific formula is as follows: in, is the trust value of the supplementary data, and μ is the trust attenuation coefficient.

10. The multi-sensor fusion environment anomaly detection method under communication restricted conditions as claimed in claim 5, characterized in that: The neurons of the 1DCNN classification model adjust their weights through the trust adjustment factor. The specific formula is as follows: Where η is the learning rate, is the gradient of the loss function L with respect to the weight W.

Images